diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000006/group_0/index_tensor([3113, 3113])_group_0_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000006/group_0/index_tensor([3113, 3113])_group_0_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..699abdb213155af187133dcddc8cd456cbadde96 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000006/group_0/index_tensor([3113, 3113])_group_0_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "shot, shot from a slightly upward angle, captures an extreme close-up of the walls of a massive, three-meter-long bronze coffin. The ancient bronze's surface is covered in varying shades of patina and the marks of age. The camera slowly pulls back and pans upward, gradually revealing an intricately carved nine-headed serpent encircling a sun disc. The serpent's scales, layered and lifelike, pulsate ominously with a dark, blood-red glow. These eyes, set in unpolished dark red gemstones, provide the only source of light in this shrouded, gloomy underground chamber, casting eerie, swaying shadows on the surrounding stone walls. As the camera continues to pull out, it finally stabilizes in a medium shot, revealing the coffin in its entirety and a young man in a simple black robe lying on its lid. At that moment, the long eyelashes on the boy's delicate yet pale face began to tremble violently, as if struggling to escape from a bottomless nightmare. Then, his eyes fluttered open, revealing utter terror and confusion. Almost instantly, through the black robes on his chest, a ball of blue flame pulsed silently and violently from the area where his heart lay. This mysterious, cold light illuminated the outline of his chest clearly. Then, the flame rapidly contracted and condensed, eventually transforming into a palm-sized miniature figure composed entirely of blue flame. This flaming figure, arms crossed, calmly \"sat\" on the boy's left rib cage. The entire scene, interwoven with blood-red and dark blue light, created an atmosphere of ultimate horror and fantasy.", + "seed": 32440, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000006/group_0/index_3113_seed_32440_dp_0_rank_0_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 0, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.05698383674025535, + "vlm_logprob": -0.05698383674025535, + "vlm_yes_no_margin": 3.275, + "vlm_token_credit": -0.05698383674025535, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a shrouded, gloomy underground chamber with stone walls?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a massive three-meter-long bronze coffin with patina and age marks visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a nine-headed serpent encircling a sun disc appear on the coffin with scales pulsating in a dark blood-red glow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a young man in a simple black robe lying on the coffin lid, whose eyes open to reveal terror and a blue flame forming a miniature figure on his chest?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start with an extreme close-up from a slightly upward angle, then slowly pull back and pan upward to stabilize in a medium shot?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.16413405537605286 + ], + "opposite_logprob_tokens": [ + -1.9141340255737305 + ], + "logprob_reward": -0.16413405537605286, + "yes_no_margin_reward": 1.75, + "token_credit_reward": -0.16413405537605286, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2479586899280548, + 0.0847151130437851, + 0.15199397504329681, + 0.17234349250793457, + 0.04462919756770134, + 0.05589982867240906, + 0.06636454910039902, + 0.045525602996349335, + 0.08031091094017029, + 0.050258662551641464 + ], + "credit_map": [ + 0.021359026432037354, + 0.024695169180631638, + 0.025282906368374825, + 0.03908742219209671, + 0.028684593737125397, + 0.027448974549770355, + 0.042804256081581116, + 0.038596343249082565, + 0.004657309502363205, + 0.009248688817024231, + 0.009467615745961666, + 0.012025842443108559, + 0.01092117466032505, + 0.01789868250489235, + 0.00954330526292324, + 0.010952495038509369, + 0.010785352438688278, + 0.01505284383893013, + 0.021363699808716774, + 0.029183082282543182, + 0.01664266176521778, + 0.016229767352342606, + 0.023941906169056892, + 0.01879465952515602, + 0.014812792651355267, + 0.021640408784151077, + 0.03260880708694458, + 0.028147855773568153, + 0.013620232231914997, + 0.017276247963309288, + 0.01837548054754734, + 0.025861678645014763, + 0.004932998679578304, + 0.004941936116665602, + 0.0047035603784024715, + 0.006004405207931995, + 0.004927494563162327, + 0.00538376159965992, + 0.007626325357705355, + 0.006108715198934078, + 0.005181531887501478, + 0.007357817143201828, + 0.00666943471878767, + 0.007503902539610863, + 0.006286549847573042, + 0.008381654508411884, + 0.0075100562535226345, + 0.00700888317078352, + 0.01010708138346672, + 0.005684368312358856, + 0.008247699588537216, + 0.007539332378655672, + 0.006927164271473885, + 0.010591957718133926, + 0.008298257365822792, + 0.008968683890998363, + 0.0044919452629983425, + 0.00524303549900651, + 0.007019971497356892, + 0.00417290348559618, + 0.005390588194131851, + 0.0066407653503119946, + 0.005312307737767696, + 0.007254088297486305, + 0.00798104703426361, + 0.007620568852871656, + 0.015082057565450668, + 0.00819462165236473, + 0.010632121004164219, + 0.01039830595254898, + 0.0067341867834329605, + 0.013668001629412174, + 0.005883499979972839, + 0.0050530871376395226, + 0.009632276371121407, + 0.004629254806786776, + 0.004330979660153389, + 0.007938427850604057, + 0.004483131226152182, + 0.008308006450533867 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 2 + ], + "selected_reward": -0.16413405537605286 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a shrouded, gloomy underground chamber with stone walls?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a massive three-meter-long bronze coffin with patina and age marks visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a nine-headed serpent encircling a sun disc appear on the coffin with scales pulsating in a dark blood-red glow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a young man in a simple black robe lying on the coffin lid, whose eyes open to reveal terror and a blue flame forming a miniature figure on his chest?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start with an extreme close-up from a slightly upward angle, then slowly pull back and pan upward to stabilize in a medium shot?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01421845331788063 + ], + "opposite_logprob_tokens": [ + -4.264218330383301 + ], + "logprob_reward": -0.01421845331788063, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.01421845331788063, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2257089763879776, + 0.09332840889692307, + 0.1758851706981659, + 0.14636191725730896, + 0.04878934100270271, + 0.1294247955083847, + 0.05517563968896866, + 0.04044871777296066, + 0.04782388359308243, + 0.037053171545267105 + ], + "credit_map": [ + 0.015731824561953545, + 0.033098164945840836, + 0.018404576927423477, + 0.043657295405864716, + 0.027599945664405823, + 0.023242011666297913, + 0.03755951300263405, + 0.026415646076202393, + 0.004929024260491133, + 0.012124115601181984, + 0.013465243391692638, + 0.012344344519078732, + 0.012744246050715446, + 0.014806506223976612, + 0.008871996775269508, + 0.014042936265468597, + 0.013029586523771286, + 0.014330398291349411, + 0.03236708417534828, + 0.03448102995753288, + 0.017151283100247383, + 0.013427062891423702, + 0.02951228804886341, + 0.02158643864095211, + 0.011142552830278873, + 0.015187112614512444, + 0.023789234459400177, + 0.028111277148127556, + 0.011863673105835915, + 0.020526114851236343, + 0.0163351409137249, + 0.019406814128160477, + 0.004381297621876001, + 0.005937689915299416, + 0.004625966772437096, + 0.005794735625386238, + 0.004622474778443575, + 0.006354267708957195, + 0.009695926681160927, + 0.0073769791051745415, + 0.011135561391711235, + 0.009997118264436722, + 0.016069039702415466, + 0.01656561717391014, + 0.010629640892148018, + 0.01984885148704052, + 0.015864690765738487, + 0.02931428700685501, + 0.0072458949871361256, + 0.005011534318327904, + 0.00503159873187542, + 0.00650454219430685, + 0.006368390284478664, + 0.007938309572637081, + 0.009299809113144875, + 0.007775561884045601, + 0.0029258313588798046, + 0.006360854022204876, + 0.0046878112480044365, + 0.00634576054289937, + 0.004676011856645346, + 0.005408794153481722, + 0.004909804556518793, + 0.005133849568665028, + 0.004119625315070152, + 0.004364461172372103, + 0.009511647745966911, + 0.004319887142628431, + 0.008196922019124031, + 0.006982324179261923, + 0.0048546153120696545, + 0.005474401172250509, + 0.0034998063929378986, + 0.004349042195826769, + 0.007476279512047768, + 0.0036617598962038755, + 0.004393713548779488, + 0.005481519736349583, + 0.0036974227987229824, + 0.004493627231568098 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 3 + ], + "selected_reward": -0.01421845331788063 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a shrouded, gloomy underground chamber with stone walls?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a massive three-meter-long bronze coffin with patina and age marks visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a nine-headed serpent encircling a sun disc appear on the coffin with scales pulsating in a dark blood-red glow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a young man in a simple black robe lying on the coffin lid, whose eyes open to reveal terror and a blue flame forming a miniature figure on his chest?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start with an extreme close-up from a slightly upward angle, then slowly pull back and pan upward to stabilize in a medium shot?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.026360128074884415 + ], + "opposite_logprob_tokens": [ + -3.651360034942627 + ], + "logprob_reward": -0.026360128074884415, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.026360128074884415, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.19801947474479675, + 0.12274807691574097, + 0.1591079831123352, + 0.08164574950933456, + 0.03381817787885666, + 0.1705612987279892, + 0.06070896238088608, + 0.06037318333983421, + 0.08362408727407455, + 0.029393035918474197 + ], + "credit_map": [ + 0.01574629172682762, + 0.023119663819670677, + 0.019612936303019524, + 0.024406859651207924, + 0.02340072952210903, + 0.025664938613772392, + 0.02791459485888481, + 0.03815343976020813, + 0.00793753657490015, + 0.014446791261434555, + 0.02123728021979332, + 0.017508918419480324, + 0.014889853075146675, + 0.014235427603125572, + 0.013642650097608566, + 0.018849611282348633, + 0.00787336751818657, + 0.010942413471639156, + 0.025903744623064995, + 0.03325033187866211, + 0.016100965440273285, + 0.014255317859351635, + 0.03115641325712204, + 0.019625429064035416, + 0.007170860655605793, + 0.010108313523232937, + 0.008573680184781551, + 0.011811777949333191, + 0.006708297412842512, + 0.01132955215871334, + 0.009132135659456253, + 0.016811123117804527, + 0.003237226279452443, + 0.0040566399693489075, + 0.003677317639812827, + 0.004509286489337683, + 0.00453681917861104, + 0.0041972012259066105, + 0.0053529138676822186, + 0.004250768572092056, + 0.014420608058571815, + 0.012920734472572803, + 0.01863858476281166, + 0.02208676189184189, + 0.014209156855940819, + 0.026475075632333755, + 0.021335313096642494, + 0.040475036948919296, + 0.004153987392783165, + 0.005787503439933062, + 0.005218456033617258, + 0.005451610311865807, + 0.006972027942538261, + 0.009090196341276169, + 0.015351254492998123, + 0.008683924563229084, + 0.006087192799896002, + 0.007844596169888973, + 0.011121024377644062, + 0.005628035869449377, + 0.007386037614196539, + 0.00962954480201006, + 0.0058597116731107235, + 0.006817033980041742, + 0.007643257733434439, + 0.008847502060234547, + 0.018630167469382286, + 0.007393091917037964, + 0.010430736467242241, + 0.013938906602561474, + 0.006555201485753059, + 0.010185216553509235, + 0.0027896538376808167, + 0.003926737699657679, + 0.004481171257793903, + 0.0036073485389351845, + 0.004204180091619492, + 0.0037572591099888086, + 0.0034191294107586145, + 0.0032075541093945503 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 5, + 2 + ], + "selected_reward": -0.026360128074884415 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a shrouded, gloomy underground chamber with stone walls?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a massive three-meter-long bronze coffin with patina and age marks visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a nine-headed serpent encircling a sun disc appear on the coffin with scales pulsating in a dark blood-red glow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a young man in a simple black robe lying on the coffin lid, whose eyes open to reveal terror and a blue flame forming a miniature figure on his chest?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start with an extreme close-up from a slightly upward angle, then slowly pull back and pan upward to stabilize in a medium shot?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.06203282251954079 + ], + "opposite_logprob_tokens": [ + -2.81203293800354 + ], + "logprob_reward": -0.06203282251954079, + "yes_no_margin_reward": 2.75, + "token_credit_reward": -0.06203282251954079, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14723694324493408, + 0.14034676551818848, + 0.11048281937837601, + 0.1129879355430603, + 0.16246527433395386, + 0.09484736621379852, + 0.07995157688856125, + 0.05165376886725426, + 0.06750813871622086, + 0.032519396394491196 + ], + "credit_map": [ + 0.013390562497079372, + 0.01590067148208618, + 0.014692926779389381, + 0.020942723378539085, + 0.019438503310084343, + 0.0172264464199543, + 0.01799279823899269, + 0.0276523157954216, + 0.010121236555278301, + 0.010267534293234348, + 0.018962640315294266, + 0.020491257309913635, + 0.017904367297887802, + 0.022071316838264465, + 0.020089780911803246, + 0.020438628271222115, + 0.008947025053203106, + 0.00904039479792118, + 0.015120060183107853, + 0.024905269965529442, + 0.009789947420358658, + 0.009627974592149258, + 0.019921332597732544, + 0.01313081569969654, + 0.005892343819141388, + 0.007091626524925232, + 0.013019190169870853, + 0.01475504506379366, + 0.007745688781142235, + 0.02791735529899597, + 0.012221304699778557, + 0.024345381185412407, + 0.008470501750707626, + 0.011730516329407692, + 0.009127107448875904, + 0.019558271393179893, + 0.016209915280342102, + 0.02197590097784996, + 0.03472506254911423, + 0.040667999535799026, + 0.007905990816652775, + 0.00935368798673153, + 0.010776061564683914, + 0.012648622505366802, + 0.008222315460443497, + 0.014298595488071442, + 0.011451652273535728, + 0.020190440118312836, + 0.004493407905101776, + 0.006616643164306879, + 0.0055697523057460785, + 0.007415283937007189, + 0.009347336366772652, + 0.012006798759102821, + 0.02004166506230831, + 0.01446069311350584, + 0.006003805436193943, + 0.006970065645873547, + 0.008451724424958229, + 0.004215320106595755, + 0.005644220858812332, + 0.008517814800143242, + 0.004996992647647858, + 0.006853824015706778, + 0.005148631986230612, + 0.0064020911231637, + 0.01484326645731926, + 0.005320712458342314, + 0.01070774532854557, + 0.009883998893201351, + 0.00802578218281269, + 0.0071759093552827835, + 0.0023391556460410357, + 0.0043601710349321365, + 0.00463667931035161, + 0.002762545133009553, + 0.005876379553228617, + 0.004282971378415823, + 0.004946137312799692, + 0.003315360751003027 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 0, + 1 + ], + "selected_reward": -0.06203282251954079 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a shrouded, gloomy underground chamber with stone walls?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a massive three-meter-long bronze coffin with patina and age marks visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a nine-headed serpent encircling a sun disc appear on the coffin with scales pulsating in a dark blood-red glow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a young man in a simple black robe lying on the coffin lid, whose eyes open to reveal terror and a blue flame forming a miniature figure on his chest?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start with an extreme close-up from a slightly upward angle, then slowly pull back and pan upward to stabilize in a medium shot?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01817372441291809 + ], + "opposite_logprob_tokens": [ + -4.018173694610596 + ], + "logprob_reward": -0.01817372441291809, + "yes_no_margin_reward": 4.0, + "token_credit_reward": -0.01817372441291809, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1369323581457138, + 0.1607331782579422, + 0.08785786479711533, + 0.16080491244792938, + 0.11596302688121796, + 0.15260471403598785, + 0.05790219455957413, + 0.03546345233917236, + 0.06507238000631332, + 0.02666594088077545 + ], + "credit_map": [ + 0.01313120499253273, + 0.013957664370536804, + 0.01565733179450035, + 0.01750021055340767, + 0.018710114061832428, + 0.015632227063179016, + 0.017976952716708183, + 0.024366650730371475, + 0.009276323951780796, + 0.02091570943593979, + 0.027192359790205956, + 0.01942097395658493, + 0.019993852823972702, + 0.02164355292916298, + 0.01670822873711586, + 0.02558218501508236, + 0.004947753623127937, + 0.007325278129428625, + 0.012063871137797832, + 0.01756303571164608, + 0.01006276998668909, + 0.008518231101334095, + 0.015963943675160408, + 0.011412976309657097, + 0.012809797190129757, + 0.012192669324576855, + 0.023036694154143333, + 0.01706520840525627, + 0.011760484427213669, + 0.04170754551887512, + 0.013731002807617188, + 0.02850150503218174, + 0.01077782828360796, + 0.013631213456392288, + 0.014230392873287201, + 0.014930633828043938, + 0.014547226019203663, + 0.014053386636078358, + 0.019388779997825623, + 0.014403571374714375, + 0.012813389301300049, + 0.011469414457678795, + 0.010260332375764847, + 0.01597275398671627, + 0.014781007543206215, + 0.02498507685959339, + 0.023819856345653534, + 0.03850288689136505, + 0.005697369109839201, + 0.005052936729043722, + 0.0055081588216125965, + 0.006461611017584801, + 0.006935645826160908, + 0.007490186952054501, + 0.010595318861305714, + 0.010160970501601696, + 0.004088130313903093, + 0.005090814083814621, + 0.0038718159776180983, + 0.003630948718637228, + 0.0041887094266712666, + 0.00559892738237977, + 0.003996965941041708, + 0.0049971407279372215, + 0.005350839346647263, + 0.006700617261230946, + 0.015111980028450489, + 0.006024002097547054, + 0.007120294496417046, + 0.011150444857776165, + 0.007356173824518919, + 0.006258029956370592, + 0.0026150436606258154, + 0.0030392163898795843, + 0.0052016680128872395, + 0.0027640361804515123, + 0.0035669647622853518, + 0.003636558074504137, + 0.0026814090088009834, + 0.0031610445585101843 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 1, + 5 + ], + "selected_reward": -0.01817372441291809 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "shot, shot from a slightly upward angle, captures an extreme close-up of the walls of a massive, three-meter-long bronze coffin. The ancient bronze's surface is covered in varying shades of patina and the marks of age. The camera slowly pulls back and pans upward, gradually revealing an intricately carved nine-headed serpent encircling a sun disc. The serpent's scales, layered and lifelike, pulsate ominously with a dark, blood-red glow. These eyes, set in unpolished dark red gemstones, provide the only source of light in this shrouded, gloomy underground chamber, casting eerie, swaying shadows on the surrounding stone walls. As the camera continues to pull out, it finally stabilizes in a medium shot, revealing the coffin in its entirety and a young man in a simple black robe lying on its lid. At that moment, the long eyelashes on the boy's delicate yet pale face began to tremble violently, as if struggling to escape from a bottomless nightmare. Then, his eyes fluttered open, revealing utter terror and confusion. Almost instantly, through the black robes on his chest, a ball of blue flame pulsed silently and violently from the area where his heart lay. This mysterious, cold light illuminated the outline of his chest clearly. Then, the flame rapidly contracted and condensed, eventually transforming into a palm-sized miniature figure composed entirely of blue flame. This flaming figure, arms crossed, calmly \"sat\" on the boy's left rib cage. The entire scene, interwoven with blood-red and dark blue light, created an atmosphere of ultimate horror and fantasy.", + "seed": 32441, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000006/group_0/index_3113_seed_32441_dp_0_rank_0_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 0, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.053478649631142613, + "vlm_logprob": -0.053478649631142613, + "vlm_yes_no_margin": 3.125, + "vlm_token_credit": -0.053478649631142613, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a shrouded, gloomy underground chamber with stone walls?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a massive three-meter-long bronze coffin with patina and age marks visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a nine-headed serpent encircling a sun disc appear on the coffin with scales pulsating in a dark blood-red glow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a young man in a simple black robe lying on the coffin lid, whose eyes open to reveal terror and a blue flame forming a miniature figure on his chest?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start with an extreme close-up from a slightly upward angle, then slowly pull back and pan upward to stabilize in a medium shot?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.1156684160232544 + ], + "opposite_logprob_tokens": [ + -2.240668296813965 + ], + "logprob_reward": -0.1156684160232544, + "yes_no_margin_reward": 2.125, + "token_credit_reward": -0.1156684160232544, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09859172254800797, + 0.04528028890490532, + 0.12279406189918518, + 0.06227048113942146, + 0.07265513390302658, + 0.10848365724086761, + 0.07194645702838898, + 0.1320492923259735, + 0.14430075883865356, + 0.14162816107273102 + ], + "credit_map": [ + 0.008594046346843243, + 0.00942866038531065, + 0.00877397507429123, + 0.013144322670996189, + 0.01593952253460884, + 0.013352668844163418, + 0.009288178756833076, + 0.020070340484380722, + 0.004099387209862471, + 0.006092830561101437, + 0.005393600091338158, + 0.005367618519812822, + 0.005921514239162207, + 0.006192604545503855, + 0.006866796407848597, + 0.005345936398953199, + 0.017983151599764824, + 0.01885937713086605, + 0.014436779543757439, + 0.01738189533352852, + 0.014988169074058533, + 0.011697134003043175, + 0.013378848321735859, + 0.014068707823753357, + 0.006190139800310135, + 0.006342862267047167, + 0.007993146777153015, + 0.0066857472993433475, + 0.008206400088965893, + 0.009511913172900677, + 0.006727126892656088, + 0.010613142512738705, + 0.005321561358869076, + 0.012015650980174541, + 0.00875039491802454, + 0.008345523849129677, + 0.007098381873220205, + 0.009747114963829517, + 0.0093338992446661, + 0.012042602524161339, + 0.014986719936132431, + 0.017124155536293983, + 0.016073592007160187, + 0.011224443092942238, + 0.0116653423756361, + 0.01342538371682167, + 0.010774834081530571, + 0.01320918370038271, + 0.005111976061016321, + 0.006211266387254, + 0.010363961569964886, + 0.013154877349734306, + 0.008172729052603245, + 0.010553082451224327, + 0.006615852937102318, + 0.011762706562876701, + 0.010403156280517578, + 0.016715602949261665, + 0.019756607711315155, + 0.015894077718257904, + 0.013133440166711807, + 0.020127063617110252, + 0.015298968181014061, + 0.020720364525914192, + 0.015218722634017467, + 0.013942800462245941, + 0.024101728573441505, + 0.016849355772137642, + 0.015039348043501377, + 0.02059621550142765, + 0.013503465801477432, + 0.0250491201877594, + 0.01178672257810831, + 0.01875012181699276, + 0.022858761250972748, + 0.0196150541305542, + 0.01656067743897438, + 0.018026692792773247, + 0.013997760601341724, + 0.020032353699207306 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 9, + 7 + ], + "selected_reward": -0.1156684160232544 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a shrouded, gloomy underground chamber with stone walls?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a massive three-meter-long bronze coffin with patina and age marks visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a nine-headed serpent encircling a sun disc appear on the coffin with scales pulsating in a dark blood-red glow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a young man in a simple black robe lying on the coffin lid, whose eyes open to reveal terror and a blue flame forming a miniature figure on his chest?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start with an extreme close-up from a slightly upward angle, then slowly pull back and pan upward to stabilize in a medium shot?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.033727146685123444 + ], + "opposite_logprob_tokens": [ + -3.4087271690368652 + ], + "logprob_reward": -0.033727146685123444, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.033727146685123444, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1474800854921341, + 0.09009486436843872, + 0.16569651663303375, + 0.07638101279735565, + 0.07435619831085205, + 0.10488442331552505, + 0.050480592995882034, + 0.07052166014909744, + 0.0417325422167778, + 0.1783721148967743 + ], + "credit_map": [ + 0.012460497207939625, + 0.01101576816290617, + 0.010569602251052856, + 0.012564855627715588, + 0.0382489375770092, + 0.020282071083784103, + 0.0107964389026165, + 0.03154192492365837, + 0.00813274271786213, + 0.012140577659010887, + 0.012723641470074654, + 0.009306768886744976, + 0.012860909104347229, + 0.010611377656459808, + 0.014317899942398071, + 0.01000094786286354, + 0.0217061135917902, + 0.03079758957028389, + 0.017393844202160835, + 0.021814098581671715, + 0.023977553471922874, + 0.01654805801808834, + 0.01748950034379959, + 0.015969762578606606, + 0.008019443601369858, + 0.009244999848306179, + 0.009916051290929317, + 0.012598874978721142, + 0.007880332879722118, + 0.009796862490475178, + 0.008094959892332554, + 0.010829486884176731, + 0.00804054643958807, + 0.011490922421216965, + 0.008697629906237125, + 0.008613809943199158, + 0.008931420743465424, + 0.009214051999151707, + 0.007896237075328827, + 0.011471576057374477, + 0.011398944072425365, + 0.009077074937522411, + 0.030558986589312553, + 0.01093536987900734, + 0.005910373758524656, + 0.01493226457387209, + 0.009660433977842331, + 0.012410970404744148, + 0.0035407652612775564, + 0.005803549662232399, + 0.0072344886139035225, + 0.011310544796288013, + 0.005315312184393406, + 0.006763082463294268, + 0.004464408848434687, + 0.006048441398888826, + 0.007108812220394611, + 0.0074609206058084965, + 0.011547902598977089, + 0.006743957754224539, + 0.008546177297830582, + 0.010686001740396023, + 0.007012354675680399, + 0.011415536515414715, + 0.0033407213632017374, + 0.005446850787848234, + 0.006445388309657574, + 0.004114835523068905, + 0.0054152715019881725, + 0.006195333786308765, + 0.004391297698020935, + 0.006382844410836697, + 0.014603523537516594, + 0.02227093279361725, + 0.027787946164608, + 0.022752992808818817, + 0.022322142496705055, + 0.02268793061375618, + 0.019885879009962082, + 0.026060763746500015 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 2, + 0 + ], + "selected_reward": -0.033727146685123444 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a shrouded, gloomy underground chamber with stone walls?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a massive three-meter-long bronze coffin with patina and age marks visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a nine-headed serpent encircling a sun disc appear on the coffin with scales pulsating in a dark blood-red glow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a young man in a simple black robe lying on the coffin lid, whose eyes open to reveal terror and a blue flame forming a miniature figure on his chest?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start with an extreme close-up from a slightly upward angle, then slowly pull back and pan upward to stabilize in a medium shot?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.029816951602697372 + ], + "opposite_logprob_tokens": [ + -3.5298168659210205 + ], + "logprob_reward": -0.029816951602697372, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.029816951602697372, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1081317588686943, + 0.03361133858561516, + 0.10867788642644882, + 0.04956959933042526, + 0.07419582456350327, + 0.17963464558124542, + 0.040991783142089844, + 0.08488486707210541, + 0.1016438901424408, + 0.21865835785865784 + ], + "credit_map": [ + 0.009198477491736412, + 0.007401364855468273, + 0.009107346646487713, + 0.010115035809576511, + 0.025037137791514397, + 0.014627699740231037, + 0.009539616294205189, + 0.023105088621377945, + 0.003443219466134906, + 0.0026141006965190172, + 0.0073757898062467575, + 0.003995899576693773, + 0.0041585275903344154, + 0.00411252211779356, + 0.003461294574663043, + 0.004449987784028053, + 0.010313885286450386, + 0.019222434610128403, + 0.011072244495153427, + 0.015473823063075542, + 0.014994737692177296, + 0.010877428576350212, + 0.014719530008733273, + 0.012003804557025433, + 0.005790216848254204, + 0.005986453033983707, + 0.005421055015176535, + 0.004838893655687571, + 0.0056684245355427265, + 0.005986976902931929, + 0.004485816694796085, + 0.011391766369342804, + 0.01098992582410574, + 0.008599509485065937, + 0.010836982168257236, + 0.008098121732473373, + 0.01069573312997818, + 0.007981457747519016, + 0.0072124917060136795, + 0.009781599044799805, + 0.02278892882168293, + 0.023073211312294006, + 0.037610944360494614, + 0.015518114902079105, + 0.014709114097058773, + 0.024144284427165985, + 0.018139302730560303, + 0.0236507635563612, + 0.005275351461023092, + 0.005234663840383291, + 0.006060949061065912, + 0.005915331654250622, + 0.003360792063176632, + 0.005761622451245785, + 0.002974242903292179, + 0.006408832035958767, + 0.007721189875155687, + 0.00868718046694994, + 0.01082789245992899, + 0.00935478787869215, + 0.008676267229020596, + 0.011353163979947567, + 0.015405542217195034, + 0.012858848087489605, + 0.01090238243341446, + 0.008992803283035755, + 0.01673327013850212, + 0.012652753852307796, + 0.011171839199960232, + 0.014121408574283123, + 0.012722618877887726, + 0.014346812851727009, + 0.02026902697980404, + 0.02600678987801075, + 0.03298937529325485, + 0.028243232518434525, + 0.028850117698311806, + 0.02627975307404995, + 0.024080799892544746, + 0.03193925321102142 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 5, + 2 + ], + "selected_reward": -0.029816951602697372 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a shrouded, gloomy underground chamber with stone walls?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a massive three-meter-long bronze coffin with patina and age marks visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a nine-headed serpent encircling a sun disc appear on the coffin with scales pulsating in a dark blood-red glow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a young man in a simple black robe lying on the coffin lid, whose eyes open to reveal terror and a blue flame forming a miniature figure on his chest?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start with an extreme close-up from a slightly upward angle, then slowly pull back and pan upward to stabilize in a medium shot?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07000525295734406 + ], + "opposite_logprob_tokens": [ + -2.695005178451538 + ], + "logprob_reward": -0.07000525295734406, + "yes_no_margin_reward": 2.625, + "token_credit_reward": -0.07000525295734406, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1157316192984581, + 0.06855592876672745, + 0.072965107858181, + 0.12265791743993759, + 0.09149055927991867, + 0.07207851856946945, + 0.06929849833250046, + 0.10670102387666702, + 0.06236741691827774, + 0.21815341711044312 + ], + "credit_map": [ + 0.010067216120660305, + 0.008188521489501, + 0.009666931815445423, + 0.007074267137795687, + 0.02848793752491474, + 0.012632889673113823, + 0.006598918233066797, + 0.0330149382352829, + 0.006464224774390459, + 0.008267330937087536, + 0.009037074632942677, + 0.00918613001704216, + 0.010271509177982807, + 0.008479130454361439, + 0.00930399727076292, + 0.007546525914222002, + 0.007525414228439331, + 0.013375840149819851, + 0.009361369535326958, + 0.012733417563140392, + 0.008648782037198544, + 0.007036126218736172, + 0.007540664169937372, + 0.006743491627275944, + 0.012976023368537426, + 0.013920486904680729, + 0.011444390751421452, + 0.014142727479338646, + 0.01712578721344471, + 0.014320853166282177, + 0.013684847392141819, + 0.025042809545993805, + 0.006123231258243322, + 0.011808509938418865, + 0.016223635524511337, + 0.012907463125884533, + 0.008070581592619419, + 0.012254606932401657, + 0.012053443118929863, + 0.012049093842506409, + 0.008876366540789604, + 0.008142655715346336, + 0.019064296036958694, + 0.006619075778871775, + 0.005062287673354149, + 0.010184481739997864, + 0.006054192781448364, + 0.008075159043073654, + 0.005251520313322544, + 0.008982870727777481, + 0.01175861619412899, + 0.011413326486945152, + 0.007493038196116686, + 0.010813611559569836, + 0.005383624695241451, + 0.008201895281672478, + 0.011742202565073967, + 0.022884657606482506, + 0.011577857658267021, + 0.010845509357750416, + 0.013765116222202778, + 0.015564405359327793, + 0.008158027194440365, + 0.012163244187831879, + 0.004378373268991709, + 0.00680886022746563, + 0.0137167414650321, + 0.00748337060213089, + 0.004658943507820368, + 0.010265529155731201, + 0.0059377271682024, + 0.009117872454226017, + 0.014444509521126747, + 0.024454576894640923, + 0.03565526008605957, + 0.022410893812775612, + 0.037734273821115494, + 0.028308577835559845, + 0.02538335509598255, + 0.029761992394924164 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 3, + 0 + ], + "selected_reward": -0.07000525295734406 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a shrouded, gloomy underground chamber with stone walls?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a massive three-meter-long bronze coffin with patina and age marks visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a nine-headed serpent encircling a sun disc appear on the coffin with scales pulsating in a dark blood-red glow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a young man in a simple black robe lying on the coffin lid, whose eyes open to reveal terror and a blue flame forming a miniature figure on his chest?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start with an extreme close-up from a slightly upward angle, then slowly pull back and pan upward to stabilize in a medium shot?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.018175480887293816 + ], + "opposite_logprob_tokens": [ + -4.0181756019592285 + ], + "logprob_reward": -0.018175480887293816, + "yes_no_margin_reward": 4.0, + "token_credit_reward": -0.018175480887293816, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06141975522041321, + 0.08520656079053879, + 0.040606722235679626, + 0.048984725028276443, + 0.04808066785335541, + 0.10889657586812973, + 0.06328842043876648, + 0.41249361634254456, + 0.06198297068476677, + 0.06904000043869019 + ], + "credit_map": [ + 0.006245036143809557, + 0.005186185240745544, + 0.00632345350459218, + 0.0059964885003864765, + 0.012953458353877068, + 0.007041512057185173, + 0.004924872890114784, + 0.012748746201395988, + 0.007136896718293428, + 0.008497877977788448, + 0.008663029409945011, + 0.009915534406900406, + 0.012881250120699406, + 0.013389715924859047, + 0.010667792521417141, + 0.014054466038942337, + 0.004479290917515755, + 0.005282361526042223, + 0.005069767590612173, + 0.0042191981337964535, + 0.005345606245100498, + 0.005110639613121748, + 0.0040574464946985245, + 0.007042408920824528, + 0.006296329665929079, + 0.005907920654863119, + 0.006086005363613367, + 0.007916366681456566, + 0.005033024586737156, + 0.005382120609283447, + 0.004490592982620001, + 0.00787236075848341, + 0.0049238442443311214, + 0.00664563337340951, + 0.006496844347566366, + 0.006119138561189175, + 0.005854859482496977, + 0.006122292019426823, + 0.004956386983394623, + 0.006961670238524675, + 0.011268057860434055, + 0.012594918720424175, + 0.01890084706246853, + 0.009327057749032974, + 0.010670196264982224, + 0.015807775780558586, + 0.011798003688454628, + 0.018529724329710007, + 0.005221368744969368, + 0.006680609192699194, + 0.008900580927729607, + 0.008708115667104721, + 0.009791363961994648, + 0.009038724936544895, + 0.00505616283044219, + 0.009891496039927006, + 0.03527373820543289, + 0.047156620770692825, + 0.06701620668172836, + 0.04359441250562668, + 0.04210861027240753, + 0.0660887211561203, + 0.04148212447762489, + 0.06977319717407227, + 0.005928772501647472, + 0.006318471860140562, + 0.011460250243544579, + 0.0054523758590221405, + 0.006598195526748896, + 0.010530161671340466, + 0.007179519161581993, + 0.008515221998095512, + 0.005374804604798555, + 0.008388768881559372, + 0.011016708798706532, + 0.010300188325345516, + 0.00834512896835804, + 0.008980292826890945, + 0.006942220497876406, + 0.009691886603832245 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 5, + 1 + ], + "selected_reward": -0.018175480887293816 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "shot, shot from a slightly upward angle, captures an extreme close-up of the walls of a massive, three-meter-long bronze coffin. The ancient bronze's surface is covered in varying shades of patina and the marks of age. The camera slowly pulls back and pans upward, gradually revealing an intricately carved nine-headed serpent encircling a sun disc. The serpent's scales, layered and lifelike, pulsate ominously with a dark, blood-red glow. These eyes, set in unpolished dark red gemstones, provide the only source of light in this shrouded, gloomy underground chamber, casting eerie, swaying shadows on the surrounding stone walls. As the camera continues to pull out, it finally stabilizes in a medium shot, revealing the coffin in its entirety and a young man in a simple black robe lying on its lid. At that moment, the long eyelashes on the boy's delicate yet pale face began to tremble violently, as if struggling to escape from a bottomless nightmare. Then, his eyes fluttered open, revealing utter terror and confusion. Almost instantly, through the black robes on his chest, a ball of blue flame pulsed silently and violently from the area where his heart lay. This mysterious, cold light illuminated the outline of his chest clearly. Then, the flame rapidly contracted and condensed, eventually transforming into a palm-sized miniature figure composed entirely of blue flame. This flaming figure, arms crossed, calmly \"sat\" on the boy's left rib cage. The entire scene, interwoven with blood-red and dark blue light, created an atmosphere of ultimate horror and fantasy.", + "seed": 32442, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000006/group_0/index_3113_seed_32442_dp_1_rank_2_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 0, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.08953378237783909, + "vlm_logprob": -0.08953378237783909, + "vlm_yes_no_margin": 2.8, + "vlm_token_credit": -0.08953378237783909, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a shrouded, gloomy underground chamber with stone walls?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a massive three-meter-long bronze coffin with patina and age marks visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a nine-headed serpent encircling a sun disc appear on the coffin with scales pulsating in a dark blood-red glow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a young man in a simple black robe lying on the coffin lid, whose eyes open to reveal terror and a blue flame forming a miniature figure on his chest?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start with an extreme close-up from a slightly upward angle, then slowly pull back and pan upward to stabilize in a medium shot?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.09257940202951431 + ], + "opposite_logprob_tokens": [ + -2.4675793647766113 + ], + "logprob_reward": -0.09257940202951431, + "yes_no_margin_reward": 2.375, + "token_credit_reward": -0.09257940202951431, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.524706244468689, + 0.0748220756649971, + 0.20867572724819183, + 0.03294020891189575, + 0.023499930277466774, + 0.03191541135311127, + 0.03788824751973152, + 0.019244611263275146, + 0.03475194051861763, + 0.01155555434525013 + ], + "credit_map": [ + 0.12032919377088547, + 0.01730353571474552, + 0.20402824878692627, + 0.033214911818504333, + 0.021656552329659462, + 0.09718473255634308, + 0.008651669137179852, + 0.022337401285767555, + 0.011157078668475151, + 0.00862918607890606, + 0.013090502470731735, + 0.010754894465208054, + 0.008296642452478409, + 0.008506076410412788, + 0.007866413332521915, + 0.006521278060972691, + 0.02852058596909046, + 0.02303299494087696, + 0.021870069205760956, + 0.029213914647698402, + 0.02164183184504509, + 0.037170398980379105, + 0.019984696060419083, + 0.027241230010986328, + 0.004478910937905312, + 0.005779000464826822, + 0.006154455244541168, + 0.0046445149928331375, + 0.002903885440900922, + 0.00341654266230762, + 0.0028746253810822964, + 0.0026882716920226812, + 0.00295698968693614, + 0.0030341623350977898, + 0.0028391883242875338, + 0.0029435805045068264, + 0.002367030829191208, + 0.004148745909333229, + 0.00294390763156116, + 0.002266325056552887, + 0.0034386715851724148, + 0.0038023649249225855, + 0.004194663837552071, + 0.005121982656419277, + 0.004167516250163317, + 0.0025036539882421494, + 0.005516950041055679, + 0.0031696069054305553, + 0.0039878417737782, + 0.009077356196939945, + 0.004769742023199797, + 0.00509423203766346, + 0.00406601233407855, + 0.00401546573266387, + 0.0040308707393705845, + 0.0028467276133596897, + 0.0023392029106616974, + 0.003087189281359315, + 0.002343076514080167, + 0.0021729085128754377, + 0.002154405228793621, + 0.0027439293917268515, + 0.002556681400164962, + 0.0018472191877663136, + 0.006462431512773037, + 0.003415957558900118, + 0.012768585234880447, + 0.002692030044272542, + 0.0024084479082375765, + 0.003467933041974902, + 0.0017488207668066025, + 0.0017877353820949793, + 0.0017403652891516685, + 0.0021874101366847754, + 0.001547692110762, + 0.0010434597497805953, + 0.001444838591851294, + 0.0014515607617795467, + 0.001301193842664361, + 0.0008390332222916186 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 1 + ], + "selected_reward": -0.09257940202951431 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a shrouded, gloomy underground chamber with stone walls?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a massive three-meter-long bronze coffin with patina and age marks visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a nine-headed serpent encircling a sun disc appear on the coffin with scales pulsating in a dark blood-red glow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a young man in a simple black robe lying on the coffin lid, whose eyes open to reveal terror and a blue flame forming a miniature figure on his chest?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start with an extreme close-up from a slightly upward angle, then slowly pull back and pan upward to stabilize in a medium shot?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.029840555042028427 + ], + "opposite_logprob_tokens": [ + -3.5298404693603516 + ], + "logprob_reward": -0.029840555042028427, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.029840555042028427, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.44828522205352783, + 0.051193516701459885, + 0.23541758954524994, + 0.031244801357388496, + 0.03481521084904671, + 0.03406743332743645, + 0.058384351432323456, + 0.04954405128955841, + 0.0323668047785759, + 0.024680985137820244 + ], + "credit_map": [ + 0.10542421042919159, + 0.012437784112989902, + 0.1772693693637848, + 0.028814805671572685, + 0.01810608059167862, + 0.07925067096948624, + 0.00883340835571289, + 0.018148906528949738, + 0.006717136595398188, + 0.00673674838617444, + 0.00786500982940197, + 0.00832442194223404, + 0.007930154912173748, + 0.004585845861583948, + 0.0047876290045678616, + 0.0042465669102966785, + 0.03265795111656189, + 0.027066707611083984, + 0.02426034025847912, + 0.032047729939222336, + 0.026092825457453728, + 0.04023117199540138, + 0.022290769964456558, + 0.0307700764387846, + 0.004965599626302719, + 0.004191604442894459, + 0.005346762947738171, + 0.004016929771751165, + 0.003523648949339986, + 0.0025886869989335537, + 0.0032124468125402927, + 0.0033991215750575066, + 0.0042949458584189415, + 0.004810792859643698, + 0.003985848743468523, + 0.0046087391674518585, + 0.0035089473240077496, + 0.0053758141584694386, + 0.005248601548373699, + 0.0029815202578902245, + 0.0034526067320257425, + 0.003755903569981456, + 0.004183285403996706, + 0.005286507774144411, + 0.004521152004599571, + 0.003028484294191003, + 0.0060912542976439, + 0.003748240415006876, + 0.006452823523432016, + 0.013501894660294056, + 0.007399719208478928, + 0.007873672991991043, + 0.006365951616317034, + 0.0043099792674183846, + 0.006995226256549358, + 0.00548508670181036, + 0.007277199998497963, + 0.0060399738140404224, + 0.013057205826044083, + 0.004402767401188612, + 0.004918225109577179, + 0.0044439248740673065, + 0.005279608070850372, + 0.004125144332647324, + 0.00506560318171978, + 0.006615986116230488, + 0.004878554027527571, + 0.002658974612131715, + 0.0037528597749769688, + 0.0034509499091655016, + 0.0031650757882744074, + 0.0027788011357188225, + 0.004155546426773071, + 0.004864648915827274, + 0.0029197947587817907, + 0.0028343647718429565, + 0.0029700701124966145, + 0.002945820102468133, + 0.002621180610731244, + 0.0013695599045604467 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 6 + ], + "selected_reward": -0.029840555042028427 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a shrouded, gloomy underground chamber with stone walls?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a massive three-meter-long bronze coffin with patina and age marks visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a nine-headed serpent encircling a sun disc appear on the coffin with scales pulsating in a dark blood-red glow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a young man in a simple black robe lying on the coffin lid, whose eyes open to reveal terror and a blue flame forming a miniature figure on his chest?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start with an extreme close-up from a slightly upward angle, then slowly pull back and pan upward to stabilize in a medium shot?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.05497393012046814 + ], + "opposite_logprob_tokens": [ + -2.929973840713501 + ], + "logprob_reward": -0.05497393012046814, + "yes_no_margin_reward": 2.875, + "token_credit_reward": -0.05497393012046814, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.6124646067619324, + 0.047567129135131836, + 0.07630458474159241, + 0.023261385038495064, + 0.015385454520583153, + 0.046691831201314926, + 0.035758282989263535, + 0.04003186523914337, + 0.08249689638614655, + 0.020037855952978134 + ], + "credit_map": [ + 0.13856762647628784, + 0.029185667634010315, + 0.22864104807376862, + 0.03926971182227135, + 0.02436082251369953, + 0.11315140873193741, + 0.014289112761616707, + 0.02499924786388874, + 0.006470542401075363, + 0.010151003487408161, + 0.006270971614867449, + 0.00671343645080924, + 0.006552935112267733, + 0.003575126640498638, + 0.0035261071752756834, + 0.00430701207369566, + 0.009255817160010338, + 0.012031464837491512, + 0.008589699864387512, + 0.009566409513354301, + 0.009267590008676052, + 0.011870399117469788, + 0.007036483846604824, + 0.008686723187565804, + 0.0034316242672502995, + 0.003744990797713399, + 0.003864753758534789, + 0.0033993450924754143, + 0.0025010909885168076, + 0.0015428977785632014, + 0.002530711470171809, + 0.002245973562821746, + 0.0018378400709480047, + 0.002260400913655758, + 0.001664388575591147, + 0.001786794513463974, + 0.0018721697852015495, + 0.0024084087926894426, + 0.002093818038702011, + 0.0014616356929764152, + 0.005091833416372538, + 0.005549551919102669, + 0.006113463547080755, + 0.007030222099274397, + 0.005266386549919844, + 0.005190203431993723, + 0.007812227122485638, + 0.004637946840375662, + 0.004282047040760517, + 0.005046840291470289, + 0.004879004787653685, + 0.005095224361866713, + 0.00448956061154604, + 0.004384312313050032, + 0.003964110743254423, + 0.003617184702306986, + 0.0052370717748999596, + 0.006470384541898966, + 0.006298571825027466, + 0.003539060475304723, + 0.00387552916072309, + 0.004136298317462206, + 0.006076691672205925, + 0.00439826212823391, + 0.009992855601012707, + 0.011253082193434238, + 0.028546074405312538, + 0.006156967952847481, + 0.0062509560957551, + 0.009944654069840908, + 0.004245193209499121, + 0.00610713055357337, + 0.0029548739548772573, + 0.0037284644786268473, + 0.0023086508736014366, + 0.0018681208603084087, + 0.0027711361180990934, + 0.002422401448711753, + 0.0025516292080283165, + 0.001432583318091929 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 2 + ], + "selected_reward": -0.05497393012046814 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a shrouded, gloomy underground chamber with stone walls?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a massive three-meter-long bronze coffin with patina and age marks visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a nine-headed serpent encircling a sun disc appear on the coffin with scales pulsating in a dark blood-red glow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a young man in a simple black robe lying on the coffin lid, whose eyes open to reveal terror and a blue flame forming a miniature figure on his chest?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start with an extreme close-up from a slightly upward angle, then slowly pull back and pan upward to stabilize in a medium shot?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.2520946264266968 + ], + "opposite_logprob_tokens": [ + -1.5020946264266968 + ], + "logprob_reward": -0.2520946264266968, + "yes_no_margin_reward": 1.25, + "token_credit_reward": -0.2520946264266968, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.410101979970932, + 0.06226573511958122, + 0.13507485389709473, + 0.0985114574432373, + 0.02209652028977871, + 0.10352946072816849, + 0.04517275094985962, + 0.05927274003624916, + 0.041050057858228683, + 0.02292446419596672 + ], + "credit_map": [ + 0.09428879618644714, + 0.014604225754737854, + 0.15965187549591064, + 0.02635246515274048, + 0.01636580191552639, + 0.0724102109670639, + 0.00885067693889141, + 0.017577946186065674, + 0.0081262718886137, + 0.009713559411466122, + 0.010397041216492653, + 0.009425278753042221, + 0.007494692225009203, + 0.005347511265426874, + 0.006911231204867363, + 0.004850140307098627, + 0.018665287643671036, + 0.014999120496213436, + 0.014255990274250507, + 0.019140109419822693, + 0.013700970448553562, + 0.02315991185605526, + 0.013281397521495819, + 0.017872057855129242, + 0.01442838367074728, + 0.015515041537582874, + 0.02225104719400406, + 0.013844582252204418, + 0.008044110611081123, + 0.007533229887485504, + 0.009130630642175674, + 0.0077644232660532, + 0.00316358869895339, + 0.002930794842541218, + 0.004149993881583214, + 0.002868743846192956, + 0.0018827804597094655, + 0.002545219147577882, + 0.002660313853994012, + 0.0018950847443193197, + 0.010318290442228317, + 0.011945432983338833, + 0.013014702126383781, + 0.013387095183134079, + 0.015236634761095047, + 0.009455012157559395, + 0.01892775483429432, + 0.011244519613683224, + 0.006037434097379446, + 0.007341902237385511, + 0.005284562241286039, + 0.006113565526902676, + 0.005589127540588379, + 0.004647398367524147, + 0.005229982081800699, + 0.004928775131702423, + 0.007955537177622318, + 0.00768447108566761, + 0.012451736256480217, + 0.004891423042863607, + 0.0059265438467264175, + 0.006634704302996397, + 0.008263789117336273, + 0.005464531946927309, + 0.00553029915317893, + 0.006513261701911688, + 0.007336913142353296, + 0.0029210930224508047, + 0.005015204194933176, + 0.004341955296695232, + 0.00511702848598361, + 0.004274300299584866, + 0.0037227277643978596, + 0.0035386639647185802, + 0.0033768173307180405, + 0.002744590863585472, + 0.002549251541495323, + 0.0026341131888329983, + 0.0028442973271012306, + 0.0015140007017180324 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 5 + ], + "selected_reward": -0.2520946264266968 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a shrouded, gloomy underground chamber with stone walls?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a massive three-meter-long bronze coffin with patina and age marks visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a nine-headed serpent encircling a sun disc appear on the coffin with scales pulsating in a dark blood-red glow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a young man in a simple black robe lying on the coffin lid, whose eyes open to reveal terror and a blue flame forming a miniature figure on his chest?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start with an extreme close-up from a slightly upward angle, then slowly pull back and pan upward to stabilize in a medium shot?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.018180398270487785 + ], + "opposite_logprob_tokens": [ + -4.0181803703308105 + ], + "logprob_reward": -0.018180398270487785, + "yes_no_margin_reward": 4.0, + "token_credit_reward": -0.018180398270487785, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.4984378218650818, + 0.03087613731622696, + 0.17100997269153595, + 0.026563918218016624, + 0.022655503824353218, + 0.07776866108179092, + 0.04904411733150482, + 0.0307184886187315, + 0.037492748349905014, + 0.05543268844485283 + ], + "credit_map": [ + 0.11352486163377762, + 0.017845219001173973, + 0.19217215478420258, + 0.03105204366147518, + 0.019662905484437943, + 0.09257038682699203, + 0.010781880468130112, + 0.020828379318118095, + 0.004243447445333004, + 0.003367916215211153, + 0.005000373348593712, + 0.004590215161442757, + 0.0038789273239672184, + 0.003557987045496702, + 0.003837713971734047, + 0.0023995598312467337, + 0.02324981987476349, + 0.019552769139409065, + 0.01738397218286991, + 0.02486390620470047, + 0.017842715606093407, + 0.029039395973086357, + 0.01586158759891987, + 0.023215793073177338, + 0.003494017058983445, + 0.003287685103714466, + 0.004896273370832205, + 0.003801667829975486, + 0.0028323831502348185, + 0.0017886570421978831, + 0.0038110760506242514, + 0.0026521580293774605, + 0.002526700496673584, + 0.0032131781335920095, + 0.0023158143740147352, + 0.0034745235461741686, + 0.002976738614961505, + 0.0034000780433416367, + 0.002865420188754797, + 0.0018830486806109548, + 0.009177103638648987, + 0.009789646603167057, + 0.009811058640480042, + 0.009148282930254936, + 0.010456006042659283, + 0.00835332926362753, + 0.012366952374577522, + 0.008666280657052994, + 0.004664686508476734, + 0.008945080451667309, + 0.004846355877816677, + 0.0063202339224517345, + 0.005998952314257622, + 0.00880561675876379, + 0.00516984099522233, + 0.004293349105864763, + 0.003722812980413437, + 0.00717464042827487, + 0.003108358709141612, + 0.0032199257984757423, + 0.0028518608305603266, + 0.003198802238330245, + 0.003663833485916257, + 0.0037782543804496527, + 0.0037758282851427794, + 0.006203128024935722, + 0.00996478833258152, + 0.002679624827578664, + 0.0038979703094810247, + 0.00502753583714366, + 0.003064210992306471, + 0.0028796617407351732, + 0.012552715837955475, + 0.008784737437963486, + 0.009300097823143005, + 0.005711199715733528, + 0.006393492687493563, + 0.005304616875946522, + 0.004651250317692757, + 0.002734574256464839 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 5 + ], + "selected_reward": -0.018180398270487785 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "shot, shot from a slightly upward angle, captures an extreme close-up of the walls of a massive, three-meter-long bronze coffin. The ancient bronze's surface is covered in varying shades of patina and the marks of age. The camera slowly pulls back and pans upward, gradually revealing an intricately carved nine-headed serpent encircling a sun disc. The serpent's scales, layered and lifelike, pulsate ominously with a dark, blood-red glow. These eyes, set in unpolished dark red gemstones, provide the only source of light in this shrouded, gloomy underground chamber, casting eerie, swaying shadows on the surrounding stone walls. As the camera continues to pull out, it finally stabilizes in a medium shot, revealing the coffin in its entirety and a young man in a simple black robe lying on its lid. At that moment, the long eyelashes on the boy's delicate yet pale face began to tremble violently, as if struggling to escape from a bottomless nightmare. Then, his eyes fluttered open, revealing utter terror and confusion. Almost instantly, through the black robes on his chest, a ball of blue flame pulsed silently and violently from the area where his heart lay. This mysterious, cold light illuminated the outline of his chest clearly. Then, the flame rapidly contracted and condensed, eventually transforming into a palm-sized miniature figure composed entirely of blue flame. This flaming figure, arms crossed, calmly \"sat\" on the boy's left rib cage. The entire scene, interwoven with blood-red and dark blue light, created an atmosphere of ultimate horror and fantasy.", + "seed": 32443, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000006/group_0/index_3113_seed_32443_dp_1_rank_2_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 0, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.02369696842506528, + "vlm_logprob": -0.02369696842506528, + "vlm_yes_no_margin": 3.925, + "vlm_token_credit": -0.02369696842506528, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a shrouded, gloomy underground chamber with stone walls?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a massive three-meter-long bronze coffin with patina and age marks visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a nine-headed serpent encircling a sun disc appear on the coffin with scales pulsating in a dark blood-red glow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a young man in a simple black robe lying on the coffin lid, whose eyes open to reveal terror and a blue flame forming a miniature figure on his chest?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start with an extreme close-up from a slightly upward angle, then slowly pull back and pan upward to stabilize in a medium shot?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.031526800245046616 + ], + "opposite_logprob_tokens": [ + -3.531526803970337 + ], + "logprob_reward": -0.031526800245046616, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.031526800245046616, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.32699161767959595, + 0.2519164979457855, + 0.024490399286150932, + 0.07299580425024033, + 0.07461967319250107, + 0.028460659086704254, + 0.05597863346338272, + 0.05362226814031601, + 0.07373026758432388, + 0.037194155156612396 + ], + "credit_map": [ + 0.047402456402778625, + 0.013947024941444397, + 0.031294990330934525, + 0.036804813891649246, + 0.014524663798511028, + 0.024494461715221405, + 0.0245368555188179, + 0.13398633897304535, + 0.017056060954928398, + 0.012572132982313633, + 0.09034394472837448, + 0.08249212056398392, + 0.013612783513963223, + 0.006434157025068998, + 0.01650305651128292, + 0.012902227230370045, + 0.002726408652961254, + 0.003218915080651641, + 0.002023090608417988, + 0.0035414404701441526, + 0.003711947239935398, + 0.002137409755960107, + 0.004183423239737749, + 0.002947764005511999, + 0.008254158310592175, + 0.00605325773358345, + 0.022170793265104294, + 0.016146818175911903, + 0.005791453644633293, + 0.004241081885993481, + 0.005149648059159517, + 0.0051885927096009254, + 0.006500754971057177, + 0.009193903766572475, + 0.011064933612942696, + 0.017160564661026, + 0.007485669571906328, + 0.005876635201275349, + 0.010800198651850224, + 0.006537014618515968, + 0.0021804452408105135, + 0.0035896070767194033, + 0.003328927792608738, + 0.005119127221405506, + 0.0030436471570283175, + 0.001969900680705905, + 0.0055733006447553635, + 0.0036557018756866455, + 0.006669806782156229, + 0.006391939707100391, + 0.007323395926505327, + 0.007958317175507545, + 0.005525517743080854, + 0.0067727454006671906, + 0.00904807262122631, + 0.006288839504122734, + 0.005280342418700457, + 0.003794444026425481, + 0.014682242646813393, + 0.00533671211451292, + 0.0025406780187040567, + 0.0040582530200481415, + 0.004745017271488905, + 0.01318457629531622, + 0.008900165557861328, + 0.0037049197126179934, + 0.024788497015833855, + 0.003666629083454609, + 0.002797764027491212, + 0.02067447453737259, + 0.002787224482744932, + 0.006410595029592514, + 0.006996798794716597, + 0.0031229222659021616, + 0.007910034619271755, + 0.002096206182613969, + 0.002760868752375245, + 0.004692601505666971, + 0.002284460002556443, + 0.007330264896154404 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 4 + ], + "selected_reward": -0.031526800245046616 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a shrouded, gloomy underground chamber with stone walls?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a massive three-meter-long bronze coffin with patina and age marks visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a nine-headed serpent encircling a sun disc appear on the coffin with scales pulsating in a dark blood-red glow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a young man in a simple black robe lying on the coffin lid, whose eyes open to reveal terror and a blue flame forming a miniature figure on his chest?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start with an extreme close-up from a slightly upward angle, then slowly pull back and pan upward to stabilize in a medium shot?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02983083575963974 + ], + "opposite_logprob_tokens": [ + -3.5298309326171875 + ], + "logprob_reward": -0.02983083575963974, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.02983083575963974, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.37660619616508484, + 0.1538257896900177, + 0.04858649894595146, + 0.16607950627803802, + 0.03815315663814545, + 0.03241276741027832, + 0.04413311183452606, + 0.07975345849990845, + 0.02473812736570835, + 0.03571135178208351 + ], + "credit_map": [ + 0.053390245884656906, + 0.01421080157160759, + 0.03185681626200676, + 0.04432312771677971, + 0.017003601416945457, + 0.027280015870928764, + 0.028335751965641975, + 0.1602058708667755, + 0.00974652636796236, + 0.009800095111131668, + 0.044062212109565735, + 0.04252129793167114, + 0.013886244036257267, + 0.00792111735790968, + 0.01391553319990635, + 0.011972776614129543, + 0.0057761175557971, + 0.004447407554835081, + 0.004804819822311401, + 0.006591413635760546, + 0.005826565902680159, + 0.00406298041343689, + 0.007668087258934975, + 0.009409108199179173, + 0.027620239183306694, + 0.013249031268060207, + 0.05818001553416252, + 0.029821667820215225, + 0.013275998644530773, + 0.006516137160360813, + 0.00999408308416605, + 0.00742235267534852, + 0.0033869845792651176, + 0.005667014047503471, + 0.005427468568086624, + 0.006570805795490742, + 0.005708201322704554, + 0.003531578229740262, + 0.004814666695892811, + 0.003046440426260233, + 0.0033849584870040417, + 0.004146901424974203, + 0.0050593954510986805, + 0.005754182580858469, + 0.003026083577424288, + 0.002131811110302806, + 0.0036001058761030436, + 0.005309332627803087, + 0.0038238607812672853, + 0.004966041073203087, + 0.006105012726038694, + 0.007058652117848396, + 0.0038941113743931055, + 0.00458752503618598, + 0.007454495877027512, + 0.006243416573852301, + 0.007756393402814865, + 0.005494097247719765, + 0.018167583271861076, + 0.011066392064094543, + 0.005384358111768961, + 0.005002911202609539, + 0.01329135149717331, + 0.013590374030172825, + 0.0031341176945716143, + 0.002663303166627884, + 0.0030082641169428825, + 0.002385679166764021, + 0.0014017465291544795, + 0.005979853216558695, + 0.0021642830688506365, + 0.004000884015113115, + 0.004538371227681637, + 0.004185734782367945, + 0.007020282093435526, + 0.0025960588827729225, + 0.0037687956355512142, + 0.0045079197734594345, + 0.002721241442486644, + 0.006372953299432993 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 1 + ], + "selected_reward": -0.02983083575963974 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a shrouded, gloomy underground chamber with stone walls?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a massive three-meter-long bronze coffin with patina and age marks visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a nine-headed serpent encircling a sun disc appear on the coffin with scales pulsating in a dark blood-red glow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a young man in a simple black robe lying on the coffin lid, whose eyes open to reveal terror and a blue flame forming a miniature figure on his chest?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start with an extreme close-up from a slightly upward angle, then slowly pull back and pan upward to stabilize in a medium shot?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005258894059807062 + ], + "opposite_logprob_tokens": [ + -5.255259037017822 + ], + "logprob_reward": -0.005258894059807062, + "yes_no_margin_reward": 5.25, + "token_credit_reward": -0.005258894059807062, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.37846067547798157, + 0.08545927703380585, + 0.09070322662591934, + 0.10819680988788605, + 0.06603960692882538, + 0.04540468007326126, + 0.05516602843999863, + 0.02248172089457512, + 0.1218399852514267, + 0.026248067617416382 + ], + "credit_map": [ + 0.05053260922431946, + 0.02548152580857277, + 0.038946229964494705, + 0.03725739195942879, + 0.036604370921850204, + 0.03144785016775131, + 0.03650415316224098, + 0.12168655544519424, + 0.006840132642537355, + 0.00941129308193922, + 0.017836058512330055, + 0.019227521494030952, + 0.008482294157147408, + 0.007369306869804859, + 0.007963442243635654, + 0.00832922849804163, + 0.006768776103854179, + 0.008806491270661354, + 0.009387129917740822, + 0.012463887222111225, + 0.011507835239171982, + 0.007849062792956829, + 0.012396909296512604, + 0.0215231254696846, + 0.01564720831811428, + 0.006781589239835739, + 0.03837994113564491, + 0.02212553843855858, + 0.00781785324215889, + 0.004826649092137814, + 0.005705167539417744, + 0.006912859156727791, + 0.004908359609544277, + 0.006960434839129448, + 0.011434154585003853, + 0.011918051168322563, + 0.007697942666709423, + 0.005643327720463276, + 0.01171596348285675, + 0.005761369131505489, + 0.003843148937448859, + 0.004805097822099924, + 0.006346309091895819, + 0.009903266094624996, + 0.0040533156134188175, + 0.0035283484030514956, + 0.006210005842149258, + 0.00671518687158823, + 0.00671433936804533, + 0.005991983227431774, + 0.008919569663703442, + 0.009426921606063843, + 0.004335577134042978, + 0.0052119153551757336, + 0.007649071980267763, + 0.0069166505709290504, + 0.0040105353109538555, + 0.0023257271386682987, + 0.0026523026172071695, + 0.0030259327031672, + 0.0020979943219572306, + 0.002230771817266941, + 0.002944208914414048, + 0.0031942478381097317, + 0.01839502528309822, + 0.006998726632446051, + 0.04896806180477142, + 0.005024915095418692, + 0.00379010196775198, + 0.022548997774720192, + 0.003912660758942366, + 0.012201483361423016, + 0.00391093734651804, + 0.0031980378553271294, + 0.0031418183352798223, + 0.003942720126360655, + 0.002113433089107275, + 0.00247254129499197, + 0.0034831189550459385, + 0.003985460381954908 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 3 + ], + "selected_reward": -0.005258894059807062 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a shrouded, gloomy underground chamber with stone walls?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a massive three-meter-long bronze coffin with patina and age marks visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a nine-headed serpent encircling a sun disc appear on the coffin with scales pulsating in a dark blood-red glow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a young man in a simple black robe lying on the coffin lid, whose eyes open to reveal terror and a blue flame forming a miniature figure on his chest?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start with an extreme close-up from a slightly upward angle, then slowly pull back and pan upward to stabilize in a medium shot?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03369072452187538 + ], + "opposite_logprob_tokens": [ + -3.4086906909942627 + ], + "logprob_reward": -0.03369072452187538, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.03369072452187538, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2595336437225342, + 0.065678671002388, + 0.028371578082442284, + 0.05610715597867966, + 0.028667910024523735, + 0.10295561701059341, + 0.044699329882860184, + 0.12093464285135269, + 0.1776294708251953, + 0.11542201042175293 + ], + "credit_map": [ + 0.03905109688639641, + 0.01482754573225975, + 0.02149142324924469, + 0.02499524876475334, + 0.01737837865948677, + 0.02048255316913128, + 0.022099105641245842, + 0.099208302795887, + 0.004779044538736343, + 0.006874939892441034, + 0.015865186229348183, + 0.019199064001441002, + 0.006462995428591967, + 0.0028173683676868677, + 0.00638092216104269, + 0.003299152245745063, + 0.002714687492698431, + 0.004514830186963081, + 0.0028210405725985765, + 0.004153796937316656, + 0.003726938273757696, + 0.0029462561942636967, + 0.003626827150583267, + 0.0038672012742608786, + 0.00619919877499342, + 0.004644972272217274, + 0.014994076453149319, + 0.012139935046434402, + 0.004535342566668987, + 0.0032742153853178024, + 0.0059746685437858105, + 0.004344746470451355, + 0.0029020749498158693, + 0.003391251666471362, + 0.0042552403174340725, + 0.004368473310023546, + 0.003884066129103303, + 0.0029244152829051018, + 0.003943488467484713, + 0.0029988985043019056, + 0.009043467231094837, + 0.009701477363705635, + 0.017621774226427078, + 0.01915808394551277, + 0.006423551589250565, + 0.007312099449336529, + 0.012289095669984818, + 0.021406058222055435, + 0.0029810103587806225, + 0.004905768670141697, + 0.005888687446713448, + 0.009995917789638042, + 0.004104142542928457, + 0.002856984967365861, + 0.007062727585434914, + 0.006904092617332935, + 0.011121298186480999, + 0.005517158191651106, + 0.03861800581216812, + 0.008204386569559574, + 0.0046289884485304356, + 0.00796802993863821, + 0.008436021395027637, + 0.036440759897232056, + 0.021342258900403976, + 0.006835807580500841, + 0.06245647743344307, + 0.007768520154058933, + 0.005270868539810181, + 0.05293715000152588, + 0.006147840991616249, + 0.014870532788336277, + 0.019060859456658363, + 0.01424622442573309, + 0.018068360164761543, + 0.008748657070100307, + 0.01192348450422287, + 0.012468017637729645, + 0.010976670309901237, + 0.019929733127355576 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 7 + ], + "selected_reward": -0.03369072452187538 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a shrouded, gloomy underground chamber with stone walls?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a massive three-meter-long bronze coffin with patina and age marks visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a nine-headed serpent encircling a sun disc appear on the coffin with scales pulsating in a dark blood-red glow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a young man in a simple black robe lying on the coffin lid, whose eyes open to reveal terror and a blue flame forming a miniature figure on his chest?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start with an extreme close-up from a slightly upward angle, then slowly pull back and pan upward to stabilize in a medium shot?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.018177587538957596 + ], + "opposite_logprob_tokens": [ + -4.018177509307861 + ], + "logprob_reward": -0.018177587538957596, + "yes_no_margin_reward": 4.0, + "token_credit_reward": -0.018177587538957596, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2253846377134323, + 0.21822552382946014, + 0.024176400154829025, + 0.06088275834918022, + 0.045106302946805954, + 0.03663182258605957, + 0.04398920387029648, + 0.07865066826343536, + 0.22900280356407166, + 0.03794983774423599 + ], + "credit_map": [ + 0.032966192811727524, + 0.00920855812728405, + 0.018733568489551544, + 0.02747628279030323, + 0.010450311005115509, + 0.017593326047062874, + 0.01668315939605236, + 0.09227324277162552, + 0.014988668262958527, + 0.009694065898656845, + 0.08264745771884918, + 0.07653539627790451, + 0.005872136913239956, + 0.005490096751600504, + 0.01388926524668932, + 0.009108437225222588, + 0.0027312308084219694, + 0.0033742228988558054, + 0.0022233328782022, + 0.003168217372149229, + 0.0029541784897446632, + 0.002411654917523265, + 0.0038165876176208258, + 0.0034969737753272057, + 0.007614472880959511, + 0.006909552495926619, + 0.011986568570137024, + 0.007957958616316319, + 0.0062402826733887196, + 0.005144002381712198, + 0.008487778715789318, + 0.0065421429462730885, + 0.0035523422993719578, + 0.006482966244220734, + 0.006245372351258993, + 0.009237781167030334, + 0.00545860268175602, + 0.0039581540040671825, + 0.0055677019990980625, + 0.004603381734341383, + 0.004423444624990225, + 0.004435700364410877, + 0.004208710510283709, + 0.0067445975728333, + 0.00304229324683547, + 0.0028910741675645113, + 0.004898764658719301, + 0.005987234879285097, + 0.004137955140322447, + 0.003990691155195236, + 0.006380901206284761, + 0.0064620571210980415, + 0.005234785843640566, + 0.006128537468612194, + 0.005390480626374483, + 0.006263794377446175, + 0.00714681763201952, + 0.0037605094257742167, + 0.02635236270725727, + 0.0043028914369642735, + 0.0032335652504116297, + 0.004674102179706097, + 0.005246309097856283, + 0.02393411472439766, + 0.028050929307937622, + 0.009911173023283482, + 0.08315140753984451, + 0.01103976834565401, + 0.007784008979797363, + 0.06266630440950394, + 0.01116910669952631, + 0.015230100601911545, + 0.006912334356456995, + 0.003964543808251619, + 0.007557865232229233, + 0.002655062125995755, + 0.002847088733687997, + 0.0040684971027076244, + 0.0036992651876062155, + 0.006245177239179611 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 0, + 1 + ], + "selected_reward": -0.018177587538957596 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000006/group_1/index_tensor([7097, 7097])_group_1_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000006/group_1/index_tensor([7097, 7097])_group_1_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..587b6725c76857e0688ad06aa10d9a976fd1e48d --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000006/group_1/index_tensor([7097, 7097])_group_1_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "An extreme close-up shot, captured using high-speed photography, closely tracks a spherical particle of Tieguanyin tea being released from the air. Tumbling through the air at high speed, the particle's surface is not smooth, but rather covered with intricate textures and subtle wrinkles resulting from its tightly curled shape. Its dark green base is interspersed with subtle dark shadows and pale yellow-green highlights. Bright, focused studio lighting, directed from above and sideways, creates sharp highlights and deep shadows on the particle's curved surface, dramatically enhancing its three-dimensionality and texture. Surrounding it, the background is a blur of countless similarly scattered tea particles. With a very shallow depth of field, these particles transform into a blurred, flowing sea of dark green light, forming a dynamic, directional trajectory with the main particle. The camera uses an impactful face-to-face shot effect to capture the dynamic moment when the tea leaf suddenly rushes towards the lens from the depths of the picture, almost brushing the lens, and then quickly flies out of the field of vision to the side. The whole process is full of extreme dynamic beauty and strong visual impact, as if time is frozen in this one thousandth of a second, creating a focused and powerful atmosphere.", + "seed": 16269, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000006/group_1/index_7097_seed_16269_dp_2_rank_4_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 1, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.08463812698610126, + "vlm_logprob": -0.08463812698610126, + "vlm_yes_no_margin": 3.65, + "vlm_token_credit": -0.08463812698610126, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a spherical particle of Tieguanyin tea?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the particle's surface covered with intricate textures and subtle wrinkles rather than being smooth?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the particle display a dark green base with pale yellow-green highlights and deep shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a blurred sea of countless similarly scattered tea particles creating a directional trajectory?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera capture the particle rushing towards the lens from the depths and then flying out to the side?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.06578949093818665 + ], + "opposite_logprob_tokens": [ + -2.9407894611358643 + ], + "logprob_reward": -0.06578949093818665, + "yes_no_margin_reward": 2.875, + "token_credit_reward": -0.06578949093818665, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12318412959575653, + 0.10546969622373581, + 0.036974091082811356, + 0.08688604086637497, + 0.03209947794675827, + 0.05135563760995865, + 0.09146542102098465, + 0.059816788882017136, + 0.31831493973731995, + 0.09443371742963791 + ], + "credit_map": [ + 0.013838659971952438, + 0.010294529609382153, + 0.018841026350855827, + 0.01692754589021206, + 0.01383547205477953, + 0.01923605427145958, + 0.01732298731803894, + 0.012887862510979176, + 0.009369875304400921, + 0.006693262606859207, + 0.03275955840945244, + 0.010835758410394192, + 0.0071581569500267506, + 0.011249203234910965, + 0.020774366334080696, + 0.006629520561546087, + 0.0035782530903816223, + 0.0025040653999894857, + 0.006211573723703623, + 0.0060379658825695515, + 0.002314451616257429, + 0.006050285417586565, + 0.0040687378495931625, + 0.006208760663866997, + 0.006018093321472406, + 0.005933799315243959, + 0.027608750388026237, + 0.008204602636396885, + 0.005942337214946747, + 0.01658591255545616, + 0.006795526947826147, + 0.009797033853828907, + 0.0028955936431884766, + 0.0027457273099571466, + 0.005014925729483366, + 0.005116920452564955, + 0.003330254927277565, + 0.0059262653812766075, + 0.0037095160223543644, + 0.003360277507454157, + 0.005166101735085249, + 0.003958228975534439, + 0.008591928519308567, + 0.008018468506634235, + 0.00510946661233902, + 0.009220287203788757, + 0.005115100182592869, + 0.0061760577373206615, + 0.006052019540220499, + 0.007866968400776386, + 0.009053491987287998, + 0.022003553807735443, + 0.009032437577843666, + 0.010912146419286728, + 0.008922649547457695, + 0.017622163519263268, + 0.005036358255892992, + 0.005227431654930115, + 0.009042530320584774, + 0.013014049269258976, + 0.0060830689035356045, + 0.007891963236033916, + 0.006663534324616194, + 0.006857860367745161, + 0.037244051694869995, + 0.042325615882873535, + 0.042220793664455414, + 0.045837968587875366, + 0.05340338498353958, + 0.03100043535232544, + 0.03103575110435486, + 0.03524695709347725, + 0.008043741807341576, + 0.018318800255656242, + 0.006768702995032072, + 0.010189689695835114, + 0.02554253861308098, + 0.010779643431305885, + 0.006791569292545319, + 0.007999034598469734 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 0, + 1 + ], + "selected_reward": -0.06578949093818665 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a spherical particle of Tieguanyin tea?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the particle's surface covered with intricate textures and subtle wrinkles rather than being smooth?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the particle display a dark green base with pale yellow-green highlights and deep shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a blurred sea of countless similarly scattered tea particles creating a directional trajectory?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera capture the particle rushing towards the lens from the depths and then flying out to the side?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0036256073508411646 + ], + "opposite_logprob_tokens": [ + -5.628625392913818 + ], + "logprob_reward": -0.0036256073508411646, + "yes_no_margin_reward": 5.625, + "token_credit_reward": -0.0036256073508411646, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1656627506017685, + 0.30204737186431885, + 0.03136489912867546, + 0.019094469025731087, + 0.018881548196077347, + 0.02053433656692505, + 0.04784267395734787, + 0.32699474692344666, + 0.03340364620089531, + 0.03417356312274933 + ], + "credit_map": [ + 0.014533334411680698, + 0.014059902168810368, + 0.024070141837000847, + 0.022898780182003975, + 0.024264205247163773, + 0.025048958137631416, + 0.021709226071834564, + 0.0190782081335783, + 0.02572658285498619, + 0.014660719782114029, + 0.09854037314653397, + 0.030979547649621964, + 0.01604599505662918, + 0.03066159598529339, + 0.06604461371898651, + 0.01938796043395996, + 0.0035293863620609045, + 0.003055079374462366, + 0.004458119161427021, + 0.00528078293427825, + 0.0030521342996507883, + 0.004436072427779436, + 0.0031598317436873913, + 0.004393491428345442, + 0.001574763678945601, + 0.0011870330199599266, + 0.0058736479841172695, + 0.0022307068575173616, + 0.001164484303444624, + 0.0024142556358128786, + 0.0015841275453567505, + 0.0030654517468065023, + 0.0017387268599122763, + 0.002318104961887002, + 0.0018319558585062623, + 0.003247683634981513, + 0.0026252137031406164, + 0.002798094181343913, + 0.0019876838196069, + 0.0023340857587754726, + 0.002324779285117984, + 0.0018344041891396046, + 0.0045742494985461235, + 0.0025703515857458115, + 0.0023071584291756153, + 0.0029056882485747337, + 0.001966570969671011, + 0.00205113529227674, + 0.002342009451240301, + 0.004188239108771086, + 0.00348937907256186, + 0.012859463691711426, + 0.004902461543679237, + 0.003924272488802671, + 0.004382486920803785, + 0.011754360049962997, + 0.022768259048461914, + 0.04209181293845177, + 0.027554918080568314, + 0.06522440910339355, + 0.050368551164865494, + 0.037656694650650024, + 0.0379086434841156, + 0.043421465903520584, + 0.0024461844004690647, + 0.00410398468375206, + 0.005224164575338364, + 0.004105060826987028, + 0.004714286886155605, + 0.004188730847090483, + 0.004421658348292112, + 0.004199576564133167, + 0.0026339732576161623, + 0.004850998520851135, + 0.003797265002503991, + 0.003764843801036477, + 0.006476749666035175, + 0.006674096453934908, + 0.002112040761858225, + 0.003863594029098749 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 1, + 0 + ], + "selected_reward": -0.0036256073508411646 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a spherical particle of Tieguanyin tea?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the particle's surface covered with intricate textures and subtle wrinkles rather than being smooth?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the particle display a dark green base with pale yellow-green highlights and deep shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a blurred sea of countless similarly scattered tea particles creating a directional trajectory?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera capture the particle rushing towards the lens from the depths and then flying out to the side?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0024924185127019882 + ], + "opposite_logprob_tokens": [ + -6.002492427825928 + ], + "logprob_reward": -0.0024924185127019882, + "yes_no_margin_reward": 6.0, + "token_credit_reward": -0.0024924185127019882, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.04756997153162956, + 0.0842788890004158, + 0.025461232289671898, + 0.24105191230773926, + 0.01670278050005436, + 0.011925533413887024, + 0.0624452605843544, + 0.38587692379951477, + 0.06937836110591888, + 0.055309075862169266 + ], + "credit_map": [ + 0.005129520781338215, + 0.004798922222107649, + 0.005223497282713652, + 0.007998322136700153, + 0.0059483954682946205, + 0.0060937306843698025, + 0.00711720110848546, + 0.0052603851072490215, + 0.007237030193209648, + 0.004195402842015028, + 0.028038034215569496, + 0.008221179246902466, + 0.004410159308463335, + 0.009009972214698792, + 0.018186798319220543, + 0.0049803173169493675, + 0.0027470807544887066, + 0.0020410676952451468, + 0.004631997551769018, + 0.003239196492359042, + 0.0016623293049633503, + 0.0052096666768193245, + 0.002697467803955078, + 0.003232428338378668, + 0.010685812681913376, + 0.004694490693509579, + 0.07628097385168076, + 0.0157464649528265, + 0.005586053244769573, + 0.08932899683713913, + 0.01748824492096901, + 0.02124088443815708, + 0.001681418507359922, + 0.0017317452002316713, + 0.0019357142737135291, + 0.002418959978967905, + 0.0013747926568612456, + 0.0031232195906341076, + 0.0016424619825556874, + 0.002794469939544797, + 0.0012228265404701233, + 0.0012340067187324166, + 0.001446554553695023, + 0.0016170776216313243, + 0.0012647559633478522, + 0.001901888637803495, + 0.0014435868943110108, + 0.0017948378808796406, + 0.004629623144865036, + 0.0044387890957295895, + 0.006125548388808966, + 0.016552763059735298, + 0.004082364961504936, + 0.005480223800987005, + 0.005515935365110636, + 0.015620015561580658, + 0.026090126484632492, + 0.050054896622896194, + 0.033074937760829926, + 0.07837080210447311, + 0.05963640660047531, + 0.044587746262550354, + 0.044409655034542084, + 0.049652379006147385, + 0.006459864322096109, + 0.00971541740000248, + 0.010209331288933754, + 0.008219041861593723, + 0.0095479441806674, + 0.008410404436290264, + 0.008300384506583214, + 0.008515974506735802, + 0.0029354803264141083, + 0.009002427570521832, + 0.004143287427723408, + 0.005941235926002264, + 0.011813495308160782, + 0.008909659460186958, + 0.004593485500663519, + 0.007970008999109268 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 3, + 1 + ], + "selected_reward": -0.0024924185127019882 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a spherical particle of Tieguanyin tea?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the particle's surface covered with intricate textures and subtle wrinkles rather than being smooth?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the particle display a dark green base with pale yellow-green highlights and deep shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a blurred sea of countless similarly scattered tea particles creating a directional trajectory?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera capture the particle rushing towards the lens from the depths and then flying out to the side?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.06999792158603668 + ], + "opposite_logprob_tokens": [ + -2.694998025894165 + ], + "logprob_reward": -0.06999792158603668, + "yes_no_margin_reward": 2.625, + "token_credit_reward": -0.06999792158603668, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.16764366626739502, + 0.35171690583229065, + 0.07542484998703003, + 0.16285261511802673, + 0.05012580007314682, + 0.022936590015888214, + 0.05364816635847092, + 0.0819651186466217, + 0.019532732665538788, + 0.014153538271784782 + ], + "credit_map": [ + 0.01559774111956358, + 0.014800664968788624, + 0.02349969930946827, + 0.022616805508732796, + 0.026028726249933243, + 0.023412086069583893, + 0.02169567346572876, + 0.019992265850305557, + 0.03063398413360119, + 0.020389636978507042, + 0.11303450912237167, + 0.036014460027217865, + 0.02106340415775776, + 0.03865398094058037, + 0.06996231526136398, + 0.02196457050740719, + 0.007538513280451298, + 0.008682038635015488, + 0.011463157832622528, + 0.012059656903147697, + 0.008153638802468777, + 0.010970101691782475, + 0.007603115402162075, + 0.008954625576734543, + 0.007611122447997332, + 0.003448309376835823, + 0.05480295419692993, + 0.011990922503173351, + 0.0044072517193853855, + 0.058120403438806534, + 0.012850700877606869, + 0.009620939381420612, + 0.005272307433187962, + 0.005973863881081343, + 0.004880203865468502, + 0.007501148618757725, + 0.006529451347887516, + 0.008373181335628033, + 0.005143855698406696, + 0.00645178509876132, + 0.0024678322952240705, + 0.001683040987700224, + 0.005417022854089737, + 0.0033718047197908163, + 0.001709338161163032, + 0.0035372290294617414, + 0.0018737755017355084, + 0.002876546001061797, + 0.0027158053126186132, + 0.004532076884061098, + 0.005070962943136692, + 0.01484639197587967, + 0.004868207033723593, + 0.005918745882809162, + 0.005359413102269173, + 0.010336562059819698, + 0.0057374597527086735, + 0.010206038132309914, + 0.007485777605324984, + 0.016814734786748886, + 0.011833042837679386, + 0.009620513767004013, + 0.009495805017650127, + 0.010771742090582848, + 0.002379147568717599, + 0.002676701173186302, + 0.002322947606444359, + 0.0027145680505782366, + 0.0032785406801849604, + 0.0018839644035324454, + 0.0018823196878656745, + 0.0023945404682308435, + 0.000973211950622499, + 0.001161930849775672, + 0.0018597051966935396, + 0.0014676593709737062, + 0.0015716826310381293, + 0.003250607056543231, + 0.0011527552269399166, + 0.002715985756367445 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 3 + ], + "selected_reward": -0.06999792158603668 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a spherical particle of Tieguanyin tea?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the particle's surface covered with intricate textures and subtle wrinkles rather than being smooth?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the particle display a dark green base with pale yellow-green highlights and deep shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a blurred sea of countless similarly scattered tea particles creating a directional trajectory?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera capture the particle rushing towards the lens from the depths and then flying out to the side?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.28128519654273987 + ], + "opposite_logprob_tokens": [ + -1.4062851667404175 + ], + "logprob_reward": -0.28128519654273987, + "yes_no_margin_reward": 1.125, + "token_credit_reward": -0.28128519654273987, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07334048300981522, + 0.26763683557510376, + 0.03619758039712906, + 0.24481074512004852, + 0.020104385912418365, + 0.029136834666132927, + 0.10857396572828293, + 0.14083679020404816, + 0.05464210733771324, + 0.024720313027501106 + ], + "credit_map": [ + 0.00658481614664197, + 0.006966037210077047, + 0.009542376734316349, + 0.012479972094297409, + 0.00918606761842966, + 0.008536895737051964, + 0.008870300836861134, + 0.011174016632139683, + 0.022809073328971863, + 0.016246601939201355, + 0.08379394561052322, + 0.02877678908407688, + 0.017199404537677765, + 0.027521628886461258, + 0.05372180417180061, + 0.017567569389939308, + 0.0023285825736820698, + 0.006122000515460968, + 0.0034999812487512827, + 0.004876190796494484, + 0.005695743020623922, + 0.005688002798706293, + 0.003676327411085367, + 0.004310751799494028, + 0.014664046466350555, + 0.0076897540129721165, + 0.09664993733167648, + 0.02386600337922573, + 0.006600257474929094, + 0.05559523031115532, + 0.01729462295770645, + 0.022450914606451988, + 0.0016867859521880746, + 0.002662970218807459, + 0.001841047778725624, + 0.003387824399396777, + 0.002730573993176222, + 0.0036140596494078636, + 0.0019294384401291609, + 0.002251686993986368, + 0.0035097994841635227, + 0.0018036500550806522, + 0.007017347030341625, + 0.0036294700112193823, + 0.001875736634247005, + 0.005261949263513088, + 0.0026839959900826216, + 0.003354886081069708, + 0.006317704916000366, + 0.007533014751970768, + 0.012424742802977562, + 0.02977263368666172, + 0.007687812205404043, + 0.011820153333246708, + 0.011926069855690002, + 0.02109183929860592, + 0.009372860193252563, + 0.017786851152777672, + 0.012624084949493408, + 0.028963793069124222, + 0.020841071382164955, + 0.016532208770513535, + 0.016626860946416855, + 0.018089070916175842, + 0.006929950788617134, + 0.007214355748146772, + 0.007660668343305588, + 0.0071422806940972805, + 0.008110333234071732, + 0.005949641112238169, + 0.005046495236456394, + 0.0065883854404091835, + 0.0017401077784597874, + 0.0021517807617783546, + 0.003277298528701067, + 0.002524660900235176, + 0.002688951790332794, + 0.00549734104424715, + 0.0017805240349844098, + 0.005059646908193827 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 3, + 7 + ], + "selected_reward": -0.28128519654273987 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "An extreme close-up shot, captured using high-speed photography, closely tracks a spherical particle of Tieguanyin tea being released from the air. Tumbling through the air at high speed, the particle's surface is not smooth, but rather covered with intricate textures and subtle wrinkles resulting from its tightly curled shape. Its dark green base is interspersed with subtle dark shadows and pale yellow-green highlights. Bright, focused studio lighting, directed from above and sideways, creates sharp highlights and deep shadows on the particle's curved surface, dramatically enhancing its three-dimensionality and texture. Surrounding it, the background is a blur of countless similarly scattered tea particles. With a very shallow depth of field, these particles transform into a blurred, flowing sea of dark green light, forming a dynamic, directional trajectory with the main particle. The camera uses an impactful face-to-face shot effect to capture the dynamic moment when the tea leaf suddenly rushes towards the lens from the depths of the picture, almost brushing the lens, and then quickly flies out of the field of vision to the side. The whole process is full of extreme dynamic beauty and strong visual impact, as if time is frozen in this one thousandth of a second, creating a focused and powerful atmosphere.", + "seed": 16270, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000006/group_1/index_7097_seed_16270_dp_2_rank_4_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 1, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.05903882356360555, + "vlm_logprob": -0.05903882356360555, + "vlm_yes_no_margin": 4.05, + "vlm_token_credit": -0.05903882356360555, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a spherical particle of Tieguanyin tea?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the particle's surface covered with intricate textures and subtle wrinkles rather than being smooth?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the particle display a dark green base with pale yellow-green highlights and deep shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a blurred sea of countless similarly scattered tea particles creating a directional trajectory?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera capture the particle rushing towards the lens from the depths and then flying out to the side?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.026506412774324417 + ], + "opposite_logprob_tokens": [ + -3.9015064239501953 + ], + "logprob_reward": -0.026506412774324417, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.026506412774324417, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1684008538722992, + 0.0868685245513916, + 0.03965812548995018, + 0.03115883097052574, + 0.06209040805697441, + 0.13643869757652283, + 0.14866746962070465, + 0.14025767147541046, + 0.11861759424209595, + 0.06784185767173767 + ], + "credit_map": [ + 0.01648142747581005, + 0.01439325138926506, + 0.025236280634999275, + 0.031638965010643005, + 0.017041945829987526, + 0.009263066574931145, + 0.027942728251218796, + 0.026403186842799187, + 0.010167351923882961, + 0.01051343884319067, + 0.012632954865694046, + 0.014291028492152691, + 0.00883454643189907, + 0.0052563901990652084, + 0.01558496244251728, + 0.009587856940925121, + 0.0040021068416535854, + 0.0056457119062542915, + 0.0051421779207885265, + 0.005796030163764954, + 0.004578087013214827, + 0.0023830565623939037, + 0.006652031559497118, + 0.0054589249193668365, + 0.004163335543125868, + 0.005900869611650705, + 0.004423791542649269, + 0.004263430833816528, + 0.003493202617391944, + 0.0015146308578550816, + 0.004134287126362324, + 0.00326528656296432, + 0.009834849275648594, + 0.009073327295482159, + 0.009391585364937782, + 0.007648829836398363, + 0.006730593275278807, + 0.0041524083353579044, + 0.006923666689544916, + 0.008335145190358162, + 0.014846871607005596, + 0.009208664298057556, + 0.023359257727861404, + 0.01318951416760683, + 0.007220791187137365, + 0.00989503413438797, + 0.011866782791912556, + 0.04685176536440849, + 0.011800063773989677, + 0.021963920444250107, + 0.017919139936566353, + 0.022325821220874786, + 0.019122524186968803, + 0.008513002656400204, + 0.027375904843211174, + 0.019647080451250076, + 0.009742612950503826, + 0.018076298758387566, + 0.014115573838353157, + 0.01889250986278057, + 0.03515039384365082, + 0.009478098712861538, + 0.019981518387794495, + 0.014820670709013939, + 0.014187488704919815, + 0.01744067668914795, + 0.015430913306772709, + 0.012374812737107277, + 0.018965834751725197, + 0.006832044571638107, + 0.01689949817955494, + 0.01648632436990738, + 0.008828955702483654, + 0.007399100344628096, + 0.008017714135348797, + 0.007849892601370811, + 0.00747278006747365, + 0.01296419557183981, + 0.007359790615737438, + 0.00794942956417799 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 6, + 7 + ], + "selected_reward": -0.026506412774324417 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a spherical particle of Tieguanyin tea?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the particle's surface covered with intricate textures and subtle wrinkles rather than being smooth?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the particle display a dark green base with pale yellow-green highlights and deep shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a blurred sea of countless similarly scattered tea particles creating a directional trajectory?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera capture the particle rushing towards the lens from the depths and then flying out to the side?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.00362691399641335 + ], + "opposite_logprob_tokens": [ + -5.628626823425293 + ], + "logprob_reward": -0.00362691399641335, + "yes_no_margin_reward": 5.625, + "token_credit_reward": -0.00362691399641335, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.21321804821491241, + 0.07496155798435211, + 0.06481719762086868, + 0.091302290558815, + 0.06292563676834106, + 0.052093978971242905, + 0.12097504734992981, + 0.06017014756798744, + 0.16452166438102722, + 0.09501437097787857 + ], + "credit_map": [ + 0.01391201838850975, + 0.027840202674269676, + 0.027005519717931747, + 0.048856232315301895, + 0.022773800417780876, + 0.011532974429428577, + 0.03237472474575043, + 0.02892257459461689, + 0.006881539709866047, + 0.011231782846152782, + 0.011739050969481468, + 0.010202610865235329, + 0.009045867249369621, + 0.004301941022276878, + 0.009888297878205776, + 0.01167046558111906, + 0.006835337728261948, + 0.009087895974516869, + 0.010255642235279083, + 0.00851633120328188, + 0.006107973866164684, + 0.003866741666570306, + 0.009503981098532677, + 0.010643293149769306, + 0.00902131013572216, + 0.013366635888814926, + 0.011393003165721893, + 0.013644077815115452, + 0.009494622237980366, + 0.004557149950414896, + 0.013202447444200516, + 0.01662304811179638, + 0.00675604771822691, + 0.008204340934753418, + 0.009914535097777843, + 0.007406529504805803, + 0.006567004602402449, + 0.0037435900885611773, + 0.005634814966470003, + 0.014698773622512817, + 0.005153815262019634, + 0.008668904192745686, + 0.006513228639960289, + 0.008692802861332893, + 0.006214556284248829, + 0.003140924032777548, + 0.007241427898406982, + 0.006468318402767181, + 0.015039808116853237, + 0.014378408901393414, + 0.017530221492052078, + 0.015487266704440117, + 0.012551906518638134, + 0.006699963007122278, + 0.01827915385365486, + 0.0210083220154047, + 0.005404745694249868, + 0.006070051342248917, + 0.008186573162674904, + 0.00889906007796526, + 0.007868415676057339, + 0.004171388223767281, + 0.009814378805458546, + 0.009755536913871765, + 0.015174658969044685, + 0.023349905386567116, + 0.01603364758193493, + 0.021336736157536507, + 0.02901027724146843, + 0.011765695177018642, + 0.02648744359612465, + 0.02136329561471939, + 0.010322077199816704, + 0.011509314179420471, + 0.01217776257544756, + 0.011933334171772003, + 0.010453264228999615, + 0.010110155679285526, + 0.011261544190347195, + 0.017246916890144348 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 6 + ], + "selected_reward": -0.00362691399641335 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a spherical particle of Tieguanyin tea?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the particle's surface covered with intricate textures and subtle wrinkles rather than being smooth?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the particle display a dark green base with pale yellow-green highlights and deep shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a blurred sea of countless similarly scattered tea particles creating a directional trajectory?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera capture the particle rushing towards the lens from the depths and then flying out to the side?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0015161458868533373 + ], + "opposite_logprob_tokens": [ + -6.501516342163086 + ], + "logprob_reward": -0.0015161458868533373, + "yes_no_margin_reward": 6.5, + "token_credit_reward": -0.0015161458868533373, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2854650318622589, + 0.08521202951669693, + 0.0401579774916172, + 0.06620059907436371, + 0.05931789055466652, + 0.16770976781845093, + 0.053855013102293015, + 0.09738975018262863, + 0.09251680970191956, + 0.05217522755265236 + ], + "credit_map": [ + 0.03202735632658005, + 0.01952410116791725, + 0.051188062876462936, + 0.050415024161338806, + 0.023630637675523758, + 0.017304053530097008, + 0.03623617812991142, + 0.055139586329460144, + 0.007237134501338005, + 0.010534949600696564, + 0.012213070876896381, + 0.011726006865501404, + 0.009912321344017982, + 0.005326827522367239, + 0.013043255545198917, + 0.015218454413115978, + 0.00490004476159811, + 0.004838975146412849, + 0.006068718153983355, + 0.004921216983348131, + 0.004717076662927866, + 0.002487857360392809, + 0.00466926721855998, + 0.0075548142194747925, + 0.0054975850507617, + 0.011327305808663368, + 0.008932355791330338, + 0.010042553767561913, + 0.00705290911719203, + 0.003161132801324129, + 0.008719590492546558, + 0.011467157863080502, + 0.0058351499028503895, + 0.009843090549111366, + 0.008436082862317562, + 0.007007802836596966, + 0.007625997997820377, + 0.003943589050322771, + 0.009301841259002686, + 0.007324327249079943, + 0.0186654943972826, + 0.01012971717864275, + 0.028926661238074303, + 0.016239240765571594, + 0.008967872709035873, + 0.01238623820245266, + 0.01446782611310482, + 0.057926710695028305, + 0.006455150432884693, + 0.007545949425548315, + 0.009049656800925732, + 0.006689407397061586, + 0.005440252833068371, + 0.0034371553920209408, + 0.008098999969661236, + 0.007138434797525406, + 0.0065175388008356094, + 0.013424952514469624, + 0.008314436301589012, + 0.012717279605567455, + 0.026598967611789703, + 0.006697239354252815, + 0.013459840789437294, + 0.009659484028816223, + 0.011729954741895199, + 0.011244483292102814, + 0.014490838162600994, + 0.009751621633768082, + 0.010769546963274479, + 0.00594835402444005, + 0.01268445048481226, + 0.01589754782617092, + 0.005274592898786068, + 0.005410726647824049, + 0.008189328946173191, + 0.0056624445132911205, + 0.005543361883610487, + 0.004544539377093315, + 0.005689641460776329, + 0.01186058484017849 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 5, + 7 + ], + "selected_reward": -0.0015161458868533373 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a spherical particle of Tieguanyin tea?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the particle's surface covered with intricate textures and subtle wrinkles rather than being smooth?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the particle display a dark green base with pale yellow-green highlights and deep shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a blurred sea of countless similarly scattered tea particles creating a directional trajectory?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera capture the particle rushing towards the lens from the depths and then flying out to the side?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.06202274188399315 + ], + "opposite_logprob_tokens": [ + -2.8120226860046387 + ], + "logprob_reward": -0.06202274188399315, + "yes_no_margin_reward": 2.75, + "token_credit_reward": -0.06202274188399315, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14851361513137817, + 0.15243305265903473, + 0.11151809990406036, + 0.035380490124225616, + 0.044268976897001266, + 0.06443418562412262, + 0.13993772864341736, + 0.170949786901474, + 0.04566074162721634, + 0.08690329641103745 + ], + "credit_map": [ + 0.013843825086951256, + 0.01158264558762312, + 0.025236451998353004, + 0.026975154876708984, + 0.013854792341589928, + 0.008387400768697262, + 0.0236055850982666, + 0.025027766823768616, + 0.020245885476469994, + 0.016790835186839104, + 0.022339502349495888, + 0.019205641001462936, + 0.014027826488018036, + 0.009845498949289322, + 0.019003795459866524, + 0.030974073335528374, + 0.017340287566184998, + 0.0121354591101408, + 0.016659222543239594, + 0.016725648194551468, + 0.008986241184175014, + 0.0070689027197659016, + 0.013892222195863724, + 0.018710115924477577, + 0.005782052408903837, + 0.004941860213875771, + 0.00519210658967495, + 0.004271430429071188, + 0.004197877831757069, + 0.0021823905408382416, + 0.004833039827644825, + 0.003979734610766172, + 0.007503849919885397, + 0.006840439047664404, + 0.005509589798748493, + 0.005673674866557121, + 0.005802900064736605, + 0.0027335770428180695, + 0.005326831713318825, + 0.004878111183643341, + 0.008475874550640583, + 0.005444727372378111, + 0.009650521911680698, + 0.007936620153486729, + 0.005247277673333883, + 0.004327990580350161, + 0.006718548480421305, + 0.016632623970508575, + 0.01904882863163948, + 0.011697856709361076, + 0.0253989826887846, + 0.018765199929475784, + 0.014786063693463802, + 0.008814658038318157, + 0.017581472173333168, + 0.02384466491639614, + 0.01557338610291481, + 0.01843208074569702, + 0.024945290759205818, + 0.02227349393069744, + 0.03229367360472679, + 0.010508407838642597, + 0.02544029802083969, + 0.021483169868588448, + 0.007093300111591816, + 0.004895547870546579, + 0.006055204197764397, + 0.004963545594364405, + 0.0049530332908034325, + 0.0032647675834596157, + 0.005738875363022089, + 0.008696469478309155, + 0.00875492300838232, + 0.011435117572546005, + 0.011716900393366814, + 0.013052756898105145, + 0.011778527870774269, + 0.005971899256110191, + 0.015141633339226246, + 0.00905153900384903 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 1, + 0 + ], + "selected_reward": -0.06202274188399315 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a spherical particle of Tieguanyin tea?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the particle's surface covered with intricate textures and subtle wrinkles rather than being smooth?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the particle display a dark green base with pale yellow-green highlights and deep shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a blurred sea of countless similarly scattered tea particles creating a directional trajectory?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera capture the particle rushing towards the lens from the depths and then flying out to the side?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.20152190327644348 + ], + "opposite_logprob_tokens": [ + -1.701521873474121 + ], + "logprob_reward": -0.20152190327644348, + "yes_no_margin_reward": 1.5, + "token_credit_reward": -0.20152190327644348, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1776910424232483, + 0.07827802747488022, + 0.1104438304901123, + 0.06076882407069206, + 0.06083277240395546, + 0.13500995934009552, + 0.09118098765611649, + 0.12587067484855652, + 0.056770239025354385, + 0.10315363854169846 + ], + "credit_map": [ + 0.021299393847584724, + 0.017682809382677078, + 0.020022263750433922, + 0.027554593980312347, + 0.024165045469999313, + 0.012464180588722229, + 0.03285308554768562, + 0.02164965495467186, + 0.009354328736662865, + 0.008812680840492249, + 0.010779615491628647, + 0.012597132474184036, + 0.009770386852324009, + 0.005061169620603323, + 0.013497717678546906, + 0.00840499997138977, + 0.01503853127360344, + 0.011740687303245068, + 0.015526865608990192, + 0.01642671599984169, + 0.00944273266941309, + 0.0074590048752725124, + 0.012777375988662243, + 0.022031916305422783, + 0.007606695871800184, + 0.007317688781768084, + 0.009324342012405396, + 0.007627153303474188, + 0.008399460464715958, + 0.0035658201668411493, + 0.008607573807239532, + 0.008320087566971779, + 0.009611746296286583, + 0.009226116351783276, + 0.008298544213175774, + 0.00835275836288929, + 0.007052547764033079, + 0.003423751797527075, + 0.007909503765404224, + 0.006957801524549723, + 0.014621616341173649, + 0.009006760083138943, + 0.02311493642628193, + 0.014883486554026604, + 0.008503378368914127, + 0.009826665744185448, + 0.01169193722307682, + 0.043361175805330276, + 0.013652666471898556, + 0.010165673680603504, + 0.013974013738334179, + 0.012555438093841076, + 0.009898816235363483, + 0.005001853685826063, + 0.012050842866301537, + 0.013881689868867397, + 0.009768350049853325, + 0.01585145853459835, + 0.015166257508099079, + 0.017540106549859047, + 0.02731597237288952, + 0.008258289657533169, + 0.01959078200161457, + 0.012379460968077183, + 0.007522217463701963, + 0.006044063251465559, + 0.007230408955365419, + 0.006564674433320761, + 0.005634321365505457, + 0.00345289777033031, + 0.007819095626473427, + 0.012502556666731834, + 0.011490555480122566, + 0.016032624989748, + 0.014075378887355328, + 0.015994280576705933, + 0.012020707130432129, + 0.006633542012423277, + 0.015091116540133953, + 0.011815441772341728 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 5, + 7 + ], + "selected_reward": -0.20152190327644348 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "An extreme close-up shot, captured using high-speed photography, closely tracks a spherical particle of Tieguanyin tea being released from the air. Tumbling through the air at high speed, the particle's surface is not smooth, but rather covered with intricate textures and subtle wrinkles resulting from its tightly curled shape. Its dark green base is interspersed with subtle dark shadows and pale yellow-green highlights. Bright, focused studio lighting, directed from above and sideways, creates sharp highlights and deep shadows on the particle's curved surface, dramatically enhancing its three-dimensionality and texture. Surrounding it, the background is a blur of countless similarly scattered tea particles. With a very shallow depth of field, these particles transform into a blurred, flowing sea of dark green light, forming a dynamic, directional trajectory with the main particle. The camera uses an impactful face-to-face shot effect to capture the dynamic moment when the tea leaf suddenly rushes towards the lens from the depths of the picture, almost brushing the lens, and then quickly flies out of the field of vision to the side. The whole process is full of extreme dynamic beauty and strong visual impact, as if time is frozen in this one thousandth of a second, creating a focused and powerful atmosphere.", + "seed": 16271, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000006/group_1/index_7097_seed_16271_dp_3_rank_6_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 1, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.017830786923877896, + "vlm_logprob": -0.017830786923877896, + "vlm_yes_no_margin": 4.6, + "vlm_token_credit": -0.017830786923877896, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a spherical particle of Tieguanyin tea?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the particle's surface covered with intricate textures and subtle wrinkles rather than being smooth?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the particle display a dark green base with pale yellow-green highlights and deep shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a blurred sea of countless similarly scattered tea particles creating a directional trajectory?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera capture the particle rushing towards the lens from the depths and then flying out to the side?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03841247782111168 + ], + "opposite_logprob_tokens": [ + -3.538412570953369 + ], + "logprob_reward": -0.03841247782111168, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.03841247782111168, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15251803398132324, + 0.08940853923559189, + 0.08671226352453232, + 0.43640685081481934, + 0.0659184381365776, + 0.05311129242181778, + 0.04193400219082832, + 0.029242422431707382, + 0.02084442228078842, + 0.02390380948781967 + ], + "credit_map": [ + 0.014675509184598923, + 0.01655225083231926, + 0.01533535122871399, + 0.02564144693315029, + 0.025273913517594337, + 0.012605053372681141, + 0.022466991096735, + 0.01996750198304653, + 0.006832982879132032, + 0.007890947163105011, + 0.011968401260674, + 0.007024206221103668, + 0.013183347880840302, + 0.016101373359560966, + 0.01023044902831316, + 0.016176821663975716, + 0.00759621849283576, + 0.005845655221492052, + 0.019402842968702316, + 0.008522900752723217, + 0.006009662989526987, + 0.017385683953762054, + 0.006368839647620916, + 0.015580441802740097, + 0.036579009145498276, + 0.018331089988350868, + 0.03727887198328972, + 0.03231155127286911, + 0.030462190508842468, + 0.16625840961933136, + 0.04762690141797066, + 0.06755878031253815, + 0.004331029020249844, + 0.007557784207165241, + 0.006794353481382132, + 0.006554102059453726, + 0.008031882345676422, + 0.009249348193407059, + 0.010983405634760857, + 0.012416521087288857, + 0.0038303560577332973, + 0.006582616828382015, + 0.007453543599694967, + 0.004757438320666552, + 0.005824625957757235, + 0.006788506638258696, + 0.01116555742919445, + 0.006708639208227396, + 0.0036750435829162598, + 0.003225658554583788, + 0.0029714861884713173, + 0.004666893742978573, + 0.002672565868124366, + 0.009396643377840519, + 0.003415249288082123, + 0.011910459026694298, + 0.00343745993450284, + 0.0030297073535621166, + 0.0035476721823215485, + 0.0034739612601697445, + 0.005350645165890455, + 0.003276458941400051, + 0.0035496107302606106, + 0.0035769029054790735, + 0.0044171931222081184, + 0.003290712833404541, + 0.0021875000093132257, + 0.003671063808724284, + 0.0015853127697482705, + 0.0018276741029694676, + 0.0014216703129932284, + 0.002443291712552309, + 0.0032587999012321234, + 0.004250878933817148, + 0.0025348926428705454, + 0.003102092305198312, + 0.0022506914101541042, + 0.002335682976990938, + 0.002221236703917384, + 0.003949533216655254 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 0, + 1 + ], + "selected_reward": -0.03841247782111168 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a spherical particle of Tieguanyin tea?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the particle's surface covered with intricate textures and subtle wrinkles rather than being smooth?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the particle display a dark green base with pale yellow-green highlights and deep shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a blurred sea of countless similarly scattered tea particles creating a directional trajectory?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera capture the particle rushing towards the lens from the depths and then flying out to the side?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005256284959614277 + ], + "opposite_logprob_tokens": [ + -5.255256175994873 + ], + "logprob_reward": -0.005256284959614277, + "yes_no_margin_reward": 5.25, + "token_credit_reward": -0.005256284959614277, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.30583426356315613, + 0.17642083764076233, + 0.09684260189533234, + 0.16567625105381012, + 0.059908702969551086, + 0.08707045763731003, + 0.03523264825344086, + 0.03204791620373726, + 0.023782311007380486, + 0.017183994874358177 + ], + "credit_map": [ + 0.023390725255012512, + 0.03519606590270996, + 0.03158438205718994, + 0.04719046503305435, + 0.047585442662239075, + 0.03526100888848305, + 0.05049661919474602, + 0.035129573196172714, + 0.01513280626386404, + 0.02253374643623829, + 0.021194957196712494, + 0.023484619334340096, + 0.025136882439255714, + 0.020011570304632187, + 0.028746258467435837, + 0.020179999992251396, + 0.012101415544748306, + 0.010384645313024521, + 0.016783274710178375, + 0.01053616777062416, + 0.009636107832193375, + 0.017056699842214584, + 0.009065422229468822, + 0.011278873309493065, + 0.009841367602348328, + 0.005495362915098667, + 0.011644070036709309, + 0.009294731542468071, + 0.00939091108739376, + 0.06454099714756012, + 0.015490305610001087, + 0.0399785116314888, + 0.003476653713732958, + 0.0064248559065163136, + 0.004338408820331097, + 0.008068772964179516, + 0.009383116848766804, + 0.0071938796900212765, + 0.010337886400520802, + 0.010685130022466183, + 0.0062462808564305305, + 0.009798812679946423, + 0.00964739266782999, + 0.012946457602083683, + 0.011517415754497051, + 0.013321032747626305, + 0.01271608006209135, + 0.010876987129449844, + 0.003895254572853446, + 0.005675706081092358, + 0.0034126832615584135, + 0.0058417427353560925, + 0.002768521197140217, + 0.003725738963112235, + 0.004104601684957743, + 0.005808401852846146, + 0.003045499324798584, + 0.002196267247200012, + 0.0051631261594593525, + 0.004131962079554796, + 0.004966665059328079, + 0.00479107117280364, + 0.0032851588912308216, + 0.004468168132007122, + 0.004597845952957869, + 0.004008419346064329, + 0.002832991536706686, + 0.0035536387003958225, + 0.0017825090326368809, + 0.002308053895831108, + 0.001566189108416438, + 0.0031326627358794212, + 0.002003244822844863, + 0.002806885400786996, + 0.002620718674734235, + 0.0024319766089320183, + 0.0020717126317322254, + 0.0013330356450751424, + 0.0018454082310199738, + 0.0020710129756480455 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 3 + ], + "selected_reward": -0.005256284959614277 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a spherical particle of Tieguanyin tea?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the particle's surface covered with intricate textures and subtle wrinkles rather than being smooth?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the particle display a dark green base with pale yellow-green highlights and deep shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a blurred sea of countless similarly scattered tea particles creating a directional trajectory?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera capture the particle rushing towards the lens from the depths and then flying out to the side?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0015148365637287498 + ], + "opposite_logprob_tokens": [ + -6.501514911651611 + ], + "logprob_reward": -0.0015148365637287498, + "yes_no_margin_reward": 6.5, + "token_credit_reward": -0.0015148365637287498, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.18988309800624847, + 0.053663261234760284, + 0.10418744385242462, + 0.23489327728748322, + 0.06175458803772926, + 0.17429016530513763, + 0.045841511338949203, + 0.03808432072401047, + 0.026182694360613823, + 0.07121962308883667 + ], + "credit_map": [ + 0.016390996053814888, + 0.023785222321748734, + 0.020866848528385162, + 0.025104617699980736, + 0.027666470035910606, + 0.029896961525082588, + 0.024313731119036674, + 0.021858258172869682, + 0.0046537998132407665, + 0.009508349932730198, + 0.00587061420083046, + 0.005899316631257534, + 0.006972691044211388, + 0.0065759168937802315, + 0.008039923384785652, + 0.006142647471278906, + 0.01155497133731842, + 0.008536089211702347, + 0.011472528800368309, + 0.008465378545224667, + 0.007264537271112204, + 0.019441813230514526, + 0.008623365312814713, + 0.028828758746385574, + 0.018090369179844856, + 0.013958810828626156, + 0.02631804347038269, + 0.02076256275177002, + 0.020897258073091507, + 0.07801065593957901, + 0.02660328894853592, + 0.030252287164330482, + 0.003425360657274723, + 0.006401405669748783, + 0.005594986025243998, + 0.009511282667517662, + 0.008917107246816158, + 0.007346792612224817, + 0.010035120882093906, + 0.010522534139454365, + 0.011125867255032063, + 0.018574047833681107, + 0.020278537645936012, + 0.026314832270145416, + 0.020599864423274994, + 0.028438547626137733, + 0.024731522426009178, + 0.02422693930566311, + 0.00479380926117301, + 0.006008075084537268, + 0.006491863168776035, + 0.005109330173581839, + 0.004952752497047186, + 0.005254960153251886, + 0.0073434095829725266, + 0.005887310020625591, + 0.005880215670913458, + 0.00343636074103415, + 0.006465984974056482, + 0.005461558233946562, + 0.003537745214998722, + 0.005240416619926691, + 0.004101565573364496, + 0.003960475325584412, + 0.005059356335550547, + 0.0037646035198122263, + 0.0032387010287493467, + 0.0038152295164763927, + 0.0023659346625208855, + 0.0025328933261334896, + 0.002491486957296729, + 0.002914488548412919, + 0.010047786869108677, + 0.011960227973759174, + 0.007009067106992006, + 0.008332314901053905, + 0.006179856136441231, + 0.008065360598266125, + 0.005395516753196716, + 0.014229491353034973 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 0, + 5 + ], + "selected_reward": -0.0015148365637287498 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a spherical particle of Tieguanyin tea?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the particle's surface covered with intricate textures and subtle wrinkles rather than being smooth?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the particle display a dark green base with pale yellow-green highlights and deep shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a blurred sea of countless similarly scattered tea particles creating a directional trajectory?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera capture the particle rushing towards the lens from the depths and then flying out to the side?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.029784319922327995 + ], + "opposite_logprob_tokens": [ + -3.5297842025756836 + ], + "logprob_reward": -0.029784319922327995, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.029784319922327995, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.28522157669067383, + 0.11069591343402863, + 0.08742710202932358, + 0.24297749996185303, + 0.04692719876766205, + 0.050869252532720566, + 0.0855158343911171, + 0.04215926676988602, + 0.014611675404012203, + 0.03359464555978775 + ], + "credit_map": [ + 0.021564248949289322, + 0.03686200827360153, + 0.026828953996300697, + 0.04292958602309227, + 0.053505171090364456, + 0.0284404419362545, + 0.04297474026679993, + 0.03211642801761627, + 0.009598207660019398, + 0.012798749841749668, + 0.012919683009386063, + 0.009739934466779232, + 0.014847423881292343, + 0.021349424496293068, + 0.009443901479244232, + 0.01999858394265175, + 0.009838780388236046, + 0.005991013254970312, + 0.021089063957333565, + 0.008701620623469353, + 0.01025483664125204, + 0.01417939830571413, + 0.007920562289655209, + 0.009451819583773613, + 0.01408174354583025, + 0.005628571845591068, + 0.01524332631379366, + 0.013087202794849873, + 0.009651045314967632, + 0.10095265507698059, + 0.017523929476737976, + 0.0668090358376503, + 0.0022562462836503983, + 0.0035095764324069023, + 0.005349286831915379, + 0.00578027218580246, + 0.00533512607216835, + 0.005616662558168173, + 0.007776231970638037, + 0.011303798295557499, + 0.0035015142057090998, + 0.00857123825699091, + 0.004169107414782047, + 0.0051514459773898125, + 0.005414231214672327, + 0.009814341552555561, + 0.007560696918517351, + 0.006686674430966377, + 0.008719067089259624, + 0.011184267699718475, + 0.009322261437773705, + 0.013086806051433086, + 0.008793213404715061, + 0.010383224114775658, + 0.011898070573806763, + 0.012128930538892746, + 0.004166707396507263, + 0.0036248858086764812, + 0.004143585450947285, + 0.005986346863210201, + 0.007491931784898043, + 0.006007825024425983, + 0.004254593513906002, + 0.006483391392976046, + 0.002629598369821906, + 0.002022805158048868, + 0.0014612474478781223, + 0.002546835457906127, + 0.0013734904350712895, + 0.0013248634058982134, + 0.0014015138149261475, + 0.001851322129368782, + 0.006728488020598888, + 0.005836906377226114, + 0.003807179629802704, + 0.0031089158728718758, + 0.0027168444357812405, + 0.0033630274701863527, + 0.0020695016719400883, + 0.0059637813828885555 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 1 + ], + "selected_reward": -0.029784319922327995 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a spherical particle of Tieguanyin tea?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the particle's surface covered with intricate textures and subtle wrinkles rather than being smooth?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the particle display a dark green base with pale yellow-green highlights and deep shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a blurred sea of countless similarly scattered tea particles creating a directional trajectory?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera capture the particle rushing towards the lens from the depths and then flying out to the side?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.014186015352606773 + ], + "opposite_logprob_tokens": [ + -4.264185905456543 + ], + "logprob_reward": -0.014186015352606773, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.014186015352606773, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.19221393764019012, + 0.061892665922641754, + 0.16818219423294067, + 0.06262125819921494, + 0.09243857860565186, + 0.062001075595617294, + 0.12027439475059509, + 0.10208360850811005, + 0.0876641497015953, + 0.05062814801931381 + ], + "credit_map": [ + 0.018891669809818268, + 0.022821860387921333, + 0.0211196169257164, + 0.023252686485648155, + 0.02976340427994728, + 0.02969225123524666, + 0.023840826004743576, + 0.0228316318243742, + 0.007672098930925131, + 0.00902640912681818, + 0.00905485637485981, + 0.00807915534824133, + 0.005368394311517477, + 0.008460862562060356, + 0.00648335600271821, + 0.007747533731162548, + 0.03104698285460472, + 0.011884761974215508, + 0.030564021319150925, + 0.017478175461292267, + 0.010920110158622265, + 0.01968318782746792, + 0.017701899632811546, + 0.02890305407345295, + 0.007070057559758425, + 0.004241999238729477, + 0.007597006391733885, + 0.0059514716267585754, + 0.005015787202864885, + 0.007539000827819109, + 0.00851079449057579, + 0.016695141792297363, + 0.009104692377150059, + 0.006609153933823109, + 0.01119379885494709, + 0.008135398849844933, + 0.007774996571242809, + 0.014554421417415142, + 0.010470946319401264, + 0.024595163762569427, + 0.0047945003025233746, + 0.007931774482131004, + 0.0060211364179849625, + 0.006787819787859917, + 0.006162509322166443, + 0.009379220195114613, + 0.01061383169144392, + 0.010310283862054348, + 0.009513944387435913, + 0.018045293167233467, + 0.010456657968461514, + 0.015051573514938354, + 0.011304936371743679, + 0.019976196810603142, + 0.016038652509450912, + 0.01988714188337326, + 0.013136363588273525, + 0.010524203069508076, + 0.015237499959766865, + 0.011978795751929283, + 0.009015257470309734, + 0.015244681388139725, + 0.008607719093561172, + 0.018339091911911964, + 0.018488913774490356, + 0.013035842217504978, + 0.010550366714596748, + 0.01709575019776821, + 0.007159315049648285, + 0.007663808297365904, + 0.006135975942015648, + 0.007534171920269728, + 0.006230491679161787, + 0.008183061145246029, + 0.00563054857775569, + 0.0068635414354503155, + 0.005401551723480225, + 0.005165586248040199, + 0.004297500941902399, + 0.008855867199599743 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 6 + ], + "selected_reward": -0.014186015352606773 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "An extreme close-up shot, captured using high-speed photography, closely tracks a spherical particle of Tieguanyin tea being released from the air. Tumbling through the air at high speed, the particle's surface is not smooth, but rather covered with intricate textures and subtle wrinkles resulting from its tightly curled shape. Its dark green base is interspersed with subtle dark shadows and pale yellow-green highlights. Bright, focused studio lighting, directed from above and sideways, creates sharp highlights and deep shadows on the particle's curved surface, dramatically enhancing its three-dimensionality and texture. Surrounding it, the background is a blur of countless similarly scattered tea particles. With a very shallow depth of field, these particles transform into a blurred, flowing sea of dark green light, forming a dynamic, directional trajectory with the main particle. The camera uses an impactful face-to-face shot effect to capture the dynamic moment when the tea leaf suddenly rushes towards the lens from the depths of the picture, almost brushing the lens, and then quickly flies out of the field of vision to the side. The whole process is full of extreme dynamic beauty and strong visual impact, as if time is frozen in this one thousandth of a second, creating a focused and powerful atmosphere.", + "seed": 16272, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000006/group_1/index_7097_seed_16272_dp_3_rank_6_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 1, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.047123228572309016, + "vlm_logprob": -0.047123228572309016, + "vlm_yes_no_margin": 3.925, + "vlm_token_credit": -0.047123228572309016, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a spherical particle of Tieguanyin tea?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the particle's surface covered with intricate textures and subtle wrinkles rather than being smooth?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the particle display a dark green base with pale yellow-green highlights and deep shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a blurred sea of countless similarly scattered tea particles creating a directional trajectory?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera capture the particle rushing towards the lens from the depths and then flying out to the side?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04668755456805229 + ], + "opposite_logprob_tokens": [ + -3.296687602996826 + ], + "logprob_reward": -0.04668755456805229, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.04668755456805229, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.5072577595710754, + 0.1537197381258011, + 0.023745806887745857, + 0.04971100762486458, + 0.08249996602535248, + 0.030852368101477623, + 0.024891648441553116, + 0.03501186519861221, + 0.045454055070877075, + 0.046855781227350235 + ], + "credit_map": [ + 0.03278415650129318, + 0.011646748520433903, + 0.18117177486419678, + 0.02771545574069023, + 0.013155165128409863, + 0.20418478548526764, + 0.013798185624182224, + 0.022801509127020836, + 0.013559545390307903, + 0.00474113505333662, + 0.06110319867730141, + 0.009613178670406342, + 0.0041337087750434875, + 0.04423772916197777, + 0.008513525128364563, + 0.007817727513611317, + 0.0028158100321888924, + 0.001604020013473928, + 0.004448344465345144, + 0.0027964182663708925, + 0.002381277736276388, + 0.003321747062727809, + 0.003758807433769107, + 0.002619381295517087, + 0.00483139930292964, + 0.003510355018079281, + 0.008134162053465843, + 0.006698747165501118, + 0.004057965707033873, + 0.005962243303656578, + 0.009360087104141712, + 0.007156048435717821, + 0.016635945066809654, + 0.0032482894603163004, + 0.021761944517493248, + 0.01136753149330616, + 0.0031380399595946074, + 0.006882785819470882, + 0.009939577430486679, + 0.009525866247713566, + 0.00596458837389946, + 0.002022404922172427, + 0.006415488664060831, + 0.004070070572197437, + 0.0022921122144907713, + 0.0028484093490988016, + 0.0036893088836222887, + 0.0035499888472259045, + 0.0021123234182596207, + 0.0033751744776964188, + 0.0038898035418242216, + 0.002886771457269788, + 0.0032789609394967556, + 0.0020193271338939667, + 0.004194316919893026, + 0.003134970786049962, + 0.0052335248328745365, + 0.0038394464645534754, + 0.00372177641838789, + 0.004371805116534233, + 0.003474070457741618, + 0.004619226790964603, + 0.0056188879534602165, + 0.004133129026740789, + 0.004786706995218992, + 0.0021594895515590906, + 0.01427440345287323, + 0.0038478635251522064, + 0.0030696881003677845, + 0.003882124787196517, + 0.003566236002370715, + 0.00986754521727562, + 0.005462069064378738, + 0.004005067981779575, + 0.008589598350226879, + 0.004285838454961777, + 0.005700176116079092, + 0.007112350780516863, + 0.006324447691440582, + 0.005376230459660292 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 4 + ], + "selected_reward": -0.04668755456805229 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a spherical particle of Tieguanyin tea?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the particle's surface covered with intricate textures and subtle wrinkles rather than being smooth?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the particle display a dark green base with pale yellow-green highlights and deep shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a blurred sea of countless similarly scattered tea particles creating a directional trajectory?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera capture the particle rushing towards the lens from the depths and then flying out to the side?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.003204451175406575 + ], + "opposite_logprob_tokens": [ + -5.753204345703125 + ], + "logprob_reward": -0.003204451175406575, + "yes_no_margin_reward": 5.75, + "token_credit_reward": -0.003204451175406575, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1834113895893097, + 0.2361302375793457, + 0.0233499426394701, + 0.17078296840190887, + 0.06564156711101532, + 0.11628470569849014, + 0.033797696232795715, + 0.02102033607661724, + 0.037730928510427475, + 0.11185019463300705 + ], + "credit_map": [ + 0.010466289706528187, + 0.005834371317178011, + 0.05191463604569435, + 0.009582724422216415, + 0.007276356220245361, + 0.07327862083911896, + 0.006719478406012058, + 0.018338918685913086, + 0.021249588578939438, + 0.009588408283889294, + 0.09240464866161346, + 0.014598976820707321, + 0.0076150414533913136, + 0.06246389448642731, + 0.014961007982492447, + 0.01324867457151413, + 0.0020427219569683075, + 0.002564535476267338, + 0.002350348047912121, + 0.003844232764095068, + 0.002471144311130047, + 0.0024236897006630898, + 0.0037939606700092554, + 0.0038593101780861616, + 0.013352284207940102, + 0.009719897992908955, + 0.018244585022330284, + 0.03151998668909073, + 0.012845196761190891, + 0.031544946134090424, + 0.03905140981078148, + 0.014504666440188885, + 0.01245090365409851, + 0.0024596478324383497, + 0.01931874454021454, + 0.007936575450003147, + 0.00241544540040195, + 0.0053601013496518135, + 0.007434100843966007, + 0.00826604850590229, + 0.011493696831166744, + 0.007383827585726976, + 0.02676764503121376, + 0.009263666346669197, + 0.007730917073786259, + 0.008363494649529457, + 0.008291145786643028, + 0.03699029982089996, + 0.0029290360398590565, + 0.003828702261671424, + 0.006312903016805649, + 0.003971582278609276, + 0.004742208402603865, + 0.002366675529628992, + 0.006032239180058241, + 0.003614348592236638, + 0.0029359899926930666, + 0.0022142089437693357, + 0.004292200785130262, + 0.0022838159929960966, + 0.0014999774284660816, + 0.0020535828080028296, + 0.0025831968523561954, + 0.0031573630403727293, + 0.003162133041769266, + 0.0012160588521510363, + 0.013583466410636902, + 0.0025789656210690737, + 0.0016965561080724, + 0.0028309556655585766, + 0.001595312962308526, + 0.01106747891753912, + 0.016938183456659317, + 0.009723869152367115, + 0.028291985392570496, + 0.009909914806485176, + 0.0073713515885174274, + 0.009708426892757416, + 0.007419648580253124, + 0.02248680777847767 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 3 + ], + "selected_reward": -0.003204451175406575 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a spherical particle of Tieguanyin tea?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the particle's surface covered with intricate textures and subtle wrinkles rather than being smooth?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the particle display a dark green base with pale yellow-green highlights and deep shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a blurred sea of countless similarly scattered tea particles creating a directional trajectory?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera capture the particle rushing towards the lens from the depths and then flying out to the side?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0028229409363120794 + ], + "opposite_logprob_tokens": [ + -5.8778228759765625 + ], + "logprob_reward": -0.0028229409363120794, + "yes_no_margin_reward": 5.875, + "token_credit_reward": -0.0028229409363120794, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2721051275730133, + 0.21725907921791077, + 0.022504674270749092, + 0.035019759088754654, + 0.1328597515821457, + 0.10852336138486862, + 0.02794516645371914, + 0.04730949550867081, + 0.08649686723947525, + 0.04997667297720909 + ], + "credit_map": [ + 0.01775592379271984, + 0.007042634766548872, + 0.08943306654691696, + 0.014175524935126305, + 0.006766477599740028, + 0.11221619695425034, + 0.01053932961076498, + 0.01417598221451044, + 0.01585794799029827, + 0.006699007004499435, + 0.08015716075897217, + 0.013653933070600033, + 0.007201768457889557, + 0.06982902437448502, + 0.012695252895355225, + 0.011164991185069084, + 0.002483862219378352, + 0.0020008408464491367, + 0.004358063917607069, + 0.002569776028394699, + 0.0028089575935155153, + 0.002426953287795186, + 0.003071935148909688, + 0.002784287091344595, + 0.004335524048656225, + 0.002679342171177268, + 0.007095621898770332, + 0.0039598047733306885, + 0.0024328615982085466, + 0.004173426888883114, + 0.005111520644277334, + 0.0052316575311124325, + 0.02978232502937317, + 0.004730162676423788, + 0.041769839823246, + 0.011398200877010822, + 0.004724235739558935, + 0.011874702759087086, + 0.011642199009656906, + 0.016938084736466408, + 0.013779147528111935, + 0.006339444313198328, + 0.019709227606654167, + 0.007475448306649923, + 0.0064620873890817165, + 0.00959604699164629, + 0.007804086431860924, + 0.0373578704893589, + 0.004240489099174738, + 0.0023570088669657707, + 0.00420231930911541, + 0.0025814329273998737, + 0.0027029102202504873, + 0.002608437556773424, + 0.004667365457862616, + 0.004585204180330038, + 0.00584378931671381, + 0.0035041607916355133, + 0.010106232017278671, + 0.004360746592283249, + 0.0027158993761986494, + 0.006460448727011681, + 0.006983374711126089, + 0.007334844674915075, + 0.008396835997700691, + 0.0038516658823937178, + 0.027713067829608917, + 0.006474901456385851, + 0.005769021809101105, + 0.005651860032230616, + 0.006147648207843304, + 0.02249186486005783, + 0.008407273329794407, + 0.005450317170470953, + 0.0066203526221215725, + 0.006186975631862879, + 0.005228836089372635, + 0.006848610006272793, + 0.006194339599460363, + 0.005039968993514776 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 4 + ], + "selected_reward": -0.0028229409363120794 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a spherical particle of Tieguanyin tea?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the particle's surface covered with intricate textures and subtle wrinkles rather than being smooth?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the particle display a dark green base with pale yellow-green highlights and deep shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a blurred sea of countless similarly scattered tea particles creating a directional trajectory?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera capture the particle rushing towards the lens from the depths and then flying out to the side?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07000036537647247 + ], + "opposite_logprob_tokens": [ + -2.695000410079956 + ], + "logprob_reward": -0.07000036537647247, + "yes_no_margin_reward": 2.625, + "token_credit_reward": -0.07000036537647247, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.45107680559158325, + 0.027464384213089943, + 0.010908585973083973, + 0.03317459672689438, + 0.038044799119234085, + 0.32748132944107056, + 0.008682472631335258, + 0.04076297581195831, + 0.040139030665159225, + 0.022265031933784485 + ], + "credit_map": [ + 0.02722271904349327, + 0.008775033056735992, + 0.1565050631761551, + 0.02394806779921055, + 0.010462885722517967, + 0.19282908737659454, + 0.010660046711564064, + 0.02067387104034424, + 0.0021708367858082056, + 0.0017815344035625458, + 0.005182195920497179, + 0.003038628725335002, + 0.001703033340163529, + 0.007208713795989752, + 0.0036858844105154276, + 0.0026935553178191185, + 0.00098532740958035, + 0.0011770777637138963, + 0.0011801447253674269, + 0.0019930112175643444, + 0.0011176432017236948, + 0.0008952957578003407, + 0.002210958395153284, + 0.001349127385765314, + 0.003294798079878092, + 0.0022971322759985924, + 0.003984484355896711, + 0.0037702450063079596, + 0.0028528941329568624, + 0.007919009774923325, + 0.004546753130853176, + 0.004509279504418373, + 0.00789530761539936, + 0.0019771000370383263, + 0.006059623789042234, + 0.005256205331534147, + 0.0017993187066167593, + 0.003528565401211381, + 0.00483186449855566, + 0.006696807220578194, + 0.03428592532873154, + 0.020323820412158966, + 0.07219722121953964, + 0.026097187772393227, + 0.022686759009957314, + 0.02425885945558548, + 0.023478863760828972, + 0.10415268689393997, + 0.0014181636506691575, + 0.0009098751470446587, + 0.001323638018220663, + 0.0011485269060358405, + 0.0006506236968562007, + 0.0006160052726045251, + 0.0013837837614119053, + 0.0012318558292463422, + 0.006242315750569105, + 0.004107464104890823, + 0.005597327370196581, + 0.0044156769290566444, + 0.003687803167849779, + 0.00524853216484189, + 0.005829124245792627, + 0.005634728819131851, + 0.003717886283993721, + 0.0019263048889115453, + 0.012477663345634937, + 0.003203639294952154, + 0.002187769627198577, + 0.0030659460462629795, + 0.0027711002621799707, + 0.010788720101118088, + 0.0034116460010409355, + 0.0025824301410466433, + 0.004591032862663269, + 0.0020817008335143328, + 0.001870390959084034, + 0.002783983014523983, + 0.002511371858417988, + 0.002432475332170725 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 5, + 7 + ], + "selected_reward": -0.07000036537647247 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a spherical particle of Tieguanyin tea?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the particle's surface covered with intricate textures and subtle wrinkles rather than being smooth?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the particle display a dark green base with pale yellow-green highlights and deep shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a blurred sea of countless similarly scattered tea particles creating a directional trajectory?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera capture the particle rushing towards the lens from the depths and then flying out to the side?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.11290083080530167 + ], + "opposite_logprob_tokens": [ + -2.237900733947754 + ], + "logprob_reward": -0.11290083080530167, + "yes_no_margin_reward": 2.125, + "token_credit_reward": -0.11290083080530167, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.26762959361076355, + 0.04228740558028221, + 0.03837387263774872, + 0.03713011369109154, + 0.08106481283903122, + 0.07412217557430267, + 0.04380146786570549, + 0.08498188853263855, + 0.05564713105559349, + 0.27496159076690674 + ], + "credit_map": [ + 0.01885981298983097, + 0.006833922583609819, + 0.09199352562427521, + 0.015198743902146816, + 0.007297915872186422, + 0.09842600673437119, + 0.009347549639642239, + 0.019672119989991188, + 0.004858844447880983, + 0.003292310517281294, + 0.00813630037009716, + 0.0035973989870399237, + 0.003052927553653717, + 0.01066888589411974, + 0.0050005498342216015, + 0.0036801870446652174, + 0.004087585490196943, + 0.00375612685456872, + 0.00686511117964983, + 0.006767347455024719, + 0.003211866831406951, + 0.00498144282028079, + 0.0047451527789235115, + 0.003959238529205322, + 0.004533783998340368, + 0.0039050858467817307, + 0.006837122142314911, + 0.00450510298833251, + 0.002951845293864608, + 0.00457709701731801, + 0.005233483389019966, + 0.00458659278228879, + 0.019966524094343185, + 0.0029168555047363043, + 0.023595429956912994, + 0.008641952648758888, + 0.0036046721506863832, + 0.005887580569833517, + 0.007869483903050423, + 0.008582310751080513, + 0.011961445212364197, + 0.003992996644228697, + 0.02157818153500557, + 0.006277923472225666, + 0.004202013835310936, + 0.0054830824956297874, + 0.0070511712692677975, + 0.013575365766882896, + 0.008159147575497627, + 0.004931287840008736, + 0.010606053285300732, + 0.003674406325444579, + 0.003896857611835003, + 0.0033062337897717953, + 0.005473775789141655, + 0.003753704484552145, + 0.01262457761913538, + 0.009761129505932331, + 0.011932305991649628, + 0.009892638772726059, + 0.007898476906120777, + 0.009864584542810917, + 0.013035167008638382, + 0.009973006322979927, + 0.005099811591207981, + 0.004454025533050299, + 0.009029611013829708, + 0.0069597638212144375, + 0.006034765392541885, + 0.0071258507668972015, + 0.007342000026255846, + 0.009601299650967121, + 0.0467766672372818, + 0.030584506690502167, + 0.050632722675800323, + 0.026887424290180206, + 0.025202250108122826, + 0.029825275763869286, + 0.02027849853038788, + 0.04477424547076225 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 0, + 7 + ], + "selected_reward": -0.11290083080530167 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000006/group_10/index_tensor([9163, 9163])_group_10_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000006/group_10/index_tensor([9163, 9163])_group_10_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..2101e91abeb3ff353af8e62178d10cc2304ea749 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000006/group_10/index_tensor([9163, 9163])_group_10_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "wide-angle, static time-lapse video captured from above recreates the scenery of Ciqikou, Chongqing, in a distinctive 8-bit pixel art style. Ancient stone steps, composed of rough gray and brown pixels, wind down, flanked by traditional wooden buildings constructed from dark brown pixels. Their iconic eaves and corners are simplified into hard pixel lines, and their windows are simple, glowing squares. On these narrow streets, an abstract crowd, composed of tiny pixels in vibrant colors, surges like a tide, forming a steady, swift flow of pixels down the stone steps. The sky background is a gradient of several blue hues. Several large, white, pixelated clouds drift past in unnatural, blocky formations, casting shifting, pixelated shadows on the ground. The color palette of the entire picture is strictly limited, and the relationship between light and shadow is expressed through the brightness and darkness of pixel blocks, creating a strong sense of retro and digital feeling. With a unique dynamic beauty, it perfectly reproduces the vibrant and nostalgic atmosphere of early electronic games.", + "seed": 24868, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000006/group_10/index_9163_seed_24868_dp_20_rank_40_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 10, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.043411473440937695, + "vlm_logprob": -0.043411473440937695, + "vlm_yes_no_margin": 4.3, + "vlm_token_credit": -0.043411473440937695, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video presented in an 8-bit pixel art style?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are ancient stone steps made of rough gray and brown pixels visible winding down the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are traditional wooden buildings with dark brown pixels, simplified eaves, and glowing square windows present?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is an abstract crowd of tiny vibrant pixels flowing swiftly down the stone steps?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the sky a gradient of blue hues with large white pixelated clouds casting shifting shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.1805073767900467 + ], + "opposite_logprob_tokens": [ + -1.8055074214935303 + ], + "logprob_reward": -0.1805073767900467, + "yes_no_margin_reward": 1.625, + "token_credit_reward": -0.1805073767900467, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.048631634563207626, + 0.03342123329639435, + 0.12631744146347046, + 0.1451122909784317, + 0.024025505408644676, + 0.0939091369509697, + 0.060138002038002014, + 0.41115567088127136, + 0.018666431307792664, + 0.03862259164452553 + ], + "credit_map": [ + 0.004543338902294636, + 0.005594104528427124, + 0.008933989331126213, + 0.00659953011199832, + 0.005615098401904106, + 0.004587836097925901, + 0.006809644866734743, + 0.005948096513748169, + 0.003349999664351344, + 0.0028021351899951696, + 0.006284572184085846, + 0.004332622978836298, + 0.002666073851287365, + 0.0021364910062402487, + 0.008370406925678253, + 0.0034789363853633404, + 0.01183102373033762, + 0.013882752507925034, + 0.013938989490270615, + 0.03433087840676308, + 0.011850559152662754, + 0.005413173232227564, + 0.02553616836667061, + 0.009533905424177647, + 0.013797111809253693, + 0.013714431785047054, + 0.02277388423681259, + 0.016728347167372704, + 0.01614837534725666, + 0.010756643489003181, + 0.01724311150610447, + 0.03395039960741997, + 0.0029960458632558584, + 0.002550148405134678, + 0.003070682752877474, + 0.0032131837215274572, + 0.002504311501979828, + 0.0019518558401614428, + 0.0029685785993933678, + 0.004770701751112938, + 0.007769464515149593, + 0.016928644850850105, + 0.006897278595715761, + 0.009602558799088001, + 0.015661731362342834, + 0.01731540448963642, + 0.011615192517638206, + 0.008118868805468082, + 0.004043309949338436, + 0.006630691699683666, + 0.008907480165362358, + 0.010755338706076145, + 0.007360957097262144, + 0.0038254032842814922, + 0.013559971936047077, + 0.005054854322224855, + 0.030941221863031387, + 0.03251020982861519, + 0.05412480980157852, + 0.060548316687345505, + 0.04069551080465317, + 0.02720891870558262, + 0.11848288774490356, + 0.04664381965994835, + 0.0019493545405566692, + 0.0021719904616475105, + 0.0029799742624163628, + 0.002591195050626993, + 0.0029008532874286175, + 0.0011086518643423915, + 0.0027758360374718904, + 0.0021885784808546305, + 0.0020158158149570227, + 0.013163179159164429, + 0.0015411163913086057, + 0.002062497427687049, + 0.01362556405365467, + 0.0018869721097871661, + 0.002567160176113248, + 0.0017602953594177961 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 3, + 2 + ], + "selected_reward": -0.1805073767900467 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video presented in an 8-bit pixel art style?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are ancient stone steps made of rough gray and brown pixels visible winding down the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are traditional wooden buildings with dark brown pixels, simplified eaves, and glowing square windows present?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is an abstract crowd of tiny vibrant pixels flowing swiftly down the stone steps?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the sky a gradient of blue hues with large white pixelated clouds casting shifting shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.011081461794674397 + ], + "opposite_logprob_tokens": [ + -4.511081695556641 + ], + "logprob_reward": -0.011081461794674397, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.011081461794674397, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0814228504896164, + 0.0235136765986681, + 0.11737164855003357, + 0.05859045684337616, + 0.023047994822263718, + 0.09317819774150848, + 0.03370574861764908, + 0.31121304631233215, + 0.08086837828159332, + 0.17708800733089447 + ], + "credit_map": [ + 0.008212062530219555, + 0.013240006752312183, + 0.010602463036775589, + 0.007247528992593288, + 0.014537980780005455, + 0.008520431816577911, + 0.008736344054341316, + 0.01032603532075882, + 0.003592395456507802, + 0.0022633280605077744, + 0.004612734541296959, + 0.002641543047502637, + 0.0021260601934045553, + 0.0015073582762852311, + 0.003737220773473382, + 0.0030330349691212177, + 0.011272558942437172, + 0.012732886709272861, + 0.01226387545466423, + 0.03183566406369209, + 0.010074379853904247, + 0.004940271843224764, + 0.02438604086637497, + 0.009865972213447094, + 0.005132526159286499, + 0.0050577945075929165, + 0.01052802987396717, + 0.009096878580749035, + 0.005400587804615498, + 0.0038589087780565023, + 0.008080674335360527, + 0.011435053311288357, + 0.002608278999105096, + 0.0023227869533002377, + 0.003095365362241864, + 0.0038143519777804613, + 0.0019155546324327588, + 0.001247055479325354, + 0.004457700066268444, + 0.0035869013518095016, + 0.007886729203164577, + 0.0180834848433733, + 0.007074774708598852, + 0.009087792597711086, + 0.01688595861196518, + 0.016025982797145844, + 0.011058744974434376, + 0.00707473186776042, + 0.0025982835795730352, + 0.003616079455241561, + 0.004685394000262022, + 0.0067857108078897, + 0.003923739772289991, + 0.001967091578990221, + 0.006685271859169006, + 0.0034441782627254725, + 0.023817814886569977, + 0.025261785835027695, + 0.040764909237623215, + 0.04467019811272621, + 0.031464848667383194, + 0.019666453823447227, + 0.09033717960119247, + 0.0352298729121685, + 0.004697734024375677, + 0.008499612100422382, + 0.00983477383852005, + 0.01617242954671383, + 0.008215204812586308, + 0.004813905339688063, + 0.020676685497164726, + 0.00795803777873516, + 0.007416433654725552, + 0.06760458648204803, + 0.004791293293237686, + 0.0060617937706410885, + 0.06986957043409348, + 0.0069696116261184216, + 0.008079244755208492, + 0.006295463535934687 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 9, + 2 + ], + "selected_reward": -0.011081461794674397 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video presented in an 8-bit pixel art style?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are ancient stone steps made of rough gray and brown pixels visible winding down the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are traditional wooden buildings with dark brown pixels, simplified eaves, and glowing square windows present?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is an abstract crowd of tiny vibrant pixels flowing swiftly down the stone steps?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the sky a gradient of blue hues with large white pixelated clouds casting shifting shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0015111466636881232 + ], + "opposite_logprob_tokens": [ + -6.501511096954346 + ], + "logprob_reward": -0.0015111466636881232, + "yes_no_margin_reward": 6.5, + "token_credit_reward": -0.0015111466636881232, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0671696662902832, + 0.03265916928648949, + 0.0885842815041542, + 0.08855540305376053, + 0.10720357298851013, + 0.0722307339310646, + 0.04625687748193741, + 0.2792508006095886, + 0.024614181369543076, + 0.19347530603408813 + ], + "credit_map": [ + 0.006521311588585377, + 0.007081715390086174, + 0.010948152281343937, + 0.007979186251759529, + 0.008259052410721779, + 0.008420621044933796, + 0.007736629340797663, + 0.010222998447716236, + 0.005844180937856436, + 0.0033114159014075994, + 0.006229412741959095, + 0.003939319867640734, + 0.0026117831002920866, + 0.00217737490311265, + 0.002769840881228447, + 0.005775841884315014, + 0.007710707373917103, + 0.009709404781460762, + 0.00939043890684843, + 0.024297619238495827, + 0.008569472469389439, + 0.0036687033716589212, + 0.018266692757606506, + 0.006971247959882021, + 0.008467637933790684, + 0.008465319871902466, + 0.013872972689568996, + 0.01029337476938963, + 0.009590194560587406, + 0.00654881214722991, + 0.011345657519996166, + 0.01997142843902111, + 0.014911344274878502, + 0.010145667009055614, + 0.017642704769968987, + 0.014677346684038639, + 0.009860522113740444, + 0.005962113384157419, + 0.014527608640491962, + 0.01947627402842045, + 0.005823238752782345, + 0.012492616660892963, + 0.005718611646443605, + 0.008126141503453255, + 0.011438131332397461, + 0.012564104981720448, + 0.009685015305876732, + 0.006382871884852648, + 0.003458983963355422, + 0.004033501259982586, + 0.010032194666564465, + 0.008289134129881859, + 0.004516798537224531, + 0.0023038815706968307, + 0.008717769756913185, + 0.004904615692794323, + 0.0212051123380661, + 0.022100437432527542, + 0.035634126514196396, + 0.0401679128408432, + 0.02872897870838642, + 0.018290681764483452, + 0.08246754854917526, + 0.030656002461910248, + 0.0020469294395297766, + 0.0028686190489679575, + 0.0027719412464648485, + 0.004007722716778517, + 0.0028620592784136534, + 0.0014507115120068192, + 0.005959570873528719, + 0.0026466278359293938, + 0.012838010676205158, + 0.07078038156032562, + 0.007281216327100992, + 0.00926393736153841, + 0.0703577771782875, + 0.007261045277118683, + 0.00793583132326603, + 0.007757105864584446 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 9, + 4 + ], + "selected_reward": -0.0015111466636881232 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video presented in an 8-bit pixel art style?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are ancient stone steps made of rough gray and brown pixels visible winding down the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are traditional wooden buildings with dark brown pixels, simplified eaves, and glowing square windows present?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is an abstract crowd of tiny vibrant pixels flowing swiftly down the stone steps?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the sky a gradient of blue hues with large white pixelated clouds casting shifting shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.014187542721629143 + ], + "opposite_logprob_tokens": [ + -4.264187335968018 + ], + "logprob_reward": -0.014187542721629143, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.014187542721629143, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09098973870277405, + 0.04630376026034355, + 0.10026436299085617, + 0.06548001617193222, + 0.04746399447321892, + 0.06153791397809982, + 0.1370742917060852, + 0.29284846782684326, + 0.03949996829032898, + 0.11853757500648499 + ], + "credit_map": [ + 0.008641167543828487, + 0.012546997517347336, + 0.01692366600036621, + 0.01383089367300272, + 0.010311739519238472, + 0.005655664950609207, + 0.012174373492598534, + 0.010905223898589611, + 0.006449196022003889, + 0.004307977389544249, + 0.006600933615118265, + 0.006556915584951639, + 0.002936526434496045, + 0.002686553867533803, + 0.011912129819393158, + 0.004853520076721907, + 0.009934348054230213, + 0.011437901295721531, + 0.010615966282784939, + 0.026232292875647545, + 0.008743299171328545, + 0.0043161725625395775, + 0.02050524577498436, + 0.008479118347167969, + 0.005222611129283905, + 0.007241465151309967, + 0.009693973697721958, + 0.007843328639864922, + 0.007539105601608753, + 0.004513497464358807, + 0.01313481479883194, + 0.010291212238371372, + 0.003974016290158033, + 0.004384947009384632, + 0.005496917758136988, + 0.012162731029093266, + 0.0044000535272061825, + 0.003826495725661516, + 0.006887801457196474, + 0.0063310242258012295, + 0.006449027452617884, + 0.013835887424647808, + 0.0050821625627577305, + 0.00510057108476758, + 0.015213615261018276, + 0.007305005565285683, + 0.005164701957255602, + 0.0033869321923702955, + 0.009695022366940975, + 0.015963058918714523, + 0.01439584419131279, + 0.02650395780801773, + 0.019054776057600975, + 0.009229632094502449, + 0.030898289754986763, + 0.011333690024912357, + 0.0208344254642725, + 0.024155182763934135, + 0.040762510150671005, + 0.04269808158278465, + 0.027579255402088165, + 0.019269239157438278, + 0.08327620476484299, + 0.034273531287908554, + 0.0034796318504959345, + 0.00366517947986722, + 0.005043543875217438, + 0.007505268789827824, + 0.004865828435868025, + 0.0019061168422922492, + 0.008967972360551357, + 0.004066423047333956, + 0.004636878613382578, + 0.04574308916926384, + 0.003185891080647707, + 0.004376044496893883, + 0.04721221327781677, + 0.004754892084747553, + 0.004715824034065008, + 0.003912725951522589 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 6, + 9 + ], + "selected_reward": -0.014187542721629143 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video presented in an 8-bit pixel art style?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are ancient stone steps made of rough gray and brown pixels visible winding down the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are traditional wooden buildings with dark brown pixels, simplified eaves, and glowing square windows present?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is an abstract crowd of tiny vibrant pixels flowing swiftly down the stone steps?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the sky a gradient of blue hues with large white pixelated clouds casting shifting shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.009769839234650135 + ], + "opposite_logprob_tokens": [ + -4.634769916534424 + ], + "logprob_reward": -0.009769839234650135, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.009769839234650135, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.04435906186699867, + 0.029330916702747345, + 0.060701943933963776, + 0.05996491387486458, + 0.04976488649845123, + 0.03216439485549927, + 0.06909302622079849, + 0.5735353827476501, + 0.025985684245824814, + 0.05509982258081436 + ], + "credit_map": [ + 0.005308885592967272, + 0.007117447443306446, + 0.004024700727313757, + 0.004828978795558214, + 0.00955214910209179, + 0.004359251819550991, + 0.004943972919136286, + 0.004223670344799757, + 0.004250822588801384, + 0.0028959920164197683, + 0.004991017282009125, + 0.0034986939281225204, + 0.002868078416213393, + 0.0023981300182640553, + 0.0038779848255217075, + 0.00455019436776638, + 0.005549134686589241, + 0.0074762688018381596, + 0.00631074421107769, + 0.014306106604635715, + 0.00947509054094553, + 0.00306110386736691, + 0.010273943655192852, + 0.004249548073858023, + 0.005817110650241375, + 0.005482835695147514, + 0.010119116865098476, + 0.006766874343156815, + 0.006709123030304909, + 0.004560100380331278, + 0.007628480438143015, + 0.012881267815828323, + 0.005422315560281277, + 0.005277569871395826, + 0.005831341724842787, + 0.0070880441926419735, + 0.006222745403647423, + 0.00419868528842926, + 0.008578499779105186, + 0.007145676761865616, + 0.002689668443053961, + 0.0060613034293055534, + 0.002540683839470148, + 0.003730516880750656, + 0.005328445229679346, + 0.0054878415539860725, + 0.003745240857824683, + 0.002580690896138549, + 0.004618268925696611, + 0.006624436471611261, + 0.010265648365020752, + 0.01241739559918642, + 0.007954798638820648, + 0.00452754320576787, + 0.017038753256201744, + 0.005646172445267439, + 0.04359247907996178, + 0.04695920646190643, + 0.07702169567346573, + 0.08280228078365326, + 0.057266656309366226, + 0.03586288169026375, + 0.16446897387504578, + 0.0655612200498581, + 0.003013169625774026, + 0.00282668205909431, + 0.003893923945724964, + 0.004518176428973675, + 0.002728415885940194, + 0.0016763347666710615, + 0.004677978344261646, + 0.0026509982999414206, + 0.0028607926797121763, + 0.020419709384441376, + 0.0018049628706648946, + 0.0021305985283106565, + 0.021069174632430077, + 0.0026174974627792835, + 0.0020044506527483463, + 0.002192635554820299 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 6, + 2 + ], + "selected_reward": -0.009769839234650135 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "wide-angle, static time-lapse video captured from above recreates the scenery of Ciqikou, Chongqing, in a distinctive 8-bit pixel art style. Ancient stone steps, composed of rough gray and brown pixels, wind down, flanked by traditional wooden buildings constructed from dark brown pixels. Their iconic eaves and corners are simplified into hard pixel lines, and their windows are simple, glowing squares. On these narrow streets, an abstract crowd, composed of tiny pixels in vibrant colors, surges like a tide, forming a steady, swift flow of pixels down the stone steps. The sky background is a gradient of several blue hues. Several large, white, pixelated clouds drift past in unnatural, blocky formations, casting shifting, pixelated shadows on the ground. The color palette of the entire picture is strictly limited, and the relationship between light and shadow is expressed through the brightness and darkness of pixel blocks, creating a strong sense of retro and digital feeling. With a unique dynamic beauty, it perfectly reproduces the vibrant and nostalgic atmosphere of early electronic games.", + "seed": 24869, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000006/group_10/index_9163_seed_24869_dp_20_rank_40_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 10, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.10843967130640522, + "vlm_logprob": -0.10843967130640522, + "vlm_yes_no_margin": 3.675, + "vlm_token_credit": -0.10843967130640522, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video presented in an 8-bit pixel art style?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are ancient stone steps made of rough gray and brown pixels visible winding down the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are traditional wooden buildings with dark brown pixels, simplified eaves, and glowing square windows present?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is an abstract crowd of tiny vibrant pixels flowing swiftly down the stone steps?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the sky a gradient of blue hues with large white pixelated clouds casting shifting shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.008811874315142632 + ], + "opposite_logprob_tokens": [ + -4.883811950683594 + ], + "logprob_reward": -0.008811874315142632, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.008811874315142632, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.05688147246837616, + 0.4198969602584839, + 0.045619700103998184, + 0.051524121314287186, + 0.05389876291155815, + 0.028826745226979256, + 0.029846522957086563, + 0.1101263239979744, + 0.1318398118019104, + 0.07153955101966858 + ], + "credit_map": [ + 0.010825015604496002, + 0.007691128645092249, + 0.0044266157783567905, + 0.004746908321976662, + 0.00599760003387928, + 0.005409983452409506, + 0.005403011571615934, + 0.012381217442452908, + 0.07393699884414673, + 0.12436722218990326, + 0.011577029712498188, + 0.01594882272183895, + 0.03224952146410942, + 0.014976182021200657, + 0.028580080717802048, + 0.11826111376285553, + 0.007402629591524601, + 0.012348159216344357, + 0.0025326141621917486, + 0.004250316880643368, + 0.0034757000394165516, + 0.0024656690657138824, + 0.008508543483912945, + 0.004636068362742662, + 0.014572896994650364, + 0.010940978303551674, + 0.0030345083214342594, + 0.004405806306749582, + 0.004139422904700041, + 0.002654247684404254, + 0.006048442795872688, + 0.005727824755012989, + 0.010040500201284885, + 0.005806408356875181, + 0.004125293809920549, + 0.006784406490623951, + 0.002623075619339943, + 0.0028712970670312643, + 0.0053657847456634045, + 0.01628200151026249, + 0.0035333114210516214, + 0.003251000540331006, + 0.0033565557096153498, + 0.002965950407087803, + 0.002703308593481779, + 0.0026811903808265924, + 0.0055981227196753025, + 0.004737306386232376, + 0.00362209789454937, + 0.004124089144170284, + 0.0031873187981545925, + 0.0029908674769103527, + 0.004751040134578943, + 0.0030509454663842916, + 0.00434132432565093, + 0.0037788408808410168, + 0.021244458854198456, + 0.026683291420340538, + 0.008323480375111103, + 0.007132872007787228, + 0.013536842539906502, + 0.006399213336408138, + 0.015078122727572918, + 0.011728054843842983, + 0.029914019629359245, + 0.041366565972566605, + 0.002821305999532342, + 0.005619821138679981, + 0.007692215498536825, + 0.0033690775744616985, + 0.011791376397013664, + 0.029265446588397026, + 0.0212866123765707, + 0.006725707557052374, + 0.0031285786535590887, + 0.004652434028685093, + 0.002464479533955455, + 0.003127248492091894, + 0.004243432078510523, + 0.025911062955856323 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 8, + 7 + ], + "selected_reward": -0.008811874315142632 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video presented in an 8-bit pixel art style?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are ancient stone steps made of rough gray and brown pixels visible winding down the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are traditional wooden buildings with dark brown pixels, simplified eaves, and glowing square windows present?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is an abstract crowd of tiny vibrant pixels flowing swiftly down the stone steps?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the sky a gradient of blue hues with large white pixelated clouds casting shifting shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.05495621636509895 + ], + "opposite_logprob_tokens": [ + -2.9299561977386475 + ], + "logprob_reward": -0.05495621636509895, + "yes_no_margin_reward": 2.875, + "token_credit_reward": -0.05495621636509895, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08308874070644379, + 0.2062106877565384, + 0.0318915955722332, + 0.14535166323184967, + 0.04339813441038132, + 0.0734647884964943, + 0.07203028351068497, + 0.08869234472513199, + 0.12826654314994812, + 0.12760530412197113 + ], + "credit_map": [ + 0.011636869981884956, + 0.010359611362218857, + 0.011604594998061657, + 0.010541658848524094, + 0.008867713622748852, + 0.009814031422138214, + 0.010084537789225578, + 0.010179725475609303, + 0.0474235974252224, + 0.06851883232593536, + 0.013194684870541096, + 0.008558752946555614, + 0.017812753096222878, + 0.01203098800033331, + 0.012007292360067368, + 0.026663802564144135, + 0.005975041538476944, + 0.0036527328193187714, + 0.00360375689342618, + 0.0032322213519364595, + 0.0033856676891446114, + 0.002842080546543002, + 0.00414241710677743, + 0.005057677626609802, + 0.016257269307971, + 0.014578997157514095, + 0.013599621132016182, + 0.015535649843513966, + 0.010794242843985558, + 0.01789691299200058, + 0.018145417794585228, + 0.038543544709682465, + 0.0047813113778829575, + 0.0039040036499500275, + 0.006468473467975855, + 0.007737378589808941, + 0.0028667214792221785, + 0.004763235338032246, + 0.005627419333904982, + 0.007249589543789625, + 0.008010613732039928, + 0.005686587654054165, + 0.006793553940951824, + 0.006200158037245274, + 0.004816453438252211, + 0.006137251388281584, + 0.019219232723116875, + 0.01660093665122986, + 0.008816325105726719, + 0.008098861202597618, + 0.008756906725466251, + 0.007836098782718182, + 0.0086434381082654, + 0.006206687539815903, + 0.012861592695116997, + 0.010810376144945621, + 0.017374342307448387, + 0.02911543659865856, + 0.005561427678912878, + 0.005383001174777746, + 0.009516915306448936, + 0.004617070313543081, + 0.008652579970657825, + 0.008471581153571606, + 0.02681577391922474, + 0.03959157317876816, + 0.003963980358093977, + 0.005421874579042196, + 0.00713734794408083, + 0.004713204689323902, + 0.011559082195162773, + 0.02906370535492897, + 0.03315509855747223, + 0.010265899822115898, + 0.00954146683216095, + 0.010877261869609356, + 0.004542086739093065, + 0.007142341695725918, + 0.009781837463378906, + 0.0422993004322052 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 3, + 8 + ], + "selected_reward": -0.05495621636509895 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video presented in an 8-bit pixel art style?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are ancient stone steps made of rough gray and brown pixels visible winding down the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are traditional wooden buildings with dark brown pixels, simplified eaves, and glowing square windows present?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is an abstract crowd of tiny vibrant pixels flowing swiftly down the stone steps?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the sky a gradient of blue hues with large white pixelated clouds casting shifting shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0009191579301841557 + ], + "opposite_logprob_tokens": [ + -7.000919342041016 + ], + "logprob_reward": -0.0009191579301841557, + "yes_no_margin_reward": 7.0, + "token_credit_reward": -0.0009191579301841557, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.05381080135703087, + 0.4808633625507355, + 0.06439463049173355, + 0.0499897338449955, + 0.04781319200992584, + 0.04959949851036072, + 0.019100511446595192, + 0.11687038838863373, + 0.08476797491312027, + 0.03278990462422371 + ], + "credit_map": [ + 0.006510458420962095, + 0.006232056301087141, + 0.007875517010688782, + 0.006648389156907797, + 0.004708355758339167, + 0.0065835691057145596, + 0.0060461643151938915, + 0.009206291288137436, + 0.07660131901502609, + 0.12988996505737305, + 0.015175946988165379, + 0.018201051279902458, + 0.0347830168902874, + 0.017590675503015518, + 0.04035727679729462, + 0.148264080286026, + 0.005358078051358461, + 0.004232118837535381, + 0.0024204168003052473, + 0.005794567987322807, + 0.00438266946002841, + 0.005686863791197538, + 0.02033599466085434, + 0.016183922067284584, + 0.006616510450839996, + 0.005136317573487759, + 0.0036910935305058956, + 0.004861905239522457, + 0.003568989224731922, + 0.004970635753124952, + 0.009011511690914631, + 0.012132774107158184, + 0.007595319766551256, + 0.0062415711581707, + 0.007793564815074205, + 0.0045192972756922245, + 0.004735629539936781, + 0.003970769699662924, + 0.005840241443365812, + 0.007116798311471939, + 0.005212831310927868, + 0.0038353886920958757, + 0.0038682846352458, + 0.0036165497731417418, + 0.003326755715534091, + 0.003867975203320384, + 0.012409374117851257, + 0.013462340459227562, + 0.0027004675939679146, + 0.0023104220163077116, + 0.002301535103470087, + 0.001642382936552167, + 0.002284733112901449, + 0.0019437681185081601, + 0.003146858187392354, + 0.0027703437954187393, + 0.01813281700015068, + 0.02976658008992672, + 0.006753372959792614, + 0.0066688209772109985, + 0.009052152745425701, + 0.0060514528304338455, + 0.025232834741473198, + 0.015212351456284523, + 0.017608672380447388, + 0.025846580043435097, + 0.0026326002553105354, + 0.0041377427987754345, + 0.005122498609125614, + 0.003100020345300436, + 0.007838776335120201, + 0.018481088802218437, + 0.008035840466618538, + 0.0033649709075689316, + 0.0026325194630771875, + 0.0023221883457154036, + 0.002316530793905258, + 0.0025409983936697245, + 0.0029024132527410984, + 0.008674440905451775 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 7, + 8 + ], + "selected_reward": -0.0009191579301841557 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video presented in an 8-bit pixel art style?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are ancient stone steps made of rough gray and brown pixels visible winding down the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are traditional wooden buildings with dark brown pixels, simplified eaves, and glowing square windows present?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is an abstract crowd of tiny vibrant pixels flowing swiftly down the stone steps?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the sky a gradient of blue hues with large white pixelated clouds casting shifting shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.4288835823535919 + ], + "opposite_logprob_tokens": [ + -1.0538835525512695 + ], + "logprob_reward": -0.4288835823535919, + "yes_no_margin_reward": 0.625, + "token_credit_reward": -0.4288835823535919, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06311752647161484, + 0.5569616556167603, + 0.08097097277641296, + 0.027853185310959816, + 0.02786080166697502, + 0.04658258706331253, + 0.03160259127616882, + 0.030945969745516777, + 0.03883427754044533, + 0.09527046233415604 + ], + "credit_map": [ + 0.0135725736618042, + 0.005326441489160061, + 0.0036988521460443735, + 0.00657216552644968, + 0.004015659913420677, + 0.004400311969220638, + 0.006094885058701038, + 0.019436636939644814, + 0.10207992047071457, + 0.16816891729831696, + 0.01871207356452942, + 0.029204167425632477, + 0.04321691766381264, + 0.021470895037055016, + 0.032577481120824814, + 0.14153125882148743, + 0.005833517760038376, + 0.004293343052268028, + 0.004286739509552717, + 0.0072686015628278255, + 0.005033423658460379, + 0.005357797257602215, + 0.03006860800087452, + 0.01882893778383732, + 0.004320814274251461, + 0.003935032058507204, + 0.0020094714127480984, + 0.0029157381504774094, + 0.0022561417426913977, + 0.002450904343277216, + 0.004507611971348524, + 0.0054574706591665745, + 0.003307918319478631, + 0.002468428574502468, + 0.002209864556789398, + 0.005386936012655497, + 0.002204917836934328, + 0.00236813654191792, + 0.003475845092907548, + 0.006438754498958588, + 0.0033921163994818926, + 0.004498723428696394, + 0.0027014808729290962, + 0.0025593689642846584, + 0.0022234416101127863, + 0.001562462653964758, + 0.0166793093085289, + 0.012965678237378597, + 0.0034773037768900394, + 0.0037936391308903694, + 0.0035232824739068747, + 0.0036091420333832502, + 0.0043810345232486725, + 0.003290303284302354, + 0.005701521877199411, + 0.003826358588412404, + 0.004463663324713707, + 0.00453146081417799, + 0.002467218553647399, + 0.002320462604984641, + 0.0037566747050732374, + 0.0019061313942074776, + 0.007705364376306534, + 0.0037949939724057913, + 0.00838547758758068, + 0.011797623708844185, + 0.0011551742209121585, + 0.0017956584924831986, + 0.0024545895867049694, + 0.0014759267214685678, + 0.0038689952343702316, + 0.007900827564299107, + 0.021786190569400787, + 0.007017762865871191, + 0.00831428449600935, + 0.008863361552357674, + 0.004472020082175732, + 0.007261289749294519, + 0.009541024453938007, + 0.0280145276337862 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 9, + 2 + ], + "selected_reward": -0.4288835823535919 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video presented in an 8-bit pixel art style?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are ancient stone steps made of rough gray and brown pixels visible winding down the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are traditional wooden buildings with dark brown pixels, simplified eaves, and glowing square windows present?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is an abstract crowd of tiny vibrant pixels flowing swiftly down the stone steps?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the sky a gradient of blue hues with large white pixelated clouds casting shifting shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04862752556800842 + ], + "opposite_logprob_tokens": [ + -3.0486276149749756 + ], + "logprob_reward": -0.04862752556800842, + "yes_no_margin_reward": 3.0, + "token_credit_reward": -0.04862752556800842, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11664216220378876, + 0.2529942989349365, + 0.06902381032705307, + 0.0493396557867527, + 0.15410158038139343, + 0.0619518868625164, + 0.08360442519187927, + 0.0453699566423893, + 0.11668827384710312, + 0.05028389021754265 + ], + "credit_map": [ + 0.023767154663801193, + 0.011362253688275814, + 0.00509599270299077, + 0.011995253153145313, + 0.007986513897776604, + 0.006299356464296579, + 0.010139761492609978, + 0.039995886385440826, + 0.046601902693510056, + 0.07671721279621124, + 0.007202658802270889, + 0.009374014101922512, + 0.018892206251621246, + 0.008878065273165703, + 0.017093516886234283, + 0.06823472678661346, + 0.016668425872921944, + 0.019661422818899155, + 0.004172287415713072, + 0.00518482131883502, + 0.006788360886275768, + 0.003734918311238289, + 0.0072572859935462475, + 0.0055562895722687244, + 0.011224433779716492, + 0.006908563897013664, + 0.003202245570719242, + 0.004946818109601736, + 0.003726727794855833, + 0.0025149050634354353, + 0.011448321864008904, + 0.005367645528167486, + 0.026048263534903526, + 0.02109609916806221, + 0.023721566423773766, + 0.01553528942167759, + 0.015776878222823143, + 0.012827717699110508, + 0.018821928650140762, + 0.02027384750545025, + 0.012569909915328026, + 0.008788029663264751, + 0.005601559299975634, + 0.005201241001486778, + 0.00790778174996376, + 0.006218597758561373, + 0.009899988770484924, + 0.0057647861540317535, + 0.01236346922814846, + 0.010625041089951992, + 0.011184289120137691, + 0.007263870909810066, + 0.010526184923946857, + 0.00675330962985754, + 0.013068603351712227, + 0.01181966532021761, + 0.007795117795467377, + 0.011593891307711601, + 0.002547132084146142, + 0.0025536585599184036, + 0.004938628524541855, + 0.004273151513189077, + 0.005774469580501318, + 0.005893913097679615, + 0.025339506566524506, + 0.035222746431827545, + 0.0023799987975507975, + 0.004943555686622858, + 0.007936520501971245, + 0.003305110614746809, + 0.010423733852803707, + 0.027137111872434616, + 0.013714401982724667, + 0.00528620695695281, + 0.002569151110947132, + 0.0031437473371624947, + 0.0027106250636279583, + 0.0032298429869115353, + 0.0033829794265329838, + 0.016246935352683067 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 4, + 8 + ], + "selected_reward": -0.04862752556800842 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "wide-angle, static time-lapse video captured from above recreates the scenery of Ciqikou, Chongqing, in a distinctive 8-bit pixel art style. Ancient stone steps, composed of rough gray and brown pixels, wind down, flanked by traditional wooden buildings constructed from dark brown pixels. Their iconic eaves and corners are simplified into hard pixel lines, and their windows are simple, glowing squares. On these narrow streets, an abstract crowd, composed of tiny pixels in vibrant colors, surges like a tide, forming a steady, swift flow of pixels down the stone steps. The sky background is a gradient of several blue hues. Several large, white, pixelated clouds drift past in unnatural, blocky formations, casting shifting, pixelated shadows on the ground. The color palette of the entire picture is strictly limited, and the relationship between light and shadow is expressed through the brightness and darkness of pixel blocks, creating a strong sense of retro and digital feeling. With a unique dynamic beauty, it perfectly reproduces the vibrant and nostalgic atmosphere of early electronic games.", + "seed": 24870, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000006/group_10/index_9163_seed_24870_dp_21_rank_42_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 10, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.10218467037193477, + "vlm_logprob": -0.10218467037193477, + "vlm_yes_no_margin": 3.65, + "vlm_token_credit": -0.10218467037193477, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video presented in an 8-bit pixel art style?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are ancient stone steps made of rough gray and brown pixels visible winding down the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are traditional wooden buildings with dark brown pixels, simplified eaves, and glowing square windows present?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is an abstract crowd of tiny vibrant pixels flowing swiftly down the stone steps?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the sky a gradient of blue hues with large white pixelated clouds casting shifting shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.010702827014029026 + ], + "opposite_logprob_tokens": [ + -4.635702610015869 + ], + "logprob_reward": -0.010702827014029026, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.010702827014029026, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.005595090799033642, + 0.00048114979290403426, + 0.0006363455904647708, + 0.00043191021541133523, + 0.0029374754521995783, + 0.0013778480933979154, + 0.0013228406896814704, + 0.0005033774068579078, + 0.00020141666755080223, + 0.9865125417709351 + ], + "credit_map": [ + 0.0004334952973295003, + 0.0019652298651635647, + 0.00023892194440122694, + 0.00033704593079164624, + 0.0006098670419305563, + 0.0014608376659452915, + 0.0002727135142777115, + 0.0002769793209154159, + 6.621177453780547e-05, + 5.892480112379417e-05, + 5.633618638967164e-05, + 6.971557013457641e-05, + 8.728304965188727e-05, + 4.585937494994141e-05, + 4.9570418923394755e-05, + 4.7248569899238646e-05, + 4.277523476048373e-05, + 0.00018596157315187156, + 4.115444971830584e-05, + 0.00010395848948974162, + 8.635663107270375e-05, + 9.10190719878301e-05, + 5.3513933380600065e-05, + 3.1606166885467246e-05, + 3.553573333192617e-05, + 8.165762119460851e-05, + 4.175968570052646e-05, + 7.487376569770277e-05, + 6.43376843072474e-05, + 5.03268456668593e-05, + 5.301744386088103e-05, + 3.0401410185731947e-05, + 0.00019325206812936813, + 0.00033952161902561784, + 0.00015399718540720642, + 0.0003015018010046333, + 0.0009903019526973367, + 0.0006470480002462864, + 0.00016756683180574328, + 0.00014428573194891214, + 0.0001049513157340698, + 0.00037185847759246826, + 0.00012162682833150029, + 0.00018336840730626136, + 0.00031548907281830907, + 0.0001127684663515538, + 0.0001038670598063618, + 6.39183126622811e-05, + 0.00014100788393989205, + 0.00018259554053656757, + 0.00013948013656772673, + 0.00018112380348611623, + 0.00023949619207996875, + 0.00018297536007594317, + 0.0001613059575902298, + 9.485560440225527e-05, + 4.4944976252736524e-05, + 0.00015737442299723625, + 3.948011726606637e-05, + 5.6678713008295745e-05, + 6.967936496948823e-05, + 6.863800081191584e-05, + 4.545122283161618e-05, + 2.1130548702785745e-05, + 2.860413587768562e-05, + 1.5920544683467597e-05, + 2.605536974442657e-05, + 2.5622628527344204e-05, + 3.522141560097225e-05, + 3.148262476315722e-05, + 1.9680521290865727e-05, + 1.8829414329957217e-05, + 0.0691346749663353, + 0.05809900164604187, + 0.07341188192367554, + 0.3686661124229431, + 0.2358456403017044, + 0.04290737211704254, + 0.1107618510723114, + 0.02768593654036522 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 0, + 4 + ], + "selected_reward": -0.010702827014029026 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video presented in an 8-bit pixel art style?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are ancient stone steps made of rough gray and brown pixels visible winding down the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are traditional wooden buildings with dark brown pixels, simplified eaves, and glowing square windows present?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is an abstract crowd of tiny vibrant pixels flowing swiftly down the stone steps?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the sky a gradient of blue hues with large white pixelated clouds casting shifting shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.026351653039455414 + ], + "opposite_logprob_tokens": [ + -3.6513516902923584 + ], + "logprob_reward": -0.026351653039455414, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.026351653039455414, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.019177157431840897, + 0.0013271495699882507, + 0.0036295729223638773, + 0.0031857630237936974, + 0.00496383011341095, + 0.003190895775333047, + 0.0013751030201092362, + 0.0026034656912088394, + 0.00023721071192994714, + 0.9603098630905151 + ], + "credit_map": [ + 0.0015177421737462282, + 0.006752444431185722, + 0.000976982875727117, + 0.0012809197651222348, + 0.0023741230834275484, + 0.004585156217217445, + 0.0007971019949764013, + 0.0008926867740228772, + 0.0001371913676848635, + 0.0001612366031622514, + 0.00011779093620134518, + 0.00018899515271186829, + 0.00019931486167479306, + 0.00017852427845355123, + 0.0001643469004193321, + 0.00017974957881961018, + 0.0003815408272203058, + 0.0011592304799705744, + 0.00023948974558152258, + 0.00045229593524709344, + 0.0005770179559476674, + 0.00032413468579761684, + 0.000314974196953699, + 0.00018088870274368674, + 0.00036580878077074885, + 0.0005552633083425462, + 0.00021746184211224318, + 0.000357127923052758, + 0.00017935544019564986, + 0.0011521315900608897, + 0.00017340069462079555, + 0.00018521345918998122, + 0.00039452267810702324, + 0.00042779932846315205, + 0.0004724420723505318, + 0.0004657285171560943, + 0.0011713900603353977, + 0.0011471085017547011, + 0.00041621955460868776, + 0.00046861922601237893, + 0.00032961301621980965, + 0.0006744169513694942, + 0.000320423801895231, + 0.0008650586823932827, + 0.00029909907607361674, + 0.00024285929976031184, + 0.00029259646544232965, + 0.00016682874411344528, + 0.00017323586507700384, + 0.00017492724873591214, + 0.00011790316784754395, + 0.0001886348327388987, + 0.00023382292420137674, + 0.0002012462355196476, + 0.00016427991795353591, + 0.00012105289351893589, + 0.00023243387113325298, + 0.00047539977822452784, + 0.00026071962201967835, + 0.0005840758676640689, + 0.000399098003981635, + 0.0002333784505026415, + 0.00026904710102826357, + 0.00014931289479136467, + 2.1784604541608132e-05, + 3.151172859361395e-05, + 2.7729140128940344e-05, + 4.4187003368278965e-05, + 3.1588217098033056e-05, + 3.0387531296582893e-05, + 2.913376738433726e-05, + 2.088872315653134e-05, + 0.06716194748878479, + 0.05661821737885475, + 0.07148972153663635, + 0.3587508499622345, + 0.22965887188911438, + 0.04176479950547218, + 0.10794229805469513, + 0.02692318893969059 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 0, + 4 + ], + "selected_reward": -0.026351653039455414 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video presented in an 8-bit pixel art style?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are ancient stone steps made of rough gray and brown pixels visible winding down the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are traditional wooden buildings with dark brown pixels, simplified eaves, and glowing square windows present?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is an abstract crowd of tiny vibrant pixels flowing swiftly down the stone steps?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the sky a gradient of blue hues with large white pixelated clouds casting shifting shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0019396792631596327 + ], + "opposite_logprob_tokens": [ + -6.25193977355957 + ], + "logprob_reward": -0.0019396792631596327, + "yes_no_margin_reward": 6.25, + "token_credit_reward": -0.0019396792631596327, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.03384476155042648, + 0.005613991990685463, + 0.012915940023958683, + 0.00297814654186368, + 0.008312109857797623, + 0.009529955685138702, + 0.0043772440403699875, + 0.007778705097734928, + 0.0022294530645012856, + 0.9124197363853455 + ], + "credit_map": [ + 0.00341324252076447, + 0.01060149259865284, + 0.001749455463141203, + 0.002356990473344922, + 0.0030357760842889547, + 0.009207472205162048, + 0.0016962880035862327, + 0.0017840468790382147, + 0.0005448361625894904, + 0.0004139299562666565, + 0.000518993881996721, + 0.0008361884974874556, + 0.00044927894487045705, + 0.0011129362974315882, + 0.0005777705227956176, + 0.0011600578436627984, + 0.0015667795669287443, + 0.001100759720429778, + 0.0013045186642557383, + 0.003104069037362933, + 0.0006336792139336467, + 0.0021042164880782366, + 0.0017013319302350283, + 0.0014005848206579685, + 0.00030171286198310554, + 0.00027936603873968124, + 0.00031905685318633914, + 0.0008471101755276322, + 0.0003294672933407128, + 0.0002801257651299238, + 0.0004051167343277484, + 0.00021619067410938442, + 0.0004541982489172369, + 0.00105016038287431, + 0.0004286571638658643, + 0.0010411194525659084, + 0.0029134214855730534, + 0.001454817713238299, + 0.0005651568644680083, + 0.00040457857539877295, + 0.0007012116257101297, + 0.0025358283892273903, + 0.0010232131462544203, + 0.000815733103081584, + 0.002334938384592533, + 0.0008132238872349262, + 0.0007376753492280841, + 0.0005681318580172956, + 0.0006933559197932482, + 0.0004967439454048872, + 0.00038134833448566496, + 0.0005567925400100648, + 0.0005418352084234357, + 0.0007560277008451521, + 0.0005412200116552413, + 0.0004099206707905978, + 0.0006894048419781029, + 0.00211941241286695, + 0.0005355903413146734, + 0.0009026566985994577, + 0.0012229050043970346, + 0.001262607634998858, + 0.0006282844115048647, + 0.0004178434901405126, + 0.00026059692027047276, + 0.0001604048302397132, + 0.00019195134518668056, + 0.0002855071797966957, + 0.000307818001601845, + 0.00047566115972585976, + 0.0002944342268165201, + 0.0002530792844481766, + 0.06386730819940567, + 0.05388817563652992, + 0.06804806739091873, + 0.3407056927680969, + 0.21785351634025574, + 0.03973883390426636, + 0.10256064683198929, + 0.025757471099495888 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 0, + 2 + ], + "selected_reward": -0.0019396792631596327 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video presented in an 8-bit pixel art style?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are ancient stone steps made of rough gray and brown pixels visible winding down the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are traditional wooden buildings with dark brown pixels, simplified eaves, and glowing square windows present?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is an abstract crowd of tiny vibrant pixels flowing swiftly down the stone steps?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the sky a gradient of blue hues with large white pixelated clouds casting shifting shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.4288922846317291 + ], + "opposite_logprob_tokens": [ + -1.0538922548294067 + ], + "logprob_reward": -0.4288922846317291, + "yes_no_margin_reward": 0.625, + "token_credit_reward": -0.4288922846317291, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.018419580534100533, + 0.004153583664447069, + 0.00694044167175889, + 0.00954767968505621, + 0.00413744430989027, + 0.0025548157282173634, + 0.0051297214813530445, + 0.0014095607912167907, + 0.0013252206845209002, + 0.9463819265365601 + ], + "credit_map": [ + 0.0015269516734406352, + 0.00607627909630537, + 0.0009118318557739258, + 0.001250299159437418, + 0.00205485662445426, + 0.004609591327607632, + 0.0009999416070058942, + 0.0009898269781842828, + 0.0005626145866699517, + 0.000551303441170603, + 0.0004556843778118491, + 0.0007579373777844012, + 0.0005925589939579368, + 0.00038684398168697953, + 0.00048189423978328705, + 0.0003647459961939603, + 0.0008871363825164735, + 0.0010582798859104514, + 0.0006132699782028794, + 0.0014408810529857874, + 0.0007255298551172018, + 0.0007752740639261901, + 0.000936337688472122, + 0.0005037324735894799, + 0.0009503999608568847, + 0.0018013204680755734, + 0.0006442755693569779, + 0.0009206830291077495, + 0.0007243035943247378, + 0.0034769680351018906, + 0.0004876762395724654, + 0.0005420519737526774, + 0.0005415461491793394, + 0.00031780896824784577, + 0.00045028430758975446, + 0.0006880390574224293, + 0.0004955619806423783, + 0.0008635112317278981, + 0.0003957324370276183, + 0.0003849599452223629, + 0.0002783452218864113, + 0.00038371660048142076, + 0.00031588441925123334, + 0.0006860030698589981, + 0.00026221125153824687, + 0.00016112063894979656, + 0.00034211293677799404, + 0.00012542138574644923, + 0.0009334403439424932, + 0.0006995895528234541, + 0.00044988462468609214, + 0.0007560537196695805, + 0.0010721362195909023, + 0.0004107716085854918, + 0.00048719302867539227, + 0.0003206516266800463, + 0.00018725414702203125, + 0.00017011842282954603, + 9.529660019325092e-05, + 0.0002195195120293647, + 0.00023577880347147584, + 0.0002489125181455165, + 0.0001575615315232426, + 9.51191977947019e-05, + 0.00014925737923476845, + 0.00012850800703745335, + 0.00013799074804410338, + 0.0003728141018655151, + 0.00013485272938851267, + 0.00014728769019711763, + 0.00013255390513222665, + 0.00012195594899822026, + 0.06649129092693329, + 0.05558263137936592, + 0.07064084708690643, + 0.3537924289703369, + 0.2254064381122589, + 0.04125560075044632, + 0.10662371665239334, + 0.026588918641209602 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 0, + 3 + ], + "selected_reward": -0.4288922846317291 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video presented in an 8-bit pixel art style?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are ancient stone steps made of rough gray and brown pixels visible winding down the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are traditional wooden buildings with dark brown pixels, simplified eaves, and glowing square windows present?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is an abstract crowd of tiny vibrant pixels flowing swiftly down the stone steps?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the sky a gradient of blue hues with large white pixelated clouds casting shifting shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04303690791130066 + ], + "opposite_logprob_tokens": [ + -3.168036937713623 + ], + "logprob_reward": -0.04303690791130066, + "yes_no_margin_reward": 3.125, + "token_credit_reward": -0.04303690791130066, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.02943374775350094, + 0.010265730321407318, + 0.01032963115721941, + 0.010756176896393299, + 0.043387722223997116, + 0.013955322094261646, + 0.0063977912068367004, + 0.021536706015467644, + 0.0018277831841260195, + 0.8521094918251038 + ], + "credit_map": [ + 0.002328914823010564, + 0.0077744401060044765, + 0.0017096103401854634, + 0.0024198798928409815, + 0.004994508344680071, + 0.00679677864536643, + 0.0017736433073878288, + 0.001635969732888043, + 0.0008300017798319459, + 0.002104293555021286, + 0.0008754351874813437, + 0.0014805191894993186, + 0.002226456068456173, + 0.001057573826983571, + 0.0010724442545324564, + 0.0006190052372403443, + 0.0010248807957395911, + 0.0029767693486064672, + 0.0007719900459051132, + 0.0013983991229906678, + 0.0014008966973051429, + 0.0016386572970077395, + 0.0006352361524477601, + 0.0004827998927794397, + 0.00129014125559479, + 0.002085361396893859, + 0.0010310296202078462, + 0.00241650128737092, + 0.0013727884506806731, + 0.0011749863624572754, + 0.0008613410755060613, + 0.0005240267491899431, + 0.0023717291187494993, + 0.005122191272675991, + 0.0022092631552368402, + 0.00345875252969563, + 0.015069101005792618, + 0.010868730023503304, + 0.002765376353636384, + 0.0015225756214931607, + 0.0011558207916095853, + 0.0026437907945364714, + 0.0015704820398241282, + 0.0028181516099721193, + 0.0023806968238204718, + 0.001163878827355802, + 0.0013935576425865293, + 0.0008289420511573553, + 0.0007231032359413803, + 0.0009044169564731419, + 0.0006026732735335827, + 0.000805904739536345, + 0.0013945429818704724, + 0.0008665966452099383, + 0.0006006591720506549, + 0.0004998937365598977, + 0.002145713195204735, + 0.002739277668297291, + 0.002030365401878953, + 0.002326756017282605, + 0.003361026756465435, + 0.005512205883860588, + 0.001995573053136468, + 0.001425783964805305, + 0.00015592121053487062, + 0.00026615336537361145, + 0.00022508995607495308, + 0.0003386807511560619, + 0.00021427668980322778, + 0.00026034796610474586, + 0.0002106245228787884, + 0.0001566884311614558, + 0.059549830853939056, + 0.05029134452342987, + 0.06335500627756119, + 0.3179663121700287, + 0.2041652798652649, + 0.037089306861162186, + 0.09579567611217499, + 0.02389659732580185 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 4, + 0 + ], + "selected_reward": -0.04303690791130066 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "wide-angle, static time-lapse video captured from above recreates the scenery of Ciqikou, Chongqing, in a distinctive 8-bit pixel art style. Ancient stone steps, composed of rough gray and brown pixels, wind down, flanked by traditional wooden buildings constructed from dark brown pixels. Their iconic eaves and corners are simplified into hard pixel lines, and their windows are simple, glowing squares. On these narrow streets, an abstract crowd, composed of tiny pixels in vibrant colors, surges like a tide, forming a steady, swift flow of pixels down the stone steps. The sky background is a gradient of several blue hues. Several large, white, pixelated clouds drift past in unnatural, blocky formations, casting shifting, pixelated shadows on the ground. The color palette of the entire picture is strictly limited, and the relationship between light and shadow is expressed through the brightness and darkness of pixel blocks, creating a strong sense of retro and digital feeling. With a unique dynamic beauty, it perfectly reproduces the vibrant and nostalgic atmosphere of early electronic games.", + "seed": 24871, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000006/group_10/index_9163_seed_24871_dp_21_rank_42_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 10, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.06957986064953729, + "vlm_logprob": -0.06957986064953729, + "vlm_yes_no_margin": 4.675, + "vlm_token_credit": -0.06957986064953729, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video presented in an 8-bit pixel art style?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are ancient stone steps made of rough gray and brown pixels visible winding down the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are traditional wooden buildings with dark brown pixels, simplified eaves, and glowing square windows present?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is an abstract crowd of tiny vibrant pixels flowing swiftly down the stone steps?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the sky a gradient of blue hues with large white pixelated clouds casting shifting shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0015545914648100734 + ], + "opposite_logprob_tokens": [ + -7.001554489135742 + ], + "logprob_reward": -0.0015545914648100734, + "yes_no_margin_reward": 7.0, + "token_credit_reward": -0.0015545914648100734, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2825494110584259, + 0.14797528088092804, + 0.046439964324235916, + 0.07891877740621567, + 0.2129467874765396, + 0.03815798833966255, + 0.03148912265896797, + 0.08447124809026718, + 0.050082772970199585, + 0.026968596503138542 + ], + "credit_map": [ + 0.046287547796964645, + 0.03039964847266674, + 0.03476782515645027, + 0.028211891651153564, + 0.06070232763886452, + 0.02605646476149559, + 0.031573839485645294, + 0.024549877271056175, + 0.011621730402112007, + 0.027219250798225403, + 0.01313599944114685, + 0.016746019944548607, + 0.020040379837155342, + 0.01433285977691412, + 0.011964438483119011, + 0.032914601266384125, + 0.004904750268906355, + 0.008464456535875797, + 0.005218311212956905, + 0.0045823222026228905, + 0.0072634401731193066, + 0.004004760645329952, + 0.00645435368642211, + 0.005547571461647749, + 0.010552458465099335, + 0.014148465357720852, + 0.013363394886255264, + 0.006980251520872116, + 0.010612302459776402, + 0.006498163100332022, + 0.007654455956071615, + 0.009109281934797764, + 0.011652868241071701, + 0.01747417263686657, + 0.023799581453204155, + 0.02468685433268547, + 0.020406845957040787, + 0.015673518180847168, + 0.03880538046360016, + 0.0604475773870945, + 0.0029551496263593435, + 0.0034565934911370277, + 0.0029676775448024273, + 0.00585121801123023, + 0.004184759687632322, + 0.002347977366298437, + 0.01040700450539589, + 0.0059876092709600925, + 0.0023498295340687037, + 0.004666741006076336, + 0.003176619065925479, + 0.004191334825009108, + 0.0056104836985468864, + 0.0023591499775648117, + 0.004317065700888634, + 0.004817896988242865, + 0.005588206462562084, + 0.018217798322439194, + 0.006560028996318579, + 0.009709946811199188, + 0.021299371495842934, + 0.006614759564399719, + 0.009104895405471325, + 0.007376238238066435, + 0.008215993642807007, + 0.00832329224795103, + 0.004466368351131678, + 0.003780361730605364, + 0.010740114375948906, + 0.005280081182718277, + 0.0037274484056979418, + 0.005549114663153887, + 0.002646199194714427, + 0.0037231293972581625, + 0.0041145943105220795, + 0.003066345350816846, + 0.0035303602926433086, + 0.0028029971290379763, + 0.0027952962554991245, + 0.004289673175662756 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 1 + ], + "selected_reward": -0.0015545914648100734 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video presented in an 8-bit pixel art style?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are ancient stone steps made of rough gray and brown pixels visible winding down the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are traditional wooden buildings with dark brown pixels, simplified eaves, and glowing square windows present?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is an abstract crowd of tiny vibrant pixels flowing swiftly down the stone steps?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the sky a gradient of blue hues with large white pixelated clouds casting shifting shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.016066374257206917 + ], + "opposite_logprob_tokens": [ + -4.141066551208496 + ], + "logprob_reward": -0.016066374257206917, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.016066374257206917, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.16809768974781036, + 0.25365543365478516, + 0.0710684284567833, + 0.045850835740566254, + 0.17963778972625732, + 0.056325823068618774, + 0.10629787296056747, + 0.052779123187065125, + 0.0349385030567646, + 0.031348418444395065 + ], + "credit_map": [ + 0.01802939735352993, + 0.014299379661679268, + 0.02921016700565815, + 0.020425284281373024, + 0.021303271874785423, + 0.01744028553366661, + 0.018978340551257133, + 0.02841157093644142, + 0.021140217781066895, + 0.05072076991200447, + 0.026814162731170654, + 0.029512688517570496, + 0.034645237028598785, + 0.025833649560809135, + 0.024511443451046944, + 0.04047726094722748, + 0.006305451039224863, + 0.012216535396873951, + 0.008006202057003975, + 0.008494737558066845, + 0.012197988107800484, + 0.004751737229526043, + 0.009748877957463264, + 0.009346907027065754, + 0.004510377533733845, + 0.011712965555489063, + 0.005512819159775972, + 0.003556188428774476, + 0.008435931988060474, + 0.0032396421302109957, + 0.005114161875098944, + 0.0037687502335757017, + 0.008866925723850727, + 0.015458334237337112, + 0.01999479904770851, + 0.024296440184116364, + 0.01720867119729519, + 0.012098289094865322, + 0.03313840553164482, + 0.04857592657208443, + 0.0053432960994541645, + 0.004885384347289801, + 0.006066171918064356, + 0.008086719550192356, + 0.006229258142411709, + 0.0037174811586737633, + 0.014541799202561378, + 0.007455708459019661, + 0.006164077669382095, + 0.015325246378779411, + 0.008585071191191673, + 0.01575811207294464, + 0.020941926166415215, + 0.005333621520549059, + 0.017125461250543594, + 0.01706436090171337, + 0.004550084937363863, + 0.005929218605160713, + 0.005552603863179684, + 0.004934364929795265, + 0.006995916832238436, + 0.006257630418986082, + 0.005701942835003138, + 0.012857360765337944, + 0.0076378765515983105, + 0.002817921806126833, + 0.0031553138978779316, + 0.002531324280425906, + 0.008464577607810497, + 0.0033748364076018333, + 0.002533440012484789, + 0.004423210863023996, + 0.002409785520285368, + 0.002826076466590166, + 0.0029687935020774603, + 0.006256114691495895, + 0.003214831929653883, + 0.0021959953010082245, + 0.00728969182819128, + 0.004187128506600857 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 4, + 0 + ], + "selected_reward": -0.016066374257206917 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video presented in an 8-bit pixel art style?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are ancient stone steps made of rough gray and brown pixels visible winding down the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are traditional wooden buildings with dark brown pixels, simplified eaves, and glowing square windows present?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is an abstract crowd of tiny vibrant pixels flowing swiftly down the stone steps?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the sky a gradient of blue hues with large white pixelated clouds casting shifting shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0008112476789392531 + ], + "opposite_logprob_tokens": [ + -7.1258111000061035 + ], + "logprob_reward": -0.0008112476789392531, + "yes_no_margin_reward": 7.125, + "token_credit_reward": -0.0008112476789392531, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11308257281780243, + 0.2605450749397278, + 0.05146078020334244, + 0.07812897115945816, + 0.19598032534122467, + 0.061291784048080444, + 0.054841503500938416, + 0.09562856703996658, + 0.030921341851353645, + 0.05811907723546028 + ], + "credit_map": [ + 0.019050903618335724, + 0.01448769774287939, + 0.013291734270751476, + 0.012329650111496449, + 0.02285238727927208, + 0.01102655939757824, + 0.008985266089439392, + 0.01105837244540453, + 0.023303624242544174, + 0.033456750214099884, + 0.02457391656935215, + 0.029404692351818085, + 0.03211093321442604, + 0.019043823704123497, + 0.025947241112589836, + 0.07270407676696777, + 0.006010301411151886, + 0.012047523632645607, + 0.004900787025690079, + 0.0045797862112522125, + 0.010456413961946964, + 0.004512302577495575, + 0.004546350799500942, + 0.004407310392707586, + 0.011885950341820717, + 0.01654832437634468, + 0.00992755126208067, + 0.0063813370652496815, + 0.01048742514103651, + 0.0069845374673604965, + 0.007649374194443226, + 0.008264470845460892, + 0.009635519236326218, + 0.012021445669233799, + 0.027513116598129272, + 0.014351682737469673, + 0.016515880823135376, + 0.011139734648168087, + 0.021185841411352158, + 0.08361710608005524, + 0.005737427622079849, + 0.005904223304241896, + 0.010070673190057278, + 0.007629069034010172, + 0.006093262229114771, + 0.0067303841933608055, + 0.007801436819136143, + 0.01132530253380537, + 0.002958685625344515, + 0.010105694644153118, + 0.0034610347356647253, + 0.006483727600425482, + 0.012601222842931747, + 0.0032791635021567345, + 0.007568716071546078, + 0.008383261039853096, + 0.008187547326087952, + 0.014218644239008427, + 0.009222371503710747, + 0.007244190201163292, + 0.020664194598793983, + 0.011357134208083153, + 0.008190513588488102, + 0.01654396764934063, + 0.00508513581007719, + 0.00466883322224021, + 0.0030234919395297766, + 0.0024573642294853926, + 0.005638596136122942, + 0.0034635018091648817, + 0.0029154098592698574, + 0.0036690097767859697, + 0.005193152464926243, + 0.006882649380713701, + 0.007445069495588541, + 0.005694968160241842, + 0.006949939299374819, + 0.008403307758271694, + 0.00638315686956048, + 0.011166838929057121 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 4, + 0 + ], + "selected_reward": -0.0008112476789392531 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video presented in an 8-bit pixel art style?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are ancient stone steps made of rough gray and brown pixels visible winding down the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are traditional wooden buildings with dark brown pixels, simplified eaves, and glowing square windows present?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is an abstract crowd of tiny vibrant pixels flowing swiftly down the stone steps?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the sky a gradient of blue hues with large white pixelated clouds casting shifting shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.3134136199951172 + ], + "opposite_logprob_tokens": [ + -1.3134136199951172 + ], + "logprob_reward": -0.3134136199951172, + "yes_no_margin_reward": 1.0, + "token_credit_reward": -0.3134136199951172, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10874740034341812, + 0.22194716334342957, + 0.05033329501748085, + 0.07262811809778214, + 0.2296089380979538, + 0.05797016620635986, + 0.11425110697746277, + 0.09710643440485, + 0.023936692625284195, + 0.0234706848859787 + ], + "credit_map": [ + 0.01475713774561882, + 0.011567948386073112, + 0.014799036085605621, + 0.011651420965790749, + 0.019435150548815727, + 0.010403629392385483, + 0.01222192496061325, + 0.013911145739257336, + 0.018090205267071724, + 0.042034611105918884, + 0.020405413582921028, + 0.023610884323716164, + 0.03072836808860302, + 0.021854368969798088, + 0.018155427649617195, + 0.047067880630493164, + 0.005004323553293943, + 0.010248586535453796, + 0.004729936830699444, + 0.006753893569111824, + 0.007339244242757559, + 0.0037769749760627747, + 0.006994402036070824, + 0.005485924426466227, + 0.006092055235058069, + 0.013476865366101265, + 0.01432330533862114, + 0.009073144756257534, + 0.009952171705663204, + 0.005624649580568075, + 0.007829893380403519, + 0.006256026681512594, + 0.011379534378647804, + 0.020367247983813286, + 0.025430642068386078, + 0.027549071237444878, + 0.02112283557653427, + 0.015425385907292366, + 0.04132382199168205, + 0.06701039522886276, + 0.005222102161496878, + 0.004802617710083723, + 0.003249299479648471, + 0.008408167399466038, + 0.006217428017407656, + 0.004511087667196989, + 0.017240215092897415, + 0.008319244720041752, + 0.0068307495675981045, + 0.01961136795580387, + 0.007744017988443375, + 0.015088130719959736, + 0.024633493274450302, + 0.006108410190790892, + 0.01715126633644104, + 0.01708366721868515, + 0.010187545791268349, + 0.015495507046580315, + 0.00956541858613491, + 0.005928497761487961, + 0.02660227380692959, + 0.008492644876241684, + 0.007132594473659992, + 0.013701947405934334, + 0.005056542344391346, + 0.0020965756848454475, + 0.0014374430757015944, + 0.0011728368699550629, + 0.007046699523925781, + 0.0026944163255393505, + 0.0022042954806238413, + 0.0022278802935034037, + 0.002578162821009755, + 0.0021529688965529203, + 0.0023676941636949778, + 0.002791526261717081, + 0.0032950376626104116, + 0.0024696686305105686, + 0.004009179305285215, + 0.003806446213275194 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 1, + 6 + ], + "selected_reward": -0.3134136199951172 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video presented in an 8-bit pixel art style?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are ancient stone steps made of rough gray and brown pixels visible winding down the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are traditional wooden buildings with dark brown pixels, simplified eaves, and glowing square windows present?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is an abstract crowd of tiny vibrant pixels flowing swiftly down the stone steps?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the sky a gradient of blue hues with large white pixelated clouds casting shifting shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.016053469851613045 + ], + "opposite_logprob_tokens": [ + -4.141053676605225 + ], + "logprob_reward": -0.016053469851613045, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.016053469851613045, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2619107961654663, + 0.2179897576570511, + 0.0853322371840477, + 0.11571010202169418, + 0.04708561673760414, + 0.037717562168836594, + 0.05576157197356224, + 0.06319953501224518, + 0.04557047039270401, + 0.06972236931324005 + ], + "credit_map": [ + 0.03802785649895668, + 0.04418459162116051, + 0.03508034721016884, + 0.023913942277431488, + 0.03326534107327461, + 0.03802173584699631, + 0.017202245071530342, + 0.03221474587917328, + 0.024982459843158722, + 0.05056887865066528, + 0.016202399507164955, + 0.015493625774979591, + 0.029042979702353477, + 0.02565469592809677, + 0.017455600202083588, + 0.03858911618590355, + 0.013003181666135788, + 0.0200212299823761, + 0.0062065222300589085, + 0.008372418582439423, + 0.015095138922333717, + 0.00782780721783638, + 0.009292183443903923, + 0.005513755138963461, + 0.015107561834156513, + 0.015769658610224724, + 0.014933749102056026, + 0.01131940633058548, + 0.01980971358716488, + 0.009374191984534264, + 0.014439158141613007, + 0.014956658706068993, + 0.0038695652037858963, + 0.004825651179999113, + 0.004246193449944258, + 0.003484428161755204, + 0.005353497341275215, + 0.0033342039678245783, + 0.005701402667909861, + 0.016270671039819717, + 0.003406546078622341, + 0.004725997801870108, + 0.0035888615529984236, + 0.005659645423293114, + 0.004629496950656176, + 0.0022776599507778883, + 0.008547118864953518, + 0.0048822322860360146, + 0.004260537214577198, + 0.014496738091111183, + 0.004082187544554472, + 0.0031159683130681515, + 0.017508987337350845, + 0.0037982030771672726, + 0.0034939569886773825, + 0.0050049894489347935, + 0.009111832827329636, + 0.009658187627792358, + 0.0038604671135544777, + 0.006253689993172884, + 0.0186753049492836, + 0.0037331453058868647, + 0.006144230719655752, + 0.0057626753114163876, + 0.012375268153846264, + 0.0042459783144295216, + 0.0022935958113521338, + 0.0017496694345027208, + 0.01565723866224289, + 0.0034200805239379406, + 0.0023989526089280844, + 0.0034296861849725246, + 0.0075732627883553505, + 0.013393254019320011, + 0.0053019095212221146, + 0.007277652155607939, + 0.013607935979962349, + 0.005618932191282511, + 0.009944823570549488, + 0.007004591170698404 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 3 + ], + "selected_reward": -0.016053469851613045 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000006/group_11/index_tensor([5821, 5821])_group_11_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000006/group_11/index_tensor([5821, 5821])_group_11_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..9a06c916fe4cbd2877fe2ca3cc23c0f346f29cdc --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000006/group_11/index_tensor([5821, 5821])_group_11_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "Shot from a low angle, the camera slowly pans to the right, following the determined steps of a traveler ascending the ancient Great Wall. He is a middle-aged man, his bronzed face etched with the marks of frost. His deep eyes gaze forward, his short hair rippling gently in the sharp mountain breeze. He wears a faded khaki, multi-pocketed jacket, dusted with the dust of travel. On his back rests a large, well-worn crimson backpack. His body leans forward against the steep incline, each step steady and powerful on the vast, weathered, uneven stone steps. The golden evening sun slants down at a sharp angle, framing his profile and body with a brilliant light and casting long, dramatic shadows on his resolute figure and the undulating walls. The backdrop is the majestic Great Wall, winding its way to the horizon. Ancient beacon towers appear solemn and lonely against the backdrop of distant dark green mountains. The sky is a deep sapphire blue, dotted with clouds tinged orange-red by the setting sun. The entire scene is imbued with the rich color saturation and intense contrast of light and shadow typical of a Steve McCurry photograph. Through a documentary lens, the film captures the moment when human will intersects with the grandeur of historical landscape, creating an atmosphere of solemnity, loneliness, and power.", + "seed": 30562, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000006/group_11/index_5821_seed_30562_dp_22_rank_44_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 11, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.10786786885000765, + "vlm_logprob": -0.10786786885000765, + "vlm_yes_no_margin": 3.375, + "vlm_token_credit": -0.10786786885000765, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a middle-aged man with short hair and a bronzed face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man wearing a faded khaki multi-pocketed jacket and a large crimson backpack?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the man ascending steep, weathered stone steps of the Great Wall with steady steps?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera start at a low angle and slowly pan to the right following the man?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit by golden evening sun casting long shadows against a deep sapphire blue sky with orange-red clouds?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02380155399441719 + ], + "opposite_logprob_tokens": [ + -3.898801565170288 + ], + "logprob_reward": -0.02380155399441719, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.02380155399441719, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06388842314481735, + 0.10036695003509521, + 0.05623970553278923, + 0.10321951657533646, + 0.05646537244319916, + 0.03452293947339058, + 0.14400707185268402, + 0.2928822338581085, + 0.06787155568599701, + 0.0805361419916153 + ], + "credit_map": [ + 0.008000435307621956, + 0.006494782865047455, + 0.010513467714190483, + 0.008485429920256138, + 0.006215569097548723, + 0.010798859409987926, + 0.007395960856229067, + 0.005983924493193626, + 0.017075911164283752, + 0.018734252080321312, + 0.020806483924388885, + 0.007802720181643963, + 0.009581397287547588, + 0.010571705177426338, + 0.008594620041549206, + 0.007199870888143778, + 0.0067094918340444565, + 0.005294161383062601, + 0.010230672545731068, + 0.005641956813633442, + 0.007640170399099588, + 0.008570410311222076, + 0.007321191485971212, + 0.004831662401556969, + 0.009055908769369125, + 0.009255335666239262, + 0.012706173583865166, + 0.012936640530824661, + 0.015547344461083412, + 0.019439121708273888, + 0.012016071006655693, + 0.012262936681509018, + 0.004009595140814781, + 0.004914665129035711, + 0.008462768979370594, + 0.010216888971626759, + 0.007262952160090208, + 0.006016230210661888, + 0.008013328537344933, + 0.007568948902189732, + 0.004032941535115242, + 0.003243162529543042, + 0.006713332608342171, + 0.005273482296615839, + 0.004916145000606775, + 0.0034428508952260017, + 0.003011751687154174, + 0.00388927455060184, + 0.012498644180595875, + 0.016923358663916588, + 0.024915775284171104, + 0.024205535650253296, + 0.02225159853696823, + 0.012223266996443272, + 0.013750068843364716, + 0.01723882369697094, + 0.02817864529788494, + 0.06293486803770065, + 0.034853965044021606, + 0.03668328747153282, + 0.028202179819345474, + 0.049439072608947754, + 0.031454138457775116, + 0.021136127412319183, + 0.004785310011357069, + 0.007567836903035641, + 0.008593141101300716, + 0.00910233985632658, + 0.005739162210375071, + 0.008353669196367264, + 0.011686703190207481, + 0.012043401598930359, + 0.007488895673304796, + 0.010149762965738773, + 0.011457620188593864, + 0.01962374523282051, + 0.007266804110258818, + 0.013113278895616531, + 0.005498991347849369, + 0.005937056615948677 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 6, + 3 + ], + "selected_reward": -0.02380155399441719 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a middle-aged man with short hair and a bronzed face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man wearing a faded khaki multi-pocketed jacket and a large crimson backpack?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the man ascending steep, weathered stone steps of the Great Wall with steady steps?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera start at a low angle and slowly pan to the right following the man?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit by golden evening sun casting long shadows against a deep sapphire blue sky with orange-red clouds?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.007655093912035227 + ], + "opposite_logprob_tokens": [ + -4.882655143737793 + ], + "logprob_reward": -0.007655093912035227, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.007655093912035227, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.04976170137524605, + 0.05844511464238167, + 0.030133146792650223, + 0.04577131196856499, + 0.03342146426439285, + 0.022974397987127304, + 0.037776291370391846, + 0.6305990219116211, + 0.050853829830884933, + 0.04026365280151367 + ], + "credit_map": [ + 0.005678645800799131, + 0.009359003975987434, + 0.007924661971628666, + 0.006949078291654587, + 0.004829756915569305, + 0.007741084787994623, + 0.004059352912008762, + 0.003220113692805171, + 0.0053459168411791325, + 0.006616650614887476, + 0.006767926272004843, + 0.007694015745073557, + 0.012135972268879414, + 0.005028369836509228, + 0.009354570880532265, + 0.0055016870610415936, + 0.003906699363142252, + 0.003295749658718705, + 0.005094493739306927, + 0.0024455715902149677, + 0.0039777858182787895, + 0.004999329801648855, + 0.003533847164362669, + 0.0028796677943319082, + 0.005337016191333532, + 0.0030176991131156683, + 0.011448191478848457, + 0.00436647143214941, + 0.003781276987865567, + 0.006773678120225668, + 0.005704663693904877, + 0.005342319142073393, + 0.002396646421402693, + 0.0030845366418361664, + 0.0043917326256632805, + 0.006090602837502956, + 0.005907888989895582, + 0.003014989662915468, + 0.004103143699467182, + 0.004431923385709524, + 0.0029326409567147493, + 0.00303287198767066, + 0.003193378448486328, + 0.003135557286441326, + 0.003054297063499689, + 0.002624280285090208, + 0.0019462078344076872, + 0.0030551638919860125, + 0.0034073651768267155, + 0.005421855486929417, + 0.004838781431317329, + 0.00497241597622633, + 0.007481940556317568, + 0.004117356147617102, + 0.004561843816190958, + 0.002974735340103507, + 0.05634479224681854, + 0.10850195586681366, + 0.11042041331529617, + 0.06481541693210602, + 0.047813158482313156, + 0.15204104781150818, + 0.051892198622226715, + 0.03877006471157074, + 0.0037938959430903196, + 0.004124466795474291, + 0.007556293625384569, + 0.0066662742756307125, + 0.0038118483498692513, + 0.005998307839035988, + 0.009298598393797874, + 0.009604142047464848, + 0.00401345593854785, + 0.005176607985049486, + 0.006562121212482452, + 0.009792652912437916, + 0.0024482100270688534, + 0.00590130127966404, + 0.002984558930620551, + 0.0033847445156425238 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 1, + 8 + ], + "selected_reward": -0.007655093912035227 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a middle-aged man with short hair and a bronzed face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man wearing a faded khaki multi-pocketed jacket and a large crimson backpack?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the man ascending steep, weathered stone steps of the Great Wall with steady steps?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera start at a low angle and slowly pan to the right following the man?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit by golden evening sun casting long shadows against a deep sapphire blue sky with orange-red clouds?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.2521781623363495 + ], + "opposite_logprob_tokens": [ + -1.5021781921386719 + ], + "logprob_reward": -0.2521781623363495, + "yes_no_margin_reward": 1.25, + "token_credit_reward": -0.2521781623363495, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08540716767311096, + 0.08366349339485168, + 0.12560075521469116, + 0.07708384096622467, + 0.12849850952625275, + 0.0799761712551117, + 0.09267384558916092, + 0.16630354523658752, + 0.10303893685340881, + 0.05775374546647072 + ], + "credit_map": [ + 0.008834099397063255, + 0.00808922853320837, + 0.01029897015541792, + 0.014737701043486595, + 0.010490079410374165, + 0.017229167744517326, + 0.008830919861793518, + 0.0068970052525401115, + 0.010095709003508091, + 0.01396945584565401, + 0.008815542794764042, + 0.00849068071693182, + 0.014928898774087429, + 0.009696759283542633, + 0.011274183169007301, + 0.006392267998307943, + 0.02035762183368206, + 0.012313378043472767, + 0.029769133776426315, + 0.011288593523204327, + 0.014133886434137821, + 0.014529483392834663, + 0.012124070897698402, + 0.011084598489105701, + 0.008336633443832397, + 0.005566930864006281, + 0.01673552393913269, + 0.010693417862057686, + 0.006558992899954319, + 0.011405834928154945, + 0.009646477177739143, + 0.008140025660395622, + 0.009369425475597382, + 0.009186781011521816, + 0.024999165907502174, + 0.018771028146147728, + 0.02137494459748268, + 0.015593279153108597, + 0.014259789139032364, + 0.014944110065698624, + 0.012620768509805202, + 0.007780293468385935, + 0.014054177328944206, + 0.010330929420888424, + 0.0072059305384755135, + 0.010890979319810867, + 0.006840529851615429, + 0.010252561420202255, + 0.008538587018847466, + 0.016386939212679863, + 0.013891263864934444, + 0.014128296636044979, + 0.01328132301568985, + 0.007743275258690119, + 0.00813083816319704, + 0.010573319159448147, + 0.014780445024371147, + 0.026943111792206764, + 0.025923622772097588, + 0.016591371968388557, + 0.013943769969046116, + 0.04323355481028557, + 0.01369256991893053, + 0.01119508408010006, + 0.008097020909190178, + 0.007008017506450415, + 0.014195584692060947, + 0.013231592252850533, + 0.007697792258113623, + 0.01319321058690548, + 0.019403887912631035, + 0.020211834460496902, + 0.005286054220050573, + 0.006992500741034746, + 0.007223627995699644, + 0.012009627185761929, + 0.007348822895437479, + 0.005645001772791147, + 0.005885424092411995, + 0.007362689822912216 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 4, + 2 + ], + "selected_reward": -0.2521781623363495 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a middle-aged man with short hair and a bronzed face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man wearing a faded khaki multi-pocketed jacket and a large crimson backpack?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the man ascending steep, weathered stone steps of the Great Wall with steady steps?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera start at a low angle and slowly pan to the right following the man?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit by golden evening sun casting long shadows against a deep sapphire blue sky with orange-red clouds?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.2520887851715088 + ], + "opposite_logprob_tokens": [ + -1.5020887851715088 + ], + "logprob_reward": -0.2520887851715088, + "yes_no_margin_reward": 1.25, + "token_credit_reward": -0.2520887851715088, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08787167817354202, + 0.222413569688797, + 0.14170919358730316, + 0.11740943789482117, + 0.050012171268463135, + 0.0464567206799984, + 0.07475647330284119, + 0.09098052978515625, + 0.07030900567770004, + 0.09808123111724854 + ], + "credit_map": [ + 0.0070252628065645695, + 0.011721321381628513, + 0.010474711656570435, + 0.009483708068728447, + 0.011768462136387825, + 0.016878241673111916, + 0.013170839287340641, + 0.007349124178290367, + 0.03949655964970589, + 0.03804759308695793, + 0.05604715272784233, + 0.015488733537495136, + 0.015880752354860306, + 0.018811553716659546, + 0.02455923706293106, + 0.014081993140280247, + 0.022171657532453537, + 0.014823568053543568, + 0.03503456339240074, + 0.012857789173722267, + 0.01677648164331913, + 0.015271805226802826, + 0.012572365812957287, + 0.012200969271361828, + 0.013039488345384598, + 0.00983821414411068, + 0.01930849626660347, + 0.013873348012566566, + 0.009845901280641556, + 0.016070611774921417, + 0.013779372908174992, + 0.021654007956385612, + 0.003181705716997385, + 0.004053423181176186, + 0.008435814641416073, + 0.008379390463232994, + 0.006464995909482241, + 0.006385030224919319, + 0.006626815069466829, + 0.006484998390078545, + 0.005749226082116365, + 0.0052580805495381355, + 0.006394071504473686, + 0.007045120932161808, + 0.004002288915216923, + 0.008814004249870777, + 0.003325970144942403, + 0.00586796086281538, + 0.009167983196675777, + 0.012618576176464558, + 0.01025032065808773, + 0.009834018535912037, + 0.010282693430781364, + 0.012080090120434761, + 0.005642276722937822, + 0.004880513530224562, + 0.007496627978980541, + 0.014460109174251556, + 0.014572241343557835, + 0.01011656690388918, + 0.007278729230165482, + 0.01681755669414997, + 0.011898307129740715, + 0.008340395987033844, + 0.0056702494621276855, + 0.009814589284360409, + 0.010177558287978172, + 0.009647045284509659, + 0.006683270912617445, + 0.008479807525873184, + 0.007289818953722715, + 0.012546669691801071, + 0.008066467009484768, + 0.011155678890645504, + 0.012639005668461323, + 0.028290843591094017, + 0.006308547221124172, + 0.01726236194372177, + 0.006931813899427652, + 0.007426518015563488 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 2, + 3 + ], + "selected_reward": -0.2520887851715088 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a middle-aged man with short hair and a bronzed face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man wearing a faded khaki multi-pocketed jacket and a large crimson backpack?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the man ascending steep, weathered stone steps of the Great Wall with steady steps?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera start at a low angle and slowly pan to the right following the man?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit by golden evening sun casting long shadows against a deep sapphire blue sky with orange-red clouds?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0036157488357275724 + ], + "opposite_logprob_tokens": [ + -5.628615856170654 + ], + "logprob_reward": -0.0036157488357275724, + "yes_no_margin_reward": 5.625, + "token_credit_reward": -0.0036157488357275724, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07208545506000519, + 0.07472027838230133, + 0.22844372689723969, + 0.02984732761979103, + 0.04326479509472847, + 0.06421827524900436, + 0.10195788741111755, + 0.24547599256038666, + 0.05368546023964882, + 0.08630075305700302 + ], + "credit_map": [ + 0.008031054399907589, + 0.009281524457037449, + 0.013386706821620464, + 0.010531038977205753, + 0.007739588618278503, + 0.010858618654310703, + 0.007129676640033722, + 0.00512724881991744, + 0.009901707991957664, + 0.008245885372161865, + 0.016645485535264015, + 0.007521091494709253, + 0.008931150659918785, + 0.007175704929977655, + 0.01018709596246481, + 0.006112154573202133, + 0.04149267449975014, + 0.02504700981080532, + 0.05908197537064552, + 0.021590689197182655, + 0.02126765064895153, + 0.016405683010816574, + 0.023897545412182808, + 0.019660484045743942, + 0.0029439497739076614, + 0.0027671894058585167, + 0.005372891668230295, + 0.0034564833622425795, + 0.0032858694903552532, + 0.005313436035066843, + 0.003326292149722576, + 0.003381215501576662, + 0.0035644678864628077, + 0.003138410160318017, + 0.008717261254787445, + 0.00815917830914259, + 0.005521944724023342, + 0.0037669294979423285, + 0.004826772026717663, + 0.005569834727793932, + 0.006315445993095636, + 0.0079696299508214, + 0.0076721953228116035, + 0.010789946652948856, + 0.0073990714736282825, + 0.008030303753912449, + 0.006717263255268335, + 0.009324414655566216, + 0.010542452335357666, + 0.01451725885272026, + 0.015260308980941772, + 0.0165295097976923, + 0.013903175480663776, + 0.009256001561880112, + 0.009089288301765919, + 0.012859893962740898, + 0.022028958424925804, + 0.04529827833175659, + 0.04293067753314972, + 0.025698229670524597, + 0.01899547688663006, + 0.05536383017897606, + 0.01968667097389698, + 0.015473861247301102, + 0.004056501667946577, + 0.006329535972326994, + 0.008229728788137436, + 0.008362493477761745, + 0.004425396211445332, + 0.005981993395835161, + 0.00670680170878768, + 0.009593009017407894, + 0.006642920430749655, + 0.012755976058542728, + 0.011040093377232552, + 0.01977269910275936, + 0.007927672937512398, + 0.013212758116424084, + 0.008218260481953621, + 0.006730369292199612 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 2, + 6 + ], + "selected_reward": -0.0036157488357275724 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "Shot from a low angle, the camera slowly pans to the right, following the determined steps of a traveler ascending the ancient Great Wall. He is a middle-aged man, his bronzed face etched with the marks of frost. His deep eyes gaze forward, his short hair rippling gently in the sharp mountain breeze. He wears a faded khaki, multi-pocketed jacket, dusted with the dust of travel. On his back rests a large, well-worn crimson backpack. His body leans forward against the steep incline, each step steady and powerful on the vast, weathered, uneven stone steps. The golden evening sun slants down at a sharp angle, framing his profile and body with a brilliant light and casting long, dramatic shadows on his resolute figure and the undulating walls. The backdrop is the majestic Great Wall, winding its way to the horizon. Ancient beacon towers appear solemn and lonely against the backdrop of distant dark green mountains. The sky is a deep sapphire blue, dotted with clouds tinged orange-red by the setting sun. The entire scene is imbued with the rich color saturation and intense contrast of light and shadow typical of a Steve McCurry photograph. Through a documentary lens, the film captures the moment when human will intersects with the grandeur of historical landscape, creating an atmosphere of solemnity, loneliness, and power.", + "seed": 30563, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000006/group_11/index_5821_seed_30563_dp_22_rank_44_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 11, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.1072194998152554, + "vlm_logprob": -0.1072194998152554, + "vlm_yes_no_margin": 3.225, + "vlm_token_credit": -0.1072194998152554, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a middle-aged man with short hair and a bronzed face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man wearing a faded khaki multi-pocketed jacket and a large crimson backpack?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the man ascending steep, weathered stone steps of the Great Wall with steady steps?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera start at a low angle and slowly pan to the right following the man?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit by golden evening sun casting long shadows against a deep sapphire blue sky with orange-red clouds?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.026920655742287636 + ], + "opposite_logprob_tokens": [ + -3.7769205570220947 + ], + "logprob_reward": -0.026920655742287636, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.026920655742287636, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15813274681568146, + 0.0821879655122757, + 0.0553157739341259, + 0.12657439708709717, + 0.08990124613046646, + 0.06300318241119385, + 0.13943085074424744, + 0.04947385936975479, + 0.15193825960159302, + 0.08404173702001572 + ], + "credit_map": [ + 0.016994670033454895, + 0.020993007346987724, + 0.01600700244307518, + 0.027308616787195206, + 0.02393566630780697, + 0.015152694657444954, + 0.019141361117362976, + 0.018599733710289, + 0.010146069340407848, + 0.012157243676483631, + 0.010730582289397717, + 0.009123637340962887, + 0.009720571339130402, + 0.007589840795844793, + 0.00830565020442009, + 0.014414372853934765, + 0.007375480141490698, + 0.009388189762830734, + 0.005657436326146126, + 0.011166510172188282, + 0.005805468652397394, + 0.00537278363481164, + 0.005820577498525381, + 0.004729327280074358, + 0.013617097400128841, + 0.017567679286003113, + 0.009847389534115791, + 0.015483373776078224, + 0.018559537827968597, + 0.02214195393025875, + 0.01880582608282566, + 0.010551529005169868, + 0.009648538194596767, + 0.01878100261092186, + 0.00865184422582388, + 0.012513912282884121, + 0.014071579091250896, + 0.008640705607831478, + 0.009418618865311146, + 0.008175044320523739, + 0.010644097812473774, + 0.007649016100913286, + 0.008875634521245956, + 0.009177031926810741, + 0.0071344394236803055, + 0.0052932496182620525, + 0.006665046792477369, + 0.0075646680779755116, + 0.016450341790914536, + 0.02323957160115242, + 0.012995507568120956, + 0.034947216510772705, + 0.01779184676706791, + 0.00805919524282217, + 0.012239821255207062, + 0.013707342557609081, + 0.0058246455155313015, + 0.010762136429548264, + 0.005675387103110552, + 0.007607342209666967, + 0.006064750719815493, + 0.0031028869561851025, + 0.004215904511511326, + 0.006220803130418062, + 0.04342851787805557, + 0.015357342548668385, + 0.018047455698251724, + 0.018865834921598434, + 0.012641286477446556, + 0.015350173227488995, + 0.012619808316230774, + 0.01562783680856228, + 0.015808789059519768, + 0.011204397305846214, + 0.010113217867910862, + 0.007816554047167301, + 0.009659555740654469, + 0.0072870138101279736, + 0.009159423410892487, + 0.012992791831493378 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 6 + ], + "selected_reward": -0.026920655742287636 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a middle-aged man with short hair and a bronzed face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man wearing a faded khaki multi-pocketed jacket and a large crimson backpack?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the man ascending steep, weathered stone steps of the Great Wall with steady steps?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera start at a low angle and slowly pan to the right following the man?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit by golden evening sun casting long shadows against a deep sapphire blue sky with orange-red clouds?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.009817647747695446 + ], + "opposite_logprob_tokens": [ + -4.634817600250244 + ], + "logprob_reward": -0.009817647747695446, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.009817647747695446, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09295022487640381, + 0.1017313301563263, + 0.038121920078992844, + 0.06754841655492783, + 0.14242389798164368, + 0.11737537384033203, + 0.09149312227964401, + 0.2031944990158081, + 0.08760715276002884, + 0.05755409225821495 + ], + "credit_map": [ + 0.011800871230661869, + 0.01899341680109501, + 0.010935400612652302, + 0.00902977678924799, + 0.01088676042854786, + 0.007511436939239502, + 0.010591251775622368, + 0.01320131029933691, + 0.021217184141278267, + 0.009382392279803753, + 0.009701086208224297, + 0.006066216621547937, + 0.007090526167303324, + 0.02838156931102276, + 0.0074819037690758705, + 0.01241045631468296, + 0.007082832045853138, + 0.007967229001224041, + 0.002871881239116192, + 0.005745780188590288, + 0.0034605730324983597, + 0.00541147543117404, + 0.003229643451049924, + 0.0023525080177932978, + 0.008425701409578323, + 0.011476079002022743, + 0.008996973745524883, + 0.007514216937124729, + 0.00738010136410594, + 0.008993303403258324, + 0.009822316467761993, + 0.004939725156873465, + 0.015082272700965405, + 0.037083692848682404, + 0.009752003476023674, + 0.022214747965335846, + 0.016781466081738472, + 0.009195620194077492, + 0.020759638398885727, + 0.01155443862080574, + 0.015942534431815147, + 0.02190522849559784, + 0.014997719787061214, + 0.018590068444609642, + 0.014197849668562412, + 0.010061200708150864, + 0.011298309080302715, + 0.010382466949522495, + 0.011396495625376701, + 0.011576276272535324, + 0.008497156202793121, + 0.02511771395802498, + 0.009629922918975353, + 0.006278336979448795, + 0.007797406055033207, + 0.011199815198779106, + 0.05721553415060043, + 0.011952933855354786, + 0.05881774052977562, + 0.021299900487065315, + 0.008692821487784386, + 0.018099740147590637, + 0.017331216484308243, + 0.009784610942006111, + 0.020711829885840416, + 0.009046453051269054, + 0.010416930541396141, + 0.009062312543392181, + 0.007988967001438141, + 0.008556505665183067, + 0.008887497708201408, + 0.012936660088598728, + 0.005427490919828415, + 0.01081667747348547, + 0.00963562075048685, + 0.00952962040901184, + 0.006077601574361324, + 0.004027056507766247, + 0.005065783392637968, + 0.006974244024604559 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 4, + 5 + ], + "selected_reward": -0.009817647747695446 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a middle-aged man with short hair and a bronzed face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man wearing a faded khaki multi-pocketed jacket and a large crimson backpack?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the man ascending steep, weathered stone steps of the Great Wall with steady steps?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera start at a low angle and slowly pan to the right following the man?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit by golden evening sun casting long shadows against a deep sapphire blue sky with orange-red clouds?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.3135293424129486 + ], + "opposite_logprob_tokens": [ + -1.313529372215271 + ], + "logprob_reward": -0.3135293424129486, + "yes_no_margin_reward": 1.0, + "token_credit_reward": -0.3135293424129486, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14504307508468628, + 0.30458736419677734, + 0.03416324034333229, + 0.039397936314344406, + 0.040876034647226334, + 0.08324075490236282, + 0.13241006433963776, + 0.09379781782627106, + 0.07520288974046707, + 0.051280856132507324 + ], + "credit_map": [ + 0.020737195387482643, + 0.03368605300784111, + 0.011171325109899044, + 0.011417752131819725, + 0.012947877869009972, + 0.02041575498878956, + 0.012969786301255226, + 0.02169732004404068, + 0.0939299613237381, + 0.030296042561531067, + 0.010355055332183838, + 0.007997347973287106, + 0.010510607622563839, + 0.12212119251489639, + 0.018078958615660667, + 0.011298220604658127, + 0.0040619391947984695, + 0.0049679032526910305, + 0.0041343774646520615, + 0.006907631643116474, + 0.0037044549826532602, + 0.00334717845544219, + 0.0041379500180482864, + 0.002901802770793438, + 0.0046305106952786446, + 0.006583653390407562, + 0.00467508053407073, + 0.004595689009875059, + 0.004480667877942324, + 0.005605537910014391, + 0.005824357736855745, + 0.0030024396255612373, + 0.0057616326957941055, + 0.00854003056883812, + 0.0036666214000433683, + 0.004383452236652374, + 0.005232013761997223, + 0.003508687950670719, + 0.004917258862406015, + 0.004866340663284063, + 0.017706358805298805, + 0.006812834180891514, + 0.01584099419414997, + 0.0077554951421916485, + 0.006533673033118248, + 0.008999372832477093, + 0.006441934499889612, + 0.013150093145668507, + 0.015113424509763718, + 0.01950349658727646, + 0.01272688526660204, + 0.03299808129668236, + 0.016625545918941498, + 0.008941082283854485, + 0.011513659730553627, + 0.014987880364060402, + 0.018320556730031967, + 0.00780015904456377, + 0.015355781652033329, + 0.01296889130026102, + 0.009739937260746956, + 0.008636287413537502, + 0.007414285093545914, + 0.013561922125518322, + 0.012240713462233543, + 0.008432109840214252, + 0.010410946793854237, + 0.00657079229131341, + 0.009743287228047848, + 0.006649414077401161, + 0.007857849821448326, + 0.01329778041690588, + 0.008221580646932125, + 0.005064371973276138, + 0.005221736151725054, + 0.005398513749241829, + 0.006986618973314762, + 0.004018594045192003, + 0.006461839657276869, + 0.009907604195177555 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 6 + ], + "selected_reward": -0.3135293424129486 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a middle-aged man with short hair and a bronzed face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man wearing a faded khaki multi-pocketed jacket and a large crimson backpack?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the man ascending steep, weathered stone steps of the Great Wall with steady steps?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera start at a low angle and slowly pan to the right following the man?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit by golden evening sun casting long shadows against a deep sapphire blue sky with orange-red clouds?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.17988139390945435 + ], + "opposite_logprob_tokens": [ + -1.8048813343048096 + ], + "logprob_reward": -0.17988139390945435, + "yes_no_margin_reward": 1.625, + "token_credit_reward": -0.17988139390945435, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07734867930412292, + 0.25678542256355286, + 0.04965353012084961, + 0.09588366001844406, + 0.09007018059492111, + 0.0964433029294014, + 0.05376400053501129, + 0.03132164850831032, + 0.19898340106010437, + 0.04974614083766937 + ], + "credit_map": [ + 0.007114775478839874, + 0.008021894842386246, + 0.007107895333319902, + 0.014428005553781986, + 0.012840328738093376, + 0.006235308945178986, + 0.01238124817609787, + 0.009219224564731121, + 0.07714191824197769, + 0.02883071079850197, + 0.009065568447113037, + 0.007566335145384073, + 0.011628606356680393, + 0.09779077023267746, + 0.015811186283826828, + 0.008950318209826946, + 0.0062537360936403275, + 0.008763366378843784, + 0.005104894284158945, + 0.006587936542928219, + 0.007449310272932053, + 0.004317550454288721, + 0.006127157248556614, + 0.005049582105129957, + 0.01229598093777895, + 0.023070573806762695, + 0.007582368794828653, + 0.010870527476072311, + 0.010637798346579075, + 0.009818512946367264, + 0.012728107161819935, + 0.008879795670509338, + 0.01033036969602108, + 0.024477168917655945, + 0.006847655400633812, + 0.012697434984147549, + 0.010546619072556496, + 0.006038455758243799, + 0.012109735049307346, + 0.007022742182016373, + 0.010380247607827187, + 0.022690773010253906, + 0.012181072495877743, + 0.01361792627722025, + 0.012356515973806381, + 0.009526043199002743, + 0.009546947665512562, + 0.006143772974610329, + 0.00619920389726758, + 0.007822021842002869, + 0.006348734721541405, + 0.012892459519207478, + 0.006409995723515749, + 0.003176333149895072, + 0.004584847018122673, + 0.006330403499305248, + 0.0027757338248193264, + 0.0041254982352256775, + 0.004800783935934305, + 0.008255754597485065, + 0.002910053124651313, + 0.0017417488852515817, + 0.0029476098716259003, + 0.003764466615393758, + 0.023407097905874252, + 0.031231360509991646, + 0.043931666761636734, + 0.02600097842514515, + 0.026134822517633438, + 0.014482459984719753, + 0.018136657774448395, + 0.01565835252404213, + 0.010218696668744087, + 0.007571329828351736, + 0.007225235924124718, + 0.004584995564073324, + 0.005751882214099169, + 0.004232238046824932, + 0.004802897106856108, + 0.0053588650189340115 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 8, + 5 + ], + "selected_reward": -0.17988139390945435 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a middle-aged man with short hair and a bronzed face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man wearing a faded khaki multi-pocketed jacket and a large crimson backpack?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the man ascending steep, weathered stone steps of the Great Wall with steady steps?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera start at a low angle and slowly pan to the right following the man?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit by golden evening sun casting long shadows against a deep sapphire blue sky with orange-red clouds?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005948459263890982 + ], + "opposite_logprob_tokens": [ + -5.130948543548584 + ], + "logprob_reward": -0.005948459263890982, + "yes_no_margin_reward": 5.125, + "token_credit_reward": -0.005948459263890982, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14085626602172852, + 0.06391230970621109, + 0.21266868710517883, + 0.3221195638179779, + 0.045363154262304306, + 0.05717355012893677, + 0.05081961676478386, + 0.018501346930861473, + 0.054253995418548584, + 0.03433149680495262 + ], + "credit_map": [ + 0.019099919125437737, + 0.026674799621105194, + 0.012037854641675949, + 0.018972916528582573, + 0.011522040702402592, + 0.015864167362451553, + 0.013702739030122757, + 0.022981857880949974, + 0.011025198735296726, + 0.008804595097899437, + 0.006712604314088821, + 0.005404896568506956, + 0.007594931870698929, + 0.007664418779313564, + 0.006134788040071726, + 0.010570880956947803, + 0.03356902301311493, + 0.020283162593841553, + 0.027874654158949852, + 0.02989358827471733, + 0.017111297696828842, + 0.04124027118086815, + 0.025176871567964554, + 0.017519833520054817, + 0.05472569912672043, + 0.11055267602205276, + 0.022584091871976852, + 0.02337932400405407, + 0.031497254967689514, + 0.03420994430780411, + 0.02045704424381256, + 0.024713536724448204, + 0.005635255482047796, + 0.008422219194471836, + 0.003988997079432011, + 0.008538134396076202, + 0.004503562115132809, + 0.0042562405578792095, + 0.006077202036976814, + 0.003941547125577927, + 0.00717639084905386, + 0.013079404830932617, + 0.007280728779733181, + 0.009194038808345795, + 0.0062849693931639194, + 0.004612503107637167, + 0.0049558160826563835, + 0.004589700140058994, + 0.006495864596217871, + 0.006926460191607475, + 0.0051084221340715885, + 0.010861341841518879, + 0.005593206267803907, + 0.004619138315320015, + 0.004101892467588186, + 0.00711329560726881, + 0.0024705585092306137, + 0.002598769497126341, + 0.0019688329193741083, + 0.002634275471791625, + 0.002514839870855212, + 0.0016013026470318437, + 0.002287559676915407, + 0.0024252107832580805, + 0.0068456693552434444, + 0.008030429482460022, + 0.011656823568046093, + 0.007590387016534805, + 0.005182846449315548, + 0.004186565056443214, + 0.00506630539894104, + 0.005694975610822439, + 0.00411201361566782, + 0.0076411874033510685, + 0.006249300669878721, + 0.004238223657011986, + 0.0036817356012761593, + 0.0026981714181602, + 0.00298019265756011, + 0.0027306759729981422 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 2, + 0 + ], + "selected_reward": -0.005948459263890982 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "Shot from a low angle, the camera slowly pans to the right, following the determined steps of a traveler ascending the ancient Great Wall. He is a middle-aged man, his bronzed face etched with the marks of frost. His deep eyes gaze forward, his short hair rippling gently in the sharp mountain breeze. He wears a faded khaki, multi-pocketed jacket, dusted with the dust of travel. On his back rests a large, well-worn crimson backpack. His body leans forward against the steep incline, each step steady and powerful on the vast, weathered, uneven stone steps. The golden evening sun slants down at a sharp angle, framing his profile and body with a brilliant light and casting long, dramatic shadows on his resolute figure and the undulating walls. The backdrop is the majestic Great Wall, winding its way to the horizon. Ancient beacon towers appear solemn and lonely against the backdrop of distant dark green mountains. The sky is a deep sapphire blue, dotted with clouds tinged orange-red by the setting sun. The entire scene is imbued with the rich color saturation and intense contrast of light and shadow typical of a Steve McCurry photograph. Through a documentary lens, the film captures the moment when human will intersects with the grandeur of historical landscape, creating an atmosphere of solemnity, loneliness, and power.", + "seed": 30564, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000006/group_11/index_5821_seed_30564_dp_23_rank_46_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 11, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.04693056400865316, + "vlm_logprob": -0.04693056400865316, + "vlm_yes_no_margin": 3.575, + "vlm_token_credit": -0.04693056400865316, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a middle-aged man with short hair and a bronzed face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man wearing a faded khaki multi-pocketed jacket and a large crimson backpack?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the man ascending steep, weathered stone steps of the Great Wall with steady steps?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera start at a low angle and slowly pan to the right following the man?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit by golden evening sun casting long shadows against a deep sapphire blue sky with orange-red clouds?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.08070243149995804 + ], + "opposite_logprob_tokens": [ + -2.580702543258667 + ], + "logprob_reward": -0.08070243149995804, + "yes_no_margin_reward": 2.5, + "token_credit_reward": -0.08070243149995804, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13042782247066498, + 0.13008493185043335, + 0.030343234539031982, + 0.04898803308606148, + 0.05534880608320236, + 0.21632172167301178, + 0.08654599636793137, + 0.10554559528827667, + 0.1320665031671524, + 0.06432738900184631 + ], + "credit_map": [ + 0.012471533380448818, + 0.021881455555558205, + 0.008481576107442379, + 0.0120569858700037, + 0.027603358030319214, + 0.02224167250096798, + 0.013951689936220646, + 0.011739554814994335, + 0.005355701316148043, + 0.047315459698438644, + 0.0033351583406329155, + 0.005476245190948248, + 0.05140816792845726, + 0.007115527987480164, + 0.006011355202645063, + 0.004067311063408852, + 0.002804260468110442, + 0.0050539905205369, + 0.00277585256844759, + 0.0032689515501260757, + 0.00452266214415431, + 0.0052802665159106255, + 0.0027705735992640257, + 0.003866677638143301, + 0.005733445752412081, + 0.004874394740909338, + 0.005068627651780844, + 0.008055719546973705, + 0.005888475105166435, + 0.006990342400968075, + 0.006976727396249771, + 0.0054002986289560795, + 0.006439471151679754, + 0.0054766028188169, + 0.006837083492428064, + 0.008009642362594604, + 0.006732489448040724, + 0.007155886385589838, + 0.007294314447790384, + 0.007403318304568529, + 0.023261895403265953, + 0.02359522506594658, + 0.0233553908765316, + 0.029860036447644234, + 0.021325577050447464, + 0.022064942866563797, + 0.03584032133221626, + 0.03701833635568619, + 0.009295365773141384, + 0.013949167914688587, + 0.008683552034199238, + 0.012398850172758102, + 0.009794244542717934, + 0.009694572538137436, + 0.009763251058757305, + 0.012966998852789402, + 0.009163259528577328, + 0.011162001639604568, + 0.011459286324679852, + 0.013543426059186459, + 0.010015349835157394, + 0.00877130962908268, + 0.01585312932729721, + 0.025577830150723457, + 0.013985134661197662, + 0.028140541166067123, + 0.013428423553705215, + 0.023795273154973984, + 0.01619277521967888, + 0.010717550292611122, + 0.013622443191707134, + 0.01218436285853386, + 0.006189391948282719, + 0.00684776809066534, + 0.008863076567649841, + 0.007351886946707964, + 0.009060867130756378, + 0.004629687871783972, + 0.005768160801380873, + 0.015616550110280514 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 8, + 0 + ], + "selected_reward": -0.08070243149995804 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a middle-aged man with short hair and a bronzed face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man wearing a faded khaki multi-pocketed jacket and a large crimson backpack?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the man ascending steep, weathered stone steps of the Great Wall with steady steps?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera start at a low angle and slowly pan to the right following the man?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit by golden evening sun casting long shadows against a deep sapphire blue sky with orange-red clouds?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005275732837617397 + ], + "opposite_logprob_tokens": [ + -5.255275726318359 + ], + "logprob_reward": -0.005275732837617397, + "yes_no_margin_reward": 5.25, + "token_credit_reward": -0.005275732837617397, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09181243181228638, + 0.13049133121967316, + 0.03288125991821289, + 0.038733307272195816, + 0.07950123399496078, + 0.15445959568023682, + 0.11072271317243576, + 0.12517638504505157, + 0.1751558929681778, + 0.06106584146618843 + ], + "credit_map": [ + 0.008529174141585827, + 0.012699625454843044, + 0.007473313715308905, + 0.012886500917375088, + 0.010377433151006699, + 0.013325827196240425, + 0.016419226303696632, + 0.010101335123181343, + 0.005865324288606644, + 0.04388849809765816, + 0.004428851418197155, + 0.00630780728533864, + 0.04976396635174751, + 0.0076241204515099525, + 0.0060243019834160805, + 0.006588470656424761, + 0.0027435272932052612, + 0.005269528832286596, + 0.003299704287201166, + 0.003710114164277911, + 0.005036641843616962, + 0.004566450137645006, + 0.0035731547977775335, + 0.004682137630879879, + 0.005492253229022026, + 0.003892194014042616, + 0.004290672484785318, + 0.004895171616226435, + 0.003862825920805335, + 0.006189744919538498, + 0.00436309864744544, + 0.0057473452761769295, + 0.010772752575576305, + 0.00784254353493452, + 0.011217284947633743, + 0.013746727257966995, + 0.007968046702444553, + 0.007806531619280577, + 0.011974925175309181, + 0.008172420784831047, + 0.014333829283714294, + 0.016035934910178185, + 0.01957070827484131, + 0.02156323753297329, + 0.01355462521314621, + 0.016547461971640587, + 0.02452671341598034, + 0.0283270925283432, + 0.011189397424459457, + 0.012805317528545856, + 0.01150460448116064, + 0.012839232571423054, + 0.011736857704818249, + 0.012512629851698875, + 0.01856156997382641, + 0.01957309991121292, + 0.010969039984047413, + 0.01856866665184498, + 0.012309201993048191, + 0.01776253804564476, + 0.014846575446426868, + 0.013897011987864971, + 0.013257864862680435, + 0.023565486073493958, + 0.01680656149983406, + 0.03732443228363991, + 0.018109770491719246, + 0.028719516471028328, + 0.020563611760735512, + 0.013483745977282524, + 0.016794908791780472, + 0.023353347554802895, + 0.005257862154394388, + 0.006921632215380669, + 0.007990550249814987, + 0.008650037460029125, + 0.00761431036517024, + 0.004732369910925627, + 0.006853323895484209, + 0.01304575800895691 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 5, + 1 + ], + "selected_reward": -0.005275732837617397 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a middle-aged man with short hair and a bronzed face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man wearing a faded khaki multi-pocketed jacket and a large crimson backpack?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the man ascending steep, weathered stone steps of the Great Wall with steady steps?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera start at a low angle and slowly pan to the right following the man?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit by golden evening sun casting long shadows against a deep sapphire blue sky with orange-red clouds?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07900431752204895 + ], + "opposite_logprob_tokens": [ + -2.5790042877197266 + ], + "logprob_reward": -0.07900431752204895, + "yes_no_margin_reward": 2.5, + "token_credit_reward": -0.07900431752204895, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1389034241437912, + 0.07506353408098221, + 0.024698184803128242, + 0.05245887488126755, + 0.06578195095062256, + 0.12166048586368561, + 0.1061340942978859, + 0.10019401460886002, + 0.26765692234039307, + 0.04744857922196388 + ], + "credit_map": [ + 0.012259148992598057, + 0.019836898893117905, + 0.009197126142680645, + 0.017290282994508743, + 0.02268427610397339, + 0.021583300083875656, + 0.02350708097219467, + 0.012545304372906685, + 0.005138284992426634, + 0.01828032173216343, + 0.0045057861134409904, + 0.006176797207444906, + 0.02102065458893776, + 0.005011359695345163, + 0.008182121440768242, + 0.00674819853156805, + 0.00247203279286623, + 0.004353814758360386, + 0.0024617360904812813, + 0.002955283038318157, + 0.004569641780108213, + 0.002043962012976408, + 0.003116774605587125, + 0.0027249366976320744, + 0.005503260996192694, + 0.005668144207447767, + 0.004608395975083113, + 0.005046492908149958, + 0.004711623303592205, + 0.012755488976836205, + 0.005346229765564203, + 0.008819232694804668, + 0.010449526831507683, + 0.0076936716213822365, + 0.007863848470151424, + 0.009813406504690647, + 0.006136043928563595, + 0.006974763236939907, + 0.00940788071602583, + 0.007442797999829054, + 0.012479793280363083, + 0.012286958284676075, + 0.015665479004383087, + 0.017221854999661446, + 0.010469517670571804, + 0.012393767945468426, + 0.018951378762722015, + 0.022191720083355904, + 0.012332987040281296, + 0.012186470441520214, + 0.009839667938649654, + 0.011858136393129826, + 0.010815849527716637, + 0.015031217597424984, + 0.01669776625931263, + 0.017371993511915207, + 0.0109583530575037, + 0.00993772130459547, + 0.010782813653349876, + 0.0180145762860775, + 0.009624069556593895, + 0.00869622454047203, + 0.014140348881483078, + 0.018039900809526443, + 0.02920881286263466, + 0.058807533234357834, + 0.02436545118689537, + 0.046281538903713226, + 0.03316469117999077, + 0.022256409749388695, + 0.02667948417365551, + 0.02689298987388611, + 0.004054127726703882, + 0.00477192597463727, + 0.005956448148936033, + 0.007225735578685999, + 0.0037804236635565758, + 0.0034841899760067463, + 0.00528323370963335, + 0.01289248839020729 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 0, + 5 + ], + "selected_reward": -0.07900431752204895 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a middle-aged man with short hair and a bronzed face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man wearing a faded khaki multi-pocketed jacket and a large crimson backpack?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the man ascending steep, weathered stone steps of the Great Wall with steady steps?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera start at a low angle and slowly pan to the right following the man?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit by golden evening sun casting long shadows against a deep sapphire blue sky with orange-red clouds?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.06204257160425186 + ], + "opposite_logprob_tokens": [ + -2.812042474746704 + ], + "logprob_reward": -0.06204257160425186, + "yes_no_margin_reward": 2.75, + "token_credit_reward": -0.06204257160425186, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06063898652791977, + 0.38259193301200867, + 0.08285924792289734, + 0.08999019861221313, + 0.0653698742389679, + 0.04904663935303688, + 0.04893966019153595, + 0.05667370185256004, + 0.14038881659507751, + 0.023500945419073105 + ], + "credit_map": [ + 0.00572566594928503, + 0.006693697068840265, + 0.004843854811042547, + 0.010260834358632565, + 0.0072346897795796394, + 0.007241005077958107, + 0.011682026088237762, + 0.0069572157226502895, + 0.01954514905810356, + 0.1100856140255928, + 0.015935493633151054, + 0.022917786613106728, + 0.1321491301059723, + 0.022551709786057472, + 0.03179135546088219, + 0.027615686878561974, + 0.007119724527001381, + 0.017596662044525146, + 0.005320744123309851, + 0.006482677534222603, + 0.01715458743274212, + 0.012145135551691055, + 0.009312947280704975, + 0.0077267722226679325, + 0.008057170547544956, + 0.009182496927678585, + 0.008013038896024227, + 0.010399974882602692, + 0.009363731369376183, + 0.020342258736491203, + 0.010866341181099415, + 0.013765188865363598, + 0.007231556344777346, + 0.007311078254133463, + 0.006044213660061359, + 0.006249759811908007, + 0.004721449688076973, + 0.007810303475707769, + 0.008653026074171066, + 0.017348483204841614, + 0.0053833285346627235, + 0.005038125906139612, + 0.005166225600987673, + 0.007020467426627874, + 0.007738620974123478, + 0.005821343511343002, + 0.005805306602269411, + 0.007073221728205681, + 0.004866771399974823, + 0.0045691546984016895, + 0.007023575250059366, + 0.005876197479665279, + 0.004399653524160385, + 0.004225092474371195, + 0.007982457056641579, + 0.009996756911277771, + 0.005576831288635731, + 0.00604590866714716, + 0.005420721136033535, + 0.0066446904093027115, + 0.004884711466729641, + 0.005731177981942892, + 0.004946028348058462, + 0.017423631623387337, + 0.01639174111187458, + 0.03260453790426254, + 0.013396956957876682, + 0.02436636947095394, + 0.017339499667286873, + 0.012166989967226982, + 0.014089672826230526, + 0.01003305334597826, + 0.0022170478478074074, + 0.0027880568522959948, + 0.003600707510486245, + 0.003090781159698963, + 0.0024383249692618847, + 0.001605187775567174, + 0.001986933872103691, + 0.005773904267698526 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 8, + 3 + ], + "selected_reward": -0.06204257160425186 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a middle-aged man with short hair and a bronzed face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man wearing a faded khaki multi-pocketed jacket and a large crimson backpack?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the man ascending steep, weathered stone steps of the Great Wall with steady steps?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera start at a low angle and slowly pan to the right following the man?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit by golden evening sun casting long shadows against a deep sapphire blue sky with orange-red clouds?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.007627766579389572 + ], + "opposite_logprob_tokens": [ + -4.882627964019775 + ], + "logprob_reward": -0.007627766579389572, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.007627766579389572, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.22321024537086487, + 0.12365663051605225, + 0.19193238019943237, + 0.048208992928266525, + 0.050903547555208206, + 0.11569907516241074, + 0.06286334246397018, + 0.05903077870607376, + 0.05396408215165138, + 0.0705309584736824 + ], + "credit_map": [ + 0.026749154552817345, + 0.03558558225631714, + 0.013596129603683949, + 0.023096662014722824, + 0.041882455348968506, + 0.042243920266628265, + 0.02108231745660305, + 0.01897400990128517, + 0.006358043290674686, + 0.03816789761185646, + 0.0066521321423351765, + 0.007227302994579077, + 0.04161606356501579, + 0.006559484638273716, + 0.009371384978294373, + 0.007704323623329401, + 0.030095629394054413, + 0.0420265793800354, + 0.01744261384010315, + 0.016870230436325073, + 0.030841417610645294, + 0.019301068037748337, + 0.01582842506468296, + 0.019526418298482895, + 0.006640064064413309, + 0.004746615886688232, + 0.008109004236757755, + 0.007575614377856255, + 0.004768493585288525, + 0.004927003290504217, + 0.004642557818442583, + 0.006799640133976936, + 0.007217606995254755, + 0.005135166924446821, + 0.007993927225470543, + 0.006931307725608349, + 0.0058076088316738605, + 0.00510697765275836, + 0.006342174950987101, + 0.006368774455040693, + 0.01454679574817419, + 0.011760002933442593, + 0.012317448854446411, + 0.01528256107121706, + 0.011181671172380447, + 0.011801201850175858, + 0.018469350412487984, + 0.02034003473818302, + 0.007113439030945301, + 0.007467867806553841, + 0.007101312279701233, + 0.007222683168947697, + 0.005894281901419163, + 0.006432620342820883, + 0.008846230804920197, + 0.012784910388290882, + 0.004449047148227692, + 0.007597642485052347, + 0.005108697339892387, + 0.01103740930557251, + 0.006535767111927271, + 0.005250937305390835, + 0.010840678587555885, + 0.008210599422454834, + 0.007101509720087051, + 0.008777699433267117, + 0.006150505971163511, + 0.009092800319194794, + 0.006443298887461424, + 0.0049923560582101345, + 0.004336371552199125, + 0.007069536950439215, + 0.006578791420906782, + 0.00912557728588581, + 0.00914710108190775, + 0.011778950691223145, + 0.00836371909826994, + 0.005519444588571787, + 0.007528660818934441, + 0.012488714419305325 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 1 + ], + "selected_reward": -0.007627766579389572 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "Shot from a low angle, the camera slowly pans to the right, following the determined steps of a traveler ascending the ancient Great Wall. He is a middle-aged man, his bronzed face etched with the marks of frost. His deep eyes gaze forward, his short hair rippling gently in the sharp mountain breeze. He wears a faded khaki, multi-pocketed jacket, dusted with the dust of travel. On his back rests a large, well-worn crimson backpack. His body leans forward against the steep incline, each step steady and powerful on the vast, weathered, uneven stone steps. The golden evening sun slants down at a sharp angle, framing his profile and body with a brilliant light and casting long, dramatic shadows on his resolute figure and the undulating walls. The backdrop is the majestic Great Wall, winding its way to the horizon. Ancient beacon towers appear solemn and lonely against the backdrop of distant dark green mountains. The sky is a deep sapphire blue, dotted with clouds tinged orange-red by the setting sun. The entire scene is imbued with the rich color saturation and intense contrast of light and shadow typical of a Steve McCurry photograph. Through a documentary lens, the film captures the moment when human will intersects with the grandeur of historical landscape, creating an atmosphere of solemnity, loneliness, and power.", + "seed": 30565, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000006/group_11/index_5821_seed_30565_dp_23_rank_46_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 11, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.10905365133658051, + "vlm_logprob": -0.10905365133658051, + "vlm_yes_no_margin": 3.125, + "vlm_token_credit": -0.10905365133658051, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a middle-aged man with short hair and a bronzed face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man wearing a faded khaki multi-pocketed jacket and a large crimson backpack?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the man ascending steep, weathered stone steps of the Great Wall with steady steps?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera start at a low angle and slowly pan to the right following the man?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit by golden evening sun casting long shadows against a deep sapphire blue sky with orange-red clouds?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03563729673624039 + ], + "opposite_logprob_tokens": [ + -3.410637378692627 + ], + "logprob_reward": -0.03563729673624039, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.03563729673624039, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0738024190068245, + 0.04860444366931915, + 0.06493394821882248, + 0.07962195575237274, + 0.10574409365653992, + 0.03995508328080177, + 0.0729091688990593, + 0.15088491141796112, + 0.32244428992271423, + 0.041099704802036285 + ], + "credit_map": [ + 0.0065284306183457375, + 0.00974809005856514, + 0.009063705801963806, + 0.012731652706861496, + 0.008412274532020092, + 0.007319292519241571, + 0.009179314598441124, + 0.010819658637046814, + 0.004021438304334879, + 0.0036932360380887985, + 0.005917111411690712, + 0.008976854383945465, + 0.007447047624737024, + 0.004867780487984419, + 0.007935605011880398, + 0.0057453676126897335, + 0.007828292436897755, + 0.011923643760383129, + 0.012850497849285603, + 0.0069512538611888885, + 0.007335934787988663, + 0.005005798768252134, + 0.009043966419994831, + 0.003994558937847614, + 0.008790824562311172, + 0.008161600679159164, + 0.014572870917618275, + 0.013114853762090206, + 0.011251194402575493, + 0.006626685615628958, + 0.008343706838786602, + 0.008760223165154457, + 0.013489257544279099, + 0.015522214584052563, + 0.014479816891252995, + 0.016036497429013252, + 0.012478293851017952, + 0.008890722878277302, + 0.014340977184474468, + 0.010506315156817436, + 0.0035641926806420088, + 0.0032535502687096596, + 0.004017678555101156, + 0.004434596747159958, + 0.0042708758264780045, + 0.003177252132445574, + 0.007766967639327049, + 0.009469971060752869, + 0.008653457276523113, + 0.0067969257943332195, + 0.008999216370284557, + 0.00741991912946105, + 0.008306203410029411, + 0.005781190935522318, + 0.013708728365600109, + 0.013243532739579678, + 0.01972709223628044, + 0.016973158344626427, + 0.022865142673254013, + 0.019854741171002388, + 0.01911497302353382, + 0.012233983725309372, + 0.017397677525877953, + 0.022718144580721855, + 0.036574479192495346, + 0.02635563164949417, + 0.08646152168512344, + 0.026437390595674515, + 0.02314860001206398, + 0.054207686334848404, + 0.023395972326397896, + 0.04586298391222954, + 0.003322941716760397, + 0.0034238682128489017, + 0.00801345705986023, + 0.005383374635130167, + 0.004835311323404312, + 0.002652810886502266, + 0.004351458512246609, + 0.00911648664623499 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 7, + 4 + ], + "selected_reward": -0.03563729673624039 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a middle-aged man with short hair and a bronzed face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man wearing a faded khaki multi-pocketed jacket and a large crimson backpack?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the man ascending steep, weathered stone steps of the Great Wall with steady steps?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera start at a low angle and slowly pan to the right following the man?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit by golden evening sun casting long shadows against a deep sapphire blue sky with orange-red clouds?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.00866025872528553 + ], + "opposite_logprob_tokens": [ + -4.758660316467285 + ], + "logprob_reward": -0.00866025872528553, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.00866025872528553, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08996772021055222, + 0.045583389699459076, + 0.1075613722205162, + 0.06425761431455612, + 0.2537526488304138, + 0.02869189716875553, + 0.10435032844543457, + 0.10744684189558029, + 0.12806878983974457, + 0.07031933218240738 + ], + "credit_map": [ + 0.012266075238585472, + 0.009751015342772007, + 0.010912015102803707, + 0.011464370414614677, + 0.009570298716425896, + 0.01184643805027008, + 0.01016145572066307, + 0.01399605069309473, + 0.005124242510646582, + 0.004118308424949646, + 0.0065935757011175156, + 0.007945154793560505, + 0.005120816640555859, + 0.00483431713655591, + 0.007858244702219963, + 0.00398873258382082, + 0.01142209768295288, + 0.02026963047683239, + 0.019110752269625664, + 0.010194185189902782, + 0.015410644933581352, + 0.00828569009900093, + 0.01562383584678173, + 0.007244538981467485, + 0.006568628828972578, + 0.008637740276753902, + 0.011043368838727474, + 0.009752136655151844, + 0.00807507336139679, + 0.004502993077039719, + 0.007863091304898262, + 0.007814586162567139, + 0.033296506851911545, + 0.03358272463083267, + 0.032122936099767685, + 0.04284871369600296, + 0.031263165175914764, + 0.020700793713331223, + 0.03266153112053871, + 0.027276288717985153, + 0.003161270171403885, + 0.0027342336252331734, + 0.005104760639369488, + 0.002764300210401416, + 0.0022161509841680527, + 0.002382010454311967, + 0.004495540168136358, + 0.005833631847053766, + 0.012853519059717655, + 0.007398529909551144, + 0.013600229285657406, + 0.010698847472667694, + 0.010808451101183891, + 0.00886896625161171, + 0.015817832201719284, + 0.02430395595729351, + 0.01309169176965952, + 0.012824064120650291, + 0.014486061409115791, + 0.014368265867233276, + 0.013879883103072643, + 0.009046429768204689, + 0.012463154271245003, + 0.017287295311689377, + 0.012257259339094162, + 0.014566057361662388, + 0.02710578776896, + 0.011915330775082111, + 0.012123046442866325, + 0.023211738094687462, + 0.013528254814445972, + 0.01336131151765585, + 0.004905742593109608, + 0.006316805258393288, + 0.012212906964123249, + 0.008886571042239666, + 0.008675344288349152, + 0.0043576862663030624, + 0.008640105836093426, + 0.01632416993379593 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 8, + 2 + ], + "selected_reward": -0.00866025872528553 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a middle-aged man with short hair and a bronzed face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man wearing a faded khaki multi-pocketed jacket and a large crimson backpack?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the man ascending steep, weathered stone steps of the Great Wall with steady steps?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera start at a low angle and slowly pan to the right following the man?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit by golden evening sun casting long shadows against a deep sapphire blue sky with orange-red clouds?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.1799207180738449 + ], + "opposite_logprob_tokens": [ + -1.8049206733703613 + ], + "logprob_reward": -0.1799207180738449, + "yes_no_margin_reward": 1.625, + "token_credit_reward": -0.1799207180738449, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10148046910762787, + 0.06809715926647186, + 0.04125372692942619, + 0.07060149312019348, + 0.1313437968492508, + 0.06719795614480972, + 0.22068659961223602, + 0.11209730058908463, + 0.11993852257728577, + 0.06730300933122635 + ], + "credit_map": [ + 0.011596946977078915, + 0.009462484158575535, + 0.012723303399980068, + 0.012950499542057514, + 0.010979913175106049, + 0.012971175834536552, + 0.010242355987429619, + 0.020553797483444214, + 0.006746217142790556, + 0.005843175109475851, + 0.007897010073065758, + 0.011656107380986214, + 0.011015974916517735, + 0.006370950490236282, + 0.010255654342472553, + 0.008312067948281765, + 0.0041888682171702385, + 0.005145110655575991, + 0.005216735880821943, + 0.0041297138668596745, + 0.008086095564067364, + 0.0032167253084480762, + 0.005417591892182827, + 0.005852884612977505, + 0.007609134074300528, + 0.00799945555627346, + 0.012525681406259537, + 0.012410469353199005, + 0.00923112127929926, + 0.005897357128560543, + 0.008131179958581924, + 0.006797099951654673, + 0.01359726581722498, + 0.01906553842127323, + 0.02081657387316227, + 0.01809624582529068, + 0.02281327359378338, + 0.01027593668550253, + 0.01575750857591629, + 0.01092146523296833, + 0.008169541135430336, + 0.006301181856542826, + 0.011496293358504772, + 0.007805098779499531, + 0.006184175610542297, + 0.005360724404454231, + 0.01020017173141241, + 0.01168077066540718, + 0.028355859220027924, + 0.013743615709245205, + 0.031636793166399, + 0.018487706780433655, + 0.021553346887230873, + 0.019515464082360268, + 0.031032593920826912, + 0.05636122450232506, + 0.01658887229859829, + 0.0107803326100111, + 0.01571527309715748, + 0.012898663058876991, + 0.012643299996852875, + 0.008844870142638683, + 0.013616707175970078, + 0.021009279415011406, + 0.01383612398058176, + 0.017180107533931732, + 0.02430061437189579, + 0.012293430045247078, + 0.014252140186727047, + 0.013832361437380314, + 0.009764716029167175, + 0.014479030855000019, + 0.005869229324162006, + 0.005482770502567291, + 0.011805545538663864, + 0.008005011826753616, + 0.0036772098392248154, + 0.004454663954675198, + 0.008249534294009209, + 0.0197590459138155 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 4, + 8 + ], + "selected_reward": -0.1799207180738449 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a middle-aged man with short hair and a bronzed face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man wearing a faded khaki multi-pocketed jacket and a large crimson backpack?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the man ascending steep, weathered stone steps of the Great Wall with steady steps?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera start at a low angle and slowly pan to the right following the man?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit by golden evening sun casting long shadows against a deep sapphire blue sky with orange-red clouds?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.3134239912033081 + ], + "opposite_logprob_tokens": [ + -1.313423991203308 + ], + "logprob_reward": -0.3134239912033081, + "yes_no_margin_reward": 1.0, + "token_credit_reward": -0.3134239912033081, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0946860983967781, + 0.06996075809001923, + 0.29335668683052063, + 0.11428669840097427, + 0.10266296565532684, + 0.04369585961103439, + 0.15635204315185547, + 0.04188981652259827, + 0.05104445293545723, + 0.032064639031887054 + ], + "credit_map": [ + 0.009142732247710228, + 0.007758943364024162, + 0.011482419446110725, + 0.009726909920573235, + 0.012809854932129383, + 0.008889990858733654, + 0.017713293433189392, + 0.017161954194307327, + 0.006049522198736668, + 0.00657646032050252, + 0.01041394378989935, + 0.015251662582159042, + 0.006112172268331051, + 0.006502037867903709, + 0.010610140860080719, + 0.008444813080132008, + 0.024782082065939903, + 0.03454374149441719, + 0.04145128279924393, + 0.031264327466487885, + 0.04852881282567978, + 0.024374932050704956, + 0.05757318064570427, + 0.03083832375705242, + 0.012093969620764256, + 0.014328626915812492, + 0.01374748069792986, + 0.015253507532179356, + 0.01533844880759716, + 0.010278641246259212, + 0.014408120885491371, + 0.018837912008166313, + 0.01709642820060253, + 0.015311970375478268, + 0.01723799668252468, + 0.009333070367574692, + 0.011542401276528835, + 0.008299588225781918, + 0.012707415036857128, + 0.01113409548997879, + 0.0039304462261497974, + 0.004365931265056133, + 0.005676275584846735, + 0.0036661175545305014, + 0.005606411024928093, + 0.003516916884109378, + 0.008001157082617283, + 0.008932603523135185, + 0.01415182277560234, + 0.00925908051431179, + 0.023875009268522263, + 0.01344096940010786, + 0.015278285369277, + 0.013551346957683563, + 0.0252548661082983, + 0.04154067113995552, + 0.00418855668976903, + 0.004577888175845146, + 0.00739828497171402, + 0.006540242582559586, + 0.004912411794066429, + 0.0031573709566146135, + 0.005456507205963135, + 0.005658555310219526, + 0.005401721224188805, + 0.008886908181011677, + 0.00621597608551383, + 0.005880618002265692, + 0.009043154306709766, + 0.004709383472800255, + 0.006235834211111069, + 0.00467086024582386, + 0.002693645190447569, + 0.004269354045391083, + 0.004290256183594465, + 0.004975058138370514, + 0.00299832527525723, + 0.00202201004140079, + 0.0051151965744793415, + 0.005700794979929924 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 6, + 3 + ], + "selected_reward": -0.3134239912033081 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a middle-aged man with short hair and a bronzed face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man wearing a faded khaki multi-pocketed jacket and a large crimson backpack?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the man ascending steep, weathered stone steps of the Great Wall with steady steps?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera start at a low angle and slowly pan to the right following the man?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit by golden evening sun casting long shadows against a deep sapphire blue sky with orange-red clouds?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.007625991944223642 + ], + "opposite_logprob_tokens": [ + -4.882626056671143 + ], + "logprob_reward": -0.007625991944223642, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.007625991944223642, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15443775057792664, + 0.1879538595676422, + 0.05845065042376518, + 0.026163842529058456, + 0.2224661260843277, + 0.04616033285856247, + 0.0391339510679245, + 0.08889156579971313, + 0.04292672872543335, + 0.1334151327610016 + ], + "credit_map": [ + 0.020572153851389885, + 0.013708679936826229, + 0.017674019560217857, + 0.016056684777140617, + 0.02101389318704605, + 0.016900543123483658, + 0.022979341447353363, + 0.02553243190050125, + 0.02449321188032627, + 0.01276878546923399, + 0.019375206902623177, + 0.03569983318448067, + 0.02174127846956253, + 0.018263759091496468, + 0.025156129151582718, + 0.030455663800239563, + 0.00554241519421339, + 0.009453901089727879, + 0.009903979487717152, + 0.00735066831111908, + 0.009725834242999554, + 0.00393537525087595, + 0.007851834408938885, + 0.004686637781560421, + 0.002946948865428567, + 0.0027277464978396893, + 0.004339498933404684, + 0.004539447836577892, + 0.003209738526493311, + 0.0021518822759389877, + 0.003015077207237482, + 0.0032335033174604177, + 0.03102141059935093, + 0.02782966010272503, + 0.03890186548233032, + 0.032051265239715576, + 0.025671035051345825, + 0.01807236857712269, + 0.02390187792479992, + 0.025016654282808304, + 0.0048429458402097225, + 0.004123402759432793, + 0.0070540523156523705, + 0.005537366960197687, + 0.0036987753119319677, + 0.0035562110133469105, + 0.007918007671833038, + 0.00942957028746605, + 0.004988248459994793, + 0.0024420504923909903, + 0.006096346769481897, + 0.0029225258622318506, + 0.0029430920258164406, + 0.003161897649988532, + 0.005007516592741013, + 0.011572272516787052, + 0.014137753285467625, + 0.006769523024559021, + 0.012132895179092884, + 0.009728221222758293, + 0.007516477257013321, + 0.007358008995652199, + 0.013204251416027546, + 0.018044432625174522, + 0.0056353043764829636, + 0.005182023160159588, + 0.008066784590482712, + 0.004100436344742775, + 0.005740817170590162, + 0.0047763641923666, + 0.00412747822701931, + 0.005297518335282803, + 0.012329796329140663, + 0.012917214073240757, + 0.025705676525831223, + 0.01649841107428074, + 0.0153459832072258, + 0.011648050509393215, + 0.015979494899511337, + 0.0229905154556036 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 1, + 0 + ], + "selected_reward": -0.007625991944223642 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000006/group_12/index_tensor([7645, 7645])_group_12_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000006/group_12/index_tensor([7645, 7645])_group_12_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..f0d8e117b791689f6931d4e220f52b73000317ff --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000006/group_12/index_tensor([7645, 7645])_group_12_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "Filmed from a static, extremely low angle, just inches from the ground, a Ford Mustang RTR SEN 5 and a Honda Civic Type R engage in a fierce, catastrophic clash. The camera's fixed position emphasizes the sheer scale and raw power of the two vehicles as their bodies collide in a brutal, inelastic collision. The Ford Mustang's chassis, painted glossy Electric Blue, showcasing its aggressive widebody kit and distinctive RTR grille, now presses fiercely against the side of the Honda Civic Type R, a machine rendered in vibrant Rally Red. The moment of impact is a frozen spectacle of destruction: the Mustang's front fender and the Civic's driver's door deform and fuse together, metal ripped and bent inward. The collision, rendered in an impactful graphic design style, unleashes an explosion of stylized effects; sharp, geometric fragments of silver and black metal fly from the collision point like graphic carriers, while oversized, starburst-shaped sparks burst forth in vibrant yellow and orange, boldly outlined. The scene is saturated with bold, high-contrast lighting that carves deep, dramatic shadows and casts sharp, shimmering highlights on the cars’ twisted surfaces, emphasizing the raw force of the collision and creating an atmosphere of tense, high-energy controlled chaos.", + "seed": 19486, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000006/group_12/index_7645_seed_19486_dp_24_rank_48_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 12, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.24084855951368808, + "vlm_logprob": -0.24084855951368808, + "vlm_yes_no_margin": 3.525, + "vlm_token_credit": -0.24084855951368808, + "vlm_yes_ratio": 0.8, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned at an extremely low angle, just inches from the ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a Ford Mustang RTR SEN 5 painted glossy Electric Blue with a widebody kit visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a Honda Civic Type R painted vibrant Rally Red visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the vehicles shown in a collision where the Mustang's front fender and the Civic's driver's door are deformed and fused together?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are sharp, geometric fragments of silver and black metal flying from the collision point?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -1.1632777452468872 + ], + "opposite_logprob_tokens": [ + -0.4132777750492096 + ], + "logprob_reward": -1.1632777452468872, + "yes_no_margin_reward": -0.75, + "token_credit_reward": -1.1632777452468872, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.21436122059822083, + 0.09275956451892853, + 0.07206219434738159, + 0.04046504199504852, + 0.07570653408765793, + 0.10184206813573837, + 0.08394471555948257, + 0.07844985276460648, + 0.1630946546792984, + 0.0773141160607338 + ], + "credit_map": [ + 0.023574214428663254, + 0.032853566110134125, + 0.029727106913924217, + 0.02418464794754982, + 0.03133029118180275, + 0.02516905963420868, + 0.027393708005547523, + 0.0201286468654871, + 0.010115370154380798, + 0.01243018638342619, + 0.010456133633852005, + 0.01023818738758564, + 0.011684638448059559, + 0.013476678170263767, + 0.013508169911801815, + 0.010850204154849052, + 0.007598790805786848, + 0.009396407753229141, + 0.009843640960752964, + 0.009588895365595818, + 0.009773002937436104, + 0.009334372356534004, + 0.008283684961497784, + 0.008243408054113388, + 0.004166085738688707, + 0.002954215044155717, + 0.006195237394422293, + 0.008893772028386593, + 0.00284664798527956, + 0.0051731448620557785, + 0.007102688308805227, + 0.003133252263069153, + 0.011569681577384472, + 0.006912435404956341, + 0.007784903980791569, + 0.008359075523912907, + 0.008122267201542854, + 0.009329278953373432, + 0.012410218827426434, + 0.011218685656785965, + 0.010563699528574944, + 0.006538804154843092, + 0.01377711072564125, + 0.01397705078125, + 0.006111632101237774, + 0.007419220171868801, + 0.014829267747700214, + 0.02862529084086418, + 0.00951000489294529, + 0.007102848496288061, + 0.015135666355490685, + 0.01585475355386734, + 0.00618283124640584, + 0.004396036732941866, + 0.014094440266489983, + 0.011668147519230843, + 0.009788540191948414, + 0.00637427531182766, + 0.008398759178817272, + 0.013856280595064163, + 0.00746080931276083, + 0.009306215681135654, + 0.01330616045743227, + 0.00995881948620081, + 0.012546899728477001, + 0.009680922143161297, + 0.010181460529565811, + 0.031847380101680756, + 0.010448602959513664, + 0.007330717518925667, + 0.04477860778570175, + 0.036280080676078796, + 0.004736380651593208, + 0.0059176357463002205, + 0.008608958683907986, + 0.006889090407639742, + 0.005690507125109434, + 0.005968871992081404, + 0.023427285254001617, + 0.016075391322374344 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 5 + ], + "selected_reward": -1.1632777452468872 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned at an extremely low angle, just inches from the ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a Ford Mustang RTR SEN 5 painted glossy Electric Blue with a widebody kit visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a Honda Civic Type R painted vibrant Rally Red visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the vehicles shown in a collision where the Mustang's front fender and the Civic's driver's door are deformed and fused together?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are sharp, geometric fragments of silver and black metal flying from the collision point?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.008690157905220985 + ], + "opposite_logprob_tokens": [ + -4.758690357208252 + ], + "logprob_reward": -0.008690157905220985, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.008690157905220985, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0979534313082695, + 0.06828448176383972, + 0.12225905805826187, + 0.03933368995785713, + 0.07333743572235107, + 0.0412185937166214, + 0.18272578716278076, + 0.07722481340169907, + 0.12819021940231323, + 0.16947242617607117 + ], + "credit_map": [ + 0.010029563680291176, + 0.013058792799711227, + 0.016018636524677277, + 0.009343702346086502, + 0.012419695034623146, + 0.015520400367677212, + 0.009507137350738049, + 0.012055502273142338, + 0.007992030121386051, + 0.010100673884153366, + 0.009690501727163792, + 0.009934546425938606, + 0.005761736072599888, + 0.008627095259726048, + 0.008966957218945026, + 0.007210944313555956, + 0.016820477321743965, + 0.011476115323603153, + 0.014441785402595997, + 0.017649304121732712, + 0.019041189923882484, + 0.013665174134075642, + 0.016477467492222786, + 0.012687546201050282, + 0.005618819035589695, + 0.010573760606348515, + 0.0030157470609992743, + 0.004461511969566345, + 0.003343875054270029, + 0.004677807912230492, + 0.004089903552085161, + 0.003552261972799897, + 0.012304689735174179, + 0.012324145995080471, + 0.007357422262430191, + 0.0074765486642718315, + 0.009048555046319962, + 0.011425377801060677, + 0.006474125199019909, + 0.006926575675606728, + 0.004548685159534216, + 0.0034819580614566803, + 0.006459272466599941, + 0.005485102999955416, + 0.002936306409537792, + 0.0034801724832504988, + 0.005551546346396208, + 0.009275550954043865, + 0.01736273802816868, + 0.019006939604878426, + 0.028198642656207085, + 0.031830236315727234, + 0.015979068353772163, + 0.013812250457704067, + 0.02847524732351303, + 0.028060654178261757, + 0.009580490179359913, + 0.01214566733688116, + 0.010583135299384594, + 0.015103999525308609, + 0.004637689329683781, + 0.005013574380427599, + 0.01023031771183014, + 0.009929937310516834, + 0.01694967970252037, + 0.015353175811469555, + 0.015856139361858368, + 0.018643051385879517, + 0.008956250734627247, + 0.00701843062415719, + 0.024195149540901184, + 0.02121833898127079, + 0.007442822679877281, + 0.007870244793593884, + 0.011923006735742092, + 0.014910727739334106, + 0.0093559380620718, + 0.009848426096141338, + 0.06503874808549881, + 0.04308252036571503 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 9, + 8 + ], + "selected_reward": -0.008690157905220985 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned at an extremely low angle, just inches from the ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a Ford Mustang RTR SEN 5 painted glossy Electric Blue with a widebody kit visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a Honda Civic Type R painted vibrant Rally Red visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the vehicles shown in a collision where the Mustang's front fender and the Civic's driver's door are deformed and fused together?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are sharp, geometric fragments of silver and black metal flying from the collision point?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01108712051063776 + ], + "opposite_logprob_tokens": [ + -4.511086940765381 + ], + "logprob_reward": -0.01108712051063776, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.01108712051063776, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08385767042636871, + 0.051089923828840256, + 0.11265160143375397, + 0.07172265648841858, + 0.03566844388842583, + 0.04923194274306297, + 0.07134944945573807, + 0.05376242473721504, + 0.09406422078609467, + 0.37660160660743713 + ], + "credit_map": [ + 0.010524838231503963, + 0.00965950544923544, + 0.01271701231598854, + 0.009778481908142567, + 0.009653143584728241, + 0.01289570052176714, + 0.008735130541026592, + 0.009893855080008507, + 0.00664859265089035, + 0.006409130524843931, + 0.005446429830044508, + 0.005675481166690588, + 0.004616532474756241, + 0.009688304737210274, + 0.0065036737360060215, + 0.006101779639720917, + 0.016103088855743408, + 0.01109800860285759, + 0.01251673698425293, + 0.016341401264071465, + 0.0165550634264946, + 0.014885183423757553, + 0.014926545321941376, + 0.010225582867860794, + 0.007670836988836527, + 0.009841552004218102, + 0.014624367468059063, + 0.01793643832206726, + 0.00458337040618062, + 0.005273772403597832, + 0.008099096827208996, + 0.0036932246293872595, + 0.004228409845381975, + 0.003348710248246789, + 0.003312063869088888, + 0.0036614462733268738, + 0.004800493363291025, + 0.009215750731527805, + 0.003785673063248396, + 0.003315896727144718, + 0.004367407876998186, + 0.002741748234257102, + 0.006527260411530733, + 0.007784002460539341, + 0.0031209182925522327, + 0.0033063595183193684, + 0.009190848097205162, + 0.01219339668750763, + 0.006983900908380747, + 0.006855558604001999, + 0.010003307834267616, + 0.010920913890004158, + 0.004359688144177198, + 0.006482882425189018, + 0.014230669476091862, + 0.011512530967593193, + 0.009195709601044655, + 0.005715343169867992, + 0.005194505676627159, + 0.008837010711431503, + 0.004558710381388664, + 0.0035105478018522263, + 0.00810875091701746, + 0.008641844615340233, + 0.009444858878850937, + 0.008537713438272476, + 0.008183525875210762, + 0.019843103364109993, + 0.006179879419505596, + 0.004725433420389891, + 0.01998327299952507, + 0.01716643013060093, + 0.016545647755265236, + 0.024629194289445877, + 0.021737899631261826, + 0.03189466893672943, + 0.022735074162483215, + 0.025694984942674637, + 0.14070376753807068, + 0.09266039729118347 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 2, + 8 + ], + "selected_reward": -0.01108712051063776 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned at an extremely low angle, just inches from the ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a Ford Mustang RTR SEN 5 painted glossy Electric Blue with a widebody kit visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a Honda Civic Type R painted vibrant Rally Red visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the vehicles shown in a collision where the Mustang's front fender and the Civic's driver's door are deformed and fused together?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are sharp, geometric fragments of silver and black metal flying from the collision point?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.012545124627649784 + ], + "opposite_logprob_tokens": [ + -4.387545108795166 + ], + "logprob_reward": -0.012545124627649784, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.012545124627649784, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15995226800441742, + 0.236824169754982, + 0.07879316061735153, + 0.033949971199035645, + 0.147939994931221, + 0.01935914158821106, + 0.04897031560540199, + 0.09574941545724869, + 0.10459830611944199, + 0.07386324554681778 + ], + "credit_map": [ + 0.02166878432035446, + 0.01919574663043022, + 0.01579231023788452, + 0.008494849316775799, + 0.03441822901368141, + 0.029441136866807938, + 0.013597165234386921, + 0.017344066873192787, + 0.025964397937059402, + 0.034531500190496445, + 0.02974252589046955, + 0.027760686352849007, + 0.0260272528976202, + 0.03216448798775673, + 0.03440120071172714, + 0.02623213641345501, + 0.008037015795707703, + 0.0057298424653708935, + 0.01258753426373005, + 0.01344326976686716, + 0.013138440437614918, + 0.008983968757092953, + 0.009485692717134953, + 0.00738739687949419, + 0.004287962801754475, + 0.007340908516198397, + 0.0028362360317260027, + 0.003182249842211604, + 0.003270321059972048, + 0.005448824726045132, + 0.00427979277446866, + 0.003303673816844821, + 0.018722854554653168, + 0.01610180176794529, + 0.020323624834418297, + 0.019895853474736214, + 0.016738925129175186, + 0.028974860906600952, + 0.015581128187477589, + 0.01160094141960144, + 0.002748528029769659, + 0.0015107300132513046, + 0.001996498554944992, + 0.002809007652103901, + 0.002008462557569146, + 0.002649547066539526, + 0.0023676962591707706, + 0.0032686740159988403, + 0.00710947485640645, + 0.004646265879273415, + 0.006740432698279619, + 0.008147473447024822, + 0.005031082313507795, + 0.004020703490823507, + 0.006711957044899464, + 0.006562928669154644, + 0.012097308412194252, + 0.009728524833917618, + 0.010190285742282867, + 0.01667349599301815, + 0.007237791083753109, + 0.009986722841858864, + 0.01604701578617096, + 0.013788276351988316, + 0.009753967635333538, + 0.004856577143073082, + 0.011153094470500946, + 0.021555159240961075, + 0.009163912385702133, + 0.006342678796499968, + 0.02368095889687538, + 0.0180919598788023, + 0.007420273032039404, + 0.006086610723286867, + 0.005717901978641748, + 0.007241236977279186, + 0.00630087498575449, + 0.006308644078671932, + 0.021080274134874344, + 0.013707439415156841 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 4 + ], + "selected_reward": -0.012545124627649784 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned at an extremely low angle, just inches from the ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a Ford Mustang RTR SEN 5 painted glossy Electric Blue with a widebody kit visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a Honda Civic Type R painted vibrant Rally Red visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the vehicles shown in a collision where the Mustang's front fender and the Civic's driver's door are deformed and fused together?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are sharp, geometric fragments of silver and black metal flying from the collision point?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0086426492780447 + ], + "opposite_logprob_tokens": [ + -4.758642673492432 + ], + "logprob_reward": -0.0086426492780447, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.0086426492780447, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12507441639900208, + 0.14336061477661133, + 0.16493478417396545, + 0.02532353065907955, + 0.27941837906837463, + 0.07392610609531403, + 0.03502139076590538, + 0.063516765832901, + 0.02974349819123745, + 0.05968056246638298 + ], + "credit_map": [ + 0.016058890148997307, + 0.014187349937856197, + 0.017201099544763565, + 0.009115568362176418, + 0.02227998524904251, + 0.020130831748247147, + 0.013253102079033852, + 0.012847580015659332, + 0.015982504934072495, + 0.024008436128497124, + 0.016991250216960907, + 0.016043758019804955, + 0.015178438276052475, + 0.019083188846707344, + 0.020388349890708923, + 0.015684688463807106, + 0.02474641613662243, + 0.014028796926140785, + 0.01121577713638544, + 0.011016874574124813, + 0.03865165263414383, + 0.03356141224503517, + 0.01730753295123577, + 0.014406324364244938, + 0.0029941636603325605, + 0.003388022305443883, + 0.00397713016718626, + 0.004982963670045137, + 0.0019086359534412622, + 0.0034706806764006615, + 0.0026146431919187307, + 0.001987291034311056, + 0.037292852997779846, + 0.02794739045202732, + 0.03798584267497063, + 0.03856887295842171, + 0.02782059833407402, + 0.04789356142282486, + 0.035169199109077454, + 0.026740070432424545, + 0.007481668144464493, + 0.00583238061517477, + 0.009700664319097996, + 0.009079836308956146, + 0.007493333891034126, + 0.00802561268210411, + 0.010671541094779968, + 0.015641070902347565, + 0.006082724779844284, + 0.003470356110483408, + 0.0029408757109194994, + 0.004523538053035736, + 0.004113595467060804, + 0.004082067869603634, + 0.004723544232547283, + 0.005084690637886524, + 0.007215013727545738, + 0.005364902317523956, + 0.007426837459206581, + 0.008913365192711353, + 0.007212826982140541, + 0.008762240409851074, + 0.010004124604165554, + 0.008617452345788479, + 0.00308767007663846, + 0.0019291017670184374, + 0.0027302035596221685, + 0.003781975945457816, + 0.0031646692659705877, + 0.0027381409890949726, + 0.006677870638668537, + 0.005633868742734194, + 0.0060441461391747, + 0.006040262524038553, + 0.004995279014110565, + 0.006490750703960657, + 0.005376559682190418, + 0.008217566646635532, + 0.013111796230077744, + 0.009404203854501247 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 2, + 1 + ], + "selected_reward": -0.0086426492780447 + } + ], + "vlm_decoded_outputs": [ + "No", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "Filmed from a static, extremely low angle, just inches from the ground, a Ford Mustang RTR SEN 5 and a Honda Civic Type R engage in a fierce, catastrophic clash. The camera's fixed position emphasizes the sheer scale and raw power of the two vehicles as their bodies collide in a brutal, inelastic collision. The Ford Mustang's chassis, painted glossy Electric Blue, showcasing its aggressive widebody kit and distinctive RTR grille, now presses fiercely against the side of the Honda Civic Type R, a machine rendered in vibrant Rally Red. The moment of impact is a frozen spectacle of destruction: the Mustang's front fender and the Civic's driver's door deform and fuse together, metal ripped and bent inward. The collision, rendered in an impactful graphic design style, unleashes an explosion of stylized effects; sharp, geometric fragments of silver and black metal fly from the collision point like graphic carriers, while oversized, starburst-shaped sparks burst forth in vibrant yellow and orange, boldly outlined. The scene is saturated with bold, high-contrast lighting that carves deep, dramatic shadows and casts sharp, shimmering highlights on the cars’ twisted surfaces, emphasizing the raw force of the collision and creating an atmosphere of tense, high-energy controlled chaos.", + "seed": 19487, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000006/group_12/index_7645_seed_19487_dp_24_rank_48_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 12, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.17792804776690901, + "vlm_logprob": -0.17792804776690901, + "vlm_yes_no_margin": 3.925, + "vlm_token_credit": -0.17792804776690901, + "vlm_yes_ratio": 0.8, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned at an extremely low angle, just inches from the ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a Ford Mustang RTR SEN 5 painted glossy Electric Blue with a widebody kit visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a Honda Civic Type R painted vibrant Rally Red visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the vehicles shown in a collision where the Mustang's front fender and the Civic's driver's door are deformed and fused together?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are sharp, geometric fragments of silver and black metal flying from the collision point?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -0.8581302165985107 + ], + "opposite_logprob_tokens": [ + -0.6081302165985107 + ], + "logprob_reward": -0.8581302165985107, + "yes_no_margin_reward": -0.25, + "token_credit_reward": -0.8581302165985107, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.5794949531555176, + 0.02878059446811676, + 0.034881722182035446, + 0.05364101007580757, + 0.060109950602054596, + 0.1024823859333992, + 0.03432518616318703, + 0.041024450212717056, + 0.02303740754723549, + 0.04222232475876808 + ], + "credit_map": [ + 0.18112149834632874, + 0.02552323043346405, + 0.044873934239149094, + 0.05029670521616936, + 0.028977200388908386, + 0.19078147411346436, + 0.023765362799167633, + 0.03415556997060776, + 0.0031530638225376606, + 0.00445777690038085, + 0.0027298901695758104, + 0.0030262477230280638, + 0.0029803006909787655, + 0.004345435183495283, + 0.0035754800774157047, + 0.004512399435043335, + 0.003416599938645959, + 0.00456648925319314, + 0.003710824763402343, + 0.002812665421515703, + 0.0030644882936030626, + 0.008509300649166107, + 0.003591458545997739, + 0.005209894850850105, + 0.008909350261092186, + 0.009031756781041622, + 0.007608457934111357, + 0.00463394820690155, + 0.004515056498348713, + 0.007627769373357296, + 0.005102900322526693, + 0.006211768835783005, + 0.010183636099100113, + 0.0039843409322202206, + 0.006257916800677776, + 0.0047112200409173965, + 0.0046187499538064, + 0.01774027571082115, + 0.005040238611400127, + 0.007573573384433985, + 0.0067181880585849285, + 0.0070173353888094425, + 0.015325993299484253, + 0.018301062285900116, + 0.008087724447250366, + 0.008414538577198982, + 0.02437334693968296, + 0.01424419041723013, + 0.004434308037161827, + 0.005748513154685497, + 0.004347392823547125, + 0.0052077206782996655, + 0.0027075994294136763, + 0.004834363702684641, + 0.0034588531125336885, + 0.003586434992030263, + 0.002493995940312743, + 0.002298843115568161, + 0.002971649868413806, + 0.0141179533675313, + 0.007306322921067476, + 0.003102209186181426, + 0.0035686837509274483, + 0.005164789035916328, + 0.0017460343660786748, + 0.0015762397088110447, + 0.0024325812701135874, + 0.005200523417443037, + 0.0022982200607657433, + 0.002201395109295845, + 0.003624916775152087, + 0.003957497421652079, + 0.003210115944966674, + 0.0020815443713217974, + 0.005084057804197073, + 0.007744498085230589, + 0.002975336043164134, + 0.004980612546205521, + 0.010068348608911037, + 0.006077810190618038 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 5, + 4 + ], + "selected_reward": -0.8581302165985107 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned at an extremely low angle, just inches from the ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a Ford Mustang RTR SEN 5 painted glossy Electric Blue with a widebody kit visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a Honda Civic Type R painted vibrant Rally Red visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the vehicles shown in a collision where the Mustang's front fender and the Civic's driver's door are deformed and fused together?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are sharp, geometric fragments of silver and black metal flying from the collision point?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005987801123410463 + ], + "opposite_logprob_tokens": [ + -5.130987644195557 + ], + "logprob_reward": -0.005987801123410463, + "yes_no_margin_reward": 5.125, + "token_credit_reward": -0.005987801123410463, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.34643012285232544, + 0.12925568222999573, + 0.03517771139740944, + 0.04496996849775314, + 0.13501371443271637, + 0.07123280316591263, + 0.0863456204533577, + 0.026693660765886307, + 0.07602539658546448, + 0.04885534569621086 + ], + "credit_map": [ + 0.09444646537303925, + 0.015128550119698048, + 0.029375417158007622, + 0.02937774918973446, + 0.023915497586131096, + 0.10716302692890167, + 0.01919809728860855, + 0.027825312688946724, + 0.013625189661979675, + 0.02174573577940464, + 0.011552184820175171, + 0.012003099545836449, + 0.012807179242372513, + 0.021239949390292168, + 0.014409306459128857, + 0.02187303639948368, + 0.002875171136111021, + 0.004218997899442911, + 0.0038737140130251646, + 0.003808618988841772, + 0.003211692674085498, + 0.008000199683010578, + 0.00433720275759697, + 0.004852111451327801, + 0.005546249449253082, + 0.0062668039463460445, + 0.007575382012873888, + 0.0037362428847700357, + 0.004235690459609032, + 0.006597655825316906, + 0.00571838952600956, + 0.005293555557727814, + 0.014604097232222557, + 0.006858668755739927, + 0.013114693574607372, + 0.015686064958572388, + 0.010487484745681286, + 0.04791233688592911, + 0.011092777363955975, + 0.01525758858770132, + 0.005597097333520651, + 0.008904943242669106, + 0.014971833676099777, + 0.009730263613164425, + 0.005356425419449806, + 0.006460658274590969, + 0.01147006917744875, + 0.008741512894630432, + 0.010230094194412231, + 0.01569019816815853, + 0.01125373225659132, + 0.011553141288459301, + 0.006284050643444061, + 0.012229529209434986, + 0.009519488550722599, + 0.009585392661392689, + 0.0019627874717116356, + 0.002002024557441473, + 0.003056108485907316, + 0.007595856208354235, + 0.003447893774136901, + 0.002398434095084667, + 0.0030479852575808764, + 0.0031825690530240536, + 0.006448092870414257, + 0.004446305800229311, + 0.008028536103665829, + 0.009202929213643074, + 0.007074587047100067, + 0.007516794838011265, + 0.014273826964199543, + 0.019034331664443016, + 0.004339604172855616, + 0.0024855455849319696, + 0.0059477537870407104, + 0.007283293176442385, + 0.0029112494084984064, + 0.004058965481817722, + 0.009698486886918545, + 0.012130443938076496 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 1 + ], + "selected_reward": -0.005987801123410463 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned at an extremely low angle, just inches from the ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a Ford Mustang RTR SEN 5 painted glossy Electric Blue with a widebody kit visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a Honda Civic Type R painted vibrant Rally Red visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the vehicles shown in a collision where the Mustang's front fender and the Civic's driver's door are deformed and fused together?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are sharp, geometric fragments of silver and black metal flying from the collision point?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0031986285466700792 + ], + "opposite_logprob_tokens": [ + -5.753198623657227 + ], + "logprob_reward": -0.0031986285466700792, + "yes_no_margin_reward": 5.75, + "token_credit_reward": -0.0031986285466700792, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2911430597305298, + 0.08250503242015839, + 0.03555919975042343, + 0.04001259431242943, + 0.15581580996513367, + 0.11997681856155396, + 0.08693380653858185, + 0.0850682184100151, + 0.044955745339393616, + 0.05802976340055466 + ], + "credit_map": [ + 0.07625524699687958, + 0.01571342535316944, + 0.02490614540874958, + 0.024652091786265373, + 0.02206234075129032, + 0.0879085510969162, + 0.01676267385482788, + 0.02288258634507656, + 0.008342033252120018, + 0.01272581983357668, + 0.007982436567544937, + 0.008109766989946365, + 0.008385686203837395, + 0.018216589465737343, + 0.00939721055328846, + 0.009345484897494316, + 0.0028746931347995996, + 0.006922046188265085, + 0.003160907654091716, + 0.0027900650165975094, + 0.0033641529735177755, + 0.007868224754929543, + 0.0036374488845467567, + 0.004941659513860941, + 0.004833625163882971, + 0.007713128812611103, + 0.005448457319289446, + 0.002833010395988822, + 0.003730370197445154, + 0.00570782320573926, + 0.0048690601252019405, + 0.004877115599811077, + 0.013228761963546276, + 0.010007570497691631, + 0.019462648779153824, + 0.045853663235902786, + 0.008123530074954033, + 0.03140193596482277, + 0.014494535513222218, + 0.013243168592453003, + 0.009537562727928162, + 0.009009144268929958, + 0.022023161873221397, + 0.02042543515563011, + 0.009097080677747726, + 0.00967987161129713, + 0.022923678159713745, + 0.017280878499150276, + 0.010150366462767124, + 0.013037468306720257, + 0.011670694686472416, + 0.01235282700508833, + 0.00631431769579649, + 0.011910825967788696, + 0.01101991068571806, + 0.010477394796907902, + 0.005786946974694729, + 0.004715580493211746, + 0.00807149987667799, + 0.025630200281739235, + 0.013579994440078735, + 0.008204552344977856, + 0.008422189392149448, + 0.01065725740045309, + 0.004955186974257231, + 0.0030127817299216986, + 0.005113889463245869, + 0.008271249942481518, + 0.0024266494438052177, + 0.003704636823385954, + 0.007371409330517054, + 0.010099940001964569, + 0.004626511130481958, + 0.0035618795081973076, + 0.005424723494797945, + 0.012168576940894127, + 0.002410524757578969, + 0.0035049603320658207, + 0.015412732027471066, + 0.010919850319623947 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 5 + ], + "selected_reward": -0.0031986285466700792 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned at an extremely low angle, just inches from the ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a Ford Mustang RTR SEN 5 painted glossy Electric Blue with a widebody kit visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a Honda Civic Type R painted vibrant Rally Red visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the vehicles shown in a collision where the Mustang's front fender and the Civic's driver's door are deformed and fused together?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are sharp, geometric fragments of silver and black metal flying from the collision point?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.012540886178612709 + ], + "opposite_logprob_tokens": [ + -4.387540817260742 + ], + "logprob_reward": -0.012540886178612709, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.012540886178612709, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.24228885769844055, + 0.07502461224794388, + 0.022185716778039932, + 0.0564069002866745, + 0.11911226063966751, + 0.12868230044841766, + 0.04939549043774605, + 0.07972973585128784, + 0.15992854535579681, + 0.06724557280540466 + ], + "credit_map": [ + 0.06259433180093765, + 0.010655071586370468, + 0.020810874179005623, + 0.020883910357952118, + 0.015130750834941864, + 0.07380162179470062, + 0.016307370737195015, + 0.022104918956756592, + 0.008707868866622448, + 0.009407277218997478, + 0.008314375765621662, + 0.008342931047081947, + 0.006975231692194939, + 0.01613602228462696, + 0.008058174513280392, + 0.009082731790840626, + 0.0023189412895590067, + 0.0024720828514546156, + 0.0028481055051088333, + 0.002850069897249341, + 0.0022569955326616764, + 0.003979015629738569, + 0.0031132095027714968, + 0.0023472970351576805, + 0.007990123704075813, + 0.009756251238286495, + 0.008127494715154171, + 0.00450084637850523, + 0.0047575184144079685, + 0.008142361417412758, + 0.006352466531097889, + 0.006779834628105164, + 0.021618476137518883, + 0.006895037833601236, + 0.008771887049078941, + 0.014464350417256355, + 0.008747268468141556, + 0.036619897931814194, + 0.008687278255820274, + 0.01330806128680706, + 0.00809050165116787, + 0.00807955116033554, + 0.023267487064003944, + 0.026991786435246468, + 0.007807810325175524, + 0.010572397150099277, + 0.02701481059193611, + 0.016857946291565895, + 0.004871412180364132, + 0.0038690960500389338, + 0.006295935250818729, + 0.009870401583611965, + 0.00429808534681797, + 0.0069270385429263115, + 0.0071054482832551, + 0.006158068776130676, + 0.0055933562107384205, + 0.004667474422603846, + 0.006684048566967249, + 0.02590078115463257, + 0.01359992753714323, + 0.005455999169498682, + 0.008112767711281776, + 0.00971537921577692, + 0.007771420758217573, + 0.006233770865947008, + 0.007090865168720484, + 0.030546752735972404, + 0.014113183133304119, + 0.011574017815291882, + 0.024952532723546028, + 0.0576460100710392, + 0.004541222471743822, + 0.004284203052520752, + 0.007582792080938816, + 0.010042647831141949, + 0.0038308785296976566, + 0.004666432738304138, + 0.017347803339362144, + 0.01494959369301796 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 5 + ], + "selected_reward": -0.012540886178612709 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned at an extremely low angle, just inches from the ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a Ford Mustang RTR SEN 5 painted glossy Electric Blue with a widebody kit visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a Honda Civic Type R painted vibrant Rally Red visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the vehicles shown in a collision where the Mustang's front fender and the Civic's driver's door are deformed and fused together?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are sharp, geometric fragments of silver and black metal flying from the collision point?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.009782706387341022 + ], + "opposite_logprob_tokens": [ + -4.634782791137695 + ], + "logprob_reward": -0.009782706387341022, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.009782706387341022, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08963015675544739, + 0.01874161884188652, + 0.07492037862539291, + 0.1198384091258049, + 0.1559760421514511, + 0.1374656856060028, + 0.02810269594192505, + 0.02686168998479843, + 0.25876814126968384, + 0.08969520777463913 + ], + "credit_map": [ + 0.015465480275452137, + 0.006037719547748566, + 0.008349630050361156, + 0.008385428227484226, + 0.009439609944820404, + 0.023623326793313026, + 0.00908743217587471, + 0.009241536259651184, + 0.00219685398042202, + 0.0031730218324810266, + 0.001678269705735147, + 0.001577049377374351, + 0.0025150172878056765, + 0.003028841456398368, + 0.001824293634854257, + 0.0027482728473842144, + 0.0077728270553052425, + 0.010658550076186657, + 0.009738098829984665, + 0.007277880795300007, + 0.008731938898563385, + 0.012999625876545906, + 0.0084385322406888, + 0.009302925318479538, + 0.01658179983496666, + 0.018433883786201477, + 0.01699485071003437, + 0.010384410619735718, + 0.012633709236979485, + 0.017053648829460144, + 0.013186335563659668, + 0.014569764956831932, + 0.012370482087135315, + 0.00810515508055687, + 0.015168001875281334, + 0.025185763835906982, + 0.010363738052546978, + 0.05297534540295601, + 0.014173443429172039, + 0.017634106799960136, + 0.011275414377450943, + 0.011448991484940052, + 0.018472488969564438, + 0.0230618417263031, + 0.012153889052569866, + 0.012370056472718716, + 0.02888520061969757, + 0.019797801971435547, + 0.0042814770713448524, + 0.0052987574599683285, + 0.0038074434269219637, + 0.0036482932046055794, + 0.001862629665993154, + 0.003421627450734377, + 0.002692220965400338, + 0.003090245882049203, + 0.002397495089098811, + 0.0021584774367511272, + 0.0028223907575011253, + 0.007245284970849752, + 0.004229433368891478, + 0.0026496928185224533, + 0.0022881331387907267, + 0.003070781473070383, + 0.01197613961994648, + 0.01203572191298008, + 0.012235619127750397, + 0.041307270526885986, + 0.02585778199136257, + 0.022535135969519615, + 0.04297678545117378, + 0.08984368294477463, + 0.00606438796967268, + 0.004148584324866533, + 0.007977055385708809, + 0.027030503377318382, + 0.005027254577726126, + 0.0077383387833833694, + 0.01645413413643837, + 0.015254941768944263 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 4, + 5 + ], + "selected_reward": -0.009782706387341022 + } + ], + "vlm_decoded_outputs": [ + "No", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "Filmed from a static, extremely low angle, just inches from the ground, a Ford Mustang RTR SEN 5 and a Honda Civic Type R engage in a fierce, catastrophic clash. The camera's fixed position emphasizes the sheer scale and raw power of the two vehicles as their bodies collide in a brutal, inelastic collision. The Ford Mustang's chassis, painted glossy Electric Blue, showcasing its aggressive widebody kit and distinctive RTR grille, now presses fiercely against the side of the Honda Civic Type R, a machine rendered in vibrant Rally Red. The moment of impact is a frozen spectacle of destruction: the Mustang's front fender and the Civic's driver's door deform and fuse together, metal ripped and bent inward. The collision, rendered in an impactful graphic design style, unleashes an explosion of stylized effects; sharp, geometric fragments of silver and black metal fly from the collision point like graphic carriers, while oversized, starburst-shaped sparks burst forth in vibrant yellow and orange, boldly outlined. The scene is saturated with bold, high-contrast lighting that carves deep, dramatic shadows and casts sharp, shimmering highlights on the cars’ twisted surfaces, emphasizing the raw force of the collision and creating an atmosphere of tense, high-energy controlled chaos.", + "seed": 19488, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000006/group_12/index_7645_seed_19488_dp_25_rank_50_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 12, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.16560031427070498, + "vlm_logprob": -0.16560031427070498, + "vlm_yes_no_margin": 3.725, + "vlm_token_credit": -0.16560031427070498, + "vlm_yes_ratio": 0.8, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned at an extremely low angle, just inches from the ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a Ford Mustang RTR SEN 5 painted glossy Electric Blue with a widebody kit visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a Honda Civic Type R painted vibrant Rally Red visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the vehicles shown in a collision where the Mustang's front fender and the Civic's driver's door are deformed and fused together?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are sharp, geometric fragments of silver and black metal flying from the collision point?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -0.7866209149360657 + ], + "opposite_logprob_tokens": [ + -0.6616209149360657 + ], + "logprob_reward": -0.7866209149360657, + "yes_no_margin_reward": -0.125, + "token_credit_reward": -0.7866209149360657, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.21450284123420715, + 0.048813849687576294, + 0.10849148035049438, + 0.2577503025531769, + 0.05084752291440964, + 0.11979721486568451, + 0.06649056822061539, + 0.04481784999370575, + 0.03216585889458656, + 0.05632253363728523 + ], + "credit_map": [ + 0.025276124477386475, + 0.04508469998836517, + 0.015687283128499985, + 0.013122767210006714, + 0.028542740270495415, + 0.035804565995931625, + 0.023554546758532524, + 0.027430115267634392, + 0.0043265363201498985, + 0.005858164280653, + 0.006546152289956808, + 0.005489804781973362, + 0.006759752985090017, + 0.004820102825760841, + 0.007832979783415794, + 0.007180357351899147, + 0.01528718788176775, + 0.014750339090824127, + 0.01732279732823372, + 0.010923733003437519, + 0.014059692621231079, + 0.015967408195137978, + 0.011111876927316189, + 0.009068449027836323, + 0.05626857653260231, + 0.020559169352054596, + 0.01901501603424549, + 0.011714519001543522, + 0.03359135985374451, + 0.08551030606031418, + 0.013813992030918598, + 0.01727738231420517, + 0.006121923215687275, + 0.00823904387652874, + 0.006871545687317848, + 0.006611375603824854, + 0.006924974266439676, + 0.008037556894123554, + 0.004043753258883953, + 0.003997350111603737, + 0.010103948414325714, + 0.010454555042088032, + 0.02223956771194935, + 0.018229641020298004, + 0.01921476610004902, + 0.016861500218510628, + 0.009815583936870098, + 0.012877656146883965, + 0.006495090667158365, + 0.0071076201274991035, + 0.008301162160933018, + 0.009711074642837048, + 0.008075866848230362, + 0.008463011123239994, + 0.009330798871815205, + 0.009005939587950706, + 0.004893422592431307, + 0.005582727957516909, + 0.004872643388807774, + 0.004764562472701073, + 0.003323944052681327, + 0.005171376746147871, + 0.007192245684564114, + 0.009016928263008595, + 0.0035675964318215847, + 0.0053678108379244804, + 0.002845312235876918, + 0.00360894575715065, + 0.005845166742801666, + 0.004955744370818138, + 0.0028465697541832924, + 0.0031287127640098333, + 0.005828140303492546, + 0.005286231637001038, + 0.007306785322725773, + 0.013614175841212273, + 0.005500273313373327, + 0.0038849760312587023, + 0.009858626872301102, + 0.005043323617428541 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 0, + 5 + ], + "selected_reward": -0.7866209149360657 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned at an extremely low angle, just inches from the ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a Ford Mustang RTR SEN 5 painted glossy Electric Blue with a widebody kit visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a Honda Civic Type R painted vibrant Rally Red visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the vehicles shown in a collision where the Mustang's front fender and the Civic's driver's door are deformed and fused together?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are sharp, geometric fragments of silver and black metal flying from the collision point?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.00867254938930273 + ], + "opposite_logprob_tokens": [ + -4.758672714233398 + ], + "logprob_reward": -0.00867254938930273, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.00867254938930273, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1988748461008072, + 0.05960565805435181, + 0.07080519944429398, + 0.11507119238376617, + 0.06903916597366333, + 0.2112654149532318, + 0.16431444883346558, + 0.03167763724923134, + 0.022536808624863625, + 0.05680963024497032 + ], + "credit_map": [ + 0.014590533450245857, + 0.02555413916707039, + 0.02725239098072052, + 0.03127140551805496, + 0.01888388581573963, + 0.015175786800682545, + 0.03169656917452812, + 0.03445014730095863, + 0.005358896218240261, + 0.006941808387637138, + 0.013310856185853481, + 0.008519146591424942, + 0.007083884906023741, + 0.004472639877349138, + 0.006226172670722008, + 0.007692254614084959, + 0.012288376688957214, + 0.009654845111072063, + 0.011073251254856586, + 0.006109070498496294, + 0.008170496672391891, + 0.007131003774702549, + 0.008770850487053394, + 0.007607302162796259, + 0.02512012980878353, + 0.009705514647066593, + 0.007293578237295151, + 0.00523904524743557, + 0.01533374935388565, + 0.03886023163795471, + 0.004908792674541473, + 0.008610156364738941, + 0.010204395279288292, + 0.005790273193269968, + 0.009780047461390495, + 0.009692307561635971, + 0.008280088193714619, + 0.011411124840378761, + 0.007773364428430796, + 0.00610757153481245, + 0.018534742295742035, + 0.0180978886783123, + 0.0464150607585907, + 0.04506884142756462, + 0.024801896885037422, + 0.01952790655195713, + 0.018663734197616577, + 0.02015533670783043, + 0.0163812767714262, + 0.019419517368078232, + 0.025777040049433708, + 0.03209099546074867, + 0.01863749325275421, + 0.01317980233579874, + 0.02444436401128769, + 0.014383960515260696, + 0.004237601067870855, + 0.003614400513470173, + 0.0030466201715171337, + 0.004762344993650913, + 0.0024881891440600157, + 0.003533438313752413, + 0.004859031643718481, + 0.005136009305715561, + 0.002583369379863143, + 0.0028419026639312506, + 0.0025401997845619917, + 0.0027661481872200966, + 0.0032747627701610327, + 0.003381331916898489, + 0.0026761051267385483, + 0.002472989959642291, + 0.006704751867800951, + 0.005816667806357145, + 0.007166357710957527, + 0.011529633775353432, + 0.005934818182140589, + 0.004171844571828842, + 0.011102518066763878, + 0.00438303779810667 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 0, + 6 + ], + "selected_reward": -0.00867254938930273 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned at an extremely low angle, just inches from the ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a Ford Mustang RTR SEN 5 painted glossy Electric Blue with a widebody kit visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a Honda Civic Type R painted vibrant Rally Red visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the vehicles shown in a collision where the Mustang's front fender and the Civic's driver's door are deformed and fused together?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are sharp, geometric fragments of silver and black metal flying from the collision point?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004102386999875307 + ], + "opposite_logprob_tokens": [ + -5.5041022300720215 + ], + "logprob_reward": -0.004102386999875307, + "yes_no_margin_reward": 5.5, + "token_credit_reward": -0.004102386999875307, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3909063935279846, + 0.11042826622724533, + 0.1411442905664444, + 0.08421564102172852, + 0.03447960689663887, + 0.04783385246992111, + 0.06281527131795883, + 0.01973542757332325, + 0.012335393577814102, + 0.09610580652952194 + ], + "credit_map": [ + 0.027287742123007774, + 0.05638505518436432, + 0.055036116391420364, + 0.05114443227648735, + 0.0393705777823925, + 0.032841380685567856, + 0.06442905217409134, + 0.06441204249858856, + 0.01051580160856247, + 0.012835984118282795, + 0.017306961119174957, + 0.012001300230622292, + 0.011925660073757172, + 0.009961459785699844, + 0.01701986975967884, + 0.018861230462789536, + 0.023554768413305283, + 0.02155749686062336, + 0.02401644177734852, + 0.013690425083041191, + 0.012240688316524029, + 0.012000721879303455, + 0.01116972230374813, + 0.022914016619324684, + 0.017953014001250267, + 0.008616583421826363, + 0.006543823052197695, + 0.004627033602446318, + 0.0108495457097888, + 0.024833379313349724, + 0.0049725864082574844, + 0.005819675512611866, + 0.0028763317968696356, + 0.003901630872860551, + 0.005803799256682396, + 0.006043949630111456, + 0.003630031831562519, + 0.003495443146675825, + 0.0057108765468001366, + 0.0030175435822457075, + 0.00513595063239336, + 0.00437798909842968, + 0.007089703343808651, + 0.006925654597580433, + 0.006603843066841364, + 0.009075159206986427, + 0.004697303753346205, + 0.003928249701857567, + 0.0071191671304404736, + 0.007152764592319727, + 0.00682769576087594, + 0.007987389340996742, + 0.009492531418800354, + 0.005486540030688047, + 0.012757943943142891, + 0.005991240032017231, + 0.0023119342513382435, + 0.0021760687232017517, + 0.0014277884038165212, + 0.0026819987688213587, + 0.001621041214093566, + 0.0025962633080780506, + 0.003121193964034319, + 0.0037991406861692667, + 0.0011404297547414899, + 0.003174984361976385, + 0.0008392322342842817, + 0.0008362582302652299, + 0.003317149356007576, + 0.0013384446501731873, + 0.0009812414646148682, + 0.0007076532347127795, + 0.009820923209190369, + 0.006202176213264465, + 0.01285321544855833, + 0.029712969437241554, + 0.008701134473085403, + 0.004931310191750526, + 0.01545490138232708, + 0.008429176174104214 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 1 + ], + "selected_reward": -0.004102386999875307 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned at an extremely low angle, just inches from the ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a Ford Mustang RTR SEN 5 painted glossy Electric Blue with a widebody kit visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a Honda Civic Type R painted vibrant Rally Red visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the vehicles shown in a collision where the Mustang's front fender and the Civic's driver's door are deformed and fused together?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are sharp, geometric fragments of silver and black metal flying from the collision point?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0160688366740942 + ], + "opposite_logprob_tokens": [ + -4.141068935394287 + ], + "logprob_reward": -0.0160688366740942, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.0160688366740942, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.21663255989551544, + 0.08231324702501297, + 0.17452970147132874, + 0.06589188426733017, + 0.091868095099926, + 0.14512883126735687, + 0.06336371600627899, + 0.04471614956855774, + 0.03641913831233978, + 0.0791366845369339 + ], + "credit_map": [ + 0.023113612085580826, + 0.0217672698199749, + 0.024615027010440826, + 0.021980293095111847, + 0.02104712836444378, + 0.03574102371931076, + 0.03671255707740784, + 0.03165563941001892, + 0.006568473298102617, + 0.012332995422184467, + 0.011960990726947784, + 0.008751749992370605, + 0.01706455647945404, + 0.00906264316290617, + 0.00830474030226469, + 0.008267085067927837, + 0.026691239327192307, + 0.02210049144923687, + 0.026575379073619843, + 0.015565683133900166, + 0.018968982622027397, + 0.015914293006062508, + 0.023821663111448288, + 0.024891944602131844, + 0.00908124353736639, + 0.011998661793768406, + 0.008286607451736927, + 0.007197806145995855, + 0.010316706262528896, + 0.007214708253741264, + 0.005548506509512663, + 0.0062476396560668945, + 0.008854428306221962, + 0.01501497719436884, + 0.01583968661725521, + 0.01443974394351244, + 0.008129725232720375, + 0.01160036027431488, + 0.00812507327646017, + 0.009864084422588348, + 0.012418573722243309, + 0.013575887307524681, + 0.03002445586025715, + 0.031030185520648956, + 0.01747441478073597, + 0.015014761127531528, + 0.013212032616138458, + 0.012378516606986523, + 0.00632955739274621, + 0.00586067046970129, + 0.0064283860847353935, + 0.008396011777222157, + 0.00911409966647625, + 0.006726925261318684, + 0.010681778192520142, + 0.009826279245316982, + 0.005543429404497147, + 0.005316733382642269, + 0.0038610189221799374, + 0.006213294807821512, + 0.0031648436561226845, + 0.005320370197296143, + 0.0092289624735713, + 0.006067492067813873, + 0.0035158947575837374, + 0.0064725265838205814, + 0.00413112435489893, + 0.005903622601181269, + 0.006452109664678574, + 0.004089978989213705, + 0.003053287509828806, + 0.0028005889616906643, + 0.007566614542156458, + 0.006343289744108915, + 0.008175699040293694, + 0.02141675539314747, + 0.008588256314396858, + 0.007733540143817663, + 0.011911014094948769, + 0.007401511538773775 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 5 + ], + "selected_reward": -0.0160688366740942 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned at an extremely low angle, just inches from the ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a Ford Mustang RTR SEN 5 painted glossy Electric Blue with a widebody kit visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a Honda Civic Type R painted vibrant Rally Red visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the vehicles shown in a collision where the Mustang's front fender and the Civic's driver's door are deformed and fused together?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are sharp, geometric fragments of silver and black metal flying from the collision point?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.012536883354187012 + ], + "opposite_logprob_tokens": [ + -4.387537002563477 + ], + "logprob_reward": -0.012536883354187012, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.012536883354187012, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.24378323554992676, + 0.04583504796028137, + 0.16535168886184692, + 0.18245694041252136, + 0.13570250570774078, + 0.07929671555757523, + 0.03352535143494606, + 0.024211924523115158, + 0.050126757472753525, + 0.039709750562906265 + ], + "credit_map": [ + 0.02527453750371933, + 0.034163426607847214, + 0.020065614953637123, + 0.023773904889822006, + 0.04319457709789276, + 0.026378104463219643, + 0.03480677306652069, + 0.03612630441784859, + 0.004541494883596897, + 0.0069980318658053875, + 0.004152352921664715, + 0.00428023049607873, + 0.009914030320942402, + 0.005256414879113436, + 0.004764717072248459, + 0.005927775055170059, + 0.0212522242218256, + 0.02549879252910614, + 0.024830028414726257, + 0.01602579653263092, + 0.025623111054301262, + 0.024501722306013107, + 0.014836861751973629, + 0.012783165089786053, + 0.04228072240948677, + 0.014503398910164833, + 0.010114889591932297, + 0.0067983753979206085, + 0.023818012326955795, + 0.06677644699811935, + 0.006373881362378597, + 0.01179121807217598, + 0.008937045931816101, + 0.015513040125370026, + 0.024737006053328514, + 0.024801425635814667, + 0.010077661834657192, + 0.016364742070436478, + 0.01773413084447384, + 0.017537463456392288, + 0.006867766845971346, + 0.007870251312851906, + 0.015379311516880989, + 0.01652168110013008, + 0.010306173004209995, + 0.007215193472802639, + 0.007543082349002361, + 0.007593255024403334, + 0.0046845669858157635, + 0.004170672036707401, + 0.0035356655716896057, + 0.004211448132991791, + 0.003686885815113783, + 0.00462182704359293, + 0.005107644479721785, + 0.0035066416021436453, + 0.0033052582293748856, + 0.004163190256804228, + 0.002650178037583828, + 0.0034095244482159615, + 0.0018313229084014893, + 0.002307239919900894, + 0.003953986801207066, + 0.0025912229903042316, + 0.004872443620115519, + 0.006413821130990982, + 0.0035409601405262947, + 0.005094831343740225, + 0.008815804496407509, + 0.011242597363889217, + 0.005327384918928146, + 0.0048189181834459305, + 0.0047005945816636086, + 0.004032433498650789, + 0.0046860515139997005, + 0.006700514350086451, + 0.003922479227185249, + 0.004140118137001991, + 0.007476726546883583, + 0.0040508294478058815 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 2 + ], + "selected_reward": -0.012536883354187012 + } + ], + "vlm_decoded_outputs": [ + "No", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "Filmed from a static, extremely low angle, just inches from the ground, a Ford Mustang RTR SEN 5 and a Honda Civic Type R engage in a fierce, catastrophic clash. The camera's fixed position emphasizes the sheer scale and raw power of the two vehicles as their bodies collide in a brutal, inelastic collision. The Ford Mustang's chassis, painted glossy Electric Blue, showcasing its aggressive widebody kit and distinctive RTR grille, now presses fiercely against the side of the Honda Civic Type R, a machine rendered in vibrant Rally Red. The moment of impact is a frozen spectacle of destruction: the Mustang's front fender and the Civic's driver's door deform and fuse together, metal ripped and bent inward. The collision, rendered in an impactful graphic design style, unleashes an explosion of stylized effects; sharp, geometric fragments of silver and black metal fly from the collision point like graphic carriers, while oversized, starburst-shaped sparks burst forth in vibrant yellow and orange, boldly outlined. The scene is saturated with bold, high-contrast lighting that carves deep, dramatic shadows and casts sharp, shimmering highlights on the cars’ twisted surfaces, emphasizing the raw force of the collision and creating an atmosphere of tense, high-energy controlled chaos.", + "seed": 19489, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000006/group_12/index_7645_seed_19489_dp_25_rank_50_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 12, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.23876794520765543, + "vlm_logprob": -0.23876794520765543, + "vlm_yes_no_margin": 3.425, + "vlm_token_credit": -0.23876794520765543, + "vlm_yes_ratio": 0.8, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned at an extremely low angle, just inches from the ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a Ford Mustang RTR SEN 5 painted glossy Electric Blue with a widebody kit visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a Honda Civic Type R painted vibrant Rally Red visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the vehicles shown in a collision where the Mustang's front fender and the Civic's driver's door are deformed and fused together?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are sharp, geometric fragments of silver and black metal flying from the collision point?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -1.14803147315979 + ], + "opposite_logprob_tokens": [ + -0.39803147315979004 + ], + "logprob_reward": -1.14803147315979, + "yes_no_margin_reward": -0.75, + "token_credit_reward": -1.14803147315979, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.19644805788993835, + 0.17161916196346283, + 0.06520919501781464, + 0.029002713039517403, + 0.08079581707715988, + 0.10480315238237381, + 0.054215483367443085, + 0.06892556697130203, + 0.11704672873020172, + 0.11193414777517319 + ], + "credit_map": [ + 0.03164055198431015, + 0.026170818135142326, + 0.022472688928246498, + 0.028266513720154762, + 0.020518865436315536, + 0.015638723969459534, + 0.022749410942196846, + 0.028990482911467552, + 0.013406097888946533, + 0.01052217185497284, + 0.01353094819933176, + 0.02064676024019718, + 0.020301194861531258, + 0.013364258222281933, + 0.0404072180390358, + 0.03944051265716553, + 0.007926837541162968, + 0.007692074403166771, + 0.006705550942569971, + 0.01004469208419323, + 0.006580720655620098, + 0.007335794623941183, + 0.009221076034009457, + 0.009702450595796108, + 0.0021961815655231476, + 0.0017654832918196917, + 0.004328952636569738, + 0.005927229765802622, + 0.003111653495579958, + 0.002158748684450984, + 0.00571562210097909, + 0.0037988414987921715, + 0.004246183671057224, + 0.009045615792274475, + 0.004917334299534559, + 0.010054555721580982, + 0.0258199330419302, + 0.00590189965441823, + 0.00612984923645854, + 0.014680443331599236, + 0.006256245542317629, + 0.006594635080546141, + 0.010862632654607296, + 0.010262598283588886, + 0.011064916849136353, + 0.02938353642821312, + 0.007792201824486256, + 0.022586388513445854, + 0.006299491971731186, + 0.004127682186663151, + 0.005890992470085621, + 0.011413074098527431, + 0.004438247065991163, + 0.006398698315024376, + 0.009547892026603222, + 0.006099405698478222, + 0.009135251864790916, + 0.009152032434940338, + 0.006057387217879295, + 0.012294212356209755, + 0.005885955411940813, + 0.007731493562459946, + 0.010882890783250332, + 0.0077863470651209354, + 0.026801958680152893, + 0.009401633404195309, + 0.005977721884846687, + 0.011732405982911587, + 0.0062001608312129974, + 0.03827664256095886, + 0.009744998998939991, + 0.008911200799047947, + 0.00836990401148796, + 0.0066821337677538395, + 0.008015048690140247, + 0.02107943780720234, + 0.004670933820307255, + 0.008645680733025074, + 0.024296166375279427, + 0.03017483837902546 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 8 + ], + "selected_reward": -1.14803147315979 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned at an extremely low angle, just inches from the ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a Ford Mustang RTR SEN 5 painted glossy Electric Blue with a widebody kit visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a Honda Civic Type R painted vibrant Rally Red visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the vehicles shown in a collision where the Mustang's front fender and the Civic's driver's door are deformed and fused together?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are sharp, geometric fragments of silver and black metal flying from the collision point?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.011113056913018227 + ], + "opposite_logprob_tokens": [ + -4.511113166809082 + ], + "logprob_reward": -0.011113056913018227, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.011113056913018227, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0869513750076294, + 0.15098346769809723, + 0.07689337432384491, + 0.05347457528114319, + 0.028752095997333527, + 0.06401344388723373, + 0.07434778660535812, + 0.09345845133066177, + 0.2865234613418579, + 0.08460195362567902 + ], + "credit_map": [ + 0.01144385989755392, + 0.010053521953523159, + 0.007342091761529446, + 0.015318538062274456, + 0.010004149749875069, + 0.010047544725239277, + 0.010235343128442764, + 0.01250632107257843, + 0.012329915538430214, + 0.007128260564059019, + 0.010992813855409622, + 0.021313907578587532, + 0.02430410496890545, + 0.012166682630777359, + 0.03652939200401306, + 0.026218393817543983, + 0.007746698800474405, + 0.007838793098926544, + 0.0075829243287444115, + 0.015218287706375122, + 0.006073470693081617, + 0.008554046973586082, + 0.012361263856291771, + 0.011517887003719807, + 0.005826651584357023, + 0.003993766847997904, + 0.006268497556447983, + 0.009026072919368744, + 0.006597620435059071, + 0.006953938864171505, + 0.00877905823290348, + 0.0060289702378213406, + 0.002214454347267747, + 0.0022841738536953926, + 0.002466806909069419, + 0.0039029966574162245, + 0.004884484689682722, + 0.005692790262401104, + 0.004012243822216988, + 0.0032941463869065046, + 0.005343548953533173, + 0.003538585500791669, + 0.0050087920390069485, + 0.006576379761099815, + 0.006126895546913147, + 0.020356938242912292, + 0.00494688656181097, + 0.01211541797965765, + 0.008719434030354023, + 0.005764664616435766, + 0.0067686522379517555, + 0.010779042728245258, + 0.008208674378693104, + 0.012189120054244995, + 0.009249741211533546, + 0.012668458744883537, + 0.011937997303903103, + 0.01519996952265501, + 0.00641587795689702, + 0.018303288146853447, + 0.008086534217000008, + 0.011845055036246777, + 0.012202056124806404, + 0.00946767721325159, + 0.0768948346376419, + 0.02102070301771164, + 0.011801432818174362, + 0.019171584397554398, + 0.01292831264436245, + 0.10848834365606308, + 0.01875278539955616, + 0.017465444281697273, + 0.009833800606429577, + 0.0058896210975945, + 0.0059306989423930645, + 0.016811225563287735, + 0.00463953847065568, + 0.006353371310979128, + 0.010997401550412178, + 0.024146290495991707 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 1, + 7 + ], + "selected_reward": -0.011113056913018227 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned at an extremely low angle, just inches from the ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a Ford Mustang RTR SEN 5 painted glossy Electric Blue with a widebody kit visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a Honda Civic Type R painted vibrant Rally Red visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the vehicles shown in a collision where the Mustang's front fender and the Civic's driver's door are deformed and fused together?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are sharp, geometric fragments of silver and black metal flying from the collision point?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.012545948848128319 + ], + "opposite_logprob_tokens": [ + -4.387546062469482 + ], + "logprob_reward": -0.012545948848128319, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.012545948848128319, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.27288031578063965, + 0.050533246248960495, + 0.07596858590841293, + 0.05481783300638199, + 0.03667222708463669, + 0.08924689143896103, + 0.071283258497715, + 0.1909973919391632, + 0.1041719913482666, + 0.053428202867507935 + ], + "credit_map": [ + 0.036530252546072006, + 0.03449159488081932, + 0.025506140664219856, + 0.06865298002958298, + 0.024543631821870804, + 0.02403082884848118, + 0.022484179586172104, + 0.03664072975516319, + 0.0035894589964300394, + 0.0035448074340820312, + 0.005449451040476561, + 0.007578062359243631, + 0.008618236519396305, + 0.004024939611554146, + 0.01161876693367958, + 0.006109526846557856, + 0.004669626709073782, + 0.007909886538982391, + 0.00802648440003395, + 0.011054839938879013, + 0.008990160189568996, + 0.008424097672104836, + 0.012805087491869926, + 0.014088408090174198, + 0.003978257533162832, + 0.004928337875753641, + 0.005042240954935551, + 0.007676227018237114, + 0.006614408455789089, + 0.008172405883669853, + 0.006868210155516863, + 0.011537749320268631, + 0.003979663830250502, + 0.003475629724562168, + 0.004669196903705597, + 0.004417101852595806, + 0.006299782544374466, + 0.005464816465973854, + 0.0039802598766982555, + 0.004385775420814753, + 0.007995521649718285, + 0.004988049156963825, + 0.006400442216545343, + 0.007465795148164034, + 0.008957109414041042, + 0.0275920070707798, + 0.006597586907446384, + 0.019250383600592613, + 0.010489270091056824, + 0.007808273658156395, + 0.008509276434779167, + 0.017170429229736328, + 0.006775032263249159, + 0.005016343668103218, + 0.010130500420928001, + 0.005384140647947788, + 0.024523256346583366, + 0.01581091806292534, + 0.013091688975691795, + 0.050977982580661774, + 0.010503159835934639, + 0.019183825701475143, + 0.03763729706406593, + 0.01926928386092186, + 0.018152711912989616, + 0.007191784214228392, + 0.006211460568010807, + 0.012462655082345009, + 0.006627487484365702, + 0.029252052307128906, + 0.0116879902780056, + 0.012585862539708614, + 0.004466656595468521, + 0.004167182836681604, + 0.0035666306503117085, + 0.010180535726249218, + 0.002989682601764798, + 0.005682151298969984, + 0.011251437477767467, + 0.011123931035399437 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 7, + 8 + ], + "selected_reward": -0.012545948848128319 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned at an extremely low angle, just inches from the ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a Ford Mustang RTR SEN 5 painted glossy Electric Blue with a widebody kit visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a Honda Civic Type R painted vibrant Rally Red visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the vehicles shown in a collision where the Mustang's front fender and the Civic's driver's door are deformed and fused together?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are sharp, geometric fragments of silver and black metal flying from the collision point?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.011075330898165703 + ], + "opposite_logprob_tokens": [ + -4.511075496673584 + ], + "logprob_reward": -0.011075330898165703, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.011075330898165703, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11722533404827118, + 0.1685149073600769, + 0.07307185977697372, + 0.028404967859387398, + 0.05421290546655655, + 0.10383016616106033, + 0.0460146963596344, + 0.21149292588233948, + 0.09444234520196915, + 0.10278987884521484 + ], + "credit_map": [ + 0.019021958112716675, + 0.017191817983984947, + 0.013877767138183117, + 0.02280336432158947, + 0.010815661400556564, + 0.008624725043773651, + 0.011366666294634342, + 0.013523373752832413, + 0.012753690592944622, + 0.005860824137926102, + 0.009277703240513802, + 0.02371942438185215, + 0.03380058333277702, + 0.013997231610119343, + 0.042342063039541245, + 0.026763373985886574, + 0.005365500692278147, + 0.009083245880901814, + 0.007697605062276125, + 0.008931760676205158, + 0.009514476172626019, + 0.011740381829440594, + 0.01081107184290886, + 0.00992781575769186, + 0.0025179656222462654, + 0.002521196845918894, + 0.002709610154852271, + 0.003907271195203066, + 0.0031479999888688326, + 0.0047754948027431965, + 0.0030980573501437902, + 0.005727371200919151, + 0.004940142389386892, + 0.006240228656679392, + 0.003320976858958602, + 0.0029119551181793213, + 0.01772993989288807, + 0.00447978638112545, + 0.005242267623543739, + 0.009347607381641865, + 0.008137237280607224, + 0.003197398269549012, + 0.008365737274289131, + 0.009529722854495049, + 0.007342051714658737, + 0.03129204362630844, + 0.012126773595809937, + 0.023839205503463745, + 0.0058970325626432896, + 0.003971444442868233, + 0.006452109199017286, + 0.008461984805762768, + 0.004029333591461182, + 0.004751151893287897, + 0.00672569964081049, + 0.005725940223783255, + 0.027577558532357216, + 0.019187042489647865, + 0.014928758144378662, + 0.05674545466899872, + 0.011758229695260525, + 0.018917355686426163, + 0.04137127846479416, + 0.02100723795592785, + 0.019123172387480736, + 0.006585588213056326, + 0.005046547390520573, + 0.008197512477636337, + 0.006180380005389452, + 0.031606920063495636, + 0.007720896042883396, + 0.00998133048415184, + 0.012252625077962875, + 0.005828110501170158, + 0.007461019791662693, + 0.015332477167248726, + 0.0049027022905647755, + 0.006179042160511017, + 0.020834660157561302, + 0.02999924309551716 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 1, + 0 + ], + "selected_reward": -0.011075330898165703 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned at an extremely low angle, just inches from the ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a Ford Mustang RTR SEN 5 painted glossy Electric Blue with a widebody kit visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a Honda Civic Type R painted vibrant Rally Red visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the vehicles shown in a collision where the Mustang's front fender and the Civic's driver's door are deformed and fused together?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are sharp, geometric fragments of silver and black metal flying from the collision point?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.011073916219174862 + ], + "opposite_logprob_tokens": [ + -4.511074066162109 + ], + "logprob_reward": -0.011073916219174862, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.011073916219174862, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07864779233932495, + 0.21864448487758636, + 0.04966197907924652, + 0.040002498775720596, + 0.08173997700214386, + 0.08775536715984344, + 0.0688946470618248, + 0.15892831981182098, + 0.1456729769706726, + 0.07005204260349274 + ], + "credit_map": [ + 0.009113268926739693, + 0.00744630116969347, + 0.0059693679213523865, + 0.009731956757605076, + 0.012093846686184406, + 0.009592019952833652, + 0.01286668423563242, + 0.011834343895316124, + 0.020305512472987175, + 0.021329360082745552, + 0.025575464591383934, + 0.029890887439250946, + 0.028423964977264404, + 0.015675168484449387, + 0.04511674866080284, + 0.03232734650373459, + 0.005273651797324419, + 0.004771860782057047, + 0.0058970157988369465, + 0.00857976172119379, + 0.0057575455866754055, + 0.005600320640951395, + 0.00710664689540863, + 0.006675174925476313, + 0.004266503266990185, + 0.0030520681757479906, + 0.003327696118503809, + 0.005058235488831997, + 0.00337284617125988, + 0.006137684918940067, + 0.00486160209402442, + 0.009925863705575466, + 0.0066536772064864635, + 0.009782147593796253, + 0.005315123125910759, + 0.007590002380311489, + 0.026763668283820152, + 0.00588240334764123, + 0.005612009204924107, + 0.014140942133963108, + 0.008876231499016285, + 0.005787550006061792, + 0.010108012706041336, + 0.00898031983524561, + 0.009034433402121067, + 0.020456228405237198, + 0.009366984479129314, + 0.015145597979426384, + 0.009763035923242569, + 0.005895761772990227, + 0.008915932849049568, + 0.014561288058757782, + 0.006118388380855322, + 0.005451663862913847, + 0.011942348442971706, + 0.006246221251785755, + 0.0222916342318058, + 0.01374041847884655, + 0.011189425364136696, + 0.04217782989144325, + 0.008630356751382351, + 0.01402642298489809, + 0.031158950179815292, + 0.01571325957775116, + 0.03226391226053238, + 0.010173900984227657, + 0.006562987808138132, + 0.012769684195518494, + 0.007551670074462891, + 0.0507911816239357, + 0.011725529097020626, + 0.013834082521498203, + 0.005389896687120199, + 0.00472194654867053, + 0.005239391233772039, + 0.013598560355603695, + 0.0038761133328080177, + 0.005993693135678768, + 0.01777207851409912, + 0.013460357673466206 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 7, + 8 + ], + "selected_reward": -0.011073916219174862 + } + ], + "vlm_decoded_outputs": [ + "No", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000006/group_13/index_tensor([2074, 2074])_group_13_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000006/group_13/index_tensor([2074, 2074])_group_13_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..808a8e574d272c6044ed2a5fbd7fcfab0c08c0a0 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000006/group_13/index_tensor([2074, 2074])_group_13_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "From a drone's bird's-eye perspective, the camera slowly and steadily drifts forward, offering a comprehensive and immersive view of a vast, vibrant natural scene. The landscape is dominated by a field of blooming red poppies. Their delicate, paper-thin petals create a brilliant sea of crimson, contrasting beautifully with the lush, dark green foliage of their stems and leaves. Organically interspersed among the poppies are clusters of small white and sunny yellow wildflowers, whose bright, cheerful colors enhance the scene's rich diversity. The entire field is bathed in the warm, soft light of the afternoon sun, deeply saturating colors and casting soft, elongated shadows that add depth and texture to the landscape. A gentle breeze sways the countless flower heads in a gentle, hypnotic rhythm, creating a tranquil and picturesque atmosphere that celebrates the botanical richness and wild beauty.", + "seed": 2221, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000006/group_13/index_2074_seed_2221_dp_26_rank_52_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 13, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.23877334961434826, + "vlm_logprob": -0.23877334961434826, + "vlm_yes_no_margin": 3.3, + "vlm_token_credit": -0.23877334961434826, + "vlm_yes_ratio": 0.8, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene viewed from a drone's bird's-eye perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera slowly and steadily drift forward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the landscape dominated by a field of blooming red poppies?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are clusters of small white and sunny yellow wildflowers interspersed among the poppies?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are the flowers swaying in a gentle breeze creating a rhythmic motion?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -0.9929942488670349 + ], + "opposite_logprob_tokens": [ + -0.4929942488670349 + ], + "logprob_reward": -0.9929942488670349, + "yes_no_margin_reward": -0.5, + "token_credit_reward": -0.9929942488670349, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.11305780708789825, + 0.13197408616542816, + 0.0670461505651474, + 0.0604604072868824, + 0.29583728313446045, + 0.046449292451143265, + 0.05878041684627533, + 0.04264072701334953, + 0.08368176221847534, + 0.10007207095623016 + ], + "credit_map": [ + 0.012119347229599953, + 0.0075098732486367226, + 0.023938680067658424, + 0.018387446179986, + 0.011648700572550297, + 0.011050792410969734, + 0.016338733956217766, + 0.012064230628311634, + 0.012038188986480236, + 0.010637528263032436, + 0.018599722534418106, + 0.011581812053918839, + 0.01445736549794674, + 0.027952872216701508, + 0.02430909126996994, + 0.012397519312798977, + 0.008575092069804668, + 0.00874777790158987, + 0.008821006864309311, + 0.008521241135895252, + 0.006536456290632486, + 0.010428105480968952, + 0.007820582017302513, + 0.007595886010676622, + 0.011772924102842808, + 0.007238131947815418, + 0.005700410809367895, + 0.005586069542914629, + 0.0054427641443908215, + 0.00664513697847724, + 0.01109025813639164, + 0.0069847097620368, + 0.036556702107191086, + 0.1103147491812706, + 0.027330683544278145, + 0.01952613890171051, + 0.021764185279607773, + 0.016646187752485275, + 0.014126580208539963, + 0.04957204312086105, + 0.004962757229804993, + 0.0061546675860881805, + 0.005607145838439465, + 0.005998522974550724, + 0.008964983746409416, + 0.004787368234246969, + 0.007035849150270224, + 0.002937995595857501, + 0.00789925828576088, + 0.004726618994027376, + 0.013207010924816132, + 0.00499067222699523, + 0.009494144469499588, + 0.006198241841048002, + 0.006289029493927956, + 0.005975436419248581, + 0.008593126200139523, + 0.005509324371814728, + 0.00339803216047585, + 0.005361628718674183, + 0.006296501960605383, + 0.005470248404890299, + 0.004110174253582954, + 0.0039016904775053263, + 0.010636056773364544, + 0.011487874202430248, + 0.01723744347691536, + 0.008984249085187912, + 0.009346779435873032, + 0.010732423514127731, + 0.008209412917494774, + 0.007047522813081741, + 0.016538746654987335, + 0.011575807817280293, + 0.00900902971625328, + 0.00801591295748949, + 0.022780602797865868, + 0.016403930261731148, + 0.008460248820483685, + 0.007287801709026098 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 1, + 0 + ], + "selected_reward": -0.9929942488670349 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene viewed from a drone's bird's-eye perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera slowly and steadily drift forward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the landscape dominated by a field of blooming red poppies?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are clusters of small white and sunny yellow wildflowers interspersed among the poppies?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are the flowers swaying in a gentle breeze creating a rhythmic motion?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.10035981237888336 + ], + "opposite_logprob_tokens": [ + -2.3503599166870117 + ], + "logprob_reward": -0.10035981237888336, + "yes_no_margin_reward": 2.25, + "token_credit_reward": -0.10035981237888336, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10820525139570236, + 0.055496782064437866, + 0.10204283893108368, + 0.04251750186085701, + 0.27340954542160034, + 0.047458209097385406, + 0.17247283458709717, + 0.05188014730811119, + 0.0799843892455101, + 0.06653252989053726 + ], + "credit_map": [ + 0.012830598279833794, + 0.010739214718341827, + 0.016075899824500084, + 0.01653405651450157, + 0.012704136781394482, + 0.01075020618736744, + 0.01896802708506584, + 0.009603110142052174, + 0.0070017059333622456, + 0.004725403618067503, + 0.007273922674357891, + 0.0056444876827299595, + 0.0076849632896482944, + 0.01031858567148447, + 0.007272541522979736, + 0.005575170740485191, + 0.015620180405676365, + 0.01409551315009594, + 0.01136816293001175, + 0.012957037426531315, + 0.010619044303894043, + 0.013979275710880756, + 0.010805459693074226, + 0.012598162516951561, + 0.007819375023245811, + 0.005144019611179829, + 0.004246369004249573, + 0.004987235181033611, + 0.004318251274526119, + 0.004744396544992924, + 0.006429064553231001, + 0.0048287915997207165, + 0.02800937183201313, + 0.06632963567972183, + 0.024254268035292625, + 0.03151245787739754, + 0.020406007766723633, + 0.021395348012447357, + 0.01473022997379303, + 0.0667722299695015, + 0.005245330277830362, + 0.006581936031579971, + 0.00425830390304327, + 0.005417914129793644, + 0.006924557499587536, + 0.005042737349867821, + 0.010288872756063938, + 0.003698561806231737, + 0.020864669233560562, + 0.02248903550207615, + 0.025447344407439232, + 0.01824142411351204, + 0.020199676975607872, + 0.029132038354873657, + 0.017825359478592873, + 0.018273280933499336, + 0.012852060608565807, + 0.007364721968770027, + 0.003944862633943558, + 0.004612396936863661, + 0.00550488056614995, + 0.00674901157617569, + 0.004812078550457954, + 0.006040133070200682, + 0.01242911722511053, + 0.009133095853030682, + 0.014887324534356594, + 0.009363195858895779, + 0.012637102045118809, + 0.008621112443506718, + 0.008919299580156803, + 0.003994139842689037, + 0.009248612448573112, + 0.008405238389968872, + 0.006337785627692938, + 0.005907702725380659, + 0.008169041015207767, + 0.010021618567407131, + 0.010592814534902573, + 0.00784971471875906 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 6, + 0 + ], + "selected_reward": -0.10035981237888336 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene viewed from a drone's bird's-eye perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera slowly and steadily drift forward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the landscape dominated by a field of blooming red poppies?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are clusters of small white and sunny yellow wildflowers interspersed among the poppies?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are the flowers swaying in a gentle breeze creating a rhythmic motion?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0003962923656217754 + ], + "opposite_logprob_tokens": [ + -7.875396251678467 + ], + "logprob_reward": -0.0003962923656217754, + "yes_no_margin_reward": 7.875, + "token_credit_reward": -0.0003962923656217754, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10254137217998505, + 0.09268385171890259, + 0.07968592643737793, + 0.07744744420051575, + 0.22872653603553772, + 0.06080125644803047, + 0.07097752392292023, + 0.05313323438167572, + 0.1551840901374817, + 0.07881879061460495 + ], + "credit_map": [ + 0.012671878561377525, + 0.012219991534948349, + 0.014966737478971481, + 0.013652032241225243, + 0.011403581127524376, + 0.012101789005100727, + 0.011913247406482697, + 0.013612110167741776, + 0.00987096969038248, + 0.008311767131090164, + 0.009164185263216496, + 0.010621445253491402, + 0.01616067811846733, + 0.013502980582416058, + 0.01597524620592594, + 0.009076585993170738, + 0.015056561678647995, + 0.012099357321858406, + 0.009420068934559822, + 0.010467607527971268, + 0.006751575507223606, + 0.009045758284628391, + 0.007848048582673073, + 0.008996953256428242, + 0.012166373431682587, + 0.009749754332005978, + 0.008018817752599716, + 0.008139563724398613, + 0.008073386736214161, + 0.00931604951620102, + 0.013130269013345242, + 0.008853236213326454, + 0.03158216550946236, + 0.08960900455713272, + 0.02191377989947796, + 0.01246429793536663, + 0.017183206975460052, + 0.017219917848706245, + 0.01011580415070057, + 0.02863834612071514, + 0.009704685769975185, + 0.006166187580674887, + 0.0072341011837124825, + 0.0052675227634608746, + 0.007567774038761854, + 0.012801007367670536, + 0.006759986747056246, + 0.005299990531057119, + 0.007377666886895895, + 0.007637044880539179, + 0.01734769716858864, + 0.005584269296377897, + 0.010790331289172173, + 0.008042152039706707, + 0.007453623227775097, + 0.006744737736880779, + 0.008553717285394669, + 0.005608622916042805, + 0.004944973159581423, + 0.0077545782551169395, + 0.005736229941248894, + 0.008643568493425846, + 0.005895097739994526, + 0.005996448919177055, + 0.02683202363550663, + 0.016170743852853775, + 0.03597139194607735, + 0.015886325389146805, + 0.016396697610616684, + 0.01713612861931324, + 0.01640540547668934, + 0.010385388508439064, + 0.014081878587603569, + 0.011223403736948967, + 0.008281846530735493, + 0.006335048470646143, + 0.008854416199028492, + 0.011534532532095909, + 0.010166644118726254, + 0.008341021835803986 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 8, + 0 + ], + "selected_reward": -0.0003962923656217754 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene viewed from a drone's bird's-eye perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera slowly and steadily drift forward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the landscape dominated by a field of blooming red poppies?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are clusters of small white and sunny yellow wildflowers interspersed among the poppies?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are the flowers swaying in a gentle breeze creating a rhythmic motion?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01107780635356903 + ], + "opposite_logprob_tokens": [ + -4.511077880859375 + ], + "logprob_reward": -0.01107780635356903, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.01107780635356903, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10052508860826492, + 0.13652001321315765, + 0.07513722777366638, + 0.11773470044136047, + 0.18932276964187622, + 0.06735136359930038, + 0.11426485329866409, + 0.06460733711719513, + 0.04332467168569565, + 0.09121201187372208 + ], + "credit_map": [ + 0.009770535863935947, + 0.01130816712975502, + 0.01289715338498354, + 0.013683429919183254, + 0.009995112195611, + 0.011266917921602726, + 0.014047624543309212, + 0.017556145787239075, + 0.014963432215154171, + 0.01012377254664898, + 0.01908816024661064, + 0.011205454356968403, + 0.027158984914422035, + 0.025975940749049187, + 0.018629658967256546, + 0.009374613873660564, + 0.013572396710515022, + 0.012701918371021748, + 0.0071614705957472324, + 0.009618920274078846, + 0.005748506635427475, + 0.009466377086937428, + 0.007521801628172398, + 0.009345839731395245, + 0.019383959472179413, + 0.013976826332509518, + 0.013236087746918201, + 0.011350523680448532, + 0.011540557257831097, + 0.013402087613940239, + 0.020940883085131645, + 0.013903763145208359, + 0.026181897148489952, + 0.06821805238723755, + 0.012983553111553192, + 0.015214777551591396, + 0.013126605190336704, + 0.015663282945752144, + 0.011180281639099121, + 0.026754314079880714, + 0.007007896434515715, + 0.006399958394467831, + 0.006132214330136776, + 0.006763783283531666, + 0.010466133244335651, + 0.009825672954320908, + 0.014012613333761692, + 0.006743088364601135, + 0.016099773347377777, + 0.015899797901511192, + 0.026347197592258453, + 0.009072704240679741, + 0.016859520226716995, + 0.012434993870556355, + 0.009968595579266548, + 0.007582271471619606, + 0.0074071078561246395, + 0.006046276539564133, + 0.006127180065959692, + 0.00957757979631424, + 0.006729409098625183, + 0.008267363533377647, + 0.009585759602487087, + 0.01086666714400053, + 0.009118679910898209, + 0.003400332760065794, + 0.009837545454502106, + 0.005100787151604891, + 0.004002420697361231, + 0.0047133504413068295, + 0.00463415402919054, + 0.0025174019392579794, + 0.013941116631031036, + 0.011836671270430088, + 0.00988409761339426, + 0.009717319160699844, + 0.010803823359310627, + 0.010052673518657684, + 0.01682868041098118, + 0.008147628046572208 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 1, + 3 + ], + "selected_reward": -0.01107780635356903 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene viewed from a drone's bird's-eye perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera slowly and steadily drift forward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the landscape dominated by a field of blooming red poppies?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are clusters of small white and sunny yellow wildflowers interspersed among the poppies?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are the flowers swaying in a gentle breeze creating a rhythmic motion?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.08903858810663223 + ], + "opposite_logprob_tokens": [ + -2.464038610458374 + ], + "logprob_reward": -0.08903858810663223, + "yes_no_margin_reward": 2.375, + "token_credit_reward": -0.08903858810663223, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1586071401834488, + 0.04634322598576546, + 0.07960066199302673, + 0.18544816970825195, + 0.18698228895664215, + 0.07637467980384827, + 0.14150866866111755, + 0.037254322320222855, + 0.06131238862872124, + 0.02656850405037403 + ], + "credit_map": [ + 0.022382352501153946, + 0.012470298446714878, + 0.021392395719885826, + 0.01400008425116539, + 0.009890345856547356, + 0.02153037115931511, + 0.018166694790124893, + 0.03877459093928337, + 0.006709865760058165, + 0.004658338148146868, + 0.004353972617536783, + 0.005645496305078268, + 0.006270064506679773, + 0.006346662063151598, + 0.00827475730329752, + 0.0040840706788003445, + 0.010140317492187023, + 0.010172349400818348, + 0.009503843262791634, + 0.010657192207872868, + 0.008239726535975933, + 0.010460400953888893, + 0.011585965752601624, + 0.008840871974825859, + 0.025572512298822403, + 0.02042182721197605, + 0.022899113595485687, + 0.022151412442326546, + 0.018661847338080406, + 0.016806025058031082, + 0.03592429682612419, + 0.023011133074760437, + 0.0279010571539402, + 0.0691659078001976, + 0.013906963169574738, + 0.018016573041677475, + 0.012204760685563087, + 0.015845878049731255, + 0.011163131333887577, + 0.018778013065457344, + 0.006743522826582193, + 0.009174900129437447, + 0.006903841160237789, + 0.009968994185328484, + 0.011571588926017284, + 0.007481249049305916, + 0.01836373470723629, + 0.006166853476315737, + 0.022166021168231964, + 0.017724603414535522, + 0.031272437423467636, + 0.012884252704679966, + 0.014598294161260128, + 0.020362628623843193, + 0.011569502763450146, + 0.010930933058261871, + 0.006018943153321743, + 0.004639071878045797, + 0.003943696152418852, + 0.004037199541926384, + 0.004480412229895592, + 0.004896036349236965, + 0.004619027022272348, + 0.004619935993105173, + 0.008513309061527252, + 0.007684425450861454, + 0.00921707320958376, + 0.009891561232507229, + 0.007678160909563303, + 0.006335803773254156, + 0.0062704323790967464, + 0.0057216244749724865, + 0.004954415373504162, + 0.0034856796264648438, + 0.0036093092057853937, + 0.0023446152918040752, + 0.0036861866246908903, + 0.0033825167920440435, + 0.003065422410145402, + 0.002040357096120715 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 3, + 0 + ], + "selected_reward": -0.08903858810663223 + } + ], + "vlm_decoded_outputs": [ + "No", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "From a drone's bird's-eye perspective, the camera slowly and steadily drifts forward, offering a comprehensive and immersive view of a vast, vibrant natural scene. The landscape is dominated by a field of blooming red poppies. Their delicate, paper-thin petals create a brilliant sea of crimson, contrasting beautifully with the lush, dark green foliage of their stems and leaves. Organically interspersed among the poppies are clusters of small white and sunny yellow wildflowers, whose bright, cheerful colors enhance the scene's rich diversity. The entire field is bathed in the warm, soft light of the afternoon sun, deeply saturating colors and casting soft, elongated shadows that add depth and texture to the landscape. A gentle breeze sways the countless flower heads in a gentle, hypnotic rhythm, creating a tranquil and picturesque atmosphere that celebrates the botanical richness and wild beauty.", + "seed": 2222, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000006/group_13/index_2074_seed_2222_dp_26_rank_52_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 13, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.12829886720282957, + "vlm_logprob": -0.12829886720282957, + "vlm_yes_no_margin": 3.4, + "vlm_token_credit": -0.12829886720282957, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene viewed from a drone's bird's-eye perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera slowly and steadily drift forward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the landscape dominated by a field of blooming red poppies?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are clusters of small white and sunny yellow wildflowers interspersed among the poppies?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are the flowers swaying in a gentle breeze creating a rhythmic motion?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.4025086462497711 + ], + "opposite_logprob_tokens": [ + -1.1525086164474487 + ], + "logprob_reward": -0.4025086462497711, + "yes_no_margin_reward": 0.75, + "token_credit_reward": -0.4025086462497711, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.4180849492549896, + 0.06924163550138474, + 0.056668832898139954, + 0.030513012781739235, + 0.03354623168706894, + 0.021072156727313995, + 0.2102300077676773, + 0.07417190819978714, + 0.058591388165950775, + 0.027879854664206505 + ], + "credit_map": [ + 0.025546720251441002, + 0.026551054790616035, + 0.03410927206277847, + 0.02998385950922966, + 0.027814224362373352, + 0.12280084192752838, + 0.122934989631176, + 0.028343994170427322, + 0.005211714189499617, + 0.006516971625387669, + 0.012146292254328728, + 0.012272465974092484, + 0.008303060196340084, + 0.014036216773092747, + 0.004970991984009743, + 0.005783919245004654, + 0.0074957432225346565, + 0.006287458352744579, + 0.0067472755908966064, + 0.007957550697028637, + 0.0060888053849339485, + 0.008158617652952671, + 0.00544483819976449, + 0.008488541468977928, + 0.003587099490687251, + 0.0035985761787742376, + 0.0038886675611138344, + 0.0032205739989876747, + 0.003964575938880444, + 0.004376317840069532, + 0.0029777497984468937, + 0.00489945150911808, + 0.004379583988338709, + 0.004294442478567362, + 0.0034270279575139284, + 0.0042235213331878185, + 0.003713376820087433, + 0.004257232882082462, + 0.00464401813223958, + 0.00460702832788229, + 0.00431540934368968, + 0.002562832087278366, + 0.002399498363956809, + 0.0031651232857257128, + 0.0020860438235104084, + 0.0026735260616987944, + 0.001909020938910544, + 0.0019607029389590025, + 0.04887622594833374, + 0.07905278354883194, + 0.013727054931223392, + 0.007486091926693916, + 0.01902863383293152, + 0.016053805127739906, + 0.008923700079321861, + 0.017081717029213905, + 0.007257574703544378, + 0.02525150403380394, + 0.0036886469461023808, + 0.0043831742368638515, + 0.01670021563768387, + 0.006290549412369728, + 0.003257709089666605, + 0.007342540193349123, + 0.006600110791623592, + 0.009505930356681347, + 0.006018275860697031, + 0.006161217112094164, + 0.006497181486338377, + 0.012870865873992443, + 0.006757883820682764, + 0.004179919604212046, + 0.0037162797525525093, + 0.004496746230870485, + 0.0028559486381709576, + 0.0035114451311528683, + 0.0036701555363833904, + 0.005263324826955795, + 0.0020111289341002703, + 0.0023548246826976538 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 6, + 7 + ], + "selected_reward": -0.4025086462497711 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene viewed from a drone's bird's-eye perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera slowly and steadily drift forward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the landscape dominated by a field of blooming red poppies?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are clusters of small white and sunny yellow wildflowers interspersed among the poppies?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are the flowers swaying in a gentle breeze creating a rhythmic motion?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07005427032709122 + ], + "opposite_logprob_tokens": [ + -2.695054292678833 + ], + "logprob_reward": -0.07005427032709122, + "yes_no_margin_reward": 2.625, + "token_credit_reward": -0.07005427032709122, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1761532574892044, + 0.0479242242872715, + 0.12425891309976578, + 0.03246795013546944, + 0.03290845453739166, + 0.14710862934589386, + 0.18050719797611237, + 0.12318006157875061, + 0.09568605571985245, + 0.03980525955557823 + ], + "credit_map": [ + 0.011189944110810757, + 0.016926394775509834, + 0.016866590827703476, + 0.018331848084926605, + 0.017875608056783676, + 0.03764329105615616, + 0.039624929428100586, + 0.017694665119051933, + 0.00432602921500802, + 0.00424712710082531, + 0.009607121348381042, + 0.0050284299068152905, + 0.005438901949673891, + 0.010935576632618904, + 0.0035043659154325724, + 0.0048366729170084, + 0.009543483145534992, + 0.011813868768513203, + 0.014302494004368782, + 0.02033948339521885, + 0.012368876487016678, + 0.023939084261655807, + 0.012503542006015778, + 0.019448071718215942, + 0.003665977157652378, + 0.003956208936870098, + 0.004144805949181318, + 0.004149878863245249, + 0.004920159466564655, + 0.004319001920521259, + 0.003698427928611636, + 0.003613488283008337, + 0.0031158921774476767, + 0.004392098169773817, + 0.002715091686695814, + 0.00555928610265255, + 0.0036739909555763006, + 0.004782210104167461, + 0.005138821434229612, + 0.0035310646053403616, + 0.02215847186744213, + 0.016878178343176842, + 0.013669711537659168, + 0.0195966437458992, + 0.01481117494404316, + 0.03309198096394539, + 0.012819153256714344, + 0.014083321206271648, + 0.03367043286561966, + 0.07321462035179138, + 0.012446519918739796, + 0.008035711012780666, + 0.01694406196475029, + 0.012922477908432484, + 0.009368408471345901, + 0.013904971070587635, + 0.012757115066051483, + 0.0327347032725811, + 0.009268277324736118, + 0.01247320231050253, + 0.031544338911771774, + 0.009110363200306892, + 0.00755144702270627, + 0.007740618661046028, + 0.008634400554001331, + 0.022727884352207184, + 0.008784339763224125, + 0.01512926910072565, + 0.008158938959240913, + 0.019984809681773186, + 0.008112072013318539, + 0.004154338501393795, + 0.004051838535815477, + 0.005191202275454998, + 0.004282968584448099, + 0.003964634612202644, + 0.0068499925546348095, + 0.00927110854536295, + 0.003327490296214819, + 0.002866026246920228 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 0, + 5 + ], + "selected_reward": -0.07005427032709122 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene viewed from a drone's bird's-eye perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera slowly and steadily drift forward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the landscape dominated by a field of blooming red poppies?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are clusters of small white and sunny yellow wildflowers interspersed among the poppies?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are the flowers swaying in a gentle breeze creating a rhythmic motion?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0009310677996836603 + ], + "opposite_logprob_tokens": [ + -7.000931262969971 + ], + "logprob_reward": -0.0009310677996836603, + "yes_no_margin_reward": 7.0, + "token_credit_reward": -0.0009310677996836603, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1837436854839325, + 0.12952715158462524, + 0.1577754020690918, + 0.09039655327796936, + 0.03703326731920242, + 0.11489266902208328, + 0.13617652654647827, + 0.06829545646905899, + 0.042971186339855194, + 0.039188094437122345 + ], + "credit_map": [ + 0.0128812026232481, + 0.012727062217891216, + 0.016526954248547554, + 0.01671520434319973, + 0.011854094453155994, + 0.047599922865629196, + 0.050021011382341385, + 0.015418220311403275, + 0.009096140041947365, + 0.008697154931724072, + 0.03362216427922249, + 0.009379372000694275, + 0.01159973256289959, + 0.033964551985263824, + 0.010435247793793678, + 0.01273277960717678, + 0.01796742156147957, + 0.01575786992907524, + 0.021164029836654663, + 0.03327956050634384, + 0.014711108058691025, + 0.01871841587126255, + 0.01599229872226715, + 0.020184693858027458, + 0.013248970732092857, + 0.009624774567782879, + 0.013561743311583996, + 0.009807806462049484, + 0.013311794027686119, + 0.01075053308159113, + 0.010373570024967194, + 0.009717359207570553, + 0.003922424279153347, + 0.0041013723239302635, + 0.0033870339393615723, + 0.00437258230522275, + 0.00523380609229207, + 0.004821007139980793, + 0.005258860066533089, + 0.00593617744743824, + 0.017086930572986603, + 0.010358176194131374, + 0.010219351388514042, + 0.010264091193675995, + 0.010176788084208965, + 0.03744850307703018, + 0.007597933989018202, + 0.011740890331566334, + 0.031985457986593246, + 0.046297360211610794, + 0.009081614203751087, + 0.005477632861584425, + 0.011525344103574753, + 0.014372718520462513, + 0.00598655641078949, + 0.011449845507740974, + 0.006957612931728363, + 0.019046729430556297, + 0.005005849525332451, + 0.004707552958279848, + 0.0168351661413908, + 0.00790297333151102, + 0.0031207518186420202, + 0.004718820098787546, + 0.006681239232420921, + 0.006615911610424519, + 0.003864907892420888, + 0.00437748059630394, + 0.005888339132070541, + 0.00853512343019247, + 0.003795604221522808, + 0.0032125841826200485, + 0.005999179091304541, + 0.0069438093341887, + 0.004432473331689835, + 0.0035154535435140133, + 0.005265513435006142, + 0.006871755700558424, + 0.0028966020327061415, + 0.0032633086666464806 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 6 + ], + "selected_reward": -0.0009310677996836603 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene viewed from a drone's bird's-eye perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera slowly and steadily drift forward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the landscape dominated by a field of blooming red poppies?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are clusters of small white and sunny yellow wildflowers interspersed among the poppies?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are the flowers swaying in a gentle breeze creating a rhythmic motion?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.007633563131093979 + ], + "opposite_logprob_tokens": [ + -4.882633686065674 + ], + "logprob_reward": -0.007633563131093979, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.007633563131093979, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.17323702573776245, + 0.09195602685213089, + 0.2451784759759903, + 0.07330790907144547, + 0.06448804587125778, + 0.08741799741983414, + 0.08255893737077713, + 0.08946745097637177, + 0.035975854843854904, + 0.05641226842999458 + ], + "credit_map": [ + 0.011246461421251297, + 0.010800093412399292, + 0.01638164557516575, + 0.02105850912630558, + 0.011434541083872318, + 0.04460449516773224, + 0.04326603189110756, + 0.014445245265960693, + 0.008967767469584942, + 0.010143072344362736, + 0.010145709849894047, + 0.016701601445674896, + 0.01249539665877819, + 0.013687822967767715, + 0.009616089053452015, + 0.010198562406003475, + 0.023673253133893013, + 0.026004161685705185, + 0.030400501564145088, + 0.026760606095194817, + 0.029771292582154274, + 0.04985785111784935, + 0.02549414336681366, + 0.03321666270494461, + 0.008826429955661297, + 0.009182038716971874, + 0.011506488546729088, + 0.008143886923789978, + 0.012953643687069416, + 0.008656824007630348, + 0.007834394462406635, + 0.006204200442880392, + 0.0075563350692391396, + 0.007104542572051287, + 0.006912533659487963, + 0.0062181586399674416, + 0.009309319779276848, + 0.0081716850399971, + 0.009216096252202988, + 0.009999371133744717, + 0.01352408342063427, + 0.014474240131676197, + 0.009379908442497253, + 0.016421882435679436, + 0.007819553837180138, + 0.011434854939579964, + 0.0071058268658816814, + 0.00725764874368906, + 0.018546832725405693, + 0.022645382210612297, + 0.006850316189229488, + 0.004140048753470182, + 0.008109678514301777, + 0.009845506399869919, + 0.0055741542018949986, + 0.006847010459750891, + 0.008989151567220688, + 0.022282088175415993, + 0.00830351933836937, + 0.008115069009363651, + 0.020945748314261436, + 0.009711120277643204, + 0.004780281335115433, + 0.006340473424643278, + 0.004024415276944637, + 0.006336349993944168, + 0.004057957790791988, + 0.005022258497774601, + 0.003969935700297356, + 0.007315830327570438, + 0.0030049625784158707, + 0.002244145842269063, + 0.006260171998292208, + 0.00874156691133976, + 0.006134398281574249, + 0.005311549641191959, + 0.007860305719077587, + 0.014297855086624622, + 0.003949792589992285, + 0.0038566221483051777 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 0, + 1 + ], + "selected_reward": -0.007633563131093979 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene viewed from a drone's bird's-eye perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera slowly and steadily drift forward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the landscape dominated by a field of blooming red poppies?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are clusters of small white and sunny yellow wildflowers interspersed among the poppies?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are the flowers swaying in a gentle breeze creating a rhythmic motion?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.16036678850650787 + ], + "opposite_logprob_tokens": [ + -1.9103667736053467 + ], + "logprob_reward": -0.16036678850650787, + "yes_no_margin_reward": 1.75, + "token_credit_reward": -0.16036678850650787, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2878829538822174, + 0.11617495864629745, + 0.2027524709701538, + 0.10639189928770065, + 0.0523531548678875, + 0.0407419428229332, + 0.045494742691516876, + 0.07120499759912491, + 0.02921568788588047, + 0.04778716713190079 + ], + "credit_map": [ + 0.013500811532139778, + 0.02032570354640484, + 0.01766051910817623, + 0.01637415960431099, + 0.02013067528605461, + 0.09647734463214874, + 0.08731304109096527, + 0.016100717708468437, + 0.009181422181427479, + 0.00868635531514883, + 0.015525013208389282, + 0.02518600970506668, + 0.008828622289001942, + 0.015014534816145897, + 0.016375716775655746, + 0.01737728901207447, + 0.019476493820548058, + 0.018644696101546288, + 0.023075183853507042, + 0.0379716232419014, + 0.022425206378102303, + 0.03503955155611038, + 0.020850224420428276, + 0.0252695195376873, + 0.012601464055478573, + 0.014927114360034466, + 0.018403103575110435, + 0.01130920834839344, + 0.01125427521765232, + 0.01441811304539442, + 0.010796144604682922, + 0.012682483531534672, + 0.004703341983258724, + 0.006023609545081854, + 0.006260863039642572, + 0.007623426150530577, + 0.006922714877873659, + 0.007508665323257446, + 0.0073356362991034985, + 0.0059748985804617405, + 0.003764575347304344, + 0.00666396226733923, + 0.004131261724978685, + 0.006166272796690464, + 0.005672041326761246, + 0.005226584151387215, + 0.0048680598847568035, + 0.004249188117682934, + 0.010877870954573154, + 0.0037273254711180925, + 0.004504753276705742, + 0.003491032402962446, + 0.0053044818341732025, + 0.010954264551401138, + 0.0027052692603319883, + 0.0039297472685575485, + 0.008362459018826485, + 0.01714271865785122, + 0.006164637394249439, + 0.006410927977412939, + 0.011830889619886875, + 0.011550845578312874, + 0.0041237687692046165, + 0.005618751514703035, + 0.0037747113965451717, + 0.003998815082013607, + 0.0031103473156690598, + 0.0036928646732121706, + 0.0039658513851463795, + 0.004947342909872532, + 0.003376177977770567, + 0.002349576912820339, + 0.004788183607161045, + 0.0062862578779459, + 0.004862628411501646, + 0.005913480184972286, + 0.008280382491648197, + 0.010260420851409435, + 0.003931312821805477, + 0.003464504610747099 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 1 + ], + "selected_reward": -0.16036678850650787 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "From a drone's bird's-eye perspective, the camera slowly and steadily drifts forward, offering a comprehensive and immersive view of a vast, vibrant natural scene. The landscape is dominated by a field of blooming red poppies. Their delicate, paper-thin petals create a brilliant sea of crimson, contrasting beautifully with the lush, dark green foliage of their stems and leaves. Organically interspersed among the poppies are clusters of small white and sunny yellow wildflowers, whose bright, cheerful colors enhance the scene's rich diversity. The entire field is bathed in the warm, soft light of the afternoon sun, deeply saturating colors and casting soft, elongated shadows that add depth and texture to the landscape. A gentle breeze sways the countless flower heads in a gentle, hypnotic rhythm, creating a tranquil and picturesque atmosphere that celebrates the botanical richness and wild beauty.", + "seed": 2223, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000006/group_13/index_2074_seed_2223_dp_27_rank_54_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 13, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.08674579310463741, + "vlm_logprob": -0.08674579310463741, + "vlm_yes_no_margin": 3.975000023841858, + "vlm_token_credit": -0.08674579310463741, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene viewed from a drone's bird's-eye perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera slowly and steadily drift forward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the landscape dominated by a field of blooming red poppies?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are clusters of small white and sunny yellow wildflowers interspersed among the poppies?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are the flowers swaying in a gentle breeze creating a rhythmic motion?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.24230147898197174 + ], + "opposite_logprob_tokens": [ + -1.6173014640808105 + ], + "logprob_reward": -0.24230147898197174, + "yes_no_margin_reward": 1.375, + "token_credit_reward": -0.24230147898197174, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1531088948249817, + 0.2510857880115509, + 0.01613408327102661, + 0.052944451570510864, + 0.037149958312511444, + 0.03449711948633194, + 0.04938525706529617, + 0.06684271991252899, + 0.14592504501342773, + 0.19292668998241425 + ], + "credit_map": [ + 0.01881273090839386, + 0.01136008184403181, + 0.013639366254210472, + 0.050957027822732925, + 0.011922081932425499, + 0.009229111485183239, + 0.01989346742630005, + 0.017295027151703835, + 0.021322965621948242, + 0.012907616794109344, + 0.015374635346233845, + 0.017855798825621605, + 0.010611656121909618, + 0.013818608596920967, + 0.08457382768392563, + 0.07462067902088165, + 0.002544859889894724, + 0.0020338434260338545, + 0.0017294837161898613, + 0.001904154778458178, + 0.0026733994018286467, + 0.001823912956751883, + 0.0017243999755010009, + 0.0017000295920297503, + 0.004798117093741894, + 0.005243266932666302, + 0.007756003178656101, + 0.017383242025971413, + 0.0029780373442918062, + 0.0033769013825803995, + 0.006362630054354668, + 0.005046251695603132, + 0.004026568029075861, + 0.007794191129505634, + 0.00500348350033164, + 0.0039403410628438, + 0.003471076488494873, + 0.005154292099177837, + 0.0034011355601251125, + 0.004358870442956686, + 0.0050062937662005424, + 0.0060197035782039165, + 0.003967387601733208, + 0.00303328363224864, + 0.004416308365762234, + 0.0048968433402478695, + 0.003907658159732819, + 0.00324964034371078, + 0.005852670408785343, + 0.0050578550435602665, + 0.0029107718728482723, + 0.006503945216536522, + 0.004407485015690327, + 0.0033458909019827843, + 0.008110334165394306, + 0.013196306303143501, + 0.0073622423224151134, + 0.006926841102540493, + 0.00952801201492548, + 0.006382714491337538, + 0.008059227839112282, + 0.004485826473683119, + 0.01645340397953987, + 0.007644451688975096, + 0.014755886048078537, + 0.01910397782921791, + 0.022586369886994362, + 0.023565536364912987, + 0.023389620706439018, + 0.012349570170044899, + 0.017856460064649582, + 0.012317625805735588, + 0.02503896877169609, + 0.03021276369690895, + 0.015312029980123043, + 0.010743505321443081, + 0.06531126797199249, + 0.012160273268818855, + 0.016765166074037552, + 0.017382726073265076 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 9, + 0 + ], + "selected_reward": -0.24230147898197174 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene viewed from a drone's bird's-eye perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera slowly and steadily drift forward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the landscape dominated by a field of blooming red poppies?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are clusters of small white and sunny yellow wildflowers interspersed among the poppies?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are the flowers swaying in a gentle breeze creating a rhythmic motion?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04309982433915138 + ], + "opposite_logprob_tokens": [ + -3.168099880218506 + ], + "logprob_reward": -0.04309982433915138, + "yes_no_margin_reward": 3.125, + "token_credit_reward": -0.04309982433915138, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10767042636871338, + 0.14565294981002808, + 0.045035019516944885, + 0.04627719894051552, + 0.0633450299501419, + 0.053796108812093735, + 0.09286194294691086, + 0.047131381928920746, + 0.2573120892047882, + 0.14091789722442627 + ], + "credit_map": [ + 0.008482815697789192, + 0.007691231090575457, + 0.013473663479089737, + 0.019815828651189804, + 0.008884833194315434, + 0.0187488105148077, + 0.016102302819490433, + 0.014470932073891163, + 0.011036751791834831, + 0.007297167554497719, + 0.008275967091321945, + 0.010604830458760262, + 0.008789950050413609, + 0.011464194394648075, + 0.04655981808900833, + 0.04162425920367241, + 0.009562812745571136, + 0.004938263911753893, + 0.005228132009506226, + 0.005592379719018936, + 0.004874840844422579, + 0.004700201563537121, + 0.005914346314966679, + 0.004224044270813465, + 0.0072686574421823025, + 0.009111783467233181, + 0.004420863930135965, + 0.008153311908245087, + 0.004632818978279829, + 0.004887001123279333, + 0.003979213070124388, + 0.00382354692555964, + 0.011089916341006756, + 0.011890620924532413, + 0.005670545157045126, + 0.005276314914226532, + 0.005080499220639467, + 0.008545036427676678, + 0.005205499939620495, + 0.010586597956717014, + 0.006642859894782305, + 0.00506650609895587, + 0.01534169353544712, + 0.004812651313841343, + 0.004067752510309219, + 0.0049162134528160095, + 0.006305886898189783, + 0.006642539985477924, + 0.013280028477311134, + 0.011219536885619164, + 0.0062080915085971355, + 0.01090766116976738, + 0.009536064229905605, + 0.005867770407348871, + 0.013697788119316101, + 0.022144995629787445, + 0.005821629893034697, + 0.005167229566723108, + 0.006535619031637907, + 0.005774592515081167, + 0.0066229840740561485, + 0.004804509226232767, + 0.007732557598501444, + 0.004672259092330933, + 0.03180612996220589, + 0.037860654294490814, + 0.044932086020708084, + 0.042679812759160995, + 0.02787497639656067, + 0.018096644431352615, + 0.03367052972316742, + 0.02039123699069023, + 0.013270698487758636, + 0.02123752236366272, + 0.015252629294991493, + 0.014764118008315563, + 0.029416874051094055, + 0.01518736407160759, + 0.01641731895506382, + 0.015371356159448624 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 1, + 9 + ], + "selected_reward": -0.04309982433915138 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene viewed from a drone's bird's-eye perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera slowly and steadily drift forward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the landscape dominated by a field of blooming red poppies?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are clusters of small white and sunny yellow wildflowers interspersed among the poppies?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are the flowers swaying in a gentle breeze creating a rhythmic motion?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0002743821241892874 + ], + "opposite_logprob_tokens": [ + -8.250274658203125 + ], + "logprob_reward": -0.0002743821241892874, + "yes_no_margin_reward": 8.25, + "token_credit_reward": -0.0002743821241892874, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14728565514087677, + 0.13260652124881744, + 0.027028996497392654, + 0.07109396904706955, + 0.02541472762823105, + 0.06692683696746826, + 0.06852307915687561, + 0.1021428108215332, + 0.2556195855140686, + 0.10335789620876312 + ], + "credit_map": [ + 0.01472858153283596, + 0.01339177880436182, + 0.017812632024288177, + 0.02718084305524826, + 0.01319084968417883, + 0.018427006900310516, + 0.023130062967538834, + 0.019423875957727432, + 0.016333498060703278, + 0.013739312067627907, + 0.008005427196621895, + 0.008656250312924385, + 0.006520268507301807, + 0.009204693138599396, + 0.03280084580183029, + 0.037346214056015015, + 0.00613575940951705, + 0.0027384425047785044, + 0.0025327964685857296, + 0.002337538870051503, + 0.004182902164757252, + 0.002816400257870555, + 0.003345305332913995, + 0.0029398498591035604, + 0.007300501223653555, + 0.008272690698504448, + 0.01024581491947174, + 0.022928545251488686, + 0.003738208906725049, + 0.004693933762609959, + 0.008849791251122952, + 0.005064471159130335, + 0.003169352887198329, + 0.005799395032227039, + 0.002205260330811143, + 0.002123812213540077, + 0.0022479272447526455, + 0.005211214069277048, + 0.0015081901801750064, + 0.00314957438968122, + 0.0074816434644162655, + 0.005275363568216562, + 0.009089899249374866, + 0.003893796354532242, + 0.003571163397282362, + 0.004586423747241497, + 0.010052552446722984, + 0.02297598496079445, + 0.007993366569280624, + 0.00576914893463254, + 0.009244936518371105, + 0.006078700534999371, + 0.004182365257292986, + 0.004697646014392376, + 0.011047156527638435, + 0.01950974389910698, + 0.011169051751494408, + 0.008469915017485619, + 0.015743285417556763, + 0.007402120158076286, + 0.008471636101603508, + 0.007098215166479349, + 0.03155006840825081, + 0.01223850715905428, + 0.030406594276428223, + 0.03885725140571594, + 0.04129759967327118, + 0.04581531509757042, + 0.030502216890454292, + 0.018047990277409554, + 0.030862435698509216, + 0.0198301263153553, + 0.015290206298232079, + 0.015840215608477592, + 0.008276223205029964, + 0.006473056506365538, + 0.03402873873710632, + 0.007544539403170347, + 0.006221313029527664, + 0.009683596901595592 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 0, + 1 + ], + "selected_reward": -0.0002743821241892874 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene viewed from a drone's bird's-eye perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera slowly and steadily drift forward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the landscape dominated by a field of blooming red poppies?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are clusters of small white and sunny yellow wildflowers interspersed among the poppies?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are the flowers swaying in a gentle breeze creating a rhythmic motion?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005255929194390774 + ], + "opposite_logprob_tokens": [ + -5.255255699157715 + ], + "logprob_reward": -0.005255929194390774, + "yes_no_margin_reward": 5.25, + "token_credit_reward": -0.005255929194390774, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10772965848445892, + 0.10016685724258423, + 0.06894592940807343, + 0.20417913794517517, + 0.048322927206754684, + 0.060625530779361725, + 0.10049238801002502, + 0.06217554211616516, + 0.22243595123291016, + 0.024926064535975456 + ], + "credit_map": [ + 0.011288661509752274, + 0.013299302197992802, + 0.01110921986401081, + 0.022367123514413834, + 0.009618064388632774, + 0.01304856687784195, + 0.015624484047293663, + 0.011374247260391712, + 0.01837572455406189, + 0.007203646004199982, + 0.006787468679249287, + 0.00709678465500474, + 0.005874004680663347, + 0.008055736310780048, + 0.016687344759702682, + 0.030086148530244827, + 0.011326699517667294, + 0.006841187831014395, + 0.006990611087530851, + 0.007060868199914694, + 0.009859255515038967, + 0.007978198118507862, + 0.009846676141023636, + 0.009042431600391865, + 0.017919447273015976, + 0.01867394894361496, + 0.03143475204706192, + 0.07323987782001495, + 0.008416041731834412, + 0.011794829741120338, + 0.027933930978178978, + 0.014766295440495014, + 0.0062810941599309444, + 0.01196212787181139, + 0.005102938506752253, + 0.005745969247072935, + 0.0037474201526492834, + 0.0052820490673184395, + 0.004082630388438702, + 0.006118698976933956, + 0.00644052866846323, + 0.006148878484964371, + 0.01656966097652912, + 0.004187739919871092, + 0.005172827281057835, + 0.006994074210524559, + 0.0047003584913909435, + 0.010411460883915424, + 0.010764526203274727, + 0.01484182570129633, + 0.01567765325307846, + 0.011608364060521126, + 0.011541458778083324, + 0.01010132860392332, + 0.011603525839745998, + 0.014353706501424313, + 0.0063489628955721855, + 0.0052780588157474995, + 0.009454580955207348, + 0.004539678804576397, + 0.005135430954396725, + 0.0054465848952531815, + 0.018701402470469475, + 0.00727084232494235, + 0.024034718051552773, + 0.046588316559791565, + 0.016600294038653374, + 0.01957681216299534, + 0.04933853819966316, + 0.02009640820324421, + 0.030899807810783386, + 0.015301056206226349, + 0.003129831049591303, + 0.004783326294273138, + 0.002388749737292528, + 0.00238374643959105, + 0.0038829846307635307, + 0.0027705784887075424, + 0.003034653142094612, + 0.0025521968491375446 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 3, + 0 + ], + "selected_reward": -0.005255929194390774 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene viewed from a drone's bird's-eye perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera slowly and steadily drift forward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the landscape dominated by a field of blooming red poppies?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are clusters of small white and sunny yellow wildflowers interspersed among the poppies?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are the flowers swaying in a gentle breeze creating a rhythmic motion?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.1427973508834839 + ], + "opposite_logprob_tokens": [ + -2.0177974700927734 + ], + "logprob_reward": -0.1427973508834839, + "yes_no_margin_reward": 1.8750001192092896, + "token_credit_reward": -0.1427973508834839, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06879256665706635, + 0.34382951259613037, + 0.02142859436571598, + 0.05953055992722511, + 0.06874881684780121, + 0.038981325924396515, + 0.1378159373998642, + 0.09835289418697357, + 0.12310925871133804, + 0.03941049799323082 + ], + "credit_map": [ + 0.007287187501788139, + 0.005720778368413448, + 0.009661910124123096, + 0.01677040383219719, + 0.006741153541952372, + 0.006659099832177162, + 0.008372912183403969, + 0.007579117547720671, + 0.022124240174889565, + 0.013927299529314041, + 0.02068527787923813, + 0.025826163589954376, + 0.015139119699597359, + 0.01692255213856697, + 0.12621690332889557, + 0.10298793762922287, + 0.004914726596325636, + 0.00356886675581336, + 0.002178886905312538, + 0.0026767973322421312, + 0.0019220474641770124, + 0.0022021078038960695, + 0.002237337175756693, + 0.0017278240993618965, + 0.006588132586330175, + 0.0072276354767382145, + 0.00717140780761838, + 0.016814179718494415, + 0.0039022336713969707, + 0.003734919475391507, + 0.006817151792347431, + 0.007274901028722525, + 0.011071208864450455, + 0.013646622188389301, + 0.006163479760289192, + 0.0069863214157521725, + 0.004554798360913992, + 0.008662578649818897, + 0.00499620009213686, + 0.012667608447372913, + 0.0034336538519710302, + 0.004718346055597067, + 0.009462463669478893, + 0.003226425964385271, + 0.0034354040399193764, + 0.00453685550019145, + 0.005703432485461235, + 0.004464747384190559, + 0.015937695279717445, + 0.014029168523848057, + 0.017185157164931297, + 0.014783679507672787, + 0.010411661118268967, + 0.010128319263458252, + 0.019666779786348343, + 0.035673484206199646, + 0.01039261557161808, + 0.006076078396290541, + 0.016648685559630394, + 0.006200716830790043, + 0.006971872411668301, + 0.006491014268249273, + 0.033776845782995224, + 0.011795060709118843, + 0.014339469373226166, + 0.027887800708413124, + 0.0090562105178833, + 0.011379786767065525, + 0.027090691030025482, + 0.0104190893471241, + 0.015224860049784184, + 0.007711343001574278, + 0.005938563030213118, + 0.00643288716673851, + 0.002514428459107876, + 0.0025444719940423965, + 0.012416034005582333, + 0.0032736097928136587, + 0.0034127228427678347, + 0.0028777779079973698 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 6, + 8 + ], + "selected_reward": -0.1427973508834839 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "From a drone's bird's-eye perspective, the camera slowly and steadily drifts forward, offering a comprehensive and immersive view of a vast, vibrant natural scene. The landscape is dominated by a field of blooming red poppies. Their delicate, paper-thin petals create a brilliant sea of crimson, contrasting beautifully with the lush, dark green foliage of their stems and leaves. Organically interspersed among the poppies are clusters of small white and sunny yellow wildflowers, whose bright, cheerful colors enhance the scene's rich diversity. The entire field is bathed in the warm, soft light of the afternoon sun, deeply saturating colors and casting soft, elongated shadows that add depth and texture to the landscape. A gentle breeze sways the countless flower heads in a gentle, hypnotic rhythm, creating a tranquil and picturesque atmosphere that celebrates the botanical richness and wild beauty.", + "seed": 2224, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000006/group_13/index_2074_seed_2224_dp_27_rank_54_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 13, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.14734934217412957, + "vlm_logprob": -0.14734934217412957, + "vlm_yes_no_margin": 3.625000011920929, + "vlm_token_credit": -0.14734934217412957, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene viewed from a drone's bird's-eye perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera slowly and steadily drift forward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the landscape dominated by a field of blooming red poppies?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are clusters of small white and sunny yellow wildflowers interspersed among the poppies?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are the flowers swaying in a gentle breeze creating a rhythmic motion?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.44410914182662964 + ], + "opposite_logprob_tokens": [ + -1.0691092014312744 + ], + "logprob_reward": -0.44410914182662964, + "yes_no_margin_reward": 0.6250000596046448, + "token_credit_reward": -0.44410914182662964, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2496257871389389, + 0.09082098305225372, + 0.038732338696718216, + 0.1953476071357727, + 0.026480993255972862, + 0.05503470450639725, + 0.037517476826906204, + 0.014757359400391579, + 0.2794346213340759, + 0.012248066253960133 + ], + "credit_map": [ + 0.027077293023467064, + 0.0170136708766222, + 0.01827194169163704, + 0.026247965171933174, + 0.019194718450307846, + 0.02996884472668171, + 0.03364379703998566, + 0.07820755988359451, + 0.009178313426673412, + 0.01157347857952118, + 0.009394449181854725, + 0.009285976178944111, + 0.01179178524762392, + 0.017694439738988876, + 0.012714726850390434, + 0.00918782502412796, + 0.002732120454311371, + 0.006492900662124157, + 0.0023475848138332367, + 0.005694161169230938, + 0.0062562511302530766, + 0.009308128617703915, + 0.004046209156513214, + 0.0018549823435023427, + 0.00929311290383339, + 0.07037220150232315, + 0.009998494759202003, + 0.007044685073196888, + 0.06471594423055649, + 0.013756214641034603, + 0.00822883378714323, + 0.011938132345676422, + 0.0025393194518983364, + 0.0031498288735747337, + 0.002125456929206848, + 0.0023674038238823414, + 0.0037144336383789778, + 0.006038073915988207, + 0.0035053507890552282, + 0.003041126998141408, + 0.0047792112454771996, + 0.00809907354414463, + 0.004955065902322531, + 0.007121027447283268, + 0.006439232267439365, + 0.0070641059428453445, + 0.004847059957683086, + 0.011729931458830833, + 0.0036234217695891857, + 0.004569261334836483, + 0.0035222203005105257, + 0.0028087759856134653, + 0.008797833696007729, + 0.008577749133110046, + 0.0025145085528492928, + 0.003103705355897546, + 0.0012455143732950091, + 0.0017637636046856642, + 0.0030218944884836674, + 0.0023279613815248013, + 0.0019287632312625647, + 0.0017570824129506946, + 0.0011875361669808626, + 0.0015248445561155677, + 0.010363221168518066, + 0.016349701210856438, + 0.0037330971099436283, + 0.004389407113194466, + 0.11493392288684845, + 0.1190449520945549, + 0.005905764177441597, + 0.0047145565040409565, + 0.0008575869724154472, + 0.0024593262933194637, + 0.0012585638323798776, + 0.0014253646368160844, + 0.0011048137675970793, + 0.002225805539637804, + 0.0013739372370764613, + 0.0015426684403792024 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 0, + 3 + ], + "selected_reward": -0.44410914182662964 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene viewed from a drone's bird's-eye perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera slowly and steadily drift forward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the landscape dominated by a field of blooming red poppies?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are clusters of small white and sunny yellow wildflowers interspersed among the poppies?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are the flowers swaying in a gentle breeze creating a rhythmic motion?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.062078870832920074 + ], + "opposite_logprob_tokens": [ + -2.8120789527893066 + ], + "logprob_reward": -0.062078870832920074, + "yes_no_margin_reward": 2.75, + "token_credit_reward": -0.062078870832920074, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3094330430030823, + 0.021985819563269615, + 0.12154027074575424, + 0.21371719241142273, + 0.07657275348901749, + 0.030878858640789986, + 0.07023324072360992, + 0.014708678238093853, + 0.12311667203903198, + 0.017813483253121376 + ], + "credit_map": [ + 0.01949348859488964, + 0.020861711353063583, + 0.020165927708148956, + 0.026468267664313316, + 0.08213339745998383, + 0.03649969398975372, + 0.02863388881087303, + 0.0751766487956047, + 0.0026397292967885733, + 0.0030461607966572046, + 0.002914276672527194, + 0.0021965461783111095, + 0.001759674516506493, + 0.003793769981712103, + 0.0031186859123408794, + 0.0025169774889945984, + 0.0074049546383321285, + 0.01404555980116129, + 0.006053676828742027, + 0.009930361062288284, + 0.025725875049829483, + 0.038841888308525085, + 0.013252115808427334, + 0.006285838782787323, + 0.010046291165053844, + 0.08069198578596115, + 0.008427600376307964, + 0.007737055420875549, + 0.07299049198627472, + 0.0134371817111969, + 0.00761660560965538, + 0.012769977562129498, + 0.007430667988955975, + 0.01654643751680851, + 0.005666016601026058, + 0.007788720075041056, + 0.013917484320700169, + 0.015165521763265133, + 0.00476906867697835, + 0.005288835149258375, + 0.002625253750011325, + 0.0034701453987509012, + 0.003867609892040491, + 0.0051064989529550076, + 0.003415748942643404, + 0.0038180227857083082, + 0.004010471981018782, + 0.00456510903313756, + 0.003294130554422736, + 0.004150791559368372, + 0.004437112249433994, + 0.008998827077448368, + 0.019383229315280914, + 0.021153120324015617, + 0.00508273346349597, + 0.0037332954816520214, + 0.001282961806282401, + 0.0025866201613098383, + 0.0018283205572515726, + 0.002159931231290102, + 0.0023137989919632673, + 0.0020463839173316956, + 0.0013696213718503714, + 0.0011210412485525012, + 0.0054947203025221825, + 0.008436550386250019, + 0.0028042008634656668, + 0.0029986926820129156, + 0.047145452350378036, + 0.04980490356683731, + 0.003468671813607216, + 0.002963478211313486, + 0.001447844784706831, + 0.0029183560982346535, + 0.0018176855519413948, + 0.0017875905614346266, + 0.0020751189440488815, + 0.0040064603090286255, + 0.0019447813974693418, + 0.0018156454898416996 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 8 + ], + "selected_reward": -0.062078870832920074 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene viewed from a drone's bird's-eye perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera slowly and steadily drift forward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the landscape dominated by a field of blooming red poppies?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are clusters of small white and sunny yellow wildflowers interspersed among the poppies?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are the flowers swaying in a gentle breeze creating a rhythmic motion?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.000350175570929423 + ], + "opposite_logprob_tokens": [ + -8.000349998474121 + ], + "logprob_reward": -0.000350175570929423, + "yes_no_margin_reward": 8.0, + "token_credit_reward": -0.000350175570929423, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.5848722457885742, + 0.027438471093773842, + 0.03272736445069313, + 0.13327860832214355, + 0.03228029981255531, + 0.03335846960544586, + 0.024343328550457954, + 0.008414133451879025, + 0.10897788405418396, + 0.014309237711131573 + ], + "credit_map": [ + 0.03874965384602547, + 0.03768308460712433, + 0.025229215621948242, + 0.03947579488158226, + 0.17346276342868805, + 0.09464932978153229, + 0.04797838255763054, + 0.12764400243759155, + 0.0026465817354619503, + 0.0033703609369695187, + 0.0031321055721491575, + 0.0025759858544915915, + 0.00277429330162704, + 0.00546336779370904, + 0.003865666687488556, + 0.00361010804772377, + 0.002765094395726919, + 0.005515242461115122, + 0.0025662570260465145, + 0.0031576473265886307, + 0.004519783426076174, + 0.009747495874762535, + 0.0027056350372731686, + 0.0017502112314105034, + 0.006297864019870758, + 0.050144877284765244, + 0.003924766555428505, + 0.0055564758367836475, + 0.04184723272919655, + 0.007568532135337591, + 0.004466630052775145, + 0.013472219929099083, + 0.003320199204608798, + 0.008439835160970688, + 0.0019568540155887604, + 0.002553805010393262, + 0.00532379187643528, + 0.007063594646751881, + 0.001624887459911406, + 0.0019973318558186293, + 0.0033450238406658173, + 0.0060404567047953606, + 0.0024520764127373695, + 0.0034806067124009132, + 0.004477012436836958, + 0.006255003158003092, + 0.0025452596601098776, + 0.004763033241033554, + 0.0012169628171250224, + 0.0017663981998339295, + 0.0009496993152424693, + 0.001255010487511754, + 0.008782556280493736, + 0.008369390852749348, + 0.0007829408859834075, + 0.0012203692458570004, + 0.0007237761165015399, + 0.0011497335508465767, + 0.0013676229864358902, + 0.0012733020121231675, + 0.001307376311160624, + 0.0008750595734454691, + 0.0008133611991070211, + 0.0009039015858434141, + 0.004946052096784115, + 0.007916487753391266, + 0.0016476090531796217, + 0.0016909370897337794, + 0.042943164706230164, + 0.045359909534454346, + 0.0021760815288871527, + 0.0022976440377533436, + 0.0013367354404181242, + 0.002533853752538562, + 0.0010554982582107186, + 0.00136734324041754, + 0.002135763643309474, + 0.0033555657137185335, + 0.0014041834510862827, + 0.0011202944442629814 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 8 + ], + "selected_reward": -0.000350175570929423 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene viewed from a drone's bird's-eye perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera slowly and steadily drift forward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the landscape dominated by a field of blooming red poppies?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are clusters of small white and sunny yellow wildflowers interspersed among the poppies?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are the flowers swaying in a gentle breeze creating a rhythmic motion?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004640049301087856 + ], + "opposite_logprob_tokens": [ + -5.379640102386475 + ], + "logprob_reward": -0.004640049301087856, + "yes_no_margin_reward": 5.375, + "token_credit_reward": -0.004640049301087856, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10805322229862213, + 0.12295831739902496, + 0.07796209305524826, + 0.12137710303068161, + 0.05237321928143501, + 0.13025496900081635, + 0.10834775865077972, + 0.03864871338009834, + 0.20282500982284546, + 0.037199635058641434 + ], + "credit_map": [ + 0.019624318927526474, + 0.00976730789989233, + 0.008598921820521355, + 0.009546559303998947, + 0.008292059414088726, + 0.012474640272557735, + 0.01270765345543623, + 0.027041753754019737, + 0.01395875308662653, + 0.018387731164693832, + 0.014356664381921291, + 0.015348058193922043, + 0.010061763226985931, + 0.017438432201743126, + 0.01986299268901348, + 0.013543913140892982, + 0.005913522560149431, + 0.008216412737965584, + 0.006085823755711317, + 0.012652360834181309, + 0.010675720870494843, + 0.02319575846195221, + 0.0067655956372618675, + 0.004456899128854275, + 0.006178067997097969, + 0.01998928375542164, + 0.0053519755601882935, + 0.0046543520875275135, + 0.04535667598247528, + 0.022177768871188164, + 0.00846028421074152, + 0.009208697825670242, + 0.005190222058445215, + 0.0075212037190794945, + 0.00803273543715477, + 0.00714525394141674, + 0.008849842473864555, + 0.006194537039846182, + 0.0046819159761071205, + 0.004757512826472521, + 0.01119834091514349, + 0.02292131632566452, + 0.010412962175905704, + 0.015064056031405926, + 0.017160972580313683, + 0.025858065113425255, + 0.010916560888290405, + 0.01672269217669964, + 0.0047743418253958225, + 0.009746873751282692, + 0.004571186378598213, + 0.009196505881845951, + 0.0331852100789547, + 0.03635404631495476, + 0.005039626732468605, + 0.005479964427649975, + 0.0028683205600827932, + 0.004651801660656929, + 0.006864370312541723, + 0.005947055295109749, + 0.005823038052767515, + 0.003648736048489809, + 0.003910788334906101, + 0.004934604279696941, + 0.0073958816938102245, + 0.011464053764939308, + 0.0020942508708685637, + 0.0023938636295497417, + 0.08520946651697159, + 0.08742745965719223, + 0.0036496385000646114, + 0.0031903842464089394, + 0.00246925326064229, + 0.0071911681443452835, + 0.004613279830664396, + 0.004878008272498846, + 0.0042938352562487125, + 0.005482333712279797, + 0.004364837892353535, + 0.003906916826963425 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 5, + 1 + ], + "selected_reward": -0.004640049301087856 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene viewed from a drone's bird's-eye perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera slowly and steadily drift forward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the landscape dominated by a field of blooming red poppies?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are clusters of small white and sunny yellow wildflowers interspersed among the poppies?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are the flowers swaying in a gentle breeze creating a rhythmic motion?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.2255684733390808 + ], + "opposite_logprob_tokens": [ + -1.6005685329437256 + ], + "logprob_reward": -0.2255684733390808, + "yes_no_margin_reward": 1.375, + "token_credit_reward": -0.2255684733390808, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.5304087400436401, + 0.06137903407216072, + 0.038612231612205505, + 0.08539202064275742, + 0.03487512469291687, + 0.02256046049296856, + 0.048264432698488235, + 0.03146625682711601, + 0.13331346213817596, + 0.013728293590247631 + ], + "credit_map": [ + 0.02756608836352825, + 0.038075003772974014, + 0.0226333849132061, + 0.02737526036798954, + 0.18668420612812042, + 0.11461440473794937, + 0.034810975193977356, + 0.0786493793129921, + 0.006321415770798922, + 0.008608980104327202, + 0.006475966889411211, + 0.006848090328276157, + 0.00940291490405798, + 0.010122762992978096, + 0.007765387650579214, + 0.005833508912473917, + 0.003330893348902464, + 0.003300004405900836, + 0.002986330771818757, + 0.00555465230718255, + 0.005041161552071571, + 0.010573424398899078, + 0.005137716885656118, + 0.0026880428194999695, + 0.004420361947268248, + 0.02817087434232235, + 0.0028761986177414656, + 0.004406188614666462, + 0.02554548718035221, + 0.005574921611696482, + 0.004371218383312225, + 0.010026772506535053, + 0.0029359457548707724, + 0.006025958340615034, + 0.004161476157605648, + 0.004080289043486118, + 0.004443979822099209, + 0.005286470055580139, + 0.0039021067786961794, + 0.004038897342979908, + 0.002256009029224515, + 0.0028172596357762814, + 0.002379395067691803, + 0.0019236825173720717, + 0.002979382174089551, + 0.002301713451743126, + 0.00260287057608366, + 0.005300149321556091, + 0.002974363276734948, + 0.003922861069440842, + 0.005624596029520035, + 0.008135540410876274, + 0.009353290311992168, + 0.009340218268334866, + 0.00423837173730135, + 0.004675189033150673, + 0.002703271806240082, + 0.0046358671970665455, + 0.005466649308800697, + 0.0041194092482328415, + 0.004558245651423931, + 0.004089657217264175, + 0.0028621433302760124, + 0.0030310116708278656, + 0.00528854550793767, + 0.008543004281818867, + 0.0015676832990720868, + 0.0018928754143416882, + 0.05457759276032448, + 0.05659651383757591, + 0.002597946673631668, + 0.002249277662485838, + 0.0014550474006682634, + 0.0023211324587464333, + 0.0009742650436237454, + 0.0009154489962384105, + 0.0021206275559961796, + 0.0036388984881341457, + 0.0010198471136391163, + 0.0012830256018787622 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 3 + ], + "selected_reward": -0.2255684733390808 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000006/group_14/index_tensor([7755, 7755])_group_14_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000006/group_14/index_tensor([7755, 7755])_group_14_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..dc8bdb491fafba408849f4188dc35fc81e91dbb3 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000006/group_14/index_tensor([7755, 7755])_group_14_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "macro close-up shot, taken from a slightly overhead perspective, the camera remains still, focusing the viewer's attention squarely on a pair of steady and deft hands meticulously carving an ice sculpture. A cool spotlight shines diagonally from above, illuminating the subject on a dark stone workbench: a crystal-clear, angular block of ice. One hand, clad in a non-slip black glove, firmly grasps the edge of the block, while the other, bare, with intense concentration, steadily and calmly manipulates a fine carving knife that gleams with a cold metallic sheen. The tip of the knife gently and precisely glides across the ice. With subtle movements of the wrist, the intricate and delicate texture of the forming ice flower is gradually outlined, each incision releasing strands of hair-like ice shavings. Light penetrates the pure ice, refracting within it as captivating specks of light. The edges of the newly carved petals capture and reflect the brilliant light, complementing the surrounding fine ice crystals. Against the backdrop of a dark background, the entire picture presents an extreme sense of tranquility and concentration, full of awe for exquisite craftsmanship and the beauty of the moment.", + "seed": 7376, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000006/group_14/index_7755_seed_7376_dp_28_rank_56_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 14, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.10901318807154894, + "vlm_logprob": -0.10901318807154894, + "vlm_yes_no_margin": 3.2, + "vlm_token_credit": -0.10901318807154894, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a macro close-up taken from a slightly overhead perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera remaining still throughout the video?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are a pair of hands visible, one wearing a non-slip black glove and the other bare?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a fine carving knife with a cold metallic sheen being used to carve a crystal-clear block of ice?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is a cool spotlight shining diagonally from above to illuminate the scene on a dark stone workbench?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04091259092092514 + ], + "opposite_logprob_tokens": [ + -5.915912628173828 + ], + "logprob_reward": -0.04091259092092514, + "yes_no_margin_reward": 5.875, + "token_credit_reward": -0.04091259092092514, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1686466932296753, + 0.12299670279026031, + 0.06609771400690079, + 0.07699950039386749, + 0.16861429810523987, + 0.05754350498318672, + 0.059440892189741135, + 0.11526679992675781, + 0.08025607466697693, + 0.08413779735565186 + ], + "credit_map": [ + 0.0313880480825901, + 0.023135004565119743, + 0.018168849870562553, + 0.023227520287036896, + 0.015158817172050476, + 0.028170129284262657, + 0.0144623639062047, + 0.014935961924493313, + 0.01502800639718771, + 0.02593296207487583, + 0.013446136377751827, + 0.017996693029999733, + 0.00868688803166151, + 0.014139658771455288, + 0.011547688394784927, + 0.016218671575188637, + 0.008236627094447613, + 0.013732648454606533, + 0.008072905242443085, + 0.009061506949365139, + 0.006460836157202721, + 0.006461170502007008, + 0.0084419846534729, + 0.0056300382129848, + 0.00985304918140173, + 0.01639002561569214, + 0.009447962045669556, + 0.010633379220962524, + 0.006311143282800913, + 0.007739155553281307, + 0.009005215018987656, + 0.007619569078087807, + 0.02647327072918415, + 0.016779018566012383, + 0.015552148222923279, + 0.01473549846559763, + 0.02351572923362255, + 0.03665649890899658, + 0.015944454818964005, + 0.01895768754184246, + 0.0065222387202084064, + 0.005557852331548929, + 0.009112691506743431, + 0.0068403370678424835, + 0.007828349247574806, + 0.00983540527522564, + 0.006363288499414921, + 0.005483344662934542, + 0.009249578230082989, + 0.0058905137702822685, + 0.007926719263195992, + 0.00433217640966177, + 0.007061023730784655, + 0.009667952544987202, + 0.006222517695277929, + 0.009090413339436054, + 0.023375527933239937, + 0.01348764356225729, + 0.014285667799413204, + 0.01753113977611065, + 0.00871735904365778, + 0.010913893580436707, + 0.009741272777318954, + 0.017214298248291016, + 0.0235929936170578, + 0.007482611574232578, + 0.0054766214452683926, + 0.0076551130041480064, + 0.005820206832140684, + 0.007351852487772703, + 0.005030050408095121, + 0.017846625298261642, + 0.009629400447010994, + 0.009790092706680298, + 0.009853501804172993, + 0.006184684578329325, + 0.015045532025396824, + 0.011137685738503933, + 0.01169939711689949, + 0.010797501541674137 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 1 + ], + "selected_reward": -0.04091259092092514 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a macro close-up taken from a slightly overhead perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera remaining still throughout the video?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are a pair of hands visible, one wearing a non-slip black glove and the other bare?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a fine carving knife with a cold metallic sheen being used to carve a crystal-clear block of ice?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is a cool spotlight shining diagonally from above to illuminate the scene on a dark stone workbench?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.3134842813014984 + ], + "opposite_logprob_tokens": [ + -1.3134843111038208 + ], + "logprob_reward": -0.3134842813014984, + "yes_no_margin_reward": 1.0, + "token_credit_reward": -0.3134842813014984, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08378968387842178, + 0.08768252283334732, + 0.058040328323841095, + 0.05604015290737152, + 0.4103912115097046, + 0.033359888941049576, + 0.07263284921646118, + 0.12602025270462036, + 0.03523758053779602, + 0.036805517971515656 + ], + "credit_map": [ + 0.010639119893312454, + 0.014819168485701084, + 0.010733976028859615, + 0.010047020390629768, + 0.011595881544053555, + 0.01032214891165495, + 0.008465690538287163, + 0.007166676688939333, + 0.012156200595200062, + 0.017488745972514153, + 0.006978548131883144, + 0.007814549840986729, + 0.006548011675477028, + 0.01532046403735876, + 0.007698602043092251, + 0.013677402399480343, + 0.006420763209462166, + 0.007285524625331163, + 0.006859659682959318, + 0.0058062439784407616, + 0.010883507318794727, + 0.008229515515267849, + 0.005093374755233526, + 0.007461739704012871, + 0.007893881760537624, + 0.01232907921075821, + 0.0058848620392382145, + 0.006788427941501141, + 0.006576066371053457, + 0.006469251587986946, + 0.004918317310512066, + 0.005180266220122576, + 0.07317915558815002, + 0.03131645917892456, + 0.020551256835460663, + 0.022570917382836342, + 0.07174571603536606, + 0.09874925017356873, + 0.0363723449409008, + 0.05590612441301346, + 0.004627994727343321, + 0.0034691873006522655, + 0.0038784535136073828, + 0.004464716650545597, + 0.00457861740142107, + 0.005819651763886213, + 0.002822795184329152, + 0.0036984726320952177, + 0.01002566609531641, + 0.006004547234624624, + 0.00790721271187067, + 0.005216024350374937, + 0.010038178414106369, + 0.013268436305224895, + 0.009138169698417187, + 0.01103461068123579, + 0.03364185616374016, + 0.009489977732300758, + 0.01236398983746767, + 0.0074610174633562565, + 0.00986423809081316, + 0.015494127757847309, + 0.012567592784762383, + 0.025137457996606827, + 0.009068304672837257, + 0.0030808551236987114, + 0.0036170436069369316, + 0.001970426645129919, + 0.003023001831024885, + 0.00400325283408165, + 0.001859736628830433, + 0.008614958263933659, + 0.003339716698974371, + 0.005205936264246702, + 0.005064366385340691, + 0.005379466339945793, + 0.005259554833173752, + 0.003990217112004757, + 0.004018011037260294, + 0.004548248834908009 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 7, + 1 + ], + "selected_reward": -0.3134842813014984 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a macro close-up taken from a slightly overhead perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera remaining still throughout the video?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are a pair of hands visible, one wearing a non-slip black glove and the other bare?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a fine carving knife with a cold metallic sheen being used to carve a crystal-clear block of ice?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is a cool spotlight shining diagonally from above to illuminate the scene on a dark stone workbench?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.08906672894954681 + ], + "opposite_logprob_tokens": [ + -2.464066743850708 + ], + "logprob_reward": -0.08906672894954681, + "yes_no_margin_reward": 2.375, + "token_credit_reward": -0.08906672894954681, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2703084349632263, + 0.16754324734210968, + 0.07604469358921051, + 0.056693125516176224, + 0.12468700110912323, + 0.07067885994911194, + 0.042001426219940186, + 0.050484392791986465, + 0.06905606389045715, + 0.0725027471780777 + ], + "credit_map": [ + 0.04251755774021149, + 0.0672399252653122, + 0.026666466146707535, + 0.04375407099723816, + 0.02228527143597603, + 0.021798569709062576, + 0.028413159772753716, + 0.017633410170674324, + 0.026598285883665085, + 0.04753146693110466, + 0.019778676331043243, + 0.022409919649362564, + 0.009418736211955547, + 0.009883904829621315, + 0.01972324773669243, + 0.012199009768664837, + 0.009415443986654282, + 0.013208760879933834, + 0.008866543881595135, + 0.008146507665514946, + 0.013577926903963089, + 0.007080912124365568, + 0.008742954581975937, + 0.007005640771239996, + 0.00722874840721488, + 0.009848111309111118, + 0.009395225904881954, + 0.008382202126085758, + 0.004991450812667608, + 0.00508152786642313, + 0.006471352186053991, + 0.005294507369399071, + 0.016200773417949677, + 0.011177032254636288, + 0.014050852507352829, + 0.008354050107300282, + 0.020766684785485268, + 0.020404882729053497, + 0.012634829618036747, + 0.021097896620631218, + 0.012027278542518616, + 0.008546293713152409, + 0.016222607344388962, + 0.007626055274158716, + 0.004613621160387993, + 0.008207354694604874, + 0.00690208887681365, + 0.00653356546536088, + 0.005973865743726492, + 0.004283461719751358, + 0.0051352581940591335, + 0.005432618781924248, + 0.004864432383328676, + 0.005432655103504658, + 0.004972486756742001, + 0.005906648933887482, + 0.00901496596634388, + 0.005360792391002178, + 0.006722168996930122, + 0.006414592266082764, + 0.0041243815794587135, + 0.005178109742701054, + 0.0069759683683514595, + 0.0066934144124388695, + 0.016573986038565636, + 0.007069522049278021, + 0.005997516214847565, + 0.006764396093785763, + 0.006417921278625727, + 0.007102225441485643, + 0.005306554492563009, + 0.013823941349983215, + 0.007292439695447683, + 0.011570927686989307, + 0.007815994322299957, + 0.00832422822713852, + 0.00946903321892023, + 0.01094144769012928, + 0.0083525525406003, + 0.00873611867427826 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 4 + ], + "selected_reward": -0.08906672894954681 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a macro close-up taken from a slightly overhead perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera remaining still throughout the video?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are a pair of hands visible, one wearing a non-slip black glove and the other bare?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a fine carving knife with a cold metallic sheen being used to carve a crystal-clear block of ice?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is a cool spotlight shining diagonally from above to illuminate the scene on a dark stone workbench?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.012549244798719883 + ], + "opposite_logprob_tokens": [ + -4.38754940032959 + ], + "logprob_reward": -0.012549244798719883, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.012549244798719883, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1499653160572052, + 0.15206627547740936, + 0.052198849618434906, + 0.06640475243330002, + 0.08763325959444046, + 0.13212138414382935, + 0.07213318347930908, + 0.16397343575954437, + 0.03721614554524422, + 0.08628741651773453 + ], + "credit_map": [ + 0.017306294292211533, + 0.021889202296733856, + 0.017798494547605515, + 0.023207714781165123, + 0.014200658537447453, + 0.022246455773711205, + 0.015850450843572617, + 0.01746605522930622, + 0.02228531613945961, + 0.034835126250982285, + 0.020194971933960915, + 0.028797900304198265, + 0.007850484922528267, + 0.010900158435106277, + 0.014990223571658134, + 0.01221209391951561, + 0.005920035298913717, + 0.009800444357097149, + 0.006847534794360399, + 0.0070632402785122395, + 0.00557640939950943, + 0.00730308098718524, + 0.005045605357736349, + 0.004642501473426819, + 0.008613428100943565, + 0.010189895518124104, + 0.009255088865756989, + 0.01145851518958807, + 0.005761123728007078, + 0.005911799147725105, + 0.0074150036089122295, + 0.007799895480275154, + 0.010725341737270355, + 0.0073904013261199, + 0.008227122016251087, + 0.006858262699097395, + 0.015457644127309322, + 0.014194408431649208, + 0.01029460970312357, + 0.014485462568700314, + 0.01334715448319912, + 0.016786888241767883, + 0.024428103119134903, + 0.021669888868927956, + 0.01303771696984768, + 0.013398832641541958, + 0.01654270850121975, + 0.012910090386867523, + 0.012565434910356998, + 0.006427603308111429, + 0.006802200339734554, + 0.00909824576228857, + 0.008195645175874233, + 0.010039693675935268, + 0.009364536963403225, + 0.009639822877943516, + 0.019796399399638176, + 0.03048863261938095, + 0.026664236560463905, + 0.02345576323568821, + 0.011930522508919239, + 0.017526187002658844, + 0.013959061354398727, + 0.02015264518558979, + 0.006821027956902981, + 0.004342969041317701, + 0.004743457771837711, + 0.0040151095017790794, + 0.003719619708135724, + 0.003179003717377782, + 0.004192420281469822, + 0.0062025366351008415, + 0.01110120490193367, + 0.010719687677919865, + 0.010454999282956123, + 0.012787421233952045, + 0.009737002663314342, + 0.010541596449911594, + 0.010289875790476799, + 0.010655625723302364 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 1, + 0 + ], + "selected_reward": -0.012549244798719883 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a macro close-up taken from a slightly overhead perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera remaining still throughout the video?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are a pair of hands visible, one wearing a non-slip black glove and the other bare?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a fine carving knife with a cold metallic sheen being used to carve a crystal-clear block of ice?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is a cool spotlight shining diagonally from above to illuminate the scene on a dark stone workbench?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.08905309438705444 + ], + "opposite_logprob_tokens": [ + -2.464053153991699 + ], + "logprob_reward": -0.08905309438705444, + "yes_no_margin_reward": 2.375, + "token_credit_reward": -0.08905309438705444, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08074931055307388, + 0.3321927785873413, + 0.048494819551706314, + 0.0608832947909832, + 0.13065281510353088, + 0.09208894520998001, + 0.034711576998233795, + 0.03904275968670845, + 0.023443805053830147, + 0.15773990750312805 + ], + "credit_map": [ + 0.009797124192118645, + 0.009258213452994823, + 0.012799595482647419, + 0.008742895908653736, + 0.009921150282025337, + 0.01113600842654705, + 0.008525275625288486, + 0.010569046251475811, + 0.04215694218873978, + 0.09457100927829742, + 0.0368698425590992, + 0.041851501911878586, + 0.019246134907007217, + 0.04223792254924774, + 0.029676219448447227, + 0.025583181530237198, + 0.006376179400831461, + 0.012178155593574047, + 0.005405445117503405, + 0.007321234326809645, + 0.00484309159219265, + 0.003848356893286109, + 0.005000002682209015, + 0.003522352548316121, + 0.009058088064193726, + 0.010312982834875584, + 0.007414042484015226, + 0.01084810495376587, + 0.004866957198828459, + 0.007865958847105503, + 0.006247853860259056, + 0.00426930096000433, + 0.01554914005100727, + 0.016372285783290863, + 0.01961875706911087, + 0.015344151295721531, + 0.013690005987882614, + 0.018935630097985268, + 0.01317292358726263, + 0.017969919368624687, + 0.013534421101212502, + 0.01189177855849266, + 0.013415314257144928, + 0.013286128640174866, + 0.010957229882478714, + 0.01206355169415474, + 0.007065138313919306, + 0.009875372052192688, + 0.00543542206287384, + 0.004096487537026405, + 0.0037680394016206264, + 0.004290487617254257, + 0.0038134416099637747, + 0.004796891938894987, + 0.003569455584511161, + 0.004941347520798445, + 0.005405357573181391, + 0.005077439360320568, + 0.005571017507463694, + 0.008105232380330563, + 0.003010022221133113, + 0.004043660592287779, + 0.0038050522562116385, + 0.0040249754674732685, + 0.004310560412704945, + 0.0025508960243314505, + 0.0027826980222016573, + 0.0025357056874781847, + 0.002234488492831588, + 0.0026874460745602846, + 0.002312206896021962, + 0.004029801581054926, + 0.022223904728889465, + 0.021847160533070564, + 0.022305991500616074, + 0.023209160193800926, + 0.01477803010493517, + 0.016981041058897972, + 0.015045332722365856, + 0.02134927362203598 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 9, + 4 + ], + "selected_reward": -0.08905309438705444 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "macro close-up shot, taken from a slightly overhead perspective, the camera remains still, focusing the viewer's attention squarely on a pair of steady and deft hands meticulously carving an ice sculpture. A cool spotlight shines diagonally from above, illuminating the subject on a dark stone workbench: a crystal-clear, angular block of ice. One hand, clad in a non-slip black glove, firmly grasps the edge of the block, while the other, bare, with intense concentration, steadily and calmly manipulates a fine carving knife that gleams with a cold metallic sheen. The tip of the knife gently and precisely glides across the ice. With subtle movements of the wrist, the intricate and delicate texture of the forming ice flower is gradually outlined, each incision releasing strands of hair-like ice shavings. Light penetrates the pure ice, refracting within it as captivating specks of light. The edges of the newly carved petals capture and reflect the brilliant light, complementing the surrounding fine ice crystals. Against the backdrop of a dark background, the entire picture presents an extreme sense of tranquility and concentration, full of awe for exquisite craftsmanship and the beauty of the moment.", + "seed": 7377, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000006/group_14/index_7755_seed_7377_dp_28_rank_56_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 14, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.13929457757622005, + "vlm_logprob": -0.13929457757622005, + "vlm_yes_no_margin": 3.075, + "vlm_token_credit": -0.13929457757622005, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a macro close-up taken from a slightly overhead perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera remaining still throughout the video?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are a pair of hands visible, one wearing a non-slip black glove and the other bare?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a fine carving knife with a cold metallic sheen being used to carve a crystal-clear block of ice?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is a cool spotlight shining diagonally from above to illuminate the scene on a dark stone workbench?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.05064544826745987 + ], + "opposite_logprob_tokens": [ + -4.925645351409912 + ], + "logprob_reward": -0.05064544826745987, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.05064544826745987, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.5883218050003052, + 0.0473928228020668, + 0.06451987475156784, + 0.03344706818461418, + 0.04028656706213951, + 0.06070874631404877, + 0.054735396057367325, + 0.020148294046521187, + 0.04142347350716591, + 0.04901600256562233 + ], + "credit_map": [ + 0.06213938817381859, + 0.03613567352294922, + 0.04382190480828285, + 0.14313319325447083, + 0.051589760929346085, + 0.0560343936085701, + 0.09552452713251114, + 0.09994293004274368, + 0.009655463509261608, + 0.00606582360342145, + 0.0056296465918421745, + 0.006470344960689545, + 0.00490057747811079, + 0.004140288103371859, + 0.004967956338077784, + 0.00556272454559803, + 0.008296401239931583, + 0.005928048864006996, + 0.008441600017249584, + 0.007385562639683485, + 0.005236447788774967, + 0.01047661155462265, + 0.006151607260107994, + 0.01260360050946474, + 0.003538938704878092, + 0.0028820871375501156, + 0.003324587596580386, + 0.004637848120182753, + 0.0029330519028007984, + 0.0056277913972735405, + 0.005535378586500883, + 0.004967383109033108, + 0.0038277343846857548, + 0.0034644596744328737, + 0.0045259990729391575, + 0.006469259038567543, + 0.003872969886288047, + 0.005958503112196922, + 0.005615535657852888, + 0.006552103906869888, + 0.005692053120583296, + 0.003821617690846324, + 0.010140837170183659, + 0.00441321125254035, + 0.003983458504080772, + 0.010285072028636932, + 0.005250198766589165, + 0.0171222984790802, + 0.007442156784236431, + 0.0034354825038462877, + 0.007689822465181351, + 0.006266773212701082, + 0.0035195902455598116, + 0.005110064055770636, + 0.006888938136398792, + 0.014382569119334221, + 0.0026266765780746937, + 0.0014204096514731646, + 0.00264475354924798, + 0.003169551258906722, + 0.002064561238512397, + 0.002249223180115223, + 0.002549371449276805, + 0.0034237478394061327, + 0.004572301637381315, + 0.0034370000939816236, + 0.006417444907128811, + 0.005943590775132179, + 0.003873699577525258, + 0.004360424820333719, + 0.0049291690811514854, + 0.007889841683208942, + 0.004592953249812126, + 0.0044161295518279076, + 0.008901027962565422, + 0.007222700398415327, + 0.005488752853125334, + 0.005981805268675089, + 0.00619569281116128, + 0.006216943264007568 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 5 + ], + "selected_reward": -0.05064544826745987 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a macro close-up taken from a slightly overhead perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera remaining still throughout the video?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are a pair of hands visible, one wearing a non-slip black glove and the other bare?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a fine carving knife with a cold metallic sheen being used to carve a crystal-clear block of ice?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is a cool spotlight shining diagonally from above to illuminate the scene on a dark stone workbench?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.5233548283576965 + ], + "opposite_logprob_tokens": [ + -0.8983548283576965 + ], + "logprob_reward": -0.5233548283576965, + "yes_no_margin_reward": 0.375, + "token_credit_reward": -0.5233548283576965, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.21151065826416016, + 0.0707361027598381, + 0.15072491765022278, + 0.04272741451859474, + 0.0334470309317112, + 0.02321746200323105, + 0.04499570652842522, + 0.03530871868133545, + 0.016638437286019325, + 0.370693564414978 + ], + "credit_map": [ + 0.026456164196133614, + 0.022923922166228294, + 0.02270517498254776, + 0.041058242321014404, + 0.01917164959013462, + 0.02689262479543686, + 0.02460799366235733, + 0.027694890275597572, + 0.00920555554330349, + 0.006498327478766441, + 0.007329628802835941, + 0.005697448737919331, + 0.0081255491822958, + 0.012744277715682983, + 0.007223861292004585, + 0.013911454007029533, + 0.01703910157084465, + 0.015658443793654442, + 0.022313689813017845, + 0.016803696751594543, + 0.010470358654856682, + 0.026376232504844666, + 0.015011959709227085, + 0.027051422744989395, + 0.0035032082814723253, + 0.004906836897134781, + 0.004792904946953058, + 0.006312170997262001, + 0.003506269073113799, + 0.007791104260832071, + 0.006247367709875107, + 0.005667553748935461, + 0.0030761605594307184, + 0.0030299925711005926, + 0.003427886636927724, + 0.004442241974174976, + 0.0034670792520046234, + 0.005285740364342928, + 0.004015757702291012, + 0.006702173035591841, + 0.0022811973467469215, + 0.0015257090562954545, + 0.004484404809772968, + 0.0024401554837822914, + 0.001473784795962274, + 0.006250425241887569, + 0.002089772140607238, + 0.0026720131281763315, + 0.0056459857150912285, + 0.003017762675881386, + 0.005209171678870916, + 0.004655696917325258, + 0.0030483694281429052, + 0.004639302380383015, + 0.005445878952741623, + 0.013333537615835667, + 0.0038425992242991924, + 0.0025532941799610853, + 0.003935586661100388, + 0.005168669857084751, + 0.004099540878087282, + 0.004225439392030239, + 0.006388561334460974, + 0.005095029715448618, + 0.0019407919608056545, + 0.001142504159361124, + 0.002120313234627247, + 0.0023095079232007265, + 0.0013079498894512653, + 0.0017941558035090566, + 0.0020086460281163454, + 0.004014568869024515, + 0.039951834827661514, + 0.03614961728453636, + 0.06574232876300812, + 0.047295331954956055, + 0.03987396880984306, + 0.04155378043651581, + 0.05793129652738571, + 0.0421953909099102 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 0, + 2 + ], + "selected_reward": -0.5233548283576965 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a macro close-up taken from a slightly overhead perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera remaining still throughout the video?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are a pair of hands visible, one wearing a non-slip black glove and the other bare?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a fine carving knife with a cold metallic sheen being used to carve a crystal-clear block of ice?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is a cool spotlight shining diagonally from above to illuminate the scene on a dark stone workbench?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.054968513548374176 + ], + "opposite_logprob_tokens": [ + -2.9299685955047607 + ], + "logprob_reward": -0.054968513548374176, + "yes_no_margin_reward": 2.875, + "token_credit_reward": -0.054968513548374176, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.4885890781879425, + 0.1239933967590332, + 0.11972080171108246, + 0.05733229219913483, + 0.025282375514507294, + 0.07551795244216919, + 0.026245437562465668, + 0.041282542049884796, + 0.015311898663640022, + 0.026724111288785934 + ], + "credit_map": [ + 0.04770772159099579, + 0.03192909434437752, + 0.03971482068300247, + 0.13687410950660706, + 0.03522258996963501, + 0.0336635559797287, + 0.06867145001888275, + 0.09480581432580948, + 0.017926741391420364, + 0.009446932002902031, + 0.014655251987278461, + 0.014252248220145702, + 0.010085863061249256, + 0.02056305482983589, + 0.011774297803640366, + 0.02528902143239975, + 0.012812103144824505, + 0.007813483476638794, + 0.03638368844985962, + 0.010905064642429352, + 0.006504818331450224, + 0.020247459411621094, + 0.005941138602793217, + 0.019113043323159218, + 0.004549218341708183, + 0.005195919889956713, + 0.009906750172376633, + 0.007098598871380091, + 0.00400191405788064, + 0.00892055593430996, + 0.007279451470822096, + 0.01037988904863596, + 0.0022805246990174055, + 0.0021911649964749813, + 0.0023967470042407513, + 0.0033517233096063137, + 0.002909768372774124, + 0.004540022928267717, + 0.0028726174496114254, + 0.004739807453006506, + 0.007130949757993221, + 0.005004251841455698, + 0.014327270910143852, + 0.005630824249237776, + 0.0041505577974021435, + 0.018668094649910927, + 0.005524045322090387, + 0.015081958845257759, + 0.0031220796518027782, + 0.0021527695935219526, + 0.002784731099382043, + 0.0025177483912557364, + 0.002154894405975938, + 0.0029473109170794487, + 0.0030908032786101103, + 0.007475101854652166, + 0.004443314857780933, + 0.0028316546231508255, + 0.005800323095172644, + 0.005807274486869574, + 0.004066633526235819, + 0.004653584212064743, + 0.0046344054862856865, + 0.009045355021953583, + 0.0014583669835701585, + 0.0011654949048534036, + 0.002365012187510729, + 0.002098072087392211, + 0.0016296925023198128, + 0.0020843108650296926, + 0.0017543418798595667, + 0.0027566079515963793, + 0.0026294924318790436, + 0.002080025617033243, + 0.004821381531655788, + 0.0032434784807264805, + 0.0037136103492230177, + 0.0034586444962769747, + 0.0038336440920829773, + 0.0029438387136906385 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 2 + ], + "selected_reward": -0.054968513548374176 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a macro close-up taken from a slightly overhead perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera remaining still throughout the video?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are a pair of hands visible, one wearing a non-slip black glove and the other bare?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a fine carving knife with a cold metallic sheen being used to carve a crystal-clear block of ice?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is a cool spotlight shining diagonally from above to illuminate the scene on a dark stone workbench?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.012546301819384098 + ], + "opposite_logprob_tokens": [ + -4.387546539306641 + ], + "logprob_reward": -0.012546301819384098, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.012546301819384098, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.35664188861846924, + 0.11683713644742966, + 0.21793299913406372, + 0.04887077212333679, + 0.07496015727519989, + 0.022727398201823235, + 0.022287292405962944, + 0.0449681431055069, + 0.029264329001307487, + 0.0655098408460617 + ], + "credit_map": [ + 0.03759485110640526, + 0.026171740144491196, + 0.03439704328775406, + 0.06446411460638046, + 0.048403315246105194, + 0.034735918045043945, + 0.06055476516485214, + 0.05032013729214668, + 0.022575631737709045, + 0.009183715097606182, + 0.013072382658720016, + 0.019997714087367058, + 0.007093522697687149, + 0.011628932319581509, + 0.019097426906228065, + 0.014187807217240334, + 0.023429561406373978, + 0.03011520579457283, + 0.027232540771365166, + 0.028682978823781013, + 0.023154359310865402, + 0.028605066239833832, + 0.029407279565930367, + 0.027306007221341133, + 0.004613188561052084, + 0.0043622408993542194, + 0.005021636839956045, + 0.007762649562209845, + 0.004128835629671812, + 0.007300932891666889, + 0.007899773307144642, + 0.007781513500958681, + 0.007511095609515905, + 0.00976630300283432, + 0.007706189062446356, + 0.013293708674609661, + 0.009020374156534672, + 0.008655907586216927, + 0.009937138296663761, + 0.009069442749023438, + 0.00252334657125175, + 0.0011957174865528941, + 0.004163875710219145, + 0.0024330024607479572, + 0.0015932716196402907, + 0.004115797113627195, + 0.002221538918092847, + 0.0044808476231992245, + 0.0029139781836420298, + 0.0014713078271597624, + 0.002625467721372843, + 0.002355848904699087, + 0.0017085329163819551, + 0.002049523638561368, + 0.002871258184313774, + 0.006291375495493412, + 0.005327422171831131, + 0.004141522105783224, + 0.0062884497456252575, + 0.005840679164975882, + 0.005247671157121658, + 0.004742107354104519, + 0.005873435642570257, + 0.00750685716047883, + 0.002896165242418647, + 0.0020124916918575764, + 0.003341230796650052, + 0.003914269153028727, + 0.002199062379077077, + 0.0034831706434488297, + 0.0035973091144114733, + 0.007820629514753819, + 0.00721507053822279, + 0.006652668118476868, + 0.011522655375301838, + 0.008221461437642574, + 0.007437617983669043, + 0.007220270112156868, + 0.010686632245779037, + 0.006553466897457838 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 1 + ], + "selected_reward": -0.012546301819384098 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a macro close-up taken from a slightly overhead perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera remaining still throughout the video?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are a pair of hands visible, one wearing a non-slip black glove and the other bare?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a fine carving knife with a cold metallic sheen being used to carve a crystal-clear block of ice?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is a cool spotlight shining diagonally from above to illuminate the scene on a dark stone workbench?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0549577958881855 + ], + "opposite_logprob_tokens": [ + -2.929957866668701 + ], + "logprob_reward": -0.0549577958881855, + "yes_no_margin_reward": 2.875, + "token_credit_reward": -0.0549577958881855, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.4356270730495453, + 0.12801092863082886, + 0.07688058912754059, + 0.066876620054245, + 0.06183967366814613, + 0.049937255680561066, + 0.024154799059033394, + 0.06561895459890366, + 0.03853081166744232, + 0.052523329854011536 + ], + "credit_map": [ + 0.04290705919265747, + 0.02324042282998562, + 0.03296005353331566, + 0.12414924800395966, + 0.026896540075540543, + 0.030214857310056686, + 0.06072571873664856, + 0.09453318268060684, + 0.033777326345443726, + 0.010027111507952213, + 0.013551813550293446, + 0.011980244889855385, + 0.007700157817453146, + 0.012532618828117847, + 0.0123219545930624, + 0.026119699701666832, + 0.010453795082867146, + 0.010806246660649776, + 0.009909148328006268, + 0.009796889498829842, + 0.010031966492533684, + 0.009080300107598305, + 0.008209790103137493, + 0.008592446334660053, + 0.007615572772920132, + 0.008253357373178005, + 0.006662790663540363, + 0.010092747397720814, + 0.007158039137721062, + 0.008819540031254292, + 0.011368836276233196, + 0.006905741058290005, + 0.0067595732398331165, + 0.007243666797876358, + 0.008285809308290482, + 0.009989629499614239, + 0.007367278914898634, + 0.007227176800370216, + 0.00748069304972887, + 0.007485843263566494, + 0.005398614797741175, + 0.003498109756037593, + 0.010021797381341457, + 0.005252390634268522, + 0.003578060306608677, + 0.007692269515246153, + 0.005197104066610336, + 0.009298913180828094, + 0.0038039609789848328, + 0.002229624893516302, + 0.003057832596823573, + 0.0024885537568479776, + 0.002443590434268117, + 0.0028731487691402435, + 0.0031288291793316603, + 0.0041292584501206875, + 0.007754100952297449, + 0.0054412889294326305, + 0.008836143650114536, + 0.007040671072900295, + 0.007645856123417616, + 0.006985300220549107, + 0.008485994301736355, + 0.013429599814116955, + 0.0035119641106575727, + 0.0035709950607270002, + 0.00615834491327405, + 0.005673828534781933, + 0.003667660988867283, + 0.005573843140155077, + 0.004136082716286182, + 0.006238090805709362, + 0.00523017905652523, + 0.004211842082440853, + 0.007528749760240316, + 0.005915510468184948, + 0.005293363239616156, + 0.007021432742476463, + 0.007308048661798239, + 0.010014206171035767 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 2 + ], + "selected_reward": -0.0549577958881855 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "macro close-up shot, taken from a slightly overhead perspective, the camera remains still, focusing the viewer's attention squarely on a pair of steady and deft hands meticulously carving an ice sculpture. A cool spotlight shines diagonally from above, illuminating the subject on a dark stone workbench: a crystal-clear, angular block of ice. One hand, clad in a non-slip black glove, firmly grasps the edge of the block, while the other, bare, with intense concentration, steadily and calmly manipulates a fine carving knife that gleams with a cold metallic sheen. The tip of the knife gently and precisely glides across the ice. With subtle movements of the wrist, the intricate and delicate texture of the forming ice flower is gradually outlined, each incision releasing strands of hair-like ice shavings. Light penetrates the pure ice, refracting within it as captivating specks of light. The edges of the newly carved petals capture and reflect the brilliant light, complementing the surrounding fine ice crystals. Against the backdrop of a dark background, the entire picture presents an extreme sense of tranquility and concentration, full of awe for exquisite craftsmanship and the beauty of the moment.", + "seed": 7378, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000006/group_14/index_7755_seed_7378_dp_29_rank_58_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 14, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.09959844313561916, + "vlm_logprob": -0.09959844313561916, + "vlm_yes_no_margin": 4.099999988079071, + "vlm_token_credit": -0.09959844313561916, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a macro close-up taken from a slightly overhead perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera remaining still throughout the video?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are a pair of hands visible, one wearing a non-slip black glove and the other bare?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a fine carving knife with a cold metallic sheen being used to carve a crystal-clear block of ice?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is a cool spotlight shining diagonally from above to illuminate the scene on a dark stone workbench?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02431791089475155 + ], + "opposite_logprob_tokens": [ + -6.899317741394043 + ], + "logprob_reward": -0.02431791089475155, + "yes_no_margin_reward": 6.875, + "token_credit_reward": -0.02431791089475155, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.25642549991607666, + 0.09931349754333496, + 0.09406200051307678, + 0.07271416485309601, + 0.06059626489877701, + 0.11178620159626007, + 0.05261524021625519, + 0.10832522809505463, + 0.08194348216056824, + 0.06221846863627434 + ], + "credit_map": [ + 0.028647346422076225, + 0.030975686386227608, + 0.033045802265405655, + 0.03806206211447716, + 0.04688704013824463, + 0.02297992818057537, + 0.030538344755768776, + 0.025289304554462433, + 0.01344206091016531, + 0.012891673482954502, + 0.01112768892198801, + 0.0109531469643116, + 0.01080953050404787, + 0.01350074540823698, + 0.01592344418168068, + 0.01066520530730486, + 0.01217743381857872, + 0.01039640512317419, + 0.010600147768855095, + 0.006593120284378529, + 0.013504100032150745, + 0.012246682308614254, + 0.011474799364805222, + 0.01706930622458458, + 0.007160026580095291, + 0.009913831949234009, + 0.006518953014165163, + 0.006820235401391983, + 0.012037060223519802, + 0.008816376328468323, + 0.010487223975360394, + 0.010960454121232033, + 0.006432891357690096, + 0.006067544687539339, + 0.009085407480597496, + 0.005471551790833473, + 0.007929102517664433, + 0.007239053957164288, + 0.008698060177266598, + 0.009672651067376137, + 0.008307201787829399, + 0.009963304735720158, + 0.014410539530217648, + 0.009892245754599571, + 0.012986991554498672, + 0.015602556057274342, + 0.01172736007720232, + 0.028896009549498558, + 0.0044436585158109665, + 0.0041852351278066635, + 0.005733747035264969, + 0.00600525364279747, + 0.007624122314155102, + 0.006466690916568041, + 0.006495966576039791, + 0.011660570278763771, + 0.010718679055571556, + 0.009376139380037785, + 0.013889887370169163, + 0.012110145762562752, + 0.014759927988052368, + 0.01654750294983387, + 0.012797861360013485, + 0.018125085160136223, + 0.009670471772551537, + 0.008287610486149788, + 0.008751895278692245, + 0.010282458737492561, + 0.012632697820663452, + 0.008284094743430614, + 0.012629684060811996, + 0.011404572054743767, + 0.006481128744781017, + 0.004806914832442999, + 0.0069410414434969425, + 0.007368795573711395, + 0.006567598786205053, + 0.0073967864736914635, + 0.013720029033720493, + 0.008936174213886261 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 5, + 7 + ], + "selected_reward": -0.02431791089475155 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a macro close-up taken from a slightly overhead perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera remaining still throughout the video?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are a pair of hands visible, one wearing a non-slip black glove and the other bare?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a fine carving knife with a cold metallic sheen being used to carve a crystal-clear block of ice?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is a cool spotlight shining diagonally from above to illuminate the scene on a dark stone workbench?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.42890530824661255 + ], + "opposite_logprob_tokens": [ + -1.0539052486419678 + ], + "logprob_reward": -0.42890530824661255, + "yes_no_margin_reward": 0.6249999403953552, + "token_credit_reward": -0.42890530824661255, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.20275677740573883, + 0.412465900182724, + 0.12491688132286072, + 0.0385618656873703, + 0.021905967965722084, + 0.025280533358454704, + 0.018134556710720062, + 0.06526832282543182, + 0.0628228560090065, + 0.02788628451526165 + ], + "credit_map": [ + 0.03045751340687275, + 0.02727508172392845, + 0.02200528047978878, + 0.0187839288264513, + 0.027959831058979034, + 0.019578538835048676, + 0.030172420665621758, + 0.026524197310209274, + 0.037886667996644974, + 0.04371698945760727, + 0.04235709458589554, + 0.05103033408522606, + 0.04958435148000717, + 0.05463528633117676, + 0.07585931569337845, + 0.05739589408040047, + 0.016102733090519905, + 0.012715330347418785, + 0.014706110581755638, + 0.01415610034018755, + 0.02169455587863922, + 0.012234917841851711, + 0.014882399700582027, + 0.01842471957206726, + 0.004013671539723873, + 0.004852659534662962, + 0.004823523107916117, + 0.004176665563136339, + 0.006472058594226837, + 0.004984217695891857, + 0.003582723904401064, + 0.005656344350427389, + 0.0021629766561090946, + 0.002482076408341527, + 0.002320702187716961, + 0.0018873481312766671, + 0.00314990128390491, + 0.0032375999726355076, + 0.0034901355393230915, + 0.0031752283684909344, + 0.0020733310375362635, + 0.001221794867888093, + 0.0029977690428495407, + 0.002288424177095294, + 0.0034153165761381388, + 0.003341803327202797, + 0.0024410118348896503, + 0.007501083426177502, + 0.0017819704953581095, + 0.002120919292792678, + 0.00235411967150867, + 0.0018109936499968171, + 0.00279513793066144, + 0.001899865921586752, + 0.0026995379012078047, + 0.002672011498361826, + 0.009099750779569149, + 0.006049866788089275, + 0.009209135547280312, + 0.008993287570774555, + 0.006896249949932098, + 0.008009935729205608, + 0.008379126898944378, + 0.008630974218249321, + 0.007678526919335127, + 0.006013123318552971, + 0.007645496632903814, + 0.0076600899919867516, + 0.00681920675560832, + 0.006852293852716684, + 0.007671887055039406, + 0.012482238933444023, + 0.004106492735445499, + 0.002073551993817091, + 0.004766383208334446, + 0.003134895348921418, + 0.0027645365335047245, + 0.0034496851731091738, + 0.003861376317217946, + 0.003729363903403282 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 2 + ], + "selected_reward": -0.42890530824661255 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a macro close-up taken from a slightly overhead perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera remaining still throughout the video?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are a pair of hands visible, one wearing a non-slip black glove and the other bare?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a fine carving knife with a cold metallic sheen being used to carve a crystal-clear block of ice?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is a cool spotlight shining diagonally from above to illuminate the scene on a dark stone workbench?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.009786365553736687 + ], + "opposite_logprob_tokens": [ + -4.634786128997803 + ], + "logprob_reward": -0.009786365553736687, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.009786365553736687, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.5906320810317993, + 0.0856073796749115, + 0.1402607411146164, + 0.02556474693119526, + 0.037400856614112854, + 0.04313154146075249, + 0.010005941614508629, + 0.036512505263090134, + 0.013304628431797028, + 0.017579620704054832 + ], + "credit_map": [ + 0.0904734879732132, + 0.07088255137205124, + 0.08331171423196793, + 0.05553990229964256, + 0.07676207274198532, + 0.05767407640814781, + 0.06571496278047562, + 0.09027331322431564, + 0.009713826701045036, + 0.009971892461180687, + 0.009410734288394451, + 0.011720601469278336, + 0.011359632946550846, + 0.009272170253098011, + 0.012104334309697151, + 0.012054177932441235, + 0.018230104818940163, + 0.014105835929512978, + 0.02172846347093582, + 0.0177770983427763, + 0.021854516118764877, + 0.013256850652396679, + 0.013243011198937893, + 0.020064856857061386, + 0.002547329990193248, + 0.0035838778130710125, + 0.0027622897177934647, + 0.001758188591338694, + 0.004185294732451439, + 0.004108414985239506, + 0.002569406758993864, + 0.004049943294376135, + 0.0036799961235374212, + 0.0032204394228756428, + 0.005161887966096401, + 0.004102211445569992, + 0.004331852309405804, + 0.006722182501107454, + 0.0052978419698774815, + 0.004884445574134588, + 0.004596516489982605, + 0.0031381503213196993, + 0.006636798847466707, + 0.003858445677906275, + 0.0054946113377809525, + 0.00796600803732872, + 0.004121122881770134, + 0.007319886237382889, + 0.001061272923834622, + 0.0007801085594110191, + 0.00132109469268471, + 0.0008121265564113855, + 0.0017813658341765404, + 0.0013122863601893187, + 0.001104464870877564, + 0.0018332204781472683, + 0.003465881571173668, + 0.0029048966243863106, + 0.004592885263264179, + 0.004241313319653273, + 0.006462136749178171, + 0.004059987608343363, + 0.004669004585593939, + 0.006116400472819805, + 0.0010525730904191732, + 0.0012266343692317605, + 0.0017849630676209927, + 0.0011100675910711288, + 0.0020123303402215242, + 0.001283812103793025, + 0.001466744695790112, + 0.0033675022423267365, + 0.002496209228411317, + 0.001693975878879428, + 0.002686095191165805, + 0.0020757904276251793, + 0.0018860279815271497, + 0.0024694749154150486, + 0.0020497231744229794, + 0.002222321927547455 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 1 + ], + "selected_reward": -0.009786365553736687 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a macro close-up taken from a slightly overhead perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera remaining still throughout the video?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are a pair of hands visible, one wearing a non-slip black glove and the other bare?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a fine carving knife with a cold metallic sheen being used to carve a crystal-clear block of ice?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is a cool spotlight shining diagonally from above to illuminate the scene on a dark stone workbench?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.008639222010970116 + ], + "opposite_logprob_tokens": [ + -4.758639335632324 + ], + "logprob_reward": -0.008639222010970116, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.008639222010970116, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.19666726887226105, + 0.2534623444080353, + 0.05089617893099785, + 0.059828981757164, + 0.05735119804739952, + 0.10815231502056122, + 0.03051670454442501, + 0.14691583812236786, + 0.04738461598753929, + 0.048824578523635864 + ], + "credit_map": [ + 0.019632931798696518, + 0.028070364147424698, + 0.028081180527806282, + 0.019863028079271317, + 0.02755039557814598, + 0.023826129734516144, + 0.026454316452145576, + 0.02318892627954483, + 0.031969934701919556, + 0.029347848147153854, + 0.028143499046564102, + 0.030120274052023888, + 0.029025327414274216, + 0.035560473799705505, + 0.0349152572453022, + 0.0343797504901886, + 0.0050308979116380215, + 0.004861417226493359, + 0.006289280951023102, + 0.0034734371583908796, + 0.012279441580176353, + 0.006490964908152819, + 0.005874243099242449, + 0.006596494000405073, + 0.004633256699889898, + 0.006746446713805199, + 0.006355669349431992, + 0.005327635444700718, + 0.00848700013011694, + 0.011847204528748989, + 0.008856646716594696, + 0.007575123105198145, + 0.004784060176461935, + 0.0057457806542515755, + 0.007531084585934877, + 0.006459256634116173, + 0.008347267284989357, + 0.006028882693499327, + 0.009027209132909775, + 0.009427654556930065, + 0.011415906250476837, + 0.009281828999519348, + 0.015596776269376278, + 0.012914965860545635, + 0.010625740513205528, + 0.014640877954661846, + 0.01154913380742073, + 0.022127090021967888, + 0.003324314719066024, + 0.0031343952286988497, + 0.004564288072288036, + 0.0029431984294205904, + 0.0040842099115252495, + 0.003183778142556548, + 0.004513690713793039, + 0.004768830724060535, + 0.020706936717033386, + 0.012604606337845325, + 0.023072099313139915, + 0.02230643481016159, + 0.01411164179444313, + 0.016186578199267387, + 0.019520822912454605, + 0.018406713381409645, + 0.004242769908159971, + 0.004870851989835501, + 0.006465720944106579, + 0.00491213146597147, + 0.006211050786077976, + 0.00523778609931469, + 0.005658987443894148, + 0.009785319678485394, + 0.00555817037820816, + 0.0051778508350253105, + 0.006495203357189894, + 0.006086614448577166, + 0.006129317916929722, + 0.005240207072347403, + 0.00819164328277111, + 0.005945573095232248 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 7 + ], + "selected_reward": -0.008639222010970116 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a macro close-up taken from a slightly overhead perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera remaining still throughout the video?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are a pair of hands visible, one wearing a non-slip black glove and the other bare?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a fine carving knife with a cold metallic sheen being used to carve a crystal-clear block of ice?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is a cool spotlight shining diagonally from above to illuminate the scene on a dark stone workbench?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.026343408972024918 + ], + "opposite_logprob_tokens": [ + -3.65134334564209 + ], + "logprob_reward": -0.026343408972024918, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.026343408972024918, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.19747872650623322, + 0.19184619188308716, + 0.055802926421165466, + 0.06498350948095322, + 0.05183704197406769, + 0.2622286379337311, + 0.02751535177230835, + 0.08794626593589783, + 0.034429240971803665, + 0.025932081043720245 + ], + "credit_map": [ + 0.019340137019753456, + 0.018256835639476776, + 0.021758297458291054, + 0.015615701675415039, + 0.046736445277929306, + 0.023029685020446777, + 0.030874375253915787, + 0.02186725102365017, + 0.0200413279235363, + 0.019759219139814377, + 0.021970897912979126, + 0.02315507084131241, + 0.02446097694337368, + 0.023551490157842636, + 0.02807770110666752, + 0.030829502269625664, + 0.009976455941796303, + 0.006298452615737915, + 0.008423784747719765, + 0.005453594960272312, + 0.00853572878986597, + 0.0047239153645932674, + 0.005620834417641163, + 0.006770157255232334, + 0.007575581315904856, + 0.0070418743416666985, + 0.007160958368331194, + 0.00880090519785881, + 0.009007720276713371, + 0.007493798620998859, + 0.009616531431674957, + 0.00828613806515932, + 0.0059021287597715855, + 0.005285970866680145, + 0.006976760923862457, + 0.004341414198279381, + 0.007624887861311436, + 0.008964912965893745, + 0.006122502498328686, + 0.006618461571633816, + 0.04343845322728157, + 0.019437232986092567, + 0.047943588346242905, + 0.021904049441218376, + 0.023317405954003334, + 0.030260073021054268, + 0.022991664707660675, + 0.052936166524887085, + 0.003007713006809354, + 0.00309378863312304, + 0.003349795239046216, + 0.0032362155616283417, + 0.003712516510859132, + 0.003240875666961074, + 0.0037920966278761625, + 0.004082350991666317, + 0.012646296992897987, + 0.01120337937027216, + 0.012120373547077179, + 0.011204231530427933, + 0.009057614021003246, + 0.01120795402675867, + 0.010845422744750977, + 0.009660995565354824, + 0.003966121003031731, + 0.002486888552084565, + 0.003822039347141981, + 0.003132895566523075, + 0.005658653099089861, + 0.0036251223646104336, + 0.004740040749311447, + 0.0069974809885025024, + 0.0026348959654569626, + 0.002738382201641798, + 0.003262758255004883, + 0.003518741112202406, + 0.0033198941964656115, + 0.002714209258556366, + 0.004413221962749958, + 0.0033299787901341915 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 0, + 1 + ], + "selected_reward": -0.026343408972024918 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "macro close-up shot, taken from a slightly overhead perspective, the camera remains still, focusing the viewer's attention squarely on a pair of steady and deft hands meticulously carving an ice sculpture. A cool spotlight shines diagonally from above, illuminating the subject on a dark stone workbench: a crystal-clear, angular block of ice. One hand, clad in a non-slip black glove, firmly grasps the edge of the block, while the other, bare, with intense concentration, steadily and calmly manipulates a fine carving knife that gleams with a cold metallic sheen. The tip of the knife gently and precisely glides across the ice. With subtle movements of the wrist, the intricate and delicate texture of the forming ice flower is gradually outlined, each incision releasing strands of hair-like ice shavings. Light penetrates the pure ice, refracting within it as captivating specks of light. The edges of the newly carved petals capture and reflect the brilliant light, complementing the surrounding fine ice crystals. Against the backdrop of a dark background, the entire picture presents an extreme sense of tranquility and concentration, full of awe for exquisite craftsmanship and the beauty of the moment.", + "seed": 7379, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000006/group_14/index_7755_seed_7379_dp_29_rank_58_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 14, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.11736297179013491, + "vlm_logprob": -0.11736297179013491, + "vlm_yes_no_margin": 4.1, + "vlm_token_credit": -0.11736297179013491, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a macro close-up taken from a slightly overhead perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera remaining still throughout the video?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are a pair of hands visible, one wearing a non-slip black glove and the other bare?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a fine carving knife with a cold metallic sheen being used to carve a crystal-clear block of ice?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is a cool spotlight shining diagonally from above to illuminate the scene on a dark stone workbench?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.021901875734329224 + ], + "opposite_logprob_tokens": [ + -6.646902084350586 + ], + "logprob_reward": -0.021901875734329224, + "yes_no_margin_reward": 6.625, + "token_credit_reward": -0.021901875734329224, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08925826102495193, + 0.05400592088699341, + 0.02643509767949581, + 0.03339826688170433, + 0.05604810267686844, + 0.03981028497219086, + 0.2235894799232483, + 0.10020073503255844, + 0.2975423336029053, + 0.07971148937940598 + ], + "credit_map": [ + 0.016553249210119247, + 0.008600546047091484, + 0.011744382791221142, + 0.0122580761089921, + 0.010743898339569569, + 0.009817062877118587, + 0.013070918619632721, + 0.006470132619142532, + 0.006345309317111969, + 0.004901635926216841, + 0.006995247211307287, + 0.007453520316630602, + 0.008041172288358212, + 0.009854632429778576, + 0.006014639977365732, + 0.0043997690081596375, + 0.003429257543757558, + 0.002214826876297593, + 0.003958382643759251, + 0.0035353631246834993, + 0.004090036265552044, + 0.0036594385746866465, + 0.003065375378355384, + 0.002482421463355422, + 0.004117798060178757, + 0.003212201874703169, + 0.004833960440009832, + 0.0033905941527336836, + 0.005140179302543402, + 0.0055923787876963615, + 0.003164524445310235, + 0.003946633078157902, + 0.012429348193109035, + 0.007528271526098251, + 0.007205743808299303, + 0.0049675325863063335, + 0.006734129041433334, + 0.007249895483255386, + 0.004245894495397806, + 0.005687290802598, + 0.005747141782194376, + 0.005369184073060751, + 0.006274787709116936, + 0.004585720598697662, + 0.005275382194668055, + 0.005333585198968649, + 0.00407315231859684, + 0.0031513329595327377, + 0.07141043990850449, + 0.08395007252693176, + 0.018548697233200073, + 0.014636097475886345, + 0.008931763470172882, + 0.01004562247544527, + 0.008762512356042862, + 0.007304301485419273, + 0.013753353618085384, + 0.0323464535176754, + 0.024032484740018845, + 0.008914175443351269, + 0.0039291828870773315, + 0.004545675590634346, + 0.006994438357651234, + 0.005684978328645229, + 0.04103109985589981, + 0.09505245089530945, + 0.06527125090360641, + 0.019010746851563454, + 0.022679250687360764, + 0.014619694091379642, + 0.021444404497742653, + 0.018433446064591408, + 0.0184575617313385, + 0.012098713777959347, + 0.008057148195803165, + 0.008588162250816822, + 0.009591197595000267, + 0.010850618593394756, + 0.006368190050125122, + 0.005699903704226017 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 6, + 7 + ], + "selected_reward": -0.021901875734329224 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a macro close-up taken from a slightly overhead perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera remaining still throughout the video?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are a pair of hands visible, one wearing a non-slip black glove and the other bare?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a fine carving knife with a cold metallic sheen being used to carve a crystal-clear block of ice?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is a cool spotlight shining diagonally from above to illuminate the scene on a dark stone workbench?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.52333664894104 + ], + "opposite_logprob_tokens": [ + -0.89833664894104 + ], + "logprob_reward": -0.52333664894104, + "yes_no_margin_reward": 0.375, + "token_credit_reward": -0.52333664894104, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12786448001861572, + 0.08787492662668228, + 0.044122304767370224, + 0.05029584467411041, + 0.05727614462375641, + 0.022422119975090027, + 0.20317122340202332, + 0.03726322203874588, + 0.33120039105415344, + 0.038509368896484375 + ], + "credit_map": [ + 0.01567295379936695, + 0.00637610349804163, + 0.015758521854877472, + 0.006331863347440958, + 0.007446633651852608, + 0.04073308780789375, + 0.026734840124845505, + 0.008810479193925858, + 0.019863341003656387, + 0.006025162059813738, + 0.012805339880287647, + 0.005655136425048113, + 0.008753043599426746, + 0.0216992124915123, + 0.007498544175177813, + 0.0055751497857272625, + 0.006145903374999762, + 0.0037829417269676924, + 0.0068878913298249245, + 0.005530890077352524, + 0.006453393492847681, + 0.0060810805298388, + 0.005231182556599379, + 0.00400901772081852, + 0.005284195765852928, + 0.005573068745434284, + 0.006179842632263899, + 0.004238574765622616, + 0.010064451023936272, + 0.009900677017867565, + 0.00399735989049077, + 0.00505767622962594, + 0.012447324581444263, + 0.009198659099638462, + 0.00880955159664154, + 0.00492130359634757, + 0.004672906827181578, + 0.008741393685340881, + 0.0045912545174360275, + 0.0038937493227422237, + 0.0038453591987490654, + 0.0024039107374846935, + 0.004023269284516573, + 0.002304626163095236, + 0.0019956042524427176, + 0.003242318518459797, + 0.0028033102862536907, + 0.0018037197878584266, + 0.06449637562036514, + 0.046164195984601974, + 0.04495018720626831, + 0.01107854675501585, + 0.005594663321971893, + 0.011564108543097973, + 0.0067199161276221275, + 0.012603215873241425, + 0.004349568858742714, + 0.006857750006020069, + 0.010574339888989925, + 0.0030496600084006786, + 0.0023678604047745466, + 0.004233624320477247, + 0.0027007809840142727, + 0.003129637334495783, + 0.030106017366051674, + 0.10555315762758255, + 0.10142864286899567, + 0.01592695713043213, + 0.019680501893162727, + 0.016706325113773346, + 0.020362121984362602, + 0.021436655893921852, + 0.008950475603342056, + 0.005767961032688618, + 0.004562846850603819, + 0.003942057956010103, + 0.0037624959368258715, + 0.004142240155488253, + 0.004191348794847727, + 0.003189945127815008 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 6, + 0 + ], + "selected_reward": -0.52333664894104 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a macro close-up taken from a slightly overhead perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera remaining still throughout the video?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are a pair of hands visible, one wearing a non-slip black glove and the other bare?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a fine carving knife with a cold metallic sheen being used to carve a crystal-clear block of ice?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is a cool spotlight shining diagonally from above to illuminate the scene on a dark stone workbench?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.014200235716998577 + ], + "opposite_logprob_tokens": [ + -4.264200210571289 + ], + "logprob_reward": -0.014200235716998577, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.014200235716998577, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3118785619735718, + 0.17883086204528809, + 0.054859377443790436, + 0.06464875489473343, + 0.03479485213756561, + 0.02692534774541855, + 0.08544839173555374, + 0.06671703606843948, + 0.10238287597894669, + 0.07351388782262802 + ], + "credit_map": [ + 0.03082951530814171, + 0.016723433509469032, + 0.0434202142059803, + 0.02051982283592224, + 0.018740247935056686, + 0.09229131042957306, + 0.06868898123502731, + 0.020665066316723824, + 0.026431746780872345, + 0.015395822934806347, + 0.02866744063794613, + 0.015005983412265778, + 0.013302173465490341, + 0.04360851272940636, + 0.024171382188796997, + 0.01224780734628439, + 0.00638593640178442, + 0.004735076799988747, + 0.00784546509385109, + 0.007746831979602575, + 0.008777495473623276, + 0.0066438037902116776, + 0.007389682810753584, + 0.005335086025297642, + 0.0079558826982975, + 0.006697025150060654, + 0.009423320181667805, + 0.006773326080292463, + 0.009998587891459465, + 0.011275339871644974, + 0.005172920413315296, + 0.007352360524237156, + 0.0059836809523403645, + 0.005531644448637962, + 0.005085257813334465, + 0.0030292822048068047, + 0.003725755726918578, + 0.004729040432721376, + 0.004096959717571735, + 0.0026132299099117517, + 0.0031158241908997297, + 0.0027297951746731997, + 0.004873800557106733, + 0.002967711305245757, + 0.003935294225811958, + 0.0030998217407613993, + 0.0027852372732013464, + 0.0034178660716861486, + 0.026807278394699097, + 0.015906961634755135, + 0.022987989708781242, + 0.004011403303593397, + 0.0019323565065860748, + 0.005382230505347252, + 0.0023206232581287622, + 0.006099547725170851, + 0.006905617192387581, + 0.016145536676049232, + 0.013777046464383602, + 0.010825151577591896, + 0.004055550321936607, + 0.0043145776726305485, + 0.006666623987257481, + 0.0040269335731863976, + 0.0186900831758976, + 0.026254158467054367, + 0.015513130463659763, + 0.006713224574923515, + 0.013680914416909218, + 0.008226146921515465, + 0.007231638301163912, + 0.0060735843144357204, + 0.02228713594377041, + 0.012306892313063145, + 0.010789165273308754, + 0.005653140600770712, + 0.005216224584728479, + 0.0072407289408147335, + 0.005236921366304159, + 0.004783675540238619 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 8 + ], + "selected_reward": -0.014200235716998577 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a macro close-up taken from a slightly overhead perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera remaining still throughout the video?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are a pair of hands visible, one wearing a non-slip black glove and the other bare?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a fine carving knife with a cold metallic sheen being used to carve a crystal-clear block of ice?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is a cool spotlight shining diagonally from above to illuminate the scene on a dark stone workbench?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004094313830137253 + ], + "opposite_logprob_tokens": [ + -5.504094123840332 + ], + "logprob_reward": -0.004094313830137253, + "yes_no_margin_reward": 5.5, + "token_credit_reward": -0.004094313830137253, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.21820783615112305, + 0.06942930072546005, + 0.1116989478468895, + 0.022373976185917854, + 0.02792244404554367, + 0.056861940771341324, + 0.33186012506484985, + 0.04741450026631355, + 0.06779653578996658, + 0.04643431678414345 + ], + "credit_map": [ + 0.03438660874962807, + 0.016770176589488983, + 0.025168893858790398, + 0.01743433065712452, + 0.020001372322440147, + 0.0535457618534565, + 0.03556295484304428, + 0.01533778291195631, + 0.01979798637330532, + 0.00789428036659956, + 0.008289283141493797, + 0.00607762997969985, + 0.00598559295758605, + 0.012537707574665546, + 0.003753853961825371, + 0.005092974752187729, + 0.014057006686925888, + 0.015334556810557842, + 0.012848573736846447, + 0.015686269849538803, + 0.01421325746923685, + 0.014948444440960884, + 0.014016473665833473, + 0.010594384744763374, + 0.0030660810880362988, + 0.003552231937646866, + 0.0028182510286569595, + 0.0028184030670672655, + 0.002309186616912484, + 0.0032997720409184694, + 0.0019759235437959433, + 0.00253413082100451, + 0.004038968123495579, + 0.004365366417914629, + 0.0038658841513097286, + 0.003743640147149563, + 0.0033559182193130255, + 0.003163248533383012, + 0.0028071701526641846, + 0.0025822531897574663, + 0.010317724198102951, + 0.00887157954275608, + 0.008714823052287102, + 0.005746366456151009, + 0.007303282152861357, + 0.006654614116996527, + 0.004589586053043604, + 0.0046639759093523026, + 0.12776489555835724, + 0.10531271994113922, + 0.03836621344089508, + 0.017728807404637337, + 0.010377207770943642, + 0.011770101264119148, + 0.008590888231992722, + 0.0119493268430233, + 0.0065128859132528305, + 0.011596497148275375, + 0.006835960783064365, + 0.008532673120498657, + 0.0029123870190232992, + 0.0030707749538123608, + 0.003971312195062637, + 0.003982021473348141, + 0.016801627352833748, + 0.01073097437620163, + 0.008509857580065727, + 0.005870141088962555, + 0.008698323741555214, + 0.006543854251503944, + 0.005869519896805286, + 0.004772253334522247, + 0.014301854185760021, + 0.006464524660259485, + 0.00700777443125844, + 0.005125564988702536, + 0.0032490568701177835, + 0.003912623506039381, + 0.0032695268746465445, + 0.003103395691141486 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 0, + 2 + ], + "selected_reward": -0.004094313830137253 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a macro close-up taken from a slightly overhead perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera remaining still throughout the video?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are a pair of hands visible, one wearing a non-slip black glove and the other bare?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a fine carving knife with a cold metallic sheen being used to carve a crystal-clear block of ice?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is a cool spotlight shining diagonally from above to illuminate the scene on a dark stone workbench?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02328178472816944 + ], + "opposite_logprob_tokens": [ + -3.7732818126678467 + ], + "logprob_reward": -0.02328178472816944, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.02328178472816944, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.38567760586738586, + 0.039645854383707047, + 0.04384298995137215, + 0.02505810745060444, + 0.028639856725931168, + 0.019568778574466705, + 0.05507627874612808, + 0.05280108377337456, + 0.3219045400619507, + 0.02778482809662819 + ], + "credit_map": [ + 0.05280690640211105, + 0.025452695786952972, + 0.044908557087183, + 0.021782221272587776, + 0.022182434797286987, + 0.11032530665397644, + 0.08463750779628754, + 0.023581985384225845, + 0.005004871636629105, + 0.0030044792219996452, + 0.007401838432997465, + 0.003017100039869547, + 0.0050948625430464745, + 0.009012360125780106, + 0.00383129995316267, + 0.0032790463883429766, + 0.006967129185795784, + 0.004821550101041794, + 0.006305756513029337, + 0.004658387042582035, + 0.00627651484683156, + 0.005295221693813801, + 0.004612169694155455, + 0.004906264133751392, + 0.0034520162735134363, + 0.004628330934792757, + 0.002934487769380212, + 0.002423779107630253, + 0.0035923488903790712, + 0.003955505322664976, + 0.002002475783228874, + 0.0020691624376922846, + 0.004711092449724674, + 0.006354943849146366, + 0.003897372167557478, + 0.002816425869241357, + 0.00304474332369864, + 0.0035380199551582336, + 0.0026372249703854322, + 0.0016400344902649522, + 0.003696745727211237, + 0.003680081106722355, + 0.00267790793441236, + 0.0016577907372266054, + 0.0021705967374145985, + 0.00226336601190269, + 0.0023437938652932644, + 0.001078498549759388, + 0.020902255550026894, + 0.00980214774608612, + 0.014178644865751266, + 0.002021011896431446, + 0.0012292590690776706, + 0.002777750138193369, + 0.0011904507409781218, + 0.0029747614171355963, + 0.005667501594871283, + 0.013905694708228111, + 0.015242584981024265, + 0.005848067346960306, + 0.002432081615552306, + 0.0033774948678910732, + 0.0027862258721143007, + 0.003541436279192567, + 0.043383270502090454, + 0.09239199757575989, + 0.0739365965127945, + 0.017918845638632774, + 0.03402397036552429, + 0.021292442455887794, + 0.01991110108792782, + 0.019046366214752197, + 0.007294076029211283, + 0.00426453398540616, + 0.0037635478656738997, + 0.0021818457171320915, + 0.00277676316909492, + 0.0026767943054437637, + 0.0024498419370502234, + 0.0023774292785674334 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 6 + ], + "selected_reward": -0.02328178472816944 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000006/group_15/index_tensor([4287, 4287])_group_15_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000006/group_15/index_tensor([4287, 4287])_group_15_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..4a2612443eb01718faf417ba258998133f4d2b69 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000006/group_15/index_tensor([4287, 4287])_group_15_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "static, eye-level long shot captures a gritty urban streetscape, bathed in a desaturated palette, contrasted by deep shadows and the harsh glare of streetlights on the wet sidewalk. A grimy city sidewalk, its cracked and uneven concrete slabs, stretches into the distance, bordered by a rain-soaked asphalt street that reflects a cold, artificial light. In the background, the muted reds and browns of a graffiti-tagged brick building fade into darkness. From the right side of the frame, a rider on a sleek, one-wheeled vehicle glides effortlessly and silently into the scene. The vehicle's matte black chassis absorbs light, while a thin cyan LED strip on the side provides a vibrant color accent. The unnamed rider, wearing a dark hooded waterproof jacket and slim-fitting overalls, leans forward, maintaining a focused balance. Swiftly and decisively, they pass a motionless pedestrian, hunched on the sidewalk, wearing a heavy, unremarkable coat, seemingly oblivious to the futuristic presence. The wheels sent up a fine, almost invisible spray from the wet ground, a subtle disturbance in the quiet, oppressive atmosphere of the city's anonymity.", + "seed": 30258, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000006/group_15/index_4287_seed_30258_dp_30_rank_60_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 15, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.020381482411175968, + "vlm_logprob": -0.020381482411175968, + "vlm_yes_no_margin": 5.1, + "vlm_token_credit": -0.020381482411175968, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a gritty urban streetscape with a cracked concrete sidewalk and a rain-soaked asphalt street?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the color palette desaturated with deep shadows and harsh streetlight glare on the wet ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a rider on a sleek, matte black one-wheeled vehicle with a thin cyan LED strip visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the rider wearing a dark hooded waterproof jacket and slim-fitting overalls while leaning forward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the rider pass a motionless pedestrian hunched on the sidewalk wearing a heavy coat?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.009970740415155888 + ], + "opposite_logprob_tokens": [ + -6.634970664978027 + ], + "logprob_reward": -0.009970740415155888, + "yes_no_margin_reward": 6.625, + "token_credit_reward": -0.009970740415155888, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.22216947376728058, + 0.04727368801832199, + 0.06012257561087608, + 0.04204412177205086, + 0.05218514800071716, + 0.07934612780809402, + 0.063148632645607, + 0.15228413045406342, + 0.13295021653175354, + 0.14847591519355774 + ], + "credit_map": [ + 0.020269040018320084, + 0.02004598267376423, + 0.021139247342944145, + 0.010549758560955524, + 0.01943373493850231, + 0.06203681603074074, + 0.017628582194447517, + 0.051066286861896515, + 0.007081381976604462, + 0.004457332659512758, + 0.0070688603445887566, + 0.002295115264132619, + 0.009022236801683903, + 0.007358848117291927, + 0.0038827103562653065, + 0.0061071948148310184, + 0.006781503092497587, + 0.007557586766779423, + 0.006201944779604673, + 0.003517706645652652, + 0.012662747874855995, + 0.008862406946718693, + 0.007261879742145538, + 0.00727679580450058, + 0.006035751663148403, + 0.005386028904467821, + 0.0022051320411264896, + 0.001634856453165412, + 0.013697410002350807, + 0.006417908705770969, + 0.0026458296924829483, + 0.004021200817078352, + 0.005171742290258408, + 0.008061778731644154, + 0.003679670626297593, + 0.005296379327774048, + 0.008386179804801941, + 0.007184509187936783, + 0.0064549739472568035, + 0.007949908263981342, + 0.009211522527039051, + 0.012120967730879784, + 0.004899553954601288, + 0.006272421684116125, + 0.010946554131805897, + 0.011271846480667591, + 0.010415295138955116, + 0.014207958243787289, + 0.007286216598004103, + 0.008289976045489311, + 0.006524926517158747, + 0.007047349587082863, + 0.01036288496106863, + 0.00835222564637661, + 0.0072812410071492195, + 0.008003802970051765, + 0.019048510119318962, + 0.04259160906076431, + 0.010566187091171741, + 0.01290108822286129, + 0.01595529541373253, + 0.028312930837273598, + 0.012571710161864758, + 0.010336784645915031, + 0.030378282070159912, + 0.011565486900508404, + 0.006204779259860516, + 0.008677465841174126, + 0.018945904448628426, + 0.04544026404619217, + 0.004790265578776598, + 0.006947755813598633, + 0.014062894508242607, + 0.01711176335811615, + 0.017395539209246635, + 0.013541272841393948, + 0.029483269900083542, + 0.028248945251107216, + 0.013535264879465103, + 0.015096964314579964 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 7, + 9 + ], + "selected_reward": -0.009970740415155888 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a gritty urban streetscape with a cracked concrete sidewalk and a rain-soaked asphalt street?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the color palette desaturated with deep shadows and harsh streetlight glare on the wet ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a rider on a sleek, matte black one-wheeled vehicle with a thin cyan LED strip visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the rider wearing a dark hooded waterproof jacket and slim-fitting overalls while leaning forward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the rider pass a motionless pedestrian hunched on the sidewalk wearing a heavy coat?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.003626557532697916 + ], + "opposite_logprob_tokens": [ + -5.628626346588135 + ], + "logprob_reward": -0.003626557532697916, + "yes_no_margin_reward": 5.625, + "token_credit_reward": -0.003626557532697916, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1883964240550995, + 0.04489723965525627, + 0.04015318304300308, + 0.12230366468429565, + 0.04336579516530037, + 0.1085343137383461, + 0.07122272253036499, + 0.10175542533397675, + 0.2171410471200943, + 0.06223028898239136 + ], + "credit_map": [ + 0.018533440306782722, + 0.02599574252963066, + 0.02188527025282383, + 0.010094455443322659, + 0.0138395419344306, + 0.054051633924245834, + 0.010590611025691032, + 0.03340570628643036, + 0.004544879775494337, + 0.005351611413061619, + 0.00664858240634203, + 0.0027441498823463917, + 0.01099751889705658, + 0.00615519005805254, + 0.0030737826600670815, + 0.005381520837545395, + 0.005047470331192017, + 0.004945931024849415, + 0.004656176548451185, + 0.0026634682435542345, + 0.005643737036734819, + 0.005767913069576025, + 0.005195715930312872, + 0.006232768762856722, + 0.019833257421851158, + 0.013543478213250637, + 0.008408410474658012, + 0.008630575612187386, + 0.03129763528704643, + 0.018298886716365814, + 0.011021057143807411, + 0.011270353570580482, + 0.00411507673561573, + 0.008200998418033123, + 0.0025679771788418293, + 0.0037131879944354296, + 0.009066501632332802, + 0.006251487880945206, + 0.004024812951683998, + 0.005425749812275171, + 0.012506578117609024, + 0.014057660475373268, + 0.007321607321500778, + 0.009883665479719639, + 0.015079227276146412, + 0.01719077117741108, + 0.01536435168236494, + 0.017130441963672638, + 0.008680741302669048, + 0.010759306140244007, + 0.006212907377630472, + 0.008161020465195179, + 0.010046297684311867, + 0.010133926756680012, + 0.007305788807570934, + 0.009922727942466736, + 0.010929932817816734, + 0.016244443133473396, + 0.009638383984565735, + 0.01066234428435564, + 0.01125249732285738, + 0.018611926585435867, + 0.013572612777352333, + 0.010843278840184212, + 0.02727750688791275, + 0.023548537865281105, + 0.02412501722574234, + 0.059067606925964355, + 0.02151235193014145, + 0.026708083227276802, + 0.02027823217213154, + 0.014623706229031086, + 0.006107651628553867, + 0.007607219275087118, + 0.006845144554972649, + 0.005409310106188059, + 0.0140924621373415, + 0.011530406773090363, + 0.004931946285068989, + 0.005706148687750101 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 0, + 3 + ], + "selected_reward": -0.003626557532697916 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a gritty urban streetscape with a cracked concrete sidewalk and a rain-soaked asphalt street?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the color palette desaturated with deep shadows and harsh streetlight glare on the wet ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a rider on a sleek, matte black one-wheeled vehicle with a thin cyan LED strip visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the rider wearing a dark hooded waterproof jacket and slim-fitting overalls while leaning forward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the rider pass a motionless pedestrian hunched on the sidewalk wearing a heavy coat?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004101555794477463 + ], + "opposite_logprob_tokens": [ + -5.504101753234863 + ], + "logprob_reward": -0.004101555794477463, + "yes_no_margin_reward": 5.5, + "token_credit_reward": -0.004101555794477463, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08215884864330292, + 0.031300485134124756, + 0.02747274748980999, + 0.022558337077498436, + 0.10323523730039597, + 0.11746438592672348, + 0.03552715852856636, + 0.029647260904312134, + 0.48855358362197876, + 0.0620819628238678 + ], + "credit_map": [ + 0.006306928116828203, + 0.010808916762471199, + 0.00909834261983633, + 0.004026324488222599, + 0.007147886790335178, + 0.024420659989118576, + 0.004724118392914534, + 0.015625664964318275, + 0.0032786934170871973, + 0.0036593445111066103, + 0.0052873603999614716, + 0.001521620200946927, + 0.006616543047130108, + 0.004377515986561775, + 0.0023918692022562027, + 0.004167537204921246, + 0.0033508706837892532, + 0.004052332602441311, + 0.002956623677164316, + 0.0014033566694706678, + 0.005466932896524668, + 0.00401423079892993, + 0.00305757368914783, + 0.003170828567817807, + 0.002842299174517393, + 0.0029721898026764393, + 0.0012325904099270701, + 0.0012957514263689518, + 0.006444993428885937, + 0.0035457992926239967, + 0.0021466079633682966, + 0.0020781042985618114, + 0.00963833462446928, + 0.018729573115706444, + 0.006806616671383381, + 0.007862533442676067, + 0.0200651828199625, + 0.014780846424400806, + 0.010982322506606579, + 0.014369831420481205, + 0.012206337414681911, + 0.020126378163695335, + 0.006673627998679876, + 0.011040866374969482, + 0.013168953359127045, + 0.017519522458314896, + 0.019153276458382607, + 0.017575418576598167, + 0.004821170121431351, + 0.005640017334371805, + 0.002881548833101988, + 0.004390562418848276, + 0.004678559955209494, + 0.004497864283621311, + 0.0035419859923422337, + 0.005075447726994753, + 0.0026477964129298925, + 0.005610727239400148, + 0.002543139737099409, + 0.003407195443287492, + 0.004000414162874222, + 0.004133608657866716, + 0.003829837078228593, + 0.0034745438024401665, + 0.13790318369865417, + 0.03967875987291336, + 0.018586233258247375, + 0.011481142602860928, + 0.04673964902758598, + 0.20188896358013153, + 0.01234707236289978, + 0.019928565248847008, + 0.006413247436285019, + 0.005768124014139175, + 0.006681022234261036, + 0.005148034542798996, + 0.01730877161026001, + 0.006996871903538704, + 0.0061603873036801815, + 0.007605500984936953 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 5, + 4 + ], + "selected_reward": -0.004101555794477463 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a gritty urban streetscape with a cracked concrete sidewalk and a rain-soaked asphalt street?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the color palette desaturated with deep shadows and harsh streetlight glare on the wet ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a rider on a sleek, matte black one-wheeled vehicle with a thin cyan LED strip visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the rider wearing a dark hooded waterproof jacket and slim-fitting overalls while leaning forward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the rider pass a motionless pedestrian hunched on the sidewalk wearing a heavy coat?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.00525094848126173 + ], + "opposite_logprob_tokens": [ + -5.255250930786133 + ], + "logprob_reward": -0.00525094848126173, + "yes_no_margin_reward": 5.25, + "token_credit_reward": -0.00525094848126173, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08199181407690048, + 0.06688451766967773, + 0.023517325520515442, + 0.040468018501996994, + 0.09613479673862457, + 0.07266144454479218, + 0.0861545205116272, + 0.08341939747333527, + 0.3639602065086365, + 0.08480799943208694 + ], + "credit_map": [ + 0.006027023773640394, + 0.008558424189686775, + 0.010133219882845879, + 0.00463440315797925, + 0.009243910200893879, + 0.022756217047572136, + 0.004813574254512787, + 0.01582503505051136, + 0.00529311690479517, + 0.004952611867338419, + 0.006944490130990744, + 0.002975436393171549, + 0.013509811833500862, + 0.010648621246218681, + 0.006922450382262468, + 0.015637967735528946, + 0.0029640384018421173, + 0.004816381260752678, + 0.0018526501953601837, + 0.0011350222630426288, + 0.004416659940034151, + 0.003542873542755842, + 0.002048065420240164, + 0.002741631818935275, + 0.0047362265177071095, + 0.0046955435536801815, + 0.002243201481178403, + 0.0019080373458564281, + 0.012980137951672077, + 0.007371086627244949, + 0.0025320297572761774, + 0.004001753404736519, + 0.009848559275269508, + 0.016357533633708954, + 0.005269576329737902, + 0.007358368486166, + 0.02131529711186886, + 0.014587866142392159, + 0.008355389349162579, + 0.013042198494076729, + 0.007995669730007648, + 0.012948676943778992, + 0.003983396105468273, + 0.006646200083196163, + 0.012903593480587006, + 0.011592455208301544, + 0.007084976881742477, + 0.009506470523774624, + 0.01142964418977499, + 0.013167228549718857, + 0.0069509283639490604, + 0.011979149654507637, + 0.012495260685682297, + 0.010770929977297783, + 0.00863688625395298, + 0.010724482126533985, + 0.007396032568067312, + 0.013534944504499435, + 0.00840393640100956, + 0.009729991666972637, + 0.011833716183900833, + 0.014149205759167671, + 0.01192533690482378, + 0.006446225568652153, + 0.10320281237363815, + 0.029509861022233963, + 0.013652246445417404, + 0.007948816753923893, + 0.03660976514220238, + 0.14910803735256195, + 0.009712585248053074, + 0.014216039329767227, + 0.008361317217350006, + 0.00746088707819581, + 0.009127859957516193, + 0.006871911697089672, + 0.023875419050455093, + 0.011359155178070068, + 0.006961209233850241, + 0.010790232568979263 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 4, + 6 + ], + "selected_reward": -0.00525094848126173 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a gritty urban streetscape with a cracked concrete sidewalk and a rain-soaked asphalt street?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the color palette desaturated with deep shadows and harsh streetlight glare on the wet ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a rider on a sleek, matte black one-wheeled vehicle with a thin cyan LED strip visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the rider wearing a dark hooded waterproof jacket and slim-fitting overalls while leaning forward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the rider pass a motionless pedestrian hunched on the sidewalk wearing a heavy coat?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07895760983228683 + ], + "opposite_logprob_tokens": [ + -2.5789575576782227 + ], + "logprob_reward": -0.07895760983228683, + "yes_no_margin_reward": 2.5, + "token_credit_reward": -0.07895760983228683, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1165064126253128, + 0.08689410239458084, + 0.08674246817827225, + 0.025089122354984283, + 0.061411600559949875, + 0.026595737785100937, + 0.09936664998531342, + 0.07765782624483109, + 0.3264825940132141, + 0.093253493309021 + ], + "credit_map": [ + 0.011892917566001415, + 0.012810572981834412, + 0.012141712009906769, + 0.005855176597833633, + 0.011137735098600388, + 0.03027177043259144, + 0.008151117712259293, + 0.024245407432317734, + 0.008986921049654484, + 0.010443365201354027, + 0.00867233145982027, + 0.003477595280855894, + 0.024630658328533173, + 0.015866046771407127, + 0.0042378767393529415, + 0.010579297319054604, + 0.010350571945309639, + 0.014859705232083797, + 0.006763658486306667, + 0.004145926795899868, + 0.02274566888809204, + 0.013297326862812042, + 0.007523426320403814, + 0.0070561873726546764, + 0.0037490881513804197, + 0.0029971953481435776, + 0.0013606032589450479, + 0.0015023701125755906, + 0.007349039427936077, + 0.003921989817172289, + 0.002045438624918461, + 0.002163397381082177, + 0.00609562685713172, + 0.009688097983598709, + 0.0033514490351080894, + 0.005328291095793247, + 0.014391127042472363, + 0.00969311036169529, + 0.005057189147919416, + 0.007806710433214903, + 0.0028156607877463102, + 0.004153684712946415, + 0.001976113533601165, + 0.003130338154733181, + 0.005326638929545879, + 0.004123365506529808, + 0.0027098366990685463, + 0.002360099460929632, + 0.01036880537867546, + 0.016627706587314606, + 0.006212291773408651, + 0.011822064407169819, + 0.01931784488260746, + 0.014025187119841576, + 0.00983762089163065, + 0.01115512941032648, + 0.007006806321442127, + 0.017127783969044685, + 0.007233361713588238, + 0.011880025267601013, + 0.006576530169695616, + 0.013344811275601387, + 0.005931290332227945, + 0.00855721440166235, + 0.0924987867474556, + 0.02691614255309105, + 0.011964492499828339, + 0.007375657092779875, + 0.03168392926454544, + 0.13390576839447021, + 0.008897561579942703, + 0.013240276835858822, + 0.008424198254942894, + 0.006859524175524712, + 0.007943195290863514, + 0.006365463603287935, + 0.028131945058703423, + 0.013877342455089092, + 0.007284744642674923, + 0.014367073774337769 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 0, + 6 + ], + "selected_reward": -0.07895760983228683 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "static, eye-level long shot captures a gritty urban streetscape, bathed in a desaturated palette, contrasted by deep shadows and the harsh glare of streetlights on the wet sidewalk. A grimy city sidewalk, its cracked and uneven concrete slabs, stretches into the distance, bordered by a rain-soaked asphalt street that reflects a cold, artificial light. In the background, the muted reds and browns of a graffiti-tagged brick building fade into darkness. From the right side of the frame, a rider on a sleek, one-wheeled vehicle glides effortlessly and silently into the scene. The vehicle's matte black chassis absorbs light, while a thin cyan LED strip on the side provides a vibrant color accent. The unnamed rider, wearing a dark hooded waterproof jacket and slim-fitting overalls, leans forward, maintaining a focused balance. Swiftly and decisively, they pass a motionless pedestrian, hunched on the sidewalk, wearing a heavy, unremarkable coat, seemingly oblivious to the futuristic presence. The wheels sent up a fine, almost invisible spray from the wet ground, a subtle disturbance in the quiet, oppressive atmosphere of the city's anonymity.", + "seed": 30259, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000006/group_15/index_4287_seed_30259_dp_30_rank_60_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 15, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.020060433074831963, + "vlm_logprob": -0.020060433074831963, + "vlm_yes_no_margin": 4.5, + "vlm_token_credit": -0.020060433074831963, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a gritty urban streetscape with a cracked concrete sidewalk and a rain-soaked asphalt street?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the color palette desaturated with deep shadows and harsh streetlight glare on the wet ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a rider on a sleek, matte black one-wheeled vehicle with a thin cyan LED strip visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the rider wearing a dark hooded waterproof jacket and slim-fitting overalls while leaning forward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the rider pass a motionless pedestrian hunched on the sidewalk wearing a heavy coat?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.014578958973288536 + ], + "opposite_logprob_tokens": [ + -5.139578819274902 + ], + "logprob_reward": -0.014578958973288536, + "yes_no_margin_reward": 5.125, + "token_credit_reward": -0.014578958973288536, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1902412325143814, + 0.11300206929445267, + 0.03231506794691086, + 0.043593693524599075, + 0.1110677644610405, + 0.0365644134581089, + 0.08442578464746475, + 0.07696660608053207, + 0.08550325781106949, + 0.2263200730085373 + ], + "credit_map": [ + 0.016194220632314682, + 0.020030131563544273, + 0.037371307611465454, + 0.02224278450012207, + 0.031571466475725174, + 0.02412525750696659, + 0.02374555915594101, + 0.014960511587560177, + 0.0061667063273489475, + 0.004456908442080021, + 0.005748784635215998, + 0.024237874895334244, + 0.00618725223466754, + 0.022317836061120033, + 0.03759823739528656, + 0.00628846837207675, + 0.003460264764726162, + 0.0054012201726436615, + 0.0032129990868270397, + 0.0044550164602696896, + 0.004155258648097515, + 0.004023226443678141, + 0.004164543934166431, + 0.0034425384365022182, + 0.0033015268854796886, + 0.007064195815473795, + 0.0038686729967594147, + 0.00531603442505002, + 0.006050160154700279, + 0.005314685869961977, + 0.006949187256395817, + 0.005729231052100658, + 0.010341901332139969, + 0.01753143034875393, + 0.009308910928666592, + 0.022853348404169083, + 0.015713803470134735, + 0.010757800191640854, + 0.011725340038537979, + 0.012835228815674782, + 0.002604154171422124, + 0.0037447733338922262, + 0.004186999052762985, + 0.005655942019075155, + 0.004694943781942129, + 0.005100929643958807, + 0.005134737119078636, + 0.0054419334046542645, + 0.007260574493557215, + 0.007935752160847187, + 0.010801048018038273, + 0.007117554545402527, + 0.009169777855277061, + 0.011282521300017834, + 0.01718512549996376, + 0.013673431240022182, + 0.0054865493439137936, + 0.008097363635897636, + 0.003714828286319971, + 0.0031490649562329054, + 0.026144491508603096, + 0.004587654024362564, + 0.006090885493904352, + 0.019695768132805824, + 0.005525927525013685, + 0.008055552840232849, + 0.006019306369125843, + 0.005280310288071632, + 0.020505191758275032, + 0.017605438828468323, + 0.006940646097064018, + 0.01557089015841484, + 0.022757017984986305, + 0.0188826285302639, + 0.03980027884244919, + 0.01817397214472294, + 0.013528364710509777, + 0.017817486077547073, + 0.032435718923807144, + 0.0629245936870575 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 0, + 1 + ], + "selected_reward": -0.014578958973288536 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a gritty urban streetscape with a cracked concrete sidewalk and a rain-soaked asphalt street?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the color palette desaturated with deep shadows and harsh streetlight glare on the wet ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a rider on a sleek, matte black one-wheeled vehicle with a thin cyan LED strip visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the rider wearing a dark hooded waterproof jacket and slim-fitting overalls while leaning forward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the rider pass a motionless pedestrian hunched on the sidewalk wearing a heavy coat?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.00979462917894125 + ], + "opposite_logprob_tokens": [ + -4.63479471206665 + ], + "logprob_reward": -0.00979462917894125, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.00979462917894125, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11237231642007828, + 0.2746301591396332, + 0.13965030014514923, + 0.052652545273303986, + 0.056343261152505875, + 0.04398215189576149, + 0.055778101086616516, + 0.05163361877202988, + 0.10957181453704834, + 0.10338569432497025 + ], + "credit_map": [ + 0.011725208722054958, + 0.011240477673709393, + 0.023772113025188446, + 0.014629468321800232, + 0.017456259578466415, + 0.00989777036011219, + 0.013176361098885536, + 0.010474668815732002, + 0.014203298836946487, + 0.010673063807189465, + 0.0070012034848332405, + 0.04143338277935982, + 0.018899869173765182, + 0.07649456709623337, + 0.09426646679639816, + 0.011658322997391224, + 0.012623432092368603, + 0.034543294459581375, + 0.014549307525157928, + 0.020394880324602127, + 0.011666995473206043, + 0.013368642888963223, + 0.020636048167943954, + 0.011867705732584, + 0.0023689051158726215, + 0.003655740525573492, + 0.0025787916965782642, + 0.004299421329051256, + 0.004917671903967857, + 0.004731308203190565, + 0.01618693768978119, + 0.0139137739315629, + 0.004905556328594685, + 0.009128150530159473, + 0.004637548699975014, + 0.007684770040214062, + 0.005710024386644363, + 0.006873094476759434, + 0.006889595650136471, + 0.010514527559280396, + 0.0032169134356081486, + 0.005573120433837175, + 0.006485870573669672, + 0.0057594128884375095, + 0.004997586831450462, + 0.005085332784801722, + 0.0072973319329321384, + 0.0055665867403149605, + 0.004573337268084288, + 0.0067055425606667995, + 0.005778399296104908, + 0.005994999781250954, + 0.006554553285241127, + 0.006305298302322626, + 0.01139792986214161, + 0.008468042127788067, + 0.005853415001183748, + 0.005858573131263256, + 0.005388255696743727, + 0.0039414712227880955, + 0.008814462460577488, + 0.008125018328428268, + 0.007340675685554743, + 0.006311748642474413, + 0.007039125543087721, + 0.011141185648739338, + 0.007260932587087154, + 0.006336365360766649, + 0.030537458136677742, + 0.02471928671002388, + 0.007716421969234943, + 0.014821046032011509, + 0.012427007779479027, + 0.014815078116953373, + 0.012315589934587479, + 0.015050388872623444, + 0.010711549781262875, + 0.009476106613874435, + 0.012634871527552605, + 0.015955109149217606 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 2, + 0 + ], + "selected_reward": -0.00979462917894125 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a gritty urban streetscape with a cracked concrete sidewalk and a rain-soaked asphalt street?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the color palette desaturated with deep shadows and harsh streetlight glare on the wet ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a rider on a sleek, matte black one-wheeled vehicle with a thin cyan LED strip visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the rider wearing a dark hooded waterproof jacket and slim-fitting overalls while leaning forward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the rider pass a motionless pedestrian hunched on the sidewalk wearing a heavy coat?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.008642058819532394 + ], + "opposite_logprob_tokens": [ + -4.758642196655273 + ], + "logprob_reward": -0.008642058819532394, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.008642058819532394, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15627582371234894, + 0.12964916229248047, + 0.018998315557837486, + 0.16042238473892212, + 0.1415909230709076, + 0.0167771615087986, + 0.02463889680802822, + 0.1251295804977417, + 0.09369853883981705, + 0.1328192502260208 + ], + "credit_map": [ + 0.00995423924177885, + 0.0257562343031168, + 0.015835285186767578, + 0.012337099760770798, + 0.04568982869386673, + 0.025005590170621872, + 0.013349873945116997, + 0.008347667753696442, + 0.007051022723317146, + 0.007402333896607161, + 0.006632505916059017, + 0.02614886686205864, + 0.007865491323173046, + 0.02468382939696312, + 0.04224969446659088, + 0.007615406531840563, + 0.0014947685413062572, + 0.0028106041718274355, + 0.0019978561904281378, + 0.0026438022032380104, + 0.0024891484063118696, + 0.002066840184852481, + 0.002491465536877513, + 0.0030038300901651382, + 0.006979656871408224, + 0.01686709187924862, + 0.0057043833658099174, + 0.011246532201766968, + 0.018947459757328033, + 0.010037443600594997, + 0.04796832427382469, + 0.04267149046063423, + 0.013381017372012138, + 0.024454154074192047, + 0.011428851634263992, + 0.03400494530797005, + 0.017527222633361816, + 0.010709471069276333, + 0.01235099695622921, + 0.01773424819111824, + 0.0013809021329507232, + 0.002041153609752655, + 0.0017884403932839632, + 0.002129964530467987, + 0.002070851158350706, + 0.002203534124419093, + 0.002582517685368657, + 0.0025797963608056307, + 0.0024449285119771957, + 0.0031698637176305056, + 0.005566248204559088, + 0.002627146663144231, + 0.0025271873455494642, + 0.0022629990708082914, + 0.002831379184499383, + 0.0032091415487229824, + 0.0066459206864237785, + 0.012179282493889332, + 0.007574156858026981, + 0.004659558180719614, + 0.04143216833472252, + 0.00662233354523778, + 0.008271658793091774, + 0.03774448484182358, + 0.006718189921230078, + 0.01006149873137474, + 0.009868141263723373, + 0.008442585356533527, + 0.024846289306879044, + 0.01944616809487343, + 0.0068604848347604275, + 0.007455179467797279, + 0.012771711684763432, + 0.007407899014651775, + 0.027092864736914635, + 0.009430887177586555, + 0.007139757741242647, + 0.010040984489023685, + 0.01916423626244068, + 0.03977089002728462 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 0, + 4 + ], + "selected_reward": -0.008642058819532394 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a gritty urban streetscape with a cracked concrete sidewalk and a rain-soaked asphalt street?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the color palette desaturated with deep shadows and harsh streetlight glare on the wet ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a rider on a sleek, matte black one-wheeled vehicle with a thin cyan LED strip visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the rider wearing a dark hooded waterproof jacket and slim-fitting overalls while leaning forward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the rider pass a motionless pedestrian hunched on the sidewalk wearing a heavy coat?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0052520157769322395 + ], + "opposite_logprob_tokens": [ + -5.255251884460449 + ], + "logprob_reward": -0.0052520157769322395, + "yes_no_margin_reward": 5.25, + "token_credit_reward": -0.0052520157769322395, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14205415546894073, + 0.05872340500354767, + 0.03349139913916588, + 0.0977650061249733, + 0.09035855531692505, + 0.04280347377061844, + 0.02287125401198864, + 0.21819037199020386, + 0.14551717042922974, + 0.14822514355182648 + ], + "credit_map": [ + 0.013696812093257904, + 0.014054118655622005, + 0.03284132480621338, + 0.017842072993516922, + 0.01718832738697529, + 0.014687657356262207, + 0.0188056081533432, + 0.012938247062265873, + 0.0057352464646101, + 0.005407336167991161, + 0.00492823077365756, + 0.011085054837167263, + 0.006285001523792744, + 0.014357227832078934, + 0.005114986561238766, + 0.0058103264309465885, + 0.002539997920393944, + 0.006753816734999418, + 0.0026649178471416235, + 0.004087294917553663, + 0.004223331343382597, + 0.004767910577356815, + 0.004476071801036596, + 0.003978061024099588, + 0.005590009968727827, + 0.01112035010010004, + 0.004922717809677124, + 0.007951820269227028, + 0.013234936632215977, + 0.00840080063790083, + 0.024791941046714783, + 0.021752437576651573, + 0.00821189396083355, + 0.015235495753586292, + 0.007179617416113615, + 0.02136199176311493, + 0.011538312770426273, + 0.007131724618375301, + 0.007851406931877136, + 0.011848119087517262, + 0.00380252324976027, + 0.004648951813578606, + 0.004305575042963028, + 0.00574995344504714, + 0.004451801534742117, + 0.007242241874337196, + 0.0055828941985964775, + 0.007019539829343557, + 0.002384691033512354, + 0.0023484081029891968, + 0.004977313801646233, + 0.00267829024232924, + 0.0024088609497994184, + 0.0019875946454703808, + 0.003132416168227792, + 0.002953679533675313, + 0.01289828959852457, + 0.0201435387134552, + 0.010279993526637554, + 0.01106514222919941, + 0.07159997522830963, + 0.014742112718522549, + 0.021223559975624084, + 0.05623774975538254, + 0.008624596521258354, + 0.015274173580110073, + 0.008086089976131916, + 0.008886996656656265, + 0.0432104766368866, + 0.03136201947927475, + 0.011055858805775642, + 0.01901698298752308, + 0.017683565616607666, + 0.01700620725750923, + 0.024523558095097542, + 0.01608661375939846, + 0.01414128765463829, + 0.014802842400968075, + 0.018837522715330124, + 0.02514355070888996 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 9, + 8 + ], + "selected_reward": -0.0052520157769322395 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a gritty urban streetscape with a cracked concrete sidewalk and a rain-soaked asphalt street?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the color palette desaturated with deep shadows and harsh streetlight glare on the wet ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a rider on a sleek, matte black one-wheeled vehicle with a thin cyan LED strip visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the rider wearing a dark hooded waterproof jacket and slim-fitting overalls while leaning forward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the rider pass a motionless pedestrian hunched on the sidewalk wearing a heavy coat?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.06203450262546539 + ], + "opposite_logprob_tokens": [ + -2.8120346069335938 + ], + "logprob_reward": -0.06203450262546539, + "yes_no_margin_reward": 2.75, + "token_credit_reward": -0.06203450262546539, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.49354156851768494, + 0.03860337287187576, + 0.013791066594421864, + 0.04445795714855194, + 0.07553624361753464, + 0.03130669891834259, + 0.03309865668416023, + 0.1574506312608719, + 0.02625458315014839, + 0.08595921844244003 + ], + "credit_map": [ + 0.056386567652225494, + 0.06913866102695465, + 0.04311731457710266, + 0.03674418106675148, + 0.15036159753799438, + 0.07511094957590103, + 0.03728047385811806, + 0.025401819497346878, + 0.0036405851133167744, + 0.0038215152453631163, + 0.002751709660515189, + 0.0036777465138584375, + 0.0036622623447328806, + 0.01060589961707592, + 0.00737405801191926, + 0.003069595666602254, + 0.0012127307709306479, + 0.003261828562244773, + 0.0012721252860501409, + 0.0015027238987386227, + 0.0018867016769945621, + 0.0016449429094791412, + 0.0017653341637924314, + 0.0012446800246834755, + 0.0018467005575075746, + 0.0034609537106007338, + 0.002367202192544937, + 0.004089042078703642, + 0.003929748199880123, + 0.0035862200893461704, + 0.013550507836043835, + 0.011627581901848316, + 0.006401760503649712, + 0.011358004994690418, + 0.006193371955305338, + 0.017052508890628815, + 0.010926531627774239, + 0.007295280694961548, + 0.007027380168437958, + 0.009281406179070473, + 0.0026019983924925327, + 0.00276272208429873, + 0.0031849578954279423, + 0.00552512938156724, + 0.003107046941295266, + 0.005418261047452688, + 0.003900340525433421, + 0.0048062438145279884, + 0.0034370648209005594, + 0.002695895731449127, + 0.0064783659763634205, + 0.003301304066553712, + 0.00361051457002759, + 0.0036439914256334305, + 0.005868032108992338, + 0.0040634856559336185, + 0.0079100476577878, + 0.014674492180347443, + 0.0071544828824698925, + 0.006218372378498316, + 0.053406018763780594, + 0.009075804613530636, + 0.01205519400537014, + 0.046956222504377365, + 0.001742438180372119, + 0.002598351566120982, + 0.0020864401012659073, + 0.0022178927902132273, + 0.005692369304597378, + 0.005295076407492161, + 0.002452168148010969, + 0.004169846419245005, + 0.008678143844008446, + 0.006173949688673019, + 0.019458739086985588, + 0.005629092920571566, + 0.003443721681833267, + 0.005928873084485531, + 0.012136510573327541, + 0.024510186165571213 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 7, + 9 + ], + "selected_reward": -0.06203450262546539 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "static, eye-level long shot captures a gritty urban streetscape, bathed in a desaturated palette, contrasted by deep shadows and the harsh glare of streetlights on the wet sidewalk. A grimy city sidewalk, its cracked and uneven concrete slabs, stretches into the distance, bordered by a rain-soaked asphalt street that reflects a cold, artificial light. In the background, the muted reds and browns of a graffiti-tagged brick building fade into darkness. From the right side of the frame, a rider on a sleek, one-wheeled vehicle glides effortlessly and silently into the scene. The vehicle's matte black chassis absorbs light, while a thin cyan LED strip on the side provides a vibrant color accent. The unnamed rider, wearing a dark hooded waterproof jacket and slim-fitting overalls, leans forward, maintaining a focused balance. Swiftly and decisively, they pass a motionless pedestrian, hunched on the sidewalk, wearing a heavy, unremarkable coat, seemingly oblivious to the futuristic presence. The wheels sent up a fine, almost invisible spray from the wet ground, a subtle disturbance in the quiet, oppressive atmosphere of the city's anonymity.", + "seed": 30260, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000006/group_15/index_4287_seed_30260_dp_31_rank_62_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 15, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.018906574882566928, + "vlm_logprob": -0.018906574882566928, + "vlm_yes_no_margin": 4.5, + "vlm_token_credit": -0.018906574882566928, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a gritty urban streetscape with a cracked concrete sidewalk and a rain-soaked asphalt street?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the color palette desaturated with deep shadows and harsh streetlight glare on the wet ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a rider on a sleek, matte black one-wheeled vehicle with a thin cyan LED strip visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the rider wearing a dark hooded waterproof jacket and slim-fitting overalls while leaning forward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the rider pass a motionless pedestrian hunched on the sidewalk wearing a heavy coat?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.019420597702264786 + ], + "opposite_logprob_tokens": [ + -5.269420623779297 + ], + "logprob_reward": -0.019420597702264786, + "yes_no_margin_reward": 5.25, + "token_credit_reward": -0.019420597702264786, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08021719753742218, + 0.09134526550769806, + 0.06796735525131226, + 0.052323587238788605, + 0.04148421809077263, + 0.132148876786232, + 0.0758557841181755, + 0.10517498850822449, + 0.2554488182067871, + 0.09803390502929688 + ], + "credit_map": [ + 0.012291030958294868, + 0.008954747579991817, + 0.006503945682197809, + 0.01015318650752306, + 0.011439728550612926, + 0.01169659849256277, + 0.009586346335709095, + 0.00959161389619112, + 0.010912765748798847, + 0.011327758431434631, + 0.005850385874509811, + 0.011065438389778137, + 0.014335117302834988, + 0.012839037925004959, + 0.013769269920885563, + 0.011245490051805973, + 0.006366207264363766, + 0.005849035456776619, + 0.005056098569184542, + 0.008913138881325722, + 0.008972376585006714, + 0.010708080604672432, + 0.01281800027936697, + 0.009284417144954205, + 0.004589317832142115, + 0.00686543108895421, + 0.007721273694187403, + 0.009323072619736195, + 0.0053003160282969475, + 0.00815832894295454, + 0.004706658888608217, + 0.005659190472215414, + 0.003408831311389804, + 0.006360616069287062, + 0.0038053193129599094, + 0.0036646025255322456, + 0.008966797031462193, + 0.005626982543617487, + 0.004938344471156597, + 0.004712723195552826, + 0.006913293618708849, + 0.014976331032812595, + 0.015804631635546684, + 0.018908077850937843, + 0.019040336832404137, + 0.013417798094451427, + 0.022166121751070023, + 0.020922303199768066, + 0.008120222017168999, + 0.008948298171162605, + 0.007760312408208847, + 0.010719522833824158, + 0.011055183596909046, + 0.009396848268806934, + 0.007989509962499142, + 0.0118658896535635, + 0.011627046391367912, + 0.010255434550344944, + 0.01036812923848629, + 0.012011030688881874, + 0.016030002385377884, + 0.013664099387824535, + 0.009734779596328735, + 0.02148446813225746, + 0.020539404824376106, + 0.03618679195642471, + 0.02424049936234951, + 0.025315243750810623, + 0.026611411944031715, + 0.06896551698446274, + 0.0162249356508255, + 0.03736502304673195, + 0.00951175857335329, + 0.018469313159585, + 0.005835958290845156, + 0.010161151178181171, + 0.010016065090894699, + 0.018423430621623993, + 0.011562212370336056, + 0.014054026454687119 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 5, + 7 + ], + "selected_reward": -0.019420597702264786 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a gritty urban streetscape with a cracked concrete sidewalk and a rain-soaked asphalt street?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the color palette desaturated with deep shadows and harsh streetlight glare on the wet ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a rider on a sleek, matte black one-wheeled vehicle with a thin cyan LED strip visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the rider wearing a dark hooded waterproof jacket and slim-fitting overalls while leaning forward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the rider pass a motionless pedestrian hunched on the sidewalk wearing a heavy coat?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.008647731505334377 + ], + "opposite_logprob_tokens": [ + -4.758647918701172 + ], + "logprob_reward": -0.008647731505334377, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.008647731505334377, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15824469923973083, + 0.13650277256965637, + 0.057112857699394226, + 0.06772608309984207, + 0.051489248871803284, + 0.09423967450857162, + 0.07226405292749405, + 0.10832884907722473, + 0.15181100368499756, + 0.10228077322244644 + ], + "credit_map": [ + 0.020352721214294434, + 0.017214173451066017, + 0.014704570174217224, + 0.014121122658252716, + 0.02039981633424759, + 0.02651866525411606, + 0.01864914782345295, + 0.02628447487950325, + 0.014355955645442009, + 0.01835421845316887, + 0.012112045660614967, + 0.01680019497871399, + 0.022676395252346992, + 0.019921818748116493, + 0.017119640484452248, + 0.0151625145226717, + 0.005243292078375816, + 0.004641357809305191, + 0.0035719959996640682, + 0.01119820773601532, + 0.009481100365519524, + 0.007595505099743605, + 0.009066607803106308, + 0.006314794532954693, + 0.005269826389849186, + 0.007358338683843613, + 0.0071855634450912476, + 0.013558234088122845, + 0.009791585616767406, + 0.008609729818999767, + 0.009630569256842136, + 0.006322232075035572, + 0.0035011463332921267, + 0.006177553441375494, + 0.005089378450065851, + 0.005767918657511473, + 0.008992298506200314, + 0.007360232062637806, + 0.008343202993273735, + 0.006257515400648117, + 0.00575249781832099, + 0.012553524225950241, + 0.011310549452900887, + 0.009305913932621479, + 0.014837822876870632, + 0.011671529151499271, + 0.015432716347277164, + 0.013375119306147099, + 0.008084370754659176, + 0.007991193793714046, + 0.008409744128584862, + 0.009811167605221272, + 0.01132061704993248, + 0.008024774491786957, + 0.009302394464612007, + 0.00931979063898325, + 0.011650759726762772, + 0.00892628263682127, + 0.011063647456467152, + 0.01784849725663662, + 0.01634472794830799, + 0.01115741953253746, + 0.015886956825852394, + 0.015450557693839073, + 0.01133915688842535, + 0.021412869915366173, + 0.013276220299303532, + 0.013384244404733181, + 0.016025688499212265, + 0.04520934447646141, + 0.008290269412100315, + 0.02287321537733078, + 0.006841047201305628, + 0.016517961397767067, + 0.006118303630501032, + 0.010686492547392845, + 0.014299275353550911, + 0.01946393772959709, + 0.013541778549551964, + 0.014811976812779903 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 1 + ], + "selected_reward": -0.008647731505334377 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a gritty urban streetscape with a cracked concrete sidewalk and a rain-soaked asphalt street?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the color palette desaturated with deep shadows and harsh streetlight glare on the wet ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a rider on a sleek, matte black one-wheeled vehicle with a thin cyan LED strip visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the rider wearing a dark hooded waterproof jacket and slim-fitting overalls while leaning forward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the rider pass a motionless pedestrian hunched on the sidewalk wearing a heavy coat?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.006739741191267967 + ], + "opposite_logprob_tokens": [ + -5.006739616394043 + ], + "logprob_reward": -0.006739741191267967, + "yes_no_margin_reward": 5.0, + "token_credit_reward": -0.006739741191267967, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1036292314529419, + 0.08886533975601196, + 0.13726241886615753, + 0.04876367747783661, + 0.054861120879650116, + 0.07964786142110825, + 0.08747491985559464, + 0.10550043731927872, + 0.17959174513816833, + 0.11440322548151016 + ], + "credit_map": [ + 0.01015468779951334, + 0.011549529619514942, + 0.006827167700976133, + 0.010058438405394554, + 0.026097305119037628, + 0.016792399808764458, + 0.007882275618612766, + 0.014267432503402233, + 0.008551027625799179, + 0.01203744113445282, + 0.004719221498817205, + 0.009757231920957565, + 0.018976332619786263, + 0.012529583647847176, + 0.01171216368675232, + 0.010582340881228447, + 0.010450639761984348, + 0.008401176892220974, + 0.00920358207076788, + 0.018134726211428642, + 0.014271631836891174, + 0.024233847856521606, + 0.028316808864474297, + 0.024250013753771782, + 0.005462665110826492, + 0.007730242796242237, + 0.004990443121641874, + 0.007693620398640633, + 0.005557605531066656, + 0.005965270567685366, + 0.006091126706451178, + 0.005272702779620886, + 0.006705202162265778, + 0.006186472252011299, + 0.005813662428408861, + 0.005178346298635006, + 0.011164183728396893, + 0.008022910915315151, + 0.00502484105527401, + 0.006765500642359257, + 0.006549593526870012, + 0.009904526174068451, + 0.009876294061541557, + 0.010061673820018768, + 0.010409307666122913, + 0.008514869026839733, + 0.012655091471970081, + 0.011676506139338017, + 0.009026888757944107, + 0.006752803921699524, + 0.006053101271390915, + 0.013711929321289062, + 0.012919038534164429, + 0.008599074557423592, + 0.014411436393857002, + 0.016000641509890556, + 0.011529143899679184, + 0.00919201411306858, + 0.012921267189085484, + 0.01932087540626526, + 0.013112070970237255, + 0.011047800071537495, + 0.017960915341973305, + 0.010416347533464432, + 0.014879359863698483, + 0.02646751143038273, + 0.018108690157532692, + 0.019999049603939056, + 0.0198386050760746, + 0.038011714816093445, + 0.014782420359551907, + 0.027504393830895424, + 0.007444890681654215, + 0.016573071479797363, + 0.007116016931831837, + 0.014898731373250484, + 0.014078127220273018, + 0.016371063888072968, + 0.022017117589712143, + 0.01590421423316002 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 2, + 9 + ], + "selected_reward": -0.006739741191267967 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a gritty urban streetscape with a cracked concrete sidewalk and a rain-soaked asphalt street?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the color palette desaturated with deep shadows and harsh streetlight glare on the wet ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a rider on a sleek, matte black one-wheeled vehicle with a thin cyan LED strip visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the rider wearing a dark hooded waterproof jacket and slim-fitting overalls while leaning forward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the rider pass a motionless pedestrian hunched on the sidewalk wearing a heavy coat?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.011079221963882446 + ], + "opposite_logprob_tokens": [ + -4.51107931137085 + ], + "logprob_reward": -0.011079221963882446, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.011079221963882446, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1014418825507164, + 0.07207155227661133, + 0.08101803809404373, + 0.09511061012744904, + 0.07100922614336014, + 0.12745708227157593, + 0.05214114114642143, + 0.054600100964307785, + 0.13424170017242432, + 0.2109086811542511 + ], + "credit_map": [ + 0.009399572387337685, + 0.011943451128900051, + 0.0063325692899525166, + 0.00938449427485466, + 0.024050572887063026, + 0.01641811989247799, + 0.008506862446665764, + 0.015406235121190548, + 0.006689964793622494, + 0.008401468396186829, + 0.004544736351817846, + 0.008049702271819115, + 0.015636485069990158, + 0.009898284450173378, + 0.010161015205085278, + 0.00868989247828722, + 0.005704905837774277, + 0.006256175693124533, + 0.0050575630739331245, + 0.0089609045535326, + 0.011747844517230988, + 0.013765589334070683, + 0.01581207662820816, + 0.013712982647120953, + 0.006773685105144978, + 0.01003358419984579, + 0.00953867007046938, + 0.013600998558104038, + 0.00897996872663498, + 0.01616244576871395, + 0.018466519191861153, + 0.011554744094610214, + 0.005214024800807238, + 0.007002140395343304, + 0.005537012126296759, + 0.005839131306856871, + 0.019852906465530396, + 0.012438916601240635, + 0.007241363171488047, + 0.00788373127579689, + 0.007121921516954899, + 0.014968523755669594, + 0.017114900052547455, + 0.016380026936531067, + 0.021125927567481995, + 0.01192621048539877, + 0.019118351861834526, + 0.01970123127102852, + 0.005610560532659292, + 0.005541118327528238, + 0.004569181241095066, + 0.007931291125714779, + 0.008415184915065765, + 0.005189680028706789, + 0.008639541454613209, + 0.006244581658393145, + 0.00526763079687953, + 0.007056862581521273, + 0.005760962143540382, + 0.006715437863022089, + 0.010659542866051197, + 0.006169781554490328, + 0.006947477348148823, + 0.006022402551025152, + 0.00918958242982626, + 0.020693305879831314, + 0.01243650633841753, + 0.014910003170371056, + 0.014893651008605957, + 0.03289691358804703, + 0.008861836045980453, + 0.02035989984869957, + 0.01401663850992918, + 0.040171604603528976, + 0.013488607481122017, + 0.023006493225693703, + 0.028446557000279427, + 0.03176436200737953, + 0.03171173855662346, + 0.02830267697572708 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 8, + 5 + ], + "selected_reward": -0.011079221963882446 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a gritty urban streetscape with a cracked concrete sidewalk and a rain-soaked asphalt street?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the color palette desaturated with deep shadows and harsh streetlight glare on the wet ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a rider on a sleek, matte black one-wheeled vehicle with a thin cyan LED strip visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the rider wearing a dark hooded waterproof jacket and slim-fitting overalls while leaning forward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the rider pass a motionless pedestrian hunched on the sidewalk wearing a heavy coat?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04864558205008507 + ], + "opposite_logprob_tokens": [ + -3.048645496368408 + ], + "logprob_reward": -0.04864558205008507, + "yes_no_margin_reward": 3.0, + "token_credit_reward": -0.04864558205008507, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1787540465593338, + 0.20157715678215027, + 0.06442885845899582, + 0.05708654969930649, + 0.02650383859872818, + 0.06211204454302788, + 0.044450219720602036, + 0.1187821552157402, + 0.07134222984313965, + 0.1749628782272339 + ], + "credit_map": [ + 0.010816490277647972, + 0.024241719394922256, + 0.00985924806445837, + 0.012786810286343098, + 0.06247229874134064, + 0.03362244367599487, + 0.013086038641631603, + 0.011869005858898163, + 0.024990534409880638, + 0.031396862119436264, + 0.01023193635046482, + 0.01633555069565773, + 0.04949366673827171, + 0.030660994350910187, + 0.018542755395174026, + 0.01992487721145153, + 0.004747071303427219, + 0.006287555675953627, + 0.0034364566672593355, + 0.0074025415815413, + 0.01103094220161438, + 0.011085757985711098, + 0.009328343905508518, + 0.011110186576843262, + 0.004512906074523926, + 0.00503868144005537, + 0.003551843808963895, + 0.007670716382563114, + 0.010215317830443382, + 0.008131202310323715, + 0.010147704742848873, + 0.00781817827373743, + 0.0026879797223955393, + 0.002820350928232074, + 0.0030242796055972576, + 0.0023393079172819853, + 0.003397800726816058, + 0.00431784288957715, + 0.0032633382361382246, + 0.0046529402025043964, + 0.003927531186491251, + 0.009717735461890697, + 0.007445639930665493, + 0.0111001031473279, + 0.007575256749987602, + 0.009282972663640976, + 0.006152156740427017, + 0.006910651922225952, + 0.005497721489518881, + 0.004997578449547291, + 0.004077307879924774, + 0.00767893809825182, + 0.006660667713731527, + 0.004959438927471638, + 0.0042196763679385185, + 0.006358889397233725, + 0.011931275017559528, + 0.01138896495103836, + 0.007902644574642181, + 0.013366293162107468, + 0.014114359393715858, + 0.013311208225786686, + 0.009750344790518284, + 0.037017062306404114, + 0.004400573205202818, + 0.008074727840721607, + 0.004884330090135336, + 0.0058464352041482925, + 0.00796525925397873, + 0.022814588621258736, + 0.006348336115479469, + 0.01100798137485981, + 0.012929706834256649, + 0.023113787174224854, + 0.012730136513710022, + 0.017677443102002144, + 0.022240683436393738, + 0.02724369801580906, + 0.03274654969573021, + 0.026280883699655533 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 9 + ], + "selected_reward": -0.04864558205008507 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "static, eye-level long shot captures a gritty urban streetscape, bathed in a desaturated palette, contrasted by deep shadows and the harsh glare of streetlights on the wet sidewalk. A grimy city sidewalk, its cracked and uneven concrete slabs, stretches into the distance, bordered by a rain-soaked asphalt street that reflects a cold, artificial light. In the background, the muted reds and browns of a graffiti-tagged brick building fade into darkness. From the right side of the frame, a rider on a sleek, one-wheeled vehicle glides effortlessly and silently into the scene. The vehicle's matte black chassis absorbs light, while a thin cyan LED strip on the side provides a vibrant color accent. The unnamed rider, wearing a dark hooded waterproof jacket and slim-fitting overalls, leans forward, maintaining a focused balance. Swiftly and decisively, they pass a motionless pedestrian, hunched on the sidewalk, wearing a heavy, unremarkable coat, seemingly oblivious to the futuristic presence. The wheels sent up a fine, almost invisible spray from the wet ground, a subtle disturbance in the quiet, oppressive atmosphere of the city's anonymity.", + "seed": 30261, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000006/group_15/index_4287_seed_30261_dp_31_rank_62_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 15, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.023586494382470846, + "vlm_logprob": -0.023586494382470846, + "vlm_yes_no_margin": 4.425, + "vlm_token_credit": -0.023586494382470846, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a gritty urban streetscape with a cracked concrete sidewalk and a rain-soaked asphalt street?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the color palette desaturated with deep shadows and harsh streetlight glare on the wet ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a rider on a sleek, matte black one-wheeled vehicle with a thin cyan LED strip visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the rider wearing a dark hooded waterproof jacket and slim-fitting overalls while leaning forward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the rider pass a motionless pedestrian hunched on the sidewalk wearing a heavy coat?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.014574728906154633 + ], + "opposite_logprob_tokens": [ + -5.1395745277404785 + ], + "logprob_reward": -0.014574728906154633, + "yes_no_margin_reward": 5.125, + "token_credit_reward": -0.014574728906154633, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.25536292791366577, + 0.06580802798271179, + 0.03557240962982178, + 0.023686468601226807, + 0.03844910487532616, + 0.05717269703745842, + 0.03721042349934578, + 0.0545930415391922, + 0.04943859577178955, + 0.38270631432533264 + ], + "credit_map": [ + 0.015155031345784664, + 0.016663318499922752, + 0.010542117059230804, + 0.011476976796984673, + 0.02359159104526043, + 0.0877874493598938, + 0.014311704784631729, + 0.07583475857973099, + 0.005304170772433281, + 0.006679651793092489, + 0.003510880982503295, + 0.004202826414257288, + 0.013241125270724297, + 0.01571660302579403, + 0.009077340364456177, + 0.008075431920588017, + 0.0038602803833782673, + 0.0032927654683589935, + 0.004715977236628532, + 0.002779018133878708, + 0.004280697554349899, + 0.006463967263698578, + 0.004031125456094742, + 0.0061485785990953445, + 0.0027043058071285486, + 0.0025324916932731867, + 0.0036102363374084234, + 0.0024272638838738203, + 0.004054544027894735, + 0.003983234986662865, + 0.001846902770921588, + 0.0025274904910475016, + 0.003744659014046192, + 0.005293190013617277, + 0.003299748757854104, + 0.004738969728350639, + 0.008234868757426739, + 0.007651729974895716, + 0.0030311939772218466, + 0.0024547483772039413, + 0.005289311520755291, + 0.01113610528409481, + 0.00620906800031662, + 0.011166561394929886, + 0.004533117171376944, + 0.007502011023461819, + 0.006035953760147095, + 0.00530056981369853, + 0.003997080028057098, + 0.005313057452440262, + 0.002588114235550165, + 0.004716614726930857, + 0.00490714842453599, + 0.006421220023185015, + 0.005509128328412771, + 0.0037580642383545637, + 0.005747293122112751, + 0.011811554431915283, + 0.003271546447649598, + 0.0057235537096858025, + 0.00888446718454361, + 0.008877691812813282, + 0.006049457471817732, + 0.004227478057146072, + 0.004037672653794289, + 0.006817297078669071, + 0.003275540191680193, + 0.005277157761156559, + 0.01188700646162033, + 0.0070855263620615005, + 0.005544575862586498, + 0.005513820331543684, + 0.011212470009922981, + 0.10002151131629944, + 0.00969765055924654, + 0.00823489110916853, + 0.16300050914287567, + 0.03997156769037247, + 0.009758204221725464, + 0.040809523314237595 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 0, + 1 + ], + "selected_reward": -0.014574728906154633 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a gritty urban streetscape with a cracked concrete sidewalk and a rain-soaked asphalt street?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the color palette desaturated with deep shadows and harsh streetlight glare on the wet ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a rider on a sleek, matte black one-wheeled vehicle with a thin cyan LED strip visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the rider wearing a dark hooded waterproof jacket and slim-fitting overalls while leaning forward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the rider pass a motionless pedestrian hunched on the sidewalk wearing a heavy coat?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005960309412330389 + ], + "opposite_logprob_tokens": [ + -5.130960464477539 + ], + "logprob_reward": -0.005960309412330389, + "yes_no_margin_reward": 5.125, + "token_credit_reward": -0.005960309412330389, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.22181065380573273, + 0.051635969430208206, + 0.1458415538072586, + 0.05699330195784569, + 0.08606573194265366, + 0.1664758324623108, + 0.04426911473274231, + 0.06250237673521042, + 0.0571562685072422, + 0.1072491854429245 + ], + "credit_map": [ + 0.013320032507181168, + 0.016953244805336, + 0.008168076165020466, + 0.010610739700496197, + 0.0241848211735487, + 0.07269387692213058, + 0.013202665373682976, + 0.06267719715833664, + 0.004773853346705437, + 0.006905036978423595, + 0.0027989749796688557, + 0.003237625118345022, + 0.012190037406980991, + 0.010201050899922848, + 0.0054059261456131935, + 0.0061234645545482635, + 0.014595179818570614, + 0.01839509978890419, + 0.016817990690469742, + 0.010924953036010265, + 0.02290031872689724, + 0.030262550339102745, + 0.012005051597952843, + 0.01994040608406067, + 0.006333187688142061, + 0.014155915006995201, + 0.011554907076060772, + 0.0037400173023343086, + 0.005493860226124525, + 0.00820479728281498, + 0.0021712377201765776, + 0.005339385475963354, + 0.008004139177501202, + 0.009336617775261402, + 0.007789329160004854, + 0.007788754068315029, + 0.0214547086507082, + 0.01589386537671089, + 0.007465956266969442, + 0.008332367055118084, + 0.019131921231746674, + 0.02929108217358589, + 0.021348292008042336, + 0.03542120009660721, + 0.014981674961745739, + 0.023823607712984085, + 0.01264128740876913, + 0.009836765937507153, + 0.00463853869587183, + 0.007571211084723473, + 0.0034523496869951487, + 0.004393693991005421, + 0.005756573285907507, + 0.006687560584396124, + 0.007486571557819843, + 0.004282612819224596, + 0.006451252847909927, + 0.01209578849375248, + 0.004425066523253918, + 0.006210795603692532, + 0.012538657523691654, + 0.010919219814240932, + 0.006002470385283232, + 0.003859121585264802, + 0.00556740490719676, + 0.007467940449714661, + 0.004386466462165117, + 0.006093727890402079, + 0.014655872248113155, + 0.008447523228824139, + 0.005132556892931461, + 0.005404784344136715, + 0.008003723807632923, + 0.021086497232317924, + 0.0070764594711363316, + 0.005144651513546705, + 0.020931880921125412, + 0.015493009239435196, + 0.007788029033690691, + 0.02172493189573288 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 5, + 2 + ], + "selected_reward": -0.005960309412330389 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a gritty urban streetscape with a cracked concrete sidewalk and a rain-soaked asphalt street?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the color palette desaturated with deep shadows and harsh streetlight glare on the wet ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a rider on a sleek, matte black one-wheeled vehicle with a thin cyan LED strip visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the rider wearing a dark hooded waterproof jacket and slim-fitting overalls while leaning forward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the rider pass a motionless pedestrian hunched on the sidewalk wearing a heavy coat?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.008643004111945629 + ], + "opposite_logprob_tokens": [ + -4.75864315032959 + ], + "logprob_reward": -0.008643004111945629, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.008643004111945629, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0689338818192482, + 0.03692389652132988, + 0.02259283885359764, + 0.04120625555515289, + 0.08180610835552216, + 0.3208346366882324, + 0.09779571741819382, + 0.10247471183538437, + 0.1110973060131073, + 0.11633466929197311 + ], + "credit_map": [ + 0.004937875550240278, + 0.006662414874881506, + 0.0031059293542057276, + 0.004966397304087877, + 0.015263793990015984, + 0.01494584046304226, + 0.006276573054492474, + 0.01277505699545145, + 0.0034296177327632904, + 0.004148315638303757, + 0.0024113510735332966, + 0.002125037834048271, + 0.008790728636085987, + 0.007767303381115198, + 0.004501889459788799, + 0.003749655093997717, + 0.0022903617937117815, + 0.0019917008467018604, + 0.0031271481420844793, + 0.0013294598320499063, + 0.0037778900004923344, + 0.0046135731972754, + 0.002345268614590168, + 0.0031174386385828257, + 0.003968492615967989, + 0.01209060288965702, + 0.009195344522595406, + 0.00273579778149724, + 0.003816303564235568, + 0.00455841701477766, + 0.0020680599845945835, + 0.0027732360176742077, + 0.00856055598706007, + 0.01034020259976387, + 0.006710477638989687, + 0.008731923997402191, + 0.012332424521446228, + 0.013249136507511139, + 0.01277895923703909, + 0.009102432057261467, + 0.03207537159323692, + 0.050530605018138885, + 0.049360647797584534, + 0.07859652489423752, + 0.023080939427018166, + 0.04171324148774147, + 0.023851828649640083, + 0.021625466644763947, + 0.009535207413136959, + 0.01426349114626646, + 0.01062057726085186, + 0.012052709236741066, + 0.010950401425361633, + 0.012337425723671913, + 0.019607143476605415, + 0.008428771048784256, + 0.010388313792645931, + 0.014804191887378693, + 0.0055708675645291805, + 0.010376004502177238, + 0.022083235904574394, + 0.018915386870503426, + 0.012921654619276524, + 0.007415051106363535, + 0.012494032271206379, + 0.012165399268269539, + 0.00957534834742546, + 0.013786912895739079, + 0.02446100488305092, + 0.015715179964900017, + 0.010715020820498466, + 0.012184408493340015, + 0.005262944847345352, + 0.016462435945868492, + 0.004658987745642662, + 0.0033960642758756876, + 0.04016807675361633, + 0.015622321516275406, + 0.006370685528963804, + 0.024393156170845032 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 9, + 8 + ], + "selected_reward": -0.008643004111945629 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a gritty urban streetscape with a cracked concrete sidewalk and a rain-soaked asphalt street?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the color palette desaturated with deep shadows and harsh streetlight glare on the wet ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a rider on a sleek, matte black one-wheeled vehicle with a thin cyan LED strip visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the rider wearing a dark hooded waterproof jacket and slim-fitting overalls while leaning forward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the rider pass a motionless pedestrian hunched on the sidewalk wearing a heavy coat?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.00978624727576971 + ], + "opposite_logprob_tokens": [ + -4.634786128997803 + ], + "logprob_reward": -0.00978624727576971, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.00978624727576971, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15753737092018127, + 0.04081893339753151, + 0.02810807153582573, + 0.0455646775662899, + 0.13474497199058533, + 0.18816740810871124, + 0.05471665412187576, + 0.07409778982400894, + 0.08024603873491287, + 0.19599802792072296 + ], + "credit_map": [ + 0.013782849535346031, + 0.019608136266469955, + 0.006166359875351191, + 0.008129038847982883, + 0.027050476521253586, + 0.04124363884329796, + 0.010181781835854053, + 0.03137510269880295, + 0.00346004799939692, + 0.005973119754344225, + 0.002381371334195137, + 0.0026189659256488085, + 0.008673261851072311, + 0.009456479921936989, + 0.004010362084954977, + 0.0042453305795788765, + 0.002562049077823758, + 0.002805853495374322, + 0.0036345464177429676, + 0.002338652266189456, + 0.004899261984974146, + 0.005617764312773943, + 0.002739525167271495, + 0.003510420909151435, + 0.005476182326674461, + 0.0112839937210083, + 0.007032075896859169, + 0.003037642687559128, + 0.005091871600598097, + 0.006394118536263704, + 0.002292030258104205, + 0.004956768359988928, + 0.012027506716549397, + 0.01653674989938736, + 0.012707704678177834, + 0.016731226816773415, + 0.017395805567502975, + 0.020246505737304688, + 0.024689188227057457, + 0.01441030390560627, + 0.02160835638642311, + 0.031968019902706146, + 0.026751834899187088, + 0.03733949363231659, + 0.015338833443820477, + 0.02858482301235199, + 0.01445296686142683, + 0.012123098596930504, + 0.006000253837555647, + 0.009471829980611801, + 0.0053412881679832935, + 0.005567141342908144, + 0.005609256215393543, + 0.008801070973277092, + 0.009419495239853859, + 0.004506323486566544, + 0.007201568689197302, + 0.013843893073499203, + 0.0037977909669280052, + 0.006880421657115221, + 0.017097502946853638, + 0.013648701831698418, + 0.007074322085827589, + 0.004553595557808876, + 0.007908389903604984, + 0.0115969218313694, + 0.004492743872106075, + 0.00677618058398366, + 0.02293163165450096, + 0.012660566717386246, + 0.006706637796014547, + 0.007172978017479181, + 0.0074351998046040535, + 0.03467908129096031, + 0.007637308444827795, + 0.007030131295323372, + 0.06726185977458954, + 0.023683007806539536, + 0.011550907045602798, + 0.036720555275678635 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 5, + 0 + ], + "selected_reward": -0.00978624727576971 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a gritty urban streetscape with a cracked concrete sidewalk and a rain-soaked asphalt street?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the color palette desaturated with deep shadows and harsh streetlight glare on the wet ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a rider on a sleek, matte black one-wheeled vehicle with a thin cyan LED strip visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the rider wearing a dark hooded waterproof jacket and slim-fitting overalls while leaning forward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the rider pass a motionless pedestrian hunched on the sidewalk wearing a heavy coat?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07896818220615387 + ], + "opposite_logprob_tokens": [ + -2.5789682865142822 + ], + "logprob_reward": -0.07896818220615387, + "yes_no_margin_reward": 2.5, + "token_credit_reward": -0.07896818220615387, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11833296716213226, + 0.037651460617780685, + 0.04329230263829231, + 0.03323906287550926, + 0.053035665303468704, + 0.12556672096252441, + 0.038826897740364075, + 0.06148844584822655, + 0.10008607059717178, + 0.38848042488098145 + ], + "credit_map": [ + 0.00689469650387764, + 0.010670415125787258, + 0.004603445064276457, + 0.007280592806637287, + 0.015027591027319431, + 0.036113377660512924, + 0.00952957384288311, + 0.028213275596499443, + 0.0030588642694056034, + 0.005612248554825783, + 0.0019397878786548972, + 0.0022566451225429773, + 0.008994135074317455, + 0.006863623857498169, + 0.0037536618765443563, + 0.005172493867576122, + 0.0031978581100702286, + 0.0041481973603367805, + 0.007432583253830671, + 0.002908321563154459, + 0.007996268570423126, + 0.007929415442049503, + 0.0038428360130637884, + 0.0058368234895169735, + 0.0034787592012435198, + 0.0105739189311862, + 0.0052743167616426945, + 0.0015054339310154319, + 0.004889637231826782, + 0.0032475197222083807, + 0.0011616592528298497, + 0.0031078183092176914, + 0.005500558763742447, + 0.009803097695112228, + 0.005708911921828985, + 0.006542900111526251, + 0.00887326430529356, + 0.008349589072167873, + 0.004233349580317736, + 0.004023991525173187, + 0.01566332019865513, + 0.02305414341390133, + 0.014840999618172646, + 0.017864983528852463, + 0.01213009376078844, + 0.02203555777668953, + 0.011127883568406105, + 0.008849738165736198, + 0.0039047659374773502, + 0.004622165113687515, + 0.003888737875968218, + 0.003476392477750778, + 0.005375535227358341, + 0.0077140615321695805, + 0.0065419855527579784, + 0.003303257282823324, + 0.005376141052693129, + 0.012978620827198029, + 0.0035388388205319643, + 0.007326369173824787, + 0.011678824201226234, + 0.010810492560267448, + 0.00580595200881362, + 0.003973207902163267, + 0.007658353075385094, + 0.015664568170905113, + 0.006079947110265493, + 0.008572719991207123, + 0.025874294340610504, + 0.014602303504943848, + 0.010653204284608364, + 0.010980687104165554, + 0.02380942367017269, + 0.11639687418937683, + 0.01587854139506817, + 0.015354925766587257, + 0.08803008496761322, + 0.050165869295597076, + 0.02399756945669651, + 0.05484714359045029 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 5, + 0 + ], + "selected_reward": -0.07896818220615387 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000006/group_2/index_tensor([1599, 1599])_group_2_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000006/group_2/index_tensor([1599, 1599])_group_2_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..2aaf712f98556748220ed04835c776fd82b77514 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000006/group_2/index_tensor([1599, 1599])_group_2_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "smoothly panning drone shot follows a small boat as it navigates an abstract world composed entirely of vibrant blocks of color and bold black lines, à la Henri Matisse. The image below unfolds not as a physical body of water, but rather as a vast, deconstructed map of the Forbidden City complex, composed of large, flat areas of color: cadmium yellow representing palace rooftops, vermilion areas symbolizing city walls, and geometric shapes of azure and emerald green representing courtyards, all clearly separated by decisive, expressive black outlines. Across this sea of color, a symbolic vessel, minimalist and defined by a few black, calligraphic strokes, its hull a pure block of ultramarine blue, glides forward with a steady, rhythmic rhythm. Each stroke of the oars creates stylized ripples of simple lines across the flat surface of color, a movement of grace and deliberation. The entire picture is free from traditional light and shadow and perspective, presenting a purely flat decorative beauty. The light is evenly distributed on each color block, creating a tranquil, dreamy and artistically meaningful atmosphere, like a symbolic journey through history and modern art.", + "seed": 23460, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000006/group_2/index_1599_seed_23460_dp_4_rank_8_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 2, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.041162051260471344, + "vlm_logprob": -0.041162051260471344, + "vlm_yes_no_margin": 4.75, + "vlm_token_credit": -0.041162051260471344, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene an abstract world composed of vibrant color blocks and bold black lines in a Henri Matisse style?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the background a deconstructed map of the Forbidden City with cadmium yellow rooftops, vermilion walls, and azure/emerald green courtyards?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is there a minimalist boat with a pure ultramarine blue hull and black calligraphic strokes gliding across the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the boat create stylized ripples of simple lines as it moves, without traditional light, shadow, or perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot a smooth pan following the boat's forward movement?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0016381428577005863 + ], + "opposite_logprob_tokens": [ + -7.001637935638428 + ], + "logprob_reward": -0.0016381428577005863, + "yes_no_margin_reward": 7.0, + "token_credit_reward": -0.0016381428577005863, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.20700840651988983, + 0.1640753149986267, + 0.03276330232620239, + 0.07668118178844452, + 0.0503237284719944, + 0.050943318754434586, + 0.08379665017127991, + 0.17577044665813446, + 0.08958204835653305, + 0.06905560940504074 + ], + "credit_map": [ + 0.029957609251141548, + 0.020225178450345993, + 0.028202788904309273, + 0.027314428240060806, + 0.020742539316415787, + 0.02210649661719799, + 0.03192087635397911, + 0.026538487523794174, + 0.018941767513751984, + 0.015148268081247807, + 0.013996713794767857, + 0.01502182800322771, + 0.04319982975721359, + 0.027674011886119843, + 0.013866343535482883, + 0.01622655801475048, + 0.00367885734885931, + 0.006759440526366234, + 0.0034873366821557283, + 0.002778461202979088, + 0.003115680767223239, + 0.0034172560553997755, + 0.0038129487074911594, + 0.005713319405913353, + 0.006915553007274866, + 0.008680474944412708, + 0.008731873705983162, + 0.011917823925614357, + 0.005183234345167875, + 0.00476716598495841, + 0.01644831895828247, + 0.014036732725799084, + 0.004806133918464184, + 0.004544729366898537, + 0.007394708693027496, + 0.004770766943693161, + 0.007899103686213493, + 0.004349132999777794, + 0.00645726453512907, + 0.01010188926011324, + 0.005889182444661856, + 0.004842129535973072, + 0.005587393417954445, + 0.011558394879102707, + 0.00468952301889658, + 0.0053555285558104515, + 0.007760545238852501, + 0.005260622128844261, + 0.009257053025066853, + 0.009673400782048702, + 0.012071794830262661, + 0.009033406153321266, + 0.010985528118908405, + 0.011835752986371517, + 0.0075333090499043465, + 0.013406402431428432, + 0.021570423617959023, + 0.016337180510163307, + 0.021356983110308647, + 0.010895111598074436, + 0.014401563443243504, + 0.025987721979618073, + 0.02465914376080036, + 0.04056231305003166, + 0.009708838537335396, + 0.010115377604961395, + 0.025491628795862198, + 0.012016607448458672, + 0.0075754448771476746, + 0.00835037138313055, + 0.007815181277692318, + 0.008508599363267422, + 0.011413443833589554, + 0.006149264983832836, + 0.009172525256872177, + 0.006202151533216238, + 0.007756279781460762, + 0.010364881716668606, + 0.006942295003682375, + 0.011054771021008492 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 7, + 1 + ], + "selected_reward": -0.0016381428577005863 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene an abstract world composed of vibrant color blocks and bold black lines in a Henri Matisse style?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the background a deconstructed map of the Forbidden City with cadmium yellow rooftops, vermilion walls, and azure/emerald green courtyards?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is there a minimalist boat with a pure ultramarine blue hull and black calligraphic strokes gliding across the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the boat create stylized ripples of simple lines as it moves, without traditional light, shadow, or perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot a smooth pan following the boat's forward movement?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.016088075935840607 + ], + "opposite_logprob_tokens": [ + -4.141088008880615 + ], + "logprob_reward": -0.016088075935840607, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.016088075935840607, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.16882212460041046, + 0.09904574602842331, + 0.0373709574341774, + 0.06412597000598907, + 0.0610324963927269, + 0.10404948145151138, + 0.10410363227128983, + 0.18103353679180145, + 0.07557621598243713, + 0.10483985394239426 + ], + "credit_map": [ + 0.01859797164797783, + 0.02154277078807354, + 0.0178330447524786, + 0.025613106787204742, + 0.021759631112217903, + 0.01304412167519331, + 0.027361810207366943, + 0.02306966483592987, + 0.012916305102407932, + 0.013412474654614925, + 0.017022885382175446, + 0.013280261307954788, + 0.009009716100990772, + 0.0074082957580685616, + 0.013919644057750702, + 0.012076156213879585, + 0.003916894551366568, + 0.008532651700079441, + 0.0033727195113897324, + 0.0036971494555473328, + 0.004155331291258335, + 0.0044305697083473206, + 0.0034821107983589172, + 0.005783530883491039, + 0.006245073862373829, + 0.007669476792216301, + 0.012161455117166042, + 0.009158234111964703, + 0.005475303623825312, + 0.004951893351972103, + 0.008789058774709702, + 0.009675468318164349, + 0.008878074586391449, + 0.008584121242165565, + 0.0054015787318348885, + 0.005262670107185841, + 0.006811394821852446, + 0.006681577302515507, + 0.006662643048912287, + 0.012750438414514065, + 0.009546379558742046, + 0.006680753082036972, + 0.011098665185272694, + 0.027732325717806816, + 0.010340328328311443, + 0.00879114679992199, + 0.02025759592652321, + 0.009602281264960766, + 0.012098095379769802, + 0.012112516909837723, + 0.018607109785079956, + 0.013020365498960018, + 0.009755323641002178, + 0.015082795172929764, + 0.009205259382724762, + 0.014222163707017899, + 0.02482115849852562, + 0.017704201862215996, + 0.017918165773153305, + 0.010532267391681671, + 0.026633314788341522, + 0.017369017004966736, + 0.018721504136919975, + 0.04733390733599663, + 0.012949934229254723, + 0.00870379339903593, + 0.01626739278435707, + 0.008201979100704193, + 0.007778239902108908, + 0.00634795380756259, + 0.007061107084155083, + 0.00826582033187151, + 0.010212249122560024, + 0.008839926682412624, + 0.014440027996897697, + 0.015101289376616478, + 0.008934006094932556, + 0.008607733994722366, + 0.016627077013254166, + 0.022077541798353195 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 0, + 9 + ], + "selected_reward": -0.016088075935840607 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene an abstract world composed of vibrant color blocks and bold black lines in a Henri Matisse style?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the background a deconstructed map of the Forbidden City with cadmium yellow rooftops, vermilion walls, and azure/emerald green courtyards?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is there a minimalist boat with a pure ultramarine blue hull and black calligraphic strokes gliding across the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the boat create stylized ripples of simple lines as it moves, without traditional light, shadow, or perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot a smooth pan following the boat's forward movement?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004636727273464203 + ], + "opposite_logprob_tokens": [ + -5.379636764526367 + ], + "logprob_reward": -0.004636727273464203, + "yes_no_margin_reward": 5.375, + "token_credit_reward": -0.004636727273464203, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3886505365371704, + 0.07577482610940933, + 0.02070801518857479, + 0.21136531233787537, + 0.035777147859334946, + 0.02906891144812107, + 0.02455315925180912, + 0.06984717398881912, + 0.022588692605495453, + 0.12166617810726166 + ], + "credit_map": [ + 0.040026433765888214, + 0.04367263615131378, + 0.0438094325363636, + 0.06754828989505768, + 0.04308086633682251, + 0.025774145498871803, + 0.07099489867687225, + 0.05374384671449661, + 0.006600324530154467, + 0.008227030746638775, + 0.007237846031785011, + 0.007963680662214756, + 0.018659573048353195, + 0.012073544785380363, + 0.007141462527215481, + 0.007871362380683422, + 0.002181061776354909, + 0.0034156511537730694, + 0.00243963859975338, + 0.0022383269388228655, + 0.0023879276122897863, + 0.0026720361784100533, + 0.0021303582470864058, + 0.003243016777560115, + 0.013691345229744911, + 0.020715391263365746, + 0.026955774053931236, + 0.036082085222005844, + 0.010675405152142048, + 0.012700633145868778, + 0.052986107766628265, + 0.03755856305360794, + 0.005343410186469555, + 0.0028900490142405033, + 0.005867875646799803, + 0.002362990751862526, + 0.00514874467626214, + 0.0020902787800878286, + 0.004893433302640915, + 0.007180365733802319, + 0.0026169191114604473, + 0.0028634315822273493, + 0.0027858486864715815, + 0.0066260602325201035, + 0.00289901508949697, + 0.002514560241252184, + 0.005094378255307674, + 0.003668698715046048, + 0.0027862712740898132, + 0.0028362483717501163, + 0.0035859106574207544, + 0.0027500162832438946, + 0.0029742722399532795, + 0.00369890034198761, + 0.002004254376515746, + 0.003917286638170481, + 0.007236122619360685, + 0.006818390917032957, + 0.007516474463045597, + 0.005566931329667568, + 0.00930842850357294, + 0.011277501471340656, + 0.008651943877339363, + 0.013471380807459354, + 0.003389491466805339, + 0.0030495210085064173, + 0.003616559784859419, + 0.002394637558609247, + 0.003246790962293744, + 0.002319824881851673, + 0.0018329797312617302, + 0.0027388855814933777, + 0.008503235876560211, + 0.007287313695997, + 0.0248783677816391, + 0.015650931745767593, + 0.010332011617720127, + 0.011467678472399712, + 0.017358729615807533, + 0.02618790976703167 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 9 + ], + "selected_reward": -0.004636727273464203 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene an abstract world composed of vibrant color blocks and bold black lines in a Henri Matisse style?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the background a deconstructed map of the Forbidden City with cadmium yellow rooftops, vermilion walls, and azure/emerald green courtyards?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is there a minimalist boat with a pure ultramarine blue hull and black calligraphic strokes gliding across the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the boat create stylized ripples of simple lines as it moves, without traditional light, shadow, or perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot a smooth pan following the boat's forward movement?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.003612422849982977 + ], + "opposite_logprob_tokens": [ + -5.628612518310547 + ], + "logprob_reward": -0.003612422849982977, + "yes_no_margin_reward": 5.625, + "token_credit_reward": -0.003612422849982977, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2786732614040375, + 0.09089040011167526, + 0.05873335525393486, + 0.06699557602405548, + 0.07066220790147781, + 0.04406930133700371, + 0.08143381774425507, + 0.18897733092308044, + 0.025921309366822243, + 0.09364347904920578 + ], + "credit_map": [ + 0.03177236020565033, + 0.031734898686409, + 0.03614388406276703, + 0.041100628674030304, + 0.03105814941227436, + 0.02046489715576172, + 0.048571884632110596, + 0.03782656788825989, + 0.011735326610505581, + 0.00854705274105072, + 0.012178203091025352, + 0.011679992079734802, + 0.01331600546836853, + 0.010256371460855007, + 0.010708000510931015, + 0.012469443492591381, + 0.007212632801383734, + 0.008469589054584503, + 0.005895469803363085, + 0.0050164926797151566, + 0.007556207478046417, + 0.008079232648015022, + 0.006168590858578682, + 0.010335138067603111, + 0.006567276548594236, + 0.008564494550228119, + 0.008926319889724255, + 0.00809838529676199, + 0.006868139375001192, + 0.0044013638980686665, + 0.012952674180269241, + 0.010616922751069069, + 0.006128546316176653, + 0.009002462029457092, + 0.01087863277643919, + 0.004004352260380983, + 0.010730165988206863, + 0.005263687577098608, + 0.007144646253436804, + 0.017509711906313896, + 0.004966626409441233, + 0.004684098530560732, + 0.004368979018181562, + 0.005837808828800917, + 0.006015780847519636, + 0.007448466029018164, + 0.004950750153511763, + 0.005796792916953564, + 0.008280821144580841, + 0.009345046244561672, + 0.009293642826378345, + 0.007468685507774353, + 0.012196103110909462, + 0.014335820451378822, + 0.007791185285896063, + 0.01272251084446907, + 0.0182506050914526, + 0.023079048842191696, + 0.02004111185669899, + 0.016031265258789062, + 0.027989426627755165, + 0.03122660703957081, + 0.02790340594947338, + 0.024455860257148743, + 0.0027963793836534023, + 0.003864970989525318, + 0.0050743501633405685, + 0.002568705938756466, + 0.003823072649538517, + 0.002199843293055892, + 0.0022892255801707506, + 0.0033047636970877647, + 0.012014352716505527, + 0.005559257231652737, + 0.017316250130534172, + 0.00971959624439478, + 0.007985100150108337, + 0.01075043436139822, + 0.012618070468306541, + 0.01768041029572487 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 7, + 9 + ], + "selected_reward": -0.003612422849982977 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene an abstract world composed of vibrant color blocks and bold black lines in a Henri Matisse style?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the background a deconstructed map of the Forbidden City with cadmium yellow rooftops, vermilion walls, and azure/emerald green courtyards?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is there a minimalist boat with a pure ultramarine blue hull and black calligraphic strokes gliding across the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the boat create stylized ripples of simple lines as it moves, without traditional light, shadow, or perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot a smooth pan following the boat's forward movement?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.17983488738536835 + ], + "opposite_logprob_tokens": [ + -1.8048348426818848 + ], + "logprob_reward": -0.17983488738536835, + "yes_no_margin_reward": 1.625, + "token_credit_reward": -0.17983488738536835, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.18025998771190643, + 0.07662925869226456, + 0.04082693159580231, + 0.10257140547037125, + 0.04849197342991829, + 0.049261778593063354, + 0.06730543076992035, + 0.34341561794281006, + 0.04609007015824318, + 0.045147497206926346 + ], + "credit_map": [ + 0.02404686249792576, + 0.020008813589811325, + 0.023944076150655746, + 0.024056531488895416, + 0.01859605684876442, + 0.020919792354106903, + 0.026522083207964897, + 0.022165779024362564, + 0.007329883985221386, + 0.005823000334203243, + 0.014278188347816467, + 0.010545386001467705, + 0.011752183549106121, + 0.007285822182893753, + 0.009516797959804535, + 0.010098002851009369, + 0.003982891794294119, + 0.006153264082968235, + 0.004831823520362377, + 0.004057316575199366, + 0.00507726427167654, + 0.0062471069395542145, + 0.004332901909947395, + 0.0061443643644452095, + 0.006654908414930105, + 0.008879409171640873, + 0.012833785265684128, + 0.021580355241894722, + 0.006320019718259573, + 0.005489620380103588, + 0.02481280267238617, + 0.01600051298737526, + 0.005899115931242704, + 0.004974191077053547, + 0.009219586849212646, + 0.004566452000290155, + 0.009096483699977398, + 0.003773302538320422, + 0.005362419877201319, + 0.005600422155112028, + 0.006504382006824017, + 0.004648929461836815, + 0.006308570969849825, + 0.007915875874459743, + 0.005960486363619566, + 0.006073803175240755, + 0.005762811750173569, + 0.006086922250688076, + 0.0076295859180390835, + 0.006975132972002029, + 0.010584264062345028, + 0.005644278135150671, + 0.007026360370218754, + 0.011071575805544853, + 0.007711377926170826, + 0.010662856511771679, + 0.04429745301604271, + 0.040251921862363815, + 0.028928186744451523, + 0.0210198312997818, + 0.057453278452157974, + 0.0453726202249527, + 0.03282380476593971, + 0.07326853275299072, + 0.004682461731135845, + 0.005941241979598999, + 0.005550390109419823, + 0.011002580635249615, + 0.005340874660760164, + 0.0034451503306627274, + 0.005085424520075321, + 0.005041949916630983, + 0.004058776888996363, + 0.004317843820899725, + 0.005407316610217094, + 0.006024096626788378, + 0.004875605925917625, + 0.004696613177657127, + 0.006703417748212814, + 0.00906382780522108 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 0, + 3 + ], + "selected_reward": -0.17983488738536835 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "smoothly panning drone shot follows a small boat as it navigates an abstract world composed entirely of vibrant blocks of color and bold black lines, à la Henri Matisse. The image below unfolds not as a physical body of water, but rather as a vast, deconstructed map of the Forbidden City complex, composed of large, flat areas of color: cadmium yellow representing palace rooftops, vermilion areas symbolizing city walls, and geometric shapes of azure and emerald green representing courtyards, all clearly separated by decisive, expressive black outlines. Across this sea of color, a symbolic vessel, minimalist and defined by a few black, calligraphic strokes, its hull a pure block of ultramarine blue, glides forward with a steady, rhythmic rhythm. Each stroke of the oars creates stylized ripples of simple lines across the flat surface of color, a movement of grace and deliberation. The entire picture is free from traditional light and shadow and perspective, presenting a purely flat decorative beauty. The light is evenly distributed on each color block, creating a tranquil, dreamy and artistically meaningful atmosphere, like a symbolic journey through history and modern art.", + "seed": 23461, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000006/group_2/index_1599_seed_23461_dp_4_rank_8_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 2, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.04092226237989962, + "vlm_logprob": -0.04092226237989962, + "vlm_yes_no_margin": 4.775, + "vlm_token_credit": -0.04092226237989962, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene an abstract world composed of vibrant color blocks and bold black lines in a Henri Matisse style?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the background a deconstructed map of the Forbidden City with cadmium yellow rooftops, vermilion walls, and azure/emerald green courtyards?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is there a minimalist boat with a pure ultramarine blue hull and black calligraphic strokes gliding across the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the boat create stylized ripples of simple lines as it moves, without traditional light, shadow, or perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot a smooth pan following the boat's forward movement?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.002097907941788435 + ], + "opposite_logprob_tokens": [ + -7.002098083496094 + ], + "logprob_reward": -0.002097907941788435, + "yes_no_margin_reward": 7.0, + "token_credit_reward": -0.002097907941788435, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14745816588401794, + 0.07526827603578568, + 0.08787641674280167, + 0.03519471734762192, + 0.11174483597278595, + 0.15129320323467255, + 0.1331719011068344, + 0.13185769319534302, + 0.05527803674340248, + 0.07085669040679932 + ], + "credit_map": [ + 0.023477213457226753, + 0.025675304234027863, + 0.017329562455415726, + 0.014041277579963207, + 0.02673502452671528, + 0.012926115654408932, + 0.010609807446599007, + 0.01666385494172573, + 0.008846646174788475, + 0.011931400746107101, + 0.006388600450009108, + 0.007236533798277378, + 0.008683334104716778, + 0.01012942660599947, + 0.0050851390697062016, + 0.01696719229221344, + 0.010706113651394844, + 0.011764397844672203, + 0.00808463990688324, + 0.008743892423808575, + 0.012845848686993122, + 0.01169662456959486, + 0.009097558446228504, + 0.01493733562529087, + 0.00346832862123847, + 0.003777578705921769, + 0.0032240122091025114, + 0.005962612573057413, + 0.003503544721752405, + 0.0050490680150687695, + 0.004095369949936867, + 0.006114202085882425, + 0.015819473192095757, + 0.006961551494896412, + 0.03387265279889107, + 0.014679682441055775, + 0.00763359572738409, + 0.01073108147829771, + 0.0072846501134335995, + 0.014762149192392826, + 0.020030418410897255, + 0.012917923741042614, + 0.0335860401391983, + 0.025870306417346, + 0.010374568402767181, + 0.01432832796126604, + 0.015022875741124153, + 0.01916273683309555, + 0.01490817405283451, + 0.025149669498205185, + 0.017111731693148613, + 0.012671767733991146, + 0.025979170575737953, + 0.011382954195141792, + 0.005627328995615244, + 0.02034110762178898, + 0.013357576914131641, + 0.018706195056438446, + 0.01692093536257744, + 0.024353589862585068, + 0.020487213507294655, + 0.017778471112251282, + 0.00890359841287136, + 0.011350126937031746, + 0.0072657261043787, + 0.006864994298666716, + 0.005941914394497871, + 0.00845580454915762, + 0.0072959610261023045, + 0.0054969568736851215, + 0.006452715490013361, + 0.007503959350287914, + 0.007430174853652716, + 0.005852391477674246, + 0.011028492823243141, + 0.012341566383838654, + 0.0062489500269293785, + 0.007137414999306202, + 0.008976370096206665, + 0.01184132695198059 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 0, + 6 + ], + "selected_reward": -0.002097907941788435 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene an abstract world composed of vibrant color blocks and bold black lines in a Henri Matisse style?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the background a deconstructed map of the Forbidden City with cadmium yellow rooftops, vermilion walls, and azure/emerald green courtyards?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is there a minimalist boat with a pure ultramarine blue hull and black calligraphic strokes gliding across the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the boat create stylized ripples of simple lines as it moves, without traditional light, shadow, or perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot a smooth pan following the boat's forward movement?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.012561134994029999 + ], + "opposite_logprob_tokens": [ + -4.387561321258545 + ], + "logprob_reward": -0.012561134994029999, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.012561134994029999, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2657870054244995, + 0.0940031185746193, + 0.1987895518541336, + 0.0513039231300354, + 0.06583673506975174, + 0.07726383954286575, + 0.08366167545318604, + 0.06616993248462677, + 0.03708520904183388, + 0.06009897589683533 + ], + "credit_map": [ + 0.04344968497753143, + 0.024180244654417038, + 0.03575030714273453, + 0.03688524663448334, + 0.0369802787899971, + 0.03837382793426514, + 0.025856178253889084, + 0.02431124821305275, + 0.010549189522862434, + 0.01770245097577572, + 0.006872177589684725, + 0.007137954235076904, + 0.012394442223012447, + 0.014249157160520554, + 0.0064162323251366615, + 0.018681515008211136, + 0.014253964647650719, + 0.0156855545938015, + 0.021008118987083435, + 0.013439597561955452, + 0.052410319447517395, + 0.03754089027643204, + 0.01622808538377285, + 0.02822304517030716, + 0.0027044068556278944, + 0.003708773758262396, + 0.006236235611140728, + 0.005012197885662317, + 0.005891538690775633, + 0.006042353343218565, + 0.004514632746577263, + 0.017193788662552834, + 0.006995393428951502, + 0.005502146668732166, + 0.01287130732089281, + 0.00780859962105751, + 0.00535437511280179, + 0.007917970418930054, + 0.005058059468865395, + 0.014328891411423683, + 0.010378020815551281, + 0.007945560850203037, + 0.011651306413114071, + 0.013322893530130386, + 0.006615230347961187, + 0.006662510335445404, + 0.010134774260222912, + 0.01055354718118906, + 0.012889494188129902, + 0.01453804224729538, + 0.009411402978003025, + 0.010697931982576847, + 0.01264926791191101, + 0.005986731965094805, + 0.0041158078238368034, + 0.013373004272580147, + 0.0123108746483922, + 0.010684509761631489, + 0.008005284704267979, + 0.010888496413826942, + 0.007747463416308165, + 0.004787650424987078, + 0.004009065683931112, + 0.0077365911565721035, + 0.00600013229995966, + 0.004159227944910526, + 0.003772458992898464, + 0.006547850091010332, + 0.0044230008497834206, + 0.0030681302305310965, + 0.004361498635262251, + 0.004752913024276495, + 0.005498289130628109, + 0.003065550932660699, + 0.014165609143674374, + 0.01007875520735979, + 0.0036058162804692984, + 0.005901767872273922, + 0.011236976832151413, + 0.006546215619891882 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 1 + ], + "selected_reward": -0.012561134994029999 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene an abstract world composed of vibrant color blocks and bold black lines in a Henri Matisse style?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the background a deconstructed map of the Forbidden City with cadmium yellow rooftops, vermilion walls, and azure/emerald green courtyards?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is there a minimalist boat with a pure ultramarine blue hull and black calligraphic strokes gliding across the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the boat create stylized ripples of simple lines as it moves, without traditional light, shadow, or perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot a smooth pan following the boat's forward movement?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.007633563131093979 + ], + "opposite_logprob_tokens": [ + -4.882633686065674 + ], + "logprob_reward": -0.007633563131093979, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.007633563131093979, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.16939722001552582, + 0.04347361624240875, + 0.3482309877872467, + 0.04721171408891678, + 0.11219741404056549, + 0.07290703058242798, + 0.049255628138780594, + 0.05514400079846382, + 0.05308303236961365, + 0.04909931495785713 + ], + "credit_map": [ + 0.022635042667388916, + 0.019224287942051888, + 0.02116643637418747, + 0.016517171636223793, + 0.02319713495671749, + 0.02636118046939373, + 0.015279431827366352, + 0.0250165406614542, + 0.007606942672282457, + 0.007190903183072805, + 0.003816417418420315, + 0.004842138383537531, + 0.005295938812196255, + 0.005447708535939455, + 0.0030453484505414963, + 0.006228222977370024, + 0.036689866334199905, + 0.04532777518033981, + 0.036722876131534576, + 0.03448844701051712, + 0.06872649490833282, + 0.053592197597026825, + 0.029631566256284714, + 0.04305179789662361, + 0.005881453398615122, + 0.004213806241750717, + 0.006049650721251965, + 0.006584196351468563, + 0.004829778335988522, + 0.006057684309780598, + 0.004122943617403507, + 0.009472203440964222, + 0.015571489930152893, + 0.006284685805439949, + 0.03164447471499443, + 0.015904434025287628, + 0.005957538262009621, + 0.009589106775820255, + 0.00695745600387454, + 0.020288242027163506, + 0.011187488213181496, + 0.0062148612923920155, + 0.011576234363019466, + 0.012949749827384949, + 0.006301400251686573, + 0.006963563151657581, + 0.009113608859479427, + 0.008600139059126377, + 0.006508991587907076, + 0.0035707505885511637, + 0.005994406994432211, + 0.0055261217057704926, + 0.005738737061619759, + 0.007277580443769693, + 0.0032114938367158175, + 0.01142754964530468, + 0.014035230502486229, + 0.007576270028948784, + 0.006151976063847542, + 0.00649360753595829, + 0.00634212838485837, + 0.004810268525034189, + 0.0036735928151756525, + 0.006060935091227293, + 0.009490491822361946, + 0.0057772258296608925, + 0.0061413622461259365, + 0.010595825500786304, + 0.004625079687684774, + 0.003661674913018942, + 0.006611178629100323, + 0.006180199328809977, + 0.004429399501532316, + 0.0037063707131892443, + 0.011354422196745872, + 0.006864064838737249, + 0.0034559560008347034, + 0.005111583974212408, + 0.007797319907695055, + 0.0063802022486925125 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 0, + 4 + ], + "selected_reward": -0.007633563131093979 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene an abstract world composed of vibrant color blocks and bold black lines in a Henri Matisse style?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the background a deconstructed map of the Forbidden City with cadmium yellow rooftops, vermilion walls, and azure/emerald green courtyards?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is there a minimalist boat with a pure ultramarine blue hull and black calligraphic strokes gliding across the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the boat create stylized ripples of simple lines as it moves, without traditional light, shadow, or perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot a smooth pan following the boat's forward movement?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0024873053189367056 + ], + "opposite_logprob_tokens": [ + -6.0024871826171875 + ], + "logprob_reward": -0.0024873053189367056, + "yes_no_margin_reward": 6.0, + "token_credit_reward": -0.0024873053189367056, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1473030000925064, + 0.061070870608091354, + 0.1357840746641159, + 0.03318975493311882, + 0.061920881271362305, + 0.08344486355781555, + 0.07881718873977661, + 0.09450028836727142, + 0.14082185924053192, + 0.1631472259759903 + ], + "credit_map": [ + 0.024082116782665253, + 0.014953876845538616, + 0.016007818281650543, + 0.015601875260472298, + 0.028009597212076187, + 0.014208255335688591, + 0.017151353880763054, + 0.017288116738200188, + 0.009739408269524574, + 0.007351291365921497, + 0.005971693433821201, + 0.008671889081597328, + 0.006261974573135376, + 0.008634188212454319, + 0.00394729059189558, + 0.010493135079741478, + 0.016486717388033867, + 0.013948232866823673, + 0.02273150160908699, + 0.019719259813427925, + 0.022655796259641647, + 0.013674262911081314, + 0.012553349137306213, + 0.014014962129294872, + 0.002862833673134446, + 0.00342855928465724, + 0.003616166999563575, + 0.005120842717587948, + 0.005313596222549677, + 0.005380311980843544, + 0.0037716832011938095, + 0.003695762250572443, + 0.007556264754384756, + 0.00423487788066268, + 0.017042426392436028, + 0.008125846274197102, + 0.003792368108406663, + 0.00744579778984189, + 0.0044781360775232315, + 0.009245163761079311, + 0.014458166435360909, + 0.008160091005265713, + 0.012584084644913673, + 0.015266062691807747, + 0.007503936067223549, + 0.006888554897159338, + 0.008957949467003345, + 0.009626023471355438, + 0.008488694205880165, + 0.011706901714205742, + 0.008500783704221249, + 0.006673660594969988, + 0.016844401136040688, + 0.011172686703503132, + 0.004326803144067526, + 0.011103258468210697, + 0.008988364599645138, + 0.013539120554924011, + 0.010426342487335205, + 0.01593022048473358, + 0.015582989901304245, + 0.016351234167814255, + 0.0052828737534582615, + 0.008399134501814842, + 0.03664952516555786, + 0.01328484620898962, + 0.017901139333844185, + 0.018661828711628914, + 0.012710176408290863, + 0.0134933702647686, + 0.013525181449949741, + 0.014595791697502136, + 0.018245365470647812, + 0.015006671659648418, + 0.024014025926589966, + 0.030061082914471626, + 0.013824975118041039, + 0.012902890332043171, + 0.024158043786883354, + 0.024934174492955208 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 0, + 8 + ], + "selected_reward": -0.0024873053189367056 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene an abstract world composed of vibrant color blocks and bold black lines in a Henri Matisse style?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the background a deconstructed map of the Forbidden City with cadmium yellow rooftops, vermilion walls, and azure/emerald green courtyards?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is there a minimalist boat with a pure ultramarine blue hull and black calligraphic strokes gliding across the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the boat create stylized ripples of simple lines as it moves, without traditional light, shadow, or perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot a smooth pan following the boat's forward movement?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.179831400513649 + ], + "opposite_logprob_tokens": [ + -1.8048313856124878 + ], + "logprob_reward": -0.179831400513649, + "yes_no_margin_reward": 1.625, + "token_credit_reward": -0.179831400513649, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12600077688694, + 0.0674791932106018, + 0.1796310842037201, + 0.07600359618663788, + 0.14061616361141205, + 0.14194494485855103, + 0.0886542871594429, + 0.06466749310493469, + 0.06487803906202316, + 0.050124481320381165 + ], + "credit_map": [ + 0.018612239509820938, + 0.013740108348429203, + 0.0173436738550663, + 0.01193644106388092, + 0.016759753227233887, + 0.022678934037685394, + 0.011112488806247711, + 0.013817132450640202, + 0.010964768007397652, + 0.009175071492791176, + 0.007046577986329794, + 0.008442387916147709, + 0.008832328021526337, + 0.00783487781882286, + 0.004251881502568722, + 0.010931297205388546, + 0.02285555750131607, + 0.017157018184661865, + 0.02184361219406128, + 0.029970534145832062, + 0.025856461375951767, + 0.019762590527534485, + 0.020098313689231873, + 0.02208697237074375, + 0.007358020171523094, + 0.006634166929870844, + 0.014190121553838253, + 0.007940178737044334, + 0.009302317164838314, + 0.010247414000332355, + 0.006380124483257532, + 0.013951249420642853, + 0.019869118928909302, + 0.007140158675611019, + 0.04199586436152458, + 0.01857537031173706, + 0.007658143062144518, + 0.012501207180321217, + 0.00848014373332262, + 0.024396145716309547, + 0.01724839210510254, + 0.01950456015765667, + 0.02474963665008545, + 0.024155857041478157, + 0.012574327178299427, + 0.013481509871780872, + 0.013990078121423721, + 0.016240576282143593, + 0.011669646948575974, + 0.007412501145154238, + 0.013886582106351852, + 0.013568567112088203, + 0.007054000627249479, + 0.009161031804978848, + 0.0055139693431556225, + 0.020387982949614525, + 0.007702181115746498, + 0.012370471842586994, + 0.00712271872907877, + 0.008517510257661343, + 0.007616657763719559, + 0.009400324895977974, + 0.004018065519630909, + 0.007919558323919773, + 0.008976360782980919, + 0.006014612503349781, + 0.007680181879550219, + 0.0131046362221241, + 0.008304575458168983, + 0.005876914598047733, + 0.0064584631472826, + 0.008462291210889816, + 0.004303645342588425, + 0.0037066421937197447, + 0.01003783568739891, + 0.0066404324024915695, + 0.005346669815480709, + 0.0057870433665812016, + 0.007094816770404577, + 0.0072073908522725105 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 5, + 4 + ], + "selected_reward": -0.179831400513649 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "smoothly panning drone shot follows a small boat as it navigates an abstract world composed entirely of vibrant blocks of color and bold black lines, à la Henri Matisse. The image below unfolds not as a physical body of water, but rather as a vast, deconstructed map of the Forbidden City complex, composed of large, flat areas of color: cadmium yellow representing palace rooftops, vermilion areas symbolizing city walls, and geometric shapes of azure and emerald green representing courtyards, all clearly separated by decisive, expressive black outlines. Across this sea of color, a symbolic vessel, minimalist and defined by a few black, calligraphic strokes, its hull a pure block of ultramarine blue, glides forward with a steady, rhythmic rhythm. Each stroke of the oars creates stylized ripples of simple lines across the flat surface of color, a movement of grace and deliberation. The entire picture is free from traditional light and shadow and perspective, presenting a purely flat decorative beauty. The light is evenly distributed on each color block, creating a tranquil, dreamy and artistically meaningful atmosphere, like a symbolic journey through history and modern art.", + "seed": 23462, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000006/group_2/index_1599_seed_23462_dp_5_rank_10_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 2, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.026456461800262333, + "vlm_logprob": -0.026456461800262333, + "vlm_yes_no_margin": 4.8, + "vlm_token_credit": -0.026456461800262333, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene an abstract world composed of vibrant color blocks and bold black lines in a Henri Matisse style?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the background a deconstructed map of the Forbidden City with cadmium yellow rooftops, vermilion walls, and azure/emerald green courtyards?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is there a minimalist boat with a pure ultramarine blue hull and black calligraphic strokes gliding across the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the boat create stylized ripples of simple lines as it moves, without traditional light, shadow, or perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot a smooth pan following the boat's forward movement?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0031207927968353033 + ], + "opposite_logprob_tokens": [ + -6.7531208992004395 + ], + "logprob_reward": -0.0031207927968353033, + "yes_no_margin_reward": 6.75, + "token_credit_reward": -0.0031207927968353033, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.04417414218187332, + 0.11415775120258331, + 0.05032152310013771, + 0.05938125401735306, + 0.017773592844605446, + 0.045231062918901443, + 0.5448374152183533, + 0.03542483225464821, + 0.04773782938718796, + 0.04096056893467903 + ], + "credit_map": [ + 0.005776228848844767, + 0.003343892516568303, + 0.004363212734460831, + 0.008147124201059341, + 0.004115674179047346, + 0.003938617184758186, + 0.005763604771345854, + 0.008725782856345177, + 0.017413266003131866, + 0.009644047357141972, + 0.010700919665396214, + 0.02617097645998001, + 0.008163833059370518, + 0.00819076132029295, + 0.017105456441640854, + 0.016768481582403183, + 0.012419945560395718, + 0.014415965415537357, + 0.0033033008221536875, + 0.004053897690027952, + 0.003440016880631447, + 0.0016756333643570542, + 0.004542900715023279, + 0.006469855550676584, + 0.016472209244966507, + 0.010500929318368435, + 0.008372276090085506, + 0.00421911058947444, + 0.00510048121213913, + 0.003042736556380987, + 0.006014552433043718, + 0.005658954847604036, + 0.002250388264656067, + 0.0018313744803890586, + 0.0016338340938091278, + 0.0024360704701393843, + 0.0014399882638826966, + 0.0013221196131780744, + 0.0040651713497936726, + 0.002794646192342043, + 0.01341682206839323, + 0.0028901079203933477, + 0.002129378030076623, + 0.0027698155026882887, + 0.0041874488815665245, + 0.004137156065553427, + 0.011031024158000946, + 0.004669307731091976, + 0.024652613326907158, + 0.23122146725654602, + 0.01689988188445568, + 0.01993725076317787, + 0.02990560047328472, + 0.17442120611667633, + 0.017284339293837547, + 0.03051506169140339, + 0.003059318521991372, + 0.002865756396204233, + 0.002057003788650036, + 0.0036020195111632347, + 0.007036881987005472, + 0.0070005967281758785, + 0.004044962581247091, + 0.005758289713412523, + 0.0043181730434298515, + 0.014117594808340073, + 0.00430002948269248, + 0.005117905791848898, + 0.004332044627517462, + 0.003711050609126687, + 0.006928782444447279, + 0.004912242759019136, + 0.0061883097514510155, + 0.004253683611750603, + 0.005435955710709095, + 0.0048005543649196625, + 0.003485433291643858, + 0.005422609392553568, + 0.0067810495384037495, + 0.004592971410602331 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 1, + 3 + ], + "selected_reward": -0.0031207927968353033 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene an abstract world composed of vibrant color blocks and bold black lines in a Henri Matisse style?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the background a deconstructed map of the Forbidden City with cadmium yellow rooftops, vermilion walls, and azure/emerald green courtyards?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is there a minimalist boat with a pure ultramarine blue hull and black calligraphic strokes gliding across the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the boat create stylized ripples of simple lines as it moves, without traditional light, shadow, or perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot a smooth pan following the boat's forward movement?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03373256325721741 + ], + "opposite_logprob_tokens": [ + -3.4087326526641846 + ], + "logprob_reward": -0.03373256325721741, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.03373256325721741, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1213182881474495, + 0.03789578005671501, + 0.08442465215921402, + 0.033272627741098404, + 0.05968732759356499, + 0.09778761863708496, + 0.36567074060440063, + 0.03353440761566162, + 0.030853286385536194, + 0.13555534183979034 + ], + "credit_map": [ + 0.011453848332166672, + 0.0076053813099861145, + 0.011640832759439945, + 0.013646812178194523, + 0.007822645828127861, + 0.010401502251625061, + 0.017900118604302406, + 0.040847133845090866, + 0.0053826021030545235, + 0.0033357006032019854, + 0.004587170202285051, + 0.005140821915119886, + 0.0035268354695290327, + 0.003828617511317134, + 0.005640467628836632, + 0.006453557871282101, + 0.014076702296733856, + 0.01871812343597412, + 0.00719854049384594, + 0.005851705092936754, + 0.0108495419844985, + 0.006895800586789846, + 0.011053293943405151, + 0.009780935943126678, + 0.0042038653045892715, + 0.007383392192423344, + 0.004107913002371788, + 0.003074102569371462, + 0.0043571097776293755, + 0.002716448623687029, + 0.004968356806784868, + 0.002461435738950968, + 0.007868916727602482, + 0.005784475710242987, + 0.004007128532975912, + 0.01148148998618126, + 0.0050721545703709126, + 0.0065838065929710865, + 0.012792467139661312, + 0.0060968827456235886, + 0.026983939111232758, + 0.006400455720722675, + 0.004951593466103077, + 0.006457941140979528, + 0.010126148350536823, + 0.007142600603401661, + 0.027501340955495834, + 0.008223584853112698, + 0.02178158238530159, + 0.1195850521326065, + 0.02593507431447506, + 0.031381551176309586, + 0.02246107906103134, + 0.06885385513305664, + 0.027476489543914795, + 0.04819602891802788, + 0.003119595116004348, + 0.003180490341037512, + 0.002892425050958991, + 0.0035095184575766325, + 0.007113284431397915, + 0.005776154808700085, + 0.003210260532796383, + 0.004732673987746239, + 0.003092869184911251, + 0.004169297870248556, + 0.004232958424836397, + 0.004721039440482855, + 0.0034778700210154057, + 0.0032571631018072367, + 0.003974434919655323, + 0.003927652258425951, + 0.016661757603287697, + 0.01034799125045538, + 0.024345282465219498, + 0.019325848668813705, + 0.012062749825417995, + 0.018319619819521904, + 0.019993921741843224, + 0.014498160220682621 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 9, + 0 + ], + "selected_reward": -0.03373256325721741 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene an abstract world composed of vibrant color blocks and bold black lines in a Henri Matisse style?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the background a deconstructed map of the Forbidden City with cadmium yellow rooftops, vermilion walls, and azure/emerald green courtyards?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is there a minimalist boat with a pure ultramarine blue hull and black calligraphic strokes gliding across the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the boat create stylized ripples of simple lines as it moves, without traditional light, shadow, or perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot a smooth pan following the boat's forward movement?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0036168177612125874 + ], + "opposite_logprob_tokens": [ + -5.628616809844971 + ], + "logprob_reward": -0.0036168177612125874, + "yes_no_margin_reward": 5.625, + "token_credit_reward": -0.0036168177612125874, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07683338969945908, + 0.431424081325531, + 0.09778237342834473, + 0.023279612883925438, + 0.010197081603109837, + 0.049164921045303345, + 0.24561180174350739, + 0.02343720942735672, + 0.014483192004263401, + 0.02778635174036026 + ], + "credit_map": [ + 0.010085228830575943, + 0.005627125967293978, + 0.006593223661184311, + 0.011151810176670551, + 0.00498639652505517, + 0.005887183360755444, + 0.011342779733240604, + 0.021159643307328224, + 0.07118259370326996, + 0.03856462985277176, + 0.039211124181747437, + 0.1051117479801178, + 0.030236518010497093, + 0.0321558341383934, + 0.059664737433195114, + 0.05529690533876419, + 0.018346963450312614, + 0.024274930357933044, + 0.006907770410180092, + 0.004654577001929283, + 0.01417741272598505, + 0.007031934801489115, + 0.009656514041125774, + 0.012732263654470444, + 0.004772954620420933, + 0.0022617096547037363, + 0.004723998252302408, + 0.002388833789154887, + 0.0017914911732077599, + 0.001519565237686038, + 0.0031642126850783825, + 0.002656846074387431, + 0.001366649754345417, + 0.0013108465354889631, + 0.001051414874382317, + 0.001277587260119617, + 0.0010030949488282204, + 0.0011869381414726377, + 0.001647454802878201, + 0.001353095518425107, + 0.012472840957343578, + 0.002634408650919795, + 0.002462126547470689, + 0.0031341020949184895, + 0.004306493792682886, + 0.0037773754447698593, + 0.016207102686166763, + 0.004170471336692572, + 0.011582322418689728, + 0.1014632061123848, + 0.008340245112776756, + 0.01056718174368143, + 0.014188416302204132, + 0.0744204893708229, + 0.010391214862465858, + 0.014658714644610882, + 0.0022979869972914457, + 0.002146010519936681, + 0.0016989227151498199, + 0.00206014490686357, + 0.004486855585128069, + 0.004160325974225998, + 0.0021709054708480835, + 0.004416055977344513, + 0.0010953366290777922, + 0.0018232606817036867, + 0.0018779400270432234, + 0.0014472291804850101, + 0.001430118572898209, + 0.0015131898690015078, + 0.0029827149119228125, + 0.002313402248546481, + 0.004574574530124664, + 0.0025946921668946743, + 0.003278449410572648, + 0.0033654889557510614, + 0.0025412568356841803, + 0.003955339081585407, + 0.0038777964655309916, + 0.0035987519659101963 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 6, + 2 + ], + "selected_reward": -0.0036168177612125874 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene an abstract world composed of vibrant color blocks and bold black lines in a Henri Matisse style?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the background a deconstructed map of the Forbidden City with cadmium yellow rooftops, vermilion walls, and azure/emerald green courtyards?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is there a minimalist boat with a pure ultramarine blue hull and black calligraphic strokes gliding across the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the boat create stylized ripples of simple lines as it moves, without traditional light, shadow, or perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot a smooth pan following the boat's forward movement?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.002816402819007635 + ], + "opposite_logprob_tokens": [ + -5.877816200256348 + ], + "logprob_reward": -0.002816402819007635, + "yes_no_margin_reward": 5.875, + "token_credit_reward": -0.002816402819007635, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14354337751865387, + 0.08331859111785889, + 0.036759886890649796, + 0.03668344393372536, + 0.02464231662452221, + 0.03893071413040161, + 0.5504239797592163, + 0.03776136040687561, + 0.025483226403594017, + 0.02245319075882435 + ], + "credit_map": [ + 0.025216665118932724, + 0.008740353398025036, + 0.016874827444553375, + 0.042857516556978226, + 0.00729419244453311, + 0.007425711955875158, + 0.016569845378398895, + 0.018564263358712196, + 0.016308920457959175, + 0.009453414008021355, + 0.007468230556696653, + 0.022823985666036606, + 0.004783692304044962, + 0.004511786159127951, + 0.010851124301552773, + 0.0071174283511936665, + 0.005022967234253883, + 0.004061330575495958, + 0.004864762537181377, + 0.0032231395598500967, + 0.0033492797520011663, + 0.005336114205420017, + 0.006651566829532385, + 0.004250723868608475, + 0.009561530314385891, + 0.004116271622478962, + 0.006259148009121418, + 0.0034650301095098257, + 0.0033122552558779716, + 0.0020741838961839676, + 0.005271406844258308, + 0.002623613690957427, + 0.002537604421377182, + 0.0019714643713086843, + 0.002025832887738943, + 0.00635757390409708, + 0.001315536443144083, + 0.0027105228509753942, + 0.004274951759725809, + 0.0034488271921873093, + 0.005095669999718666, + 0.003554085735231638, + 0.003751038806512952, + 0.0034923136699944735, + 0.003922013100236654, + 0.005011982284486294, + 0.010612381622195244, + 0.0034912293776869774, + 0.026220250874757767, + 0.24213463068008423, + 0.011455941013991833, + 0.016982551664114, + 0.028928177431225777, + 0.16962392628192902, + 0.01102826464921236, + 0.04405014216899872, + 0.003053132677450776, + 0.00257267733104527, + 0.002376358024775982, + 0.0022667290177196264, + 0.010462208651006222, + 0.006302459165453911, + 0.0031368809286504984, + 0.007590909022837877, + 0.0023073724005371332, + 0.004003488924354315, + 0.004113205708563328, + 0.003443240886554122, + 0.0025439613964408636, + 0.002141382312402129, + 0.003993513528257608, + 0.0029370582196861506, + 0.003974398132413626, + 0.002408712636679411, + 0.002113804453983903, + 0.001921233139000833, + 0.0020666548516601324, + 0.0027453936636447906, + 0.0045354487374424934, + 0.0026875431649386883 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 0, + 1 + ], + "selected_reward": -0.002816402819007635 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene an abstract world composed of vibrant color blocks and bold black lines in a Henri Matisse style?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the background a deconstructed map of the Forbidden City with cadmium yellow rooftops, vermilion walls, and azure/emerald green courtyards?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is there a minimalist boat with a pure ultramarine blue hull and black calligraphic strokes gliding across the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the boat create stylized ripples of simple lines as it moves, without traditional light, shadow, or perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot a smooth pan following the boat's forward movement?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.08899573236703873 + ], + "opposite_logprob_tokens": [ + -2.4639956951141357 + ], + "logprob_reward": -0.08899573236703873, + "yes_no_margin_reward": 2.375, + "token_credit_reward": -0.08899573236703873, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1131773293018341, + 0.15617604553699493, + 0.07812682539224625, + 0.08820328116416931, + 0.027187280356884003, + 0.03736470639705658, + 0.3614750802516937, + 0.036865003407001495, + 0.0632999911904335, + 0.03812440112233162 + ], + "credit_map": [ + 0.01894887536764145, + 0.006670412141829729, + 0.01275752391666174, + 0.031733203679323196, + 0.005526164546608925, + 0.004807576537132263, + 0.014400957152247429, + 0.01833261549472809, + 0.022660525515675545, + 0.013294865377247334, + 0.018621306866407394, + 0.03316745162010193, + 0.012139663100242615, + 0.012748565524816513, + 0.021120447665452957, + 0.022423231974244118, + 0.018737399950623512, + 0.022749025374650955, + 0.0053897215984761715, + 0.007053795736283064, + 0.0066679129377007484, + 0.0038322664331644773, + 0.006948087364435196, + 0.006748616229742765, + 0.021311111748218536, + 0.015872864052653313, + 0.014426214620471, + 0.006361272651702166, + 0.009339636191725731, + 0.005498010199517012, + 0.009042012505233288, + 0.006352158263325691, + 0.002646341919898987, + 0.002223801100626588, + 0.002159612951800227, + 0.004555477295070887, + 0.002058479469269514, + 0.002479333896189928, + 0.006966153159737587, + 0.00409808149561286, + 0.01060489285737276, + 0.0028328944463282824, + 0.002247472759336233, + 0.0022370649967342615, + 0.002904612571001053, + 0.0030582398176193237, + 0.010245779529213905, + 0.0032337477896362543, + 0.017139995470643044, + 0.13959570229053497, + 0.01364809088408947, + 0.015382825396955013, + 0.03129299730062485, + 0.11184358596801758, + 0.015815479680895805, + 0.016756409779191017, + 0.0035841455683112144, + 0.0031953800935298204, + 0.0032929894514381886, + 0.003678827313706279, + 0.006183118559420109, + 0.00666626775637269, + 0.004266167059540749, + 0.005998107139021158, + 0.00592086324468255, + 0.0057118735276162624, + 0.009273028932511806, + 0.007842417806386948, + 0.006892230361700058, + 0.007980295456945896, + 0.00976463221013546, + 0.009914644062519073, + 0.006985070183873177, + 0.003926639445126057, + 0.0048202467150986195, + 0.0038621944840997458, + 0.0034026294015347958, + 0.0062270392663776875, + 0.004549977835267782, + 0.004350604023784399 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 1, + 0 + ], + "selected_reward": -0.08899573236703873 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "smoothly panning drone shot follows a small boat as it navigates an abstract world composed entirely of vibrant blocks of color and bold black lines, à la Henri Matisse. The image below unfolds not as a physical body of water, but rather as a vast, deconstructed map of the Forbidden City complex, composed of large, flat areas of color: cadmium yellow representing palace rooftops, vermilion areas symbolizing city walls, and geometric shapes of azure and emerald green representing courtyards, all clearly separated by decisive, expressive black outlines. Across this sea of color, a symbolic vessel, minimalist and defined by a few black, calligraphic strokes, its hull a pure block of ultramarine blue, glides forward with a steady, rhythmic rhythm. Each stroke of the oars creates stylized ripples of simple lines across the flat surface of color, a movement of grace and deliberation. The entire picture is free from traditional light and shadow and perspective, presenting a purely flat decorative beauty. The light is evenly distributed on each color block, creating a tranquil, dreamy and artistically meaningful atmosphere, like a symbolic journey through history and modern art.", + "seed": 23463, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000006/group_2/index_1599_seed_23463_dp_5_rank_10_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 2, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.02959311455488205, + "vlm_logprob": -0.02959311455488205, + "vlm_yes_no_margin": 4.999999976158142, + "vlm_token_credit": -0.02959311455488205, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene an abstract world composed of vibrant color blocks and bold black lines in a Henri Matisse style?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the background a deconstructed map of the Forbidden City with cadmium yellow rooftops, vermilion walls, and azure/emerald green courtyards?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is there a minimalist boat with a pure ultramarine blue hull and black calligraphic strokes gliding across the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the boat create stylized ripples of simple lines as it moves, without traditional light, shadow, or perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot a smooth pan following the boat's forward movement?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.002230182057246566 + ], + "opposite_logprob_tokens": [ + -6.502230167388916 + ], + "logprob_reward": -0.002230182057246566, + "yes_no_margin_reward": 6.5, + "token_credit_reward": -0.002230182057246566, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0923372209072113, + 0.19303369522094727, + 0.058514248579740524, + 0.025475146248936653, + 0.2280399650335312, + 0.06240784749388695, + 0.030249396339058876, + 0.05360794439911842, + 0.06371945142745972, + 0.1926150768995285 + ], + "credit_map": [ + 0.015658969059586525, + 0.02293405309319496, + 0.007138613611459732, + 0.006592750549316406, + 0.011711654253304005, + 0.01070206519216299, + 0.006110131274908781, + 0.011488976888358593, + 0.01928338035941124, + 0.019200840964913368, + 0.009964967146515846, + 0.00844020489603281, + 0.056439947336912155, + 0.05205787718296051, + 0.009943821467459202, + 0.017702646553516388, + 0.008122691884636879, + 0.007001965306699276, + 0.004458856303244829, + 0.005936225410550833, + 0.012927249073982239, + 0.009456410072743893, + 0.005087321158498526, + 0.0055235279724001884, + 0.002424525562673807, + 0.002920796861872077, + 0.0031720842234790325, + 0.0023985914885997772, + 0.004075745586305857, + 0.00446516927331686, + 0.0024205672089010477, + 0.0035976676736027002, + 0.016673294827342033, + 0.013634459115564823, + 0.011138085275888443, + 0.013372705318033695, + 0.028523722663521767, + 0.08777550607919693, + 0.01189386285841465, + 0.04502832517027855, + 0.007929996587336063, + 0.006584047805517912, + 0.004205926787108183, + 0.005721906200051308, + 0.011327759362757206, + 0.006849487777799368, + 0.005810142960399389, + 0.013978577218949795, + 0.0039418949745595455, + 0.0053897639736533165, + 0.0038167869206517935, + 0.003142958739772439, + 0.0024866322055459023, + 0.00612837728112936, + 0.00205467757768929, + 0.0032883051317185163, + 0.0036134999245405197, + 0.0034239275846630335, + 0.006164493039250374, + 0.0038510423619300127, + 0.014585182070732117, + 0.01132231019437313, + 0.004360178019851446, + 0.006287307944148779, + 0.004136779811233282, + 0.0087633291259408, + 0.0029788746032863855, + 0.004203493706882, + 0.0208860095590353, + 0.01767229661345482, + 0.0022931641433387995, + 0.002785508520901203, + 0.0225466750562191, + 0.055023226886987686, + 0.008912910707294941, + 0.015074311755597591, + 0.028736431151628494, + 0.022147223353385925, + 0.01779332011938095, + 0.022380996495485306 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 1, + 9 + ], + "selected_reward": -0.002230182057246566 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene an abstract world composed of vibrant color blocks and bold black lines in a Henri Matisse style?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the background a deconstructed map of the Forbidden City with cadmium yellow rooftops, vermilion walls, and azure/emerald green courtyards?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is there a minimalist boat with a pure ultramarine blue hull and black calligraphic strokes gliding across the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the boat create stylized ripples of simple lines as it moves, without traditional light, shadow, or perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot a smooth pan following the boat's forward movement?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.014211871661245823 + ], + "opposite_logprob_tokens": [ + -4.264211654663086 + ], + "logprob_reward": -0.014211871661245823, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.014211871661245823, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.054211460053920746, + 0.09493748843669891, + 0.04192177206277847, + 0.05632467567920685, + 0.09684059023857117, + 0.09933613985776901, + 0.02350451983511448, + 0.10584089905023575, + 0.0642814040184021, + 0.36280110478401184 + ], + "credit_map": [ + 0.008585960604250431, + 0.00952309649437666, + 0.005288152489811182, + 0.0038489263970404863, + 0.007184040267020464, + 0.006027013994753361, + 0.004930674098432064, + 0.008823595009744167, + 0.009310915134847164, + 0.008245970122516155, + 0.005217109341174364, + 0.005641384050250053, + 0.026000995188951492, + 0.025838280096650124, + 0.005301084369421005, + 0.00938175618648529, + 0.006184163969010115, + 0.005421446170657873, + 0.00382658326998353, + 0.004679364152252674, + 0.006809558719396591, + 0.006035367958247662, + 0.00371625367552042, + 0.005249032750725746, + 0.0072521306574344635, + 0.007669325917959213, + 0.003938727080821991, + 0.00600182730704546, + 0.0071683721616864204, + 0.011596537195146084, + 0.004540076479315758, + 0.00815767515450716, + 0.008982698433101177, + 0.02722262591123581, + 0.004563676659017801, + 0.005597040057182312, + 0.021851610392332077, + 0.013917980715632439, + 0.004214833956211805, + 0.010490114800632, + 0.013448759913444519, + 0.010102460160851479, + 0.007359715178608894, + 0.008469299413263798, + 0.018653078004717827, + 0.009946121834218502, + 0.007920545525848866, + 0.023436158895492554, + 0.0021586394868791103, + 0.0036845197901129723, + 0.002453547203913331, + 0.0025673643685877323, + 0.0032046157866716385, + 0.004097099881619215, + 0.0016418222803622484, + 0.0036969126667827368, + 0.0034895427525043488, + 0.005770176183432341, + 0.00540571054443717, + 0.0055757262744009495, + 0.03904378041625023, + 0.03829983249306679, + 0.003141040215268731, + 0.005115083884447813, + 0.0029462117236107588, + 0.004574055317789316, + 0.0024331274908035994, + 0.0025038840249180794, + 0.02575031854212284, + 0.019875453785061836, + 0.003506952663883567, + 0.002691402332857251, + 0.0429295152425766, + 0.10500426590442657, + 0.01645429991185665, + 0.029224297031760216, + 0.055828630924224854, + 0.04206390678882599, + 0.03223289176821709, + 0.039063286036252975 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 7, + 5 + ], + "selected_reward": -0.014211871661245823 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene an abstract world composed of vibrant color blocks and bold black lines in a Henri Matisse style?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the background a deconstructed map of the Forbidden City with cadmium yellow rooftops, vermilion walls, and azure/emerald green courtyards?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is there a minimalist boat with a pure ultramarine blue hull and black calligraphic strokes gliding across the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the boat create stylized ripples of simple lines as it moves, without traditional light, shadow, or perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot a smooth pan following the boat's forward movement?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.001714071026071906 + ], + "opposite_logprob_tokens": [ + -6.37671422958374 + ], + "logprob_reward": -0.001714071026071906, + "yes_no_margin_reward": 6.375, + "token_credit_reward": -0.001714071026071906, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1469050794839859, + 0.16728442907333374, + 0.051192305982112885, + 0.05368330702185631, + 0.14039084315299988, + 0.03407161310315132, + 0.041563794016838074, + 0.07794259488582611, + 0.21686649322509766, + 0.0700996071100235 + ], + "credit_map": [ + 0.0198033656924963, + 0.014377872459590435, + 0.01794748194515705, + 0.021545732393860817, + 0.01624874770641327, + 0.022123273462057114, + 0.014189167879521847, + 0.02066941373050213, + 0.024036815389990807, + 0.020278189331293106, + 0.010072541423141956, + 0.008069384843111038, + 0.040217913687229156, + 0.03934171423316002, + 0.00885705091059208, + 0.01641080155968666, + 0.004805312491953373, + 0.008098701946437359, + 0.005066067446023226, + 0.006737337913364172, + 0.007222464773803949, + 0.007401123642921448, + 0.004533953964710236, + 0.007327340543270111, + 0.0063332500867545605, + 0.00681728171184659, + 0.0032287882640957832, + 0.0037513021379709244, + 0.00842469371855259, + 0.013819260522723198, + 0.004034914076328278, + 0.007273809053003788, + 0.009872333146631718, + 0.008710469119250774, + 0.007418093737214804, + 0.012534036301076412, + 0.01579224318265915, + 0.047050584107637405, + 0.008407872170209885, + 0.030605191364884377, + 0.0035813050344586372, + 0.0035326755605638027, + 0.004532504826784134, + 0.004845006391406059, + 0.00335173262283206, + 0.004657165613025427, + 0.005141811911016703, + 0.004429410211741924, + 0.004794702865183353, + 0.0054266792722046375, + 0.004577171988785267, + 0.00533345527946949, + 0.0038190451450645924, + 0.009570474736392498, + 0.0029277647845447063, + 0.005114497616887093, + 0.007075757719576359, + 0.005707078613340855, + 0.010326612740755081, + 0.012693156488239765, + 0.014647378586232662, + 0.01485426165163517, + 0.005345008801668882, + 0.007293335627764463, + 0.015558280982077122, + 0.011313769966363907, + 0.006315501872450113, + 0.005857814569026232, + 0.09664487093687057, + 0.06849919259548187, + 0.006565593648701906, + 0.006111457012593746, + 0.006918685976415873, + 0.016174286603927612, + 0.004581252112984657, + 0.007845013402402401, + 0.008441590704023838, + 0.008398385718464851, + 0.010720568709075451, + 0.007019814103841782 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 1, + 0 + ], + "selected_reward": -0.001714071026071906 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene an abstract world composed of vibrant color blocks and bold black lines in a Henri Matisse style?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the background a deconstructed map of the Forbidden City with cadmium yellow rooftops, vermilion walls, and azure/emerald green courtyards?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is there a minimalist boat with a pure ultramarine blue hull and black calligraphic strokes gliding across the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the boat create stylized ripples of simple lines as it moves, without traditional light, shadow, or perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot a smooth pan following the boat's forward movement?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.002816402819007635 + ], + "opposite_logprob_tokens": [ + -5.877816200256348 + ], + "logprob_reward": -0.002816402819007635, + "yes_no_margin_reward": 5.875, + "token_credit_reward": -0.002816402819007635, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07150207459926605, + 0.12200022488832474, + 0.020762447267770767, + 0.041651807725429535, + 0.033827684819698334, + 0.021249057725071907, + 0.08502036333084106, + 0.06479033082723618, + 0.5052502751350403, + 0.033945679664611816 + ], + "credit_map": [ + 0.014605901204049587, + 0.016640447080135345, + 0.006484499666839838, + 0.005704544018954039, + 0.00788589846342802, + 0.006832841318100691, + 0.004400623496621847, + 0.008947332389652729, + 0.018657496199011803, + 0.02951011434197426, + 0.010328373871743679, + 0.007297406438738108, + 0.011654874309897423, + 0.024603573605418205, + 0.00605576578527689, + 0.013892626389861107, + 0.0036556324921548367, + 0.0027820677496492863, + 0.001419801963493228, + 0.0017179553396999836, + 0.00351556739769876, + 0.002825595671311021, + 0.0016403914196416736, + 0.003205436747521162, + 0.0052789258770644665, + 0.0054968055337667465, + 0.0036454799119383097, + 0.004057086538523436, + 0.006732319016009569, + 0.007673549931496382, + 0.0029961979016661644, + 0.005771444644778967, + 0.003089700825512409, + 0.0036336935590952635, + 0.0022917594760656357, + 0.0023707423824816942, + 0.00535315228626132, + 0.01161625050008297, + 0.0019300151616334915, + 0.0035423743538558483, + 0.0030596614815294743, + 0.0026636344846338034, + 0.002117817522957921, + 0.001855997834354639, + 0.003319942392408848, + 0.003199169412255287, + 0.0025183542165905237, + 0.00251448224298656, + 0.009492000564932823, + 0.011809371411800385, + 0.008708049543201923, + 0.007637468632310629, + 0.012372137978672981, + 0.014226586557924747, + 0.005494451150298119, + 0.01528030913323164, + 0.005298171192407608, + 0.004657333716750145, + 0.009685029275715351, + 0.006876641884446144, + 0.006108366418629885, + 0.0120680658146739, + 0.007156809326261282, + 0.012939924374222755, + 0.012680667452514172, + 0.021539010107517242, + 0.012651223689317703, + 0.01041905116289854, + 0.22599560022354126, + 0.19211634993553162, + 0.016955610364675522, + 0.012892832979559898, + 0.004233356099575758, + 0.009177452884614468, + 0.0032900723163038492, + 0.003171200631186366, + 0.004675160627812147, + 0.0037148459814488888, + 0.00336093339137733, + 0.0023226593621075153 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 1, + 6 + ], + "selected_reward": -0.002816402819007635 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene an abstract world composed of vibrant color blocks and bold black lines in a Henri Matisse style?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the background a deconstructed map of the Forbidden City with cadmium yellow rooftops, vermilion walls, and azure/emerald green courtyards?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is there a minimalist boat with a pure ultramarine blue hull and black calligraphic strokes gliding across the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the boat create stylized ripples of simple lines as it moves, without traditional light, shadow, or perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot a smooth pan following the boat's forward movement?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.12699304521083832 + ], + "opposite_logprob_tokens": [ + -2.12699294090271 + ], + "logprob_reward": -0.12699304521083832, + "yes_no_margin_reward": 1.9999998807907104, + "token_credit_reward": -0.12699304521083832, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.05706215277314186, + 0.03656694293022156, + 0.01950695365667343, + 0.025402173399925232, + 0.016163090243935585, + 0.012091448530554771, + 0.04678453132510185, + 0.05465162545442581, + 0.5309484601020813, + 0.200822651386261 + ], + "credit_map": [ + 0.009288465604186058, + 0.007113260217010975, + 0.007057253271341324, + 0.008288686163723469, + 0.00483620073646307, + 0.006627016235142946, + 0.006312557030469179, + 0.007538709323853254, + 0.0031691810581833124, + 0.008152688853442669, + 0.002987024374306202, + 0.0025243673007935286, + 0.009646564722061157, + 0.004709778819233179, + 0.002354113617911935, + 0.0030232241842895746, + 0.0032055967021733522, + 0.0029469795990735292, + 0.001409263932146132, + 0.002066029468551278, + 0.0037734510842710733, + 0.0025060786865651608, + 0.0018266600091010332, + 0.0017728930106386542, + 0.0028868233785033226, + 0.0025771588552743196, + 0.0029107765294611454, + 0.0033853077329695225, + 0.0030854560900479555, + 0.004203306045383215, + 0.002715908922255039, + 0.0036374335177242756, + 0.0012739263474941254, + 0.0025474780704826117, + 0.0013014479773119092, + 0.0017956168157979846, + 0.0012352506164461374, + 0.0033968992065638304, + 0.0014294155407696962, + 0.0031830554362386465, + 0.001718531595543027, + 0.001573610701598227, + 0.000758104957640171, + 0.0007637104718014598, + 0.0021686244290322065, + 0.002063141204416752, + 0.0009326536091975868, + 0.0021130708046257496, + 0.005787580739706755, + 0.007391020655632019, + 0.005373912397772074, + 0.004255910869687796, + 0.00421261927112937, + 0.007023922633379698, + 0.0032470282167196274, + 0.009492532350122929, + 0.003023302648216486, + 0.003221044782549143, + 0.005139186047017574, + 0.005576552357524633, + 0.01028604805469513, + 0.014850768260657787, + 0.004604886751621962, + 0.007949832826852798, + 0.013527599163353443, + 0.02164578065276146, + 0.013298191130161285, + 0.010454781353473663, + 0.23893988132476807, + 0.20422133803367615, + 0.016160782426595688, + 0.012700045481324196, + 0.02324880100786686, + 0.0579354465007782, + 0.008912647143006325, + 0.0159236341714859, + 0.02961151674389839, + 0.025536978617310524, + 0.019681744277477264, + 0.019971858710050583 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 9, + 0 + ], + "selected_reward": -0.12699304521083832 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000006/group_3/index_tensor([4380, 4380])_group_3_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000006/group_3/index_tensor([4380, 4380])_group_3_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..a4e06e5c82742c824123b54209163909cf830c16 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000006/group_3/index_tensor([4380, 4380])_group_3_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "slow tracking shot, shot from a low angle, captures a pristine white seagull soaring beneath a vast, azure sky. Brilliant sunlight streams down, illuminating each feather with a pearly sheen, almost blending into the sky. Its massive wings gracefully spread to their full extent, a few dark feathers at the wingtips, like a calligrapher's delicate finishing strokes, adding depth to the pure white. Rather than flapping violently, it gracefully glides through the air, relying on invisible currents of air, its body only occasionally making subtle adjustments to maintain balance, its streamlined form tracing a graceful, smooth arc. Sunlight filters through the translucent edges of its wings, outlining its sharp outline and casting a soft shadow on its belly. Its piercing black eyes gaze into the distance. The vibrant, saturated colors of the entire image, the stark contrast between the deep blue of the sky and the pure white of the seagull, create an atmosphere of profound tranquility, vastness, and a sense of unfettered freedom.", + "seed": 30900, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000006/group_3/index_4380_seed_30900_dp_6_rank_12_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 3, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.20749060919042678, + "vlm_logprob": -0.20749060919042678, + "vlm_yes_no_margin": 3.3, + "vlm_token_credit": -0.20749060919042678, + "vlm_yes_ratio": 0.8, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a pristine white seagull?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the seagull soaring beneath a vast azure sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the camera shot from a low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the seagull gliding smoothly with its wings fully spread rather than flapping violently?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video feature a slow tracking shot following the seagull?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -0.9013515710830688 + ], + "opposite_logprob_tokens": [ + -0.5263515710830688 + ], + "logprob_reward": -0.9013515710830688, + "yes_no_margin_reward": -0.375, + "token_credit_reward": -0.9013515710830688, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.12881071865558624, + 0.13585758209228516, + 0.017251335084438324, + 0.12278229743242264, + 0.20151415467262268, + 0.1328647881746292, + 0.04032472148537636, + 0.015274004079401493, + 0.15656863152980804, + 0.04875171184539795 + ], + "credit_map": [ + 0.005595283582806587, + 0.006718473508954048, + 0.009594818577170372, + 0.005164012312889099, + 0.04646795615553856, + 0.04225511848926544, + 0.007910455577075481, + 0.005104620009660721, + 0.003414977341890335, + 0.0045911772176623344, + 0.011132050305604935, + 0.0038272179663181305, + 0.05756441503763199, + 0.04581369832158089, + 0.004037979058921337, + 0.005476091057062149, + 0.0022512651048600674, + 0.0021050546783953905, + 0.0019540993962436914, + 0.0022105774842202663, + 0.002151982858777046, + 0.0019366107881069183, + 0.002327868714928627, + 0.002313880482688546, + 0.010879178531467915, + 0.01735016331076622, + 0.013678314164280891, + 0.015513540245592594, + 0.02392977848649025, + 0.015890145674347878, + 0.014189401641488075, + 0.011351791210472584, + 0.01000464241951704, + 0.010548385791480541, + 0.011647769249975681, + 0.01187816634774208, + 0.07227130234241486, + 0.05995556339621544, + 0.013494753278791904, + 0.011713595129549503, + 0.00581906083971262, + 0.0114799989387393, + 0.00818499643355608, + 0.005337724927812815, + 0.04983435571193695, + 0.040886323899030685, + 0.005451389122754335, + 0.00587095832452178, + 0.00391496391966939, + 0.004978025797754526, + 0.0051331729628145695, + 0.006535560358315706, + 0.005125633906573057, + 0.0044037639163434505, + 0.005473223980516195, + 0.004760384559631348, + 0.0016896171728149056, + 0.001860554562881589, + 0.0019765212200582027, + 0.001799404970370233, + 0.001963532529771328, + 0.0025164431426674128, + 0.0018404745496809483, + 0.001627459074370563, + 0.009124995209276676, + 0.010056623257696629, + 0.01087728887796402, + 0.012224980629980564, + 0.050462160259485245, + 0.04905329644680023, + 0.008117695339024067, + 0.006651615723967552, + 0.005612287204712629, + 0.005519409663975239, + 0.005165995564311743, + 0.005183935631066561, + 0.006770811043679714, + 0.006005607545375824, + 0.009291821159422398, + 0.005201853346079588 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 8, + 1 + ], + "selected_reward": -0.9013515710830688 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a pristine white seagull?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the seagull soaring beneath a vast azure sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the camera shot from a low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the seagull gliding smoothly with its wings fully spread rather than flapping violently?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video feature a slow tracking shot following the seagull?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.00173049361910671 + ], + "opposite_logprob_tokens": [ + -6.376730442047119 + ], + "logprob_reward": -0.00173049361910671, + "yes_no_margin_reward": 6.375, + "token_credit_reward": -0.00173049361910671, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.22531405091285706, + 0.07427025586366653, + 0.024805555120110512, + 0.14332206547260284, + 0.19740651547908783, + 0.05308815464377403, + 0.025940291583538055, + 0.035202864557504654, + 0.1619318127632141, + 0.05871845409274101 + ], + "credit_map": [ + 0.010495711117982864, + 0.012704347260296345, + 0.017046237364411354, + 0.008272130973637104, + 0.08113281428813934, + 0.07227201759815216, + 0.014561380259692669, + 0.008829410187900066, + 0.005581614561378956, + 0.005888612475246191, + 0.009474103339016438, + 0.006578301079571247, + 0.011808246374130249, + 0.015287784859538078, + 0.005601935088634491, + 0.01404966413974762, + 0.0030765177216380835, + 0.0030672072898596525, + 0.002920192200690508, + 0.0031593386083841324, + 0.0034726618323475122, + 0.0030785654671490192, + 0.0027557366993278265, + 0.003275335533544421, + 0.01152844913303852, + 0.018520811572670937, + 0.015997154638171196, + 0.01671922393143177, + 0.03243707865476608, + 0.019713126122951508, + 0.015799883753061295, + 0.012606346979737282, + 0.009017262607812881, + 0.01075491402298212, + 0.011976299807429314, + 0.01363685168325901, + 0.06787408888339996, + 0.05598617345094681, + 0.015053893439471722, + 0.013107030652463436, + 0.004431871231645346, + 0.007359218783676624, + 0.009532852098345757, + 0.004378916695713997, + 0.00999442208558321, + 0.008669419214129448, + 0.004115095362067223, + 0.004606361035257578, + 0.001978188054636121, + 0.003062488744035363, + 0.0027957605198025703, + 0.0031091116834431887, + 0.004065164364874363, + 0.004960610996931791, + 0.003539708210155368, + 0.0024292587768286467, + 0.003939641173928976, + 0.0038628994952887297, + 0.005077118054032326, + 0.004206423182040453, + 0.004727364517748356, + 0.005569923669099808, + 0.0047057019546628, + 0.0031137943733483553, + 0.009565690532326698, + 0.010580272413790226, + 0.010340671986341476, + 0.012328054755926132, + 0.0517079159617424, + 0.05090473219752312, + 0.008598306216299534, + 0.007906174287199974, + 0.0062636034563183784, + 0.006865423638373613, + 0.005387536715716124, + 0.00617643678560853, + 0.00785190798342228, + 0.007049233186990023, + 0.012841382063925266, + 0.0062829297967255116 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 8 + ], + "selected_reward": -0.00173049361910671 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a pristine white seagull?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the seagull soaring beneath a vast azure sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the camera shot from a low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the seagull gliding smoothly with its wings fully spread rather than flapping violently?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video feature a slow tracking shot following the seagull?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07898899912834167 + ], + "opposite_logprob_tokens": [ + -2.578989028930664 + ], + "logprob_reward": -0.07898899912834167, + "yes_no_margin_reward": 2.5, + "token_credit_reward": -0.07898899912834167, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07076563686132431, + 0.16446931660175323, + 0.08728928864002228, + 0.02909553050994873, + 0.0595708042383194, + 0.4189874231815338, + 0.07920968532562256, + 0.032927367836236954, + 0.024026161059737206, + 0.03365875780582428 + ], + "credit_map": [ + 0.005298680160194635, + 0.00433660252019763, + 0.005088400561362505, + 0.0036427651066333055, + 0.02266908437013626, + 0.023166237398982048, + 0.0032388512045145035, + 0.0033250143751502037, + 0.0054438430815935135, + 0.005347983445972204, + 0.017761781811714172, + 0.00462538655847311, + 0.07082947343587875, + 0.05033522844314575, + 0.004378634039312601, + 0.005746997892856598, + 0.014264694415032864, + 0.0081580625846982, + 0.011708837002515793, + 0.009400094859302044, + 0.012937543913722038, + 0.010593727231025696, + 0.01265630405396223, + 0.00757003016769886, + 0.002363668754696846, + 0.0027244272641837597, + 0.0037217189092189074, + 0.0031582058873027563, + 0.006254173815250397, + 0.004048176109790802, + 0.0039028488099575043, + 0.0029223132878541946, + 0.005897372029721737, + 0.005128360353410244, + 0.0059861415065824986, + 0.005583716090768576, + 0.014536600559949875, + 0.011804533191025257, + 0.00465280469506979, + 0.005981278605759144, + 0.02205396443605423, + 0.05435817316174507, + 0.05018055438995361, + 0.028948800638318062, + 0.1121615320444107, + 0.08878853172063828, + 0.03681071102619171, + 0.025685150176286697, + 0.008403263986110687, + 0.008363836444914341, + 0.009560481645166874, + 0.010278131812810898, + 0.012641467154026031, + 0.011798427440226078, + 0.009481744840741158, + 0.008682332932949066, + 0.004205143544822931, + 0.0046568200923502445, + 0.003761995118111372, + 0.0037680126260966063, + 0.004542825277894735, + 0.004626401234418154, + 0.003777514211833477, + 0.003588656662032008, + 0.002197297289967537, + 0.0025710277259349823, + 0.0022989020217210054, + 0.002353479852899909, + 0.005578040610998869, + 0.005037527531385422, + 0.0021505896002054214, + 0.0018392970087006688, + 0.0038303870242089033, + 0.0031768169719725847, + 0.004084191285073757, + 0.003662337316200137, + 0.0046601309441030025, + 0.004208185710012913, + 0.00612918846309185, + 0.003907524049282074 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 1, + 2 + ], + "selected_reward": -0.07898899912834167 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a pristine white seagull?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the seagull soaring beneath a vast azure sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the camera shot from a low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the seagull gliding smoothly with its wings fully spread rather than flapping violently?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video feature a slow tracking shot following the seagull?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.006744122598320246 + ], + "opposite_logprob_tokens": [ + -5.006743907928467 + ], + "logprob_reward": -0.006744122598320246, + "yes_no_margin_reward": 5.0, + "token_credit_reward": -0.006744122598320246, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12306258082389832, + 0.34040358662605286, + 0.057855427265167236, + 0.1075323298573494, + 0.07685568928718567, + 0.08743917942047119, + 0.04133206233382225, + 0.03825419023633003, + 0.09761326760053635, + 0.029651718214154243 + ], + "credit_map": [ + 0.005463140085339546, + 0.00715560233220458, + 0.008570553734898567, + 0.005006985738873482, + 0.04421647638082504, + 0.041168201714754105, + 0.006925871595740318, + 0.004555754829198122, + 0.00879589281976223, + 0.010906017385423183, + 0.031654827296733856, + 0.009085697121918201, + 0.1474374383687973, + 0.11164425313472748, + 0.009862534701824188, + 0.011016896925866604, + 0.006638573948293924, + 0.00538936723023653, + 0.008083818480372429, + 0.005154785234481096, + 0.01217296440154314, + 0.009986675344407558, + 0.005466050934046507, + 0.004963194485753775, + 0.006797426380217075, + 0.008123301900923252, + 0.009454821236431599, + 0.0077504487708210945, + 0.030077695846557617, + 0.026803279295563698, + 0.01058642566204071, + 0.007938937284052372, + 0.004744459874927998, + 0.0044472538866102695, + 0.0050952923484146595, + 0.006870817393064499, + 0.02287641167640686, + 0.019500814378261566, + 0.007531990297138691, + 0.005788652691990137, + 0.003004268975928426, + 0.004838159307837486, + 0.005489342845976353, + 0.004756084643304348, + 0.03264222666621208, + 0.028541794046759605, + 0.004304671660065651, + 0.0038626256864517927, + 0.005826650187373161, + 0.0035541937686502934, + 0.004156429320573807, + 0.004529143683612347, + 0.006266905926167965, + 0.005708595272153616, + 0.005347596947103739, + 0.0059425486251711845, + 0.004815396852791309, + 0.004447659011930227, + 0.004671074450016022, + 0.004815795924514532, + 0.005357799585908651, + 0.004463509190827608, + 0.004351441282778978, + 0.005331513471901417, + 0.00625233631581068, + 0.007716102059930563, + 0.0070646111853420734, + 0.008684234693646431, + 0.028553633019328117, + 0.027835896238684654, + 0.0067976247519254684, + 0.0047088321298360825, + 0.0034748732578009367, + 0.003573746420443058, + 0.0028913591522723436, + 0.0029901606030762196, + 0.003950581420212984, + 0.0032183609437197447, + 0.005747887305915356, + 0.003804747713729739 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 3 + ], + "selected_reward": -0.006744122598320246 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a pristine white seagull?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the seagull soaring beneath a vast azure sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the camera shot from a low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the seagull gliding smoothly with its wings fully spread rather than flapping violently?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video feature a slow tracking shot following the seagull?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.048637859523296356 + ], + "opposite_logprob_tokens": [ + -3.048637866973877 + ], + "logprob_reward": -0.048637859523296356, + "yes_no_margin_reward": 3.0, + "token_credit_reward": -0.048637859523296356, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.175496444106102, + 0.2463684380054474, + 0.08055946975946426, + 0.20330102741718292, + 0.038798313587903976, + 0.05419595167040825, + 0.040755853056907654, + 0.07208810746669769, + 0.06680633872747421, + 0.021630125120282173 + ], + "credit_map": [ + 0.010850345715880394, + 0.012065968476235867, + 0.011946843937039375, + 0.007322961930185556, + 0.057841841131448746, + 0.056989409029483795, + 0.008507215417921543, + 0.00997184868901968, + 0.007436348590999842, + 0.008655400946736336, + 0.02020750567317009, + 0.007962891831994057, + 0.10162682831287384, + 0.0807364284992218, + 0.007759089581668377, + 0.011983937583863735, + 0.014060795307159424, + 0.005997306201606989, + 0.013108929619193077, + 0.006774401292204857, + 0.014981674961745739, + 0.0116726728156209, + 0.007048195693641901, + 0.006915474776178598, + 0.01028394978493452, + 0.022324547171592712, + 0.016785085201263428, + 0.019680896773934364, + 0.0661485567688942, + 0.04661278799176216, + 0.00992080383002758, + 0.011544373817741871, + 0.0049702925607562065, + 0.004462468437850475, + 0.004822984337806702, + 0.004064994864165783, + 0.006237821187824011, + 0.005856092553585768, + 0.0037945262156426907, + 0.0045891315676271915, + 0.004763968754559755, + 0.007317348383367062, + 0.005955835338681936, + 0.005509744398295879, + 0.012906263582408428, + 0.00972959864884615, + 0.004282800015062094, + 0.0037303848657757044, + 0.0033807549625635147, + 0.004180245567113161, + 0.006564468145370483, + 0.005533421412110329, + 0.006035934668034315, + 0.0051221586763858795, + 0.0060295588336884975, + 0.0039093089289963245, + 0.008797915652394295, + 0.008233107626438141, + 0.00964129064232111, + 0.010222062468528748, + 0.009554333984851837, + 0.008161455392837524, + 0.010645641945302486, + 0.006832302547991276, + 0.00404603686183691, + 0.0049714576452970505, + 0.004003917798399925, + 0.00541719188913703, + 0.020313367247581482, + 0.020118923857808113, + 0.00482920603826642, + 0.0031062327325344086, + 0.0020749946124851704, + 0.0028738598339259624, + 0.0019164837431162596, + 0.002521080896258354, + 0.003717879531905055, + 0.003557700663805008, + 0.0024601786863058805, + 0.002507945988327265 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 3, + 0 + ], + "selected_reward": -0.048637859523296356 + } + ], + "vlm_decoded_outputs": [ + "No", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "slow tracking shot, shot from a low angle, captures a pristine white seagull soaring beneath a vast, azure sky. Brilliant sunlight streams down, illuminating each feather with a pearly sheen, almost blending into the sky. Its massive wings gracefully spread to their full extent, a few dark feathers at the wingtips, like a calligrapher's delicate finishing strokes, adding depth to the pure white. Rather than flapping violently, it gracefully glides through the air, relying on invisible currents of air, its body only occasionally making subtle adjustments to maintain balance, its streamlined form tracing a graceful, smooth arc. Sunlight filters through the translucent edges of its wings, outlining its sharp outline and casting a soft shadow on its belly. Its piercing black eyes gaze into the distance. The vibrant, saturated colors of the entire image, the stark contrast between the deep blue of the sky and the pure white of the seagull, create an atmosphere of profound tranquility, vastness, and a sense of unfettered freedom.", + "seed": 30901, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000006/group_3/index_4380_seed_30901_dp_6_rank_12_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 3, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.06422861905302853, + "vlm_logprob": -0.06422861905302853, + "vlm_yes_no_margin": 3.85, + "vlm_token_credit": -0.06422861905302853, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a pristine white seagull?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the seagull soaring beneath a vast azure sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the camera shot from a low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the seagull gliding smoothly with its wings fully spread rather than flapping violently?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video feature a slow tracking shot following the seagull?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.1022089421749115 + ], + "opposite_logprob_tokens": [ + -2.3522088527679443 + ], + "logprob_reward": -0.1022089421749115, + "yes_no_margin_reward": 2.25, + "token_credit_reward": -0.1022089421749115, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.17307907342910767, + 0.18248268961906433, + 0.19325928390026093, + 0.2211974859237671, + 0.07141867280006409, + 0.0250144861638546, + 0.07742124050855637, + 0.02508886717259884, + 0.01754140853881836, + 0.013496770523488522 + ], + "credit_map": [ + 0.018728362396359444, + 0.03321666270494461, + 0.020906580612063408, + 0.01774231716990471, + 0.021842489019036293, + 0.02740255557000637, + 0.019088586792349815, + 0.014151527546346188, + 0.02236047014594078, + 0.015185389667749405, + 0.02378379926085472, + 0.02111634612083435, + 0.01867007650434971, + 0.028196534141898155, + 0.01710347831249237, + 0.03606659919023514, + 0.021444834768772125, + 0.02958589605987072, + 0.024229053407907486, + 0.018584908917546272, + 0.022546546533703804, + 0.03566060960292816, + 0.02019863948225975, + 0.021008826792240143, + 0.027368733659386635, + 0.021477065980434418, + 0.02934480831027031, + 0.02664629928767681, + 0.02704494073987007, + 0.022441457957029343, + 0.020943446084856987, + 0.04593075439333916, + 0.006338360253721476, + 0.008093530312180519, + 0.009918341413140297, + 0.011239610612392426, + 0.007060493342578411, + 0.014355793595314026, + 0.005720443092286587, + 0.008692112751305103, + 0.0042734211310744286, + 0.0042337412014603615, + 0.004375006537884474, + 0.002421028446406126, + 0.0024830554611980915, + 0.00222053867764771, + 0.0025792939122766256, + 0.0024284033570438623, + 0.005921181757003069, + 0.009662718512117863, + 0.011822117492556572, + 0.009616457857191563, + 0.006960898172110319, + 0.016233352944254875, + 0.008568854071199894, + 0.008635659702122211, + 0.0025605172850191593, + 0.003421880304813385, + 0.0035582997370511293, + 0.004686216823756695, + 0.0019170549931004643, + 0.002946223597973585, + 0.0032294737175107002, + 0.002769204555079341, + 0.001908222446218133, + 0.0019031129777431488, + 0.002079959725961089, + 0.0028721727430820465, + 0.0021187132224440575, + 0.0022351802326738834, + 0.0017721621552482247, + 0.0026518884114921093, + 0.0013611778849735856, + 0.0014597265981137753, + 0.0016525209648534656, + 0.0017460091039538383, + 0.001953365048393607, + 0.001599721610546112, + 0.001713722595013678, + 0.002010528463870287 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 2, + 1 + ], + "selected_reward": -0.1022089421749115 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a pristine white seagull?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the seagull soaring beneath a vast azure sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the camera shot from a low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the seagull gliding smoothly with its wings fully spread rather than flapping violently?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video feature a slow tracking shot following the seagull?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0017275185091421008 + ], + "opposite_logprob_tokens": [ + -6.37672758102417 + ], + "logprob_reward": -0.0017275185091421008, + "yes_no_margin_reward": 6.375, + "token_credit_reward": -0.0017275185091421008, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.20832811295986176, + 0.20754921436309814, + 0.13166949152946472, + 0.07403561472892761, + 0.091697558760643, + 0.04138709977269173, + 0.082948237657547, + 0.03741101175546646, + 0.050945281982421875, + 0.0740283727645874 + ], + "credit_map": [ + 0.019859563559293747, + 0.03223658725619316, + 0.02097277343273163, + 0.014485200867056847, + 0.022698258981108665, + 0.04415486380457878, + 0.017125248908996582, + 0.03679561987519264, + 0.031930748373270035, + 0.014655593782663345, + 0.025223305448889732, + 0.025051433593034744, + 0.014764794148504734, + 0.03134604170918465, + 0.021861769258975983, + 0.042715538293123245, + 0.01745413988828659, + 0.028429772704839706, + 0.012833804823458195, + 0.010887962765991688, + 0.012363508343696594, + 0.021221527829766273, + 0.011318503879010677, + 0.017160281538963318, + 0.012147271074354649, + 0.0077039627358317375, + 0.009346865117549896, + 0.007764007896184921, + 0.01048866007477045, + 0.00621116254478693, + 0.00886060856282711, + 0.011513074859976768, + 0.01285587903112173, + 0.01015101931989193, + 0.01579972356557846, + 0.01373849343508482, + 0.008710849098861217, + 0.00811237283051014, + 0.011288001202046871, + 0.011041214689612389, + 0.00687082065269351, + 0.006758378352969885, + 0.007908427156507969, + 0.00427084369584918, + 0.0029538199305534363, + 0.003892352804541588, + 0.0040553174912929535, + 0.004677136428654194, + 0.013354183174669743, + 0.009777367115020752, + 0.011760720051825047, + 0.008446011692285538, + 0.0079807685688138, + 0.009625785052776337, + 0.00950656458735466, + 0.01249683927744627, + 0.004976671189069748, + 0.005109356250613928, + 0.0051829325966537, + 0.006271183956414461, + 0.0034937586169689894, + 0.0031156535260379314, + 0.004938841797411442, + 0.004322616383433342, + 0.008099767379462719, + 0.005759903229773045, + 0.0064161187037825584, + 0.00866688508540392, + 0.006606039125472307, + 0.005122588016092777, + 0.005025125574320555, + 0.005248852074146271, + 0.010791996493935585, + 0.008118837140500546, + 0.010009223595261574, + 0.011175856925547123, + 0.007945002987980843, + 0.008536751382052898, + 0.00841918122023344, + 0.009031524881720543 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 2 + ], + "selected_reward": -0.0017275185091421008 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a pristine white seagull?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the seagull soaring beneath a vast azure sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the camera shot from a low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the seagull gliding smoothly with its wings fully spread rather than flapping violently?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video feature a slow tracking shot following the seagull?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.17990349233150482 + ], + "opposite_logprob_tokens": [ + -1.804903507232666 + ], + "logprob_reward": -0.17990349233150482, + "yes_no_margin_reward": 1.625, + "token_credit_reward": -0.17990349233150482, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.32982927560806274, + 0.12969422340393066, + 0.11650960892438889, + 0.07294955849647522, + 0.06974176317453384, + 0.03739845007658005, + 0.07670211791992188, + 0.04725322499871254, + 0.08643213659524918, + 0.03348959609866142 + ], + "credit_map": [ + 0.030389226973056793, + 0.06102239713072777, + 0.04298567771911621, + 0.03265080600976944, + 0.026010457426309586, + 0.07655123621225357, + 0.023166531696915627, + 0.037052955478429794, + 0.01893860101699829, + 0.012059150263667107, + 0.016890203580260277, + 0.014807633124291897, + 0.015944959595799446, + 0.01850530505180359, + 0.014066956005990505, + 0.018481420353055, + 0.018599089235067368, + 0.017529340460896492, + 0.017604628577828407, + 0.012263554148375988, + 0.010176572017371655, + 0.01628694497048855, + 0.01172417402267456, + 0.012325300835072994, + 0.015397593379020691, + 0.008853393606841564, + 0.010065825656056404, + 0.006979349069297314, + 0.009604380466043949, + 0.007596045266836882, + 0.005071387160569429, + 0.009381589479744434, + 0.0080897631123662, + 0.007685596588999033, + 0.012433374300599098, + 0.007854671217501163, + 0.005394049920141697, + 0.011609755456447601, + 0.007187772076576948, + 0.009486780501902103, + 0.005720740649849176, + 0.005529556889086962, + 0.005227849818766117, + 0.005863159894943237, + 0.003947458229959011, + 0.004004061222076416, + 0.0030665176454931498, + 0.004039101302623749, + 0.015955539420247078, + 0.008625454269349575, + 0.011525758542120457, + 0.010087073780596256, + 0.005595549941062927, + 0.008857998065650463, + 0.006742863450199366, + 0.009311879053711891, + 0.007596999406814575, + 0.007495108991861343, + 0.0069120279513299465, + 0.006951681803911924, + 0.005432430189102888, + 0.00398798706009984, + 0.004177922382950783, + 0.004699068609625101, + 0.020144063979387283, + 0.011664737947285175, + 0.01248554140329361, + 0.010181956924498081, + 0.006772072520107031, + 0.006006050854921341, + 0.008113935589790344, + 0.011063776910305023, + 0.0036446149460971355, + 0.003993344027549028, + 0.0043380381539464, + 0.005225753877311945, + 0.004713799338787794, + 0.004340703599154949, + 0.0034134669695049524, + 0.003819875419139862 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 2 + ], + "selected_reward": -0.17990349233150482 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a pristine white seagull?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the seagull soaring beneath a vast azure sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the camera shot from a low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the seagull gliding smoothly with its wings fully spread rather than flapping violently?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video feature a slow tracking shot following the seagull?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0036200247704982758 + ], + "opposite_logprob_tokens": [ + -5.628620147705078 + ], + "logprob_reward": -0.0036200247704982758, + "yes_no_margin_reward": 5.625, + "token_credit_reward": -0.0036200247704982758, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1448969691991806, + 0.14709071815013885, + 0.23987270891666412, + 0.11760956794023514, + 0.06700242310762405, + 0.06475203484296799, + 0.09347782284021378, + 0.05067797750234604, + 0.037208445370197296, + 0.03741135075688362 + ], + "credit_map": [ + 0.011764010414481163, + 0.027118423953652382, + 0.016628770157694817, + 0.011174265295267105, + 0.013298922218382359, + 0.03466071933507919, + 0.01096373237669468, + 0.019288118928670883, + 0.01483568362891674, + 0.012484796345233917, + 0.017432676628232002, + 0.01822061836719513, + 0.012660232372581959, + 0.020087461918592453, + 0.024522356688976288, + 0.02684689499437809, + 0.02463432028889656, + 0.050012391060590744, + 0.023175274953246117, + 0.01528448797762394, + 0.022930823266506195, + 0.05645762383937836, + 0.019138861447572708, + 0.0282389335334301, + 0.010109934024512768, + 0.012317296117544174, + 0.014050996862351894, + 0.014493642374873161, + 0.017916442826390266, + 0.014303909614682198, + 0.015089777298271656, + 0.019327577203512192, + 0.005938686430454254, + 0.007744141388684511, + 0.009385728277266026, + 0.010169782675802708, + 0.00789098720997572, + 0.00897198636084795, + 0.007864976301789284, + 0.009036133997142315, + 0.006319742184132338, + 0.010263021104037762, + 0.009490014053881168, + 0.007724334951490164, + 0.006694588344544172, + 0.01013758685439825, + 0.007172200828790665, + 0.006950542330741882, + 0.007569475099444389, + 0.01235946360975504, + 0.013811186887323856, + 0.01033499650657177, + 0.008923892863094807, + 0.0182951707392931, + 0.012428742833435535, + 0.009754898957908154, + 0.005906959529966116, + 0.007560278754681349, + 0.007174326106905937, + 0.006694374140352011, + 0.004683660343289375, + 0.006489780265837908, + 0.004215160850435495, + 0.007953437976539135, + 0.0039015754591673613, + 0.004748076666146517, + 0.004611359443515539, + 0.005861012265086174, + 0.004010290373116732, + 0.004838905762881041, + 0.0039204624481499195, + 0.005316763650625944, + 0.004826711490750313, + 0.004475447349250317, + 0.0049970438703894615, + 0.004990789107978344, + 0.004775309469550848, + 0.00498832855373621, + 0.0037378997076302767, + 0.0046198247000575066 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 1, + 0 + ], + "selected_reward": -0.0036200247704982758 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a pristine white seagull?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the seagull soaring beneath a vast azure sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the camera shot from a low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the seagull gliding smoothly with its wings fully spread rather than flapping violently?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video feature a slow tracking shot following the seagull?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03368311747908592 + ], + "opposite_logprob_tokens": [ + -3.4086830615997314 + ], + "logprob_reward": -0.03368311747908592, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.03368311747908592, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.31708139181137085, + 0.1472679227590561, + 0.08836603909730911, + 0.04695681482553482, + 0.12337048351764679, + 0.04315761849284172, + 0.07042385637760162, + 0.04889391362667084, + 0.05921432375907898, + 0.05526763200759888 + ], + "credit_map": [ + 0.03297048807144165, + 0.04563872143626213, + 0.0571708157658577, + 0.02788335457444191, + 0.03259294852614403, + 0.06463965028524399, + 0.03219827637076378, + 0.023987138643860817, + 0.018687397241592407, + 0.01052142959088087, + 0.018615733832120895, + 0.015664219856262207, + 0.014409593306481838, + 0.020520996302366257, + 0.015703393146395683, + 0.03314515948295593, + 0.012110965326428413, + 0.015988219529390335, + 0.0074293045327067375, + 0.007921302691102028, + 0.006791900843381882, + 0.020269129425287247, + 0.007423404138535261, + 0.01043181587010622, + 0.0075514186173677444, + 0.0049341656267642975, + 0.007188891991972923, + 0.004620109219104052, + 0.005266353487968445, + 0.00458323210477829, + 0.004227953962981701, + 0.008584688417613506, + 0.009227118454873562, + 0.01305358111858368, + 0.01751597225666046, + 0.015545399859547615, + 0.01254244428128004, + 0.025217005982995033, + 0.014424464665353298, + 0.015844499692320824, + 0.004118419252336025, + 0.007666308898478746, + 0.00848442129790783, + 0.005212152376770973, + 0.00476415129378438, + 0.0026866497937589884, + 0.004581248387694359, + 0.005644267424941063, + 0.005343072582036257, + 0.008123904466629028, + 0.010012389160692692, + 0.008848256431519985, + 0.008642636239528656, + 0.014040756970643997, + 0.009246387518942356, + 0.006166453938931227, + 0.0048942542634904385, + 0.006655355915427208, + 0.006844347808510065, + 0.004797718487679958, + 0.003937640693038702, + 0.006273339502513409, + 0.006383118219673634, + 0.009108140133321285, + 0.0070270937867462635, + 0.006312775891274214, + 0.008004004135727882, + 0.00866714958101511, + 0.007692623883485794, + 0.009098992682993412, + 0.004295121878385544, + 0.008116563782095909, + 0.006111462134867907, + 0.007247167639434338, + 0.00828169472515583, + 0.005852384492754936, + 0.006229674443602562, + 0.0077763330191373825, + 0.0062145059928298, + 0.007554409559816122 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 4 + ], + "selected_reward": -0.03368311747908592 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "slow tracking shot, shot from a low angle, captures a pristine white seagull soaring beneath a vast, azure sky. Brilliant sunlight streams down, illuminating each feather with a pearly sheen, almost blending into the sky. Its massive wings gracefully spread to their full extent, a few dark feathers at the wingtips, like a calligrapher's delicate finishing strokes, adding depth to the pure white. Rather than flapping violently, it gracefully glides through the air, relying on invisible currents of air, its body only occasionally making subtle adjustments to maintain balance, its streamlined form tracing a graceful, smooth arc. Sunlight filters through the translucent edges of its wings, outlining its sharp outline and casting a soft shadow on its belly. Its piercing black eyes gaze into the distance. The vibrant, saturated colors of the entire image, the stark contrast between the deep blue of the sky and the pure white of the seagull, create an atmosphere of profound tranquility, vastness, and a sense of unfettered freedom.", + "seed": 30902, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000006/group_3/index_4380_seed_30902_dp_7_rank_14_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 3, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.21161839938722551, + "vlm_logprob": -0.21161839938722551, + "vlm_yes_no_margin": 3.25, + "vlm_token_credit": -0.21161839938722551, + "vlm_yes_ratio": 0.8, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a pristine white seagull?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the seagull soaring beneath a vast azure sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the camera shot from a low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the seagull gliding smoothly with its wings fully spread rather than flapping violently?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video feature a slow tracking shot following the seagull?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -0.9001715779304504 + ], + "opposite_logprob_tokens": [ + -0.5251715779304504 + ], + "logprob_reward": -0.9001715779304504, + "yes_no_margin_reward": -0.375, + "token_credit_reward": -0.9001715779304504, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.0469287633895874, + 0.016420533880591393, + 0.2189388871192932, + 0.03845949470996857, + 0.04413255304098129, + 0.06324560940265656, + 0.07412928342819214, + 0.10738315433263779, + 0.37397515773773193, + 0.016386570408940315 + ], + "credit_map": [ + 0.006067043170332909, + 0.005574851296842098, + 0.005872155074030161, + 0.004817829001694918, + 0.007141031790524721, + 0.005300131160765886, + 0.006690274924039841, + 0.005465442314743996, + 0.0016577683854848146, + 0.0015253021847456694, + 0.0017055926145985723, + 0.001616047229617834, + 0.0016411484684795141, + 0.0014540412230417132, + 0.005031054373830557, + 0.0017895770724862814, + 0.0820493996143341, + 0.025302840396761894, + 0.008060392923653126, + 0.006775285582989454, + 0.007370490115135908, + 0.006527864374220371, + 0.007219414692372084, + 0.07563318312168121, + 0.0068883709609508514, + 0.005792404059320688, + 0.006215277127921581, + 0.004115013405680656, + 0.0032019815407693386, + 0.0030038258992135525, + 0.0031927917152643204, + 0.006049827206879854, + 0.006579998414963484, + 0.005797014106065035, + 0.004665800370275974, + 0.00516117550432682, + 0.005130639765411615, + 0.0060365949757397175, + 0.0047727893106639385, + 0.005988533142954111, + 0.009111510589718819, + 0.009484182111918926, + 0.007465030532330275, + 0.007742148824036121, + 0.006672744173556566, + 0.006683687679469585, + 0.007625064346939325, + 0.008461236022412777, + 0.01024587918072939, + 0.010618120431900024, + 0.008246304467320442, + 0.00992707721889019, + 0.01019134558737278, + 0.00794336199760437, + 0.011058413423597813, + 0.005898771807551384, + 0.015793636441230774, + 0.013337192125618458, + 0.01169496402144432, + 0.012788001447916031, + 0.015914225950837135, + 0.017188023775815964, + 0.013543570414185524, + 0.007123530842363834, + 0.052184779196977615, + 0.05181761831045151, + 0.02787887305021286, + 0.047257617115974426, + 0.0644950419664383, + 0.0767911896109581, + 0.027796708047389984, + 0.025753295049071312, + 0.002330504124984145, + 0.0027237411122769117, + 0.001718362676911056, + 0.0023274337872862816, + 0.0021334069315344095, + 0.0018590280087664723, + 0.0015777349472045898, + 0.0017163570737466216 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 2, + 7 + ], + "selected_reward": -0.9001715779304504 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a pristine white seagull?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the seagull soaring beneath a vast azure sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the camera shot from a low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the seagull gliding smoothly with its wings fully spread rather than flapping violently?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video feature a slow tracking shot following the seagull?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.002211983548477292 + ], + "opposite_logprob_tokens": [ + -6.127212047576904 + ], + "logprob_reward": -0.002211983548477292, + "yes_no_margin_reward": 6.125, + "token_credit_reward": -0.002211983548477292, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08131586760282516, + 0.05110209435224533, + 0.29962655901908875, + 0.07123810797929764, + 0.12441714853048325, + 0.04716052860021591, + 0.06410116702318192, + 0.14260618388652802, + 0.05561637133359909, + 0.06281594187021255 + ], + "credit_map": [ + 0.01083298772573471, + 0.010020979680120945, + 0.009147094562649727, + 0.009362502954900265, + 0.011383632197976112, + 0.01027261558920145, + 0.010634270496666431, + 0.00966177973896265, + 0.007688439916819334, + 0.005930607207119465, + 0.004688377026468515, + 0.00489909527823329, + 0.0061159804463386536, + 0.006244868505746126, + 0.008681032806634903, + 0.006853692699223757, + 0.10721301287412643, + 0.034206174314022064, + 0.010632147081196308, + 0.009410512633621693, + 0.010193126276135445, + 0.009564537554979324, + 0.012659089639782906, + 0.10574793070554733, + 0.015518812462687492, + 0.0164633896201849, + 0.008844726718962193, + 0.008304070681333542, + 0.0057750483974814415, + 0.005906206090003252, + 0.004708876833319664, + 0.005716981366276741, + 0.022421235218644142, + 0.015266749076545238, + 0.010727792046964169, + 0.013716096989810467, + 0.01262335292994976, + 0.0160831231623888, + 0.011806471273303032, + 0.02177233248949051, + 0.007189328782260418, + 0.0064518083818256855, + 0.00500663835555315, + 0.005755005404353142, + 0.004787071142345667, + 0.0050400556065142155, + 0.005545148625969887, + 0.007385471370071173, + 0.008501608856022358, + 0.00941803865134716, + 0.007267701905220747, + 0.006893293000757694, + 0.010569892823696136, + 0.007679997012019157, + 0.0073012118227779865, + 0.006469422951340675, + 0.018885059282183647, + 0.016996774822473526, + 0.01173404324799776, + 0.015275971032679081, + 0.02272237092256546, + 0.020824426785111427, + 0.02359064854681492, + 0.012576892971992493, + 0.008869029581546783, + 0.007622046396136284, + 0.006027041468769312, + 0.005699432920664549, + 0.010948479175567627, + 0.00910003948956728, + 0.004103119485080242, + 0.0032471795566380024, + 0.008790751919150352, + 0.0085124047473073, + 0.007929808460175991, + 0.007473473437130451, + 0.011255597695708275, + 0.009529964998364449, + 0.005359197501093149, + 0.003964744973927736 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 7, + 4 + ], + "selected_reward": -0.002211983548477292 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a pristine white seagull?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the seagull soaring beneath a vast azure sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the camera shot from a low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the seagull gliding smoothly with its wings fully spread rather than flapping violently?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video feature a slow tracking shot following the seagull?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.08904818445444107 + ], + "opposite_logprob_tokens": [ + -2.464048147201538 + ], + "logprob_reward": -0.08904818445444107, + "yes_no_margin_reward": 2.375, + "token_credit_reward": -0.08904818445444107, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.029810039326548576, + 0.014377118088304996, + 0.07210521399974823, + 0.04122505336999893, + 0.4233458340167999, + 0.015319284982979298, + 0.019102972000837326, + 0.1311868131160736, + 0.1861463040113449, + 0.0673813745379448 + ], + "credit_map": [ + 0.006787619087845087, + 0.0037892365362495184, + 0.002471141517162323, + 0.0036106896586716175, + 0.002121306722983718, + 0.0025313396472483873, + 0.004567157011479139, + 0.0039315493777394295, + 0.002094365656375885, + 0.0015920448349788785, + 0.0012729421723634005, + 0.0016866918886080384, + 0.0012775694485753775, + 0.0012044947361573577, + 0.003083732444792986, + 0.0021652772556990385, + 0.016464462503790855, + 0.01063202042132616, + 0.0061087231151759624, + 0.007929690182209015, + 0.006838786415755749, + 0.006485101766884327, + 0.007204077206552029, + 0.010442354716360569, + 0.010822129435837269, + 0.009679348208010197, + 0.004322113003581762, + 0.00426898617297411, + 0.0032541989348828793, + 0.0031201187521219254, + 0.0023229848593473434, + 0.003435172140598297, + 0.06813029199838638, + 0.048992712050676346, + 0.04565051198005676, + 0.045088306069374084, + 0.04568647965788841, + 0.057278044521808624, + 0.0440690778195858, + 0.06845041364431381, + 0.003000404452905059, + 0.002681031124666333, + 0.0016902958741411567, + 0.0022203484550118446, + 0.0012158120516687632, + 0.0010173621121793985, + 0.0016538078198209405, + 0.001840222510509193, + 0.0031140188220888376, + 0.003401195164769888, + 0.002668356988579035, + 0.0025564369279891253, + 0.0023443419486284256, + 0.0018454859964549541, + 0.0016042955685406923, + 0.0015688405837863684, + 0.018170524388551712, + 0.01632487215101719, + 0.012551669962704182, + 0.014439759775996208, + 0.019546423107385635, + 0.019844379276037216, + 0.019354069605469704, + 0.010955113917589188, + 0.026802165433764458, + 0.023654399439692497, + 0.01459075603634119, + 0.02117171324789524, + 0.032852549105882645, + 0.04061312973499298, + 0.013689672574400902, + 0.012771925888955593, + 0.009874648414552212, + 0.01075085811316967, + 0.007806509733200073, + 0.009386913850903511, + 0.012066365219652653, + 0.006843909155577421, + 0.006036180071532726, + 0.0046159871853888035 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 8, + 7 + ], + "selected_reward": -0.08904818445444107 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a pristine white seagull?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the seagull soaring beneath a vast azure sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the camera shot from a low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the seagull gliding smoothly with its wings fully spread rather than flapping violently?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video feature a slow tracking shot following the seagull?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.00464076129719615 + ], + "opposite_logprob_tokens": [ + -5.379640579223633 + ], + "logprob_reward": -0.00464076129719615, + "yes_no_margin_reward": 5.375, + "token_credit_reward": -0.00464076129719615, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07542845606803894, + 0.048469264060258865, + 0.21828609704971313, + 0.08074545860290527, + 0.11808958649635315, + 0.07326114922761917, + 0.06386817991733551, + 0.09867987781763077, + 0.04760641232132912, + 0.17556548118591309 + ], + "credit_map": [ + 0.00793150719255209, + 0.009596102871000767, + 0.010478212498128414, + 0.009123919531702995, + 0.008612498641014099, + 0.008083068765699863, + 0.009921198710799217, + 0.011681945994496346, + 0.004897946026176214, + 0.005369608756154776, + 0.00515358429402113, + 0.006355558522045612, + 0.004868011921644211, + 0.006424285005778074, + 0.009176632389426231, + 0.006223639473319054, + 0.07528331875801086, + 0.029686370864510536, + 0.010771487839519978, + 0.013265972025692463, + 0.00923711434006691, + 0.010341491550207138, + 0.01232021301984787, + 0.057380132377147675, + 0.009873904287815094, + 0.011929265223443508, + 0.008411997929215431, + 0.009966142475605011, + 0.006781038362532854, + 0.0070095788687467575, + 0.007498099468648434, + 0.01927543431520462, + 0.015090861357748508, + 0.01470224466174841, + 0.01682405173778534, + 0.01901629939675331, + 0.01490175910294056, + 0.01187783945351839, + 0.012179557234048843, + 0.013496979139745235, + 0.014015167020261288, + 0.011661119759082794, + 0.010650108568370342, + 0.0104396166279912, + 0.006086188368499279, + 0.005455253645777702, + 0.006742458324879408, + 0.008211235515773296, + 0.009484000504016876, + 0.00860722828656435, + 0.0056732697412371635, + 0.007977514527738094, + 0.008938590064644814, + 0.007037613540887833, + 0.010548366233706474, + 0.0056016044691205025, + 0.011730246245861053, + 0.011096496134996414, + 0.006743430159986019, + 0.013239400461316109, + 0.01604006066918373, + 0.019349902868270874, + 0.01349576935172081, + 0.006984568201005459, + 0.007253375835716724, + 0.006806270219385624, + 0.0035066900309175253, + 0.005378519184887409, + 0.00936032459139824, + 0.007709847297519445, + 0.0037045397330075502, + 0.0038868519477546215, + 0.02329876273870468, + 0.01972692273557186, + 0.016353653743863106, + 0.016928553581237793, + 0.037358663976192474, + 0.0412321612238884, + 0.010205413214862347, + 0.010461356490850449 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 9, + 4 + ], + "selected_reward": -0.00464076129719615 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a pristine white seagull?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the seagull soaring beneath a vast azure sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the camera shot from a low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the seagull gliding smoothly with its wings fully spread rather than flapping violently?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video feature a slow tracking shot following the seagull?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.06201948970556259 + ], + "opposite_logprob_tokens": [ + -2.8120195865631104 + ], + "logprob_reward": -0.06201948970556259, + "yes_no_margin_reward": 2.75, + "token_credit_reward": -0.06201948970556259, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.05954485759139061, + 0.02799782156944275, + 0.14628399908542633, + 0.050557512789964676, + 0.09685727953910828, + 0.07059905678033829, + 0.03112625516951084, + 0.13298550248146057, + 0.141377255320549, + 0.24267049133777618 + ], + "credit_map": [ + 0.007462759967893362, + 0.006602477747946978, + 0.005499829538166523, + 0.008271646685898304, + 0.0063926661387085915, + 0.007331705652177334, + 0.010275512002408504, + 0.007708261720836163, + 0.003323948010802269, + 0.0028641200624406338, + 0.0028243246488273144, + 0.002604901324957609, + 0.002726808888837695, + 0.003831542795524001, + 0.005503622815012932, + 0.004318550694733858, + 0.04436710104346275, + 0.02258479595184326, + 0.010056947357952595, + 0.012900360859930515, + 0.008088212460279465, + 0.009038678370416164, + 0.00920126773416996, + 0.030046634376049042, + 0.013219024986028671, + 0.013510825112462044, + 0.005256033502519131, + 0.003763421904295683, + 0.00360663328319788, + 0.004822699818760157, + 0.002141415374353528, + 0.004237454850226641, + 0.02671271376311779, + 0.02130967751145363, + 0.005828665569424629, + 0.006368796341121197, + 0.005459442734718323, + 0.006640492007136345, + 0.007682580966502428, + 0.01685490645468235, + 0.011702511459589005, + 0.011142156086862087, + 0.007950357161462307, + 0.008226646110415459, + 0.0068750083446502686, + 0.0061934287659823895, + 0.008118866011500359, + 0.010390079580247402, + 0.004752674605697393, + 0.004811639431864023, + 0.005440127570182085, + 0.004222249612212181, + 0.0026715376880019903, + 0.002579385181888938, + 0.002747605787590146, + 0.003901035524904728, + 0.01884007826447487, + 0.014044244773685932, + 0.014726197347044945, + 0.01817443035542965, + 0.016985174268484116, + 0.020819127559661865, + 0.019801951944828033, + 0.00959428958594799, + 0.019616052508354187, + 0.018232669681310654, + 0.010591607540845871, + 0.017115764319896698, + 0.02573033981025219, + 0.029945990070700645, + 0.010583614930510521, + 0.009561211802065372, + 0.03492521122097969, + 0.029273835942149162, + 0.03371696174144745, + 0.02814747393131256, + 0.043262846767902374, + 0.04107050970196724, + 0.019435446709394455, + 0.01283818669617176 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 2, + 8 + ], + "selected_reward": -0.06201948970556259 + } + ], + "vlm_decoded_outputs": [ + "No", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "slow tracking shot, shot from a low angle, captures a pristine white seagull soaring beneath a vast, azure sky. Brilliant sunlight streams down, illuminating each feather with a pearly sheen, almost blending into the sky. Its massive wings gracefully spread to their full extent, a few dark feathers at the wingtips, like a calligrapher's delicate finishing strokes, adding depth to the pure white. Rather than flapping violently, it gracefully glides through the air, relying on invisible currents of air, its body only occasionally making subtle adjustments to maintain balance, its streamlined form tracing a graceful, smooth arc. Sunlight filters through the translucent edges of its wings, outlining its sharp outline and casting a soft shadow on its belly. Its piercing black eyes gaze into the distance. The vibrant, saturated colors of the entire image, the stark contrast between the deep blue of the sky and the pure white of the seagull, create an atmosphere of profound tranquility, vastness, and a sense of unfettered freedom.", + "seed": 30903, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000006/group_3/index_4380_seed_30903_dp_7_rank_14_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 3, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.29424117363523694, + "vlm_logprob": -0.29424117363523694, + "vlm_yes_no_margin": 3.075, + "vlm_token_credit": -0.29424117363523694, + "vlm_yes_ratio": 0.8, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a pristine white seagull?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the seagull soaring beneath a vast azure sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the camera shot from a low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the seagull gliding smoothly with its wings fully spread rather than flapping violently?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video feature a slow tracking shot following the seagull?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -1.3159018754959106 + ], + "opposite_logprob_tokens": [ + -0.31590190529823303 + ], + "logprob_reward": -1.3159018754959106, + "yes_no_margin_reward": -1.0, + "token_credit_reward": -1.3159018754959106, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.023920567706227303, + 0.8505386710166931, + 0.04989432543516159, + 0.053271204233169556, + 0.001718863844871521, + 0.0015175966545939445, + 0.0018160403706133366, + 0.008663006126880646, + 0.004012199584394693, + 0.004647565074265003 + ], + "credit_map": [ + 0.000859651539940387, + 0.0029112868942320347, + 0.0004197196685709059, + 0.0005162445595487952, + 0.010969162918627262, + 0.0059618097729980946, + 0.00045325615792535245, + 0.0018294363981112838, + 0.01907281018793583, + 0.05854891240596771, + 0.006063897628337145, + 0.006087868940085173, + 0.4046062231063843, + 0.3325556218624115, + 0.00684279203414917, + 0.016760487109422684, + 0.0026926633436232805, + 0.003642046358436346, + 0.000862872286234051, + 0.0008724427898414433, + 0.022577350959181786, + 0.011846335604786873, + 0.0011063921265304089, + 0.0062942178919911385, + 0.00803162157535553, + 0.002090120455250144, + 0.0020746937952935696, + 0.0022825812920928, + 0.014054710976779461, + 0.017087573185563087, + 0.0017690687673166394, + 0.005880835000425577, + 0.00018070198711939156, + 0.0001889224222395569, + 0.0003356766828801483, + 0.0001830477121984586, + 0.00014592359366361052, + 0.0002051601477432996, + 0.0002690089459065348, + 0.00021042236767243594, + 0.00011890121822943911, + 0.00021827290765941143, + 0.00025553061277605593, + 0.00017463788390159607, + 0.00013774940453004092, + 0.0002491913619451225, + 0.0002351271832594648, + 0.00012818614777643234, + 0.00017336955352220684, + 0.00015490017540287226, + 0.0004249467747285962, + 0.0001238993281731382, + 0.0001421708584530279, + 0.00046387920156121254, + 0.00018923166499007493, + 0.00014364269736688584, + 0.0004570302553474903, + 0.00026515996432863176, + 0.0030160490423440933, + 0.00045265170047059655, + 0.0005007608560845256, + 0.0031323812436312437, + 0.00035255582770332694, + 0.00048641799367032945, + 0.0003544835199136287, + 0.0005218045553192496, + 0.0005532051436603069, + 0.0004720143333543092, + 0.0006112766568548977, + 0.0005398392677307129, + 0.0003001652657985687, + 0.0006594111328013241, + 0.0008172334055416286, + 0.00016339060675818473, + 0.0016482098726555705, + 0.00015478202840313315, + 0.00010783562174765393, + 0.0012533859116956592, + 6.451077933888882e-05, + 0.00043821672443300486 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 3, + 2 + ], + "selected_reward": -1.3159018754959106 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a pristine white seagull?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the seagull soaring beneath a vast azure sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the camera shot from a low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the seagull gliding smoothly with its wings fully spread rather than flapping violently?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video feature a slow tracking shot following the seagull?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0015282867243513465 + ], + "opposite_logprob_tokens": [ + -6.501528263092041 + ], + "logprob_reward": -0.0015282867243513465, + "yes_no_margin_reward": 6.5, + "token_credit_reward": -0.0015282867243513465, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.024189241230487823, + 0.879194438457489, + 0.043259017169475555, + 0.03323741257190704, + 0.0032450465951114893, + 0.0018788052257150412, + 0.0013406635262072086, + 0.008309932425618172, + 0.00396775035187602, + 0.0013777200365439057 + ], + "credit_map": [ + 0.0009075928246602416, + 0.002942525316029787, + 0.0005401277448982, + 0.0006503379554487765, + 0.010852042585611343, + 0.006584570277482271, + 0.0005595076363533735, + 0.001152537646703422, + 0.01963847503066063, + 0.06071760132908821, + 0.006275651045143604, + 0.006318948697298765, + 0.41868260502815247, + 0.3432091772556305, + 0.0070862541906535625, + 0.017265718430280685, + 0.002087036147713661, + 0.003244417253881693, + 0.0006788859609514475, + 0.0006426072213798761, + 0.019867869094014168, + 0.010251030325889587, + 0.0008277805172838271, + 0.005659391172230244, + 0.005557303316891193, + 0.0014586499892175198, + 0.0014560805866494775, + 0.0016292163636535406, + 0.007318106479942799, + 0.010125193744897842, + 0.0013108011335134506, + 0.004382063169032335, + 0.0002541775465942919, + 0.0002956166281364858, + 0.0006539491005241871, + 0.00047027584514580667, + 0.000326995417708531, + 0.00046596789616160095, + 0.0004964845720678568, + 0.00028157950146123767, + 0.0001458376063965261, + 0.00021900134743191302, + 0.00039381784154102206, + 0.00017157100955955684, + 0.00016898912144824862, + 0.000326909008435905, + 0.0002536805404815823, + 0.00019899864855688065, + 0.00013333099195733666, + 0.00011006493150489405, + 0.00020097494416404516, + 0.00020269611559342593, + 0.0001400670880684629, + 0.0001903028169181198, + 0.0002164679317502305, + 0.0001467587280785665, + 0.00042777013732120395, + 0.00018747017020359635, + 0.003282662248238921, + 0.00036745049874298275, + 0.0003936290158890188, + 0.0030208686366677284, + 0.0002689223038032651, + 0.00036115944385528564, + 0.0004149650048930198, + 0.0004471067222766578, + 0.0006187278195284307, + 0.00040705426363274455, + 0.0005655530258081853, + 0.0006167604005895555, + 0.00029400072526186705, + 0.0006035824189893901, + 0.0002458677627146244, + 8.537271060049534e-05, + 0.00027878445689566433, + 7.014347647782415e-05, + 9.032364323502406e-05, + 0.00035205381573177874, + 5.9038611652795225e-05, + 0.00019613554468378425 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 2, + 3 + ], + "selected_reward": -0.0015282867243513465 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a pristine white seagull?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the seagull soaring beneath a vast azure sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the camera shot from a low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the seagull gliding smoothly with its wings fully spread rather than flapping violently?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video feature a slow tracking shot following the seagull?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.08905211091041565 + ], + "opposite_logprob_tokens": [ + -2.464052200317383 + ], + "logprob_reward": -0.08905211091041565, + "yes_no_margin_reward": 2.375, + "token_credit_reward": -0.08905211091041565, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.04719538614153862, + 0.4722571074962616, + 0.26223987340927124, + 0.09705229848623276, + 0.014230575412511826, + 0.009031164459884167, + 0.009791111573576927, + 0.03709567338228226, + 0.011275481432676315, + 0.03983129933476448 + ], + "credit_map": [ + 0.0021247947588562965, + 0.005714342463761568, + 0.0013596265343949199, + 0.0023098019883036613, + 0.019044486805796623, + 0.0120287099853158, + 0.0020612734369933605, + 0.0025523516815155745, + 0.01156772393733263, + 0.03412103280425072, + 0.0035106393042951822, + 0.0033302619121968746, + 0.2262667715549469, + 0.18082281947135925, + 0.00385336228646338, + 0.008784503675997257, + 0.01065981574356556, + 0.021355217322707176, + 0.003380495123565197, + 0.0028443161863833666, + 0.12227899581193924, + 0.06041747331619263, + 0.004583540838211775, + 0.03672002628445625, + 0.005348679609596729, + 0.004640971310436726, + 0.0030781009700149298, + 0.0010253547225147486, + 0.044827524572610855, + 0.028453808277845383, + 0.001817203825339675, + 0.00786064937710762, + 0.0010896367020905018, + 0.0017901207320392132, + 0.0021340730600059032, + 0.002360581886023283, + 0.0016119916690513492, + 0.0019104031380265951, + 0.0023336834274232388, + 0.0010000846814364195, + 0.0008419712539762259, + 0.0011004912666976452, + 0.0014934191713109612, + 0.0013389440719038248, + 0.000895296863745898, + 0.001422307570464909, + 0.0012002375442534685, + 0.0007384967757388949, + 0.0009419948910363019, + 0.0010418504243716598, + 0.0014368604170158505, + 0.0013834278797730803, + 0.0011343605583533645, + 0.0014885927084833384, + 0.0013797793071717024, + 0.000984245096333325, + 0.0016153109027072787, + 0.000536032603122294, + 0.01587624102830887, + 0.0011273561976850033, + 0.0008540390990674496, + 0.014891527593135834, + 0.0005865959683433175, + 0.0016085694078356028, + 0.0009119416936300695, + 0.0012228375999256968, + 0.0013795135309919715, + 0.0010015690932050347, + 0.00147021294105798, + 0.0023531026672571898, + 0.0007349483203142881, + 0.0022013564594089985, + 0.006851051468402147, + 0.0015421034768223763, + 0.01413536537438631, + 0.0011272741248831153, + 0.0008957333047874272, + 0.01053470280021429, + 0.0006178472540341318, + 0.004127221647650003 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 2, + 3 + ], + "selected_reward": -0.08905211091041565 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a pristine white seagull?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the seagull soaring beneath a vast azure sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the camera shot from a low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the seagull gliding smoothly with its wings fully spread rather than flapping violently?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video feature a slow tracking shot following the seagull?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.00979085173457861 + ], + "opposite_logprob_tokens": [ + -4.634790897369385 + ], + "logprob_reward": -0.00979085173457861, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.00979085173457861, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.003891001222655177, + 0.8885776996612549, + 0.01891680434346199, + 0.0268479622900486, + 0.0016370383091270924, + 0.0010268846526741982, + 0.006659745238721371, + 0.004887384362518787, + 0.00249849003739655, + 0.0450570248067379 + ], + "credit_map": [ + 0.0003193114825990051, + 0.0006094931741245091, + 0.0004116087802685797, + 0.0004660513950511813, + 0.0008782354998402297, + 0.0006214347085915506, + 0.00028288166504353285, + 0.0003019842552021146, + 0.02023436687886715, + 0.0622137226164341, + 0.006369931623339653, + 0.0062710875645279884, + 0.4237768054008484, + 0.3454693853855133, + 0.007211716845631599, + 0.017030617222189903, + 0.0012926555937156081, + 0.0013562497915700078, + 0.0003405431052669883, + 0.0003205373650416732, + 0.008424586616456509, + 0.004546613898128271, + 0.00038827842217870057, + 0.002247338416054845, + 0.0030334542971104383, + 0.0009159145411103964, + 0.0008682937477715313, + 0.0009471988305449486, + 0.009187348186969757, + 0.009242599830031395, + 0.0007559087825939059, + 0.0018972421530634165, + 0.0001528137072455138, + 0.0002313997392775491, + 0.00023984011204447597, + 0.0002679622848518193, + 0.00013679079711437225, + 0.000181915849680081, + 0.0002507796452846378, + 0.00017553605721332133, + 9.193870209855959e-05, + 0.00012676922779064626, + 0.0001328993821516633, + 0.00016698642866685987, + 0.00011296714365016669, + 0.00012031571532133967, + 0.00016791728558018804, + 0.0001070907455869019, + 0.0006512760301120579, + 0.00036705739330500364, + 0.002113939030095935, + 0.0002809582802001387, + 0.0003588471154216677, + 0.0022680216934531927, + 0.00033283428638242185, + 0.00028681097319349647, + 0.00024363098782487214, + 0.00013625185238197446, + 0.0016315225511789322, + 0.00026862238883040845, + 0.0002814830222632736, + 0.0018037796253338456, + 0.0001832817360991612, + 0.00033881180570460856, + 0.00026022910606116056, + 0.0002536314132157713, + 0.00044380538747645915, + 0.00027392207994125783, + 0.000343071180395782, + 0.00036138491122983396, + 0.00020555147784762084, + 0.000356894510332495, + 0.007576957810670137, + 0.001751078525558114, + 0.016016865149140358, + 0.0012639494379982352, + 0.0010661293053999543, + 0.012044712901115417, + 0.0007436225423589349, + 0.004593708086758852 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 9, + 3 + ], + "selected_reward": -0.00979085173457861 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a pristine white seagull?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the seagull soaring beneath a vast azure sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the camera shot from a low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the seagull gliding smoothly with its wings fully spread rather than flapping violently?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video feature a slow tracking shot following the seagull?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.054932743310928345 + ], + "opposite_logprob_tokens": [ + -2.9299328327178955 + ], + "logprob_reward": -0.054932743310928345, + "yes_no_margin_reward": 2.875, + "token_credit_reward": -0.054932743310928345, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.010539487935602665, + 0.8886388540267944, + 0.034281618893146515, + 0.031482595950365067, + 0.003905967343598604, + 0.001839750912040472, + 0.0020959940738976, + 0.004712159279733896, + 0.004525722935795784, + 0.01797778718173504 + ], + "credit_map": [ + 0.0006452418747358024, + 0.0016724009765312076, + 0.0004675240197684616, + 0.00044086319394409657, + 0.00401034951210022, + 0.0010240019764751196, + 0.00045154205872677267, + 0.0018275654874742031, + 0.020187094807624817, + 0.06128271669149399, + 0.0063291676342487335, + 0.0063269794918596745, + 0.42266297340393066, + 0.3470417261123657, + 0.007089657709002495, + 0.01771865412592888, + 0.0017363684019073844, + 0.0029503630939871073, + 0.00044298009015619755, + 0.0005212336545810103, + 0.015606740489602089, + 0.006313050165772438, + 0.0007572215981781483, + 0.005953662563115358, + 0.001640444970689714, + 0.001926485332660377, + 0.0011617777636274695, + 0.0005264018545858562, + 0.014224760234355927, + 0.005981010384857655, + 0.000700238742865622, + 0.005321479868143797, + 0.0002754797169473022, + 0.000465946679469198, + 0.0004223667783662677, + 0.0006514597916975617, + 0.00048456902732141316, + 0.0005589188076555729, + 0.0006266803247854114, + 0.00042054642108269036, + 0.00016491285350639373, + 0.0002385957632213831, + 0.00029837325564585626, + 0.00017998612020164728, + 0.00020650865917559713, + 0.00034265892463736236, + 0.0002433897607261315, + 0.0001653256476856768, + 0.000260900822468102, + 0.0001800858590286225, + 0.00035592037602327764, + 0.0002286714006913826, + 0.0002041573025053367, + 0.00048097435501404107, + 0.00022874206479173154, + 0.00015654195158276707, + 0.0002524294250179082, + 9.774087811820209e-05, + 0.0015289249131456017, + 0.00013974733883515, + 0.0001531664456706494, + 0.0019986575935035944, + 7.91999336797744e-05, + 0.000462293071905151, + 0.0005549623165279627, + 0.0005380793008953333, + 0.0009997368324548006, + 0.0005165736074559391, + 0.0004194231878500432, + 0.0006026154733262956, + 0.00032172727514989674, + 0.0005726053495891392, + 0.0030435153748840094, + 0.0007212267373688519, + 0.00639286870136857, + 0.0005069124163128436, + 0.00042220880277454853, + 0.004997292999178171, + 0.0002640915336087346, + 0.001629671431146562 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 2, + 3 + ], + "selected_reward": -0.054932743310928345 + } + ], + "vlm_decoded_outputs": [ + "No", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000006/group_4/index_tensor([8270, 8270])_group_4_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000006/group_4/index_tensor([8270, 8270])_group_4_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..a2b675c7d756c8c20daa04bf5213c79c924a1562 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000006/group_4/index_tensor([8270, 8270])_group_4_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "drone shot, captured from a wide angle and from the side and rear, follows a heavily armored ancient warrior as he rides a jet-black, muscular steed, galloping along the Great Wall, which winds its way through the mountains. The warrior wears a phoenix-winged helmet with red tassels that dance wildly in the wind. His intricately carved bronze armor, adorned with mountain patterns, gleams with a cool metallic sheen as he moves, and the beast-headed shoulder guards adorning his shoulders project a majestic presence. His bronzed face is stern and resolute. He leans forward slightly, gripping the dark leather and brass-trimmed reins. His crimson cape, swirling in the wind, ripples like a flaming banner. The steed's hooves thunder, each powerful thump pounding against the weathered, massive stone pavement, stirring up a thick cloud of yellow dust and fine gravel, leaving a billowing trail behind him. It's dusk, and slanting golden sunlight paints the endless, towering mountains and the gray city walls that undulate like a giant dragon with a magnificent warm hue. Backlighting frames the swirling dust and mist with a shimmering golden edge, lengthening the lonely shadows of the men and horses cast on the city walls. The entire scene, framed by the grand composition, is imbued with a sense of epic grandeur and the urgency and loneliness of a mission, creating an atmosphere of solemnity and power.", + "seed": 20271, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000006/group_4/index_8270_seed_20271_dp_8_rank_16_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 4, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.025424876250326633, + "vlm_logprob": -0.025424876250326633, + "vlm_yes_no_margin": 3.975, + "vlm_token_credit": -0.025424876250326633, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a drone shot following the subject from the side and rear?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a heavily armored ancient warrior riding a jet-black muscular steed?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the warrior wear a phoenix-winged helmet with red tassels and intricately carved bronze armor with mountain patterns?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene set on the Great Wall winding through mountains at dusk with slanting golden sunlight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the steed gallop up the wall, kicking up a thick cloud of yellow dust and fine gravel behind it?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0360950343310833 + ], + "opposite_logprob_tokens": [ + -3.661095142364502 + ], + "logprob_reward": -0.0360950343310833, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.0360950343310833, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10585454106330872, + 0.14918997883796692, + 0.035538289695978165, + 0.03782178461551666, + 0.04143989086151123, + 0.031250547617673874, + 0.04189445078372955, + 0.1479753702878952, + 0.22062213718891144, + 0.18841300904750824 + ], + "credit_map": [ + 0.01290522888302803, + 0.011693636886775494, + 0.026702702045440674, + 0.012755650095641613, + 0.011715762317180634, + 0.014495151117444038, + 0.006887628231197596, + 0.008698787540197372, + 0.022776534780859947, + 0.012927843257784843, + 0.022180894389748573, + 0.008133940398693085, + 0.01552295871078968, + 0.05062321573495865, + 0.008334431797266006, + 0.008690153248608112, + 0.004085440654307604, + 0.005355833098292351, + 0.004764647223055363, + 0.003017229260876775, + 0.005085294600576162, + 0.005951059050858021, + 0.003699947614222765, + 0.0035788393579423428, + 0.004555886145681143, + 0.0052794646471738815, + 0.00747234933078289, + 0.0030341350939124823, + 0.0069039673544466496, + 0.004849553108215332, + 0.003104406874626875, + 0.002622022293508053, + 0.009434741921722889, + 0.003361478913575411, + 0.007207025773823261, + 0.0028388281352818012, + 0.0049322606064379215, + 0.004892940167337656, + 0.0031000180169939995, + 0.00567259918898344, + 0.004566336516290903, + 0.003375427098944783, + 0.00372853665612638, + 0.0029989066533744335, + 0.0031198030337691307, + 0.005725752096623182, + 0.002496168715879321, + 0.005239615682512522, + 0.0046114856377244, + 0.005530559457838535, + 0.00523228757083416, + 0.004419639240950346, + 0.006846279371529818, + 0.006710860878229141, + 0.003577937139198184, + 0.004965402651578188, + 0.015559680759906769, + 0.025629499927163124, + 0.016847534105181694, + 0.011355898343026638, + 0.022261586040258408, + 0.024739010259509087, + 0.010999491438269615, + 0.020582662895321846, + 0.008821156807243824, + 0.009232607670128345, + 0.028866244480013847, + 0.007922596298158169, + 0.03036116249859333, + 0.06766457110643387, + 0.006179964169859886, + 0.0615738183259964, + 0.010053351521492004, + 0.005393967032432556, + 0.07237285375595093, + 0.00796546321362257, + 0.010310962796211243, + 0.010205796919763088, + 0.006288627628237009, + 0.06582198292016983 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 9, + 1 + ], + "selected_reward": -0.0360950343310833 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a drone shot following the subject from the side and rear?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a heavily armored ancient warrior riding a jet-black muscular steed?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the warrior wear a phoenix-winged helmet with red tassels and intricately carved bronze armor with mountain patterns?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene set on the Great Wall winding through mountains at dusk with slanting golden sunlight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the steed gallop up the wall, kicking up a thick cloud of yellow dust and fine gravel behind it?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.043063513934612274 + ], + "opposite_logprob_tokens": [ + -3.1680634021759033 + ], + "logprob_reward": -0.043063513934612274, + "yes_no_margin_reward": 3.125, + "token_credit_reward": -0.043063513934612274, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10331273078918457, + 0.06396263837814331, + 0.08208669722080231, + 0.06044887751340866, + 0.09443433582782745, + 0.030032508075237274, + 0.059848081320524216, + 0.026587067171931267, + 0.049721721559762955, + 0.42956528067588806 + ], + "credit_map": [ + 0.016670117154717445, + 0.013669256120920181, + 0.01472887396812439, + 0.010194629430770874, + 0.0116281034424901, + 0.019401181489229202, + 0.009086165577173233, + 0.007934405468404293, + 0.009702634997665882, + 0.006965259090065956, + 0.010759701021015644, + 0.004253358580172062, + 0.006727991625666618, + 0.015366332605481148, + 0.005922453943639994, + 0.004264905117452145, + 0.008139587938785553, + 0.0080940630286932, + 0.00915368739515543, + 0.0054927607998251915, + 0.017885809764266014, + 0.016842510551214218, + 0.008621352724730968, + 0.007856925949454308, + 0.007984288968145847, + 0.007495480123907328, + 0.006430847570300102, + 0.004138976335525513, + 0.01172107458114624, + 0.014547123573720455, + 0.004265759140253067, + 0.003865331644192338, + 0.013077705167233944, + 0.011918693780899048, + 0.015683960169553757, + 0.007734545040875673, + 0.012842230498790741, + 0.012054109945893288, + 0.010135289281606674, + 0.010987803339958191, + 0.0034488108940422535, + 0.0024346434511244297, + 0.005573736969381571, + 0.001998761435970664, + 0.0028464847709983587, + 0.006329797208309174, + 0.0022583394311368465, + 0.005141937639564276, + 0.005174302961677313, + 0.009878910146653652, + 0.005519077647477388, + 0.005582103505730629, + 0.014473224990069866, + 0.006317498628050089, + 0.006115322932600975, + 0.006787638179957867, + 0.00310114654712379, + 0.005255231633782387, + 0.0027159973978996277, + 0.0031192521564662457, + 0.00393733661621809, + 0.0034461459144949913, + 0.0028652024921029806, + 0.0021467532496899366, + 0.0029608288314193487, + 0.004448368214070797, + 0.006634608842432499, + 0.0020873446483165026, + 0.006674835458397865, + 0.012898892164230347, + 0.0021486144978553057, + 0.011868229135870934, + 0.02467067539691925, + 0.008366669528186321, + 0.17347437143325806, + 0.016022413969039917, + 0.023381877690553665, + 0.022866342216730118, + 0.011100051924586296, + 0.149682879447937 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 0, + 4 + ], + "selected_reward": -0.043063513934612274 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a drone shot following the subject from the side and rear?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a heavily armored ancient warrior riding a jet-black muscular steed?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the warrior wear a phoenix-winged helmet with red tassels and intricately carved bronze armor with mountain patterns?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene set on the Great Wall winding through mountains at dusk with slanting golden sunlight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the steed gallop up the wall, kicking up a thick cloud of yellow dust and fine gravel behind it?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.020590247586369514 + ], + "opposite_logprob_tokens": [ + -3.895590305328369 + ], + "logprob_reward": -0.020590247586369514, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.020590247586369514, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06696051359176636, + 0.05660991370677948, + 0.046137213706970215, + 0.042910654097795486, + 0.05404379963874817, + 0.053232911974191666, + 0.05654015764594078, + 0.030504390597343445, + 0.07716808468103409, + 0.5158924460411072 + ], + "credit_map": [ + 0.008784291334450245, + 0.010528362356126308, + 0.007458762731403112, + 0.0071921395137906075, + 0.009924130514264107, + 0.009509598836302757, + 0.006205444224178791, + 0.0073577845469117165, + 0.007885819301009178, + 0.004208986181765795, + 0.008705670945346355, + 0.0034471959806978703, + 0.005905485711991787, + 0.019891470670700073, + 0.0028944392688572407, + 0.003670845413580537, + 0.0052103316411376, + 0.010346837341785431, + 0.0044191814959049225, + 0.003402415895834565, + 0.007920113392174244, + 0.005975143983960152, + 0.005344507750123739, + 0.003518678480759263, + 0.003784423228353262, + 0.005610376596450806, + 0.004103309474885464, + 0.004171245731413364, + 0.010708796791732311, + 0.007691292557865381, + 0.003846685867756605, + 0.002994521288201213, + 0.008312512189149857, + 0.005889797583222389, + 0.010575919412076473, + 0.0040427315980196, + 0.00716050248593092, + 0.006807727739214897, + 0.004586352966725826, + 0.00666824821382761, + 0.005648534744977951, + 0.005470061209052801, + 0.008597261272370815, + 0.0035838333424180746, + 0.0067053791135549545, + 0.010488596744835377, + 0.003924650605767965, + 0.008814592845737934, + 0.006087177433073521, + 0.0074624535627663136, + 0.006085263565182686, + 0.005803086329251528, + 0.0073743159882724285, + 0.010028201155364513, + 0.006670473609119654, + 0.007029181346297264, + 0.0035413748119026423, + 0.006071686279028654, + 0.002809955272823572, + 0.0026545345317572355, + 0.005501306150108576, + 0.004538238979876041, + 0.002346422988921404, + 0.0030408683232963085, + 0.0036634388379752636, + 0.006021576002240181, + 0.008746704086661339, + 0.003336144145578146, + 0.01082437764853239, + 0.022883184254169464, + 0.003244092920795083, + 0.018448565155267715, + 0.02925257943570614, + 0.010261389426887035, + 0.20755767822265625, + 0.019746284931898117, + 0.02690119668841362, + 0.027674822136759758, + 0.013754544779658318, + 0.18074390292167664 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 8, + 0 + ], + "selected_reward": -0.020590247586369514 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a drone shot following the subject from the side and rear?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a heavily armored ancient warrior riding a jet-black muscular steed?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the warrior wear a phoenix-winged helmet with red tassels and intricately carved bronze armor with mountain patterns?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene set on the Great Wall winding through mountains at dusk with slanting golden sunlight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the steed gallop up the wall, kicking up a thick cloud of yellow dust and fine gravel behind it?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.023283764719963074 + ], + "opposite_logprob_tokens": [ + -3.7732837200164795 + ], + "logprob_reward": -0.023283764719963074, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.023283764719963074, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10890769213438034, + 0.037050243467092514, + 0.031704131513834, + 0.07015862315893173, + 0.06913438439369202, + 0.04750583693385124, + 0.07701926678419113, + 0.0566839724779129, + 0.16549257934093475, + 0.3363431990146637 + ], + "credit_map": [ + 0.017089832574129105, + 0.011820903047919273, + 0.02322537451982498, + 0.00984430406242609, + 0.011795050464570522, + 0.019707750529050827, + 0.007011251524090767, + 0.008413229137659073, + 0.006620835047215223, + 0.0059602404944598675, + 0.004490204155445099, + 0.002457062480971217, + 0.003760562976822257, + 0.009321990422904491, + 0.002435604576021433, + 0.002003742614760995, + 0.004394032061100006, + 0.004338840022683144, + 0.004379018675535917, + 0.0024399978574365377, + 0.005433831363916397, + 0.003987708128988743, + 0.003162052482366562, + 0.003568649524822831, + 0.010140758007764816, + 0.011686853133141994, + 0.011331661604344845, + 0.005816776771098375, + 0.012160798534750938, + 0.010048174299299717, + 0.005297280848026276, + 0.0036763246171176434, + 0.010815992020070553, + 0.006412365473806858, + 0.011492338962852955, + 0.004947790410369635, + 0.008216856978833675, + 0.00869509857147932, + 0.006092827767133713, + 0.012461117468774319, + 0.006935551762580872, + 0.00480954023078084, + 0.004429253749549389, + 0.0029667632188647985, + 0.006011656019836664, + 0.013598449528217316, + 0.0027378229424357414, + 0.006016802974045277, + 0.008297593332827091, + 0.010276076383888721, + 0.006360435858368874, + 0.0046918983571231365, + 0.018708089366555214, + 0.012796725146472454, + 0.005712478421628475, + 0.010175971314311028, + 0.006175583694130182, + 0.011170120909810066, + 0.005677843000739813, + 0.005051648244261742, + 0.008852953091263771, + 0.00847031269222498, + 0.0052934796549379826, + 0.005992032587528229, + 0.006155125796794891, + 0.0083662960678339, + 0.020892972126603127, + 0.005725923925638199, + 0.022342076525092125, + 0.05155456066131592, + 0.004679636564105749, + 0.04577600210905075, + 0.026685066521167755, + 0.0102152144536376, + 0.12851326167583466, + 0.013528021052479744, + 0.018181566148996353, + 0.01864687353372574, + 0.010222130455076694, + 0.11035105586051941 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 8, + 0 + ], + "selected_reward": -0.023283764719963074 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a drone shot following the subject from the side and rear?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a heavily armored ancient warrior riding a jet-black muscular steed?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the warrior wear a phoenix-winged helmet with red tassels and intricately carved bronze armor with mountain patterns?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene set on the Great Wall winding through mountains at dusk with slanting golden sunlight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the steed gallop up the wall, kicking up a thick cloud of yellow dust and fine gravel behind it?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004091820679605007 + ], + "opposite_logprob_tokens": [ + -5.504091739654541 + ], + "logprob_reward": -0.004091820679605007, + "yes_no_margin_reward": 5.5, + "token_credit_reward": -0.004091820679605007, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10635840147733688, + 0.045180901885032654, + 0.021653056144714355, + 0.05701504647731781, + 0.10651478916406631, + 0.03642716631293297, + 0.126181960105896, + 0.03189555183053017, + 0.04705486819148064, + 0.4217182397842407 + ], + "credit_map": [ + 0.015539023093879223, + 0.013898794539272785, + 0.016889795660972595, + 0.012267690151929855, + 0.015355050563812256, + 0.014972478151321411, + 0.008366886526346207, + 0.00906867254525423, + 0.00518229603767395, + 0.004044137895107269, + 0.008950580842792988, + 0.0031268226448446512, + 0.004919876344501972, + 0.01344246231019497, + 0.0029295256827026606, + 0.002585198963060975, + 0.0030510560609400272, + 0.0030294808093458414, + 0.0026903627440333366, + 0.001777784200385213, + 0.003900731448084116, + 0.0033288393169641495, + 0.0024389170575886965, + 0.0014358853222802281, + 0.00983851496130228, + 0.007394011598080397, + 0.010735791176557541, + 0.0044880821369588375, + 0.007925158366560936, + 0.008506163023412228, + 0.004004110582172871, + 0.004123212303966284, + 0.020021071657538414, + 0.009880464524030685, + 0.02223704569041729, + 0.008974811993539333, + 0.01104937493801117, + 0.01612173393368721, + 0.007542894687503576, + 0.010687397792935371, + 0.005770890507847071, + 0.003207409754395485, + 0.0042175911366939545, + 0.0022328533232212067, + 0.00444948673248291, + 0.009693543426692486, + 0.00267535331659019, + 0.004180040210485458, + 0.010951547883450985, + 0.01836743950843811, + 0.009432490915060043, + 0.009180350229144096, + 0.03653080761432648, + 0.019115900620818138, + 0.008683167397975922, + 0.013920257799327374, + 0.003415487939491868, + 0.004391555208712816, + 0.0038670666981488466, + 0.0020087873563170433, + 0.00431937538087368, + 0.00721709942445159, + 0.0018547414802014828, + 0.004821438807994127, + 0.005521205719560385, + 0.00426842737942934, + 0.008553946390748024, + 0.0038589707110077143, + 0.004418083932250738, + 0.007762218359857798, + 0.002282301662489772, + 0.010389714501798153, + 0.023238614201545715, + 0.007553120143711567, + 0.1712695211172104, + 0.01651252619922161, + 0.022478044033050537, + 0.02220860682427883, + 0.010756377130746841, + 0.1477014273405075 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 6, + 4 + ], + "selected_reward": -0.004091820679605007 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "drone shot, captured from a wide angle and from the side and rear, follows a heavily armored ancient warrior as he rides a jet-black, muscular steed, galloping along the Great Wall, which winds its way through the mountains. The warrior wears a phoenix-winged helmet with red tassels that dance wildly in the wind. His intricately carved bronze armor, adorned with mountain patterns, gleams with a cool metallic sheen as he moves, and the beast-headed shoulder guards adorning his shoulders project a majestic presence. His bronzed face is stern and resolute. He leans forward slightly, gripping the dark leather and brass-trimmed reins. His crimson cape, swirling in the wind, ripples like a flaming banner. The steed's hooves thunder, each powerful thump pounding against the weathered, massive stone pavement, stirring up a thick cloud of yellow dust and fine gravel, leaving a billowing trail behind him. It's dusk, and slanting golden sunlight paints the endless, towering mountains and the gray city walls that undulate like a giant dragon with a magnificent warm hue. Backlighting frames the swirling dust and mist with a shimmering golden edge, lengthening the lonely shadows of the men and horses cast on the city walls. The entire scene, framed by the grand composition, is imbued with a sense of epic grandeur and the urgency and loneliness of a mission, creating an atmosphere of solemnity and power.", + "seed": 20272, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000006/group_4/index_8270_seed_20272_dp_8_rank_16_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 4, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.05342442747205496, + "vlm_logprob": -0.05342442747205496, + "vlm_yes_no_margin": 3.175, + "vlm_token_credit": -0.05342442747205496, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a drone shot following the subject from the side and rear?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a heavily armored ancient warrior riding a jet-black muscular steed?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the warrior wear a phoenix-winged helmet with red tassels and intricately carved bronze armor with mountain patterns?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene set on the Great Wall winding through mountains at dusk with slanting golden sunlight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the steed gallop up the wall, kicking up a thick cloud of yellow dust and fine gravel behind it?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07436845451593399 + ], + "opposite_logprob_tokens": [ + -2.824368476867676 + ], + "logprob_reward": -0.07436845451593399, + "yes_no_margin_reward": 2.75, + "token_credit_reward": -0.07436845451593399, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.18078018724918365, + 0.049340613186359406, + 0.049994636327028275, + 0.03968038409948349, + 0.07806917279958725, + 0.057315096259117126, + 0.055812686681747437, + 0.03796929866075516, + 0.4394510090351105, + 0.011587012559175491 + ], + "credit_map": [ + 0.012685606256127357, + 0.04692106693983078, + 0.013931147754192352, + 0.012467442080378532, + 0.014402925968170166, + 0.007404942996799946, + 0.050414904952049255, + 0.022552145645022392, + 0.004157642833888531, + 0.006255677435547113, + 0.004635806195437908, + 0.005498848389834166, + 0.004340797662734985, + 0.004896790254861116, + 0.013677556067705154, + 0.0058774929493665695, + 0.004630596376955509, + 0.01069545652717352, + 0.007494450081139803, + 0.004293503239750862, + 0.004517985507845879, + 0.005164495203644037, + 0.007393199019134045, + 0.005804950837045908, + 0.0045518227852880955, + 0.006168873980641365, + 0.008388252928853035, + 0.004139711149036884, + 0.0039891935884952545, + 0.003866303712129593, + 0.004459635820239782, + 0.004116591531783342, + 0.010431942529976368, + 0.011768706142902374, + 0.008709711953997612, + 0.009592303074896336, + 0.010938277468085289, + 0.014049380086362362, + 0.006340006832033396, + 0.0062388465739786625, + 0.006273583509027958, + 0.00992567278444767, + 0.01077488623559475, + 0.004483488388359547, + 0.005524095147848129, + 0.011091905646026134, + 0.005519067868590355, + 0.0037223966792225838, + 0.00717524578794837, + 0.007485933601856232, + 0.0059526958502829075, + 0.005194223485887051, + 0.0090559720993042, + 0.010338112711906433, + 0.004894358571618795, + 0.005716146435588598, + 0.00491492310538888, + 0.0074957674369215965, + 0.004291072022169828, + 0.004554138518869877, + 0.004339142236858606, + 0.004087452311068773, + 0.005542191211134195, + 0.002744609722867608, + 0.03662734478712082, + 0.07117890566587448, + 0.04973001778125763, + 0.05377034842967987, + 0.04332536831498146, + 0.08321741968393326, + 0.056658849120140076, + 0.04494274780154228, + 0.0018883395241573453, + 0.0020222216844558716, + 0.0012399713741615415, + 0.0009759289096109569, + 0.0016971585573628545, + 0.0011914934730157256, + 0.0013581052189692855, + 0.0012137934099882841 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 0, + 4 + ], + "selected_reward": -0.07436845451593399 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a drone shot following the subject from the side and rear?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a heavily armored ancient warrior riding a jet-black muscular steed?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the warrior wear a phoenix-winged helmet with red tassels and intricately carved bronze armor with mountain patterns?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene set on the Great Wall winding through mountains at dusk with slanting golden sunlight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the steed gallop up the wall, kicking up a thick cloud of yellow dust and fine gravel behind it?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.10031269490718842 + ], + "opposite_logprob_tokens": [ + -2.3503127098083496 + ], + "logprob_reward": -0.10031269490718842, + "yes_no_margin_reward": 2.25, + "token_credit_reward": -0.10031269490718842, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08055981993675232, + 0.1435844749212265, + 0.09467577189207077, + 0.036381129175424576, + 0.021235311403870583, + 0.03388961777091026, + 0.08470112085342407, + 0.04597833752632141, + 0.4421030580997467, + 0.016891270875930786 + ], + "credit_map": [ + 0.0070386845618486404, + 0.008669160306453705, + 0.01223727222532034, + 0.006492582615464926, + 0.013061587698757648, + 0.005028535611927509, + 0.01458472665399313, + 0.013447273522615433, + 0.012576395645737648, + 0.02249361015856266, + 0.015751687809824944, + 0.0180728230625391, + 0.014237243682146072, + 0.010885119438171387, + 0.03273896127939224, + 0.0168286319822073, + 0.010029182769358158, + 0.022183042019605637, + 0.014331125654280186, + 0.00891813263297081, + 0.009172996506094933, + 0.008547822013497353, + 0.008475312031805515, + 0.013018161058425903, + 0.004600140731781721, + 0.0068992143496870995, + 0.005256801377981901, + 0.003962811082601547, + 0.005015967879444361, + 0.003710633609443903, + 0.003602983895689249, + 0.0033325774129480124, + 0.002568094292655587, + 0.003629129845649004, + 0.0030420683324337006, + 0.0014741943450644612, + 0.0026968042366206646, + 0.0033072796650230885, + 0.001847592880949378, + 0.0026701479218900204, + 0.004085661843419075, + 0.005533058661967516, + 0.004070849157869816, + 0.003575259819626808, + 0.0049858237616717815, + 0.004538478329777718, + 0.0043333545327186584, + 0.002767134690657258, + 0.01199977844953537, + 0.013909672386944294, + 0.009633950889110565, + 0.011059179902076721, + 0.011373067274689674, + 0.009971758350729942, + 0.007570783607661724, + 0.009182930923998356, + 0.005072079598903656, + 0.01041359081864357, + 0.004229321144521236, + 0.004915099591016769, + 0.007591903675347567, + 0.005933569278568029, + 0.004334317520260811, + 0.0034884510096162558, + 0.03380126133561134, + 0.07609926909208298, + 0.05349113047122955, + 0.044739775359630585, + 0.043203044682741165, + 0.0932513177394867, + 0.05109366774559021, + 0.04642362520098686, + 0.001763605629093945, + 0.0023886605631560087, + 0.0015474016545340419, + 0.002544742776080966, + 0.0023644128814339638, + 0.0018198222387582064, + 0.0026911350432783365, + 0.0017714890418574214 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 1, + 2 + ], + "selected_reward": -0.10031269490718842 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a drone shot following the subject from the side and rear?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a heavily armored ancient warrior riding a jet-black muscular steed?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the warrior wear a phoenix-winged helmet with red tassels and intricately carved bronze armor with mountain patterns?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene set on the Great Wall winding through mountains at dusk with slanting golden sunlight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the steed gallop up the wall, kicking up a thick cloud of yellow dust and fine gravel behind it?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.014205760322511196 + ], + "opposite_logprob_tokens": [ + -4.2642059326171875 + ], + "logprob_reward": -0.014205760322511196, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.014205760322511196, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2897535562515259, + 0.17525433003902435, + 0.07410724461078644, + 0.03853578492999077, + 0.11904866993427277, + 0.04267062246799469, + 0.06869586557149887, + 0.05711067467927933, + 0.11643040925264359, + 0.018392877653241158 + ], + "credit_map": [ + 0.024320200085639954, + 0.0712854191660881, + 0.04181641340255737, + 0.012028664350509644, + 0.038610029965639114, + 0.01475831400603056, + 0.05842730775475502, + 0.028507210314273834, + 0.016370302066206932, + 0.02460433542728424, + 0.01583302579820156, + 0.01526353508234024, + 0.01844147779047489, + 0.012804499827325344, + 0.04898030310869217, + 0.02295684441924095, + 0.007871215231716633, + 0.010025907307863235, + 0.01599854603409767, + 0.005487017333507538, + 0.007289263419806957, + 0.00844134483486414, + 0.008448178879916668, + 0.010545768775045872, + 0.0043315589427948, + 0.006737579591572285, + 0.006058902479708195, + 0.0037372824735939503, + 0.005451429169625044, + 0.0042157587595283985, + 0.004477426875382662, + 0.0035258440766483545, + 0.014609514735639095, + 0.01416972279548645, + 0.017775041982531548, + 0.010108018293976784, + 0.017796820029616356, + 0.02358277700841427, + 0.008849474601447582, + 0.012157297693192959, + 0.005060699302703142, + 0.010499711148440838, + 0.0040685394778847694, + 0.0051958682015538216, + 0.0051008788868784904, + 0.00426497170701623, + 0.0052046701312065125, + 0.003275286639109254, + 0.0110018877312541, + 0.01184770092368126, + 0.007267928682267666, + 0.006523603573441505, + 0.00967127364128828, + 0.009663408622145653, + 0.004788768943399191, + 0.007931292057037354, + 0.007375694811344147, + 0.005884120240807533, + 0.005457928404211998, + 0.006231990642845631, + 0.008909638039767742, + 0.005366363096982241, + 0.010883874259889126, + 0.007001062855124474, + 0.01409732922911644, + 0.013487229123711586, + 0.01042648870497942, + 0.012824236415326595, + 0.015634959563612938, + 0.01217184029519558, + 0.01937728002667427, + 0.018411049619317055, + 0.0026732790283858776, + 0.0032623123843222857, + 0.0021880227141082287, + 0.0019911655690521, + 0.0025245705619454384, + 0.0013719573616981506, + 0.0025391848757863045, + 0.0018423845758661628 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 4 + ], + "selected_reward": -0.014205760322511196 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a drone shot following the subject from the side and rear?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a heavily armored ancient warrior riding a jet-black muscular steed?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the warrior wear a phoenix-winged helmet with red tassels and intricately carved bronze armor with mountain patterns?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene set on the Great Wall winding through mountains at dusk with slanting golden sunlight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the steed gallop up the wall, kicking up a thick cloud of yellow dust and fine gravel behind it?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.054951928555965424 + ], + "opposite_logprob_tokens": [ + -2.9299519062042236 + ], + "logprob_reward": -0.054951928555965424, + "yes_no_margin_reward": 2.875, + "token_credit_reward": -0.054951928555965424, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.16271159052848816, + 0.20586079359054565, + 0.07147997617721558, + 0.12872709333896637, + 0.01583251543343067, + 0.11192005127668381, + 0.03759288415312767, + 0.031218504533171654, + 0.21966899931430817, + 0.014987525530159473 + ], + "credit_map": [ + 0.014667211100459099, + 0.030553575605154037, + 0.025941019877791405, + 0.009450992569327354, + 0.024791553616523743, + 0.013126778416335583, + 0.020452218130230904, + 0.023728258907794952, + 0.016667520627379417, + 0.016550257802009583, + 0.016094962134957314, + 0.012023872695863247, + 0.022084912285208702, + 0.014278829097747803, + 0.07507915794849396, + 0.0330812968313694, + 0.006071961484849453, + 0.00912017747759819, + 0.014813889749348164, + 0.004728259053081274, + 0.006002840120345354, + 0.010044035501778126, + 0.010423743166029453, + 0.010275081731379032, + 0.01323196291923523, + 0.02055908739566803, + 0.021040046587586403, + 0.017267338931560516, + 0.018603168427944183, + 0.012301173992455006, + 0.012980127707123756, + 0.012744196690618992, + 0.0017745496006682515, + 0.0020945484284311533, + 0.002150186337530613, + 0.0015475294785574079, + 0.002482274081557989, + 0.0022396096028387547, + 0.002002137713134289, + 0.00154168123845011, + 0.012231146916747093, + 0.008591757155954838, + 0.022139079868793488, + 0.00927599985152483, + 0.008947540074586868, + 0.023927360773086548, + 0.015602873638272285, + 0.011204296723008156, + 0.005144895985722542, + 0.006688457448035479, + 0.004067874047905207, + 0.004325470887124538, + 0.0046594333834946156, + 0.005584387574344873, + 0.0027894414961338043, + 0.004332925658673048, + 0.004721547942608595, + 0.004628621973097324, + 0.0030068648047745228, + 0.003166500711813569, + 0.005057551898062229, + 0.004366897512227297, + 0.0037314496003091335, + 0.002539073582738638, + 0.019334958866238594, + 0.03649817034602165, + 0.024567652493715286, + 0.025269463658332825, + 0.022196386009454727, + 0.04229205474257469, + 0.026564979925751686, + 0.02294536866247654, + 0.0022365476470440626, + 0.0032811167184263468, + 0.00174910353962332, + 0.00147829984780401, + 0.001708118012174964, + 0.0013644990976899862, + 0.0017438166541978717, + 0.0014260252937674522 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 1, + 0 + ], + "selected_reward": -0.054951928555965424 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a drone shot following the subject from the side and rear?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a heavily armored ancient warrior riding a jet-black muscular steed?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the warrior wear a phoenix-winged helmet with red tassels and intricately carved bronze armor with mountain patterns?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene set on the Great Wall winding through mountains at dusk with slanting golden sunlight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the steed gallop up the wall, kicking up a thick cloud of yellow dust and fine gravel behind it?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.023283299058675766 + ], + "opposite_logprob_tokens": [ + -3.7732832431793213 + ], + "logprob_reward": -0.023283299058675766, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.023283299058675766, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06459221988916397, + 0.0511474609375, + 0.19189703464508057, + 0.04833458364009857, + 0.05916094034910202, + 0.0602155439555645, + 0.03966116905212402, + 0.03510711342096329, + 0.42869874835014343, + 0.021185170859098434 + ], + "credit_map": [ + 0.005690909922122955, + 0.01383108552545309, + 0.01050128135830164, + 0.005969175137579441, + 0.008653576485812664, + 0.004730681888759136, + 0.007121679838746786, + 0.008093830198049545, + 0.00461912527680397, + 0.005666565615683794, + 0.004999231081455946, + 0.004840826150029898, + 0.005194259807467461, + 0.0049200598150491714, + 0.014735104516148567, + 0.006172286346554756, + 0.020861299708485603, + 0.017179543152451515, + 0.04970376938581467, + 0.011285705491900444, + 0.014396591112017632, + 0.025063419714570045, + 0.024804413318634033, + 0.02860228344798088, + 0.004938435275107622, + 0.005856214091181755, + 0.014030209742486477, + 0.003953895531594753, + 0.004935775883495808, + 0.004148469306528568, + 0.004689195193350315, + 0.005782389082014561, + 0.006495961453765631, + 0.006642184220254421, + 0.013064932078123093, + 0.005816289223730564, + 0.00700216693803668, + 0.006938606034964323, + 0.005614741705358028, + 0.007586057763546705, + 0.006090288981795311, + 0.010451260022819042, + 0.011233174242079258, + 0.00610562926158309, + 0.005479269661009312, + 0.011121880263090134, + 0.006256202235817909, + 0.003477838123217225, + 0.003746730973944068, + 0.005846097134053707, + 0.003974284045398235, + 0.00434260256588459, + 0.00705281738191843, + 0.006483221892267466, + 0.003119290340691805, + 0.00509612588211894, + 0.004423491656780243, + 0.004762249067425728, + 0.003477246267721057, + 0.004051049239933491, + 0.004567011725157499, + 0.0035059871152043343, + 0.006615306716412306, + 0.003704771865159273, + 0.038133081048727036, + 0.06916108727455139, + 0.04692531004548073, + 0.052953798323869705, + 0.04252753034234047, + 0.08090691268444061, + 0.05549658462405205, + 0.042594414204359055, + 0.0028483732603490353, + 0.004977376665920019, + 0.002634697128087282, + 0.0021491285879164934, + 0.0021379245445132256, + 0.0020786256063729525, + 0.0021953301038593054, + 0.0021637154277414083 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 2, + 0 + ], + "selected_reward": -0.023283299058675766 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "drone shot, captured from a wide angle and from the side and rear, follows a heavily armored ancient warrior as he rides a jet-black, muscular steed, galloping along the Great Wall, which winds its way through the mountains. The warrior wears a phoenix-winged helmet with red tassels that dance wildly in the wind. His intricately carved bronze armor, adorned with mountain patterns, gleams with a cool metallic sheen as he moves, and the beast-headed shoulder guards adorning his shoulders project a majestic presence. His bronzed face is stern and resolute. He leans forward slightly, gripping the dark leather and brass-trimmed reins. His crimson cape, swirling in the wind, ripples like a flaming banner. The steed's hooves thunder, each powerful thump pounding against the weathered, massive stone pavement, stirring up a thick cloud of yellow dust and fine gravel, leaving a billowing trail behind him. It's dusk, and slanting golden sunlight paints the endless, towering mountains and the gray city walls that undulate like a giant dragon with a magnificent warm hue. Backlighting frames the swirling dust and mist with a shimmering golden edge, lengthening the lonely shadows of the men and horses cast on the city walls. The entire scene, framed by the grand composition, is imbued with a sense of epic grandeur and the urgency and loneliness of a mission, creating an atmosphere of solemnity and power.", + "seed": 20273, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000006/group_4/index_8270_seed_20273_dp_9_rank_18_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 4, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.03762575937435031, + "vlm_logprob": -0.03762575937435031, + "vlm_yes_no_margin": 3.75, + "vlm_token_credit": -0.03762575937435031, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a drone shot following the subject from the side and rear?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a heavily armored ancient warrior riding a jet-black muscular steed?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the warrior wear a phoenix-winged helmet with red tassels and intricately carved bronze armor with mountain patterns?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene set on the Great Wall winding through mountains at dusk with slanting golden sunlight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the steed gallop up the wall, kicking up a thick cloud of yellow dust and fine gravel behind it?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.06101521477103233 + ], + "opposite_logprob_tokens": [ + -3.0610151290893555 + ], + "logprob_reward": -0.06101521477103233, + "yes_no_margin_reward": 3.0, + "token_credit_reward": -0.06101521477103233, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1023169681429863, + 0.12414547055959702, + 0.10699785500764847, + 0.046066831797361374, + 0.10451704263687134, + 0.13882367312908173, + 0.20699919760227203, + 0.07162465155124664, + 0.06923940777778625, + 0.029268890619277954 + ], + "credit_map": [ + 0.009064302779734135, + 0.008029739372432232, + 0.014814399182796478, + 0.021160133183002472, + 0.009094628505408764, + 0.014946219511330128, + 0.012421279214322567, + 0.0127862598747015, + 0.011564334854483604, + 0.012466546148061752, + 0.012200859375298023, + 0.018562210723757744, + 0.01698928512632847, + 0.015318512916564941, + 0.018791716545820236, + 0.01825200766324997, + 0.009986492805182934, + 0.012064616195857525, + 0.009669612161815166, + 0.015905199572443962, + 0.01448468491435051, + 0.010895638726651669, + 0.01757843606173992, + 0.01641318015754223, + 0.008245546370744705, + 0.006063250824809074, + 0.003784866537898779, + 0.007742026355117559, + 0.005352315492928028, + 0.003985520917922258, + 0.005672397091984749, + 0.005220904480665922, + 0.015076128765940666, + 0.015027721412479877, + 0.010890427976846695, + 0.023658709600567818, + 0.010183797217905521, + 0.010880433022975922, + 0.010611100122332573, + 0.00818872544914484, + 0.01271833200007677, + 0.023126669228076935, + 0.01798485592007637, + 0.026232466101646423, + 0.011445633135735989, + 0.022845279425382614, + 0.011712887324392796, + 0.012757544405758381, + 0.017585918307304382, + 0.027095014229416847, + 0.028011195361614227, + 0.02618914283812046, + 0.029366854578256607, + 0.027749162167310715, + 0.029430054128170013, + 0.021571852266788483, + 0.004372543189674616, + 0.0037793591618537903, + 0.013995303772389889, + 0.016545381397008896, + 0.0094094667583704, + 0.006941779516637325, + 0.008804072625935078, + 0.0077767460606992245, + 0.008598711341619492, + 0.007192692719399929, + 0.006232877727597952, + 0.012866867706179619, + 0.010173317976295948, + 0.0064753880724310875, + 0.011247438378632069, + 0.006452113855630159, + 0.0028341382276266813, + 0.004210189916193485, + 0.002673597540706396, + 0.005637751892209053, + 0.00386052462272346, + 0.0037980112247169018, + 0.0034678310621529818, + 0.002786847297102213 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 5, + 1 + ], + "selected_reward": -0.06101521477103233 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a drone shot following the subject from the side and rear?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a heavily armored ancient warrior riding a jet-black muscular steed?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the warrior wear a phoenix-winged helmet with red tassels and intricately carved bronze armor with mountain patterns?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene set on the Great Wall winding through mountains at dusk with slanting golden sunlight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the steed gallop up the wall, kicking up a thick cloud of yellow dust and fine gravel behind it?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0890301913022995 + ], + "opposite_logprob_tokens": [ + -2.4640302658081055 + ], + "logprob_reward": -0.0890301913022995, + "yes_no_margin_reward": 2.375, + "token_credit_reward": -0.0890301913022995, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.20588085055351257, + 0.15379656851291656, + 0.06298761069774628, + 0.03664987161755562, + 0.10388386249542236, + 0.1752513200044632, + 0.11831571906805038, + 0.051579173654317856, + 0.03611553832888603, + 0.05553943291306496 + ], + "credit_map": [ + 0.01752656325697899, + 0.024025356397032738, + 0.02541261725127697, + 0.032731425017118454, + 0.04217052087187767, + 0.0171212088316679, + 0.022159794345498085, + 0.024733366444706917, + 0.015148918144404888, + 0.012478761374950409, + 0.012186897918581963, + 0.016542602330446243, + 0.026817725971341133, + 0.016493238508701324, + 0.02059287205338478, + 0.03353554382920265, + 0.008100726641714573, + 0.00795653648674488, + 0.005677773617208004, + 0.009389286860823631, + 0.005366016179323196, + 0.006721265148371458, + 0.010513304732739925, + 0.009262696839869022, + 0.0063681346364319324, + 0.004266387317329645, + 0.0025822955649346113, + 0.006302012130618095, + 0.004294831305742264, + 0.0040583680383861065, + 0.00387735478579998, + 0.004900489933788776, + 0.011208601295948029, + 0.010438242927193642, + 0.012677803635597229, + 0.028306040912866592, + 0.014349834993481636, + 0.008354605175554752, + 0.01140034943819046, + 0.007148376200348139, + 0.018216202035546303, + 0.017382128164172173, + 0.016923926770687103, + 0.022730212658643723, + 0.03328493610024452, + 0.021507572382688522, + 0.02440907247364521, + 0.020797278732061386, + 0.012904349714517593, + 0.019792012870311737, + 0.01475413329899311, + 0.01312609389424324, + 0.01591094769537449, + 0.016079774126410484, + 0.016477761790156364, + 0.009270638227462769, + 0.004148657433688641, + 0.00409030681475997, + 0.012027622200548649, + 0.010700381360948086, + 0.007178757805377245, + 0.004253226798027754, + 0.004678529221564531, + 0.004501692485064268, + 0.00503952894359827, + 0.005661788396537304, + 0.0031904529314488173, + 0.005762690212577581, + 0.004609415773302317, + 0.004193950444459915, + 0.005049807485193014, + 0.002607901580631733, + 0.010733471252024174, + 0.00687311589717865, + 0.005758900195360184, + 0.011341676115989685, + 0.004819220397621393, + 0.00549595570191741, + 0.0056239645928144455, + 0.004893127828836441 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 5, + 1 + ], + "selected_reward": -0.0890301913022995 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a drone shot following the subject from the side and rear?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a heavily armored ancient warrior riding a jet-black muscular steed?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the warrior wear a phoenix-winged helmet with red tassels and intricately carved bronze armor with mountain patterns?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene set on the Great Wall winding through mountains at dusk with slanting golden sunlight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the steed gallop up the wall, kicking up a thick cloud of yellow dust and fine gravel behind it?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.016075054183602333 + ], + "opposite_logprob_tokens": [ + -4.141075134277344 + ], + "logprob_reward": -0.016075054183602333, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.016075054183602333, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13203546404838562, + 0.17766882479190826, + 0.08088144659996033, + 0.08852311223745346, + 0.07065722346305847, + 0.15123608708381653, + 0.060633108019828796, + 0.11471947282552719, + 0.09458505362272263, + 0.029060224071145058 + ], + "credit_map": [ + 0.014385406859219074, + 0.01028340682387352, + 0.018285343423485756, + 0.024772396311163902, + 0.014447043649852276, + 0.014351611025631428, + 0.017116980627179146, + 0.018393268808722496, + 0.015938475728034973, + 0.016305197030305862, + 0.019360436126589775, + 0.020931370556354523, + 0.02640787698328495, + 0.019700828939676285, + 0.02525666356086731, + 0.033767979592084885, + 0.016210323199629784, + 0.014837030321359634, + 0.007417932152748108, + 0.01234560739248991, + 0.007141138892620802, + 0.007916927337646484, + 0.007866444066166878, + 0.007146048825234175, + 0.02522360160946846, + 0.013257695361971855, + 0.006041905377060175, + 0.012770244851708412, + 0.007794918958097696, + 0.007356579415500164, + 0.007905598729848862, + 0.008172559551894665, + 0.008655193261802197, + 0.009919319301843643, + 0.008722185157239437, + 0.019618947058916092, + 0.004912924952805042, + 0.006233432330191135, + 0.006179383490234613, + 0.006415839307010174, + 0.014455068856477737, + 0.01728324592113495, + 0.014636821113526821, + 0.023642713204026222, + 0.023621173575520515, + 0.021858403459191322, + 0.014815835282206535, + 0.0209228303283453, + 0.005551156122237444, + 0.009011561051011086, + 0.007095824461430311, + 0.00871984101831913, + 0.006471207365393639, + 0.008806303143501282, + 0.009470234625041485, + 0.005506978835910559, + 0.008007270283997059, + 0.007215031888335943, + 0.03186768665909767, + 0.03424146771430969, + 0.010783894918859005, + 0.007861816324293613, + 0.007911613211035728, + 0.006830696016550064, + 0.007928993552923203, + 0.008233695290982723, + 0.008197112008929253, + 0.016325024887919426, + 0.014962023124098778, + 0.015984995290637016, + 0.011934074573218822, + 0.011019137687981129, + 0.00346023659221828, + 0.0049737910740077496, + 0.0029731637332588434, + 0.005875853821635246, + 0.0029828252736479044, + 0.00390408537350595, + 0.002488009864464402, + 0.0024022578727453947 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 5, + 0 + ], + "selected_reward": -0.016075054183602333 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a drone shot following the subject from the side and rear?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a heavily armored ancient warrior riding a jet-black muscular steed?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the warrior wear a phoenix-winged helmet with red tassels and intricately carved bronze armor with mountain patterns?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene set on the Great Wall winding through mountains at dusk with slanting golden sunlight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the steed gallop up the wall, kicking up a thick cloud of yellow dust and fine gravel behind it?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01606426201760769 + ], + "opposite_logprob_tokens": [ + -4.141064167022705 + ], + "logprob_reward": -0.01606426201760769, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.01606426201760769, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1359984576702118, + 0.13529665768146515, + 0.07554838806390762, + 0.1508186161518097, + 0.07320473343133926, + 0.12496448308229446, + 0.1084422841668129, + 0.07567451149225235, + 0.06620106101036072, + 0.05385078117251396 + ], + "credit_map": [ + 0.01406112965196371, + 0.01118707936257124, + 0.015443537384271622, + 0.03660505264997482, + 0.012213971465826035, + 0.014286364428699017, + 0.0175422802567482, + 0.014659031294286251, + 0.018294109031558037, + 0.020495524629950523, + 0.013414730317890644, + 0.014353238977491856, + 0.015626385807991028, + 0.02123713679611683, + 0.015115990303456783, + 0.01675954833626747, + 0.009693378582596779, + 0.00981335062533617, + 0.010453331284224987, + 0.01884784735739231, + 0.005883184261620045, + 0.007343228906393051, + 0.006817167159169912, + 0.00669690128415823, + 0.027066711336374283, + 0.02673843502998352, + 0.009495082311332226, + 0.02919882908463478, + 0.01190425269305706, + 0.015211961232125759, + 0.014949393458664417, + 0.01625395007431507, + 0.009430953301489353, + 0.012067019939422607, + 0.008084733970463276, + 0.015760665759444237, + 0.00611046701669693, + 0.00768869137391448, + 0.007538276258856058, + 0.006523930933326483, + 0.01638537459075451, + 0.024700317531824112, + 0.013631010428071022, + 0.024115102365612984, + 0.013870484195649624, + 0.01225700881332159, + 0.00943706650286913, + 0.010568119585514069, + 0.01278824545443058, + 0.0163574181497097, + 0.012780414894223213, + 0.023173891007900238, + 0.01050830539315939, + 0.010309555567800999, + 0.012386441230773926, + 0.010138013400137424, + 0.009451969526708126, + 0.009358495473861694, + 0.022472074255347252, + 0.009992910549044609, + 0.005098441615700722, + 0.007780035957694054, + 0.005761459935456514, + 0.005759133491665125, + 0.007722701411694288, + 0.012511119246482849, + 0.005974547937512398, + 0.015014947392046452, + 0.008821865543723106, + 0.006494470406323671, + 0.005409070290625095, + 0.004252336919307709, + 0.007225405424833298, + 0.009893491864204407, + 0.005886161234229803, + 0.015607183799147606, + 0.0037878050934523344, + 0.005511910188943148, + 0.003442946122959256, + 0.0024958790745586157 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 0, + 1 + ], + "selected_reward": -0.01606426201760769 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a drone shot following the subject from the side and rear?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a heavily armored ancient warrior riding a jet-black muscular steed?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the warrior wear a phoenix-winged helmet with red tassels and intricately carved bronze armor with mountain patterns?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene set on the Great Wall winding through mountains at dusk with slanting golden sunlight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the steed gallop up the wall, kicking up a thick cloud of yellow dust and fine gravel behind it?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005944074597209692 + ], + "opposite_logprob_tokens": [ + -5.13094425201416 + ], + "logprob_reward": -0.005944074597209692, + "yes_no_margin_reward": 5.125, + "token_credit_reward": -0.005944074597209692, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0847179964184761, + 0.09447839111089706, + 0.04541396349668503, + 0.0804496705532074, + 0.14857056736946106, + 0.14028556644916534, + 0.08517756313085556, + 0.04302380234003067, + 0.17922483384609222, + 0.09865763783454895 + ], + "credit_map": [ + 0.007875868119299412, + 0.008202197030186653, + 0.012353088706731796, + 0.009890931658446789, + 0.01594087854027748, + 0.009826242923736572, + 0.011083508841693401, + 0.009545285254716873, + 0.015690486878156662, + 0.015044588595628738, + 0.010491935536265373, + 0.01127941720187664, + 0.008669541217386723, + 0.01224249042570591, + 0.009799339808523655, + 0.011260596103966236, + 0.004111945629119873, + 0.0042957826517522335, + 0.0038589127361774445, + 0.0053263152949512005, + 0.005618419498205185, + 0.005032564513385296, + 0.009159930050373077, + 0.008010097779333591, + 0.02619761787354946, + 0.014788221567869186, + 0.005513519048690796, + 0.009974557906389236, + 0.006096730940043926, + 0.004542849957942963, + 0.005748603027313948, + 0.007587576750665903, + 0.013703178614377975, + 0.017901211977005005, + 0.02102830447256565, + 0.048818327486515045, + 0.014906629920005798, + 0.010008380748331547, + 0.012789391912519932, + 0.00941515527665615, + 0.012663775123655796, + 0.023378532379865646, + 0.019878992810845375, + 0.04025715962052345, + 0.011534448713064194, + 0.013025417923927307, + 0.009803740307688713, + 0.009743506088852882, + 0.008845947682857513, + 0.014304372482001781, + 0.010047726333141327, + 0.013647579587996006, + 0.008037484250962734, + 0.009075958281755447, + 0.012139751575887203, + 0.009078748524188995, + 0.003654471365734935, + 0.003993121441453695, + 0.008451320230960846, + 0.00776915205642581, + 0.00633035646751523, + 0.005436313804239035, + 0.0033581864554435015, + 0.00403088191524148, + 0.019299142062664032, + 0.025530649349093437, + 0.01577369123697281, + 0.05163069814443588, + 0.01761982962489128, + 0.015573116950690746, + 0.023285310715436935, + 0.01051241159439087, + 0.01774512603878975, + 0.016047174111008644, + 0.011081038974225521, + 0.027770979329943657, + 0.00632221344858408, + 0.007889407686889172, + 0.0068197413347661495, + 0.00498195318505168 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 4, + 5 + ], + "selected_reward": -0.005944074597209692 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "drone shot, captured from a wide angle and from the side and rear, follows a heavily armored ancient warrior as he rides a jet-black, muscular steed, galloping along the Great Wall, which winds its way through the mountains. The warrior wears a phoenix-winged helmet with red tassels that dance wildly in the wind. His intricately carved bronze armor, adorned with mountain patterns, gleams with a cool metallic sheen as he moves, and the beast-headed shoulder guards adorning his shoulders project a majestic presence. His bronzed face is stern and resolute. He leans forward slightly, gripping the dark leather and brass-trimmed reins. His crimson cape, swirling in the wind, ripples like a flaming banner. The steed's hooves thunder, each powerful thump pounding against the weathered, massive stone pavement, stirring up a thick cloud of yellow dust and fine gravel, leaving a billowing trail behind him. It's dusk, and slanting golden sunlight paints the endless, towering mountains and the gray city walls that undulate like a giant dragon with a magnificent warm hue. Backlighting frames the swirling dust and mist with a shimmering golden edge, lengthening the lonely shadows of the men and horses cast on the city walls. The entire scene, framed by the grand composition, is imbued with a sense of epic grandeur and the urgency and loneliness of a mission, creating an atmosphere of solemnity and power.", + "seed": 20274, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000006/group_4/index_8270_seed_20274_dp_9_rank_18_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 4, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.05515195094048977, + "vlm_logprob": -0.05515195094048977, + "vlm_yes_no_margin": 3.199999976158142, + "vlm_token_credit": -0.05515195094048977, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a drone shot following the subject from the side and rear?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a heavily armored ancient warrior riding a jet-black muscular steed?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the warrior wear a phoenix-winged helmet with red tassels and intricately carved bronze armor with mountain patterns?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene set on the Great Wall winding through mountains at dusk with slanting golden sunlight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the steed gallop up the wall, kicking up a thick cloud of yellow dust and fine gravel behind it?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.08089243620634079 + ], + "opposite_logprob_tokens": [ + -2.705892324447632 + ], + "logprob_reward": -0.08089243620634079, + "yes_no_margin_reward": 2.625, + "token_credit_reward": -0.08089243620634079, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.37395167350769043, + 0.12757211923599243, + 0.07678026705980301, + 0.09023645520210266, + 0.06549884378910065, + 0.07063388079404831, + 0.0466720387339592, + 0.028091808781027794, + 0.08109322935342789, + 0.039469633251428604 + ], + "credit_map": [ + 0.07735389471054077, + 0.03438716381788254, + 0.034275636076927185, + 0.039105165749788284, + 0.04932556301355362, + 0.04304540902376175, + 0.05182895064353943, + 0.044629909098148346, + 0.024423256516456604, + 0.022527307271957397, + 0.009421766735613346, + 0.011548454873263836, + 0.01918959617614746, + 0.011895637959241867, + 0.012630586512386799, + 0.015935515984892845, + 0.012357094325125217, + 0.009320391342043877, + 0.007348301354795694, + 0.007590396795421839, + 0.01389613188803196, + 0.007691747508943081, + 0.010202405042946339, + 0.008373801596462727, + 0.009916134178638458, + 0.0193211380392313, + 0.01011990662664175, + 0.011880926787853241, + 0.00862138718366623, + 0.009011590853333473, + 0.011064802296459675, + 0.010300571098923683, + 0.007746478542685509, + 0.007949868217110634, + 0.008506362326443195, + 0.006984722800552845, + 0.008864532224833965, + 0.01076532993465662, + 0.006964641623198986, + 0.007716905791312456, + 0.009248272515833378, + 0.01148471049964428, + 0.00757213169708848, + 0.00894855335354805, + 0.008933805860579014, + 0.009828520938754082, + 0.005603012163192034, + 0.009014877490699291, + 0.004593282006680965, + 0.006300697103142738, + 0.006178597453981638, + 0.004075708799064159, + 0.00915585644543171, + 0.005812270101159811, + 0.003893336746841669, + 0.0066622900776565075, + 0.00438734469935298, + 0.002870177151635289, + 0.0021618909668177366, + 0.002056570490822196, + 0.0064586070366203785, + 0.004775669425725937, + 0.00175772444345057, + 0.0036238229367882013, + 0.009519173763692379, + 0.00890375953167677, + 0.016950108110904694, + 0.006376559846103191, + 0.008030939847230911, + 0.007892316207289696, + 0.012191051617264748, + 0.011229320429265499, + 0.003929181024432182, + 0.008438020944595337, + 0.004826032090932131, + 0.0034312435891479254, + 0.006140495650470257, + 0.00345794134773314, + 0.00431571202352643, + 0.004931009374558926 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 3 + ], + "selected_reward": -0.08089243620634079 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a drone shot following the subject from the side and rear?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a heavily armored ancient warrior riding a jet-black muscular steed?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the warrior wear a phoenix-winged helmet with red tassels and intricately carved bronze armor with mountain patterns?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene set on the Great Wall winding through mountains at dusk with slanting golden sunlight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the steed gallop up the wall, kicking up a thick cloud of yellow dust and fine gravel behind it?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.12703903019428253 + ], + "opposite_logprob_tokens": [ + -2.1270389556884766 + ], + "logprob_reward": -0.12703903019428253, + "yes_no_margin_reward": 1.9999998807907104, + "token_credit_reward": -0.12703903019428253, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.4412236511707306, + 0.044369351118803024, + 0.025547491386532784, + 0.04273417219519615, + 0.031129304319620132, + 0.04600384086370468, + 0.03495434671640396, + 0.02777285873889923, + 0.04858582839369774, + 0.2576790750026703 + ], + "credit_map": [ + 0.11743558943271637, + 0.04481034353375435, + 0.03858790546655655, + 0.029737165197730064, + 0.06806132942438126, + 0.04587694630026817, + 0.03608028218150139, + 0.06063413992524147, + 0.00778423435986042, + 0.007222190033644438, + 0.003382426919415593, + 0.004374713636934757, + 0.005582430399954319, + 0.004143490456044674, + 0.006649501156061888, + 0.0052303653210401535, + 0.0032025063410401344, + 0.005704233888536692, + 0.0019746527541428804, + 0.002324732020497322, + 0.003819819772616029, + 0.0026693108957260847, + 0.003024653298780322, + 0.0028275838121771812, + 0.004087396897375584, + 0.005547199863940477, + 0.00809354055672884, + 0.006457956973463297, + 0.00369263649918139, + 0.004161414224654436, + 0.0054313549771904945, + 0.005262678489089012, + 0.0038413554430007935, + 0.004821389447897673, + 0.0032041308004409075, + 0.004385450854897499, + 0.0030619523022323847, + 0.004098261706531048, + 0.003234089817851782, + 0.004482676740735769, + 0.004918585065752268, + 0.01207794900983572, + 0.00371048622764647, + 0.005017895717173815, + 0.0055922698229551315, + 0.004828646779060364, + 0.004273562226444483, + 0.005584451369941235, + 0.0035205415915697813, + 0.007497179321944714, + 0.00296725332736969, + 0.003170793177559972, + 0.004758263006806374, + 0.005331246182322502, + 0.0030160536989569664, + 0.004693019203841686, + 0.0030601818580180407, + 0.004878268577158451, + 0.0027585183270275593, + 0.0051910304464399815, + 0.0025823272299021482, + 0.00347033329308033, + 0.0026245899498462677, + 0.003207612317055464, + 0.005405441392213106, + 0.010313833132386208, + 0.006321256048977375, + 0.004796740598976612, + 0.004862479865550995, + 0.004866723902523518, + 0.005634899716824293, + 0.006384461186826229, + 0.02181706763803959, + 0.04366927593946457, + 0.04383087530732155, + 0.026085905730724335, + 0.04003158584237099, + 0.018218770623207092, + 0.033403363078832626, + 0.030622253194451332 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 8 + ], + "selected_reward": -0.12703903019428253 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a drone shot following the subject from the side and rear?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a heavily armored ancient warrior riding a jet-black muscular steed?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the warrior wear a phoenix-winged helmet with red tassels and intricately carved bronze armor with mountain patterns?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene set on the Great Wall winding through mountains at dusk with slanting golden sunlight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the steed gallop up the wall, kicking up a thick cloud of yellow dust and fine gravel behind it?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.018201233819127083 + ], + "opposite_logprob_tokens": [ + -4.0182013511657715 + ], + "logprob_reward": -0.018201233819127083, + "yes_no_margin_reward": 4.0, + "token_credit_reward": -0.018201233819127083, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.47907888889312744, + 0.10085693746805191, + 0.057672902941703796, + 0.054176826030015945, + 0.05781925842165947, + 0.06955709308385849, + 0.04969397187232971, + 0.04541509225964546, + 0.02225821651518345, + 0.06347081065177917 + ], + "credit_map": [ + 0.12281213700771332, + 0.03285244479775429, + 0.0372890867292881, + 0.04510566219687462, + 0.07506375759840012, + 0.05581854283809662, + 0.04398311674594879, + 0.0661541223526001, + 0.0179445780813694, + 0.015251247212290764, + 0.007581403013318777, + 0.010626507923007011, + 0.012864119373261929, + 0.009311165660619736, + 0.014639324508607388, + 0.012638584710657597, + 0.005700922105461359, + 0.011748509481549263, + 0.005552936811000109, + 0.009805870242416859, + 0.006214554887264967, + 0.007506742607802153, + 0.004886523354798555, + 0.006256843451410532, + 0.005656675435602665, + 0.009556571021676064, + 0.0064749945886433125, + 0.007201267872005701, + 0.008154194802045822, + 0.005657576024532318, + 0.005531528498977423, + 0.0059440224431455135, + 0.0061895283870399, + 0.010652526281774044, + 0.006378053221851587, + 0.0070791603066027164, + 0.006381609942764044, + 0.007499129977077246, + 0.006085569970309734, + 0.0075536807999014854, + 0.007347098086029291, + 0.022577425464987755, + 0.005073683336377144, + 0.00871279276907444, + 0.006494398228824139, + 0.005942278541624546, + 0.004840083885937929, + 0.008569334633648396, + 0.006039637140929699, + 0.007984112948179245, + 0.005478346720337868, + 0.0034571595024317503, + 0.008403063751757145, + 0.008188711479306221, + 0.0031551423016935587, + 0.0069877989590168, + 0.005712814629077911, + 0.007662807125598192, + 0.0034390937071293592, + 0.0052835033275187016, + 0.007071202155202627, + 0.007011532783508301, + 0.0036698493640869856, + 0.005564287770539522, + 0.0034478832967579365, + 0.00399145670235157, + 0.0017988282488659024, + 0.0020567025057971478, + 0.0027423473075032234, + 0.003655277891084552, + 0.002059302059933543, + 0.002506418852135539, + 0.0065688774921, + 0.01040769461542368, + 0.009857687167823315, + 0.005931805819272995, + 0.008161487057805061, + 0.0088554248213768, + 0.0061425394378602505, + 0.007545295171439648 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 5 + ], + "selected_reward": -0.018201233819127083 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a drone shot following the subject from the side and rear?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a heavily armored ancient warrior riding a jet-black muscular steed?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the warrior wear a phoenix-winged helmet with red tassels and intricately carved bronze armor with mountain patterns?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene set on the Great Wall winding through mountains at dusk with slanting golden sunlight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the steed gallop up the wall, kicking up a thick cloud of yellow dust and fine gravel behind it?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02634398825466633 + ], + "opposite_logprob_tokens": [ + -3.651344060897827 + ], + "logprob_reward": -0.02634398825466633, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.02634398825466633, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.4080050587654114, + 0.08112232387065887, + 0.07273562252521515, + 0.06864910572767258, + 0.056819137185811996, + 0.0798971876502037, + 0.053906626999378204, + 0.027478545904159546, + 0.029169131070375443, + 0.12221719324588776 + ], + "credit_map": [ + 0.06372044235467911, + 0.039930857717990875, + 0.03697728365659714, + 0.027507007122039795, + 0.04930790513753891, + 0.06818229705095291, + 0.0668274387717247, + 0.05555184930562973, + 0.01101327408105135, + 0.010501643642783165, + 0.007023084443062544, + 0.008626950904726982, + 0.011531196534633636, + 0.01305258460342884, + 0.007278622593730688, + 0.012094976380467415, + 0.0076888552866876125, + 0.008405596017837524, + 0.005312962457537651, + 0.004934230819344521, + 0.011652388609945774, + 0.017124544829130173, + 0.007524090353399515, + 0.010092958807945251, + 0.008170118555426598, + 0.006411854177713394, + 0.006852172780781984, + 0.004956359975039959, + 0.010659272782504559, + 0.01482979953289032, + 0.008091102354228497, + 0.008678426966071129, + 0.008699805475771427, + 0.007395859342068434, + 0.005295636598020792, + 0.005147221498191357, + 0.007147898431867361, + 0.010551613755524158, + 0.00375856994651258, + 0.008822532370686531, + 0.017263008281588554, + 0.008652408607304096, + 0.008648937568068504, + 0.004903305321931839, + 0.009604043327271938, + 0.015250598080456257, + 0.00458433898165822, + 0.010990556329488754, + 0.00871824286878109, + 0.003434927435591817, + 0.005875778384506702, + 0.0029870341531932354, + 0.0068276142701506615, + 0.013520359061658382, + 0.0035360942129045725, + 0.009006580337882042, + 0.0030168204102665186, + 0.004216877277940512, + 0.00290830060839653, + 0.0026908053550869226, + 0.004770789761096239, + 0.00450132368132472, + 0.0020943721756339073, + 0.003279258031398058, + 0.003980927634984255, + 0.0034949316177517176, + 0.003594528418034315, + 0.0028786570765078068, + 0.0045937784016132355, + 0.004486100748181343, + 0.0024764216504991055, + 0.003663789015263319, + 0.017921701073646545, + 0.017138630151748657, + 0.011554386466741562, + 0.009507807902991772, + 0.02295243926346302, + 0.020834362134337425, + 0.009024666622281075, + 0.013283212669193745 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 1 + ], + "selected_reward": -0.02634398825466633 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a drone shot following the subject from the side and rear?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a heavily armored ancient warrior riding a jet-black muscular steed?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the warrior wear a phoenix-winged helmet with red tassels and intricately carved bronze armor with mountain patterns?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene set on the Great Wall winding through mountains at dusk with slanting golden sunlight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the steed gallop up the wall, kicking up a thick cloud of yellow dust and fine gravel behind it?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.023283066228032112 + ], + "opposite_logprob_tokens": [ + -3.773283004760742 + ], + "logprob_reward": -0.023283066228032112, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.023283066228032112, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2188924103975296, + 0.14582745730876923, + 0.06645727157592773, + 0.1298607438802719, + 0.08335147798061371, + 0.18761640787124634, + 0.037281401455402374, + 0.045404672622680664, + 0.050096042454242706, + 0.035212136805057526 + ], + "credit_map": [ + 0.02077595517039299, + 0.02653251588344574, + 0.03066037781536579, + 0.02888641133904457, + 0.02096530608832836, + 0.027164198458194733, + 0.03919799253344536, + 0.024709638208150864, + 0.02586902491748333, + 0.023146726191043854, + 0.010674760676920414, + 0.014791347086429596, + 0.023723313584923744, + 0.013909678906202316, + 0.017189037054777145, + 0.016523556783795357, + 0.00715295085683465, + 0.008953157812356949, + 0.006555074825882912, + 0.0076466514728963375, + 0.012010862119495869, + 0.007642846554517746, + 0.00949106365442276, + 0.007004650309681892, + 0.01680266112089157, + 0.030320441350340843, + 0.010125294327735901, + 0.022862784564495087, + 0.011950097978115082, + 0.012434187345206738, + 0.010429524816572666, + 0.014935742132365704, + 0.008565276861190796, + 0.010462547652423382, + 0.011395879089832306, + 0.013021135702729225, + 0.008978067897260189, + 0.012471812777221203, + 0.007129355799406767, + 0.011327393352985382, + 0.02840057760477066, + 0.03265204280614853, + 0.029917733743786812, + 0.01998988725244999, + 0.01651190221309662, + 0.01821395568549633, + 0.017222551628947258, + 0.024707749485969543, + 0.0026648316998034716, + 0.0064314366318285465, + 0.003917090129107237, + 0.0026775477454066277, + 0.008486990816891193, + 0.004729428794234991, + 0.0029914311598986387, + 0.005382641218602657, + 0.007043973542749882, + 0.005080457776784897, + 0.004863142501562834, + 0.003997602500021458, + 0.008753402158617973, + 0.007072945591062307, + 0.0035951435565948486, + 0.0049980031326413155, + 0.005831503309309483, + 0.004827702417969704, + 0.009309775196015835, + 0.004389849491417408, + 0.005869388114660978, + 0.005085288546979427, + 0.00787318591028452, + 0.006909342482686043, + 0.0037841186858713627, + 0.006749278865754604, + 0.0048276446759700775, + 0.004341011866927147, + 0.005589762236922979, + 0.002793900202959776, + 0.0028911507688462734, + 0.004235267639160156 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 5, + 1 + ], + "selected_reward": -0.023283066228032112 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000006/group_5/index_tensor([6003, 6003])_group_5_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000006/group_5/index_tensor([6003, 6003])_group_5_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..9ffc1e30b2304e1c7e2c3393c74739f789ce1f6c --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000006/group_5/index_tensor([6003, 6003])_group_5_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "static yet unwavering extreme macro shot delves into the surreal landscape of the human eye in lifelike 4K detail. The camera focuses on the scrotum, not a pure white, but a pearly, slightly yellowish-yellow surface with a damp sheen, woven through a complex network of crimson veins. These intricate, blood-filled channels, rendered with breathtaking clarity, branch like a delicate, leafless tree, some appearing as dense, protruding vessels, while others fragment into hair-like capillaries. The central action is the subtle, rhythmic pulsation within these veins; they expand and contract almost imperceptibly in a steady biological rhythm, suggesting the silent beating of a heart. Bright, unfocused, clinical lighting illuminates every tiny texture without casting harsh shadows and creating small, sharp, specular highlights along the curve of the cornea. The edge of a dark brown iris, with its intricate fibrous pattern, is visible on one side of the frame, creating a scrotal landscape reminiscent of an alien. The overall atmosphere is one of intense, almost uncomfortable intimacy and biological fascination, a visceral exploration of the inner mechanisms of life.", + "seed": 32399, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000006/group_5/index_6003_seed_32399_dp_10_rank_20_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 5, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.07726691477000713, + "vlm_logprob": -0.07726691477000713, + "vlm_yes_no_margin": 3.6, + "vlm_token_credit": -0.07726691477000713, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme macro view focusing on the internal structures of the human eye?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the central surface pearly, slightly yellowish, and damp with a sheen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are crimson veins visible branching like a leafless tree with some dense protruding vessels and hair-like capillaries?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the veins exhibit a subtle, rhythmic pulsation expanding and contracting in a steady biological rhythm?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the edge of a dark brown iris with an intricate fibrous pattern visible on one side of the frame?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03230399265885353 + ], + "opposite_logprob_tokens": [ + -3.657304048538208 + ], + "logprob_reward": -0.03230399265885353, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.03230399265885353, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.32400909066200256, + 0.07870358973741531, + 0.03974143788218498, + 0.14106817543506622, + 0.07114937901496887, + 0.06732600927352905, + 0.06655248999595642, + 0.10345514863729477, + 0.05192926898598671, + 0.0560653991997242 + ], + "credit_map": [ + 0.04440903663635254, + 0.040799397975206375, + 0.03833724558353424, + 0.04111894592642784, + 0.04422855004668236, + 0.03551279008388519, + 0.029799731448292732, + 0.04980337619781494, + 0.011952901259064674, + 0.013326648622751236, + 0.009065147489309311, + 0.011810017749667168, + 0.010283851996064186, + 0.007501951418817043, + 0.005794468801468611, + 0.00896859634667635, + 0.0041510541923344135, + 0.004880156833678484, + 0.007323880214244127, + 0.00470952270552516, + 0.004802226554602385, + 0.004836264997720718, + 0.003448773641139269, + 0.005589558742940426, + 0.027034509927034378, + 0.018801039084792137, + 0.02108578197658062, + 0.014220571145415306, + 0.014203834347426891, + 0.013719234615564346, + 0.011303081177175045, + 0.0207001231610775, + 0.005740379448980093, + 0.006906683556735516, + 0.009279890917241573, + 0.009801464155316353, + 0.010087862610816956, + 0.009498391300439835, + 0.005970732308924198, + 0.013863971456885338, + 0.0053997342474758625, + 0.006210288032889366, + 0.008393771015107632, + 0.009156209416687489, + 0.010996682569384575, + 0.007068110164254904, + 0.00649077445268631, + 0.01361043844372034, + 0.006119080353528261, + 0.0071827867068350315, + 0.011905346997082233, + 0.009629192762076855, + 0.008522274903953075, + 0.007196675520390272, + 0.005967041943222284, + 0.010030089877545834, + 0.008705316111445427, + 0.010785412043333054, + 0.0180011335760355, + 0.013133575208485126, + 0.014537439681589603, + 0.012919500470161438, + 0.008701740764081478, + 0.016671037301421165, + 0.004295930732041597, + 0.00482811639085412, + 0.007007484789937735, + 0.0054346611723303795, + 0.006732739973813295, + 0.00567986723035574, + 0.005453835241496563, + 0.012496634386479855, + 0.005002124700695276, + 0.006068442016839981, + 0.009637339971959591, + 0.0067401970736682415, + 0.006086659152060747, + 0.005675392225384712, + 0.005375797860324383, + 0.011479444801807404 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 7 + ], + "selected_reward": -0.03230399265885353 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme macro view focusing on the internal structures of the human eye?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the central surface pearly, slightly yellowish, and damp with a sheen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are crimson veins visible branching like a leafless tree with some dense protruding vessels and hair-like capillaries?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the veins exhibit a subtle, rhythmic pulsation expanding and contracting in a steady biological rhythm?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the edge of a dark brown iris with an intricate fibrous pattern visible on one side of the frame?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004667577799409628 + ], + "opposite_logprob_tokens": [ + -5.37966775894165 + ], + "logprob_reward": -0.004667577799409628, + "yes_no_margin_reward": 5.375, + "token_credit_reward": -0.004667577799409628, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.25953981280326843, + 0.23885942995548248, + 0.054972149431705475, + 0.041586775332689285, + 0.07514198869466782, + 0.045042484998703, + 0.031050190329551697, + 0.07641472667455673, + 0.1061655804514885, + 0.0712268128991127 + ], + "credit_map": [ + 0.03728829324245453, + 0.029615413397550583, + 0.03195076063275337, + 0.035229433327913284, + 0.0307497289031744, + 0.02583830989897251, + 0.023350469768047333, + 0.04551742970943451, + 0.05644785240292549, + 0.036108098924160004, + 0.025698840618133545, + 0.027432307600975037, + 0.017767421901226044, + 0.026699943467974663, + 0.019805176183581352, + 0.0288997869938612, + 0.009286446496844292, + 0.008519887924194336, + 0.0053184605203568935, + 0.008309436962008476, + 0.006093543954193592, + 0.004932025913149118, + 0.005807872395962477, + 0.006704474799335003, + 0.009623315185308456, + 0.005492132622748613, + 0.003722144989296794, + 0.004823047202080488, + 0.004765976220369339, + 0.0034721726551651955, + 0.003608154831454158, + 0.006079834420233965, + 0.008350566029548645, + 0.00795233715325594, + 0.011707725934684277, + 0.009794257581233978, + 0.009605045430362225, + 0.006936769932508469, + 0.008725818246603012, + 0.012069467455148697, + 0.0042236801236867905, + 0.0045873988419771194, + 0.005878421477973461, + 0.005892712157219648, + 0.006948552560061216, + 0.004042071755975485, + 0.004858934320509434, + 0.008610716089606285, + 0.0034551750868558884, + 0.0034615451004356146, + 0.003920815419405699, + 0.00391062255948782, + 0.006216875277459621, + 0.0028736910317093134, + 0.0032759651076048613, + 0.003935500979423523, + 0.010781330056488514, + 0.007715162821114063, + 0.008009018376469612, + 0.00807904452085495, + 0.014633921906352043, + 0.007271699607372284, + 0.006107074674218893, + 0.013817468658089638, + 0.00938662514090538, + 0.008532563224434853, + 0.007595549803227186, + 0.013200675137341022, + 0.015368440188467503, + 0.012138386256992817, + 0.009301182813942432, + 0.030642153695225716, + 0.005812841467559338, + 0.00839786697179079, + 0.00824857410043478, + 0.010503035970032215, + 0.009142798371613026, + 0.0074757169932127, + 0.008703043684363365, + 0.012942928820848465 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 8 + ], + "selected_reward": -0.004667577799409628 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme macro view focusing on the internal structures of the human eye?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the central surface pearly, slightly yellowish, and damp with a sheen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are crimson veins visible branching like a leafless tree with some dense protruding vessels and hair-like capillaries?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the veins exhibit a subtle, rhythmic pulsation expanding and contracting in a steady biological rhythm?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the edge of a dark brown iris with an intricate fibrous pattern visible on one side of the frame?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005961731541901827 + ], + "opposite_logprob_tokens": [ + -5.130961894989014 + ], + "logprob_reward": -0.005961731541901827, + "yes_no_margin_reward": 5.125, + "token_credit_reward": -0.005961731541901827, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.26161375641822815, + 0.15647609531879425, + 0.08197817206382751, + 0.10706818848848343, + 0.0978991910815239, + 0.07315286248922348, + 0.06436604261398315, + 0.06845859438180923, + 0.03548562899231911, + 0.05350148677825928 + ], + "credit_map": [ + 0.04534366726875305, + 0.03920307382941246, + 0.030561894178390503, + 0.03610075265169144, + 0.026189621537923813, + 0.028343459591269493, + 0.02516494318842888, + 0.030706331133842468, + 0.027539735659956932, + 0.019167719408869743, + 0.019975794479250908, + 0.017850534990429878, + 0.018758416175842285, + 0.015447331592440605, + 0.013161016628146172, + 0.024575557559728622, + 0.013878676109015942, + 0.010757780633866787, + 0.010354865342378616, + 0.008458446711301804, + 0.010719742625951767, + 0.008548492565751076, + 0.007006648927927017, + 0.012253515422344208, + 0.015813132748007774, + 0.01230578776448965, + 0.01638651080429554, + 0.012398073449730873, + 0.010983044281601906, + 0.011070482432842255, + 0.009418810717761517, + 0.018692340701818466, + 0.007382998708635569, + 0.009443114511668682, + 0.013435875996947289, + 0.013000113889575005, + 0.015382232144474983, + 0.01351243443787098, + 0.008054500445723534, + 0.017687924206256866, + 0.007366276811808348, + 0.007432100363075733, + 0.007229427341371775, + 0.010075888596475124, + 0.011612272821366787, + 0.008588660508394241, + 0.007247682195156813, + 0.013600556179881096, + 0.006555302068591118, + 0.005427289754152298, + 0.008592329919338226, + 0.008635922335088253, + 0.011203411035239697, + 0.00627698702737689, + 0.00634610652923584, + 0.011328687891364098, + 0.010422949679195881, + 0.007342140655964613, + 0.0056887040846049786, + 0.007814126089215279, + 0.011119089089334011, + 0.0070078750140964985, + 0.006142343394458294, + 0.01292136125266552, + 0.0033797931391745806, + 0.003172570839524269, + 0.004670307505875826, + 0.0051728603430092335, + 0.005293256137520075, + 0.003934276290237904, + 0.004289679229259491, + 0.005572883877903223, + 0.005292861256748438, + 0.005372954066842794, + 0.007114597596228123, + 0.006676392164081335, + 0.0075828032568097115, + 0.005388887599110603, + 0.0036063536535948515, + 0.012466635555028915 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 3 + ], + "selected_reward": -0.005961731541901827 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme macro view focusing on the internal structures of the human eye?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the central surface pearly, slightly yellowish, and damp with a sheen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are crimson veins visible branching like a leafless tree with some dense protruding vessels and hair-like capillaries?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the veins exhibit a subtle, rhythmic pulsation expanding and contracting in a steady biological rhythm?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the edge of a dark brown iris with an intricate fibrous pattern visible on one side of the frame?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.28136542439460754 + ], + "opposite_logprob_tokens": [ + -1.4063653945922852 + ], + "logprob_reward": -0.28136542439460754, + "yes_no_margin_reward": 1.125, + "token_credit_reward": -0.28136542439460754, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3908342123031616, + 0.13069197535514832, + 0.16530993580818176, + 0.07906545698642731, + 0.04477058723568916, + 0.044925011694431305, + 0.033470649272203445, + 0.02929842285811901, + 0.02909816801548004, + 0.052535638213157654 + ], + "credit_map": [ + 0.08017469942569733, + 0.05431317910552025, + 0.0484854057431221, + 0.04394230619072914, + 0.03185095265507698, + 0.0438685417175293, + 0.03722676634788513, + 0.05097232758998871, + 0.016748134046792984, + 0.014938977546989918, + 0.01835676096379757, + 0.017630986869335175, + 0.015715444460511208, + 0.01804259419441223, + 0.009794408455491066, + 0.01946466974914074, + 0.026363980025053024, + 0.021969124674797058, + 0.020291291177272797, + 0.017085807397961617, + 0.023343713954091072, + 0.01669919304549694, + 0.014718727208673954, + 0.024838080629706383, + 0.007562541402876377, + 0.009369049221277237, + 0.01228867657482624, + 0.010259436443448067, + 0.010550182312726974, + 0.008480074815452099, + 0.007034820504486561, + 0.013520668260753155, + 0.0069604129530489445, + 0.00516879465430975, + 0.005790897645056248, + 0.005016059149056673, + 0.005636401940137148, + 0.004563131835311651, + 0.00371099472977221, + 0.007923893630504608, + 0.0038092671893537045, + 0.005081453826278448, + 0.004218322224915028, + 0.0064012156799435616, + 0.007469631731510162, + 0.0039283898659050465, + 0.004348482005298138, + 0.009668244980275631, + 0.00323881721124053, + 0.003129604971036315, + 0.005535612814128399, + 0.0039939540438354015, + 0.0047500827349722385, + 0.003536480711773038, + 0.0033414082136005163, + 0.005944686476141214, + 0.004363165702670813, + 0.0027991144452244043, + 0.0031448157969862223, + 0.00323173007927835, + 0.005660850089043379, + 0.003225293243303895, + 0.0023451033048331738, + 0.004528348799794912, + 0.003357349894940853, + 0.003812143811956048, + 0.00462888041511178, + 0.0043378048576414585, + 0.0030634275171905756, + 0.0034505072981119156, + 0.002978254808112979, + 0.0034697968512773514, + 0.004422396887093782, + 0.006063485983759165, + 0.006935386918485165, + 0.009389122016727924, + 0.00524017633870244, + 0.005098327528685331, + 0.004689379595220089, + 0.010697356425225735 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 1 + ], + "selected_reward": -0.28136542439460754 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme macro view focusing on the internal structures of the human eye?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the central surface pearly, slightly yellowish, and damp with a sheen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are crimson veins visible branching like a leafless tree with some dense protruding vessels and hair-like capillaries?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the veins exhibit a subtle, rhythmic pulsation expanding and contracting in a steady biological rhythm?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the edge of a dark brown iris with an intricate fibrous pattern visible on one side of the frame?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.06203584745526314 + ], + "opposite_logprob_tokens": [ + -2.8120357990264893 + ], + "logprob_reward": -0.06203584745526314, + "yes_no_margin_reward": 2.75, + "token_credit_reward": -0.06203584745526314, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.19708889722824097, + 0.21779686212539673, + 0.22704298794269562, + 0.1313537359237671, + 0.042066436260938644, + 0.04005807265639305, + 0.028755836188793182, + 0.0547005720436573, + 0.017731470987200737, + 0.0434052012860775 + ], + "credit_map": [ + 0.030888309702277184, + 0.020818503573536873, + 0.03147030249238014, + 0.01884625479578972, + 0.021207498386502266, + 0.02235112339258194, + 0.016007425263524055, + 0.0354994535446167, + 0.04803769662976265, + 0.03217380493879318, + 0.028600692749023438, + 0.024782473221421242, + 0.016776273027062416, + 0.02368956059217453, + 0.017833402380347252, + 0.025902925059199333, + 0.03076174296438694, + 0.024844754487276077, + 0.030161360278725624, + 0.02303590625524521, + 0.03021838515996933, + 0.023802321404218674, + 0.02013249881565571, + 0.04408600181341171, + 0.02366478741168976, + 0.01605084165930748, + 0.018283335492014885, + 0.014031153172254562, + 0.01422154251486063, + 0.012344333343207836, + 0.01107574999332428, + 0.02168198488652706, + 0.004117111675441265, + 0.0034611462615430355, + 0.007004782557487488, + 0.005422282963991165, + 0.005296645220369101, + 0.005489049479365349, + 0.0038799913600087166, + 0.0073954216204583645, + 0.0035167571622878313, + 0.0036751653533428907, + 0.005687362980097532, + 0.004107097629457712, + 0.007275872398167849, + 0.003962538670748472, + 0.004146830644458532, + 0.007686442695558071, + 0.002705390565097332, + 0.002299056388437748, + 0.005814980249851942, + 0.0032938034273684025, + 0.003963269293308258, + 0.00373842753469944, + 0.0021724875550717115, + 0.004768418613821268, + 0.005844872444868088, + 0.004917918238788843, + 0.007846604101359844, + 0.0048823729157447815, + 0.01036621443927288, + 0.00677691213786602, + 0.004146951250731945, + 0.00991871953010559, + 0.0016023661009967327, + 0.0018553721019998193, + 0.0023136092349886894, + 0.0025721597485244274, + 0.0020593502558767796, + 0.00214669993147254, + 0.001661327201873064, + 0.003520586295053363, + 0.003568946151062846, + 0.004622094798833132, + 0.005846317857503891, + 0.007976901717483997, + 0.0054642329923808575, + 0.004300320520997047, + 0.004288002382963896, + 0.007338379975408316 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 1, + 0 + ], + "selected_reward": -0.06203584745526314 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "static yet unwavering extreme macro shot delves into the surreal landscape of the human eye in lifelike 4K detail. The camera focuses on the scrotum, not a pure white, but a pearly, slightly yellowish-yellow surface with a damp sheen, woven through a complex network of crimson veins. These intricate, blood-filled channels, rendered with breathtaking clarity, branch like a delicate, leafless tree, some appearing as dense, protruding vessels, while others fragment into hair-like capillaries. The central action is the subtle, rhythmic pulsation within these veins; they expand and contract almost imperceptibly in a steady biological rhythm, suggesting the silent beating of a heart. Bright, unfocused, clinical lighting illuminates every tiny texture without casting harsh shadows and creating small, sharp, specular highlights along the curve of the cornea. The edge of a dark brown iris, with its intricate fibrous pattern, is visible on one side of the frame, creating a scrotal landscape reminiscent of an alien. The overall atmosphere is one of intense, almost uncomfortable intimacy and biological fascination, a visceral exploration of the inner mechanisms of life.", + "seed": 32400, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000006/group_5/index_6003_seed_32400_dp_10_rank_20_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 5, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.035552076320163904, + "vlm_logprob": -0.035552076320163904, + "vlm_yes_no_margin": 4.7, + "vlm_token_credit": -0.035552076320163904, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme macro view focusing on the internal structures of the human eye?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the central surface pearly, slightly yellowish, and damp with a sheen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are crimson veins visible branching like a leafless tree with some dense protruding vessels and hair-like capillaries?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the veins exhibit a subtle, rhythmic pulsation expanding and contracting in a steady biological rhythm?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the edge of a dark brown iris with an intricate fibrous pattern visible on one side of the frame?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005821772385388613 + ], + "opposite_logprob_tokens": [ + -6.755821704864502 + ], + "logprob_reward": -0.005821772385388613, + "yes_no_margin_reward": 6.75, + "token_credit_reward": -0.005821772385388613, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11191914230585098, + 0.051899347454309464, + 0.05011323466897011, + 0.047635458409786224, + 0.039971910417079926, + 0.04680171608924866, + 0.0759962648153305, + 0.07107304036617279, + 0.10506980121135712, + 0.39952003955841064 + ], + "credit_map": [ + 0.013540096580982208, + 0.012281686998903751, + 0.014947031624615192, + 0.01958823762834072, + 0.012777273543179035, + 0.012513335794210434, + 0.013439239002764225, + 0.012832244858145714, + 0.01027386263012886, + 0.004430848639458418, + 0.01009669154882431, + 0.005335561465471983, + 0.004441248252987862, + 0.006002297159284353, + 0.005675317719578743, + 0.0056435237638652325, + 0.00796596147119999, + 0.004039854742586613, + 0.00859878771007061, + 0.005112648010253906, + 0.006238346453756094, + 0.00691522192209959, + 0.004754357039928436, + 0.00648806057870388, + 0.006109227426350117, + 0.0035713710822165012, + 0.0062184142880141735, + 0.005369058810174465, + 0.004703831393271685, + 0.009274009615182877, + 0.005063778255134821, + 0.007325771730393171, + 0.0026858835481107235, + 0.005208183545619249, + 0.0035248363856226206, + 0.0042176577262580395, + 0.004661142360419035, + 0.006673978641629219, + 0.0053358981385827065, + 0.0076643330976367, + 0.003412136109545827, + 0.0019859890453517437, + 0.003581981174647808, + 0.0019564309623092413, + 0.0033144906628876925, + 0.014402642846107483, + 0.0034835010301321745, + 0.014664553105831146, + 0.00782736111432314, + 0.004337813705205917, + 0.009013788774609566, + 0.003651076229289174, + 0.0052688634023070335, + 0.019483303651213646, + 0.006111958064138889, + 0.020302115008234978, + 0.004755285568535328, + 0.004943235777318478, + 0.006447458174079657, + 0.0036771090235561132, + 0.005418562795966864, + 0.019275501370429993, + 0.0062020947225391865, + 0.020353803411126137, + 0.01250825822353363, + 0.008255967870354652, + 0.012358355335891247, + 0.00784001313149929, + 0.008218145929276943, + 0.0248376727104187, + 0.009086142294108868, + 0.021965255960822105, + 0.010130607523024082, + 0.011354737915098667, + 0.007486918941140175, + 0.00815089326351881, + 0.03201975300908089, + 0.15560568869113922, + 0.026170026510953903, + 0.14860142767429352 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 0, + 8 + ], + "selected_reward": -0.005821772385388613 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme macro view focusing on the internal structures of the human eye?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the central surface pearly, slightly yellowish, and damp with a sheen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are crimson veins visible branching like a leafless tree with some dense protruding vessels and hair-like capillaries?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the veins exhibit a subtle, rhythmic pulsation expanding and contracting in a steady biological rhythm?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the edge of a dark brown iris with an intricate fibrous pattern visible on one side of the frame?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.011101149953901768 + ], + "opposite_logprob_tokens": [ + -4.511101245880127 + ], + "logprob_reward": -0.011101149953901768, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.011101149953901768, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3048291504383087, + 0.14554011821746826, + 0.03587765246629715, + 0.018999658524990082, + 0.022150278091430664, + 0.02968715876340866, + 0.09056930243968964, + 0.07432729005813599, + 0.016645681113004684, + 0.26137375831604004 + ], + "credit_map": [ + 0.014696613885462284, + 0.01637718454003334, + 0.013765781186521053, + 0.022234560921788216, + 0.03976670652627945, + 0.09457138925790787, + 0.03383152559399605, + 0.0695854127407074, + 0.007280842401087284, + 0.0044348519295454025, + 0.007030718494206667, + 0.004569820128381252, + 0.015695495530962944, + 0.051057010889053345, + 0.01101668644696474, + 0.04445469006896019, + 0.004238187335431576, + 0.00413895258679986, + 0.004899351391941309, + 0.003567684441804886, + 0.005137085448950529, + 0.004986667539924383, + 0.00455916952341795, + 0.004350556526333094, + 0.0016064911615103483, + 0.002200555521994829, + 0.0016315896064043045, + 0.002906043315306306, + 0.003188192844390869, + 0.002842418383806944, + 0.0025408300571143627, + 0.0020835374016314745, + 0.001589630963280797, + 0.0013934377348050475, + 0.0018199589103460312, + 0.0017390961293131113, + 0.0028454740531742573, + 0.005539424251765013, + 0.0021670465357601643, + 0.005056209396570921, + 0.002643004059791565, + 0.002152977744117379, + 0.002919028513133526, + 0.002145984210073948, + 0.0023239050060510635, + 0.007499254308640957, + 0.0023685500491410494, + 0.007634456269443035, + 0.004799349699169397, + 0.003600322874262929, + 0.005240248516201973, + 0.003167447866871953, + 0.007575449533760548, + 0.03031003661453724, + 0.006306712981313467, + 0.029569728299975395, + 0.005449154879897833, + 0.003672828199341893, + 0.005707448814064264, + 0.00341434171423316, + 0.006126086693257093, + 0.02137041836977005, + 0.006087263114750385, + 0.022499747574329376, + 0.0033718219492584467, + 0.001223970321007073, + 0.003626770805567503, + 0.0012318530352786183, + 0.001066564116626978, + 0.0019356822595000267, + 0.0018859944539144635, + 0.002303024521097541, + 0.011009975336492062, + 0.011748459190130234, + 0.009830337017774582, + 0.0100077074021101, + 0.02017786353826523, + 0.09330008178949356, + 0.01719251088798046, + 0.08810681104660034 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 1 + ], + "selected_reward": -0.011101149953901768 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme macro view focusing on the internal structures of the human eye?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the central surface pearly, slightly yellowish, and damp with a sheen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are crimson veins visible branching like a leafless tree with some dense protruding vessels and hair-like capillaries?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the veins exhibit a subtle, rhythmic pulsation expanding and contracting in a steady biological rhythm?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the edge of a dark brown iris with an intricate fibrous pattern visible on one side of the frame?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0019491974962875247 + ], + "opposite_logprob_tokens": [ + -6.251949310302734 + ], + "logprob_reward": -0.0019491974962875247, + "yes_no_margin_reward": 6.25, + "token_credit_reward": -0.0019491974962875247, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.16687937080860138, + 0.19528882205486298, + 0.11395645141601562, + 0.051988184452056885, + 0.04000450298190117, + 0.07096793502569199, + 0.06314357370138168, + 0.07107263803482056, + 0.037152234464883804, + 0.18954633176326752 + ], + "credit_map": [ + 0.023072218522429466, + 0.02055216208100319, + 0.02346641570329666, + 0.02175888419151306, + 0.015649128705263138, + 0.01433718204498291, + 0.027335934340953827, + 0.02070741541683674, + 0.010993780568242073, + 0.008995992131531239, + 0.012087026610970497, + 0.011761129833757877, + 0.021791305392980576, + 0.06185443326830864, + 0.017502455040812492, + 0.05030267313122749, + 0.011074589565396309, + 0.014569018967449665, + 0.012075317092239857, + 0.014383322559297085, + 0.01676332578063011, + 0.014390022493898869, + 0.01360053475946188, + 0.01710030809044838, + 0.005095216911286116, + 0.004917209967970848, + 0.0051625873893499374, + 0.006455426570028067, + 0.00678037665784359, + 0.009507989510893822, + 0.007510504685342312, + 0.006558866705745459, + 0.003939387388527393, + 0.003502278821542859, + 0.00429430091753602, + 0.004293026868253946, + 0.004618419334292412, + 0.006456547882407904, + 0.004605989903211594, + 0.00829454604536295, + 0.003222865052521229, + 0.006166448350995779, + 0.0030121554154902697, + 0.005841373465955257, + 0.010217170231044292, + 0.01723361760377884, + 0.007972696796059608, + 0.017301596701145172, + 0.0036103397142142057, + 0.0030891750939190388, + 0.004085447173565626, + 0.003050426719710231, + 0.005036885384470224, + 0.01831509731709957, + 0.005515485070645809, + 0.020440712571144104, + 0.006655746139585972, + 0.007485519628971815, + 0.007781116291880608, + 0.0062892381101846695, + 0.009569255635142326, + 0.013203725218772888, + 0.008236792869865894, + 0.011851237155497074, + 0.004330351948738098, + 0.002992188325151801, + 0.004150892607867718, + 0.002052637981250882, + 0.00267868977971375, + 0.00822975393384695, + 0.0035758630838245153, + 0.009141849353909492, + 0.005127602256834507, + 0.007692078594118357, + 0.005552199669182301, + 0.0057203262113034725, + 0.016626976430416107, + 0.06629987806081772, + 0.014849252998828888, + 0.06767798960208893 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 9, + 0 + ], + "selected_reward": -0.0019491974962875247 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme macro view focusing on the internal structures of the human eye?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the central surface pearly, slightly yellowish, and damp with a sheen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are crimson veins visible branching like a leafless tree with some dense protruding vessels and hair-like capillaries?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the veins exhibit a subtle, rhythmic pulsation expanding and contracting in a steady biological rhythm?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the edge of a dark brown iris with an intricate fibrous pattern visible on one side of the frame?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.1428224742412567 + ], + "opposite_logprob_tokens": [ + -2.017822504043579 + ], + "logprob_reward": -0.1428224742412567, + "yes_no_margin_reward": 1.875, + "token_credit_reward": -0.1428224742412567, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.25505512952804565, + 0.07853788137435913, + 0.1422411948442459, + 0.04987166076898575, + 0.047086361795663834, + 0.09696152061223984, + 0.07428507506847382, + 0.08329044282436371, + 0.0950508788228035, + 0.07761985063552856 + ], + "credit_map": [ + 0.029729047790169716, + 0.0221933014690876, + 0.030608205124735832, + 0.029368745163083076, + 0.031587984412908554, + 0.05085548385977745, + 0.02817438170313835, + 0.032537974417209625, + 0.008094925433397293, + 0.009895573370158672, + 0.007415772415697575, + 0.012003351002931595, + 0.00978720560669899, + 0.007591244764626026, + 0.015389334410429, + 0.008360466919839382, + 0.017041124403476715, + 0.013218129053711891, + 0.017857074737548828, + 0.01777738519012928, + 0.016371000558137894, + 0.023198502138257027, + 0.01701633259654045, + 0.019761644303798676, + 0.0038986385334283113, + 0.006077883765101433, + 0.004674142692238092, + 0.006976710166782141, + 0.005187615752220154, + 0.0061761545948684216, + 0.009950662963092327, + 0.006929850671440363, + 0.00415545841678977, + 0.004441587720066309, + 0.0037478883750736713, + 0.005639891140162945, + 0.004831655882298946, + 0.009331553243100643, + 0.005938855931162834, + 0.008999470621347427, + 0.006972495466470718, + 0.005447974894195795, + 0.008068738505244255, + 0.005134114995598793, + 0.006672591436654329, + 0.028997763991355896, + 0.006994309835135937, + 0.028673531487584114, + 0.009595165029168129, + 0.007906787097454071, + 0.011067044921219349, + 0.007055537775158882, + 0.0074564567767083645, + 0.012758098542690277, + 0.008397620171308517, + 0.010048355907201767, + 0.005201617255806923, + 0.006492744665592909, + 0.006222687195986509, + 0.004461969248950481, + 0.006980244070291519, + 0.02185313031077385, + 0.007349345367401838, + 0.024728702381253242, + 0.0024827346205711365, + 0.002617583842948079, + 0.002355615608394146, + 0.0026012419257313013, + 0.0073777916841208935, + 0.03500882163643837, + 0.0074231806211173534, + 0.03518390655517578, + 0.0036268627736717463, + 0.0032351233530789614, + 0.0030786374118179083, + 0.002870393916964531, + 0.006543908733874559, + 0.028021685779094696, + 0.00539243221282959, + 0.024850808084011078 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 5 + ], + "selected_reward": -0.1428224742412567 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme macro view focusing on the internal structures of the human eye?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the central surface pearly, slightly yellowish, and damp with a sheen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are crimson veins visible branching like a leafless tree with some dense protruding vessels and hair-like capillaries?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the veins exhibit a subtle, rhythmic pulsation expanding and contracting in a steady biological rhythm?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the edge of a dark brown iris with an intricate fibrous pattern visible on one side of the frame?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01606578752398491 + ], + "opposite_logprob_tokens": [ + -4.14106559753418 + ], + "logprob_reward": -0.01606578752398491, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.01606578752398491, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09242746978998184, + 0.24560114741325378, + 0.051006972789764404, + 0.03004576452076435, + 0.04012031480669975, + 0.050821442157030106, + 0.1458197832107544, + 0.05886105075478554, + 0.058348510414361954, + 0.22694756090641022 + ], + "credit_map": [ + 0.009888497181236744, + 0.011593172326683998, + 0.010709716007113457, + 0.016797155141830444, + 0.010139786638319492, + 0.009413476102054119, + 0.011522220447659492, + 0.012363443151116371, + 0.010683569125831127, + 0.008068609051406384, + 0.009723207913339138, + 0.008693036623299122, + 0.02827896922826767, + 0.08908861130475998, + 0.015787094831466675, + 0.07527805864810944, + 0.00675272848457098, + 0.004063333850353956, + 0.00784346368163824, + 0.005451998673379421, + 0.006297625135630369, + 0.008402797393500805, + 0.005311902612447739, + 0.006883127149194479, + 0.003933441825211048, + 0.003310392377898097, + 0.004449040163308382, + 0.003262958722189069, + 0.0028800477739423513, + 0.004622208885848522, + 0.0031738518737256527, + 0.0044138263911008835, + 0.0046965512447059155, + 0.0030378394294530153, + 0.004932533949613571, + 0.003373753745108843, + 0.0030570081435143948, + 0.0071080829948186874, + 0.004342041444033384, + 0.00957250501960516, + 0.005257369484752417, + 0.003221402643248439, + 0.005458547733724117, + 0.004309365060180426, + 0.003764835884794593, + 0.010735010728240013, + 0.0047476994805037975, + 0.013327214866876602, + 0.007649304810911417, + 0.0045508998446166515, + 0.009073645807802677, + 0.004572991281747818, + 0.011017593555152416, + 0.0502324178814888, + 0.010318586602807045, + 0.04840433597564697, + 0.00616186810657382, + 0.004442967474460602, + 0.007712068501859903, + 0.003914018627256155, + 0.004559151362627745, + 0.011503312736749649, + 0.00512654148042202, + 0.01544112153351307, + 0.006571977864950895, + 0.0031454393174499273, + 0.005934373941272497, + 0.0033585629425942898, + 0.003938662353903055, + 0.015428638085722923, + 0.004743989557027817, + 0.015226864255964756, + 0.005825356114655733, + 0.005912144202739, + 0.005514321383088827, + 0.0033415043726563454, + 0.017398221418261528, + 0.08817284554243088, + 0.01623358763754368, + 0.08454956859350204 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 9, + 6 + ], + "selected_reward": -0.01606578752398491 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "static yet unwavering extreme macro shot delves into the surreal landscape of the human eye in lifelike 4K detail. The camera focuses on the scrotum, not a pure white, but a pearly, slightly yellowish-yellow surface with a damp sheen, woven through a complex network of crimson veins. These intricate, blood-filled channels, rendered with breathtaking clarity, branch like a delicate, leafless tree, some appearing as dense, protruding vessels, while others fragment into hair-like capillaries. The central action is the subtle, rhythmic pulsation within these veins; they expand and contract almost imperceptibly in a steady biological rhythm, suggesting the silent beating of a heart. Bright, unfocused, clinical lighting illuminates every tiny texture without casting harsh shadows and creating small, sharp, specular highlights along the curve of the cornea. The edge of a dark brown iris, with its intricate fibrous pattern, is visible on one side of the frame, creating a scrotal landscape reminiscent of an alien. The overall atmosphere is one of intense, almost uncomfortable intimacy and biological fascination, a visceral exploration of the inner mechanisms of life.", + "seed": 32401, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000006/group_5/index_6003_seed_32401_dp_11_rank_22_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 5, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.07667159531265497, + "vlm_logprob": -0.07667159531265497, + "vlm_yes_no_margin": 3.874999988079071, + "vlm_token_credit": -0.07667159531265497, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme macro view focusing on the internal structures of the human eye?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the central surface pearly, slightly yellowish, and damp with a sheen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are crimson veins visible branching like a leafless tree with some dense protruding vessels and hair-like capillaries?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the veins exhibit a subtle, rhythmic pulsation expanding and contracting in a steady biological rhythm?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the edge of a dark brown iris with an intricate fibrous pattern visible on one side of the frame?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.007301905658096075 + ], + "opposite_logprob_tokens": [ + -5.7573018074035645 + ], + "logprob_reward": -0.007301905658096075, + "yes_no_margin_reward": 5.75, + "token_credit_reward": -0.007301905658096075, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.26920342445373535, + 0.3069959878921509, + 0.05262383818626404, + 0.05926185101270676, + 0.09811299294233322, + 0.039734967052936554, + 0.0541873425245285, + 0.03923977166414261, + 0.03262271359562874, + 0.048017144203186035 + ], + "credit_map": [ + 0.023507455363869667, + 0.015396823175251484, + 0.02609860897064209, + 0.015781981870532036, + 0.013170383870601654, + 0.07127965986728668, + 0.07963725924491882, + 0.024331239983439445, + 0.02637639082968235, + 0.023314710706472397, + 0.02521427720785141, + 0.02446565218269825, + 0.018676450476050377, + 0.0958155021071434, + 0.04626650735735893, + 0.04686649516224861, + 0.0075751617550849915, + 0.007044210098683834, + 0.007368018385022879, + 0.00536473048850894, + 0.00465531088411808, + 0.006686916574835777, + 0.006445783190429211, + 0.0074837044812738895, + 0.004823362920433283, + 0.005155236925929785, + 0.00738182058557868, + 0.005032730288803577, + 0.007894674316048622, + 0.009538069367408752, + 0.0077946987003088, + 0.011641254648566246, + 0.008239822462201118, + 0.012834418565034866, + 0.010706506669521332, + 0.008780320174992085, + 0.00853438675403595, + 0.01857472024857998, + 0.016377095133066177, + 0.014065721072256565, + 0.004591533448547125, + 0.004920202773064375, + 0.005968446843326092, + 0.0035077324137091637, + 0.0037395127583295107, + 0.0047561307437717915, + 0.008512419648468494, + 0.0037389860954135656, + 0.006654089316725731, + 0.004756787791848183, + 0.01029437966644764, + 0.005441590212285519, + 0.003697766689583659, + 0.005404922179877758, + 0.010360358282923698, + 0.007577451411634684, + 0.004989992827177048, + 0.004491280764341354, + 0.007744065020233393, + 0.0032714984845370054, + 0.0024400039110332727, + 0.00381451821886003, + 0.006467150058597326, + 0.006021261680871248, + 0.002970700850710273, + 0.003983682487159967, + 0.004962046165019274, + 0.003552880836650729, + 0.00322492397390306, + 0.003200560575351119, + 0.002855879021808505, + 0.007872037589550018, + 0.005644442047923803, + 0.006533329840749502, + 0.008083103224635124, + 0.005158863961696625, + 0.00477471761405468, + 0.00465792790055275, + 0.006127032916992903, + 0.007037726230919361 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 4 + ], + "selected_reward": -0.007301905658096075 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme macro view focusing on the internal structures of the human eye?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the central surface pearly, slightly yellowish, and damp with a sheen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are crimson veins visible branching like a leafless tree with some dense protruding vessels and hair-like capillaries?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the veins exhibit a subtle, rhythmic pulsation expanding and contracting in a steady biological rhythm?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the edge of a dark brown iris with an intricate fibrous pattern visible on one side of the frame?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005278815981000662 + ], + "opposite_logprob_tokens": [ + -5.255278587341309 + ], + "logprob_reward": -0.005278815981000662, + "yes_no_margin_reward": 5.25, + "token_credit_reward": -0.005278815981000662, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.4354206323623657, + 0.11644008010625839, + 0.07637618482112885, + 0.030940260738134384, + 0.097598597407341, + 0.07737772911787033, + 0.03380303457379341, + 0.043140292167663574, + 0.028505582362413406, + 0.060397617518901825 + ], + "credit_map": [ + 0.02304389700293541, + 0.022553520277142525, + 0.02439860813319683, + 0.02989320456981659, + 0.02555268071591854, + 0.16043026745319366, + 0.12803658843040466, + 0.021511845290660858, + 0.010148468427360058, + 0.011507046408951283, + 0.009368532337248325, + 0.00820864923298359, + 0.007911632768809795, + 0.0289362370967865, + 0.022647779434919357, + 0.01771172694861889, + 0.008942977525293827, + 0.007892902009189129, + 0.008852794766426086, + 0.008087268099188805, + 0.006767460610717535, + 0.010095460340380669, + 0.011850765906274319, + 0.013886549510061741, + 0.0034324966836720705, + 0.003307152772322297, + 0.004787509795278311, + 0.0029497873038053513, + 0.0043453373946249485, + 0.004251306876540184, + 0.00304809445515275, + 0.0048185717314481735, + 0.008462185971438885, + 0.008970207534730434, + 0.015139400959014893, + 0.007550397422164679, + 0.007639147341251373, + 0.019318358972668648, + 0.019020047038793564, + 0.011498846113681793, + 0.009143130853772163, + 0.006091872230172157, + 0.011352757923305035, + 0.006814731750637293, + 0.005027484614402056, + 0.011773713864386082, + 0.015767639502882957, + 0.011406393721699715, + 0.003608472878113389, + 0.006502686068415642, + 0.004455179441720247, + 0.004624910652637482, + 0.003235032083466649, + 0.0035598541144281626, + 0.004531459882855415, + 0.003285434562712908, + 0.00562146445736289, + 0.004759077914059162, + 0.006049838848412037, + 0.0050062923692166805, + 0.003751452313736081, + 0.004884518217295408, + 0.0063249352388083935, + 0.0067427074536681175, + 0.002941465238109231, + 0.0024339223746210337, + 0.0033745041582733393, + 0.0029976489022374153, + 0.0037431917153298855, + 0.004685559775680304, + 0.0033175155986100435, + 0.005011773202568293, + 0.007207322865724564, + 0.006080735940486193, + 0.010277409106492996, + 0.0065107145346701145, + 0.008143303915858269, + 0.006316893268376589, + 0.00781880784779787, + 0.008042425848543644 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 4 + ], + "selected_reward": -0.005278815981000662 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme macro view focusing on the internal structures of the human eye?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the central surface pearly, slightly yellowish, and damp with a sheen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are crimson veins visible branching like a leafless tree with some dense protruding vessels and hair-like capillaries?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the veins exhibit a subtle, rhythmic pulsation expanding and contracting in a steady biological rhythm?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the edge of a dark brown iris with an intricate fibrous pattern visible on one side of the frame?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01421516202390194 + ], + "opposite_logprob_tokens": [ + -4.264214992523193 + ], + "logprob_reward": -0.01421516202390194, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.01421516202390194, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3603024482727051, + 0.21103134751319885, + 0.116474449634552, + 0.04164613410830498, + 0.08142635226249695, + 0.08876753598451614, + 0.023991378024220467, + 0.017017513513565063, + 0.015238068997859955, + 0.04410479962825775 + ], + "credit_map": [ + 0.03193560242652893, + 0.03546326234936714, + 0.025994863361120224, + 0.02844109944999218, + 0.03577793017029762, + 0.07558032870292664, + 0.06797425448894501, + 0.05913509428501129, + 0.018665699288249016, + 0.01975618675351143, + 0.013560198247432709, + 0.014352060854434967, + 0.014272436499595642, + 0.06201356276869774, + 0.03380684554576874, + 0.034604351967573166, + 0.013755735009908676, + 0.013132956810295582, + 0.009405598975718021, + 0.009694576263427734, + 0.009136727079749107, + 0.024205947294831276, + 0.012692663818597794, + 0.02445024624466896, + 0.0029660577420145273, + 0.004755562171339989, + 0.003552880370989442, + 0.003424297319725156, + 0.0044190892949700356, + 0.008589698933064938, + 0.0058939834125339985, + 0.00804456602782011, + 0.006752028129994869, + 0.007562305778264999, + 0.007280780002474785, + 0.006218865048140287, + 0.007175948936492205, + 0.017709966748952866, + 0.01572592929005623, + 0.013000528328120708, + 0.00892786867916584, + 0.008247719146311283, + 0.007917243055999279, + 0.007500687148422003, + 0.006839032284915447, + 0.01635177806019783, + 0.020493540912866592, + 0.012489667162299156, + 0.002885863883420825, + 0.0023326780647039413, + 0.0037825412582606077, + 0.0019730492495000362, + 0.0016057268949225545, + 0.00272934976965189, + 0.004860799293965101, + 0.003821369493380189, + 0.0015745492419227958, + 0.0018230185378342867, + 0.001650968217290938, + 0.001470605144277215, + 0.001420786720700562, + 0.0024875092785805464, + 0.0025815351400524378, + 0.004008539952337742, + 0.001761624007485807, + 0.001999428728595376, + 0.002390594221651554, + 0.0018935810076072812, + 0.0018934615654870868, + 0.001966702751815319, + 0.0016240704571828246, + 0.0017086054431274533, + 0.004049165174365044, + 0.004687909036874771, + 0.004331640433520079, + 0.004224533215165138, + 0.005330475978553295, + 0.005245676264166832, + 0.005551542155444622, + 0.010683858767151833 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 2 + ], + "selected_reward": -0.01421516202390194 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme macro view focusing on the internal structures of the human eye?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the central surface pearly, slightly yellowish, and damp with a sheen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are crimson veins visible branching like a leafless tree with some dense protruding vessels and hair-like capillaries?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the veins exhibit a subtle, rhythmic pulsation expanding and contracting in a steady biological rhythm?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the edge of a dark brown iris with an intricate fibrous pattern visible on one side of the frame?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.31350451707839966 + ], + "opposite_logprob_tokens": [ + -1.3135044574737549 + ], + "logprob_reward": -0.31350451707839966, + "yes_no_margin_reward": 0.9999999403953552, + "token_credit_reward": -0.31350451707839966, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3614433705806732, + 0.39494240283966064, + 0.04429119825363159, + 0.03243795782327652, + 0.02403859980404377, + 0.014158101752400398, + 0.038975950330495834, + 0.022664103657007217, + 0.03474750369787216, + 0.032300788909196854 + ], + "credit_map": [ + 0.038483306765556335, + 0.04159310832619667, + 0.033527396619319916, + 0.03232834115624428, + 0.03624722361564636, + 0.06225619092583656, + 0.05263334885239601, + 0.06437447667121887, + 0.03532484173774719, + 0.03401675820350647, + 0.025825664401054382, + 0.026573851704597473, + 0.023689288645982742, + 0.11401424556970596, + 0.08293810486793518, + 0.05255968123674393, + 0.004708013962954283, + 0.004181636031717062, + 0.00564580550417304, + 0.0036022369749844074, + 0.003179628401994705, + 0.008665301837027073, + 0.004991644062101841, + 0.009316935203969479, + 0.004055191297084093, + 0.0030998566653579473, + 0.004839254077523947, + 0.002565105212852359, + 0.00454906839877367, + 0.004004237242043018, + 0.0038616808596998453, + 0.005463568959385157, + 0.003566168714314699, + 0.0035461257211863995, + 0.002946985187008977, + 0.002533610677346587, + 0.0020175136160105467, + 0.0029457202181220055, + 0.002170708030462265, + 0.0043117692694067955, + 0.0014232862740755081, + 0.0016428478993475437, + 0.0018676469335332513, + 0.0017948552267625928, + 0.0015980857424438, + 0.001892564818263054, + 0.0016367995413020253, + 0.002302015433087945, + 0.006575664039701223, + 0.005871153436601162, + 0.006795160938054323, + 0.0040117306634783745, + 0.0030838230159133673, + 0.0036412496119737625, + 0.0042518251575529575, + 0.00474534509703517, + 0.0024968604557216167, + 0.0030016067903488874, + 0.00314659858122468, + 0.002604896668344736, + 0.0025254383217543364, + 0.002228569006547332, + 0.002456843387335539, + 0.004203292541205883, + 0.0032626716420054436, + 0.0034281390253454447, + 0.004133631940931082, + 0.0030368450097739697, + 0.003749745897948742, + 0.005388988181948662, + 0.004051616881042719, + 0.007695868145674467, + 0.0032809448894113302, + 0.0029672281816601753, + 0.004500212147831917, + 0.0032651484943926334, + 0.003982121590524912, + 0.0038841161876916885, + 0.003962467424571514, + 0.006458552088588476 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 2 + ], + "selected_reward": -0.31350451707839966 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme macro view focusing on the internal structures of the human eye?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the central surface pearly, slightly yellowish, and damp with a sheen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are crimson veins visible branching like a leafless tree with some dense protruding vessels and hair-like capillaries?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the veins exhibit a subtle, rhythmic pulsation expanding and contracting in a steady biological rhythm?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the edge of a dark brown iris with an intricate fibrous pattern visible on one side of the frame?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.043057575821876526 + ], + "opposite_logprob_tokens": [ + -3.168057680130005 + ], + "logprob_reward": -0.043057575821876526, + "yes_no_margin_reward": 3.125, + "token_credit_reward": -0.043057575821876526, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1855718195438385, + 0.169594407081604, + 0.09481338411569595, + 0.09455986320972443, + 0.1481216996908188, + 0.14471903443336487, + 0.0302213653922081, + 0.05248170718550682, + 0.018056025728583336, + 0.06186075881123543 + ], + "credit_map": [ + 0.02058965526521206, + 0.020042238757014275, + 0.024065662175416946, + 0.013174552470445633, + 0.0171333160251379, + 0.029994284734129906, + 0.024470455944538116, + 0.03610164299607277, + 0.01719120517373085, + 0.015695994719862938, + 0.021448487415909767, + 0.0225064717233181, + 0.012749021872878075, + 0.043878864496946335, + 0.013877619989216328, + 0.02224672958254814, + 0.013794120401144028, + 0.009563453495502472, + 0.01913810335099697, + 0.006880410946905613, + 0.005715229082852602, + 0.01187918335199356, + 0.017804831266403198, + 0.010038061998784542, + 0.011883798986673355, + 0.014209011569619179, + 0.014655765146017075, + 0.006887901574373245, + 0.013658345676958561, + 0.012651021592319012, + 0.009704725816845894, + 0.01090929377824068, + 0.015796877443790436, + 0.01780472882091999, + 0.02426302433013916, + 0.013531236909329891, + 0.010145703330636024, + 0.023960860446095467, + 0.02717781811952591, + 0.015441441908478737, + 0.017258577048778534, + 0.016056841239333153, + 0.022127481177449226, + 0.014441806823015213, + 0.009785976260900497, + 0.016801761463284492, + 0.02607722580432892, + 0.022169359028339386, + 0.003936187829822302, + 0.005672249011695385, + 0.005427931435406208, + 0.003436069702729583, + 0.0027589784003794193, + 0.002169197890907526, + 0.0031505387742072344, + 0.0036702125798910856, + 0.007794070523232222, + 0.0050642224960029125, + 0.011219904758036137, + 0.003978838212788105, + 0.005506578367203474, + 0.00536306481808424, + 0.0065846689976751804, + 0.006970360409468412, + 0.0026906540151685476, + 0.0016858106246218085, + 0.003744551446288824, + 0.0020122574642300606, + 0.0015323528787121177, + 0.0023981330450624228, + 0.0023564568255096674, + 0.0016358103603124619, + 0.007380098104476929, + 0.004700642544776201, + 0.011338644661009312, + 0.0053235809318721294, + 0.0058095138520002365, + 0.007039414718747139, + 0.00909334048628807, + 0.011175524443387985 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 4 + ], + "selected_reward": -0.043057575821876526 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "static yet unwavering extreme macro shot delves into the surreal landscape of the human eye in lifelike 4K detail. The camera focuses on the scrotum, not a pure white, but a pearly, slightly yellowish-yellow surface with a damp sheen, woven through a complex network of crimson veins. These intricate, blood-filled channels, rendered with breathtaking clarity, branch like a delicate, leafless tree, some appearing as dense, protruding vessels, while others fragment into hair-like capillaries. The central action is the subtle, rhythmic pulsation within these veins; they expand and contract almost imperceptibly in a steady biological rhythm, suggesting the silent beating of a heart. Bright, unfocused, clinical lighting illuminates every tiny texture without casting harsh shadows and creating small, sharp, specular highlights along the curve of the cornea. The edge of a dark brown iris, with its intricate fibrous pattern, is visible on one side of the frame, creating a scrotal landscape reminiscent of an alien. The overall atmosphere is one of intense, almost uncomfortable intimacy and biological fascination, a visceral exploration of the inner mechanisms of life.", + "seed": 32402, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000006/group_5/index_6003_seed_32402_dp_11_rank_22_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 5, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.05941395042464137, + "vlm_logprob": -0.05941395042464137, + "vlm_yes_no_margin": 3.85, + "vlm_token_credit": -0.05941395042464137, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme macro view focusing on the internal structures of the human eye?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the central surface pearly, slightly yellowish, and damp with a sheen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are crimson veins visible branching like a leafless tree with some dense protruding vessels and hair-like capillaries?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the veins exhibit a subtle, rhythmic pulsation expanding and contracting in a steady biological rhythm?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the edge of a dark brown iris with an intricate fibrous pattern visible on one side of the frame?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.012006042525172234 + ], + "opposite_logprob_tokens": [ + -5.012005805969238 + ], + "logprob_reward": -0.012006042525172234, + "yes_no_margin_reward": 5.0, + "token_credit_reward": -0.012006042525172234, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.24873299896717072, + 0.07614605128765106, + 0.0670512467622757, + 0.12084459513425827, + 0.11364775151014328, + 0.09775480628013611, + 0.05195853114128113, + 0.15191341936588287, + 0.030803535133600235, + 0.04114711284637451 + ], + "credit_map": [ + 0.02189636416733265, + 0.034984294325113297, + 0.022491415962576866, + 0.023454248905181885, + 0.03286558389663696, + 0.03972635045647621, + 0.02969079092144966, + 0.043623924255371094, + 0.005334767512977123, + 0.010878720320761204, + 0.008570587262511253, + 0.008075425401329994, + 0.009983804076910019, + 0.011605664156377316, + 0.010635090991854668, + 0.01106197852641344, + 0.006314018741250038, + 0.0077438694424927235, + 0.011927804909646511, + 0.0063094752840697765, + 0.008141973987221718, + 0.009022679179906845, + 0.009636259637773037, + 0.007955161854624748, + 0.010279069654643536, + 0.015355261042714119, + 0.021306099370121956, + 0.014266771264374256, + 0.016297627240419388, + 0.016463009640574455, + 0.012399565428495407, + 0.014477179385721684, + 0.011237974278628826, + 0.013703743927180767, + 0.017689477652311325, + 0.013627484440803528, + 0.016787679865956306, + 0.015596487559378147, + 0.014484148472547531, + 0.010520740412175655, + 0.010572451166808605, + 0.011410295963287354, + 0.01400140579789877, + 0.011102302931249142, + 0.01429516077041626, + 0.01395388413220644, + 0.012924028560519218, + 0.009495275095105171, + 0.004215292166918516, + 0.007704244460910559, + 0.008966841734945774, + 0.005533935036510229, + 0.0070971278473734856, + 0.005985583178699017, + 0.005543542560189962, + 0.006911960896104574, + 0.014155974611639977, + 0.019249416887760162, + 0.018226664513349533, + 0.01642860472202301, + 0.022959427908062935, + 0.024944160133600235, + 0.020174315199255943, + 0.015774836763739586, + 0.0036888474132865667, + 0.004682107828557491, + 0.003293687244877219, + 0.003229525638744235, + 0.004452581517398357, + 0.0037051886320114136, + 0.003896294394508004, + 0.0038552996702492237, + 0.003178217215463519, + 0.005487553775310516, + 0.0057383943349123, + 0.004675824195146561, + 0.005447628442198038, + 0.006606861483305693, + 0.00496523454785347, + 0.005047398153692484 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 7, + 3 + ], + "selected_reward": -0.012006042525172234 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme macro view focusing on the internal structures of the human eye?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the central surface pearly, slightly yellowish, and damp with a sheen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are crimson veins visible branching like a leafless tree with some dense protruding vessels and hair-like capillaries?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the veins exhibit a subtle, rhythmic pulsation expanding and contracting in a steady biological rhythm?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the edge of a dark brown iris with an intricate fibrous pattern visible on one side of the frame?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.011116358451545238 + ], + "opposite_logprob_tokens": [ + -4.5111165046691895 + ], + "logprob_reward": -0.011116358451545238, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.011116358451545238, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.26706555485725403, + 0.08844178915023804, + 0.11623857915401459, + 0.08244146406650543, + 0.12468302994966507, + 0.06606631726026535, + 0.03690047189593315, + 0.13742031157016754, + 0.03737441450357437, + 0.043368104845285416 + ], + "credit_map": [ + 0.02075648494064808, + 0.0411883145570755, + 0.031855907291173935, + 0.023373475298285484, + 0.03419225290417671, + 0.037968192249536514, + 0.041349269449710846, + 0.03638162836432457, + 0.005693559069186449, + 0.012153828516602516, + 0.010045629926025867, + 0.007586032152175903, + 0.012216860428452492, + 0.01207080576568842, + 0.018084611743688583, + 0.010590451769530773, + 0.010976974852383137, + 0.012143879197537899, + 0.02069469913840294, + 0.011687778867781162, + 0.014626427553594112, + 0.01744813844561577, + 0.016278306022286415, + 0.012382370419800282, + 0.0077520329505205154, + 0.010618277825415134, + 0.011879919096827507, + 0.008514542132616043, + 0.01151733286678791, + 0.012188181281089783, + 0.010039481334388256, + 0.009931682609021664, + 0.010380729101598263, + 0.018059050664305687, + 0.016942040994763374, + 0.015271027572453022, + 0.017543021589517593, + 0.01744130812585354, + 0.01660875231027603, + 0.012437092140316963, + 0.006720816716551781, + 0.008684330619871616, + 0.01099075935781002, + 0.006540201138705015, + 0.008903314359486103, + 0.007709756027907133, + 0.00942165032029152, + 0.007095482666045427, + 0.0023862547241151333, + 0.00440778024494648, + 0.003850220236927271, + 0.0033497533295303583, + 0.005771156400442123, + 0.006492367945611477, + 0.0061185103841125965, + 0.004524422809481621, + 0.012131932191550732, + 0.018163161352276802, + 0.015780044719576836, + 0.014913616701960564, + 0.02050038054585457, + 0.023881470784544945, + 0.01891265995800495, + 0.0131370285525918, + 0.003314241301268339, + 0.004335702862590551, + 0.007282068952918053, + 0.0045365942642092705, + 0.0038077975623309612, + 0.005165654234588146, + 0.004325810354202986, + 0.004606542643159628, + 0.0035785131622105837, + 0.004895837977528572, + 0.006175811868160963, + 0.004851080477237701, + 0.005414426792412996, + 0.0061195301823318005, + 0.007129875011742115, + 0.005203024949878454 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 7, + 4 + ], + "selected_reward": -0.011116358451545238 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme macro view focusing on the internal structures of the human eye?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the central surface pearly, slightly yellowish, and damp with a sheen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are crimson veins visible branching like a leafless tree with some dense protruding vessels and hair-like capillaries?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the veins exhibit a subtle, rhythmic pulsation expanding and contracting in a steady biological rhythm?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the edge of a dark brown iris with an intricate fibrous pattern visible on one side of the frame?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005264822859317064 + ], + "opposite_logprob_tokens": [ + -5.255264759063721 + ], + "logprob_reward": -0.005264822859317064, + "yes_no_margin_reward": 5.25, + "token_credit_reward": -0.005264822859317064, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.20008288323879242, + 0.0822511687874794, + 0.07399247586727142, + 0.063621886074543, + 0.09947416931390762, + 0.09836982935667038, + 0.08756403625011444, + 0.17154739797115326, + 0.057278864085674286, + 0.06581734120845795 + ], + "credit_map": [ + 0.01848958246409893, + 0.03355666622519493, + 0.019567593932151794, + 0.017223669216036797, + 0.02447614260017872, + 0.027739042416214943, + 0.03337359428405762, + 0.025656595826148987, + 0.006839002948254347, + 0.012579272501170635, + 0.012325871735811234, + 0.00794614851474762, + 0.011476377956569195, + 0.009379932656884193, + 0.009731640107929707, + 0.01197291724383831, + 0.0068619949743151665, + 0.01186414249241352, + 0.0075917853973805904, + 0.0070621781051158905, + 0.01331835426390171, + 0.00896699819713831, + 0.008308878168463707, + 0.010018130764365196, + 0.005998990964144468, + 0.009821927174925804, + 0.008265023119747639, + 0.0045923893339931965, + 0.010498113930225372, + 0.008299135603010654, + 0.007379195652902126, + 0.008767103776335716, + 0.008230173029005527, + 0.015557580627501011, + 0.014517592266201973, + 0.012369243428111076, + 0.012686296366155148, + 0.013431311585009098, + 0.011927440762519836, + 0.010754522867500782, + 0.007567797787487507, + 0.012936718761920929, + 0.013967787846922874, + 0.007997180335223675, + 0.01467263326048851, + 0.015189099125564098, + 0.013825574889779091, + 0.012213028036057949, + 0.005966687574982643, + 0.011765131726861, + 0.01003684289753437, + 0.00866043008863926, + 0.01348347682505846, + 0.013417284935712814, + 0.013566975481808186, + 0.010667195543646812, + 0.01561910193413496, + 0.02168823964893818, + 0.022730370983481407, + 0.017815763130784035, + 0.02401009015738964, + 0.025604700669646263, + 0.022671492770314217, + 0.021407632157206535, + 0.005764028523117304, + 0.008107862435281277, + 0.008481541648507118, + 0.006534318905323744, + 0.006714303977787495, + 0.006962523330003023, + 0.008564173243939877, + 0.006150109227746725, + 0.005277804099023342, + 0.010397756472229958, + 0.006894465070217848, + 0.005445366725325584, + 0.011737349443137646, + 0.008388492278754711, + 0.01056956872344017, + 0.007106533274054527 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 7, + 4 + ], + "selected_reward": -0.005264822859317064 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme macro view focusing on the internal structures of the human eye?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the central surface pearly, slightly yellowish, and damp with a sheen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are crimson veins visible branching like a leafless tree with some dense protruding vessels and hair-like capillaries?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the veins exhibit a subtle, rhythmic pulsation expanding and contracting in a steady biological rhythm?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the edge of a dark brown iris with an intricate fibrous pattern visible on one side of the frame?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.22562232613563538 + ], + "opposite_logprob_tokens": [ + -1.600622296333313 + ], + "logprob_reward": -0.22562232613563538, + "yes_no_margin_reward": 1.375, + "token_credit_reward": -0.22562232613563538, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.26984095573425293, + 0.11279366165399551, + 0.21319885551929474, + 0.030347518622875214, + 0.04918351024389267, + 0.106172114610672, + 0.08999047428369522, + 0.0636378601193428, + 0.03934439644217491, + 0.025490617379546165 + ], + "credit_map": [ + 0.024177344515919685, + 0.0428931787610054, + 0.033267803490161896, + 0.024995550513267517, + 0.04512235149741173, + 0.04347052052617073, + 0.03115365467965603, + 0.0247605349868536, + 0.01161184161901474, + 0.012788426131010056, + 0.018668124452233315, + 0.010617099702358246, + 0.014662038534879684, + 0.015546195209026337, + 0.01471368595957756, + 0.014186253771185875, + 0.01641637273132801, + 0.023430870845913887, + 0.0410587340593338, + 0.02169576846063137, + 0.02567317895591259, + 0.030920377001166344, + 0.026507876813411713, + 0.027495674788951874, + 0.002693213988095522, + 0.003927386365830898, + 0.00331879872828722, + 0.0022496399469673634, + 0.004183880053460598, + 0.005275297444313765, + 0.0036872676573693752, + 0.0050120363011956215, + 0.0043634274043142796, + 0.0062059820629656315, + 0.0071009485982358456, + 0.005990819074213505, + 0.006586975883692503, + 0.005469117779284716, + 0.006182643119245768, + 0.007283602375537157, + 0.0071496213786304, + 0.016191070899367332, + 0.014832885935902596, + 0.0105651980265975, + 0.01558849960565567, + 0.015807636082172394, + 0.013345534913241863, + 0.012691662646830082, + 0.007409033831208944, + 0.013837278820574284, + 0.013151837512850761, + 0.01010254304856062, + 0.013127408921718597, + 0.012628138065338135, + 0.009574570693075657, + 0.010159659199416637, + 0.0044955299235880375, + 0.007749422453343868, + 0.0058637456968426704, + 0.0063134185038506985, + 0.010192138142883778, + 0.008686208166182041, + 0.009539653547108173, + 0.010797739028930664, + 0.0036661855410784483, + 0.004201902542263269, + 0.006536319851875305, + 0.0035949628800153732, + 0.005300691816955805, + 0.0057334075681865215, + 0.0043457564897835255, + 0.005965170916169882, + 0.002240249188616872, + 0.0037186848931014538, + 0.0028146638069301844, + 0.0027373512275516987, + 0.0030595494899898767, + 0.0048801410011947155, + 0.0031937570311129093, + 0.002846218878403306 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 1 + ], + "selected_reward": -0.22562232613563538 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme macro view focusing on the internal structures of the human eye?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the central surface pearly, slightly yellowish, and damp with a sheen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are crimson veins visible branching like a leafless tree with some dense protruding vessels and hair-like capillaries?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the veins exhibit a subtle, rhythmic pulsation expanding and contracting in a steady biological rhythm?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the edge of a dark brown iris with an intricate fibrous pattern visible on one side of the frame?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04306020215153694 + ], + "opposite_logprob_tokens": [ + -3.168060302734375 + ], + "logprob_reward": -0.04306020215153694, + "yes_no_margin_reward": 3.125, + "token_credit_reward": -0.04306020215153694, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10578754544258118, + 0.10850363224744797, + 0.10474174469709396, + 0.10097049921751022, + 0.10553932189941406, + 0.11784311383962631, + 0.027054082602262497, + 0.26177626848220825, + 0.028767507523298264, + 0.03901630640029907 + ], + "credit_map": [ + 0.00837719440460205, + 0.014773473143577576, + 0.018120599910616875, + 0.009111459366977215, + 0.013385402038693428, + 0.014265037141740322, + 0.015987643972039223, + 0.01176672987639904, + 0.009259037673473358, + 0.011855179443955421, + 0.02550681121647358, + 0.01181500218808651, + 0.01166091114282608, + 0.013256020843982697, + 0.013613102026283741, + 0.011537556536495686, + 0.010536329820752144, + 0.013255886733531952, + 0.017551211640238762, + 0.011412377469241619, + 0.012325968593358994, + 0.013231712393462658, + 0.01565476320683956, + 0.010773482732474804, + 0.008595815859735012, + 0.013644755817949772, + 0.019096262753009796, + 0.009903903119266033, + 0.01177369523793459, + 0.01177034992724657, + 0.01306179165840149, + 0.01312391459941864, + 0.009191781282424927, + 0.011796805076301098, + 0.018382027745246887, + 0.009769462049007416, + 0.010478401556611061, + 0.016461551189422607, + 0.011850981041789055, + 0.017608296126127243, + 0.009704750962555408, + 0.013901116326451302, + 0.01985907554626465, + 0.009148657321929932, + 0.017296697944402695, + 0.01597306691110134, + 0.014371956698596478, + 0.017587779089808464, + 0.002504762727767229, + 0.0028199905063956976, + 0.004725771024823189, + 0.00226282374933362, + 0.0035897493362426758, + 0.003827218431979418, + 0.0035754318814724684, + 0.0037483321502804756, + 0.022003456950187683, + 0.03075731359422207, + 0.03367402032017708, + 0.024328473955392838, + 0.03673529997467995, + 0.04065365344285965, + 0.03537510707974434, + 0.0382489375770092, + 0.0022315741516649723, + 0.005565367639064789, + 0.0039046211168169975, + 0.002810729667544365, + 0.0045320275239646435, + 0.0036868113093078136, + 0.0032441981602460146, + 0.0027921758592128754, + 0.004073139280080795, + 0.0060461158864200115, + 0.006666051689535379, + 0.0028816862031817436, + 0.006121970247477293, + 0.00575181795284152, + 0.004226239398121834, + 0.003249281784519553 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 5, + 1 + ], + "selected_reward": -0.04306020215153694 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000006/group_6/index_tensor([7633, 7633])_group_6_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000006/group_6/index_tensor([7633, 7633])_group_6_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..f01a18c217f7211764c5d1b9d855e4638aa62ee2 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000006/group_6/index_tensor([7633, 7633])_group_6_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "powerful overhead shot, the camera fixed at an extremely low angle to the ground, swaying violently as the earth trembles violently, reveals a mountain-sized beast slowly and ponderously rolling past. The entire scene is shrouded in a stark contrast of black and white. A harsh, low-angled light sweeps from the side, carving every deep groove in the beast's rocky hide into a pitch-black shadow, while its ridges of muscle and bone appear a brilliant, almost blown-out white. Each step of its towering legs sends the camera careening with a devastating jolt, sending countless rubble and dust from the ground swirling in a chaotic, swirling cloud of smoke in the light beam. The viewer's gaze is completely absorbed by a portion of its massive body, unable to see its full extent, only its unparalleled mass and power. The entire image is imbued with a sense of primal, untamed power, creating a sense of apocalyptic, suffocating oppression and epic grandeur.", + "seed": 12738, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000006/group_6/index_7633_seed_12738_dp_12_rank_24_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 6, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.24124805848114192, + "vlm_logprob": -0.24124805848114192, + "vlm_yes_no_margin": 3.449999988079071, + "vlm_token_credit": -0.24124805848114192, + "vlm_yes_ratio": 0.8, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene presented in a stark black and white contrast?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera fixed at an extremely low angle to the ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a mountain-sized beast with rocky hide visible rolling past the camera?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a harsh low-angled light sweep from the side creating pitch-black shadows and blown-out white highlights?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera sway violently and jolt as the beast steps, causing rubble and dust to swirl in a chaotic cloud?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04272375628352165 + ], + "opposite_logprob_tokens": [ + -3.2927236557006836 + ], + "logprob_reward": -0.04272375628352165, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.04272375628352165, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3210868537425995, + 0.02168484590947628, + 0.03201253339648247, + 0.0275268517434597, + 0.25202471017837524, + 0.025695258751511574, + 0.19253870844841003, + 0.08488833159208298, + 0.017631445080041885, + 0.024910513311624527 + ], + "credit_map": [ + 0.05074135214090347, + 0.02201487496495247, + 0.016423065215349197, + 0.020535124465823174, + 0.020993269979953766, + 0.11429246515035629, + 0.015612923540174961, + 0.060473743826150894, + 0.003134088357910514, + 0.003201881190761924, + 0.0025241978000849485, + 0.0027817741502076387, + 0.0024850491899996996, + 0.003686582436785102, + 0.002315386664122343, + 0.0015558861196041107, + 0.005822712555527687, + 0.006252117920666933, + 0.003071992192417383, + 0.0029147458262741566, + 0.0035337714944034815, + 0.0026176003739237785, + 0.005242674611508846, + 0.0025569198187440634, + 0.004038685001432896, + 0.0040380582213401794, + 0.0040342011488974094, + 0.0032243235036730766, + 0.003298157826066017, + 0.0039955019019544125, + 0.002686196006834507, + 0.002211727434769273, + 0.03932680934667587, + 0.022027259692549706, + 0.038221970200538635, + 0.030582761391997337, + 0.0269906148314476, + 0.023996839299798012, + 0.02701425552368164, + 0.04386419430375099, + 0.0024441308341920376, + 0.003182424930855632, + 0.0029083585832268, + 0.004157824441790581, + 0.003172490047290921, + 0.003497946774587035, + 0.0036905480083078146, + 0.002641534199938178, + 0.014523726888000965, + 0.038296427577733994, + 0.012173837050795555, + 0.0702022984623909, + 0.007926723919808865, + 0.009649024344980717, + 0.035219959914684296, + 0.004546689335256815, + 0.007833059877157211, + 0.011706243269145489, + 0.012275285087525845, + 0.01785299740731716, + 0.006694986019283533, + 0.008577660657465458, + 0.012568783015012741, + 0.007379309739917517, + 0.0025526925455778837, + 0.001254658098332584, + 0.0023735666181892157, + 0.0034167433623224497, + 0.0018983271438628435, + 0.002379803918302059, + 0.0021613535936921835, + 0.001594299334101379, + 0.0028120973147451878, + 0.003959239460527897, + 0.0039005812723189592, + 0.004255872219800949, + 0.003172531258314848, + 0.0028677373193204403, + 0.0019128081621602178, + 0.0020296480506658554 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 6 + ], + "selected_reward": -0.04272375628352165 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene presented in a stark black and white contrast?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera fixed at an extremely low angle to the ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a mountain-sized beast with rocky hide visible rolling past the camera?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a harsh low-angled light sweep from the side creating pitch-black shadows and blown-out white highlights?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera sway violently and jolt as the beast steps, causing rubble and dust to swirl in a chaotic cloud?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -1.1370625495910645 + ], + "opposite_logprob_tokens": [ + -0.3870624899864197 + ], + "logprob_reward": -1.1370625495910645, + "yes_no_margin_reward": -0.7500000596046448, + "token_credit_reward": -1.1370625495910645, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.47876253724098206, + 0.030248647555708885, + 0.06195289269089699, + 0.030800610780715942, + 0.08691234886646271, + 0.029738547280430794, + 0.13541284203529358, + 0.057066138833761215, + 0.0437653511762619, + 0.04534008726477623 + ], + "credit_map": [ + 0.08297083526849747, + 0.03396078199148178, + 0.01688539795577526, + 0.025648603215813637, + 0.02971258945763111, + 0.18058711290359497, + 0.01536720059812069, + 0.09362999349832535, + 0.004918256308883429, + 0.004535626620054245, + 0.005151855293661356, + 0.004452591761946678, + 0.0029330849647521973, + 0.0021752102766186, + 0.004107085522264242, + 0.001974936109036207, + 0.012054681777954102, + 0.011871504597365856, + 0.007756921462714672, + 0.006670450326055288, + 0.006311587058007717, + 0.006465368904173374, + 0.005966669414192438, + 0.00485570915043354, + 0.0038387165404856205, + 0.004443434067070484, + 0.003516139928251505, + 0.003199273720383644, + 0.003980362322181463, + 0.0036748466081917286, + 0.003408560762181878, + 0.004739278927445412, + 0.012718056328594685, + 0.006857593543827534, + 0.012843349017202854, + 0.01140531711280346, + 0.010606914758682251, + 0.008607746101915836, + 0.009167129173874855, + 0.01470625028014183, + 0.002355020958930254, + 0.0044565522111952305, + 0.003740560030564666, + 0.005131208803504705, + 0.0047533647157251835, + 0.0029357948806136847, + 0.003197871381416917, + 0.0031681761611253023, + 0.010183523409068584, + 0.025401966646313667, + 0.008406577631831169, + 0.0489424504339695, + 0.005975555628538132, + 0.006573478225618601, + 0.026432853192090988, + 0.0034964289516210556, + 0.005368646699935198, + 0.00871887244284153, + 0.009078392758965492, + 0.008040896616876125, + 0.005138967651873827, + 0.006720567587763071, + 0.0062361727468669415, + 0.007763626519590616, + 0.0033937750849872828, + 0.004185547586530447, + 0.004613860510289669, + 0.005654656793922186, + 0.010962328873574734, + 0.005238854326307774, + 0.005811681970953941, + 0.0039046418387442827, + 0.004262719303369522, + 0.0032651948276907206, + 0.011990262195467949, + 0.004648458678275347, + 0.004625079222023487, + 0.007478635758161545, + 0.0050170267932116985, + 0.004052709322422743 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 6, + 4 + ], + "selected_reward": -1.1370625495910645 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene presented in a stark black and white contrast?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera fixed at an extremely low angle to the ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a mountain-sized beast with rocky hide visible rolling past the camera?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a harsh low-angled light sweep from the side creating pitch-black shadows and blown-out white highlights?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera sway violently and jolt as the beast steps, causing rubble and dust to swirl in a chaotic cloud?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.002820563269779086 + ], + "opposite_logprob_tokens": [ + -5.8778204917907715 + ], + "logprob_reward": -0.002820563269779086, + "yes_no_margin_reward": 5.875, + "token_credit_reward": -0.002820563269779086, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.264659583568573, + 0.03258334845304489, + 0.05554893985390663, + 0.02348621003329754, + 0.07831830531358719, + 0.08325182646512985, + 0.2958521544933319, + 0.07684915512800217, + 0.03592517599463463, + 0.05352521687746048 + ], + "credit_map": [ + 0.04313051700592041, + 0.019095391035079956, + 0.013404673896729946, + 0.016856443136930466, + 0.015010048635303974, + 0.09635947644710541, + 0.01184089295566082, + 0.048962149769067764, + 0.004725639708340168, + 0.005908609833568335, + 0.0038674643728882074, + 0.0038811771664768457, + 0.0031741380225867033, + 0.0023419889621436596, + 0.005538376048207283, + 0.003145957598462701, + 0.008579968474805355, + 0.010980106890201569, + 0.007166014518588781, + 0.004921553190797567, + 0.00983965303748846, + 0.004484761040657759, + 0.004920716863125563, + 0.004656168632209301, + 0.0029829349368810654, + 0.0036028381437063217, + 0.0030595650896430016, + 0.0022739949636161327, + 0.00385070382617414, + 0.0037619243375957012, + 0.0018743057735264301, + 0.0020799452904611826, + 0.012581286020576954, + 0.00867030955851078, + 0.011225249618291855, + 0.009477779269218445, + 0.007884438149631023, + 0.006823553238064051, + 0.008739706128835678, + 0.01291598379611969, + 0.007683670613914728, + 0.014866665937006474, + 0.008949213661253452, + 0.01547005120664835, + 0.011241423897445202, + 0.0083989929407835, + 0.009108610451221466, + 0.0075332047417759895, + 0.02102472260594368, + 0.05327572673559189, + 0.01797584630548954, + 0.10930080711841583, + 0.012298927642405033, + 0.013851548545062542, + 0.06095878407359123, + 0.0071658301167190075, + 0.00655078561976552, + 0.007180829532444477, + 0.011753799393773079, + 0.014570636674761772, + 0.0064004287123680115, + 0.0101622911170125, + 0.00887740682810545, + 0.01135298516601324, + 0.005196926183998585, + 0.002990077016875148, + 0.005256094038486481, + 0.004877551458775997, + 0.00527863809838891, + 0.004304472357034683, + 0.003850536420941353, + 0.004170883446931839, + 0.004921354353427887, + 0.007884673774242401, + 0.011191660538315773, + 0.006149529945105314, + 0.006898515857756138, + 0.00733467610552907, + 0.004958512727171183, + 0.004186298698186874 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 0, + 5 + ], + "selected_reward": -0.002820563269779086 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene presented in a stark black and white contrast?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera fixed at an extremely low angle to the ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a mountain-sized beast with rocky hide visible rolling past the camera?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a harsh low-angled light sweep from the side creating pitch-black shadows and blown-out white highlights?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera sway violently and jolt as the beast steps, causing rubble and dust to swirl in a chaotic cloud?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.011088181287050247 + ], + "opposite_logprob_tokens": [ + -4.5110883712768555 + ], + "logprob_reward": -0.011088181287050247, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.011088181287050247, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3172999322414398, + 0.07120346277952194, + 0.07436147332191467, + 0.04634903743863106, + 0.06678774952888489, + 0.025323571637272835, + 0.2317744493484497, + 0.07506745308637619, + 0.050028104335069656, + 0.04180474951863289 + ], + "credit_map": [ + 0.05276684835553169, + 0.02092362940311432, + 0.013250118121504784, + 0.017493147403001785, + 0.01874157227575779, + 0.11903522163629532, + 0.013260980136692524, + 0.061828408390283585, + 0.010693463496863842, + 0.010340576991438866, + 0.009545684792101383, + 0.009702830575406551, + 0.006075193174183369, + 0.0060723102651536465, + 0.012822332791984081, + 0.005951072555035353, + 0.013241345062851906, + 0.014082292094826698, + 0.009322044439613819, + 0.008746902458369732, + 0.008329087868332863, + 0.007888533174991608, + 0.007052213419228792, + 0.00569906085729599, + 0.006898507010191679, + 0.005653932690620422, + 0.006353473756462336, + 0.006043917965143919, + 0.00546882301568985, + 0.007094913627952337, + 0.005064701195806265, + 0.0037707677111029625, + 0.010676831938326359, + 0.0064377314411103725, + 0.010606803931295872, + 0.006263480521738529, + 0.007612934336066246, + 0.006223075557500124, + 0.008184214122593403, + 0.010782673954963684, + 0.00256490265019238, + 0.003130426863208413, + 0.0030198819003999233, + 0.004598635248839855, + 0.003048940096050501, + 0.003411287674680352, + 0.003506786422803998, + 0.002042712178081274, + 0.015678659081459045, + 0.04035034030675888, + 0.013819881714880466, + 0.0867089256644249, + 0.009472486563026905, + 0.010769369080662727, + 0.04938306659460068, + 0.005591708701103926, + 0.006639536004513502, + 0.014353934675455093, + 0.010353600606322289, + 0.013637631200253963, + 0.006512884516268969, + 0.00726955384016037, + 0.010041091591119766, + 0.00625921506434679, + 0.00502591161057353, + 0.0047996933571994305, + 0.007087557576596737, + 0.006742906291037798, + 0.00630749948322773, + 0.006439782679080963, + 0.008290682919323444, + 0.005334070418030024, + 0.0046746679581701756, + 0.0037524413783103228, + 0.008267341181635857, + 0.006133877672255039, + 0.004447214305400848, + 0.004766612313687801, + 0.005931499879807234, + 0.0038310966920107603 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 6, + 7 + ], + "selected_reward": -0.011088181287050247 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene presented in a stark black and white contrast?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera fixed at an extremely low angle to the ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a mountain-sized beast with rocky hide visible rolling past the camera?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a harsh low-angled light sweep from the side creating pitch-black shadows and blown-out white highlights?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera sway violently and jolt as the beast steps, causing rubble and dust to swirl in a chaotic cloud?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.012545241974294186 + ], + "opposite_logprob_tokens": [ + -4.387545108795166 + ], + "logprob_reward": -0.012545241974294186, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.012545241974294186, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.416776567697525, + 0.028521519154310226, + 0.053743746131658554, + 0.0531056746840477, + 0.04658791422843933, + 0.04409930109977722, + 0.11070679128170013, + 0.1218743845820427, + 0.07280575484037399, + 0.05177835747599602 + ], + "credit_map": [ + 0.07232124358415604, + 0.029871806502342224, + 0.017030643299221992, + 0.022731294855475426, + 0.024811992421746254, + 0.15657128393650055, + 0.013148722238838673, + 0.08028959482908249, + 0.0046142167411744595, + 0.004621956497430801, + 0.004639703314751387, + 0.004320433828979731, + 0.00269387848675251, + 0.00206292443908751, + 0.0036081066355109215, + 0.001960298279300332, + 0.01068192906677723, + 0.011844680644571781, + 0.005532254930585623, + 0.004890559706836939, + 0.006197100970894098, + 0.0046471706591546535, + 0.005324732977896929, + 0.004625316243618727, + 0.007352198474109173, + 0.008109071291983128, + 0.006802637130022049, + 0.006907513365149498, + 0.006439062766730785, + 0.005540120881050825, + 0.006737191695719957, + 0.00521787628531456, + 0.008284744806587696, + 0.003716828767210245, + 0.005104255862534046, + 0.005055882036685944, + 0.004564235452562571, + 0.004034114070236683, + 0.0060719093307852745, + 0.009755945764482021, + 0.004467795602977276, + 0.007253273390233517, + 0.004638431593775749, + 0.008582502603530884, + 0.0055734640918672085, + 0.004369422793388367, + 0.005213741213083267, + 0.004000670742243528, + 0.008348227478563786, + 0.020455513149499893, + 0.007395279593765736, + 0.039370346814394, + 0.005399268586188555, + 0.005841306876391172, + 0.0206746868789196, + 0.0032221602741628885, + 0.011064247228205204, + 0.012431832030415535, + 0.018152402713894844, + 0.025289852172136307, + 0.009465977549552917, + 0.01494403276592493, + 0.014147256501019001, + 0.01637878268957138, + 0.0069859386421740055, + 0.006794363725930452, + 0.010156645439565182, + 0.010619232431054115, + 0.006239474285393953, + 0.00997867714613676, + 0.010167050175368786, + 0.011864373460412025, + 0.006564343348145485, + 0.010596058331429958, + 0.005881180055439472, + 0.007404342759400606, + 0.007552123162895441, + 0.005051754415035248, + 0.005427902098745108, + 0.003300652140751481 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 7, + 6 + ], + "selected_reward": -0.012545241974294186 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "No", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "powerful overhead shot, the camera fixed at an extremely low angle to the ground, swaying violently as the earth trembles violently, reveals a mountain-sized beast slowly and ponderously rolling past. The entire scene is shrouded in a stark contrast of black and white. A harsh, low-angled light sweeps from the side, carving every deep groove in the beast's rocky hide into a pitch-black shadow, while its ridges of muscle and bone appear a brilliant, almost blown-out white. Each step of its towering legs sends the camera careening with a devastating jolt, sending countless rubble and dust from the ground swirling in a chaotic, swirling cloud of smoke in the light beam. The viewer's gaze is completely absorbed by a portion of its massive body, unable to see its full extent, only its unparalleled mass and power. The entire image is imbued with a sense of primal, untamed power, creating a sense of apocalyptic, suffocating oppression and epic grandeur.", + "seed": 12739, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000006/group_6/index_7633_seed_12739_dp_12_rank_24_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 6, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.1338301807641983, + "vlm_logprob": -0.1338301807641983, + "vlm_yes_no_margin": 3.475, + "vlm_token_credit": -0.1338301807641983, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene presented in a stark black and white contrast?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera fixed at an extremely low angle to the ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a mountain-sized beast with rocky hide visible rolling past the camera?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a harsh low-angled light sweep from the side creating pitch-black shadows and blown-out white highlights?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera sway violently and jolt as the beast steps, causing rubble and dust to swirl in a chaotic cloud?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.06609684228897095 + ], + "opposite_logprob_tokens": [ + -2.816096782684326 + ], + "logprob_reward": -0.06609684228897095, + "yes_no_margin_reward": 2.75, + "token_credit_reward": -0.06609684228897095, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.33798712491989136, + 0.11441231518983841, + 0.0848049521446228, + 0.05208883434534073, + 0.07466202974319458, + 0.046245235949754715, + 0.04682057350873947, + 0.03737584128975868, + 0.0741521418094635, + 0.13145090639591217 + ], + "credit_map": [ + 0.021846702322363853, + 0.028901848942041397, + 0.02817424014210701, + 0.03428926691412926, + 0.038826875388622284, + 0.04727713763713837, + 0.08398078382015228, + 0.054690275341272354, + 0.013818125240504742, + 0.012038475833833218, + 0.020396534353494644, + 0.01837097853422165, + 0.011447116732597351, + 0.015004701912403107, + 0.010007055476307869, + 0.013329320587217808, + 0.009513500146567822, + 0.013669470325112343, + 0.010099018923938274, + 0.010347348637878895, + 0.009514903649687767, + 0.0181503277271986, + 0.00597543315961957, + 0.007534951437264681, + 0.006288918666541576, + 0.006639121565967798, + 0.006472384091466665, + 0.007721067406237125, + 0.0052091628313064575, + 0.006012761499732733, + 0.008867200464010239, + 0.004878216423094273, + 0.007627116050571203, + 0.008756617084145546, + 0.008870013989508152, + 0.010005510412156582, + 0.010555285029113293, + 0.007195968180894852, + 0.015387149527668953, + 0.006264363415539265, + 0.0060788728296756744, + 0.007627489045262337, + 0.0037127481773495674, + 0.0061012860387563705, + 0.004813832230865955, + 0.00669952854514122, + 0.004456392489373684, + 0.006755087058991194, + 0.0047653052024543285, + 0.005365816410630941, + 0.0070382715202867985, + 0.009770272299647331, + 0.0048789381980896, + 0.003901461139321327, + 0.005906497593969107, + 0.005194014869630337, + 0.008899472653865814, + 0.005046808626502752, + 0.009763777256011963, + 0.002680366626009345, + 0.0028277626261115074, + 0.0034580554347485304, + 0.002056973986327648, + 0.002642625942826271, + 0.009526200592517853, + 0.0111362524330616, + 0.014663079753518105, + 0.006849455647170544, + 0.006840583868324757, + 0.014082271605730057, + 0.005453674588352442, + 0.005600630305707455, + 0.014989547431468964, + 0.023754864931106567, + 0.03166645020246506, + 0.013629695400595665, + 0.008474588394165039, + 0.020414240658283234, + 0.00796984788030386, + 0.010551674291491508 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 1 + ], + "selected_reward": -0.06609684228897095 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene presented in a stark black and white contrast?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera fixed at an extremely low angle to the ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a mountain-sized beast with rocky hide visible rolling past the camera?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a harsh low-angled light sweep from the side creating pitch-black shadows and blown-out white highlights?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera sway violently and jolt as the beast steps, causing rubble and dust to swirl in a chaotic cloud?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.5761573314666748 + ], + "opposite_logprob_tokens": [ + -0.8261573314666748 + ], + "logprob_reward": -0.5761573314666748, + "yes_no_margin_reward": 0.25, + "token_credit_reward": -0.5761573314666748, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15525095164775848, + 0.13133692741394043, + 0.13442325592041016, + 0.0668071061372757, + 0.12881967425346375, + 0.0755055621266365, + 0.034456655383110046, + 0.11523705720901489, + 0.07301561534404755, + 0.08514722436666489 + ], + "credit_map": [ + 0.011623316444456577, + 0.016879713162779808, + 0.014394275844097137, + 0.01464283000677824, + 0.020818209275603294, + 0.02133084647357464, + 0.03281304985284805, + 0.022748716175556183, + 0.02133120782673359, + 0.01747989095747471, + 0.013725544326007366, + 0.011308461427688599, + 0.01786777190864086, + 0.021202290430665016, + 0.015027591958642006, + 0.013394155539572239, + 0.011250299401581287, + 0.014543584547936916, + 0.029719563201069832, + 0.02339717373251915, + 0.014142573811113834, + 0.014521593227982521, + 0.012855715118348598, + 0.013992740772664547, + 0.007855227217078209, + 0.010987195186316967, + 0.009938057512044907, + 0.009549573995172977, + 0.007491424214094877, + 0.006664292886853218, + 0.008993326686322689, + 0.005328014958649874, + 0.01291781198233366, + 0.014336843974888325, + 0.015395060181617737, + 0.014832606539130211, + 0.019748369231820107, + 0.012718490324914455, + 0.02886572666466236, + 0.010004767216742039, + 0.010978050529956818, + 0.014426206238567829, + 0.006803466007113457, + 0.009561587125062943, + 0.008804162964224815, + 0.007815657183527946, + 0.007604438345879316, + 0.009511993266642094, + 0.0038758849259465933, + 0.004417958203703165, + 0.003146257484331727, + 0.0032922369427978992, + 0.004296043887734413, + 0.005771016702055931, + 0.0038561217952519655, + 0.005801134742796421, + 0.017243467271327972, + 0.01066363975405693, + 0.01472944300621748, + 0.009590874426066875, + 0.025122832506895065, + 0.010974377393722534, + 0.018582776188850403, + 0.00832964014261961, + 0.008655883371829987, + 0.009106701239943504, + 0.013560357503592968, + 0.010167362168431282, + 0.007028303109109402, + 0.010279710404574871, + 0.0049162860959768295, + 0.009301004000008106, + 0.009320047684013844, + 0.014777247793972492, + 0.01912144012749195, + 0.00960028637200594, + 0.006711395923048258, + 0.012347606010735035, + 0.006412281654775143, + 0.006856915075331926 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 1 + ], + "selected_reward": -0.5761573314666748 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene presented in a stark black and white contrast?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera fixed at an extremely low angle to the ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a mountain-sized beast with rocky hide visible rolling past the camera?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a harsh low-angled light sweep from the side creating pitch-black shadows and blown-out white highlights?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera sway violently and jolt as the beast steps, causing rubble and dust to swirl in a chaotic cloud?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005951184779405594 + ], + "opposite_logprob_tokens": [ + -5.130951404571533 + ], + "logprob_reward": -0.005951184779405594, + "yes_no_margin_reward": 5.125, + "token_credit_reward": -0.005951184779405594, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12009217590093613, + 0.24306069314479828, + 0.07241913676261902, + 0.09098813682794571, + 0.10747504234313965, + 0.11340795457363129, + 0.027929706498980522, + 0.033899471163749695, + 0.0813138484954834, + 0.10941378772258759 + ], + "credit_map": [ + 0.008671856485307217, + 0.01193020585924387, + 0.013924813829362392, + 0.014982196502387524, + 0.010426490567624569, + 0.015715325251221657, + 0.027490345761179924, + 0.01695094257593155, + 0.03336767107248306, + 0.027036147192120552, + 0.025225035846233368, + 0.02192661166191101, + 0.03689500316977501, + 0.04983511194586754, + 0.021736275404691696, + 0.027038879692554474, + 0.006071890238672495, + 0.006673200521618128, + 0.015067502856254578, + 0.011867200955748558, + 0.007375634741038084, + 0.012551160529255867, + 0.0060691796243190765, + 0.0067433686926960945, + 0.011048712767660618, + 0.011794058606028557, + 0.011745316907763481, + 0.012120661325752735, + 0.009666758589446545, + 0.009677309542894363, + 0.016548536717891693, + 0.008386790752410889, + 0.010255604051053524, + 0.011424580588936806, + 0.010112697258591652, + 0.013490802608430386, + 0.018045563250780106, + 0.008342274464666843, + 0.026714779436588287, + 0.009088746272027493, + 0.01483170036226511, + 0.02247934229671955, + 0.009400722570717335, + 0.014558383263647556, + 0.011965665966272354, + 0.013394666835665703, + 0.00995598640292883, + 0.016821499913930893, + 0.002709747292101383, + 0.002997106174007058, + 0.003911673557013273, + 0.0058758133091032505, + 0.003054465400055051, + 0.0027176279108971357, + 0.003376496024429798, + 0.003286781022325158, + 0.0048288870602846146, + 0.003194148652255535, + 0.004748850595206022, + 0.002219890709966421, + 0.0063925934955477715, + 0.005083722062408924, + 0.004285254515707493, + 0.003146127099171281, + 0.009637763723731041, + 0.012600667774677277, + 0.01396915502846241, + 0.008025825954973698, + 0.0076991175301373005, + 0.014364280737936497, + 0.006605261471122503, + 0.008411779068410397, + 0.011989576742053032, + 0.018769586458802223, + 0.01718578115105629, + 0.01157748606055975, + 0.011093610897660255, + 0.019580107182264328, + 0.008475048467516899, + 0.010742594487965107 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 5 + ], + "selected_reward": -0.005951184779405594 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene presented in a stark black and white contrast?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera fixed at an extremely low angle to the ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a mountain-sized beast with rocky hide visible rolling past the camera?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a harsh low-angled light sweep from the side creating pitch-black shadows and blown-out white highlights?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera sway violently and jolt as the beast steps, causing rubble and dust to swirl in a chaotic cloud?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.006745661608874798 + ], + "opposite_logprob_tokens": [ + -5.0067458152771 + ], + "logprob_reward": -0.006745661608874798, + "yes_no_margin_reward": 5.0, + "token_credit_reward": -0.006745661608874798, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.30402877926826477, + 0.2043481320142746, + 0.10383476316928864, + 0.06075162813067436, + 0.03648969531059265, + 0.09886085242033005, + 0.025075815618038177, + 0.06713561713695526, + 0.03655422851443291, + 0.06292048841714859 + ], + "credit_map": [ + 0.020770389586687088, + 0.03155115619301796, + 0.022431934252381325, + 0.025541020557284355, + 0.040997762233018875, + 0.04208006709814072, + 0.06912288814783096, + 0.05153356492519379, + 0.026981104165315628, + 0.02168060466647148, + 0.018683377653360367, + 0.030282167717814445, + 0.03162628412246704, + 0.03373612463474274, + 0.017826233059167862, + 0.02353222295641899, + 0.011802881956100464, + 0.017784859985113144, + 0.012459099292755127, + 0.0126125393435359, + 0.0130610391497612, + 0.01997384801506996, + 0.0086201261729002, + 0.007520370185375214, + 0.007084602490067482, + 0.009413032792508602, + 0.007249585352838039, + 0.00959588773548603, + 0.006754899397492409, + 0.006784405559301376, + 0.008939295075833797, + 0.004929916933178902, + 0.003597215050831437, + 0.0035931007005274296, + 0.0033386603463441133, + 0.004852269776165485, + 0.0054472400806844234, + 0.0034687549341470003, + 0.007135656196624041, + 0.005056797526776791, + 0.013452926650643349, + 0.018679169937968254, + 0.008381422609090805, + 0.01269101444631815, + 0.010533184744417667, + 0.012243032455444336, + 0.008617250248789787, + 0.014262854121625423, + 0.002426349790766835, + 0.0032807989045977592, + 0.0032142314594238997, + 0.004785267170518637, + 0.002669682726264, + 0.003017136361449957, + 0.002634588396176696, + 0.0030477617401629686, + 0.018152382224798203, + 0.006726042367517948, + 0.020787497982382774, + 0.00397108681499958, + 0.005289963446557522, + 0.004787790589034557, + 0.0037331371568143368, + 0.003687717020511627, + 0.004716475959867239, + 0.004782503936439753, + 0.007987926714122295, + 0.0032496757339686155, + 0.0030301196966320276, + 0.006431146524846554, + 0.0026086021680384874, + 0.0037477819714695215, + 0.0066384561359882355, + 0.009009363129734993, + 0.01121810358017683, + 0.006973418407142162, + 0.006747501902282238, + 0.010055698454380035, + 0.004817926324903965, + 0.007460025139153004 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 2 + ], + "selected_reward": -0.006745661608874798 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene presented in a stark black and white contrast?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera fixed at an extremely low angle to the ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a mountain-sized beast with rocky hide visible rolling past the camera?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a harsh low-angled light sweep from the side creating pitch-black shadows and blown-out white highlights?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera sway violently and jolt as the beast steps, causing rubble and dust to swirl in a chaotic cloud?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.014199883677065372 + ], + "opposite_logprob_tokens": [ + -4.264199733734131 + ], + "logprob_reward": -0.014199883677065372, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.014199883677065372, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12181097269058228, + 0.09195459634065628, + 0.09469760954380035, + 0.1991720199584961, + 0.05763690546154976, + 0.14538869261741638, + 0.05839649960398674, + 0.10566432029008865, + 0.06619483232498169, + 0.05908359959721565 + ], + "credit_map": [ + 0.009313582442700863, + 0.014593185856938362, + 0.012692119926214218, + 0.011921195313334465, + 0.010724183171987534, + 0.015996277332305908, + 0.026213256642222404, + 0.020357176661491394, + 0.011134815402328968, + 0.007318081334233284, + 0.014955918304622173, + 0.00918209832161665, + 0.016468096524477005, + 0.013294396921992302, + 0.010797547176480293, + 0.008803640492260456, + 0.008117470890283585, + 0.014413201250135899, + 0.014746362343430519, + 0.011614483781158924, + 0.012778948061168194, + 0.013863460160791874, + 0.00865680817514658, + 0.01050687488168478, + 0.019977379590272903, + 0.029405774548649788, + 0.029183363541960716, + 0.03271295502781868, + 0.019769921898841858, + 0.021420804783701897, + 0.03025124967098236, + 0.01645057275891304, + 0.00571422278881073, + 0.006166908424347639, + 0.006279720459133387, + 0.006978873163461685, + 0.006610436365008354, + 0.006166758481413126, + 0.009581179358065128, + 0.010138804093003273, + 0.018912531435489655, + 0.02599242702126503, + 0.012850746512413025, + 0.019353646785020828, + 0.015590552240610123, + 0.017720473930239677, + 0.012423286214470863, + 0.02254503034055233, + 0.0061895460821688175, + 0.007172699086368084, + 0.007421066518872976, + 0.010149359703063965, + 0.006542454939335585, + 0.008176561444997787, + 0.0057557933032512665, + 0.006989023648202419, + 0.016767751425504684, + 0.01239823643118143, + 0.018111584708094597, + 0.009079801850020885, + 0.009449219331145287, + 0.01719014346599579, + 0.010652784258127213, + 0.012014790438115597, + 0.00868241023272276, + 0.013355433009564877, + 0.014278178103268147, + 0.006359928287565708, + 0.004867688287049532, + 0.008659723214805126, + 0.0035792477428913116, + 0.006412222981452942, + 0.0069424984976649284, + 0.008806067518889904, + 0.011668635532259941, + 0.006189757492393255, + 0.006602956913411617, + 0.008990591391921043, + 0.004523065406829119, + 0.005360027775168419 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 5, + 0 + ], + "selected_reward": -0.014199883677065372 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "powerful overhead shot, the camera fixed at an extremely low angle to the ground, swaying violently as the earth trembles violently, reveals a mountain-sized beast slowly and ponderously rolling past. The entire scene is shrouded in a stark contrast of black and white. A harsh, low-angled light sweeps from the side, carving every deep groove in the beast's rocky hide into a pitch-black shadow, while its ridges of muscle and bone appear a brilliant, almost blown-out white. Each step of its towering legs sends the camera careening with a devastating jolt, sending countless rubble and dust from the ground swirling in a chaotic, swirling cloud of smoke in the light beam. The viewer's gaze is completely absorbed by a portion of its massive body, unable to see its full extent, only its unparalleled mass and power. The entire image is imbued with a sense of primal, untamed power, creating a sense of apocalyptic, suffocating oppression and epic grandeur.", + "seed": 12740, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000006/group_6/index_7633_seed_12740_dp_13_rank_26_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 6, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.10605955151841044, + "vlm_logprob": -0.10605955151841044, + "vlm_yes_no_margin": 3.75, + "vlm_token_credit": -0.10605955151841044, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene presented in a stark black and white contrast?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera fixed at an extremely low angle to the ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a mountain-sized beast with rocky hide visible rolling past the camera?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a harsh low-angled light sweep from the side creating pitch-black shadows and blown-out white highlights?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera sway violently and jolt as the beast steps, causing rubble and dust to swirl in a chaotic cloud?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.028548341244459152 + ], + "opposite_logprob_tokens": [ + -3.778548240661621 + ], + "logprob_reward": -0.028548341244459152, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.028548341244459152, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10163788497447968, + 0.08135681599378586, + 0.03204263374209404, + 0.025306876748800278, + 0.02488582581281662, + 0.2455478459596634, + 0.28603363037109375, + 0.040169425308704376, + 0.1078081950545311, + 0.05521087720990181 + ], + "credit_map": [ + 0.007943306118249893, + 0.010730006732046604, + 0.011056163348257542, + 0.014183338731527328, + 0.014666111208498478, + 0.013208470307290554, + 0.01813899725675583, + 0.011711484752595425, + 0.008281271904706955, + 0.011987273581326008, + 0.011836053803563118, + 0.014070109464228153, + 0.006003788206726313, + 0.012548673897981644, + 0.008656964637339115, + 0.007972667925059795, + 0.0032310455571860075, + 0.00452761584892869, + 0.004175633192062378, + 0.004521870519965887, + 0.0036934316158294678, + 0.003917474765330553, + 0.004145422019064426, + 0.0038301334716379642, + 0.0033197500742971897, + 0.004183930344879627, + 0.002991744317114353, + 0.003843131475150585, + 0.0019646731670945883, + 0.0014280213508754969, + 0.004989326000213623, + 0.0025862962938845158, + 0.003925972152501345, + 0.0033065122552216053, + 0.0023492518812417984, + 0.0035129389725625515, + 0.003345759119838476, + 0.003294882597401738, + 0.0026568020693957806, + 0.002493703505024314, + 0.03824750706553459, + 0.025354284793138504, + 0.020388979464769363, + 0.026190394535660744, + 0.03580864518880844, + 0.03542694076895714, + 0.028510410338640213, + 0.035620663315057755, + 0.02329336106777191, + 0.02422594279050827, + 0.02567797154188156, + 0.02660948969423771, + 0.047595907002687454, + 0.04072844982147217, + 0.035555336624383926, + 0.06234712898731232, + 0.004706160165369511, + 0.00588573981076479, + 0.0049410490319132805, + 0.0056253597140312195, + 0.005055226851254702, + 0.003952313680201769, + 0.005115213338285685, + 0.004888359922915697, + 0.016792191192507744, + 0.022767893970012665, + 0.010560481809079647, + 0.011612098664045334, + 0.010200089775025845, + 0.010804329998791218, + 0.00980635266751051, + 0.015264744870364666, + 0.0073593733832240105, + 0.010835953056812286, + 0.007411808706820011, + 0.007360560353845358, + 0.005799718201160431, + 0.005900165997445583, + 0.004581405781209469, + 0.005961887538433075 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 5, + 8 + ], + "selected_reward": -0.028548341244459152 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene presented in a stark black and white contrast?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera fixed at an extremely low angle to the ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a mountain-sized beast with rocky hide visible rolling past the camera?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a harsh low-angled light sweep from the side creating pitch-black shadows and blown-out white highlights?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera sway violently and jolt as the beast steps, causing rubble and dust to swirl in a chaotic cloud?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.4742940664291382 + ], + "opposite_logprob_tokens": [ + -0.9742940664291382 + ], + "logprob_reward": -0.4742940664291382, + "yes_no_margin_reward": 0.5, + "token_credit_reward": -0.4742940664291382, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09753752499818802, + 0.07295465469360352, + 0.04913659766316414, + 0.0302235409617424, + 0.030172187834978104, + 0.2582302689552307, + 0.06205825135111809, + 0.032946500927209854, + 0.3025819659233093, + 0.06415847688913345 + ], + "credit_map": [ + 0.00969734601676464, + 0.009559143334627151, + 0.010587375611066818, + 0.01431189477443695, + 0.013381238095462322, + 0.01345919631421566, + 0.012615456245839596, + 0.0139258848503232, + 0.009136339649558067, + 0.012944662012159824, + 0.009330982342362404, + 0.012126676738262177, + 0.006247039418667555, + 0.009532686322927475, + 0.00841080117970705, + 0.005225466564297676, + 0.005501176696270704, + 0.00626786332577467, + 0.005342458840459585, + 0.011215271428227425, + 0.004808768630027771, + 0.0043965610675513744, + 0.007244805805385113, + 0.0043596974574029446, + 0.004433604422956705, + 0.004452009219676256, + 0.0032895677722990513, + 0.005721601191908121, + 0.00288747763261199, + 0.0019311869982630014, + 0.004202178213745356, + 0.003305917838588357, + 0.004357532598078251, + 0.0056510563008487225, + 0.003213366726413369, + 0.004446778912097216, + 0.0027154202107340097, + 0.002647970337420702, + 0.004549592267721891, + 0.0025904744397848845, + 0.043537601828575134, + 0.02843344397842884, + 0.0172309260815382, + 0.024757718667387962, + 0.0347738191485405, + 0.03740033507347107, + 0.02808680385351181, + 0.04400962218642235, + 0.007702609058469534, + 0.006759296171367168, + 0.003930442500859499, + 0.013845804147422314, + 0.006860759109258652, + 0.006429064553231001, + 0.007964206859469414, + 0.008566072210669518, + 0.004428794141858816, + 0.004473200999200344, + 0.003943165298551321, + 0.004133893176913261, + 0.00416252575814724, + 0.003416192950680852, + 0.004093371331691742, + 0.004295365419238806, + 0.042405866086483, + 0.07851976156234741, + 0.03548463433980942, + 0.036064039915800095, + 0.03431762382388115, + 0.02143908478319645, + 0.03391316533088684, + 0.020437845960259438, + 0.00900003407150507, + 0.006390425376594067, + 0.007589193060994148, + 0.012041435576975346, + 0.007430799771100283, + 0.004379897844046354, + 0.010446406900882721, + 0.006880287081003189 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 5, + 0 + ], + "selected_reward": -0.4742940664291382 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene presented in a stark black and white contrast?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera fixed at an extremely low angle to the ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a mountain-sized beast with rocky hide visible rolling past the camera?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a harsh low-angled light sweep from the side creating pitch-black shadows and blown-out white highlights?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera sway violently and jolt as the beast steps, causing rubble and dust to swirl in a chaotic cloud?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004639219027012587 + ], + "opposite_logprob_tokens": [ + -5.379639148712158 + ], + "logprob_reward": -0.004639219027012587, + "yes_no_margin_reward": 5.375, + "token_credit_reward": -0.004639219027012587, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0828077420592308, + 0.1625043749809265, + 0.048520274460315704, + 0.02316971868276596, + 0.042934492230415344, + 0.1681205779314041, + 0.06935008615255356, + 0.1019439622759819, + 0.190535768866539, + 0.11011295020580292 + ], + "credit_map": [ + 0.00902306754142046, + 0.01060703955590725, + 0.00879185926169157, + 0.015382708050310612, + 0.010416190139949322, + 0.011925616301596165, + 0.00793500803411007, + 0.008726262487471104, + 0.018934287130832672, + 0.028551073744893074, + 0.019515659660100937, + 0.021512432023882866, + 0.02526617795228958, + 0.020403575152158737, + 0.014523609541356564, + 0.013797571882605553, + 0.004520564340054989, + 0.005439637694507837, + 0.005692807026207447, + 0.007034013047814369, + 0.006334901321679354, + 0.005111441947519779, + 0.008862694725394249, + 0.005524221807718277, + 0.0032571284100413322, + 0.00365600292570889, + 0.003229073714464903, + 0.003629210637882352, + 0.001785239321179688, + 0.0012626489624381065, + 0.003900888143107295, + 0.0024495278485119343, + 0.006247642450034618, + 0.006298399064689875, + 0.003919398877769709, + 0.0057685403153300285, + 0.0028268303722143173, + 0.005602533463388681, + 0.006897832732647657, + 0.0053733219392597675, + 0.018624553456902504, + 0.023894915357232094, + 0.01724766008555889, + 0.02035045437514782, + 0.030039969831705093, + 0.020978158339858055, + 0.02263883501291275, + 0.014346051029860973, + 0.00595896178856492, + 0.007638340815901756, + 0.007108153309673071, + 0.008291009813547134, + 0.01167601253837347, + 0.00881415605545044, + 0.00914058182388544, + 0.010722879320383072, + 0.012226106598973274, + 0.01808413304388523, + 0.00608246773481369, + 0.011700544506311417, + 0.016054192557930946, + 0.010399574413895607, + 0.016453895717859268, + 0.010943058878183365, + 0.026347465813159943, + 0.03003493882715702, + 0.01742849126458168, + 0.02365262247622013, + 0.020438438281416893, + 0.019725758582353592, + 0.02397013269364834, + 0.028937961906194687, + 0.012758288532495499, + 0.012893905863165855, + 0.012532344087958336, + 0.01914065144956112, + 0.011692238040268421, + 0.006853072438389063, + 0.022112391889095306, + 0.0121300695464015 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 5, + 1 + ], + "selected_reward": -0.004639219027012587 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene presented in a stark black and white contrast?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera fixed at an extremely low angle to the ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a mountain-sized beast with rocky hide visible rolling past the camera?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a harsh low-angled light sweep from the side creating pitch-black shadows and blown-out white highlights?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera sway violently and jolt as the beast steps, causing rubble and dust to swirl in a chaotic cloud?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0067444778978824615 + ], + "opposite_logprob_tokens": [ + -5.006744384765625 + ], + "logprob_reward": -0.0067444778978824615, + "yes_no_margin_reward": 5.0, + "token_credit_reward": -0.0067444778978824615, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07646037638187408, + 0.06127139553427696, + 0.0405772365629673, + 0.04227793216705322, + 0.09001791477203369, + 0.14033670723438263, + 0.10449028015136719, + 0.13685335218906403, + 0.1825093775987625, + 0.125205397605896 + ], + "credit_map": [ + 0.00948359165340662, + 0.0086912140250206, + 0.006942239124327898, + 0.010556702502071857, + 0.012009798549115658, + 0.010730068199336529, + 0.008818007074296474, + 0.009228757582604885, + 0.006429583765566349, + 0.010009524412453175, + 0.007214389741420746, + 0.012447858229279518, + 0.005249774549156427, + 0.007533052004873753, + 0.006473193876445293, + 0.005914017092436552, + 0.004964590072631836, + 0.004388867877423763, + 0.004455153364688158, + 0.007450919132679701, + 0.002958713797852397, + 0.003540254896506667, + 0.0063756778836250305, + 0.006443058140575886, + 0.005686670541763306, + 0.006939424201846123, + 0.006119714118540287, + 0.006926233880221844, + 0.003328474471345544, + 0.0025431348476558924, + 0.006621118169277906, + 0.004113160539418459, + 0.013478219509124756, + 0.020360039547085762, + 0.007775343954563141, + 0.014136125333607197, + 0.008572308346629143, + 0.008136291056871414, + 0.010400892235338688, + 0.007158698979765177, + 0.015968158841133118, + 0.016619479283690453, + 0.016470469534397125, + 0.017378993332386017, + 0.025209587067365646, + 0.017675120383501053, + 0.018421292304992676, + 0.012593605555593967, + 0.009535851888358593, + 0.00874913576990366, + 0.00987000111490488, + 0.010299955494701862, + 0.014589434489607811, + 0.01535829808562994, + 0.013993244618177414, + 0.02209436148405075, + 0.019901331514120102, + 0.029977302998304367, + 0.013385373167693615, + 0.020905286073684692, + 0.011699841357767582, + 0.010482852347195148, + 0.01930246315896511, + 0.011198896914720535, + 0.018501996994018555, + 0.031197277829051018, + 0.018845221027731895, + 0.02222684770822525, + 0.016628898680210114, + 0.01607091724872589, + 0.03169802203774452, + 0.027340197935700417, + 0.015647323802113533, + 0.012903698720037937, + 0.014234994538128376, + 0.022385941818356514, + 0.015919193625450134, + 0.0076971095986664295, + 0.021667730063199997, + 0.014749397523701191 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 5, + 7 + ], + "selected_reward": -0.0067444778978824615 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene presented in a stark black and white contrast?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera fixed at an extremely low angle to the ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a mountain-sized beast with rocky hide visible rolling past the camera?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a harsh low-angled light sweep from the side creating pitch-black shadows and blown-out white highlights?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera sway violently and jolt as the beast steps, causing rubble and dust to swirl in a chaotic cloud?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.016071652993559837 + ], + "opposite_logprob_tokens": [ + -4.141071796417236 + ], + "logprob_reward": -0.016071652993559837, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.016071652993559837, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.05118933320045471, + 0.07450322061777115, + 0.10231968760490417, + 0.047325197607278824, + 0.04098182171583176, + 0.07691595703363419, + 0.08016325533390045, + 0.02734084613621235, + 0.3841840922832489, + 0.11507659405469894 + ], + "credit_map": [ + 0.004864399787038565, + 0.00504670524969697, + 0.005032448563724756, + 0.0068216086365282536, + 0.007360274437814951, + 0.007177672814577818, + 0.007401483133435249, + 0.007484741974622011, + 0.008433018811047077, + 0.012138094753026962, + 0.008347960188984871, + 0.007945255376398563, + 0.007650279439985752, + 0.01050012931227684, + 0.009929712861776352, + 0.009558763355016708, + 0.010450799018144608, + 0.01114886999130249, + 0.010058525949716568, + 0.017978336662054062, + 0.010187610052525997, + 0.007810600101947784, + 0.022202366963028908, + 0.012482589110732079, + 0.005850340239703655, + 0.007673386484384537, + 0.009386099874973297, + 0.005717407446354628, + 0.005409673321992159, + 0.0036369631998240948, + 0.00594225013628602, + 0.0037090792320668697, + 0.005870799068361521, + 0.0065137953497469425, + 0.003399176523089409, + 0.005006526131182909, + 0.0038242603186517954, + 0.005461163353174925, + 0.006409830413758755, + 0.0044962698593735695, + 0.01596212573349476, + 0.006478183902800083, + 0.0077522676438093185, + 0.006349564529955387, + 0.005992045626044273, + 0.012239865958690643, + 0.006586181465536356, + 0.015555721707642078, + 0.007947683334350586, + 0.007663005962967873, + 0.006851013284176588, + 0.00813563447445631, + 0.01627492904663086, + 0.011765471659600735, + 0.010545605793595314, + 0.010979915969073772, + 0.0036645177751779556, + 0.005735927261412144, + 0.002810981823131442, + 0.0033981199376285076, + 0.0032900115475058556, + 0.0023324282374233007, + 0.0038888612762093544, + 0.002219998277723789, + 0.054241981357336044, + 0.0656324103474617, + 0.035910192877054214, + 0.042317066341638565, + 0.036278583109378815, + 0.042187463492155075, + 0.04512786865234375, + 0.06248851120471954, + 0.014567640610039234, + 0.01150435209274292, + 0.012191884219646454, + 0.019519032910466194, + 0.01169925183057785, + 0.00774014089256525, + 0.023670252412557602, + 0.014184044674038887 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 9, + 2 + ], + "selected_reward": -0.016071652993559837 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "powerful overhead shot, the camera fixed at an extremely low angle to the ground, swaying violently as the earth trembles violently, reveals a mountain-sized beast slowly and ponderously rolling past. The entire scene is shrouded in a stark contrast of black and white. A harsh, low-angled light sweeps from the side, carving every deep groove in the beast's rocky hide into a pitch-black shadow, while its ridges of muscle and bone appear a brilliant, almost blown-out white. Each step of its towering legs sends the camera careening with a devastating jolt, sending countless rubble and dust from the ground swirling in a chaotic, swirling cloud of smoke in the light beam. The viewer's gaze is completely absorbed by a portion of its massive body, unable to see its full extent, only its unparalleled mass and power. The entire image is imbued with a sense of primal, untamed power, creating a sense of apocalyptic, suffocating oppression and epic grandeur.", + "seed": 12741, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000006/group_6/index_7633_seed_12741_dp_13_rank_26_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 6, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.13451204728335142, + "vlm_logprob": -0.13451204728335142, + "vlm_yes_no_margin": 3.275, + "vlm_token_credit": -0.13451204728335142, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene presented in a stark black and white contrast?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera fixed at an extremely low angle to the ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a mountain-sized beast with rocky hide visible rolling past the camera?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a harsh low-angled light sweep from the side creating pitch-black shadows and blown-out white highlights?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera sway violently and jolt as the beast steps, causing rubble and dust to swirl in a chaotic cloud?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.16535061597824097 + ], + "opposite_logprob_tokens": [ + -1.9153506755828857 + ], + "logprob_reward": -0.16535061597824097, + "yes_no_margin_reward": 1.75, + "token_credit_reward": -0.16535061597824097, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12024852633476257, + 0.16600000858306885, + 0.07298188656568527, + 0.10599756985902786, + 0.07654876261949539, + 0.07112286239862442, + 0.028031039983034134, + 0.052887994796037674, + 0.06614581495523453, + 0.2400355488061905 + ], + "credit_map": [ + 0.01593860425055027, + 0.010199528187513351, + 0.017243722453713417, + 0.015228127129375935, + 0.010922545567154884, + 0.01655728369951248, + 0.016073178499937057, + 0.018085524439811707, + 0.03302113711833954, + 0.01469508372247219, + 0.009267302230000496, + 0.012592953629791737, + 0.049037303775548935, + 0.012203536927700043, + 0.018988458439707756, + 0.016194216907024384, + 0.008982214145362377, + 0.011769188567996025, + 0.01103390846401453, + 0.006880716420710087, + 0.009882625192403793, + 0.0079068997874856, + 0.010817980393767357, + 0.005708345212042332, + 0.012865865603089333, + 0.012993927113711834, + 0.015505432151257992, + 0.01805342175066471, + 0.01076352596282959, + 0.012248040176928043, + 0.01327905710786581, + 0.010288292542099953, + 0.01073791366070509, + 0.009006214328110218, + 0.00916810892522335, + 0.0117591992020607, + 0.005809290334582329, + 0.008723528124392033, + 0.009020641446113586, + 0.012323861941695213, + 0.006421784404665232, + 0.01335228607058525, + 0.009962236508727074, + 0.007861446589231491, + 0.007238555699586868, + 0.009038183838129044, + 0.008043618872761726, + 0.009204739704728127, + 0.0032493986655026674, + 0.003289841813966632, + 0.0029776368755847216, + 0.0025618276558816433, + 0.002663567429408431, + 0.005913954693824053, + 0.0028191180899739265, + 0.00455569289624691, + 0.005223752465099096, + 0.007433398626744747, + 0.006000147666782141, + 0.008932499215006828, + 0.005154313053935766, + 0.004920085426419973, + 0.009667815640568733, + 0.005555983167141676, + 0.006768077611923218, + 0.007057739421725273, + 0.00633394718170166, + 0.007503949571400881, + 0.00721528846770525, + 0.008738663978874683, + 0.01013548206537962, + 0.01239265687763691, + 0.032145798206329346, + 0.028860561549663544, + 0.033749617636203766, + 0.023521458730101585, + 0.03490917757153511, + 0.03799227625131607, + 0.021114341914653778, + 0.0277422945946455 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 1, + 0 + ], + "selected_reward": -0.16535061597824097 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene presented in a stark black and white contrast?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera fixed at an extremely low angle to the ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a mountain-sized beast with rocky hide visible rolling past the camera?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a harsh low-angled light sweep from the side creating pitch-black shadows and blown-out white highlights?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera sway violently and jolt as the beast steps, causing rubble and dust to swirl in a chaotic cloud?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.4743090867996216 + ], + "opposite_logprob_tokens": [ + -0.9743090867996216 + ], + "logprob_reward": -0.4743090867996216, + "yes_no_margin_reward": 0.5, + "token_credit_reward": -0.4743090867996216, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.03834031522274017, + 0.1048930212855339, + 0.5454636812210083, + 0.016252879053354263, + 0.021504200994968414, + 0.035518087446689606, + 0.1518983393907547, + 0.032462798058986664, + 0.03732268884778023, + 0.016343992203474045 + ], + "credit_map": [ + 0.00425178837031126, + 0.003223014995455742, + 0.0035887956619262695, + 0.0049712988547980785, + 0.004995169583708048, + 0.00603565713390708, + 0.006183224730193615, + 0.005091365892440081, + 0.015836071223020554, + 0.014461335726082325, + 0.008757724426686764, + 0.013254828751087189, + 0.014528430067002773, + 0.014995184727013111, + 0.01310522761195898, + 0.00995421502739191, + 0.08683735132217407, + 0.14002655446529388, + 0.08708812296390533, + 0.06085231527686119, + 0.04937048256397247, + 0.050302065908908844, + 0.04401976615190506, + 0.026967070996761322, + 0.0021554692648351192, + 0.0023186085745692253, + 0.0017661782912909985, + 0.002288896357640624, + 0.0026452266611158848, + 0.0013332999078556895, + 0.002138588111847639, + 0.0016066106036305428, + 0.0028756449464708567, + 0.0029363278299570084, + 0.002518875990062952, + 0.003066183300688863, + 0.001739076804369688, + 0.001797095057554543, + 0.004204871598631144, + 0.002366125350818038, + 0.003494087141007185, + 0.005860449746251106, + 0.004475515801459551, + 0.004648973699659109, + 0.0034906479995697737, + 0.004147600382566452, + 0.004511716775596142, + 0.004889093805104494, + 0.01915952004492283, + 0.023812739178538322, + 0.016830207780003548, + 0.017537176609039307, + 0.014947623945772648, + 0.025351693853735924, + 0.011789141222834587, + 0.02247023768723011, + 0.003474750556051731, + 0.005285507068037987, + 0.004304048139601946, + 0.004621284548193216, + 0.0035438917111605406, + 0.0037943520583212376, + 0.004338102415204048, + 0.0031008643563836813, + 0.004322814755141735, + 0.003794513875618577, + 0.0036521824076771736, + 0.003906590864062309, + 0.004998900927603245, + 0.004596504848450422, + 0.005653854925185442, + 0.006397326942533255, + 0.002526613650843501, + 0.0022054833825677633, + 0.0030321606900542974, + 0.002062232932075858, + 0.0017402205849066377, + 0.0013440672773867846, + 0.0017060069367289543, + 0.0017272067489102483 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 6, + 1 + ], + "selected_reward": -0.4743090867996216 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene presented in a stark black and white contrast?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera fixed at an extremely low angle to the ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a mountain-sized beast with rocky hide visible rolling past the camera?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a harsh low-angled light sweep from the side creating pitch-black shadows and blown-out white highlights?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera sway violently and jolt as the beast steps, causing rubble and dust to swirl in a chaotic cloud?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0036200247704982758 + ], + "opposite_logprob_tokens": [ + -5.628620147705078 + ], + "logprob_reward": -0.0036200247704982758, + "yes_no_margin_reward": 5.625, + "token_credit_reward": -0.0036200247704982758, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10132274776697159, + 0.17653919756412506, + 0.12889835238456726, + 0.09982797503471375, + 0.05726371333003044, + 0.06515154987573624, + 0.26475611329078674, + 0.026054508984088898, + 0.030837416648864746, + 0.04934845119714737 + ], + "credit_map": [ + 0.009302723221480846, + 0.010892021469771862, + 0.013390322215855122, + 0.015295114368200302, + 0.008328717201948166, + 0.016011733561754227, + 0.012892142869532108, + 0.01520997378975153, + 0.046661023050546646, + 0.015642499551177025, + 0.00750988582149148, + 0.007853412069380283, + 0.05964052304625511, + 0.013193916529417038, + 0.01168308686465025, + 0.014354840852320194, + 0.017090480774641037, + 0.02518821693956852, + 0.021586833521723747, + 0.015919556841254234, + 0.011681299656629562, + 0.010600963607430458, + 0.015926918014883995, + 0.010904080234467983, + 0.011186441406607628, + 0.011978154070675373, + 0.015994146466255188, + 0.013401278294622898, + 0.012254228815436363, + 0.00993985403329134, + 0.015923313796520233, + 0.009150552563369274, + 0.005241358652710915, + 0.0055354586802423, + 0.0066389599815011024, + 0.009772571735084057, + 0.006471529137343168, + 0.008177935145795345, + 0.008398330770432949, + 0.007027569226920605, + 0.005704921204596758, + 0.008062548004090786, + 0.007873824797570705, + 0.008601061068475246, + 0.007675187196582556, + 0.009699176996946335, + 0.00804818794131279, + 0.00948664452880621, + 0.0335443913936615, + 0.04275229945778847, + 0.029834827408194542, + 0.029716556891798973, + 0.02613195590674877, + 0.044409044086933136, + 0.021020177751779556, + 0.03734687343239784, + 0.002350958762690425, + 0.0036843069829046726, + 0.0035487802233546972, + 0.003687383607029915, + 0.003841511672362685, + 0.003371171187609434, + 0.0031846179626882076, + 0.0023857783526182175, + 0.0030747679993510246, + 0.0032886762637645006, + 0.0028789816424250603, + 0.0037600977811962366, + 0.0038129433523863554, + 0.0037138042971491814, + 0.0038947705179452896, + 0.0064133754931390285, + 0.005699431989341974, + 0.005446980707347393, + 0.006116047035902739, + 0.004936807788908482, + 0.008127695880830288, + 0.006441972684115171, + 0.0056982156820595264, + 0.006881298031657934 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 1, + 2 + ], + "selected_reward": -0.0036200247704982758 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene presented in a stark black and white contrast?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera fixed at an extremely low angle to the ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a mountain-sized beast with rocky hide visible rolling past the camera?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a harsh low-angled light sweep from the side creating pitch-black shadows and blown-out white highlights?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera sway violently and jolt as the beast steps, causing rubble and dust to swirl in a chaotic cloud?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.011086766608059406 + ], + "opposite_logprob_tokens": [ + -4.511086940765381 + ], + "logprob_reward": -0.011086766608059406, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.011086766608059406, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06150941923260689, + 0.2274584025144577, + 0.050943393260240555, + 0.13961468636989594, + 0.06250878423452377, + 0.05801943689584732, + 0.22776858508586884, + 0.08115607500076294, + 0.04248020797967911, + 0.04854093864560127 + ], + "credit_map": [ + 0.007283961866050959, + 0.006346032489091158, + 0.006498655769973993, + 0.008369958959519863, + 0.005532161798328161, + 0.010822765529155731, + 0.008905179798603058, + 0.007750716060400009, + 0.05769617110490799, + 0.020099811255931854, + 0.007679756265133619, + 0.008024157956242561, + 0.08296143263578415, + 0.010146659798920155, + 0.020445529371500015, + 0.020404895767569542, + 0.008166572079062462, + 0.0051162708550691605, + 0.007881026715040207, + 0.0035933086182922125, + 0.006962059065699577, + 0.009229034185409546, + 0.004886549897491932, + 0.005108577199280262, + 0.01967036910355091, + 0.021038087084889412, + 0.02571607194840908, + 0.013918655924499035, + 0.02224818803369999, + 0.010923953726887703, + 0.017944416031241417, + 0.00815496314316988, + 0.008822446689009666, + 0.008633745834231377, + 0.007416705135256052, + 0.008634217083454132, + 0.005253084935247898, + 0.006681468803435564, + 0.009914319962263107, + 0.007152802310883999, + 0.006459573283791542, + 0.011206337250769138, + 0.007936161011457443, + 0.005767481867223978, + 0.005654490552842617, + 0.007250017020851374, + 0.006548711564391851, + 0.0071966699324548244, + 0.029605504125356674, + 0.03508496284484863, + 0.024234147742390633, + 0.02539876103401184, + 0.021677745506167412, + 0.040271274745464325, + 0.01761552132666111, + 0.03388068825006485, + 0.008798711933195591, + 0.014291838742792606, + 0.011783859692513943, + 0.011810737662017345, + 0.007462027482688427, + 0.00846870057284832, + 0.010960372164845467, + 0.0075798314064741135, + 0.004375972785055637, + 0.004256934393197298, + 0.004281623288989067, + 0.004632454365491867, + 0.004904261790215969, + 0.005643335171043873, + 0.006806922145187855, + 0.007578709628432989, + 0.006260171066969633, + 0.005020486190915108, + 0.006866252515465021, + 0.004105847794562578, + 0.005939365364611149, + 0.0076113431714475155, + 0.005591552704572678, + 0.007145925424993038 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 1, + 3 + ], + "selected_reward": -0.011086766608059406 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene presented in a stark black and white contrast?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera fixed at an extremely low angle to the ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a mountain-sized beast with rocky hide visible rolling past the camera?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a harsh low-angled light sweep from the side creating pitch-black shadows and blown-out white highlights?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera sway violently and jolt as the beast steps, causing rubble and dust to swirl in a chaotic cloud?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.018193742260336876 + ], + "opposite_logprob_tokens": [ + -4.01819372177124 + ], + "logprob_reward": -0.018193742260336876, + "yes_no_margin_reward": 4.0, + "token_credit_reward": -0.018193742260336876, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11494211852550507, + 0.0708935335278511, + 0.32749927043914795, + 0.15209819376468658, + 0.04377950727939606, + 0.03740790858864784, + 0.08303223550319672, + 0.06207546219229698, + 0.06336788088083267, + 0.044903889298439026 + ], + "credit_map": [ + 0.014655008912086487, + 0.01592014916241169, + 0.015612630173563957, + 0.01353555265814066, + 0.011209382675588131, + 0.013454665429890156, + 0.015249049291014671, + 0.015305675566196442, + 0.011307807639241219, + 0.006883302703499794, + 0.00563404243439436, + 0.00698358379304409, + 0.015667622908949852, + 0.008546695113182068, + 0.009163249284029007, + 0.006707229185849428, + 0.045912276953458786, + 0.07326623797416687, + 0.06108500063419342, + 0.037201788276433945, + 0.04221993684768677, + 0.02395840361714363, + 0.024111101403832436, + 0.019744524732232094, + 0.018628569319844246, + 0.019878264516592026, + 0.027774961665272713, + 0.01976463943719864, + 0.016163310036063194, + 0.017024926841259003, + 0.02057493105530739, + 0.012288595549762249, + 0.004128923639655113, + 0.004850427620112896, + 0.004946975037455559, + 0.006552025675773621, + 0.003828467568382621, + 0.004584313835948706, + 0.009557832032442093, + 0.0053305416367948055, + 0.004783507902175188, + 0.007539030630141497, + 0.005627619568258524, + 0.0031421950552612543, + 0.004321701359003782, + 0.005953497253358364, + 0.0024984553456306458, + 0.003541899612173438, + 0.01060811709612608, + 0.010907277464866638, + 0.009794210083782673, + 0.008837968111038208, + 0.006222917232662439, + 0.01706327684223652, + 0.0074002426117658615, + 0.012198229320347309, + 0.0054298522882163525, + 0.007469936273992062, + 0.006524241995066404, + 0.007445048540830612, + 0.007504105567932129, + 0.0085514597594738, + 0.009610001929104328, + 0.009540818631649017, + 0.006593344267457724, + 0.006019115447998047, + 0.00665307929739356, + 0.006292922422289848, + 0.007177690509706736, + 0.008730713278055191, + 0.009792063385248184, + 0.012108951807022095, + 0.004570536315441132, + 0.004992734640836716, + 0.003967654425650835, + 0.005063518416136503, + 0.006214076653122902, + 0.004676285199820995, + 0.008289140649139881, + 0.007129942532628775 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 3, + 0 + ], + "selected_reward": -0.018193742260336876 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000006/group_7/index_tensor([2712, 2712])_group_7_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000006/group_7/index_tensor([2712, 2712])_group_7_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..eabd2787ffff9eace8ba86df6868bd0ad75b7541 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000006/group_7/index_tensor([2712, 2712])_group_7_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "An extreme close-up, static macro shot, fixed at eye level, focuses on a single eye. Within its deep brown iris, the pupil, sharply constricted by extreme fear, reflects a faint, unidentified light source. Initially, the white of the eye appears clear and moist, a few long, trembling eyelashes casting a faint shadow. Suddenly, from the corner of the eye, a thin thread of blood emerges like a flash of red lightning. With an unsettling speed, it begins to branch and spread frantically, a network of countless scarlet, branching capillaries rapidly eroding the once pristine white sclera. This process is steady and relentless, the network of blood vessels growing denser and denser, evolving from sparse lines to a crisscrossing crimson, ultimately engulfing the entire white of the eyeball, transforming it into a terrifying, biologically menacing blood-red. The entire scene, illuminated by the cold, sharp light, creates a sense of intense physical discomfort and a profoundly visually horrifying atmosphere.", + "seed": 22675, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000006/group_7/index_2712_seed_22675_dp_14_rank_28_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 7, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.009911953436676413, + "vlm_logprob": -0.009911953436676413, + "vlm_yes_no_margin": 6.525, + "vlm_token_credit": -0.009911953436676413, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up focusing on a single eye?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the eye initially appear clear and moist with long trembling eyelashes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a thin thread of blood emerge from the corner of the eye and branch out rapidly?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the blood network spread to engulf the entire white sclera, turning it blood-red?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot static and fixed at eye level throughout the sequence?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.007666332647204399 + ], + "opposite_logprob_tokens": [ + -10.25766658782959 + ], + "logprob_reward": -0.007666332647204399, + "yes_no_margin_reward": 10.25, + "token_credit_reward": -0.007666332647204399, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1339000165462494, + 0.04831770434975624, + 0.05797271057963371, + 0.1391621083021164, + 0.07232645153999329, + 0.16231606900691986, + 0.064908966422081, + 0.05540642514824867, + 0.06513932347297668, + 0.20055027306079865 + ], + "credit_map": [ + 0.02318519726395607, + 0.007011634763330221, + 0.014623790048062801, + 0.0077436622232198715, + 0.020343521609902382, + 0.013847296126186848, + 0.01212435495108366, + 0.03502056747674942, + 0.006295504979789257, + 0.0018004006706178188, + 0.005900240503251553, + 0.0025004451163113117, + 0.006488507613539696, + 0.004800331313163042, + 0.004155396018177271, + 0.016376879066228867, + 0.00716570857912302, + 0.0021718533243983984, + 0.0062170433811843395, + 0.003825685242190957, + 0.007987728342413902, + 0.00714818574488163, + 0.006211510859429836, + 0.01724499650299549, + 0.019728589802980423, + 0.005606437101960182, + 0.029946889728307724, + 0.00843114499002695, + 0.0158859696239233, + 0.016977349296212196, + 0.014953577890992165, + 0.027632150799036026, + 0.009046223014593124, + 0.00491663021966815, + 0.005798203404992819, + 0.004659145604819059, + 0.010961780324578285, + 0.012108808383345604, + 0.009003165178000927, + 0.01583249308168888, + 0.017351897433400154, + 0.007781829219311476, + 0.009234984405338764, + 0.008796569891273975, + 0.01906764507293701, + 0.045681968331336975, + 0.014636416919529438, + 0.03976475074887276, + 0.006496863905340433, + 0.0041801840998232365, + 0.005349744576960802, + 0.009151609614491463, + 0.008994154632091522, + 0.008950179442763329, + 0.008724542334675789, + 0.013061692006886005, + 0.005723257083445787, + 0.003930557053536177, + 0.004069110378623009, + 0.008127599954605103, + 0.007183580659329891, + 0.00596056692302227, + 0.012396786361932755, + 0.008014966733753681, + 0.007720315363258123, + 0.006718765012919903, + 0.004343152977526188, + 0.005960205104202032, + 0.011949868872761726, + 0.009238154627382755, + 0.009474475868046284, + 0.009734383784234524, + 0.007168370299041271, + 0.007555438205599785, + 0.006524319294840097, + 0.008692271076142788, + 0.022530287504196167, + 0.06993450224399567, + 0.06425414234399796, + 0.0138909462839365 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 5, + 3 + ], + "selected_reward": -0.007666332647204399 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up focusing on a single eye?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the eye initially appear clear and moist with long trembling eyelashes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a thin thread of blood emerge from the corner of the eye and branch out rapidly?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the blood network spread to engulf the entire white sclera, turning it blood-red?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot static and fixed at eye level throughout the sequence?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005272530950605869 + ], + "opposite_logprob_tokens": [ + -5.255272388458252 + ], + "logprob_reward": -0.005272530950605869, + "yes_no_margin_reward": 5.25, + "token_credit_reward": -0.005272530950605869, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07456603646278381, + 0.060977838933467865, + 0.09704037010669708, + 0.2745419144630432, + 0.0968145951628685, + 0.11487261950969696, + 0.07672904431819916, + 0.049949679523706436, + 0.0986359566450119, + 0.05587194859981537 + ], + "credit_map": [ + 0.006044804118573666, + 0.0030691870488226414, + 0.00720339035615325, + 0.0052448357455432415, + 0.01097168494015932, + 0.009384794160723686, + 0.008050329983234406, + 0.024597015231847763, + 0.004968765191733837, + 0.0016473981086164713, + 0.00771572720259428, + 0.0030424105934798717, + 0.008824835531413555, + 0.007196241058409214, + 0.0061542438343167305, + 0.021428214386105537, + 0.01568617857992649, + 0.007443148642778397, + 0.011233397759497166, + 0.006952112540602684, + 0.013071822933852673, + 0.014018183574080467, + 0.011228113435208797, + 0.01740741729736328, + 0.028967633843421936, + 0.016553619876503944, + 0.04184480011463165, + 0.018535878509283066, + 0.03431304544210434, + 0.040297407656908035, + 0.03270161524415016, + 0.061327893286943436, + 0.019034884870052338, + 0.005700371693819761, + 0.009212013334035873, + 0.008422603830695152, + 0.014803903177380562, + 0.01728496141731739, + 0.009041435085237026, + 0.013314418494701385, + 0.01128035131841898, + 0.008004656992852688, + 0.011714732274413109, + 0.0105539970099926, + 0.012677326798439026, + 0.021894685924053192, + 0.010136927478015423, + 0.02860993705689907, + 0.009459397755563259, + 0.004889995791018009, + 0.006057858467102051, + 0.006651243194937706, + 0.010267677716910839, + 0.012403465807437897, + 0.010346564464271069, + 0.01665283553302288, + 0.004379423800855875, + 0.004369356669485569, + 0.0034050613176077604, + 0.005265981424599886, + 0.006962763145565987, + 0.006778666749596596, + 0.012084310874342918, + 0.006704118102788925, + 0.007164953742176294, + 0.008018922060728073, + 0.006085171829909086, + 0.009551049210131168, + 0.01494415570050478, + 0.012617756612598896, + 0.020674830302596092, + 0.019579118117690086, + 0.00229979632422328, + 0.002091826405376196, + 0.0016971012810245156, + 0.0024836095981299877, + 0.00588091230019927, + 0.021813152357935905, + 0.012671537697315216, + 0.006934012286365032 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 5, + 8 + ], + "selected_reward": -0.005272530950605869 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up focusing on a single eye?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the eye initially appear clear and moist with long trembling eyelashes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a thin thread of blood emerge from the corner of the eye and branch out rapidly?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the blood network spread to engulf the entire white sclera, turning it blood-red?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot static and fixed at eye level throughout the sequence?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.00044050050200894475 + ], + "opposite_logprob_tokens": [ + -7.75044059753418 + ], + "logprob_reward": -0.00044050050200894475, + "yes_no_margin_reward": 7.75, + "token_credit_reward": -0.00044050050200894475, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.04275801405310631, + 0.03109925612807274, + 0.014246192760765553, + 0.1011415645480156, + 0.11701075732707977, + 0.1414872258901596, + 0.029419615864753723, + 0.03121872805058956, + 0.03777875006198883, + 0.4538399279117584 + ], + "credit_map": [ + 0.0050192540511488914, + 0.0017299865139648318, + 0.0040257880464196205, + 0.002238054294139147, + 0.005892347078770399, + 0.006081602536141872, + 0.00438987510278821, + 0.013381104916334152, + 0.003172175493091345, + 0.0015974954003468156, + 0.0036076714750379324, + 0.001917480956763029, + 0.003881596028804779, + 0.0036041983403265476, + 0.0029293971601873636, + 0.010389240458607674, + 0.002441337564960122, + 0.0010705911554396152, + 0.001597774331457913, + 0.0011141496943309903, + 0.002150368643924594, + 0.0022691956255584955, + 0.0015007974579930305, + 0.002101978985592723, + 0.014059357345104218, + 0.005239518359303474, + 0.030082549899816513, + 0.007095117121934891, + 0.011692285537719727, + 0.013151256367564201, + 0.008065165020525455, + 0.011756319552659988, + 0.020162565633654594, + 0.007844790816307068, + 0.011464574374258518, + 0.00863407738506794, + 0.018930289894342422, + 0.02118493989109993, + 0.010625146329402924, + 0.018164372071623802, + 0.016440894454717636, + 0.006477366201579571, + 0.026067139580845833, + 0.013594163581728935, + 0.013218388892710209, + 0.02880159765481949, + 0.013843162916600704, + 0.023044515401124954, + 0.0052785854786634445, + 0.002174904802814126, + 0.0024580834433436394, + 0.002898156875744462, + 0.004463785793632269, + 0.0037542367354035378, + 0.0042675030417740345, + 0.0041243587620556355, + 0.003281276673078537, + 0.0021598294842988253, + 0.0016546042170375586, + 0.003968620672821999, + 0.004210000857710838, + 0.00411151023581624, + 0.005881897639483213, + 0.0059509859420359135, + 0.0033599589951336384, + 0.004010265693068504, + 0.0027980573941022158, + 0.004203884396702051, + 0.006671097595244646, + 0.006496341433376074, + 0.0049206893891096115, + 0.005318454001098871, + 0.01619768515229225, + 0.011377143673598766, + 0.010705351829528809, + 0.018984198570251465, + 0.03184058889746666, + 0.1698731631040573, + 0.16646839678287506, + 0.028393397107720375 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 5, + 4 + ], + "selected_reward": -0.00044050050200894475 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up focusing on a single eye?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the eye initially appear clear and moist with long trembling eyelashes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a thin thread of blood emerge from the corner of the eye and branch out rapidly?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the blood network spread to engulf the entire white sclera, turning it blood-red?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot static and fixed at eye level throughout the sequence?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0024857595562934875 + ], + "opposite_logprob_tokens": [ + -6.002485752105713 + ], + "logprob_reward": -0.0024857595562934875, + "yes_no_margin_reward": 6.0, + "token_credit_reward": -0.0024857595562934875, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.02485102042555809, + 0.015529653057456017, + 0.006381843239068985, + 0.04844225198030472, + 0.03362668305635452, + 0.0311688631772995, + 0.04925864189863205, + 0.044315267354249954, + 0.032491303980350494, + 0.7139344811439514 + ], + "credit_map": [ + 0.0030113954562693834, + 0.0008545331074856222, + 0.0029572302009910345, + 0.0010685193119570613, + 0.003336906200274825, + 0.0034194786567240953, + 0.0023809969425201416, + 0.007821961306035519, + 0.0015199005138128996, + 0.0009090875391848385, + 0.0020115359220653772, + 0.0010789185762405396, + 0.002167642582207918, + 0.0020140735432505608, + 0.0013080359203740954, + 0.00452045863494277, + 0.0006591915735043585, + 0.000446846621343866, + 0.0007292507216334343, + 0.0006366462912410498, + 0.0009355637012049556, + 0.0009938686853274703, + 0.0009418986737728119, + 0.0010385768255218863, + 0.00562236737459898, + 0.002520240843296051, + 0.01320509985089302, + 0.0033683639485388994, + 0.006222833879292011, + 0.006715519819408655, + 0.004558933898806572, + 0.006228889338672161, + 0.0027530579827725887, + 0.002011652570217848, + 0.004827525001019239, + 0.0031794141978025436, + 0.003309152089059353, + 0.00467173894867301, + 0.003515503602102399, + 0.00935863982886076, + 0.005278212018311024, + 0.0022917960304766893, + 0.0032707194332033396, + 0.002423733938485384, + 0.0036381003446877003, + 0.005009412299841642, + 0.004632867407053709, + 0.004624020308256149, + 0.005652592051774263, + 0.003878683550283313, + 0.0056878249160945415, + 0.007220863364636898, + 0.006310689263045788, + 0.006639225874096155, + 0.005562019534409046, + 0.008306744508445263, + 0.006127096246927977, + 0.003533132839947939, + 0.0032036341726779938, + 0.0075982799753546715, + 0.005463475361466408, + 0.005498668644577265, + 0.005758283659815788, + 0.007132697384804487, + 0.002256938023492694, + 0.0024555723648518324, + 0.002033218741416931, + 0.0035993398632854223, + 0.004744667559862137, + 0.0046357931569218636, + 0.006013580597937107, + 0.006752190180122852, + 0.024511637166142464, + 0.017641494050621986, + 0.01621067523956299, + 0.03028947301208973, + 0.051319729536771774, + 0.2756888270378113, + 0.2480509728193283, + 0.050221726298332214 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 6, + 3 + ], + "selected_reward": -0.0024857595562934875 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up focusing on a single eye?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the eye initially appear clear and moist with long trembling eyelashes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a thin thread of blood emerge from the corner of the eye and branch out rapidly?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the blood network spread to engulf the entire white sclera, turning it blood-red?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot static and fixed at eye level throughout the sequence?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03369464352726936 + ], + "opposite_logprob_tokens": [ + -3.4086947441101074 + ], + "logprob_reward": -0.03369464352726936, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.03369464352726936, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.05272287502884865, + 0.05245727300643921, + 0.07776662707328796, + 0.0938660278916359, + 0.07041969895362854, + 0.047499895095825195, + 0.11478698253631592, + 0.04497445374727249, + 0.05340271443128586, + 0.3921034336090088 + ], + "credit_map": [ + 0.005787822883576155, + 0.003729181131348014, + 0.007131462916731834, + 0.004957578610628843, + 0.007409951183944941, + 0.00903245434165001, + 0.006178932264447212, + 0.008495491929352283, + 0.004940972663462162, + 0.0029034523759037256, + 0.007625080645084381, + 0.005279477685689926, + 0.008268307894468307, + 0.008818128146231174, + 0.006156090646982193, + 0.008465762250125408, + 0.013439049944281578, + 0.005714644677937031, + 0.010425514541566372, + 0.006042241584509611, + 0.010042460635304451, + 0.010522464290261269, + 0.009330842643976212, + 0.012249406427145004, + 0.01217009499669075, + 0.005222751758992672, + 0.02214594930410385, + 0.00725469458848238, + 0.011161412112414837, + 0.011519589461386204, + 0.009447717107832432, + 0.014943809248507023, + 0.005533264484256506, + 0.0044872211292386055, + 0.008365853689610958, + 0.007935615256428719, + 0.010071348398923874, + 0.008158284239470959, + 0.00878099538385868, + 0.017087116837501526, + 0.006149989552795887, + 0.003183826105669141, + 0.009270568378269672, + 0.004379698075354099, + 0.004778595175594091, + 0.006500298157334328, + 0.006209647282958031, + 0.007027273066341877, + 0.011577898636460304, + 0.008848407305777073, + 0.008531895466148853, + 0.012142081744968891, + 0.019507035613059998, + 0.01853533647954464, + 0.01480970811098814, + 0.02083461545407772, + 0.00501256762072444, + 0.003544148989021778, + 0.0034595176111906767, + 0.004758565686643124, + 0.007502354681491852, + 0.00812464114278555, + 0.005965194199234247, + 0.006607462652027607, + 0.003384786657989025, + 0.0046179634518921375, + 0.001953753409907222, + 0.003245370928198099, + 0.007953246124088764, + 0.004201256204396486, + 0.013163415715098381, + 0.014882917515933514, + 0.015254690311849117, + 0.011072985827922821, + 0.009740425273776054, + 0.017186390236020088, + 0.029386093840003014, + 0.13881346583366394, + 0.14481860399246216, + 0.02583080530166626 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 6, + 3 + ], + "selected_reward": -0.03369464352726936 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "An extreme close-up, static macro shot, fixed at eye level, focuses on a single eye. Within its deep brown iris, the pupil, sharply constricted by extreme fear, reflects a faint, unidentified light source. Initially, the white of the eye appears clear and moist, a few long, trembling eyelashes casting a faint shadow. Suddenly, from the corner of the eye, a thin thread of blood emerges like a flash of red lightning. With an unsettling speed, it begins to branch and spread frantically, a network of countless scarlet, branching capillaries rapidly eroding the once pristine white sclera. This process is steady and relentless, the network of blood vessels growing denser and denser, evolving from sparse lines to a crisscrossing crimson, ultimately engulfing the entire white of the eyeball, transforming it into a terrifying, biologically menacing blood-red. The entire scene, illuminated by the cold, sharp light, creates a sense of intense physical discomfort and a profoundly visually horrifying atmosphere.", + "seed": 22676, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000006/group_7/index_2712_seed_22676_dp_14_rank_28_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 7, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.007857586047612131, + "vlm_logprob": -0.007857586047612131, + "vlm_yes_no_margin": 6.7, + "vlm_token_credit": -0.007857586047612131, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up focusing on a single eye?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the eye initially appear clear and moist with long trembling eyelashes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a thin thread of blood emerge from the corner of the eye and branch out rapidly?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the blood network spread to engulf the entire white sclera, turning it blood-red?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot static and fixed at eye level throughout the sequence?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.00599787337705493 + ], + "opposite_logprob_tokens": [ + -10.005997657775879 + ], + "logprob_reward": -0.00599787337705493, + "yes_no_margin_reward": 10.0, + "token_credit_reward": -0.00599787337705493, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.4287412464618683, + 0.042940400540828705, + 0.1172713041305542, + 0.07512163370847702, + 0.046091437339782715, + 0.07731626182794571, + 0.05531188100576401, + 0.0333704836666584, + 0.04998740553855896, + 0.07384787499904633 + ], + "credit_map": [ + 0.03814682736992836, + 0.00759574631229043, + 0.13921545445919037, + 0.025545744225382805, + 0.019664287567138672, + 0.03883391618728638, + 0.023812850937247276, + 0.13592645525932312, + 0.004374445416033268, + 0.0018597214948385954, + 0.010480859316885471, + 0.003518949495628476, + 0.0027021830901503563, + 0.004124389961361885, + 0.004972922615706921, + 0.010906935669481754, + 0.008641512133181095, + 0.0023242824245244265, + 0.03148568421602249, + 0.00643336633220315, + 0.007044774945825338, + 0.011136257089674473, + 0.0068955873139202595, + 0.04330984875559807, + 0.00725486408919096, + 0.00232927268370986, + 0.022172439843416214, + 0.00597583781927824, + 0.003410755656659603, + 0.007715531624853611, + 0.005424696486443281, + 0.020838242024183273, + 0.0035148272290825844, + 0.0029837486799806356, + 0.006214261054992676, + 0.009718934074044228, + 0.00361258233897388, + 0.0046331267803907394, + 0.008071037009358406, + 0.007342924363911152, + 0.006347090937197208, + 0.0046242536045610905, + 0.0074374177493155, + 0.019980214536190033, + 0.006746968720108271, + 0.019510501995682716, + 0.006251659709960222, + 0.006418162956833839, + 0.006480323616415262, + 0.004254729021340609, + 0.0049621593207120895, + 0.004976912401616573, + 0.005517683923244476, + 0.009444424882531166, + 0.006051408592611551, + 0.013624249957501888, + 0.004243775736540556, + 0.0033934658858925104, + 0.002438620198518038, + 0.0030090068466961384, + 0.00472851749509573, + 0.005196907091885805, + 0.004135238006711006, + 0.006224954966455698, + 0.005434208083897829, + 0.004485591780394316, + 0.009000788442790508, + 0.0043920790776610374, + 0.007433055900037289, + 0.00607062503695488, + 0.004988412372767925, + 0.008182642981410027, + 0.008447055704891682, + 0.0140748992562294, + 0.005286156199872494, + 0.007016527466475964, + 0.012433221563696861, + 0.008530071005225182, + 0.009865530766546726, + 0.00819441583007574 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 5 + ], + "selected_reward": -0.00599787337705493 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up focusing on a single eye?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the eye initially appear clear and moist with long trembling eyelashes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a thin thread of blood emerge from the corner of the eye and branch out rapidly?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the blood network spread to engulf the entire white sclera, turning it blood-red?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot static and fixed at eye level throughout the sequence?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.00527063338086009 + ], + "opposite_logprob_tokens": [ + -5.255270481109619 + ], + "logprob_reward": -0.00527063338086009, + "yes_no_margin_reward": 5.25, + "token_credit_reward": -0.00527063338086009, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06625837087631226, + 0.15372587740421295, + 0.12236636877059937, + 0.13019950687885284, + 0.21531538665294647, + 0.15275682508945465, + 0.04234731197357178, + 0.030904356390237808, + 0.034280236810445786, + 0.05184570699930191 + ], + "credit_map": [ + 0.005614039488136768, + 0.0027681791689246893, + 0.008664477616548538, + 0.006947027053683996, + 0.0052758632227778435, + 0.006278888322412968, + 0.009175103157758713, + 0.02153479866683483, + 0.01584765501320362, + 0.004505290649831295, + 0.04385383427143097, + 0.011698664166033268, + 0.008002130314707756, + 0.014924285933375359, + 0.010213486850261688, + 0.04468053579330444, + 0.00945553369820118, + 0.002541094087064266, + 0.0336589552462101, + 0.005965321324765682, + 0.0072444104589521885, + 0.012163391336798668, + 0.0076326304115355015, + 0.0437050387263298, + 0.012136410921812057, + 0.0027857222594320774, + 0.038911543786525726, + 0.00855164509266615, + 0.006119607016444206, + 0.012537441216409206, + 0.008546644821763039, + 0.040610503405332565, + 0.025790490210056305, + 0.011544736102223396, + 0.05264740809798241, + 0.03733641281723976, + 0.016657566651701927, + 0.016640322282910347, + 0.02628243714570999, + 0.028416013345122337, + 0.01447073183953762, + 0.004405638203024864, + 0.009705721400678158, + 0.028267377987504005, + 0.008439738303422928, + 0.04462409391999245, + 0.009894202463328838, + 0.03294932842254639, + 0.004226334393024445, + 0.0034550009295344353, + 0.0045329732820391655, + 0.0035713673569262028, + 0.004599587991833687, + 0.007076803129166365, + 0.004820931237190962, + 0.010064315050840378, + 0.0039623696357011795, + 0.0021160405594855547, + 0.003216581651940942, + 0.0028401429299265146, + 0.003207002067938447, + 0.005649150814861059, + 0.004629833623766899, + 0.0052832383662462234, + 0.003787616966292262, + 0.0033924137242138386, + 0.004372315481305122, + 0.0027786639984697104, + 0.0069536250084638596, + 0.004279152024537325, + 0.0051405541598796844, + 0.003575897542759776, + 0.004705651197582483, + 0.006189109291881323, + 0.0035048681311309338, + 0.0055227321572601795, + 0.011293170973658562, + 0.006377135869115591, + 0.007121381349861622, + 0.0071316612884402275 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 1, + 5 + ], + "selected_reward": -0.00527063338086009 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up focusing on a single eye?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the eye initially appear clear and moist with long trembling eyelashes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a thin thread of blood emerge from the corner of the eye and branch out rapidly?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the blood network spread to engulf the entire white sclera, turning it blood-red?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot static and fixed at eye level throughout the sequence?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0003438596613705158 + ], + "opposite_logprob_tokens": [ + -8.000344276428223 + ], + "logprob_reward": -0.0003438596613705158, + "yes_no_margin_reward": 8.0, + "token_credit_reward": -0.0003438596613705158, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08122894167900085, + 0.03947899863123894, + 0.023018304258584976, + 0.12297244369983673, + 0.061521224677562714, + 0.41562747955322266, + 0.10814113914966583, + 0.04993743076920509, + 0.044970955699682236, + 0.053103066980838776 + ], + "credit_map": [ + 0.006854241713881493, + 0.002206486649811268, + 0.023256687447428703, + 0.005417905747890472, + 0.0047458759509027, + 0.00742425536736846, + 0.004897100850939751, + 0.026426389813423157, + 0.0038143000565469265, + 0.0012787791201844811, + 0.011802257969975471, + 0.003402870148420334, + 0.002093878109008074, + 0.003923851530998945, + 0.002805237425491214, + 0.010357821360230446, + 0.0019622573163360357, + 0.001350250095129013, + 0.00490417517721653, + 0.0028385347686707973, + 0.001302960212342441, + 0.0025964672677218914, + 0.001903473399579525, + 0.006160185672342777, + 0.011851122602820396, + 0.0023416541516780853, + 0.039450503885746, + 0.007436725776642561, + 0.005401240661740303, + 0.011519752442836761, + 0.0071389395743608475, + 0.037832506000995636, + 0.00557824457064271, + 0.0033729178830981255, + 0.011762544512748718, + 0.012094598263502121, + 0.004962545819580555, + 0.00530206086114049, + 0.011581288650631905, + 0.006867019925266504, + 0.04831632226705551, + 0.019800683483481407, + 0.06382817775011063, + 0.11115188151597977, + 0.021343812346458435, + 0.0635259598493576, + 0.02781100757420063, + 0.059849631041288376, + 0.011604104191064835, + 0.007998501881957054, + 0.012927158735692501, + 0.009456546977162361, + 0.009121346287429333, + 0.015844225883483887, + 0.01144736260175705, + 0.02974189631640911, + 0.007222128100693226, + 0.004972750786691904, + 0.004032743629068136, + 0.0042031933553516865, + 0.006229782942682505, + 0.007824419997632504, + 0.0057061584666371346, + 0.009746253490447998, + 0.004180018324404955, + 0.0036715841852128506, + 0.0070395153015851974, + 0.00433764886111021, + 0.0074087935499846935, + 0.0068145436234772205, + 0.004366650711745024, + 0.00715220021083951, + 0.0049796393141150475, + 0.005489241797477007, + 0.0048081050626933575, + 0.0056299674324691296, + 0.01074167899787426, + 0.005938839167356491, + 0.008559618145227432, + 0.006955981720238924 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 3, + 6 + ], + "selected_reward": -0.0003438596613705158 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up focusing on a single eye?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the eye initially appear clear and moist with long trembling eyelashes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a thin thread of blood emerge from the corner of the eye and branch out rapidly?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the blood network spread to engulf the entire white sclera, turning it blood-red?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot static and fixed at eye level throughout the sequence?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0013338964199647307 + ], + "opposite_logprob_tokens": [ + -6.626333713531494 + ], + "logprob_reward": -0.0013338964199647307, + "yes_no_margin_reward": 6.625, + "token_credit_reward": -0.0013338964199647307, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.23623788356781006, + 0.028515752404928207, + 0.05087587609887123, + 0.13165752589702606, + 0.062215421348810196, + 0.12307016551494598, + 0.14502200484275818, + 0.060863394290208817, + 0.053249359130859375, + 0.10829268395900726 + ], + "credit_map": [ + 0.02117428369820118, + 0.004748844541609287, + 0.0786830484867096, + 0.014003670774400234, + 0.010256804525852203, + 0.02094271220266819, + 0.013440405949950218, + 0.07298809289932251, + 0.0024373363703489304, + 0.0015600118786096573, + 0.007566908840090036, + 0.003722462570294738, + 0.0017617219127714634, + 0.002766847610473633, + 0.003751958254724741, + 0.004948503337800503, + 0.005144878290593624, + 0.0030407721642404795, + 0.014216854237020016, + 0.003996169660240412, + 0.003562734229490161, + 0.0047973766922950745, + 0.006265748292207718, + 0.009851336479187012, + 0.012535964138805866, + 0.002164545003324747, + 0.04201823100447655, + 0.008122486062347889, + 0.005833802279084921, + 0.012332594022154808, + 0.007102996110916138, + 0.04154689610004425, + 0.0052971988916397095, + 0.004517244175076485, + 0.012542655691504478, + 0.007598270196467638, + 0.006108816247433424, + 0.007010670844465494, + 0.010351026430726051, + 0.008789528161287308, + 0.011011471040546894, + 0.007049704436212778, + 0.010850486345589161, + 0.03975837677717209, + 0.006636687554419041, + 0.01550219301134348, + 0.010527539066970348, + 0.021733691915869713, + 0.013102278113365173, + 0.01988779380917549, + 0.016313455998897552, + 0.009729119017720222, + 0.03266582638025284, + 0.021928224712610245, + 0.01294626109302044, + 0.01844903640449047, + 0.012695888988673687, + 0.004419815726578236, + 0.006548137869685888, + 0.005025732330977917, + 0.0067389062605798244, + 0.008077447302639484, + 0.008779533207416534, + 0.008577925153076649, + 0.005334945395588875, + 0.0038787424564361572, + 0.007083097938448191, + 0.005671870894730091, + 0.006906183436512947, + 0.007267247419804335, + 0.006735061760991812, + 0.010372204706072807, + 0.011275695636868477, + 0.014172232709825039, + 0.0059638358652591705, + 0.01090640015900135, + 0.02268250472843647, + 0.012331989593803883, + 0.015024903230369091, + 0.015935111790895462 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 6, + 3 + ], + "selected_reward": -0.0013338964199647307 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up focusing on a single eye?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the eye initially appear clear and moist with long trembling eyelashes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a thin thread of blood emerge from the corner of the eye and branch out rapidly?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the blood network spread to engulf the entire white sclera, turning it blood-red?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot static and fixed at eye level throughout the sequence?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.026341667398810387 + ], + "opposite_logprob_tokens": [ + -3.651341676712036 + ], + "logprob_reward": -0.026341667398810387, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.026341667398810387, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.20084595680236816, + 0.14395982027053833, + 0.06698790192604065, + 0.07480279356241226, + 0.07607326656579971, + 0.19171929359436035, + 0.05007464811205864, + 0.03517725691199303, + 0.038726843893527985, + 0.12163224071264267 + ], + "credit_map": [ + 0.017274819314479828, + 0.005103984382003546, + 0.05635003373026848, + 0.019773779436945915, + 0.009150435216724873, + 0.01842164620757103, + 0.014417421072721481, + 0.060353834182024, + 0.013014843687415123, + 0.0032342521008104086, + 0.0463101826608181, + 0.011349808424711227, + 0.0069895097985863686, + 0.012909779325127602, + 0.008430778048932552, + 0.041720662266016006, + 0.0059802778996527195, + 0.0021329112350940704, + 0.02029075287282467, + 0.004401199985295534, + 0.0032653366215527058, + 0.006059850100427866, + 0.0045736744068562984, + 0.02028389647603035, + 0.006984308827668428, + 0.0018229469424113631, + 0.023718055337667465, + 0.0053184241987764835, + 0.003381164977326989, + 0.00664689764380455, + 0.004464427474886179, + 0.022466566413640976, + 0.006424324121326208, + 0.0043631913140416145, + 0.013489367440342903, + 0.010764732025563717, + 0.006017141975462437, + 0.006343997083604336, + 0.016496757045388222, + 0.012173754163086414, + 0.019952982664108276, + 0.007386266253888607, + 0.010021466761827469, + 0.013425840996205807, + 0.01180466078221798, + 0.05362942814826965, + 0.014258259907364845, + 0.06124036759138107, + 0.004443827550858259, + 0.004873570054769516, + 0.006022205576300621, + 0.005598628893494606, + 0.007473418023437262, + 0.00619012676179409, + 0.00534167792648077, + 0.010131195187568665, + 0.003723332891240716, + 0.003360395785421133, + 0.0037193610332906246, + 0.002582920715212822, + 0.005851816851645708, + 0.005814759060740471, + 0.0053447093814611435, + 0.004779960494488478, + 0.004337598104029894, + 0.0030878849793225527, + 0.006147668696939945, + 0.0033892723731696606, + 0.005042319651693106, + 0.004851778037846088, + 0.003864309983327985, + 0.008006014861166477, + 0.01342741772532463, + 0.01850709319114685, + 0.0065328204073011875, + 0.011790807358920574, + 0.02621408738195896, + 0.015786854550242424, + 0.017471764236688614, + 0.01190139353275299 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 5, + 1 + ], + "selected_reward": -0.026341667398810387 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "An extreme close-up, static macro shot, fixed at eye level, focuses on a single eye. Within its deep brown iris, the pupil, sharply constricted by extreme fear, reflects a faint, unidentified light source. Initially, the white of the eye appears clear and moist, a few long, trembling eyelashes casting a faint shadow. Suddenly, from the corner of the eye, a thin thread of blood emerges like a flash of red lightning. With an unsettling speed, it begins to branch and spread frantically, a network of countless scarlet, branching capillaries rapidly eroding the once pristine white sclera. This process is steady and relentless, the network of blood vessels growing denser and denser, evolving from sparse lines to a crisscrossing crimson, ultimately engulfing the entire white of the eyeball, transforming it into a terrifying, biologically menacing blood-red. The entire scene, illuminated by the cold, sharp light, creates a sense of intense physical discomfort and a profoundly visually horrifying atmosphere.", + "seed": 22677, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000006/group_7/index_2712_seed_22677_dp_15_rank_30_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 7, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.010979942372068762, + "vlm_logprob": -0.010979942372068762, + "vlm_yes_no_margin": 6.025, + "vlm_token_credit": -0.010979942372068762, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up focusing on a single eye?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the eye initially appear clear and moist with long trembling eyelashes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a thin thread of blood emerge from the corner of the eye and branch out rapidly?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the blood network spread to engulf the entire white sclera, turning it blood-red?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot static and fixed at eye level throughout the sequence?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.003707798197865486 + ], + "opposite_logprob_tokens": [ + -9.378707885742188 + ], + "logprob_reward": -0.003707798197865486, + "yes_no_margin_reward": 9.375, + "token_credit_reward": -0.003707798197865486, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06521112471818924, + 0.16268014907836914, + 0.026527225971221924, + 0.039501581341028214, + 0.03754788637161255, + 0.03572807461023331, + 0.11759719252586365, + 0.4254293441772461, + 0.04575292021036148, + 0.044024500995874405 + ], + "credit_map": [ + 0.007575920317322016, + 0.005893663037568331, + 0.011379610747098923, + 0.011752280406653881, + 0.004213991109281778, + 0.008035018108785152, + 0.008482100442051888, + 0.007878539152443409, + 0.0184188075363636, + 0.007985880598425865, + 0.05661113187670708, + 0.016127383336424828, + 0.008214667439460754, + 0.02250058576464653, + 0.0135651221498847, + 0.01925657130777836, + 0.004084532614797354, + 0.002009978052228689, + 0.004818237852305174, + 0.0034689223393797874, + 0.002480811672285199, + 0.0036920944694429636, + 0.0027606815565377474, + 0.003211966948583722, + 0.006204153876751661, + 0.0030903301667422056, + 0.003847035812214017, + 0.005116940010339022, + 0.0035696704871952534, + 0.004802434705197811, + 0.008072410710155964, + 0.004798602312803268, + 0.0045404089614748955, + 0.003634257009252906, + 0.005161500070244074, + 0.0032539020758122206, + 0.003946727607399225, + 0.005699304863810539, + 0.006326806731522083, + 0.004984978586435318, + 0.0071940310299396515, + 0.0033558851573616266, + 0.003099362598732114, + 0.002509022131562233, + 0.0045844074338674545, + 0.005329517647624016, + 0.004337416496127844, + 0.0053184316493570805, + 0.011091954074800014, + 0.01655324175953865, + 0.008853122591972351, + 0.01813485100865364, + 0.016817962750792503, + 0.014847652986645699, + 0.012847509235143661, + 0.018450886011123657, + 0.05035798251628876, + 0.07734254747629166, + 0.0509113185107708, + 0.04233769699931145, + 0.06501616537570953, + 0.05849538370966911, + 0.044892117381095886, + 0.036076147109270096, + 0.004300135653465986, + 0.007921300828456879, + 0.0029674107208848, + 0.0067746578715741634, + 0.005515580531209707, + 0.008376040495932102, + 0.005483134184032679, + 0.004414656665176153, + 0.003939613234251738, + 0.00567098380997777, + 0.0023251385428011417, + 0.008813486434519291, + 0.004528043791651726, + 0.008512715809047222, + 0.004127779509872198, + 0.006106738932430744 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 1, + 6 + ], + "selected_reward": -0.003707798197865486 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up focusing on a single eye?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the eye initially appear clear and moist with long trembling eyelashes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a thin thread of blood emerge from the corner of the eye and branch out rapidly?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the blood network spread to engulf the entire white sclera, turning it blood-red?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot static and fixed at eye level throughout the sequence?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.011107869446277618 + ], + "opposite_logprob_tokens": [ + -4.511107921600342 + ], + "logprob_reward": -0.011107869446277618, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.011107869446277618, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10437210649251938, + 0.6169665455818176, + 0.026982296258211136, + 0.025469811633229256, + 0.06655262410640717, + 0.030743617564439774, + 0.045444462448358536, + 0.044447824358940125, + 0.02508554421365261, + 0.013935133814811707 + ], + "credit_map": [ + 0.009035213850438595, + 0.010639745742082596, + 0.014506016857922077, + 0.015783766284585, + 0.010138708166778088, + 0.016201557591557503, + 0.012564281933009624, + 0.01550281886011362, + 0.07206978648900986, + 0.030119040980935097, + 0.2091437578201294, + 0.05857642740011215, + 0.029654910787940025, + 0.08921052515506744, + 0.05884610116481781, + 0.06934603303670883, + 0.0023627106565982103, + 0.002212344668805599, + 0.005759368650615215, + 0.004131746478378773, + 0.0024862538557499647, + 0.0034619069192558527, + 0.0032436989713460207, + 0.0033242644276469946, + 0.005038872826844454, + 0.001929774065501988, + 0.0025060363113880157, + 0.0029526727739721537, + 0.0021120894234627485, + 0.003075507003813982, + 0.004677229095250368, + 0.0031776311807334423, + 0.004378196317702532, + 0.0030275932513177395, + 0.006171586457639933, + 0.0035146751906722784, + 0.004421866964548826, + 0.02105160988867283, + 0.00916755385696888, + 0.014819537289440632, + 0.003307493170723319, + 0.0038926969282329082, + 0.0027372711338102818, + 0.0058580683544278145, + 0.002285416005179286, + 0.004169899504631758, + 0.0030703505035489798, + 0.0054224212653934956, + 0.005479430314153433, + 0.0069834087044000626, + 0.002888978458940983, + 0.004222378600388765, + 0.006858406588435173, + 0.008125676773488522, + 0.005304624792188406, + 0.005581555888056755, + 0.007464216556400061, + 0.008032761514186859, + 0.0031163329258561134, + 0.003780726110562682, + 0.006053312681615353, + 0.006567112170159817, + 0.005334651097655296, + 0.0040987110696733, + 0.002201762283220887, + 0.003189442213624716, + 0.001669493387453258, + 0.003306330880150199, + 0.002770406426861882, + 0.007129480596631765, + 0.0023807468824088573, + 0.0024378809612244368, + 0.001365224365144968, + 0.0024405471049249172, + 0.001177433179691434, + 0.002347423927858472, + 0.0013332890812307596, + 0.002173541346564889, + 0.0018283684039488435, + 0.0012693062890321016 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 4 + ], + "selected_reward": -0.011107869446277618 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up focusing on a single eye?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the eye initially appear clear and moist with long trembling eyelashes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a thin thread of blood emerge from the corner of the eye and branch out rapidly?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the blood network spread to engulf the entire white sclera, turning it blood-red?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot static and fixed at eye level throughout the sequence?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0004406196530908346 + ], + "opposite_logprob_tokens": [ + -7.75044059753418 + ], + "logprob_reward": -0.0004406196530908346, + "yes_no_margin_reward": 7.75, + "token_credit_reward": -0.0004406196530908346, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12035246938467026, + 0.22447440028190613, + 0.06879168003797531, + 0.05696402117609978, + 0.124514639377594, + 0.07285255193710327, + 0.11417941004037857, + 0.06756069511175156, + 0.06940677016973495, + 0.08090328425168991 + ], + "credit_map": [ + 0.009426400996744633, + 0.008284899406135082, + 0.018543493002653122, + 0.012816345319151878, + 0.013812691904604435, + 0.025734592229127884, + 0.01308454293757677, + 0.018649518489837646, + 0.026825839653611183, + 0.012444846332073212, + 0.07002517580986023, + 0.023353837430477142, + 0.011832287535071373, + 0.03399796783924103, + 0.021548522636294365, + 0.024445947259664536, + 0.008298143744468689, + 0.0049121929332613945, + 0.01209250558167696, + 0.008087058551609516, + 0.005005750339478254, + 0.0117416400462389, + 0.008734642527997494, + 0.009919759817421436, + 0.008226368576288223, + 0.003667820245027542, + 0.008526323363184929, + 0.00507646519690752, + 0.004180464893579483, + 0.008561077527701855, + 0.011967831291258335, + 0.006757673807442188, + 0.008130851201713085, + 0.006732203997671604, + 0.013514485210180283, + 0.007240528240799904, + 0.006756575312465429, + 0.033053331077098846, + 0.02090921252965927, + 0.028177456930279732, + 0.013625160790979862, + 0.008256824687123299, + 0.0062418668530881405, + 0.009323366917669773, + 0.007307748310267925, + 0.01046318095177412, + 0.010232064872980118, + 0.007402344606816769, + 0.011704609729349613, + 0.01852838322520256, + 0.008798770606517792, + 0.016299499198794365, + 0.018679840490221977, + 0.018859421834349632, + 0.010595879517495632, + 0.01071301568299532, + 0.008524935692548752, + 0.01207717601209879, + 0.007433444261550903, + 0.006091302260756493, + 0.009289226494729519, + 0.009921885095536709, + 0.00767352432012558, + 0.006549205631017685, + 0.006014641840010881, + 0.007168774027377367, + 0.0047881947830319405, + 0.008372482843697071, + 0.006829215679317713, + 0.022593749687075615, + 0.005690286867320538, + 0.007949421182274818, + 0.006726516410708427, + 0.01491544395685196, + 0.005183008499443531, + 0.014415135607123375, + 0.008221707306802273, + 0.008183300495147705, + 0.014682709239423275, + 0.00857547577470541 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 4, + 0 + ], + "selected_reward": -0.0004406196530908346 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up focusing on a single eye?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the eye initially appear clear and moist with long trembling eyelashes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a thin thread of blood emerge from the corner of the eye and branch out rapidly?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the blood network spread to engulf the entire white sclera, turning it blood-red?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot static and fixed at eye level throughout the sequence?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005947629921138287 + ], + "opposite_logprob_tokens": [ + -5.130947589874268 + ], + "logprob_reward": -0.005947629921138287, + "yes_no_margin_reward": 5.125, + "token_credit_reward": -0.005947629921138287, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.04214324429631233, + 0.06172658130526543, + 0.05816614627838135, + 0.09099903702735901, + 0.29830268025398254, + 0.13055957853794098, + 0.11467970907688141, + 0.06474357098340988, + 0.04366661235690117, + 0.09501287341117859 + ], + "credit_map": [ + 0.003290899097919464, + 0.0036534545943140984, + 0.006647251546382904, + 0.00845804251730442, + 0.002904890337958932, + 0.00667283171787858, + 0.004476829431951046, + 0.006039042957127094, + 0.00684450613334775, + 0.0028366229962557554, + 0.021116789430379868, + 0.00591863552108407, + 0.0028431126847863197, + 0.0094609335064888, + 0.00477497698739171, + 0.007931003347039223, + 0.007967941462993622, + 0.004255346953868866, + 0.011530167423188686, + 0.008351124823093414, + 0.004630535840988159, + 0.007882552221417427, + 0.005488197784870863, + 0.008060277439653873, + 0.01581624709069729, + 0.006437003146857023, + 0.01221294142305851, + 0.007061307784169912, + 0.008245362900197506, + 0.015673842281103134, + 0.01482933759689331, + 0.010723000392317772, + 0.017234303057193756, + 0.010342057794332504, + 0.03649373725056648, + 0.01567802205681801, + 0.022828124463558197, + 0.08966992050409317, + 0.038597267121076584, + 0.06745922565460205, + 0.02348768524825573, + 0.01341189444065094, + 0.01692681945860386, + 0.012936003506183624, + 0.013355888426303864, + 0.017701106145977974, + 0.012848662212491035, + 0.019891515374183655, + 0.009902428835630417, + 0.01466988306492567, + 0.009201062843203545, + 0.014873853884637356, + 0.024864105507731438, + 0.017085306346416473, + 0.012193704023957253, + 0.011889362707734108, + 0.008143424987792969, + 0.011104215867817402, + 0.007326263003051281, + 0.007791901472955942, + 0.00726578664034605, + 0.009404166601598263, + 0.00791895855218172, + 0.0057888515293598175, + 0.0041186194866895676, + 0.00604892335832119, + 0.003978088963776827, + 0.005931420251727104, + 0.00455701956525445, + 0.008271776139736176, + 0.006702437065541744, + 0.004058328922837973, + 0.007384893950074911, + 0.015447461046278477, + 0.006166566628962755, + 0.016016246750950813, + 0.01013889443129301, + 0.013783670961856842, + 0.015018203295767307, + 0.011056942865252495 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 5, + 6 + ], + "selected_reward": -0.005947629921138287 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up focusing on a single eye?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the eye initially appear clear and moist with long trembling eyelashes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a thin thread of blood emerge from the corner of the eye and branch out rapidly?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the blood network spread to engulf the entire white sclera, turning it blood-red?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot static and fixed at eye level throughout the sequence?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03369579464197159 + ], + "opposite_logprob_tokens": [ + -3.408695697784424 + ], + "logprob_reward": -0.03369579464197159, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.03369579464197159, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10780742019414902, + 0.11302395164966583, + 0.043371908366680145, + 0.05110981687903404, + 0.04025537520647049, + 0.03182133287191391, + 0.07819877564907074, + 0.46033576130867004, + 0.045969508588314056, + 0.028106169775128365 + ], + "credit_map": [ + 0.00794996228069067, + 0.01244242675602436, + 0.012269571423530579, + 0.020655963569879532, + 0.010016846470534801, + 0.013741392642259598, + 0.0183972530066967, + 0.012334004044532776, + 0.012295679189264774, + 0.0052046142518520355, + 0.03980393335223198, + 0.01039411686360836, + 0.0057859867811203, + 0.014727448113262653, + 0.01136395800858736, + 0.013448220677673817, + 0.006889064330607653, + 0.0028903745114803314, + 0.006965445354580879, + 0.0050643072463572025, + 0.0035287931095808744, + 0.00704334257170558, + 0.00511795561760664, + 0.005872627720236778, + 0.00890660472214222, + 0.004256437532603741, + 0.006521731149405241, + 0.005310078151524067, + 0.003734336234629154, + 0.00904110074043274, + 0.0064460961148142815, + 0.006893437821418047, + 0.0055653005838394165, + 0.00361657259054482, + 0.005255411379039288, + 0.0034016405697911978, + 0.005958513356745243, + 0.004727999679744244, + 0.00817299447953701, + 0.003556944662705064, + 0.004669675603508949, + 0.002810821635648608, + 0.0030159135349094868, + 0.002543415641412139, + 0.004563516471534967, + 0.004723947029560804, + 0.005424268543720245, + 0.004069775342941284, + 0.00887210387736559, + 0.0077296472154557705, + 0.005413975566625595, + 0.008374431170523167, + 0.010373619385063648, + 0.0110549945384264, + 0.010924529284238815, + 0.015455476939678192, + 0.05408845469355583, + 0.08011819422245026, + 0.05888201668858528, + 0.04795156419277191, + 0.0677935928106308, + 0.06147022172808647, + 0.04725191369652748, + 0.04277981445193291, + 0.0049471864476799965, + 0.00689783226698637, + 0.0032893584575504065, + 0.008106459863483906, + 0.004851197823882103, + 0.00830418523401022, + 0.005014415364712477, + 0.004558875225484371, + 0.002671099966391921, + 0.004808347672224045, + 0.0015888033667579293, + 0.004613132681697607, + 0.0031725631561130285, + 0.0030971283558756113, + 0.004205930046737194, + 0.003949164412915707 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 1, + 0 + ], + "selected_reward": -0.03369579464197159 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "An extreme close-up, static macro shot, fixed at eye level, focuses on a single eye. Within its deep brown iris, the pupil, sharply constricted by extreme fear, reflects a faint, unidentified light source. Initially, the white of the eye appears clear and moist, a few long, trembling eyelashes casting a faint shadow. Suddenly, from the corner of the eye, a thin thread of blood emerges like a flash of red lightning. With an unsettling speed, it begins to branch and spread frantically, a network of countless scarlet, branching capillaries rapidly eroding the once pristine white sclera. This process is steady and relentless, the network of blood vessels growing denser and denser, evolving from sparse lines to a crisscrossing crimson, ultimately engulfing the entire white of the eyeball, transforming it into a terrifying, biologically menacing blood-red. The entire scene, illuminated by the cold, sharp light, creates a sense of intense physical discomfort and a profoundly visually horrifying atmosphere.", + "seed": 22678, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000006/group_7/index_2712_seed_22678_dp_15_rank_30_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 7, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.01571867293678224, + "vlm_logprob": -0.01571867293678224, + "vlm_yes_no_margin": 5.775, + "vlm_token_credit": -0.01571867293678224, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up focusing on a single eye?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the eye initially appear clear and moist with long trembling eyelashes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a thin thread of blood emerge from the corner of the eye and branch out rapidly?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the blood network spread to engulf the entire white sclera, turning it blood-red?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot static and fixed at eye level throughout the sequence?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004742089658975601 + ], + "opposite_logprob_tokens": [ + -9.254741668701172 + ], + "logprob_reward": -0.004742089658975601, + "yes_no_margin_reward": 9.25, + "token_credit_reward": -0.004742089658975601, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.05398149788379669, + 0.06412417441606522, + 0.06554745137691498, + 0.09293931722640991, + 0.0417168065905571, + 0.2555055320262909, + 0.06373101472854614, + 0.08561941236257553, + 0.12363780289888382, + 0.1531970202922821 + ], + "credit_map": [ + 0.009022623300552368, + 0.004213461186736822, + 0.009476255625486374, + 0.0055764950811862946, + 0.004209598060697317, + 0.007562112063169479, + 0.00666479067876935, + 0.007256161421537399, + 0.006377999670803547, + 0.0034067677333950996, + 0.007446395233273506, + 0.005107908509671688, + 0.0058614760637283325, + 0.017869511619210243, + 0.00709410710260272, + 0.010960012674331665, + 0.006095858756452799, + 0.003302690340206027, + 0.00924063939601183, + 0.0036747169215232134, + 0.005483060609549284, + 0.018734987825155258, + 0.008337912149727345, + 0.010677591897547245, + 0.006340282037854195, + 0.002694802125915885, + 0.009915060363709927, + 0.004789820872247219, + 0.006866296753287315, + 0.017336545512080193, + 0.02256224863231182, + 0.022434256970882416, + 0.00419282354414463, + 0.0019237310625612736, + 0.007059686351567507, + 0.0030772481113672256, + 0.003491712501272559, + 0.006650164257735014, + 0.007683359086513519, + 0.007638080976903439, + 0.06881175935268402, + 0.010217491537332535, + 0.11339082568883896, + 0.018382716923952103, + 0.010304825380444527, + 0.013731931336224079, + 0.009242895990610123, + 0.011423101648688316, + 0.01044473797082901, + 0.008425439707934856, + 0.00898914597928524, + 0.010131251066923141, + 0.0076066432520747185, + 0.004084228537976742, + 0.006874121259897947, + 0.0071754478849470615, + 0.013284039683640003, + 0.009122993797063828, + 0.011304697021842003, + 0.015524856746196747, + 0.005986867006868124, + 0.007644359953701496, + 0.009718798100948334, + 0.013032795861363411, + 0.020423879846930504, + 0.01814214512705803, + 0.015052735805511475, + 0.012064672075212002, + 0.01503364834934473, + 0.01198667287826538, + 0.016597984358668327, + 0.014336061663925648, + 0.02318291738629341, + 0.024859581142663956, + 0.0124208340421319, + 0.026476074010133743, + 0.015662146732211113, + 0.013056996278464794, + 0.016955288127064705, + 0.020583178848028183 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 9, + 8 + ], + "selected_reward": -0.004742089658975601 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up focusing on a single eye?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the eye initially appear clear and moist with long trembling eyelashes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a thin thread of blood emerge from the corner of the eye and branch out rapidly?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the blood network spread to engulf the entire white sclera, turning it blood-red?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot static and fixed at eye level throughout the sequence?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.016106843948364258 + ], + "opposite_logprob_tokens": [ + -4.141106605529785 + ], + "logprob_reward": -0.016106843948364258, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.016106843948364258, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09677959978580475, + 0.14921090006828308, + 0.20651721954345703, + 0.06340464949607849, + 0.025461265817284584, + 0.14493489265441895, + 0.05042525380849838, + 0.07313722372055054, + 0.08140046149492264, + 0.10872851312160492 + ], + "credit_map": [ + 0.01330804917961359, + 0.006669749040156603, + 0.012681064195930958, + 0.008564166724681854, + 0.008799857459962368, + 0.02075771614909172, + 0.013761099427938461, + 0.01223789807409048, + 0.021655796095728874, + 0.009114278480410576, + 0.02601533755660057, + 0.011138509027659893, + 0.014620327390730381, + 0.027319028973579407, + 0.018819743767380714, + 0.020527873188257217, + 0.029402563348412514, + 0.011415902525186539, + 0.03699871897697449, + 0.018398428335785866, + 0.016380829736590385, + 0.041687123477458954, + 0.025072719901800156, + 0.027160944417119026, + 0.004075544886291027, + 0.002048810012638569, + 0.007193379569798708, + 0.003458610735833645, + 0.006060910876840353, + 0.011709761805832386, + 0.015343154780566692, + 0.013514475896954536, + 0.0032034243922680616, + 0.0014488582964986563, + 0.004669380374252796, + 0.0018055734690278769, + 0.00234409817494452, + 0.0038083114195615053, + 0.0040923673659563065, + 0.004089250694960356, + 0.03975615277886391, + 0.006799611262977123, + 0.06198396533727646, + 0.01068758126348257, + 0.006588524207472801, + 0.008208508603274822, + 0.005130636505782604, + 0.005779930390417576, + 0.008331689983606339, + 0.007129752542823553, + 0.005578846670687199, + 0.007440867368131876, + 0.0054216571152210236, + 0.0049905176274478436, + 0.004877001047134399, + 0.006654927507042885, + 0.022144267335534096, + 0.008454347960650921, + 0.012477440759539604, + 0.008772841654717922, + 0.0039137485437095165, + 0.00416971230879426, + 0.005655234679579735, + 0.007549629546701908, + 0.016700629144906998, + 0.012084558606147766, + 0.007475003600120544, + 0.007728128228336573, + 0.007036044728010893, + 0.0070199365727603436, + 0.007941320538520813, + 0.01541484147310257, + 0.029210776090621948, + 0.013222265057265759, + 0.010231856256723404, + 0.013492905534803867, + 0.010370180010795593, + 0.005589951761066914, + 0.0075820391066372395, + 0.01902853697538376 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 1, + 5 + ], + "selected_reward": -0.016106843948364258 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up focusing on a single eye?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the eye initially appear clear and moist with long trembling eyelashes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a thin thread of blood emerge from the corner of the eye and branch out rapidly?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the blood network spread to engulf the entire white sclera, turning it blood-red?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot static and fixed at eye level throughout the sequence?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0004406196530908346 + ], + "opposite_logprob_tokens": [ + -7.75044059753418 + ], + "logprob_reward": -0.0004406196530908346, + "yes_no_margin_reward": 7.75, + "token_credit_reward": -0.0004406196530908346, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.037162818014621735, + 0.04170360788702965, + 0.06385014206171036, + 0.08271889388561249, + 0.06560288369655609, + 0.12197640538215637, + 0.08164114505052567, + 0.2869858741760254, + 0.15348714590072632, + 0.06487110257148743 + ], + "credit_map": [ + 0.004514896310865879, + 0.002846467075869441, + 0.005309829022735357, + 0.004836116451770067, + 0.0033726415131241083, + 0.006835986860096455, + 0.004028689116239548, + 0.005418192595243454, + 0.004851136356592178, + 0.002480636816471815, + 0.006636872887611389, + 0.0029182136058807373, + 0.004178779665380716, + 0.008083339780569077, + 0.006832502782344818, + 0.005722124595195055, + 0.008923163637518883, + 0.0036257971078157425, + 0.011337749660015106, + 0.005692129954695702, + 0.0052073742263019085, + 0.012693914584815502, + 0.008200844749808311, + 0.008169162087142467, + 0.007102692034095526, + 0.0034178677015006542, + 0.009071153588593006, + 0.004225573968142271, + 0.007097476627677679, + 0.014377414248883724, + 0.018114598467946053, + 0.019312115386128426, + 0.005946867633610964, + 0.0026915569324046373, + 0.008381837978959084, + 0.0043425001204013824, + 0.006563194096088409, + 0.01045627798885107, + 0.01467179786413908, + 0.012548853643238544, + 0.021527094766497612, + 0.006940142251551151, + 0.03664977848529816, + 0.013185949064791203, + 0.008413600735366344, + 0.01226236391812563, + 0.012253079563379288, + 0.010744399391114712, + 0.013348991982638836, + 0.009507639333605766, + 0.011804932728409767, + 0.015064300037920475, + 0.006324708461761475, + 0.006749941501766443, + 0.008605633862316608, + 0.010234998539090157, + 0.1070120707154274, + 0.01733660139143467, + 0.043054960668087006, + 0.016054661944508553, + 0.010842598043382168, + 0.014830413274466991, + 0.0158316008746624, + 0.06202298030257225, + 0.02732057124376297, + 0.01937771774828434, + 0.02152758277952671, + 0.015868889167904854, + 0.015355855226516724, + 0.016381684690713882, + 0.022320304065942764, + 0.015334527008235455, + 0.009408339858055115, + 0.005142276640981436, + 0.01019730232656002, + 0.01772509515285492, + 0.004714637063443661, + 0.00533225666731596, + 0.004941313527524471, + 0.007409887388348579 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 8, + 5 + ], + "selected_reward": -0.0004406196530908346 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up focusing on a single eye?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the eye initially appear clear and moist with long trembling eyelashes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a thin thread of blood emerge from the corner of the eye and branch out rapidly?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the blood network spread to engulf the entire white sclera, turning it blood-red?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot static and fixed at eye level throughout the sequence?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.008640403859317303 + ], + "opposite_logprob_tokens": [ + -4.758640289306641 + ], + "logprob_reward": -0.008640403859317303, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.008640403859317303, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.045840002596378326, + 0.011021214537322521, + 0.029066428542137146, + 0.049503009766340256, + 0.027709806337952614, + 0.24943894147872925, + 0.12233050167560577, + 0.09415193647146225, + 0.11164877563714981, + 0.25928938388824463 + ], + "credit_map": [ + 0.005120330024510622, + 0.0030002437997609377, + 0.005980658810585737, + 0.0035374262370169163, + 0.003832465736195445, + 0.012641936540603638, + 0.003906614612787962, + 0.007820322178304195, + 0.001416141982190311, + 0.0008049110765568912, + 0.00209934264421463, + 0.0011195234255865216, + 0.0010783238103613257, + 0.001821526326239109, + 0.0013086677063256502, + 0.001372777158394456, + 0.002961161080747843, + 0.001468564965762198, + 0.004358218051493168, + 0.0017663101898506284, + 0.0026962957344949245, + 0.00714369909837842, + 0.004988820757716894, + 0.003683355636894703, + 0.003353941487148404, + 0.0019140207441523671, + 0.005390028469264507, + 0.002884587738662958, + 0.0029333308339118958, + 0.009479315020143986, + 0.011644739657640457, + 0.011903047561645508, + 0.0025777071714401245, + 0.0015618911711499095, + 0.003980719018727541, + 0.0025076766032725573, + 0.0024734975304454565, + 0.004056221805512905, + 0.005011802539229393, + 0.00554029131308198, + 0.06049608811736107, + 0.011602331884205341, + 0.10200679302215576, + 0.020121973007917404, + 0.012350843288004398, + 0.015032133087515831, + 0.013890441507101059, + 0.013938331045210361, + 0.018144860863685608, + 0.016079865396022797, + 0.016124049201607704, + 0.01877063699066639, + 0.012026594020426273, + 0.010122441686689854, + 0.014741110615432262, + 0.01632094196975231, + 0.015527559444308281, + 0.010813257656991482, + 0.0140727199614048, + 0.015134484507143497, + 0.006830369587987661, + 0.007968884892761707, + 0.009127230383455753, + 0.014677431434392929, + 0.020112251862883568, + 0.014318951405584812, + 0.012961039319634438, + 0.01052828785032034, + 0.010558062233030796, + 0.008826221339404583, + 0.012017345055937767, + 0.022326616570353508, + 0.05129656195640564, + 0.038166336715221405, + 0.024549027904868126, + 0.050316330045461655, + 0.0286346934735775, + 0.019619591534137726, + 0.016640901565551758, + 0.030065948143601418 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 5, + 6 + ], + "selected_reward": -0.008640403859317303 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up focusing on a single eye?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the eye initially appear clear and moist with long trembling eyelashes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a thin thread of blood emerge from the corner of the eye and branch out rapidly?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the blood network spread to engulf the entire white sclera, turning it blood-red?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot static and fixed at eye level throughout the sequence?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04866340756416321 + ], + "opposite_logprob_tokens": [ + -3.048663377761841 + ], + "logprob_reward": -0.04866340756416321, + "yes_no_margin_reward": 3.0, + "token_credit_reward": -0.04866340756416321, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08834481984376907, + 0.027760768309235573, + 0.06779011338949203, + 0.039889052510261536, + 0.04509232938289642, + 0.5311663150787354, + 0.09070876240730286, + 0.019923517480492592, + 0.054399408400058746, + 0.03492496907711029 + ], + "credit_map": [ + 0.010268539190292358, + 0.0061339158564805984, + 0.009955571964383125, + 0.008837472647428513, + 0.007461532484740019, + 0.020902274176478386, + 0.00921009760349989, + 0.015575405210256577, + 0.002681311685591936, + 0.0021806119475513697, + 0.00403167586773634, + 0.0025600637309253216, + 0.00280807469971478, + 0.005677602253854275, + 0.00481658149510622, + 0.0030048436019569635, + 0.009145981632173061, + 0.004705899395048618, + 0.00784964021295309, + 0.005845880601555109, + 0.005798359401524067, + 0.017273200675845146, + 0.005128029268234968, + 0.012043118476867676, + 0.0027037211693823338, + 0.0014716353034600616, + 0.005694414023309946, + 0.0027294072788208723, + 0.0021949252113699913, + 0.006732224021106958, + 0.009573310613632202, + 0.008789410814642906, + 0.00501207634806633, + 0.002116759540513158, + 0.007421946618705988, + 0.003010680666193366, + 0.00448819762095809, + 0.006930953823029995, + 0.008822319097816944, + 0.007289395201951265, + 0.14881186187267303, + 0.023105399683117867, + 0.2346811443567276, + 0.03942425549030304, + 0.022020645439624786, + 0.027327295392751694, + 0.01599217765033245, + 0.019803486764431, + 0.012545428238809109, + 0.012739514000713825, + 0.013176839798688889, + 0.013408243656158447, + 0.011193172074854374, + 0.006852890830487013, + 0.010642612352967262, + 0.010150063782930374, + 0.005287136882543564, + 0.0020584501326084137, + 0.0023816986940801144, + 0.0021723248064517975, + 0.001178827602416277, + 0.0015020023565739393, + 0.0014278108719736338, + 0.0039152647368609905, + 0.013551326468586922, + 0.009778821840882301, + 0.00433340622112155, + 0.005788485985249281, + 0.004161499440670013, + 0.0033549750223755836, + 0.004003455862402916, + 0.009427434764802456, + 0.005011543165892363, + 0.004373664036393166, + 0.00443314341828227, + 0.004312073811888695, + 0.004380065016448498, + 0.004012532066553831, + 0.004278822802007198, + 0.004123121965676546 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 6, + 0 + ], + "selected_reward": -0.04866340756416321 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000006/group_8/index_tensor([3529, 3529])_group_8_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000006/group_8/index_tensor([3529, 3529])_group_8_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..85656103512b23964e262b3b3e469099223c0f7d --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000006/group_8/index_tensor([3529, 3529])_group_8_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "low-angle, medium shot, filmed in extreme slow motion, slowly and steadily zooms in on a quiet corridor deep within the museum. A narrow, soft beam of light slants down from a high side window, illuminating the area perfectly, creating a dramatic, Rembrandt-esque effect of light and shadow. The focal point of the image is a silver tabby British shorthair cat, clad in a handmade, dark brown leather harness, its brass buckle gleaming discreetly in the dim light. Its body is tense, its back slightly arched, its ears pressed back, and its amber eyes wide, their pupils narrowed to a thin line of intense alert. It faces a life-size, white marble statue of a classical woman, its surface etched with the marks of time, its posture serene and timeless. The cat's mouth snaps open, baring sharp teeth. At the climax of a warning hiss, an invisible wave of sound erupts from its mouth, instantly stirring the ancient dust still within the beam. In extreme slow motion, every tiny dust particle shaken by the sound waves was ignited by the light beam, like countless suspended miniature stars, flashing an indescribable forbidden light mixed with gold and lavender, injecting a unique atmosphere of sacredness, tension and supernatural mystery into the confrontation that was about to break out in this silent space.", + "seed": 31692, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000006/group_8/index_3529_seed_31692_dp_16_rank_32_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 8, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.3359786123037338, + "vlm_logprob": -0.3359786123037338, + "vlm_yes_no_margin": 1.4, + "vlm_token_credit": -0.3359786123037338, + "vlm_yes_ratio": 0.8, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a quiet corridor deep within a museum featuring a silver tabby British shorthair cat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the cat wearing a handmade dark brown leather harness with a gleaming brass buckle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is there a life-size white marble statue of a classical woman facing the cat in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the cat snap its mouth open to bare sharp teeth and hiss at the statue?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do dust particles in the light beam ignite and flash with gold and lavender light in extreme slow motion?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -0.8288387656211853 + ], + "opposite_logprob_tokens": [ + -0.5788387656211853 + ], + "logprob_reward": -0.8288387656211853, + "yes_no_margin_reward": -0.25, + "token_credit_reward": -0.8288387656211853, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.12584169209003448, + 0.07049292325973511, + 0.261334627866745, + 0.12931424379348755, + 0.05667719617486, + 0.07693449407815933, + 0.033018577843904495, + 0.1602828949689865, + 0.06343196332454681, + 0.022671353071928024 + ], + "credit_map": [ + 0.015154226683080196, + 0.017850955948233604, + 0.010873319581151009, + 0.018856175243854523, + 0.019680604338645935, + 0.01902010850608349, + 0.011989821679890156, + 0.01241647731512785, + 0.01014089584350586, + 0.009943329729139805, + 0.006358925718814135, + 0.007136596832424402, + 0.013324206694960594, + 0.007469783537089825, + 0.006536067929118872, + 0.009583119302988052, + 0.055210620164871216, + 0.02840598113834858, + 0.06260956823825836, + 0.02676527574658394, + 0.022637831047177315, + 0.023828843608498573, + 0.023327725008130074, + 0.018548788502812386, + 0.026859063655138016, + 0.012287391349673271, + 0.033807262778282166, + 0.012028270401060581, + 0.014181002974510193, + 0.012278434820473194, + 0.00997859239578247, + 0.007894213311374187, + 0.004144761711359024, + 0.005490988027304411, + 0.004702973645180464, + 0.00678192637860775, + 0.013287996873259544, + 0.012059454806149006, + 0.004094131290912628, + 0.006114959716796875, + 0.006323040928691626, + 0.009958724491298199, + 0.008130800910294056, + 0.01366768591105938, + 0.013245778158307076, + 0.008404108695685863, + 0.008953767828643322, + 0.008250592276453972, + 0.0036199241876602173, + 0.006176329683512449, + 0.004202742129564285, + 0.0037162024527788162, + 0.0033235575538128614, + 0.004254836123436689, + 0.003700980683788657, + 0.004024006426334381, + 0.01547960564494133, + 0.01724991761147976, + 0.01772666536271572, + 0.016640212386846542, + 0.02333391085267067, + 0.03313138335943222, + 0.012536080554127693, + 0.024185124784708023, + 0.0066101206466555595, + 0.007125386968255043, + 0.007423151284456253, + 0.007897761650383472, + 0.008904117159545422, + 0.010949008166790009, + 0.00639419024810195, + 0.00812822487205267, + 0.002970887813717127, + 0.0028260466642677784, + 0.0036709595005959272, + 0.002000196138396859, + 0.002724379301071167, + 0.0032024881802499294, + 0.002621784107759595, + 0.002654612297192216 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 7, + 3 + ], + "selected_reward": -0.8288387656211853 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a quiet corridor deep within a museum featuring a silver tabby British shorthair cat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the cat wearing a handmade dark brown leather harness with a gleaming brass buckle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is there a life-size white marble statue of a classical woman facing the cat in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the cat snap its mouth open to bare sharp teeth and hiss at the statue?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do dust particles in the light beam ignite and flash with gold and lavender light in extreme slow motion?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.11301209777593613 + ], + "opposite_logprob_tokens": [ + -2.2380120754241943 + ], + "logprob_reward": -0.11301209777593613, + "yes_no_margin_reward": 2.125, + "token_credit_reward": -0.11301209777593613, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.22048258781433105, + 0.03349393233656883, + 0.1019824892282486, + 0.0666346400976181, + 0.13528740406036377, + 0.15424662828445435, + 0.029258884489536285, + 0.16233816742897034, + 0.0701255202293396, + 0.026149777695536613 + ], + "credit_map": [ + 0.04798687994480133, + 0.011589458212256432, + 0.012047921307384968, + 0.015055162832140923, + 0.01612270250916481, + 0.028736460953950882, + 0.013484423980116844, + 0.07545958459377289, + 0.005579544231295586, + 0.002291717566549778, + 0.002213261788710952, + 0.0035298573784530163, + 0.004535685293376446, + 0.006075914483517408, + 0.0035757545847445726, + 0.005692197475582361, + 0.020297735929489136, + 0.010590844787657261, + 0.015075529925525188, + 0.011010607704520226, + 0.00953433196991682, + 0.012328214012086391, + 0.01162792183458805, + 0.011517288163304329, + 0.006575527135282755, + 0.006107860244810581, + 0.0118246516212821, + 0.007239683531224728, + 0.005928155034780502, + 0.009087183512747288, + 0.0052061728201806545, + 0.014665408059954643, + 0.010876618325710297, + 0.005363421980291605, + 0.0052611129358410835, + 0.007180343382060528, + 0.0393032543361187, + 0.02210688777267933, + 0.006576673127710819, + 0.03861909732222557, + 0.009525484405457973, + 0.015060634352266788, + 0.012268364429473877, + 0.02966185100376606, + 0.019370010122656822, + 0.023239219561219215, + 0.027676401659846306, + 0.017444655299186707, + 0.003205434652045369, + 0.003161659464240074, + 0.0030958354473114014, + 0.0037089933175593615, + 0.003649764461442828, + 0.004085205961018801, + 0.004338821396231651, + 0.0040131681598722935, + 0.012650247663259506, + 0.016978513449430466, + 0.01714915968477726, + 0.018995020538568497, + 0.023232189938426018, + 0.03543619439005852, + 0.01193701010197401, + 0.02595982700586319, + 0.006385871209204197, + 0.007969665341079235, + 0.00876394473016262, + 0.008497348055243492, + 0.009905804879963398, + 0.012660348787903786, + 0.006559031084179878, + 0.00938351172953844, + 0.0025079238694161177, + 0.004087925422936678, + 0.0027021323330700397, + 0.003469439223408699, + 0.0025530641432851553, + 0.003282929537817836, + 0.003145861206576228, + 0.0044005014933645725 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 7, + 5 + ], + "selected_reward": -0.11301209777593613 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a quiet corridor deep within a museum featuring a silver tabby British shorthair cat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the cat wearing a handmade dark brown leather harness with a gleaming brass buckle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is there a life-size white marble statue of a classical woman facing the cat in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the cat snap its mouth open to bare sharp teeth and hiss at the statue?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do dust particles in the light beam ignite and flash with gold and lavender light in extreme slow motion?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.22558416426181793 + ], + "opposite_logprob_tokens": [ + -1.6005841493606567 + ], + "logprob_reward": -0.22558416426181793, + "yes_no_margin_reward": 1.375, + "token_credit_reward": -0.22558416426181793, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14383457601070404, + 0.038263529539108276, + 0.043857939541339874, + 0.11276737600564957, + 0.3096117079257965, + 0.06774459034204483, + 0.037188656628131866, + 0.13553430140018463, + 0.08866407722234726, + 0.02253320813179016 + ], + "credit_map": [ + 0.01586407609283924, + 0.017888497561216354, + 0.011475816369056702, + 0.024729875847697258, + 0.02146225795149803, + 0.022854337468743324, + 0.015863478183746338, + 0.013696247711777687, + 0.006873163860291243, + 0.003357531735673547, + 0.0033056274987757206, + 0.004400153644382954, + 0.005543655715882778, + 0.004194082226604223, + 0.004135691095143557, + 0.006453621666878462, + 0.006547721102833748, + 0.005036610644310713, + 0.007886235602200031, + 0.005210802424699068, + 0.004507523495703936, + 0.0052579729817807674, + 0.005476189311593771, + 0.003934883046895266, + 0.013038466684520245, + 0.011777865700423717, + 0.02211741916835308, + 0.016813624650239944, + 0.011451827362179756, + 0.013540156185626984, + 0.011114630848169327, + 0.012913379818201065, + 0.02785024605691433, + 0.016407329589128494, + 0.01493073534220457, + 0.012087395414710045, + 0.12005165219306946, + 0.07641154527664185, + 0.015545011498034, + 0.026327811181545258, + 0.004804463591426611, + 0.009674251079559326, + 0.006490606814622879, + 0.014293390326201916, + 0.008268674835562706, + 0.008159483782947063, + 0.00920441560447216, + 0.00684930756688118, + 0.004367887042462826, + 0.003206710098311305, + 0.00370836746878922, + 0.003242242382839322, + 0.004110909067094326, + 0.007496408186852932, + 0.004767193924635649, + 0.006288941018283367, + 0.011283811181783676, + 0.014721600338816643, + 0.014590727165341377, + 0.012842833995819092, + 0.01856035366654396, + 0.0318751260638237, + 0.009439275600016117, + 0.02222057804465294, + 0.010952310636639595, + 0.010872016660869122, + 0.012410210445523262, + 0.007633335888385773, + 0.00921973679214716, + 0.012896467000246048, + 0.008534722961485386, + 0.016145285218954086, + 0.002205506432801485, + 0.004412779118865728, + 0.0027057176921516657, + 0.00238957186229527, + 0.002777187619358301, + 0.002611241303384304, + 0.0024529285728931427, + 0.002978275064378977 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 0, + 7 + ], + "selected_reward": -0.22558416426181793 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a quiet corridor deep within a museum featuring a silver tabby British shorthair cat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the cat wearing a handmade dark brown leather harness with a gleaming brass buckle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is there a life-size white marble statue of a classical woman facing the cat in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the cat snap its mouth open to bare sharp teeth and hiss at the statue?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do dust particles in the light beam ignite and flash with gold and lavender light in extreme slow motion?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.47431930899620056 + ], + "opposite_logprob_tokens": [ + -0.974319338798523 + ], + "logprob_reward": -0.47431930899620056, + "yes_no_margin_reward": 0.5, + "token_credit_reward": -0.47431930899620056, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12070298194885254, + 0.054405003786087036, + 0.11944305151700974, + 0.08523225039243698, + 0.1900137960910797, + 0.26429787278175354, + 0.04012181982398033, + 0.07140414416790009, + 0.02788357064127922, + 0.026495572179555893 + ], + "credit_map": [ + 0.028603894636034966, + 0.004220892209559679, + 0.005089869722723961, + 0.006123824045062065, + 0.005378807429224253, + 0.018358450382947922, + 0.005321216769516468, + 0.04760602116584778, + 0.009117022156715393, + 0.006345054134726524, + 0.005568834953010082, + 0.006026691757142544, + 0.00807236135005951, + 0.005691230297088623, + 0.00558066600933671, + 0.00800313800573349, + 0.024418797343969345, + 0.011473320424556732, + 0.029268473386764526, + 0.015322417952120304, + 0.011944391764700413, + 0.009807281196117401, + 0.009087350219488144, + 0.008121003396809101, + 0.009588728658854961, + 0.011340074241161346, + 0.019047221168875694, + 0.009544340893626213, + 0.011163675226271152, + 0.009126476012170315, + 0.006863763555884361, + 0.008557968772947788, + 0.017246445640921593, + 0.007426847703754902, + 0.011921158991754055, + 0.008632785640656948, + 0.06516779214143753, + 0.021472061052918434, + 0.008994597010314465, + 0.04915211349725723, + 0.016869299113750458, + 0.029669415205717087, + 0.022700292989611626, + 0.05360386148095131, + 0.03505396097898483, + 0.035830236971378326, + 0.04193355515599251, + 0.028637241572141647, + 0.003861272009089589, + 0.006254810839891434, + 0.004940514452755451, + 0.006921869236975908, + 0.005404086317867041, + 0.004345709457993507, + 0.004115123767405748, + 0.004278429783880711, + 0.005756422411650419, + 0.006083923391997814, + 0.007329461630433798, + 0.008316329680383205, + 0.009633269160985947, + 0.017275353893637657, + 0.004823252093046904, + 0.012186123989522457, + 0.003048603655770421, + 0.003059565555304289, + 0.0035480584483593702, + 0.0034421966411173344, + 0.0037429961375892162, + 0.002840332454070449, + 0.003506735432893038, + 0.004695083014667034, + 0.0019631795585155487, + 0.003814344061538577, + 0.0025862916372716427, + 0.00289331772364676, + 0.0037221715319901705, + 0.0035397338215261698, + 0.0036079094279557467, + 0.004368622321635485 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 4, + 0 + ], + "selected_reward": -0.47431930899620056 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a quiet corridor deep within a museum featuring a silver tabby British shorthair cat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the cat wearing a handmade dark brown leather harness with a gleaming brass buckle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is there a life-size white marble statue of a classical woman facing the cat in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the cat snap its mouth open to bare sharp teeth and hiss at the statue?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do dust particles in the light beam ignite and flash with gold and lavender light in extreme slow motion?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.038138724863529205 + ], + "opposite_logprob_tokens": [ + -3.2881386280059814 + ], + "logprob_reward": -0.038138724863529205, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.038138724863529205, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.044475555419921875, + 0.03520051762461662, + 0.2239340841770172, + 0.07741574943065643, + 0.23419249057769775, + 0.04189612343907356, + 0.0558394156396389, + 0.1983443796634674, + 0.04413631930947304, + 0.044565364718437195 + ], + "credit_map": [ + 0.006503429263830185, + 0.003129965392872691, + 0.002755650319159031, + 0.005628522951155901, + 0.004646016750484705, + 0.006681487429887056, + 0.00577642722055316, + 0.009354052133858204, + 0.003576992778107524, + 0.0037562530487775803, + 0.003338507143780589, + 0.005319242365658283, + 0.004651300609111786, + 0.004651504568755627, + 0.004704347345978022, + 0.005202371161431074, + 0.04663735255599022, + 0.025926273316144943, + 0.04673944041132927, + 0.02269650436937809, + 0.020705627277493477, + 0.021731054410338402, + 0.02126973681151867, + 0.01822809688746929, + 0.0076561556197702885, + 0.007478443905711174, + 0.01407136581838131, + 0.010974026285111904, + 0.007827705703675747, + 0.009431191720068455, + 0.007548558991402388, + 0.012428304180502892, + 0.023963013663887978, + 0.009185869246721268, + 0.008574215695261955, + 0.009453529492020607, + 0.03558479994535446, + 0.0502745620906353, + 0.008538656868040562, + 0.08861783891916275, + 0.003351589199155569, + 0.0038085314445197582, + 0.004716614726930857, + 0.005900197196751833, + 0.007666498888283968, + 0.005794639699161053, + 0.005604172591120005, + 0.005053879227489233, + 0.004264951217919588, + 0.006094454322010279, + 0.004978279583156109, + 0.006154934410005808, + 0.007886799052357674, + 0.01412882748991251, + 0.006277499720454216, + 0.0060536679811775684, + 0.01493191160261631, + 0.018172750249505043, + 0.01896873489022255, + 0.028034325689077377, + 0.028456388041377068, + 0.04111295938491821, + 0.016564389690756798, + 0.03210292384028435, + 0.006703944411128759, + 0.005608555860817432, + 0.005066452547907829, + 0.005250541027635336, + 0.005938750691711903, + 0.005278319586068392, + 0.004564909730106592, + 0.005724848713725805, + 0.0048510851338505745, + 0.004787090700119734, + 0.006922291126102209, + 0.006309304852038622, + 0.004371167626231909, + 0.006389481481164694, + 0.004924706183373928, + 0.006010239943861961 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 2, + 7 + ], + "selected_reward": -0.038138724863529205 + } + ], + "vlm_decoded_outputs": [ + "No", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "low-angle, medium shot, filmed in extreme slow motion, slowly and steadily zooms in on a quiet corridor deep within the museum. A narrow, soft beam of light slants down from a high side window, illuminating the area perfectly, creating a dramatic, Rembrandt-esque effect of light and shadow. The focal point of the image is a silver tabby British shorthair cat, clad in a handmade, dark brown leather harness, its brass buckle gleaming discreetly in the dim light. Its body is tense, its back slightly arched, its ears pressed back, and its amber eyes wide, their pupils narrowed to a thin line of intense alert. It faces a life-size, white marble statue of a classical woman, its surface etched with the marks of time, its posture serene and timeless. The cat's mouth snaps open, baring sharp teeth. At the climax of a warning hiss, an invisible wave of sound erupts from its mouth, instantly stirring the ancient dust still within the beam. In extreme slow motion, every tiny dust particle shaken by the sound waves was ignited by the light beam, like countless suspended miniature stars, flashing an indescribable forbidden light mixed with gold and lavender, injecting a unique atmosphere of sacredness, tension and supernatural mystery into the confrontation that was about to break out in this silent space.", + "seed": 31693, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000006/group_8/index_3529_seed_31693_dp_16_rank_32_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 8, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.09379382766783237, + "vlm_logprob": -0.09379382766783237, + "vlm_yes_no_margin": 2.799999976158142, + "vlm_token_credit": -0.09379382766783237, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a quiet corridor deep within a museum featuring a silver tabby British shorthair cat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the cat wearing a handmade dark brown leather harness with a gleaming brass buckle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is there a life-size white marble statue of a classical woman facing the cat in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the cat snap its mouth open to bare sharp teeth and hiss at the statue?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do dust particles in the light beam ignite and flash with gold and lavender light in extreme slow motion?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.13271604478359222 + ], + "opposite_logprob_tokens": [ + -2.132715940475464 + ], + "logprob_reward": -0.13271604478359222, + "yes_no_margin_reward": 1.9999998807907104, + "token_credit_reward": -0.13271604478359222, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1744169145822525, + 0.05047035962343216, + 0.0714978575706482, + 0.1891559362411499, + 0.09800519049167633, + 0.08450645208358765, + 0.09274467825889587, + 0.08255849778652191, + 0.12479019910097122, + 0.03185399994254112 + ], + "credit_map": [ + 0.024956582114100456, + 0.016168301925063133, + 0.017459942027926445, + 0.01474906038492918, + 0.010510235093533993, + 0.03780604898929596, + 0.016454419121146202, + 0.03631231561303139, + 0.005143529269844294, + 0.0064621311612427235, + 0.00496569974347949, + 0.008407765068113804, + 0.004532855935394764, + 0.007218662183731794, + 0.006990698166191578, + 0.006749015301465988, + 0.006324801594018936, + 0.006859802175313234, + 0.007699505891650915, + 0.011001386679708958, + 0.006364002823829651, + 0.012637637555599213, + 0.008207046426832676, + 0.012403671629726887, + 0.022449718788266182, + 0.02946629747748375, + 0.02148396149277687, + 0.02584422193467617, + 0.01973194070160389, + 0.01861959882080555, + 0.027712639421224594, + 0.023847538977861404, + 0.009678971022367477, + 0.013213562779128551, + 0.01045224815607071, + 0.011996421962976456, + 0.008694457821547985, + 0.01593685708940029, + 0.014235712587833405, + 0.013796946965157986, + 0.009146683849394321, + 0.014953145757317543, + 0.009045008569955826, + 0.0145140141248703, + 0.008365179412066936, + 0.007739749737083912, + 0.008957910351455212, + 0.01178475096821785, + 0.010920470580458641, + 0.007694280240684748, + 0.0222731102257967, + 0.008969712071120739, + 0.009966178797185421, + 0.011929667554795742, + 0.011821141466498375, + 0.009170112200081348, + 0.01140476856380701, + 0.00742624094709754, + 0.01984160579741001, + 0.00824529118835926, + 0.005609237588942051, + 0.008770962245762348, + 0.008757692761719227, + 0.012502686120569706, + 0.01920407824218273, + 0.016875391826033592, + 0.033279139548540115, + 0.012236065231263638, + 0.010299929417669773, + 0.009712944738566875, + 0.010623221285641193, + 0.01255942415446043, + 0.0033799067605286837, + 0.00417941203340888, + 0.00528210774064064, + 0.0042006452567875385, + 0.003626815741881728, + 0.003441076958552003, + 0.0037996426690369844, + 0.00394438998773694 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 0, + 8 + ], + "selected_reward": -0.13271604478359222 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a quiet corridor deep within a museum featuring a silver tabby British shorthair cat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the cat wearing a handmade dark brown leather harness with a gleaming brass buckle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is there a life-size white marble statue of a classical woman facing the cat in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the cat snap its mouth open to bare sharp teeth and hiss at the statue?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do dust particles in the light beam ignite and flash with gold and lavender light in extreme slow motion?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03373406082391739 + ], + "opposite_logprob_tokens": [ + -3.408734083175659 + ], + "logprob_reward": -0.03373406082391739, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.03373406082391739, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14709767699241638, + 0.049609243869781494, + 0.07659081369638443, + 0.21330538392066956, + 0.12170959264039993, + 0.035957783460617065, + 0.07360294461250305, + 0.03587030619382858, + 0.16958652436733246, + 0.07666964083909988 + ], + "credit_map": [ + 0.016184737905859947, + 0.01903909631073475, + 0.012295466847717762, + 0.01748719997704029, + 0.01077477727085352, + 0.03333289921283722, + 0.013427219353616238, + 0.024556269869208336, + 0.006360969040542841, + 0.00821410771459341, + 0.0056115793995559216, + 0.007366017904132605, + 0.004821411333978176, + 0.0046941195614635944, + 0.006971943657845259, + 0.005569093860685825, + 0.0064591472037136555, + 0.009206054732203484, + 0.008914452977478504, + 0.012075051665306091, + 0.007363707758486271, + 0.011715728789567947, + 0.009642478078603745, + 0.011214188300073147, + 0.024527719244360924, + 0.03233940899372101, + 0.026928814128041267, + 0.026978086680173874, + 0.02420802041888237, + 0.016722150146961212, + 0.033965643495321274, + 0.02763555385172367, + 0.014088748022913933, + 0.013491101562976837, + 0.012798117473721504, + 0.017932241782546043, + 0.0136973662301898, + 0.01618855632841587, + 0.015636451542377472, + 0.017877014353871346, + 0.0035479480866342783, + 0.0035412053111940622, + 0.0039046097081154585, + 0.007803105749189854, + 0.00319571141153574, + 0.0047731841914355755, + 0.005017425399273634, + 0.004174591973423958, + 0.010340407490730286, + 0.005856203846633434, + 0.019268890842795372, + 0.00701814703643322, + 0.00702442042529583, + 0.008104476146399975, + 0.008491549640893936, + 0.007498852908611298, + 0.003234425326809287, + 0.006464212201535702, + 0.004130115732550621, + 0.006030531600117683, + 0.002995595335960388, + 0.004080882295966148, + 0.004185671918094158, + 0.004748876206576824, + 0.029147718101739883, + 0.018163461238145828, + 0.04975973442196846, + 0.01727774180471897, + 0.010236014612019062, + 0.013029658235609531, + 0.01222789566963911, + 0.01974429376423359, + 0.00860489159822464, + 0.007135362364351749, + 0.015783729031682014, + 0.007034800015389919, + 0.005945801734924316, + 0.010756843723356724, + 0.007871249690651894, + 0.013536965474486351 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 8, + 0 + ], + "selected_reward": -0.03373406082391739 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a quiet corridor deep within a museum featuring a silver tabby British shorthair cat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the cat wearing a handmade dark brown leather harness with a gleaming brass buckle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is there a life-size white marble statue of a classical woman facing the cat in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the cat snap its mouth open to bare sharp teeth and hiss at the statue?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do dust particles in the light beam ignite and flash with gold and lavender light in extreme slow motion?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.020584525540471077 + ], + "opposite_logprob_tokens": [ + -3.8955845832824707 + ], + "logprob_reward": -0.020584525540471077, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.020584525540471077, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.05326593294739723, + 0.03496577963232994, + 0.05949803441762924, + 0.5760521292686462, + 0.052484944462776184, + 0.04014439135789871, + 0.07747717201709747, + 0.051441650837659836, + 0.03953816369175911, + 0.015131823718547821 + ], + "credit_map": [ + 0.006923355627804995, + 0.005638863891363144, + 0.006816274486482143, + 0.006191931664943695, + 0.004091936629265547, + 0.008351007476449013, + 0.007297573145478964, + 0.007954983972012997, + 0.005155804567039013, + 0.005358236841857433, + 0.004863022826611996, + 0.003902364755049348, + 0.0029511917382478714, + 0.003629890503361821, + 0.004185649566352367, + 0.004919617436826229, + 0.00603554816916585, + 0.007721151690930128, + 0.005837954115122557, + 0.009492079727351665, + 0.006607792805880308, + 0.00552195543423295, + 0.010855115950107574, + 0.007426433265209198, + 0.07178240269422531, + 0.08279657363891602, + 0.07104020565748215, + 0.07486194372177124, + 0.0591670460999012, + 0.05705449730157852, + 0.08233142644166946, + 0.07701796293258667, + 0.005871470086276531, + 0.006921214517205954, + 0.005528199952095747, + 0.008127045817673206, + 0.0053213853389024734, + 0.0071114241145551205, + 0.00568007305264473, + 0.007924128323793411, + 0.003919265232980251, + 0.006448613479733467, + 0.00409476924687624, + 0.006781588774174452, + 0.0037140813656151295, + 0.005689783021807671, + 0.004766654223203659, + 0.004729635547846556, + 0.011902588419616222, + 0.004688646644353867, + 0.023181073367595673, + 0.007649419829249382, + 0.006817006040364504, + 0.007041417993605137, + 0.007178711239248514, + 0.009018301963806152, + 0.005378717556595802, + 0.0050520519725978374, + 0.011838322505354881, + 0.006691425107419491, + 0.0034165505785495043, + 0.004362018778920174, + 0.006543867755681276, + 0.008158687502145767, + 0.003672562539577484, + 0.008047213777899742, + 0.004587230738252401, + 0.0039688958786427975, + 0.003350904444232583, + 0.008243235759437084, + 0.0034710627514868975, + 0.004197055473923683, + 0.001744331675581634, + 0.0019308265764266253, + 0.002739071613177657, + 0.0014260472962632775, + 0.0014075958169996738, + 0.0022190031595528126, + 0.0019892510026693344, + 0.0016756951808929443 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 6, + 2 + ], + "selected_reward": -0.020584525540471077 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a quiet corridor deep within a museum featuring a silver tabby British shorthair cat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the cat wearing a handmade dark brown leather harness with a gleaming brass buckle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is there a life-size white marble statue of a classical woman facing the cat in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the cat snap its mouth open to bare sharp teeth and hiss at the statue?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do dust particles in the light beam ignite and flash with gold and lavender light in extreme slow motion?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.2521228790283203 + ], + "opposite_logprob_tokens": [ + -1.5021228790283203 + ], + "logprob_reward": -0.2521228790283203, + "yes_no_margin_reward": 1.25, + "token_credit_reward": -0.2521228790283203, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.05229076370596886, + 0.043688759207725525, + 0.04573877155780792, + 0.15690936148166656, + 0.07011935114860535, + 0.08919384330511093, + 0.2822297215461731, + 0.054225388914346695, + 0.1490567922592163, + 0.05654723569750786 + ], + "credit_map": [ + 0.006102642510086298, + 0.006023544352501631, + 0.004441715311259031, + 0.008062878623604774, + 0.004038634710013866, + 0.005990224424749613, + 0.007611343637108803, + 0.01001978199928999, + 0.00965784303843975, + 0.004985479637980461, + 0.0048207747749984264, + 0.0041546388529241085, + 0.003997689578682184, + 0.00472397543489933, + 0.005630864296108484, + 0.005717490334063768, + 0.004041055217385292, + 0.0051525654271245, + 0.004241527058184147, + 0.007835396565496922, + 0.0035873111337423325, + 0.00677067693322897, + 0.006756591610610485, + 0.007353649474680424, + 0.018325846642255783, + 0.025729181244969368, + 0.01696871593594551, + 0.022004935890436172, + 0.01833590678870678, + 0.013593058101832867, + 0.021186470985412598, + 0.020765256136655807, + 0.008270067162811756, + 0.007668550591915846, + 0.008606674149632454, + 0.011715419590473175, + 0.007155570667237043, + 0.0075011663138866425, + 0.008706992492079735, + 0.010494907386600971, + 0.01002945564687252, + 0.013942169956862926, + 0.011585499159991741, + 0.01592116430401802, + 0.0063895005732774734, + 0.011225929483771324, + 0.009096741676330566, + 0.011003384366631508, + 0.049537766724824905, + 0.014369461685419083, + 0.09311693161725998, + 0.025365062057971954, + 0.025144873186945915, + 0.02253555692732334, + 0.025024503469467163, + 0.02713557705283165, + 0.005397376138716936, + 0.005731058306992054, + 0.007772794459015131, + 0.007187807932496071, + 0.005100747104734182, + 0.008783026598393917, + 0.007247153203934431, + 0.0070054251700639725, + 0.024066733196377754, + 0.014337588101625443, + 0.04512900859117508, + 0.012422574684023857, + 0.008849079720675945, + 0.010777770541608334, + 0.01331366878002882, + 0.020160358399152756, + 0.006690214853733778, + 0.004788402933627367, + 0.012716904282569885, + 0.004706707317382097, + 0.004171010106801987, + 0.004765118472278118, + 0.005958620924502611, + 0.012750256806612015 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 3, + 8 + ], + "selected_reward": -0.2521228790283203 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a quiet corridor deep within a museum featuring a silver tabby British shorthair cat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the cat wearing a handmade dark brown leather harness with a gleaming brass buckle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is there a life-size white marble statue of a classical woman facing the cat in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the cat snap its mouth open to bare sharp teeth and hiss at the statue?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do dust particles in the light beam ignite and flash with gold and lavender light in extreme slow motion?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02981162816286087 + ], + "opposite_logprob_tokens": [ + -3.5298116207122803 + ], + "logprob_reward": -0.02981162816286087, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.02981162816286087, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08601296693086624, + 0.0426604226231575, + 0.028032170608639717, + 0.37196585536003113, + 0.043454959988594055, + 0.04679733142256737, + 0.07107236236333847, + 0.0721597746014595, + 0.09428298473358154, + 0.14356115460395813 + ], + "credit_map": [ + 0.013392264023423195, + 0.008609159849584103, + 0.010037967935204506, + 0.008437324315309525, + 0.006949581205844879, + 0.013148132711648941, + 0.010653827339410782, + 0.014784715138375759, + 0.009132727980613708, + 0.005328829400241375, + 0.005089325364679098, + 0.004277486819773912, + 0.003524096915498376, + 0.004378311336040497, + 0.005435347091406584, + 0.005494295619428158, + 0.002870774595066905, + 0.004476732574403286, + 0.0034184446558356285, + 0.004073414020240307, + 0.0036592171527445316, + 0.0028818552382290363, + 0.003437798237428069, + 0.0032139355316758156, + 0.04644673317670822, + 0.0541069470345974, + 0.04555601626634598, + 0.04853133484721184, + 0.039046790450811386, + 0.035551995038986206, + 0.05304190516471863, + 0.049684155732393265, + 0.004186066333204508, + 0.0056014820002019405, + 0.0037473984993994236, + 0.008609768934547901, + 0.0054619754664599895, + 0.004885499831289053, + 0.005441614892333746, + 0.005521158687770367, + 0.004104862920939922, + 0.007895062677562237, + 0.004629333969205618, + 0.006874884478747845, + 0.0032353184651583433, + 0.008016121573746204, + 0.004628312308341265, + 0.007413437590003014, + 0.006287033203989267, + 0.0088815838098526, + 0.008948029018938541, + 0.009258296340703964, + 0.0076002320274710655, + 0.00830330140888691, + 0.012285898439586163, + 0.009507984854280949, + 0.006409815978258848, + 0.007804728578776121, + 0.008359352126717567, + 0.008632561191916466, + 0.00680503249168396, + 0.010555759072303772, + 0.01234387792646885, + 0.01124864723533392, + 0.0066258832812309265, + 0.020403297618031502, + 0.007544904015958309, + 0.012029051780700684, + 0.009082683362066746, + 0.012087124399840832, + 0.0148121053352952, + 0.011697934940457344, + 0.014540824107825756, + 0.010408376343548298, + 0.02002759650349617, + 0.010586133226752281, + 0.012299710884690285, + 0.016915064305067062, + 0.02133905701339245, + 0.03744438290596008 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 9, + 8 + ], + "selected_reward": -0.02981162816286087 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "low-angle, medium shot, filmed in extreme slow motion, slowly and steadily zooms in on a quiet corridor deep within the museum. A narrow, soft beam of light slants down from a high side window, illuminating the area perfectly, creating a dramatic, Rembrandt-esque effect of light and shadow. The focal point of the image is a silver tabby British shorthair cat, clad in a handmade, dark brown leather harness, its brass buckle gleaming discreetly in the dim light. Its body is tense, its back slightly arched, its ears pressed back, and its amber eyes wide, their pupils narrowed to a thin line of intense alert. It faces a life-size, white marble statue of a classical woman, its surface etched with the marks of time, its posture serene and timeless. The cat's mouth snaps open, baring sharp teeth. At the climax of a warning hiss, an invisible wave of sound erupts from its mouth, instantly stirring the ancient dust still within the beam. In extreme slow motion, every tiny dust particle shaken by the sound waves was ignited by the light beam, like countless suspended miniature stars, flashing an indescribable forbidden light mixed with gold and lavender, injecting a unique atmosphere of sacredness, tension and supernatural mystery into the confrontation that was about to break out in this silent space.", + "seed": 31694, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000006/group_8/index_3529_seed_31694_dp_17_rank_34_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 8, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.2668081521987915, + "vlm_logprob": -0.2668081521987915, + "vlm_yes_no_margin": 1.675, + "vlm_token_credit": -0.2668081521987915, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a quiet corridor deep within a museum featuring a silver tabby British shorthair cat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the cat wearing a handmade dark brown leather harness with a gleaming brass buckle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is there a life-size white marble statue of a classical woman facing the cat in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the cat snap its mouth open to bare sharp teeth and hiss at the statue?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do dust particles in the light beam ignite and flash with gold and lavender light in extreme slow motion?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.5795711278915405 + ], + "opposite_logprob_tokens": [ + -0.8295711278915405 + ], + "logprob_reward": -0.5795711278915405, + "yes_no_margin_reward": 0.25, + "token_credit_reward": -0.5795711278915405, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.19992612302303314, + 0.07374843209981918, + 0.03056821972131729, + 0.032870981842279434, + 0.029436688870191574, + 0.04311789944767952, + 0.1788225620985031, + 0.1207740306854248, + 0.180987149477005, + 0.10974784195423126 + ], + "credit_map": [ + 0.02982083521783352, + 0.017318224534392357, + 0.026494501158595085, + 0.03043455444276333, + 0.02204393595457077, + 0.022342072799801826, + 0.02565060555934906, + 0.025821389630436897, + 0.009190216660499573, + 0.006201551761478186, + 0.010151045396924019, + 0.009430783800780773, + 0.00712504331022501, + 0.00838007777929306, + 0.008739336393773556, + 0.014530383050441742, + 0.005710277706384659, + 0.002838532207533717, + 0.004030159674584866, + 0.0035618855617940426, + 0.0023807946126908064, + 0.0035461727529764175, + 0.004346150439232588, + 0.004154243506491184, + 0.0040303785353899, + 0.0033646852243691683, + 0.003942083101719618, + 0.00513404281809926, + 0.003774805925786495, + 0.004001999739557505, + 0.004135569557547569, + 0.004487417172640562, + 0.0038491101004183292, + 0.0030456366948783398, + 0.003941111266613007, + 0.004245910327881575, + 0.0030215925071388483, + 0.0036717599723488092, + 0.0028833134565502405, + 0.004778256174176931, + 0.007175788749009371, + 0.0051889014430344105, + 0.005635728593915701, + 0.005262407008558512, + 0.004793262109160423, + 0.0037402219604700804, + 0.0058738430961966515, + 0.005447747185826302, + 0.044661372900009155, + 0.011850109323859215, + 0.018026864156126976, + 0.011538995429873466, + 0.012216292321681976, + 0.01560553815215826, + 0.010618065483868122, + 0.05430533364415169, + 0.02738136798143387, + 0.008151687681674957, + 0.026294274255633354, + 0.012681483291089535, + 0.009537503123283386, + 0.010614479891955853, + 0.018066154792904854, + 0.008047083392739296, + 0.03815664350986481, + 0.009715606458485126, + 0.041179973632097244, + 0.025308607146143913, + 0.017679201439023018, + 0.017867116257548332, + 0.014074836857616901, + 0.01700517162680626, + 0.03380777686834335, + 0.0046819839626550674, + 0.012919113971292973, + 0.006449021399021149, + 0.006596507504582405, + 0.008224393241107464, + 0.006836291402578354, + 0.030232757329940796 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 6 + ], + "selected_reward": -0.5795711278915405 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a quiet corridor deep within a museum featuring a silver tabby British shorthair cat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the cat wearing a handmade dark brown leather harness with a gleaming brass buckle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is there a life-size white marble statue of a classical woman facing the cat in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the cat snap its mouth open to bare sharp teeth and hiss at the statue?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do dust particles in the light beam ignite and flash with gold and lavender light in extreme slow motion?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07008572667837143 + ], + "opposite_logprob_tokens": [ + -2.6950857639312744 + ], + "logprob_reward": -0.07008572667837143, + "yes_no_margin_reward": 2.625, + "token_credit_reward": -0.07008572667837143, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2171354591846466, + 0.12896843254566193, + 0.03574565798044205, + 0.034018393605947495, + 0.036558981984853745, + 0.07489489763975143, + 0.06429384648799896, + 0.14244812726974487, + 0.17139539122581482, + 0.09454081952571869 + ], + "credit_map": [ + 0.03370178863406181, + 0.01759037934243679, + 0.027979286387562752, + 0.02328220009803772, + 0.02713802643120289, + 0.027911808341741562, + 0.021349510177969933, + 0.03818242996931076, + 0.021080898120999336, + 0.012576550245285034, + 0.01882156915962696, + 0.014908894896507263, + 0.01599869877099991, + 0.01522812619805336, + 0.012911086902022362, + 0.017442598938941956, + 0.005996047519147396, + 0.004119778051972389, + 0.005059986840933561, + 0.004385970998555422, + 0.003185429610311985, + 0.005079866386950016, + 0.002763823140412569, + 0.005154751241207123, + 0.004461805336177349, + 0.0022937413305044174, + 0.004939279984682798, + 0.0034983004443347454, + 0.003841871628537774, + 0.004205003846436739, + 0.0050268275663256645, + 0.005751559976488352, + 0.005515214521437883, + 0.0025344740133732557, + 0.005197268910706043, + 0.004964138381183147, + 0.003845715429633856, + 0.0032435753382742405, + 0.0056387451477348804, + 0.0056198472157120705, + 0.012118421494960785, + 0.005838553886860609, + 0.012630023062229156, + 0.0069787283428013325, + 0.00911279022693634, + 0.007595567964017391, + 0.011483446694910526, + 0.009137359447777271, + 0.012549079954624176, + 0.006330172065645456, + 0.00891350582242012, + 0.0070772552862763405, + 0.006694244686514139, + 0.008054845966398716, + 0.006068979389965534, + 0.008605756796896458, + 0.03353433310985565, + 0.009859946556389332, + 0.0313437394797802, + 0.017359498888254166, + 0.00890776701271534, + 0.009405349381268024, + 0.01993449032306671, + 0.012102997861802578, + 0.03837224468588829, + 0.007089981343597174, + 0.05507827177643776, + 0.014283264055848122, + 0.009772101417183876, + 0.012624919414520264, + 0.009040236473083496, + 0.025134360417723656, + 0.015574293211102486, + 0.010595150291919708, + 0.022393543273210526, + 0.009667840786278248, + 0.007901552133262157, + 0.009963366203010082, + 0.007604026701301336, + 0.01084104273468256 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 7 + ], + "selected_reward": -0.07008572667837143 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a quiet corridor deep within a museum featuring a silver tabby British shorthair cat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the cat wearing a handmade dark brown leather harness with a gleaming brass buckle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is there a life-size white marble statue of a classical woman facing the cat in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the cat snap its mouth open to bare sharp teeth and hiss at the statue?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do dust particles in the light beam ignite and flash with gold and lavender light in extreme slow motion?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.029807694256305695 + ], + "opposite_logprob_tokens": [ + -3.5298078060150146 + ], + "logprob_reward": -0.029807694256305695, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.029807694256305695, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15243449807167053, + 0.08132335543632507, + 0.04776342958211899, + 0.033729393035173416, + 0.0495559461414814, + 0.04104539752006531, + 0.09900344908237457, + 0.19186630845069885, + 0.2561698853969574, + 0.04710829630494118 + ], + "credit_map": [ + 0.022458914667367935, + 0.01310365553945303, + 0.021040355786681175, + 0.01725422404706478, + 0.01369389146566391, + 0.019967518746852875, + 0.017867188900709152, + 0.027048764750361443, + 0.015472480095922947, + 0.006469125393778086, + 0.011174613609910011, + 0.009643319994211197, + 0.008866026997566223, + 0.007938416674733162, + 0.007604624144732952, + 0.014154758304357529, + 0.008143825456500053, + 0.005131582263857126, + 0.0071011921390891075, + 0.006359566003084183, + 0.004634384531527758, + 0.0056150867603719234, + 0.005595075897872448, + 0.005182723980396986, + 0.005478177685290575, + 0.003051895648241043, + 0.004659328144043684, + 0.005027232225984335, + 0.0029361851047724485, + 0.0038943588733673096, + 0.004062474239617586, + 0.004619744140654802, + 0.007847369648516178, + 0.00421716645359993, + 0.008016259409487247, + 0.00678755110129714, + 0.004646020475775003, + 0.004056577105075121, + 0.007120585069060326, + 0.006864421535283327, + 0.005517391487956047, + 0.003971936646848917, + 0.005492295138537884, + 0.004425111226737499, + 0.0048016649670898914, + 0.0036671506240963936, + 0.006734347436577082, + 0.006435507442802191, + 0.018232563510537148, + 0.0069050258025527, + 0.018762873485684395, + 0.011287074536085129, + 0.011590467765927315, + 0.013216772116720676, + 0.009675287641584873, + 0.00933338887989521, + 0.04197370260953903, + 0.01149777416139841, + 0.038972366601228714, + 0.021344589069485664, + 0.014607030898332596, + 0.015881968662142754, + 0.033271193504333496, + 0.014317692257463932, + 0.05303423851728439, + 0.012406297959387302, + 0.06671950966119766, + 0.031916044652462006, + 0.021991271525621414, + 0.02181261032819748, + 0.01744886487722397, + 0.030841076746582985, + 0.009201123379170895, + 0.0034094464499503374, + 0.010403910651803017, + 0.003944902680814266, + 0.0038682103622704744, + 0.0038939588703215122, + 0.0038359928876161575, + 0.008550754748284817 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 7, + 0 + ], + "selected_reward": -0.029807694256305695 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a quiet corridor deep within a museum featuring a silver tabby British shorthair cat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the cat wearing a handmade dark brown leather harness with a gleaming brass buckle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is there a life-size white marble statue of a classical woman facing the cat in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the cat snap its mouth open to bare sharp teeth and hiss at the statue?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do dust particles in the light beam ignite and flash with gold and lavender light in extreme slow motion?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.4289379119873047 + ], + "opposite_logprob_tokens": [ + -1.0539379119873047 + ], + "logprob_reward": -0.4289379119873047, + "yes_no_margin_reward": 0.625, + "token_credit_reward": -0.4289379119873047, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08446981757879257, + 0.07458052039146423, + 0.05602899193763733, + 0.05183461308479309, + 0.04676274210214615, + 0.029141560196876526, + 0.07714954018592834, + 0.24265842139720917, + 0.3005155920982361, + 0.03685823082923889 + ], + "credit_map": [ + 0.01439050491899252, + 0.0065594809129834175, + 0.011488141492009163, + 0.010419527068734169, + 0.00794052705168724, + 0.010547737590968609, + 0.008838250301778316, + 0.01428564265370369, + 0.012158393859863281, + 0.0054622129537165165, + 0.010215443558990955, + 0.010114751756191254, + 0.008340607397258282, + 0.008481139317154884, + 0.006868437398225069, + 0.012939530424773693, + 0.007842576131224632, + 0.005267728120088577, + 0.007252267561852932, + 0.007772126700729132, + 0.006324144080281258, + 0.006940067280083895, + 0.005333338398486376, + 0.009296740405261517, + 0.006686115171760321, + 0.004700391087681055, + 0.00699158851057291, + 0.007359699346125126, + 0.005730951204895973, + 0.006913064047694206, + 0.004728366620838642, + 0.008724428713321686, + 0.006490184925496578, + 0.004183425568044186, + 0.005618717521429062, + 0.006612464319914579, + 0.006384882144629955, + 0.005596858449280262, + 0.005372442305088043, + 0.00650376221165061, + 0.005121916066855192, + 0.002632752526551485, + 0.0038501271046698093, + 0.004559362307190895, + 0.0032381939236074686, + 0.0027659388724714518, + 0.0034169384744018316, + 0.003556327661499381, + 0.0168200321495533, + 0.004613108932971954, + 0.010908893309533596, + 0.0073548476211726665, + 0.004878346808254719, + 0.007514795288443565, + 0.007059745956212282, + 0.017999760806560516, + 0.053411953151226044, + 0.014656156301498413, + 0.0497809462249279, + 0.027622096240520477, + 0.01842123456299305, + 0.019653914496302605, + 0.04126334190368652, + 0.017848772928118706, + 0.06115766987204552, + 0.015431583859026432, + 0.07404626160860062, + 0.03863398730754852, + 0.02714095078408718, + 0.026616403833031654, + 0.022822503000497818, + 0.03466622158885002, + 0.009975981898605824, + 0.00256886868737638, + 0.007373410742729902, + 0.0028184140101075172, + 0.002667598892003298, + 0.0030872218776494265, + 0.0024403149727731943, + 0.0059264181181788445 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 7, + 0 + ], + "selected_reward": -0.4289379119873047 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a quiet corridor deep within a museum featuring a silver tabby British shorthair cat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the cat wearing a handmade dark brown leather harness with a gleaming brass buckle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is there a life-size white marble statue of a classical woman facing the cat in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the cat snap its mouth open to bare sharp teeth and hiss at the statue?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do dust particles in the light beam ignite and flash with gold and lavender light in extreme slow motion?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.22563830018043518 + ], + "opposite_logprob_tokens": [ + -1.6006382703781128 + ], + "logprob_reward": -0.22563830018043518, + "yes_no_margin_reward": 1.375, + "token_credit_reward": -0.22563830018043518, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06854086369276047, + 0.070083387196064, + 0.04856162145733833, + 0.04135363921523094, + 0.03067968040704727, + 0.03292202949523926, + 0.04292892664670944, + 0.22533255815505981, + 0.22379326820373535, + 0.21580401062965393 + ], + "credit_map": [ + 0.01189301535487175, + 0.007484904024749994, + 0.009394184686243534, + 0.006131504196673632, + 0.006999560631811619, + 0.007404302712529898, + 0.00823173951357603, + 0.011001658625900745, + 0.009306886233389378, + 0.006877460516989231, + 0.009163779206573963, + 0.008475949056446552, + 0.006954311393201351, + 0.007799780927598476, + 0.007264057174324989, + 0.014241172932088375, + 0.00769093306735158, + 0.0040832324884831905, + 0.006326689384877682, + 0.007750560063868761, + 0.0056997025385499, + 0.005024118814617395, + 0.006074766628444195, + 0.005911624524742365, + 0.005209748633205891, + 0.0038023432716727257, + 0.005093212705105543, + 0.006286635994911194, + 0.004810878541320562, + 0.005587154533714056, + 0.00396609166637063, + 0.006597581319510937, + 0.004321693442761898, + 0.002591556403785944, + 0.0036139560397714376, + 0.00502575421705842, + 0.0030328838620334864, + 0.003279108088463545, + 0.004444735590368509, + 0.004369999282062054, + 0.005503882188349962, + 0.003099520457908511, + 0.006186333484947681, + 0.003143307287245989, + 0.0030254479497671127, + 0.002918119542300701, + 0.0049483394250273705, + 0.004097082186490297, + 0.00683746300637722, + 0.0028317314572632313, + 0.007089307997375727, + 0.004881755914539099, + 0.0057054972276091576, + 0.006672913674265146, + 0.005000878591090441, + 0.003909386228770018, + 0.05021582543849945, + 0.012562871910631657, + 0.05431826040148735, + 0.02329241670668125, + 0.015737788751721382, + 0.018304940313100815, + 0.03570493310689926, + 0.015195534564554691, + 0.05214430391788483, + 0.012228097766637802, + 0.048965126276016235, + 0.02924925833940506, + 0.02119905687868595, + 0.022591259330511093, + 0.018182829022407532, + 0.019233347848057747, + 0.05429985001683235, + 0.012972005642950535, + 0.03952484577894211, + 0.015801284462213516, + 0.012508979067206383, + 0.015205214731395245, + 0.019572366029024124, + 0.045919496566057205 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 8, + 9 + ], + "selected_reward": -0.22563830018043518 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "low-angle, medium shot, filmed in extreme slow motion, slowly and steadily zooms in on a quiet corridor deep within the museum. A narrow, soft beam of light slants down from a high side window, illuminating the area perfectly, creating a dramatic, Rembrandt-esque effect of light and shadow. The focal point of the image is a silver tabby British shorthair cat, clad in a handmade, dark brown leather harness, its brass buckle gleaming discreetly in the dim light. Its body is tense, its back slightly arched, its ears pressed back, and its amber eyes wide, their pupils narrowed to a thin line of intense alert. It faces a life-size, white marble statue of a classical woman, its surface etched with the marks of time, its posture serene and timeless. The cat's mouth snaps open, baring sharp teeth. At the climax of a warning hiss, an invisible wave of sound erupts from its mouth, instantly stirring the ancient dust still within the beam. In extreme slow motion, every tiny dust particle shaken by the sound waves was ignited by the light beam, like countless suspended miniature stars, flashing an indescribable forbidden light mixed with gold and lavender, injecting a unique atmosphere of sacredness, tension and supernatural mystery into the confrontation that was about to break out in this silent space.", + "seed": 31695, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000006/group_8/index_3529_seed_31695_dp_17_rank_34_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 8, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.2195039788261056, + "vlm_logprob": -0.2195039788261056, + "vlm_yes_no_margin": 2.075, + "vlm_token_credit": -0.2195039788261056, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a quiet corridor deep within a museum featuring a silver tabby British shorthair cat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the cat wearing a handmade dark brown leather harness with a gleaming brass buckle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is there a life-size white marble statue of a classical woman facing the cat in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the cat snap its mouth open to bare sharp teeth and hiss at the statue?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do dust particles in the light beam ignite and flash with gold and lavender light in extreme slow motion?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.29338082671165466 + ], + "opposite_logprob_tokens": [ + -1.418380856513977 + ], + "logprob_reward": -0.29338082671165466, + "yes_no_margin_reward": 1.125, + "token_credit_reward": -0.29338082671165466, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.18294990062713623, + 0.1521056443452835, + 0.09889914095401764, + 0.057849835604429245, + 0.07133034616708755, + 0.04327856004238129, + 0.09511210769414902, + 0.06921589374542236, + 0.06481308490037918, + 0.16444547474384308 + ], + "credit_map": [ + 0.017703663557767868, + 0.02571190521121025, + 0.022079210728406906, + 0.03356040269136429, + 0.01164684072136879, + 0.022244848310947418, + 0.02267260104417801, + 0.027330422773957253, + 0.011564953252673149, + 0.015071257948875427, + 0.008922209031879902, + 0.015872515738010406, + 0.010553418658673763, + 0.019325556233525276, + 0.03608899191021919, + 0.03470674902200699, + 0.010890849865972996, + 0.011495047248899937, + 0.011284058913588524, + 0.01820528134703636, + 0.00610427837818861, + 0.009087859652936459, + 0.012072541750967503, + 0.019759219139814377, + 0.004161038901656866, + 0.0030920200515538454, + 0.0026805687230080366, + 0.004734036047011614, + 0.003059308510273695, + 0.005982047412544489, + 0.016057513654232025, + 0.018083304166793823, + 0.0038608156610280275, + 0.003670644946396351, + 0.0047873021103441715, + 0.00622380618005991, + 0.003270824206992984, + 0.009634720161557198, + 0.017518261447548866, + 0.022363968193531036, + 0.0032589100301265717, + 0.0033566602505743504, + 0.0035002934746444225, + 0.00586219085380435, + 0.003041257616132498, + 0.01025233045220375, + 0.0062507749535143375, + 0.007756142411381006, + 0.009927848353981972, + 0.012841597199440002, + 0.010537348687648773, + 0.012474716641008854, + 0.006342573091387749, + 0.014414382167160511, + 0.012418369762599468, + 0.016155263409018517, + 0.005072803236544132, + 0.006223749369382858, + 0.007642682176083326, + 0.006820221431553364, + 0.007928447797894478, + 0.009441479109227657, + 0.01577683910727501, + 0.010309671051800251, + 0.013478650711476803, + 0.007715744432061911, + 0.006365494336932898, + 0.007412564475089312, + 0.007046330254524946, + 0.007348224520683289, + 0.007483963388949633, + 0.007962116040289402, + 0.0234467051923275, + 0.017996808513998985, + 0.01263188011944294, + 0.01586136221885681, + 0.019005872309207916, + 0.031779296696186066, + 0.021698107942938805, + 0.02202545665204525 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 1 + ], + "selected_reward": -0.29338082671165466 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a quiet corridor deep within a museum featuring a silver tabby British shorthair cat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the cat wearing a handmade dark brown leather harness with a gleaming brass buckle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is there a life-size white marble statue of a classical woman facing the cat in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the cat snap its mouth open to bare sharp teeth and hiss at the statue?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do dust particles in the light beam ignite and flash with gold and lavender light in extreme slow motion?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.055000219494104385 + ], + "opposite_logprob_tokens": [ + -2.9300003051757812 + ], + "logprob_reward": -0.055000219494104385, + "yes_no_margin_reward": 2.875, + "token_credit_reward": -0.055000219494104385, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.18874025344848633, + 0.032018955796957016, + 0.26012110710144043, + 0.1016489788889885, + 0.06527014076709747, + 0.07196982949972153, + 0.020539559423923492, + 0.055051445960998535, + 0.11793491244316101, + 0.08670477569103241 + ], + "credit_map": [ + 0.016411617398262024, + 0.024694044142961502, + 0.027933375909924507, + 0.035194527357816696, + 0.011921991594135761, + 0.0158788301050663, + 0.0311004389077425, + 0.025605425238609314, + 0.004341887775808573, + 0.003756958758458495, + 0.0037593357264995575, + 0.0033775330521166325, + 0.0025210860185325146, + 0.004805425647646189, + 0.003811454866081476, + 0.005645274184644222, + 0.021522076800465584, + 0.025656163692474365, + 0.029091395437717438, + 0.04928506538271904, + 0.012101192958652973, + 0.0257499311119318, + 0.035401687026023865, + 0.061313580721616745, + 0.004140621051192284, + 0.0036524483002722263, + 0.003198249964043498, + 0.004865663591772318, + 0.003364913398399949, + 0.011258451268076897, + 0.031878530979156494, + 0.03929009661078453, + 0.0032263961620628834, + 0.00273750233463943, + 0.002704185899347067, + 0.004244441632181406, + 0.0021663871593773365, + 0.009721161797642708, + 0.016521109268069267, + 0.02394895628094673, + 0.007064878009259701, + 0.005557378754019737, + 0.007958070375025272, + 0.008975622244179249, + 0.0036539044231176376, + 0.018623458221554756, + 0.0068380082957446575, + 0.013298509642481804, + 0.00220344215631485, + 0.002436916809529066, + 0.0025261843111366034, + 0.002468085614964366, + 0.0015621248167008162, + 0.003568303072825074, + 0.002170712687075138, + 0.0036037894897162914, + 0.004214214161038399, + 0.00834820605814457, + 0.010896091349422932, + 0.009290180169045925, + 0.003935431130230427, + 0.005276310723274946, + 0.005776572041213512, + 0.007314440328627825, + 0.013741767965257168, + 0.014409694820642471, + 0.027199432253837585, + 0.015816161409020424, + 0.009065886959433556, + 0.008335396647453308, + 0.017055261880159378, + 0.012311308644711971, + 0.007888546213507652, + 0.0124601349234581, + 0.013674696907401085, + 0.010041615925729275, + 0.012266945093870163, + 0.011600809171795845, + 0.010299487970769405, + 0.008472543209791183 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 0, + 8 + ], + "selected_reward": -0.055000219494104385 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a quiet corridor deep within a museum featuring a silver tabby British shorthair cat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the cat wearing a handmade dark brown leather harness with a gleaming brass buckle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is there a life-size white marble statue of a classical woman facing the cat in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the cat snap its mouth open to bare sharp teeth and hiss at the statue?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do dust particles in the light beam ignite and flash with gold and lavender light in extreme slow motion?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.012548656202852726 + ], + "opposite_logprob_tokens": [ + -4.387548446655273 + ], + "logprob_reward": -0.012548656202852726, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.012548656202852726, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.20554666221141815, + 0.16809222102165222, + 0.330567866563797, + 0.04699414595961571, + 0.054261140525341034, + 0.011130948550999165, + 0.028855495154857635, + 0.013075524941086769, + 0.02548106759786606, + 0.1159949004650116 + ], + "credit_map": [ + 0.012822616845369339, + 0.015146189369261265, + 0.012585153803229332, + 0.01887364499270916, + 0.009063524194061756, + 0.012127775698900223, + 0.06504195928573608, + 0.059885814785957336, + 0.014553981833159924, + 0.011822407133877277, + 0.009880317375063896, + 0.011889376677572727, + 0.008979981765151024, + 0.02258387766778469, + 0.040831100195646286, + 0.04755118116736412, + 0.02707991935312748, + 0.031314317137002945, + 0.03341396898031235, + 0.05691509321331978, + 0.02272951789200306, + 0.04139718413352966, + 0.04717876762151718, + 0.07053912431001663, + 0.0024816947989165783, + 0.002234678016975522, + 0.0016777298878878355, + 0.0029949622694402933, + 0.002069261856377125, + 0.00508571183308959, + 0.014514251612126827, + 0.015935856848955154, + 0.0023021132219582796, + 0.0022631892934441566, + 0.0018295978661626577, + 0.0027382182888686657, + 0.0016734602395445108, + 0.008748592808842659, + 0.01349878404289484, + 0.021207191050052643, + 0.0011501119006425142, + 0.0009725830750539899, + 0.0014245039783418179, + 0.0023356263991445303, + 0.0006511175888590515, + 0.0015455137472599745, + 0.0016075726598501205, + 0.0014439199585467577, + 0.0028098388575017452, + 0.004889952018857002, + 0.0024138540029525757, + 0.0027866160962730646, + 0.0018738450016826391, + 0.004503157921135426, + 0.003974806051701307, + 0.005603428464382887, + 0.001489208429120481, + 0.0011510259937494993, + 0.0019858875311911106, + 0.0016358033753931522, + 0.0010049151023849845, + 0.001566982944495976, + 0.0019490157719701529, + 0.0022926859091967344, + 0.0030928689520806074, + 0.002701446181163192, + 0.004180897958576679, + 0.0031047798693180084, + 0.002005962422117591, + 0.003232934745028615, + 0.0035124518908560276, + 0.0036497279070317745, + 0.014739882200956345, + 0.012046500109136105, + 0.009347079321742058, + 0.010728326626121998, + 0.011976515874266624, + 0.02809329144656658, + 0.016792522743344307, + 0.012270793318748474 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 0, + 1 + ], + "selected_reward": -0.012548656202852726 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a quiet corridor deep within a museum featuring a silver tabby British shorthair cat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the cat wearing a handmade dark brown leather harness with a gleaming brass buckle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is there a life-size white marble statue of a classical woman facing the cat in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the cat snap its mouth open to bare sharp teeth and hiss at the statue?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do dust particles in the light beam ignite and flash with gold and lavender light in extreme slow motion?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.5761961936950684 + ], + "opposite_logprob_tokens": [ + -0.8261961936950684 + ], + "logprob_reward": -0.5761961936950684, + "yes_no_margin_reward": 0.25, + "token_credit_reward": -0.5761961936950684, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.35191142559051514, + 0.05785386264324188, + 0.08947401493787766, + 0.07618564367294312, + 0.018229052424430847, + 0.030374161899089813, + 0.019191499799489975, + 0.041720494627952576, + 0.10608874261379242, + 0.20897118747234344 + ], + "credit_map": [ + 0.02069643698632717, + 0.03082653507590294, + 0.02935553342103958, + 0.040691401809453964, + 0.01694384589791298, + 0.020665017887949944, + 0.1006513461470604, + 0.09208124876022339, + 0.005146014504134655, + 0.0063234129920601845, + 0.005659991875290871, + 0.007231022696942091, + 0.003091521328315139, + 0.0051811253651976585, + 0.012848499231040478, + 0.012372266501188278, + 0.006885802373290062, + 0.010676027275621891, + 0.012226501479744911, + 0.01642204448580742, + 0.004600238054990768, + 0.009867826476693153, + 0.01085532084107399, + 0.01794024556875229, + 0.0022443337365984917, + 0.0024546608328819275, + 0.002011066535487771, + 0.002900015562772751, + 0.0025487393140792847, + 0.008714891970157623, + 0.024481022730469704, + 0.030830906704068184, + 0.0018040885915979743, + 0.0009589940309524536, + 0.0014286437071859837, + 0.0022616691421717405, + 0.0009387225145474076, + 0.0027460777200758457, + 0.0032192764338105917, + 0.004871579818427563, + 0.002089511835947633, + 0.0021514492109417915, + 0.002279134700074792, + 0.002891699317842722, + 0.0013497575419023633, + 0.008800497278571129, + 0.0034166346304118633, + 0.007395473308861256, + 0.0024065771140158176, + 0.002555261366069317, + 0.003153083147481084, + 0.0026914910413324833, + 0.0014802637742832303, + 0.0027270892169326544, + 0.002089577028527856, + 0.002088156994432211, + 0.004107836168259382, + 0.004266199190169573, + 0.00518173398450017, + 0.004196660593152046, + 0.003636261448264122, + 0.005922405514866114, + 0.006918915081769228, + 0.007490478456020355, + 0.010853090323507786, + 0.015317564830183983, + 0.0199771448969841, + 0.01686476171016693, + 0.0059597548097372055, + 0.008866602554917336, + 0.015497062355279922, + 0.012752749025821686, + 0.02961316518485546, + 0.0214077215641737, + 0.021174181252717972, + 0.020971452817320824, + 0.023913729935884476, + 0.04050713777542114, + 0.02944858744740486, + 0.02193518541753292 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 8 + ], + "selected_reward": -0.5761961936950684 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a quiet corridor deep within a museum featuring a silver tabby British shorthair cat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the cat wearing a handmade dark brown leather harness with a gleaming brass buckle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is there a life-size white marble statue of a classical woman facing the cat in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the cat snap its mouth open to bare sharp teeth and hiss at the statue?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do dust particles in the light beam ignite and flash with gold and lavender light in extreme slow motion?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.16039399802684784 + ], + "opposite_logprob_tokens": [ + -1.9103939533233643 + ], + "logprob_reward": -0.16039399802684784, + "yes_no_margin_reward": 1.75, + "token_credit_reward": -0.16039399802684784, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1301174908876419, + 0.05815678834915161, + 0.039529796689748764, + 0.13987873494625092, + 0.027028484269976616, + 0.07115907222032547, + 0.030385587364435196, + 0.10450377315282822, + 0.08774911612272263, + 0.3114911615848541 + ], + "credit_map": [ + 0.008955423720180988, + 0.0115352226421237, + 0.012058895081281662, + 0.015288244001567364, + 0.007034814450889826, + 0.008017357438802719, + 0.036500316113233566, + 0.030727218836545944, + 0.005420512519776821, + 0.005225020926445723, + 0.003762209787964821, + 0.005049118306487799, + 0.003967489581555128, + 0.007859531790018082, + 0.012976590543985367, + 0.013896319083869457, + 0.0034931215923279524, + 0.004072960466146469, + 0.004801155533641577, + 0.006399266421794891, + 0.0022691094782203436, + 0.0037527058739215136, + 0.006207682192325592, + 0.008533797226846218, + 0.004436742514371872, + 0.004738050978630781, + 0.0031012443359941244, + 0.005156179424375296, + 0.0046847728081047535, + 0.01613978110253811, + 0.044789869338274, + 0.05683208629488945, + 0.0012216103496029973, + 0.001173025812022388, + 0.0012940343003720045, + 0.001777116791345179, + 0.0011827648850157857, + 0.004207167308777571, + 0.006528762634843588, + 0.009644003584980965, + 0.005414207931607962, + 0.003608408384025097, + 0.004362948704510927, + 0.00726077426224947, + 0.003628344042226672, + 0.019247734919190407, + 0.00892863143235445, + 0.018708016723394394, + 0.002614237368106842, + 0.0051261200569570065, + 0.003913931082934141, + 0.004010532982647419, + 0.0019793445244431496, + 0.004287544637918472, + 0.003772520460188389, + 0.004681355319917202, + 0.008350208401679993, + 0.012751629576086998, + 0.01304910983890295, + 0.012889567762613297, + 0.009942270815372467, + 0.016462350264191628, + 0.01609225757420063, + 0.014966372400522232, + 0.008876757696270943, + 0.00938090868294239, + 0.008316604420542717, + 0.008442027494311333, + 0.006757265888154507, + 0.009985757991671562, + 0.01621410623192787, + 0.01977568306028843, + 0.036258578300476074, + 0.027291424572467804, + 0.02500329539179802, + 0.022226475179195404, + 0.03481924906373024, + 0.08666419237852097, + 0.04662233218550682, + 0.032605625689029694 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 3, + 0 + ], + "selected_reward": -0.16039399802684784 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000006/group_9/index_tensor([6606, 6606])_group_9_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000006/group_9/index_tensor([6606, 6606])_group_9_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..2e178f2a52a8479903dfc21e07875c32852576d1 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000006/group_9/index_tensor([6606, 6606])_group_9_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "In an extreme macro close-up shot from a slight downward angle, the camera remains still, focused on a professional-grade carving knife with a blade gleaming with a cool metallic sheen. Its handle, crafted from polished dark walnut, exudes a warm, textured feel. A warm, focused bench light streams down from above and sideways, illuminating the knife's tip as it steadily and deliberately slices into a block of untreated basswood, revealing the delicate grain and growth rings on the surface. With the blade's precise and fluid advance, a thin shaving of wood is gracefully released. Shot in extreme slow motion, the shaving appears like a living ribbon, slowly unfurling, curling, and swirling through the air. The delicate wood fibers within are clearly visible in the light, as if time has frozen in time. The light reflects dazzling highlights on the sharp blade, while outlining charming golden outlines on the arcs of flying sawdust. The background is completely blurred into a dark haze, leaving only this focused and poetic moment of carving. The whole picture is full of tranquility, healing and praise of exquisite craftsmanship.", + "seed": 15060, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000006/group_9/index_6606_seed_15060_dp_18_rank_36_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 9, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.06643400974571705, + "vlm_logprob": -0.06643400974571705, + "vlm_yes_no_margin": 3.8, + "vlm_token_credit": -0.06643400974571705, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme macro close-up of a professional-grade carving knife slicing into a block of untreated basswood?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the knife have a blade with a cool metallic sheen and a handle crafted from polished dark walnut?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene illuminated by a warm, focused bench light streaming from above and sideways, revealing wood grain and growth rings?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the video shot in extreme slow motion showing a thin wood shaving curling and swirling through the air with visible fibers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera completely still with a background blurred into a dark haze, while the blade reflects dazzling highlights?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.2574216425418854 + ], + "opposite_logprob_tokens": [ + -1.507421612739563 + ], + "logprob_reward": -0.2574216425418854, + "yes_no_margin_reward": 1.25, + "token_credit_reward": -0.2574216425418854, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14613084495067596, + 0.11441251635551453, + 0.23018966615200043, + 0.14087384939193726, + 0.05904793366789818, + 0.07567346096038818, + 0.05185361206531525, + 0.0429564043879509, + 0.060439031571149826, + 0.07842271775007248 + ], + "credit_map": [ + 0.0212676040828228, + 0.01593751274049282, + 0.020516186952590942, + 0.013647941872477531, + 0.023392371833324432, + 0.018408650532364845, + 0.013257274404168129, + 0.019703302532434464, + 0.01731359027326107, + 0.012745820917189121, + 0.013527728617191315, + 0.014199258759617805, + 0.01201838068664074, + 0.016435205936431885, + 0.011326310224831104, + 0.016846222802996635, + 0.03245140239596367, + 0.02462843246757984, + 0.02491648867726326, + 0.02744460478425026, + 0.0284538883715868, + 0.034022506326436996, + 0.03484204411506653, + 0.023430291563272476, + 0.015861043706536293, + 0.016179684549570084, + 0.014753833413124084, + 0.013025118038058281, + 0.022947363555431366, + 0.01513044722378254, + 0.026096932590007782, + 0.01687941513955593, + 0.0053770169615745544, + 0.008312727324664593, + 0.006782544311136007, + 0.0065702651627361774, + 0.00968501903116703, + 0.007145532872527838, + 0.007928671315312386, + 0.007246154360473156, + 0.009807266294956207, + 0.007798800244927406, + 0.008362620137631893, + 0.006420947145670652, + 0.009353915229439735, + 0.00713094137609005, + 0.014879383146762848, + 0.01191958598792553, + 0.007622122298926115, + 0.005450477357953787, + 0.004457717761397362, + 0.007305166684091091, + 0.005047906655818224, + 0.005744596011936665, + 0.009097515605390072, + 0.00712811341509223, + 0.005113800056278706, + 0.004543519113212824, + 0.004198567476123571, + 0.006139708682894707, + 0.005303745623677969, + 0.004045090172439814, + 0.0068900263868272305, + 0.006721945945173502, + 0.009037685580551624, + 0.005815914366394281, + 0.0061995843425393105, + 0.006079296115785837, + 0.007699979934841394, + 0.005765893030911684, + 0.012411859817802906, + 0.007428816985338926, + 0.0077403453178703785, + 0.008662908338010311, + 0.010067332535982132, + 0.008955015800893307, + 0.010032741352915764, + 0.008804833516478539, + 0.013830294832587242, + 0.01032924186438322 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 0, + 3 + ], + "selected_reward": -0.2574216425418854 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme macro close-up of a professional-grade carving knife slicing into a block of untreated basswood?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the knife have a blade with a cool metallic sheen and a handle crafted from polished dark walnut?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene illuminated by a warm, focused bench light streaming from above and sideways, revealing wood grain and growth rings?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the video shot in extreme slow motion showing a thin wood shaving curling and swirling through the air with visible fibers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera completely still with a background blurred into a dark haze, while the blade reflects dazzling highlights?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04869452118873596 + ], + "opposite_logprob_tokens": [ + -3.048694610595703 + ], + "logprob_reward": -0.04869452118873596, + "yes_no_margin_reward": 3.0, + "token_credit_reward": -0.04869452118873596, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10974326729774475, + 0.0946543887257576, + 0.17707107961177826, + 0.3805232644081116, + 0.03705868124961853, + 0.044433653354644775, + 0.026755699887871742, + 0.02266417071223259, + 0.030822666361927986, + 0.07627314329147339 + ], + "credit_map": [ + 0.015106527134776115, + 0.012280575931072235, + 0.01595219597220421, + 0.008999969810247421, + 0.016390172764658928, + 0.014533665962517262, + 0.011583421379327774, + 0.014896740205585957, + 0.013966570608317852, + 0.011294734664261341, + 0.011826041154563427, + 0.009817286394536495, + 0.008829833008348942, + 0.012757821008563042, + 0.009663539007306099, + 0.01649855636060238, + 0.025180790573358536, + 0.019187597557902336, + 0.0188918337225914, + 0.019986718893051147, + 0.02210143394768238, + 0.026880992576479912, + 0.026726843789219856, + 0.018114862963557243, + 0.03595205768942833, + 0.043093372136354446, + 0.0410885363817215, + 0.03834357485175133, + 0.0707784965634346, + 0.05009731650352478, + 0.058330852538347244, + 0.04283904284238815, + 0.0032828492112457752, + 0.004942236468195915, + 0.0041870358400046825, + 0.003802795894443989, + 0.006476885173469782, + 0.004781545139849186, + 0.005682908929884434, + 0.003902423894032836, + 0.006243572104722261, + 0.006427318789064884, + 0.004735411610454321, + 0.0031355638056993484, + 0.00495003629475832, + 0.007136373780667782, + 0.005701033864170313, + 0.006104340776801109, + 0.004941531457006931, + 0.003555077826604247, + 0.0024132479447871447, + 0.0017677870346233249, + 0.003375631757080555, + 0.0032173781655728817, + 0.003602230455726385, + 0.0038828151300549507, + 0.0022563531529158354, + 0.002244997536763549, + 0.0022954652085900307, + 0.002532815560698509, + 0.004149999935179949, + 0.0020323768258094788, + 0.005171712953597307, + 0.0019804479088634253, + 0.00336195295676589, + 0.0034999444615095854, + 0.0034761345013976097, + 0.003145745024085045, + 0.003028477542102337, + 0.002645590342581272, + 0.0072563751600682735, + 0.004408446606248617, + 0.007934135384857655, + 0.008097955957055092, + 0.009620963595807552, + 0.008820551447570324, + 0.009163226000964642, + 0.0063845072872936726, + 0.017101749777793884, + 0.00915005523711443 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 2, + 0 + ], + "selected_reward": -0.04869452118873596 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme macro close-up of a professional-grade carving knife slicing into a block of untreated basswood?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the knife have a blade with a cool metallic sheen and a handle crafted from polished dark walnut?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene illuminated by a warm, focused bench light streaming from above and sideways, revealing wood grain and growth rings?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the video shot in extreme slow motion showing a thin wood shaving curling and swirling through the air with visible fibers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera completely still with a background blurred into a dark haze, while the blade reflects dazzling highlights?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.003199579194188118 + ], + "opposite_logprob_tokens": [ + -5.753199577331543 + ], + "logprob_reward": -0.003199579194188118, + "yes_no_margin_reward": 5.75, + "token_credit_reward": -0.003199579194188118, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.17170801758766174, + 0.062377993017435074, + 0.34767210483551025, + 0.07452014833688736, + 0.062205031514167786, + 0.07485971599817276, + 0.06328149884939194, + 0.058171141892671585, + 0.04448153078556061, + 0.04072283208370209 + ], + "credit_map": [ + 0.02080601267516613, + 0.02133829891681671, + 0.02544294483959675, + 0.0166048351675272, + 0.022029932588338852, + 0.028837811201810837, + 0.017005616798996925, + 0.019642550498247147, + 0.01146541815251112, + 0.007737095467746258, + 0.008374042809009552, + 0.009999172762036324, + 0.0054894969798624516, + 0.006758690346032381, + 0.0053569357842206955, + 0.007197137456387281, + 0.05407022684812546, + 0.036397743970155716, + 0.04050525650382042, + 0.04373348131775856, + 0.041504763066768646, + 0.039142508059740067, + 0.054504405707120895, + 0.037813685834407806, + 0.010527783073484898, + 0.008395624347031116, + 0.007255264092236757, + 0.008018499240279198, + 0.011724026873707771, + 0.008259868249297142, + 0.013841196894645691, + 0.006497884634882212, + 0.006430651992559433, + 0.007461260538548231, + 0.007060911040753126, + 0.0055378638207912445, + 0.013182158581912518, + 0.007620496675372124, + 0.007792471442371607, + 0.0071192155592143536, + 0.01172578614205122, + 0.008388745598495007, + 0.008090472780168056, + 0.006766199134290218, + 0.007188582327216864, + 0.011136053130030632, + 0.011431996710598469, + 0.010131879709661007, + 0.009194983169436455, + 0.0069503458216786385, + 0.007318140938878059, + 0.005577781703323126, + 0.007838130928575993, + 0.00865728035569191, + 0.008038461208343506, + 0.009706375189125538, + 0.006364962551742792, + 0.005920405965298414, + 0.0047887591645121574, + 0.004913895390927792, + 0.005784680135548115, + 0.004555888939648867, + 0.01367967575788498, + 0.012162874452769756, + 0.005220852792263031, + 0.004142472520470619, + 0.004978938959538937, + 0.004115930758416653, + 0.004983110819011927, + 0.005343327764421701, + 0.008387548848986626, + 0.007309344597160816, + 0.005184223409742117, + 0.004328535404056311, + 0.005174316931515932, + 0.004561472684144974, + 0.004885938484221697, + 0.004406065680086613, + 0.007212587632238865, + 0.004969690926373005 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 0, + 5 + ], + "selected_reward": -0.003199579194188118 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme macro close-up of a professional-grade carving knife slicing into a block of untreated basswood?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the knife have a blade with a cool metallic sheen and a handle crafted from polished dark walnut?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene illuminated by a warm, focused bench light streaming from above and sideways, revealing wood grain and growth rings?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the video shot in extreme slow motion showing a thin wood shaving curling and swirling through the air with visible fibers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera completely still with a background blurred into a dark haze, while the blade reflects dazzling highlights?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.008646194823086262 + ], + "opposite_logprob_tokens": [ + -4.758646011352539 + ], + "logprob_reward": -0.008646194823086262, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.008646194823086262, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12997068464756012, + 0.07398118078708649, + 0.16068170964717865, + 0.30760204792022705, + 0.04623589664697647, + 0.04140680283308029, + 0.04779750853776932, + 0.043390654027462006, + 0.037572115659713745, + 0.11136139929294586 + ], + "credit_map": [ + 0.02557392045855522, + 0.014776565134525299, + 0.02205590344965458, + 0.012156845070421696, + 0.021252814680337906, + 0.011387249454855919, + 0.009573295712471008, + 0.013194093480706215, + 0.01163988932967186, + 0.008537117391824722, + 0.008817082270979881, + 0.009884791448712349, + 0.009195344522595406, + 0.01103104930371046, + 0.005872256588190794, + 0.00900365598499775, + 0.02298882231116295, + 0.01631806045770645, + 0.01811896823346615, + 0.022993285208940506, + 0.017493218183517456, + 0.016431907191872597, + 0.027454042807221413, + 0.01888340339064598, + 0.03226007521152496, + 0.035104937851428986, + 0.03220092132687569, + 0.037137679755687714, + 0.05138678476214409, + 0.042445432394742966, + 0.041711293160915375, + 0.03535492345690727, + 0.004320402629673481, + 0.004392087925225496, + 0.0039735701866447926, + 0.005253287497907877, + 0.009872935712337494, + 0.005375658627599478, + 0.00805220939218998, + 0.004995748866349459, + 0.007700242102146149, + 0.005441230721771717, + 0.004113036673516035, + 0.0030118899885565042, + 0.003919803537428379, + 0.005267612170428038, + 0.005977235268801451, + 0.005975750740617514, + 0.007619119714945555, + 0.0064621842466294765, + 0.004625209607183933, + 0.005123373121023178, + 0.00646630022674799, + 0.0038419896736741066, + 0.008317196741700172, + 0.005342131946235895, + 0.004502474796026945, + 0.004415508825331926, + 0.0038919427897781134, + 0.00572214275598526, + 0.0049309334717690945, + 0.0030840178951621056, + 0.010368417017161846, + 0.006475214846432209, + 0.004358421545475721, + 0.004182474687695503, + 0.003772465279325843, + 0.0036838401574641466, + 0.0034879546146839857, + 0.0043730176985263824, + 0.007812547497451305, + 0.005901395808905363, + 0.009470876306295395, + 0.01221206784248352, + 0.012679077684879303, + 0.012611457146704197, + 0.01200930867344141, + 0.00922060664743185, + 0.029532648622989655, + 0.013625354506075382 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 2, + 0 + ], + "selected_reward": -0.008646194823086262 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme macro close-up of a professional-grade carving knife slicing into a block of untreated basswood?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the knife have a blade with a cool metallic sheen and a handle crafted from polished dark walnut?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene illuminated by a warm, focused bench light streaming from above and sideways, revealing wood grain and growth rings?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the video shot in extreme slow motion showing a thin wood shaving curling and swirling through the air with visible fibers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera completely still with a background blurred into a dark haze, while the blade reflects dazzling highlights?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.014208110980689526 + ], + "opposite_logprob_tokens": [ + -4.2642083168029785 + ], + "logprob_reward": -0.014208110980689526, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.014208110980689526, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2622787654399872, + 0.08762971311807632, + 0.15117983520030975, + 0.08277920633554459, + 0.04425499588251114, + 0.10697150230407715, + 0.09532415866851807, + 0.07997050136327744, + 0.029563669115304947, + 0.06004764884710312 + ], + "credit_map": [ + 0.03720584139227867, + 0.029016312211751938, + 0.04289332032203674, + 0.02469031699001789, + 0.038109056651592255, + 0.03416566550731659, + 0.02620651014149189, + 0.02999173291027546, + 0.012483936734497547, + 0.010115505196154118, + 0.009498374536633492, + 0.013864673674106598, + 0.008632686920464039, + 0.012140735983848572, + 0.008060481399297714, + 0.01283331774175167, + 0.023540178313851357, + 0.01512639969587326, + 0.01627383939921856, + 0.017573785036802292, + 0.01891164667904377, + 0.0205832552164793, + 0.0223283339291811, + 0.016842398792505264, + 0.012406776659190655, + 0.009528188034892082, + 0.007825685665011406, + 0.009434666484594345, + 0.015599622391164303, + 0.0088001424446702, + 0.010837571695446968, + 0.008346551097929478, + 0.005082943942397833, + 0.006060101091861725, + 0.005100615322589874, + 0.0053660208359360695, + 0.006305579096078873, + 0.0049432069063186646, + 0.005983063485473394, + 0.005413464270532131, + 0.021799607202410698, + 0.01552053913474083, + 0.010911069810390472, + 0.00931799691170454, + 0.00963536649942398, + 0.009516901336610317, + 0.018228447064757347, + 0.012041568756103516, + 0.013346628285944462, + 0.013612664304673672, + 0.00800295826047659, + 0.010491326451301575, + 0.016090862452983856, + 0.007371655665338039, + 0.01376391388475895, + 0.012644155882298946, + 0.008752591907978058, + 0.00814307015389204, + 0.006127709057182074, + 0.008344053290784359, + 0.0070180087350308895, + 0.005226862151175737, + 0.021719273179769516, + 0.014638936147093773, + 0.005605234764516354, + 0.0030999279115349054, + 0.0028255865909159184, + 0.003396369982510805, + 0.002971199806779623, + 0.0027679072227329016, + 0.00489681214094162, + 0.00400063069537282, + 0.007517483551055193, + 0.0055505759082734585, + 0.006896678823977709, + 0.007623419165611267, + 0.007914959453046322, + 0.005040150601416826, + 0.01392489392310381, + 0.005579488351941109 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 5 + ], + "selected_reward": -0.014208110980689526 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "In an extreme macro close-up shot from a slight downward angle, the camera remains still, focused on a professional-grade carving knife with a blade gleaming with a cool metallic sheen. Its handle, crafted from polished dark walnut, exudes a warm, textured feel. A warm, focused bench light streams down from above and sideways, illuminating the knife's tip as it steadily and deliberately slices into a block of untreated basswood, revealing the delicate grain and growth rings on the surface. With the blade's precise and fluid advance, a thin shaving of wood is gracefully released. Shot in extreme slow motion, the shaving appears like a living ribbon, slowly unfurling, curling, and swirling through the air. The delicate wood fibers within are clearly visible in the light, as if time has frozen in time. The light reflects dazzling highlights on the sharp blade, while outlining charming golden outlines on the arcs of flying sawdust. The background is completely blurred into a dark haze, leaving only this focused and poetic moment of carving. The whole picture is full of tranquility, healing and praise of exquisite craftsmanship.", + "seed": 15061, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000006/group_9/index_6606_seed_15061_dp_18_rank_36_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 9, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.08335978463292122, + "vlm_logprob": -0.08335978463292122, + "vlm_yes_no_margin": 3.275, + "vlm_token_credit": -0.08335978463292122, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme macro close-up of a professional-grade carving knife slicing into a block of untreated basswood?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the knife have a blade with a cool metallic sheen and a handle crafted from polished dark walnut?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene illuminated by a warm, focused bench light streaming from above and sideways, revealing wood grain and growth rings?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the video shot in extreme slow motion showing a thin wood shaving curling and swirling through the air with visible fibers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera completely still with a background blurred into a dark haze, while the blade reflects dazzling highlights?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.3069969713687897 + ], + "opposite_logprob_tokens": [ + -1.4319969415664673 + ], + "logprob_reward": -0.3069969713687897, + "yes_no_margin_reward": 1.125, + "token_credit_reward": -0.3069969713687897, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12676428258419037, + 0.031398821622133255, + 0.228227898478508, + 0.16270135343074799, + 0.03074389509856701, + 0.07899970561265945, + 0.06752781569957733, + 0.1761082112789154, + 0.038235221058130264, + 0.05929278954863548 + ], + "credit_map": [ + 0.014392764307558537, + 0.020209958776831627, + 0.016126926988363266, + 0.013371041975915432, + 0.01256315317004919, + 0.021656230092048645, + 0.013019895181059837, + 0.015424303710460663, + 0.002847919473424554, + 0.006773058790713549, + 0.00303370482288301, + 0.0024901616852730513, + 0.0025024297647178173, + 0.008276517502963543, + 0.002946451073512435, + 0.0025285768788307905, + 0.0062644463032484055, + 0.03375694528222084, + 0.015361824072897434, + 0.011650331318378448, + 0.01922394149005413, + 0.08055415004491806, + 0.019113803282380104, + 0.0423024520277977, + 0.004346377216279507, + 0.01916978694498539, + 0.013693991117179394, + 0.009011998772621155, + 0.010920323431491852, + 0.054021239280700684, + 0.016533348709344864, + 0.035004280507564545, + 0.0034125633537769318, + 0.0032084276899695396, + 0.0028224317356944084, + 0.0029344975482672453, + 0.002291475422680378, + 0.006584797985851765, + 0.0034781727008521557, + 0.006011525634676218, + 0.01183344330638647, + 0.008201351389288902, + 0.008725606836378574, + 0.0064455438405275345, + 0.006795353721827269, + 0.012641629204154015, + 0.011507957242429256, + 0.012848813086748123, + 0.008763161487877369, + 0.012968222610652447, + 0.007253115531057119, + 0.006137995980679989, + 0.006332678254693747, + 0.00971545372158289, + 0.008799867704510689, + 0.0075573185458779335, + 0.016939904540777206, + 0.014796895906329155, + 0.02107066661119461, + 0.026308735832571983, + 0.016299674287438393, + 0.01737549528479576, + 0.03438764810562134, + 0.02892918325960636, + 0.004502878990024328, + 0.005827303044497967, + 0.004325603134930134, + 0.004633347503840923, + 0.0032942385878413916, + 0.007082859054207802, + 0.004541201516985893, + 0.0040277885273098946, + 0.006106866989284754, + 0.00447665574029088, + 0.003692053956910968, + 0.0036034053191542625, + 0.004928202833980322, + 0.01605386473238468, + 0.006087285000830889, + 0.014344451017677784 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 7, + 3 + ], + "selected_reward": -0.3069969713687897 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme macro close-up of a professional-grade carving knife slicing into a block of untreated basswood?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the knife have a blade with a cool metallic sheen and a handle crafted from polished dark walnut?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene illuminated by a warm, focused bench light streaming from above and sideways, revealing wood grain and growth rings?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the video shot in extreme slow motion showing a thin wood shaving curling and swirling through the air with visible fibers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera completely still with a background blurred into a dark haze, while the blade reflects dazzling highlights?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0487232469022274 + ], + "opposite_logprob_tokens": [ + -3.0487232208251953 + ], + "logprob_reward": -0.0487232469022274, + "yes_no_margin_reward": 3.0, + "token_credit_reward": -0.0487232469022274, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07258212566375732, + 0.09027747064828873, + 0.23211213946342468, + 0.06270495057106018, + 0.03209863230586052, + 0.04710683971643448, + 0.043953560292720795, + 0.24857710301876068, + 0.03802967444062233, + 0.13255755603313446 + ], + "credit_map": [ + 0.008136018179357052, + 0.012345428578555584, + 0.00805036909878254, + 0.006070405710488558, + 0.008496223017573357, + 0.016053814440965652, + 0.005747522693127394, + 0.007682341616600752, + 0.003989879973232746, + 0.014414053410291672, + 0.005905807949602604, + 0.005429955665022135, + 0.007448679767549038, + 0.02983618713915348, + 0.007895260117948055, + 0.015357640571892262, + 0.006539472844451666, + 0.03471020236611366, + 0.014679447747766972, + 0.01164115872234106, + 0.01921345852315426, + 0.08224841207265854, + 0.019888177514076233, + 0.04319179058074951, + 0.0023160178679972887, + 0.007720087189227343, + 0.005880406126379967, + 0.00355503405444324, + 0.004280606284737587, + 0.020746935158967972, + 0.00607603183016181, + 0.012129826471209526, + 0.0038836407475173473, + 0.0038113833870738745, + 0.0025762678124010563, + 0.0022601454984396696, + 0.002312257420271635, + 0.0077024344354867935, + 0.003192208707332611, + 0.006360289175063372, + 0.0069805593229830265, + 0.005781860090792179, + 0.004534391686320305, + 0.0038541904650628567, + 0.004233602900058031, + 0.007113749627023935, + 0.007225345820188522, + 0.007383136544376612, + 0.0054194265976548195, + 0.007740267552435398, + 0.00572211341932416, + 0.0039296746253967285, + 0.004991627298295498, + 0.00592327956110239, + 0.0054915547370910645, + 0.004735612776130438, + 0.02237492986023426, + 0.021990476176142693, + 0.026608332991600037, + 0.037946030497550964, + 0.022353969514369965, + 0.02406216971576214, + 0.05037565901875496, + 0.04286551475524902, + 0.0051049296744167805, + 0.0043325889855623245, + 0.0034707377199083567, + 0.003260469762608409, + 0.00344838690944016, + 0.009263133630156517, + 0.004419163800776005, + 0.004730261396616697, + 0.014155326411128044, + 0.008817324414849281, + 0.006140434183180332, + 0.0054813651368021965, + 0.009375474415719509, + 0.042813606560230255, + 0.013078038580715656, + 0.03269597887992859 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 2, + 9 + ], + "selected_reward": -0.0487232469022274 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme macro close-up of a professional-grade carving knife slicing into a block of untreated basswood?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the knife have a blade with a cool metallic sheen and a handle crafted from polished dark walnut?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene illuminated by a warm, focused bench light streaming from above and sideways, revealing wood grain and growth rings?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the video shot in extreme slow motion showing a thin wood shaving curling and swirling through the air with visible fibers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera completely still with a background blurred into a dark haze, while the blade reflects dazzling highlights?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.006752174347639084 + ], + "opposite_logprob_tokens": [ + -5.006752014160156 + ], + "logprob_reward": -0.006752174347639084, + "yes_no_margin_reward": 5.0, + "token_credit_reward": -0.006752174347639084, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07427855581045151, + 0.14043347537517548, + 0.18683907389640808, + 0.05760050192475319, + 0.059083063155412674, + 0.07101918011903763, + 0.14882707595825195, + 0.1938561350107193, + 0.021068047732114792, + 0.04699484631419182 + ], + "credit_map": [ + 0.006608756259083748, + 0.00968613289296627, + 0.008852692320942879, + 0.0072211576625704765, + 0.0071320789866149426, + 0.01778707280755043, + 0.006428984459489584, + 0.010561692528426647, + 0.00500986585393548, + 0.02590116485953331, + 0.008518538437783718, + 0.007099073380231857, + 0.010243706405162811, + 0.049524348229169846, + 0.010865813121199608, + 0.023270968347787857, + 0.009811644442379475, + 0.02907976694405079, + 0.014100746251642704, + 0.013070645742118359, + 0.015087511390447617, + 0.05845211446285248, + 0.016203952953219414, + 0.03103271685540676, + 0.0026870700530707836, + 0.009204079397022724, + 0.005331777967512608, + 0.00403384305536747, + 0.004452323075383902, + 0.016672220081090927, + 0.005093114450573921, + 0.010126082226634026, + 0.00397120974957943, + 0.007021529600024223, + 0.0035102281253784895, + 0.004460995551198721, + 0.004120394587516785, + 0.015916861593723297, + 0.006917436607182026, + 0.013164414092898369, + 0.01120554655790329, + 0.007789540570229292, + 0.00908747874200344, + 0.0073738074861466885, + 0.00595272658392787, + 0.01027703657746315, + 0.00928808469325304, + 0.010044961236417294, + 0.014887477271258831, + 0.01243206113576889, + 0.017916837707161903, + 0.01888532191514969, + 0.01660730130970478, + 0.02015962265431881, + 0.02698771469295025, + 0.020950738340616226, + 0.01878228597342968, + 0.022057585418224335, + 0.02047562226653099, + 0.027991848066449165, + 0.019180862233042717, + 0.02296331338584423, + 0.03385432809591293, + 0.02855030633509159, + 0.0028885824140161276, + 0.003600403666496277, + 0.002733549801632762, + 0.0032806925009936094, + 0.001975852297618985, + 0.0026471700984984636, + 0.0023406038526445627, + 0.0016011939151212573, + 0.0067995162680745125, + 0.005621958523988724, + 0.005264257546514273, + 0.005377272609621286, + 0.00882717501372099, + 0.0061781336553394794, + 0.0051474422216415405, + 0.0037790925707668066 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 2, + 6 + ], + "selected_reward": -0.006752174347639084 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme macro close-up of a professional-grade carving knife slicing into a block of untreated basswood?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the knife have a blade with a cool metallic sheen and a handle crafted from polished dark walnut?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene illuminated by a warm, focused bench light streaming from above and sideways, revealing wood grain and growth rings?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the video shot in extreme slow motion showing a thin wood shaving curling and swirling through the air with visible fibers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera completely still with a background blurred into a dark haze, while the blade reflects dazzling highlights?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03371607884764671 + ], + "opposite_logprob_tokens": [ + -3.4087159633636475 + ], + "logprob_reward": -0.03371607884764671, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.03371607884764671, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14200809597969055, + 0.06091698259115219, + 0.1625552624464035, + 0.18691636621952057, + 0.0561920627951622, + 0.13523924350738525, + 0.07434544712305069, + 0.06504781544208527, + 0.03281055763363838, + 0.08396810293197632 + ], + "credit_map": [ + 0.013530721887946129, + 0.02315135858952999, + 0.01617189683020115, + 0.01397328544408083, + 0.013995681889355183, + 0.03143225237727165, + 0.012368644587695599, + 0.01738426834344864, + 0.0035287512000650167, + 0.0077648707665503025, + 0.0064026727341115475, + 0.005200352519750595, + 0.0066160294227302074, + 0.017662055790424347, + 0.005750671494752169, + 0.007991586811840534, + 0.0062200636602938175, + 0.023513633757829666, + 0.012212240137159824, + 0.009897438809275627, + 0.015070701949298382, + 0.0532509982585907, + 0.013424637727439404, + 0.028965551406145096, + 0.005591176450252533, + 0.022787252441048622, + 0.016404058784246445, + 0.009840961545705795, + 0.013139490969479084, + 0.06344255805015564, + 0.017998041585087776, + 0.037712834775447845, + 0.0034608871210366488, + 0.0061965580098330975, + 0.003411708166822791, + 0.005117591004818678, + 0.004051051568239927, + 0.01486128754913807, + 0.0065804473124444485, + 0.012512537650763988, + 0.017186421900987625, + 0.013209753669798374, + 0.013880913145840168, + 0.009338769130408764, + 0.011314100585877895, + 0.026956802234053612, + 0.017783870920538902, + 0.02556862309575081, + 0.009976849891245365, + 0.009389356710016727, + 0.008865227922797203, + 0.009449327364563942, + 0.008634787984192371, + 0.010176977142691612, + 0.009796250611543655, + 0.008056669495999813, + 0.0074053676798939705, + 0.007309901062399149, + 0.009326097555458546, + 0.009005843661725521, + 0.007357366848737001, + 0.0063908277079463005, + 0.009936641901731491, + 0.008315774612128735, + 0.0036720114294439554, + 0.006953909993171692, + 0.0038375507574528456, + 0.00434667244553566, + 0.0025121907237917185, + 0.004514085594564676, + 0.0038508970756083727, + 0.0031232417095452547, + 0.013739630579948425, + 0.008760424330830574, + 0.008624985814094543, + 0.008153564296662807, + 0.012045874260365963, + 0.014579986222088337, + 0.009260588325560093, + 0.008803052827715874 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 2, + 0 + ], + "selected_reward": -0.03371607884764671 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme macro close-up of a professional-grade carving knife slicing into a block of untreated basswood?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the knife have a blade with a cool metallic sheen and a handle crafted from polished dark walnut?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene illuminated by a warm, focused bench light streaming from above and sideways, revealing wood grain and growth rings?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the video shot in extreme slow motion showing a thin wood shaving curling and swirling through the air with visible fibers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera completely still with a background blurred into a dark haze, while the blade reflects dazzling highlights?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.020610451698303223 + ], + "opposite_logprob_tokens": [ + -3.8956103324890137 + ], + "logprob_reward": -0.020610451698303223, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.020610451698303223, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06175315007567406, + 0.11221279203891754, + 0.23824533820152283, + 0.07180865854024887, + 0.03203791752457619, + 0.06020565330982208, + 0.07062218338251114, + 0.1324244886636734, + 0.09479723125696182, + 0.12589259445667267 + ], + "credit_map": [ + 0.00617557018995285, + 0.010161150246858597, + 0.007227306254208088, + 0.0071907504461705685, + 0.006754032336175442, + 0.008370231837034225, + 0.006216760259121656, + 0.009657352231442928, + 0.008560078218579292, + 0.02269955538213253, + 0.007334526162594557, + 0.009708232246339321, + 0.008833040483295918, + 0.02990657091140747, + 0.010058283805847168, + 0.015112499706447124, + 0.011782952584326267, + 0.03845452517271042, + 0.01651548407971859, + 0.01533152349293232, + 0.019261907786130905, + 0.07632336020469666, + 0.02022644504904747, + 0.04034913703799248, + 0.007021697238087654, + 0.013342753984034061, + 0.0069365291856229305, + 0.005435016471892595, + 0.006196227855980396, + 0.01783422753214836, + 0.006246546749025583, + 0.008795665577054024, + 0.004231893457472324, + 0.0033032787032425404, + 0.0034962203353643417, + 0.003937968984246254, + 0.0029259826987981796, + 0.006439673714339733, + 0.0029840373899787664, + 0.004718857817351818, + 0.008063378743827343, + 0.006460877601057291, + 0.008057162165641785, + 0.0050298492424190044, + 0.005521742627024651, + 0.012497860006988049, + 0.005568253807723522, + 0.009006530046463013, + 0.009605661034584045, + 0.013741767965257168, + 0.009072755463421345, + 0.00686886440962553, + 0.007001824676990509, + 0.010374795645475388, + 0.008795307017862797, + 0.005161200184375048, + 0.010571258142590523, + 0.01636405847966671, + 0.012487483210861683, + 0.021241914480924606, + 0.01415974646806717, + 0.013009076938033104, + 0.025519168004393578, + 0.01907178945839405, + 0.012821138836443424, + 0.01467165071517229, + 0.014410172589123249, + 0.013252743519842625, + 0.008829472586512566, + 0.012232588604092598, + 0.010464727878570557, + 0.008114730007946491, + 0.014274537563323975, + 0.007674312684684992, + 0.006545455195009708, + 0.0044929636642336845, + 0.008018848486244678, + 0.04239702969789505, + 0.01241180207580328, + 0.030077649280428886 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 7, + 9 + ], + "selected_reward": -0.020610451698303223 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "In an extreme macro close-up shot from a slight downward angle, the camera remains still, focused on a professional-grade carving knife with a blade gleaming with a cool metallic sheen. Its handle, crafted from polished dark walnut, exudes a warm, textured feel. A warm, focused bench light streams down from above and sideways, illuminating the knife's tip as it steadily and deliberately slices into a block of untreated basswood, revealing the delicate grain and growth rings on the surface. With the blade's precise and fluid advance, a thin shaving of wood is gracefully released. Shot in extreme slow motion, the shaving appears like a living ribbon, slowly unfurling, curling, and swirling through the air. The delicate wood fibers within are clearly visible in the light, as if time has frozen in time. The light reflects dazzling highlights on the sharp blade, while outlining charming golden outlines on the arcs of flying sawdust. The background is completely blurred into a dark haze, leaving only this focused and poetic moment of carving. The whole picture is full of tranquility, healing and praise of exquisite craftsmanship.", + "seed": 15062, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000006/group_9/index_6606_seed_15062_dp_19_rank_38_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 9, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.06377297630533577, + "vlm_logprob": -0.06377297630533577, + "vlm_yes_no_margin": 3.625, + "vlm_token_credit": -0.06377297630533577, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme macro close-up of a professional-grade carving knife slicing into a block of untreated basswood?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the knife have a blade with a cool metallic sheen and a handle crafted from polished dark walnut?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene illuminated by a warm, focused bench light streaming from above and sideways, revealing wood grain and growth rings?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the video shot in extreme slow motion showing a thin wood shaving curling and swirling through the air with visible fibers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera completely still with a background blurred into a dark haze, while the blade reflects dazzling highlights?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.22058787941932678 + ], + "opposite_logprob_tokens": [ + -1.7205878496170044 + ], + "logprob_reward": -0.22058787941932678, + "yes_no_margin_reward": 1.5, + "token_credit_reward": -0.22058787941932678, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08100318163633347, + 0.04078397527337074, + 0.07671567797660828, + 0.16253294050693512, + 0.12197721749544144, + 0.10517491400241852, + 0.052000079303979874, + 0.15431161224842072, + 0.1482708901166916, + 0.05722948908805847 + ], + "credit_map": [ + 0.007803693879395723, + 0.012744547799229622, + 0.017680594697594643, + 0.00819373968988657, + 0.007966523990035057, + 0.006984901148825884, + 0.007909109815955162, + 0.011720073409378529, + 0.0048179361037909985, + 0.006244046147912741, + 0.009056155569851398, + 0.0031848817598074675, + 0.0051939296536147594, + 0.005880269687622786, + 0.0034803731832653284, + 0.0029263815376907587, + 0.007664164528250694, + 0.00957503542304039, + 0.008860359899699688, + 0.007205422036349773, + 0.01106501929461956, + 0.010819470509886742, + 0.009838961064815521, + 0.01168724987655878, + 0.010760609991848469, + 0.055301520973443985, + 0.009305628016591072, + 0.008556118234992027, + 0.010773401707410812, + 0.00731535954400897, + 0.010378963313996792, + 0.050141338258981705, + 0.026165585964918137, + 0.015377804636955261, + 0.009853729046881199, + 0.009786728769540787, + 0.01049754861742258, + 0.00977841392159462, + 0.00960929598659277, + 0.03090810775756836, + 0.009084201417863369, + 0.02314620651304722, + 0.01208493858575821, + 0.01253641489893198, + 0.012770899571478367, + 0.00656978040933609, + 0.008560840040445328, + 0.020421626046299934, + 0.004152999259531498, + 0.010795681737363338, + 0.005803575273603201, + 0.005207493435591459, + 0.005237888544797897, + 0.007538069039583206, + 0.006456518080085516, + 0.00680785346776247, + 0.01218120101839304, + 0.03957265242934227, + 0.018697481602430344, + 0.01685296557843685, + 0.01680160127580166, + 0.007342513184994459, + 0.008890734985470772, + 0.03397246450185776, + 0.0184620413929224, + 0.013549063354730606, + 0.009583462961018085, + 0.02303355187177658, + 0.01778355985879898, + 0.00917038507759571, + 0.012680322863161564, + 0.04400849714875221, + 0.005530256778001785, + 0.006351412273943424, + 0.01468628365546465, + 0.006185811944305897, + 0.008129059337079525, + 0.005159483756870031, + 0.0046582771465182304, + 0.006528904195874929 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 7, + 8 + ], + "selected_reward": -0.22058787941932678 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme macro close-up of a professional-grade carving knife slicing into a block of untreated basswood?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the knife have a blade with a cool metallic sheen and a handle crafted from polished dark walnut?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene illuminated by a warm, focused bench light streaming from above and sideways, revealing wood grain and growth rings?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the video shot in extreme slow motion showing a thin wood shaving curling and swirling through the air with visible fibers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera completely still with a background blurred into a dark haze, while the blade reflects dazzling highlights?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.05503237619996071 + ], + "opposite_logprob_tokens": [ + -2.93003249168396 + ], + "logprob_reward": -0.05503237619996071, + "yes_no_margin_reward": 2.875, + "token_credit_reward": -0.05503237619996071, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11098229885101318, + 0.04760153964161873, + 0.1547916978597641, + 0.09300967305898666, + 0.15359491109848022, + 0.08949515223503113, + 0.040048666298389435, + 0.05642947927117348, + 0.12455576658248901, + 0.12949082255363464 + ], + "credit_map": [ + 0.012068510986864567, + 0.013629905879497528, + 0.02950844168663025, + 0.012962497770786285, + 0.011607961729168892, + 0.009478563442826271, + 0.01024344377219677, + 0.01148297544568777, + 0.004527759272605181, + 0.006579067092388868, + 0.009553391486406326, + 0.003912460058927536, + 0.0047226655296981335, + 0.010534903965890408, + 0.0036286739632487297, + 0.004142617806792259, + 0.013062809593975544, + 0.03555143624544144, + 0.01256607472896576, + 0.010681922547519207, + 0.015686409547924995, + 0.022233227267861366, + 0.011090798303484917, + 0.03391902893781662, + 0.0073278844356536865, + 0.029420169070363045, + 0.005232854280620813, + 0.006726002786308527, + 0.007147545926272869, + 0.005215292796492577, + 0.006946449168026447, + 0.024993473663926125, + 0.027716852724552155, + 0.026774026453495026, + 0.02070862241089344, + 0.014428111724555492, + 0.017217179760336876, + 0.014810480177402496, + 0.011966303922235966, + 0.019973324611783028, + 0.011472977697849274, + 0.02265855483710766, + 0.007571152411401272, + 0.010051485151052475, + 0.009240344166755676, + 0.005057196598500013, + 0.007684133946895599, + 0.01575930416584015, + 0.0035791657865047455, + 0.0102247204631567, + 0.003638062858954072, + 0.004070756491273642, + 0.004622486419975758, + 0.003803717205300927, + 0.0039744810201227665, + 0.00613527512177825, + 0.005478000268340111, + 0.014195188879966736, + 0.006346132140606642, + 0.005786816589534283, + 0.00829089991748333, + 0.0034443687181919813, + 0.004482812248170376, + 0.008405259810388088, + 0.017392389476299286, + 0.011916399002075195, + 0.007783080916851759, + 0.019054053351283073, + 0.011368661187589169, + 0.006718977354466915, + 0.009406490251421928, + 0.04091572016477585, + 0.012954071164131165, + 0.020016640424728394, + 0.03602200001478195, + 0.015820639207959175, + 0.014245602302253246, + 0.011360923759639263, + 0.00931191723793745, + 0.009759023785591125 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 4, + 9 + ], + "selected_reward": -0.05503237619996071 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme macro close-up of a professional-grade carving knife slicing into a block of untreated basswood?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the knife have a blade with a cool metallic sheen and a handle crafted from polished dark walnut?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene illuminated by a warm, focused bench light streaming from above and sideways, revealing wood grain and growth rings?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the video shot in extreme slow motion showing a thin wood shaving curling and swirling through the air with visible fibers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera completely still with a background blurred into a dark haze, while the blade reflects dazzling highlights?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.003629408311098814 + ], + "opposite_logprob_tokens": [ + -5.628629207611084 + ], + "logprob_reward": -0.003629408311098814, + "yes_no_margin_reward": 5.625, + "token_credit_reward": -0.003629408311098814, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2330232858657837, + 0.046326879411935806, + 0.1824345886707306, + 0.07220229506492615, + 0.08351536095142365, + 0.06646191328763962, + 0.029228704050183296, + 0.12793420255184174, + 0.12188584357500076, + 0.03698700666427612 + ], + "credit_map": [ + 0.02870306745171547, + 0.02649974264204502, + 0.07215462625026703, + 0.021912354975938797, + 0.028642799705266953, + 0.015672579407691956, + 0.016866059973835945, + 0.022572053596377373, + 0.006374410819262266, + 0.0063574411906301975, + 0.007472566794604063, + 0.0033539333380758762, + 0.006019358988851309, + 0.00727978115901351, + 0.003870889777317643, + 0.005598492920398712, + 0.013946442864835262, + 0.04680201783776283, + 0.013794959522783756, + 0.014093317091464996, + 0.015707489103078842, + 0.02773326449096203, + 0.014349786564707756, + 0.03600731119513512, + 0.00602675462141633, + 0.02182294800877571, + 0.00393751310184598, + 0.0050055342726409435, + 0.0058565703220665455, + 0.003930812235921621, + 0.0069833118468523026, + 0.018638841807842255, + 0.020497579127550125, + 0.009634096175432205, + 0.005450102500617504, + 0.004582900088280439, + 0.006595084443688393, + 0.005785420071333647, + 0.006064035464078188, + 0.024906136095523834, + 0.010717315599322319, + 0.008436223492026329, + 0.006806138902902603, + 0.00638287840411067, + 0.009023449383676052, + 0.008883276954293251, + 0.007153743878006935, + 0.009058877825737, + 0.0021368050947785378, + 0.003672051476314664, + 0.005169637035578489, + 0.005033464170992374, + 0.003359901485964656, + 0.0036089508794248104, + 0.003818212077021599, + 0.0024296799674630165, + 0.009431869722902775, + 0.03657024726271629, + 0.01007860992103815, + 0.009460576809942722, + 0.014230857603251934, + 0.0062840646132826805, + 0.0061763571575284, + 0.035701602697372437, + 0.016005484387278557, + 0.013706151396036148, + 0.007575270254164934, + 0.018599554896354675, + 0.010688893496990204, + 0.006541253067553043, + 0.009620393626391888, + 0.039148833602666855, + 0.003248287132009864, + 0.00822554249316454, + 0.0049134246073663235, + 0.004792665597051382, + 0.004290205892175436, + 0.003618141170591116, + 0.0033592507243156433, + 0.004539490211755037 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 7 + ], + "selected_reward": -0.003629408311098814 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme macro close-up of a professional-grade carving knife slicing into a block of untreated basswood?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the knife have a blade with a cool metallic sheen and a handle crafted from polished dark walnut?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene illuminated by a warm, focused bench light streaming from above and sideways, revealing wood grain and growth rings?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the video shot in extreme slow motion showing a thin wood shaving curling and swirling through the air with visible fibers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera completely still with a background blurred into a dark haze, while the blade reflects dazzling highlights?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.00979132391512394 + ], + "opposite_logprob_tokens": [ + -4.634791374206543 + ], + "logprob_reward": -0.00979132391512394, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.00979132391512394, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.054163858294487, + 0.054664094001054764, + 0.20289447903633118, + 0.07780393213033676, + 0.21412700414657593, + 0.035429760813713074, + 0.043306220322847366, + 0.06385678797960281, + 0.20723196864128113, + 0.0465218648314476 + ], + "credit_map": [ + 0.0061277528293430805, + 0.009433732368052006, + 0.012148183770477772, + 0.005273371934890747, + 0.004874409642070532, + 0.004831152502447367, + 0.0050737070851027966, + 0.006401544902473688, + 0.004765532445162535, + 0.004872940480709076, + 0.013865838758647442, + 0.004374003037810326, + 0.005988146178424358, + 0.009061361663043499, + 0.005149242002516985, + 0.006587029900401831, + 0.015475118532776833, + 0.05042107030749321, + 0.018711678683757782, + 0.009876190684735775, + 0.020857995375990868, + 0.03219159319996834, + 0.01353810727596283, + 0.04182272404432297, + 0.005829905159771442, + 0.022623352706432343, + 0.006459828466176987, + 0.005622742231935263, + 0.006244921125471592, + 0.0041540563106536865, + 0.0064919013530015945, + 0.020377222448587418, + 0.06515982747077942, + 0.010435064323246479, + 0.010888459160923958, + 0.010561500675976276, + 0.013603517785668373, + 0.010309343226253986, + 0.01520257443189621, + 0.07796671241521835, + 0.004576064646244049, + 0.004707964602857828, + 0.003998246509581804, + 0.005078612826764584, + 0.0061297789216041565, + 0.0028153706807643175, + 0.003762569511309266, + 0.004361153580248356, + 0.004592036362737417, + 0.0057027386501431465, + 0.00617843447253108, + 0.0068477969616651535, + 0.005941255949437618, + 0.005290793254971504, + 0.004899036139249802, + 0.003854126203805208, + 0.007529248017817736, + 0.00712693901732564, + 0.007603096775710583, + 0.009901764802634716, + 0.008281347341835499, + 0.004525166004896164, + 0.010328914038836956, + 0.008560316637158394, + 0.0183822400867939, + 0.03240310773253441, + 0.022449467331171036, + 0.03743961080908775, + 0.03298059105873108, + 0.016093062236905098, + 0.022815437987446785, + 0.024668456986546516, + 0.0045159002766013145, + 0.007647841703146696, + 0.006914551369845867, + 0.00588474003598094, + 0.006549082230776548, + 0.004171136766672134, + 0.0040857819840312, + 0.006752831395715475 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 8, + 2 + ], + "selected_reward": -0.00979132391512394 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme macro close-up of a professional-grade carving knife slicing into a block of untreated basswood?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the knife have a blade with a cool metallic sheen and a handle crafted from polished dark walnut?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene illuminated by a warm, focused bench light streaming from above and sideways, revealing wood grain and growth rings?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the video shot in extreme slow motion showing a thin wood shaving curling and swirling through the air with visible fibers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera completely still with a background blurred into a dark haze, while the blade reflects dazzling highlights?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.029823893681168556 + ], + "opposite_logprob_tokens": [ + -3.5298237800598145 + ], + "logprob_reward": -0.029823893681168556, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.029823893681168556, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.03925027698278427, + 0.05032893642783165, + 0.03816932439804077, + 0.09023196995258331, + 0.08046228438615799, + 0.09462901204824448, + 0.049503397196531296, + 0.2895047664642334, + 0.22118210792541504, + 0.04673794284462929 + ], + "credit_map": [ + 0.0032354583963751793, + 0.0056658657267689705, + 0.006625909358263016, + 0.004322378896176815, + 0.004743471741676331, + 0.0038750481326133013, + 0.005587543826550245, + 0.005194602068513632, + 0.0057264892384409904, + 0.008065271191298962, + 0.007006497588008642, + 0.003894092747941613, + 0.006837870925664902, + 0.009474238380789757, + 0.003612261265516281, + 0.005712214857339859, + 0.003065916011109948, + 0.008755583316087723, + 0.004027358256280422, + 0.004657423123717308, + 0.002994556212797761, + 0.004331011790782213, + 0.0036407229490578175, + 0.006696751806885004, + 0.004770447500050068, + 0.029739560559391975, + 0.0056901415809988976, + 0.004844958428293467, + 0.007115988992154598, + 0.004628603346645832, + 0.0057147773914039135, + 0.02772749960422516, + 0.018020614981651306, + 0.009062644094228745, + 0.0065628839656710625, + 0.0049120220355689526, + 0.006945616565644741, + 0.005945455748587847, + 0.007083636708557606, + 0.021929409354925156, + 0.009953421540558338, + 0.03241734579205513, + 0.005059350747615099, + 0.0051089078187942505, + 0.006084787659347057, + 0.004532179329544306, + 0.00410686107352376, + 0.02736615762114525, + 0.0040284413844347, + 0.009081995114684105, + 0.0055057210847735405, + 0.006424024701118469, + 0.006144104991108179, + 0.006136842537671328, + 0.00472346693277359, + 0.007458799984306097, + 0.022468427196145058, + 0.08871590346097946, + 0.017466170713305473, + 0.021622231230139732, + 0.026658568531274796, + 0.013025335036218166, + 0.01671757735311985, + 0.08283054828643799, + 0.030485937371850014, + 0.026699697598814964, + 0.02535431459546089, + 0.03048897162079811, + 0.03314956650137901, + 0.013185502029955387, + 0.018177013844251633, + 0.04364108666777611, + 0.0037447360809892416, + 0.008280095644295216, + 0.008496246300637722, + 0.007331567816436291, + 0.005055244546383619, + 0.00401307363063097, + 0.005085964221507311, + 0.00473101856186986 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 8, + 5 + ], + "selected_reward": -0.029823893681168556 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "In an extreme macro close-up shot from a slight downward angle, the camera remains still, focused on a professional-grade carving knife with a blade gleaming with a cool metallic sheen. Its handle, crafted from polished dark walnut, exudes a warm, textured feel. A warm, focused bench light streams down from above and sideways, illuminating the knife's tip as it steadily and deliberately slices into a block of untreated basswood, revealing the delicate grain and growth rings on the surface. With the blade's precise and fluid advance, a thin shaving of wood is gracefully released. Shot in extreme slow motion, the shaving appears like a living ribbon, slowly unfurling, curling, and swirling through the air. The delicate wood fibers within are clearly visible in the light, as if time has frozen in time. The light reflects dazzling highlights on the sharp blade, while outlining charming golden outlines on the arcs of flying sawdust. The background is completely blurred into a dark haze, leaving only this focused and poetic moment of carving. The whole picture is full of tranquility, healing and praise of exquisite craftsmanship.", + "seed": 15063, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000006/group_9/index_6606_seed_15063_dp_19_rank_38_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 9, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.08034594776108861, + "vlm_logprob": -0.08034594776108861, + "vlm_yes_no_margin": 3.375, + "vlm_token_credit": -0.08034594776108861, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme macro close-up of a professional-grade carving knife slicing into a block of untreated basswood?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the knife have a blade with a cool metallic sheen and a handle crafted from polished dark walnut?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene illuminated by a warm, focused bench light streaming from above and sideways, revealing wood grain and growth rings?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the video shot in extreme slow motion showing a thin wood shaving curling and swirling through the air with visible fibers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera completely still with a background blurred into a dark haze, while the blade reflects dazzling highlights?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.29909011721611023 + ], + "opposite_logprob_tokens": [ + -1.4240901470184326 + ], + "logprob_reward": -0.29909011721611023, + "yes_no_margin_reward": 1.125, + "token_credit_reward": -0.29909011721611023, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08863533288240433, + 0.04067165032029152, + 0.06758362054824829, + 0.07309950143098831, + 0.05167895928025246, + 0.30498945713043213, + 0.11189231276512146, + 0.0762835219502449, + 0.0639934092760086, + 0.12117226421833038 + ], + "credit_map": [ + 0.011116604320704937, + 0.008685819804668427, + 0.010498180985450745, + 0.01682194694876671, + 0.005128329619765282, + 0.006577228661626577, + 0.017259765416383743, + 0.012547442689538002, + 0.007508092559874058, + 0.004969424568116665, + 0.006137390621006489, + 0.005096115171909332, + 0.003301603952422738, + 0.0022115851752460003, + 0.006815864704549313, + 0.0046315682120621204, + 0.010421454906463623, + 0.006332890596240759, + 0.011689688079059124, + 0.009724766947329044, + 0.009038345888257027, + 0.0044253976084291935, + 0.00948272179812193, + 0.0064683533273637295, + 0.00574167724698782, + 0.004559784661978483, + 0.005890010390430689, + 0.011429470032453537, + 0.009950085543096066, + 0.0034378485288470984, + 0.01865203119814396, + 0.013438590802252293, + 0.004227849654853344, + 0.0066156829707324505, + 0.005469817668199539, + 0.009025020524859428, + 0.0046532335691154, + 0.003567917039617896, + 0.010941953398287296, + 0.007177478168159723, + 0.05583958327770233, + 0.04022959992289543, + 0.0519709549844265, + 0.031291015446186066, + 0.045693229883909225, + 0.019499044865369797, + 0.03216290473937988, + 0.028303101658821106, + 0.02442760393023491, + 0.018582478165626526, + 0.01724650152027607, + 0.008676827885210514, + 0.008243759162724018, + 0.007370206527411938, + 0.012324424460530281, + 0.015020502731204033, + 0.008531160652637482, + 0.006825899239629507, + 0.011814335361123085, + 0.012947469018399715, + 0.009369395673274994, + 0.004650858696550131, + 0.012759731151163578, + 0.009384670294821262, + 0.009043106809258461, + 0.009344217367470264, + 0.01180731039494276, + 0.007939706556499004, + 0.00441973889246583, + 0.0035294867120683193, + 0.009370513260364532, + 0.008539325557649136, + 0.008843973278999329, + 0.01690055802464485, + 0.011871680617332458, + 0.026866508647799492, + 0.01193039771169424, + 0.005134043283760548, + 0.02175571396946907, + 0.017869388684630394 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 9, + 6 + ], + "selected_reward": -0.29909011721611023 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme macro close-up of a professional-grade carving knife slicing into a block of untreated basswood?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the knife have a blade with a cool metallic sheen and a handle crafted from polished dark walnut?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene illuminated by a warm, focused bench light streaming from above and sideways, revealing wood grain and growth rings?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the video shot in extreme slow motion showing a thin wood shaving curling and swirling through the air with visible fibers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera completely still with a background blurred into a dark haze, while the blade reflects dazzling highlights?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.048736415803432465 + ], + "opposite_logprob_tokens": [ + -3.048736333847046 + ], + "logprob_reward": -0.048736415803432465, + "yes_no_margin_reward": 3.0, + "token_credit_reward": -0.048736415803432465, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11480630189180374, + 0.036890096962451935, + 0.0904541090130806, + 0.055391695350408554, + 0.04029402136802673, + 0.20033462345600128, + 0.08802579343318939, + 0.10900122672319412, + 0.2069041132926941, + 0.057897988706827164 + ], + "credit_map": [ + 0.013805540278553963, + 0.01348043978214264, + 0.01733720675110817, + 0.01728334277868271, + 0.007714836858212948, + 0.010715866461396217, + 0.019307298585772514, + 0.015161780640482903, + 0.0049765268340706825, + 0.007620805874466896, + 0.004083988256752491, + 0.0048991236835718155, + 0.003394724801182747, + 0.0019286731258034706, + 0.006020939443260431, + 0.003965315874665976, + 0.01341849472373724, + 0.009718935005366802, + 0.014856371097266674, + 0.0124197443947196, + 0.013683080673217773, + 0.006066327914595604, + 0.011867527849972248, + 0.008423627354204655, + 0.006313947960734367, + 0.0034550430718809366, + 0.005655328743159771, + 0.008727539330720901, + 0.008039629086852074, + 0.00255739432759583, + 0.012025355361402035, + 0.008617454208433628, + 0.004327928647398949, + 0.003334126900881529, + 0.006616746075451374, + 0.009331521578133106, + 0.002905127126723528, + 0.0020227008499205112, + 0.007035623770207167, + 0.004720249678939581, + 0.0361567847430706, + 0.026892200112342834, + 0.035313017666339874, + 0.019865600392222404, + 0.0302467979490757, + 0.012347404845058918, + 0.02096472680568695, + 0.018548088148236275, + 0.017282750457525253, + 0.014081603847444057, + 0.015413066372275352, + 0.007597722578793764, + 0.0074758753180503845, + 0.006659235339611769, + 0.009581810794770718, + 0.009933729656040668, + 0.010649623349308968, + 0.009174756705760956, + 0.02213173545897007, + 0.01688612811267376, + 0.014244770631194115, + 0.006963837891817093, + 0.016932938247919083, + 0.012017441913485527, + 0.02856702357530594, + 0.019260087981820107, + 0.047987114638090134, + 0.02287188731133938, + 0.023354945704340935, + 0.013001534156501293, + 0.03076794743537903, + 0.021093573421239853, + 0.004045580513775349, + 0.007873140275478363, + 0.005955550353974104, + 0.009236494079232216, + 0.008988330140709877, + 0.003340858966112137, + 0.01031983271241188, + 0.008138202130794525 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 5, + 0 + ], + "selected_reward": -0.048736415803432465 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme macro close-up of a professional-grade carving knife slicing into a block of untreated basswood?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the knife have a blade with a cool metallic sheen and a handle crafted from polished dark walnut?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene illuminated by a warm, focused bench light streaming from above and sideways, revealing wood grain and growth rings?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the video shot in extreme slow motion showing a thin wood shaving curling and swirling through the air with visible fibers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera completely still with a background blurred into a dark haze, while the blade reflects dazzling highlights?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005964694079011679 + ], + "opposite_logprob_tokens": [ + -5.130964756011963 + ], + "logprob_reward": -0.005964694079011679, + "yes_no_margin_reward": 5.125, + "token_credit_reward": -0.005964694079011679, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0629112496972084, + 0.01426628790795803, + 0.025026904419064522, + 0.025614341720938683, + 0.0324852429330349, + 0.09355608373880386, + 0.03931820020079613, + 0.13179580867290497, + 0.5053271651268005, + 0.06969873607158661 + ], + "credit_map": [ + 0.008435809053480625, + 0.01097064558416605, + 0.007662301417440176, + 0.00875591579824686, + 0.006241126451641321, + 0.0049428921192884445, + 0.008580663241446018, + 0.007321896497160196, + 0.002869711024686694, + 0.0020659607835114002, + 0.002177903428673744, + 0.0014725278597325087, + 0.0010331335943192244, + 0.0008957128738984466, + 0.0019359295256435871, + 0.0018154089339077473, + 0.0031834484543651342, + 0.0036556858103722334, + 0.003304221434518695, + 0.003331659361720085, + 0.004033101722598076, + 0.0015703197568655014, + 0.0036228715907782316, + 0.0023255946580320597, + 0.002290920354425907, + 0.0016988682327792048, + 0.0018452780786901712, + 0.003708040341734886, + 0.0031621030066162348, + 0.0014594652457162738, + 0.006670860573649406, + 0.004778802860528231, + 0.0032544457353651524, + 0.0034446308854967356, + 0.0041959090158343315, + 0.0058601186610758305, + 0.003251549554988742, + 0.001961747882887721, + 0.005989833269268274, + 0.004527001176029444, + 0.01744682714343071, + 0.012047989293932915, + 0.015518847852945328, + 0.009804613888263702, + 0.013225865550339222, + 0.006226846016943455, + 0.010040566325187683, + 0.009244514629244804, + 0.007001756690442562, + 0.005596307571977377, + 0.00604538107290864, + 0.0037750781048089266, + 0.0034546644892543554, + 0.002923210384324193, + 0.005963900126516819, + 0.004557897336781025, + 0.012849985621869564, + 0.011079136282205582, + 0.022189635783433914, + 0.019809642806649208, + 0.01628703437745571, + 0.007835852913558483, + 0.02473095990717411, + 0.017013560980558395, + 0.06678012013435364, + 0.052898794412612915, + 0.12064650654792786, + 0.05445173382759094, + 0.055788278579711914, + 0.032329827547073364, + 0.07171925157308578, + 0.05071260407567024, + 0.0046053663827478886, + 0.007960822433233261, + 0.007105688098818064, + 0.01690034382045269, + 0.009034370072185993, + 0.0042792316526174545, + 0.009718055836856365, + 0.010094855912029743 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 7, + 5 + ], + "selected_reward": -0.005964694079011679 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme macro close-up of a professional-grade carving knife slicing into a block of untreated basswood?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the knife have a blade with a cool metallic sheen and a handle crafted from polished dark walnut?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene illuminated by a warm, focused bench light streaming from above and sideways, revealing wood grain and growth rings?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the video shot in extreme slow motion showing a thin wood shaving curling and swirling through the air with visible fibers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera completely still with a background blurred into a dark haze, while the blade reflects dazzling highlights?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.014209521003067493 + ], + "opposite_logprob_tokens": [ + -4.264209747314453 + ], + "logprob_reward": -0.014209521003067493, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.014209521003067493, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.045131903141736984, + 0.03247124329209328, + 0.02906467206776142, + 0.08634538948535919, + 0.028622359037399292, + 0.11624829471111298, + 0.2067645788192749, + 0.08290091902017593, + 0.28545519709587097, + 0.08699548244476318 + ], + "credit_map": [ + 0.004937011748552322, + 0.004435990937054157, + 0.007241115439683199, + 0.007755161728709936, + 0.0026607259642332792, + 0.0032172957435250282, + 0.009381061419844627, + 0.0055035375989973545, + 0.005673219915479422, + 0.004456457681953907, + 0.006079861894249916, + 0.003960993140935898, + 0.002142413053661585, + 0.0016860663890838623, + 0.005226487293839455, + 0.003245738334953785, + 0.003713160753250122, + 0.0029924982227385044, + 0.003634638385847211, + 0.004856393672525883, + 0.004108431749045849, + 0.0025080435443669558, + 0.004074584227055311, + 0.003176918253302574, + 0.008245290257036686, + 0.0063267298974096775, + 0.005254649557173252, + 0.012165556661784649, + 0.01048301812261343, + 0.004560025408864021, + 0.023285388946533203, + 0.016024718061089516, + 0.002611408242955804, + 0.0035277216229587793, + 0.0034201438538730145, + 0.005274880211800337, + 0.0029261913150548935, + 0.0019700746051967144, + 0.005506590474396944, + 0.0033853482455015182, + 0.02321685291826725, + 0.01441268716007471, + 0.021042093634605408, + 0.011939859949052334, + 0.014614918269217014, + 0.007664615288376808, + 0.011814801953732967, + 0.011542455293238163, + 0.04280475154519081, + 0.047329965978860855, + 0.0275708194822073, + 0.01675054058432579, + 0.01628703437745571, + 0.015575619414448738, + 0.01870899461209774, + 0.021736832335591316, + 0.011643343605101109, + 0.009743727743625641, + 0.01519606914371252, + 0.01259279903024435, + 0.008907843381166458, + 0.005040611140429974, + 0.011376085691154003, + 0.008400441147387028, + 0.03482554107904434, + 0.03925701975822449, + 0.059923697263002396, + 0.030071064829826355, + 0.03290189057588577, + 0.017131390050053596, + 0.03957690671086311, + 0.03176768124103546, + 0.006124315783381462, + 0.008680169470608234, + 0.009675475768744946, + 0.01989486627280712, + 0.010319839231669903, + 0.0034419905859977007, + 0.016178002581000328, + 0.01268081460148096 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 6, + 5 + ], + "selected_reward": -0.014209521003067493 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme macro close-up of a professional-grade carving knife slicing into a block of untreated basswood?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the knife have a blade with a cool metallic sheen and a handle crafted from polished dark walnut?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene illuminated by a warm, focused bench light streaming from above and sideways, revealing wood grain and growth rings?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the video shot in extreme slow motion showing a thin wood shaving curling and swirling through the air with visible fibers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera completely still with a background blurred into a dark haze, while the blade reflects dazzling highlights?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03372899070382118 + ], + "opposite_logprob_tokens": [ + -3.408729076385498 + ], + "logprob_reward": -0.03372899070382118, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.03372899070382118, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10221647471189499, + 0.031937431544065475, + 0.061308469623327255, + 0.042345818132162094, + 0.03450722619891167, + 0.19683605432510376, + 0.08986057341098785, + 0.13059361279010773, + 0.2182793915271759, + 0.09211499243974686 + ], + "credit_map": [ + 0.014463810250163078, + 0.015256907790899277, + 0.01856633648276329, + 0.012006333097815514, + 0.0073831090703606606, + 0.009393349289894104, + 0.014313574880361557, + 0.010833041742444038, + 0.004560698289424181, + 0.004483801778405905, + 0.004082865547388792, + 0.004054499324411154, + 0.002158072078600526, + 0.0017494584899395704, + 0.006825458724051714, + 0.004022576380521059, + 0.005778454709798098, + 0.008191978558897972, + 0.006934907287359238, + 0.009774789214134216, + 0.009257926605641842, + 0.005315527785569429, + 0.009915481321513653, + 0.0061394027434289455, + 0.005171636119484901, + 0.0029168801847845316, + 0.005039711482822895, + 0.005797085352241993, + 0.007157128304243088, + 0.002669807756319642, + 0.007604576181620359, + 0.005988990422338247, + 0.004708200693130493, + 0.003914517816156149, + 0.005313906818628311, + 0.006016912870109081, + 0.0036402831319719553, + 0.0018989031668752432, + 0.004797205328941345, + 0.004217296373099089, + 0.03643718734383583, + 0.02653626911342144, + 0.029462017118930817, + 0.020370539277791977, + 0.030181804671883583, + 0.013546032831072807, + 0.021921953186392784, + 0.018380234017968178, + 0.011949693784117699, + 0.011957569047808647, + 0.014985778369009495, + 0.009656046517193317, + 0.01023541484028101, + 0.007751414552330971, + 0.012955916114151478, + 0.010368728078901768, + 0.011616479605436325, + 0.013383290730416775, + 0.027202054858207703, + 0.01885094866156578, + 0.01570664905011654, + 0.008192270062863827, + 0.022607387974858284, + 0.013034532777965069, + 0.028988873586058617, + 0.0223725363612175, + 0.05124189332127571, + 0.02245931141078472, + 0.025149250403046608, + 0.014332110993564129, + 0.032003987580537796, + 0.02173141948878765, + 0.006535104010254145, + 0.01578192412853241, + 0.008094949647784233, + 0.01696269027888775, + 0.009965012781322002, + 0.0046020084992051125, + 0.01704741083085537, + 0.013125883415341377 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 5, + 7 + ], + "selected_reward": -0.03372899070382118 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000027/group_1/index_tensor([8919, 8919])_group_1_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000027/group_1/index_tensor([8919, 8919])_group_1_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..8ccd049a62f460aa0edb48f03e8658b45d87f3ac --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000027/group_1/index_tensor([8919, 8919])_group_1_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "medium close-up shot, shot from a slightly upward angle, slowly pans forward, focusing on a middle-aged craftsman, rapt in a stone carving workshop imbued with a primitive atmosphere. Beneath his short, graying hair, beads of sweat beaded his wrinkled forehead. His eyes were fixed sharply on the towering, dark gray granite stele before him, his lips pursed in a look of intense concentration. He wore a dusty, dark blue cotton work suit, his sleeves rolled up to his elbows, revealing his muscular, veiny forearms. His left hand, rough and powerful, held a steel chisel firmly, its tip resting precisely on the pre-drawn strokes on the stele's surface. In his right hand, a wooden-handled hammer struck the end of the chisel with steady, rhythmic, and decisive strokes. Each strike is accompanied by a crisp clash, sending a burst of white stone chips flying and swirling through the air, illuminated by a beam of bright sunlight slanting in through the workshop's high windows, creating a dreamy, Tyndall-effect light column. The backdrop is the chaotic yet orderly workshop, with stones of various sizes and carving tools stacked against the wall. The overall scene is dominated by a calming color palette, dominated by the cool gray of the stone and the deep blue of the work clothes. The light imbues this focused and powerful creative scene with warmth and vitality, creating an almost sacred, tranquil and focused artistic atmosphere.", + "seed": 23383, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000027/group_1/index_8919_seed_23383_dp_2_rank_4_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 1, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.047524501010775565, + "vlm_logprob": -0.047524501010775565, + "vlm_yes_no_margin": 3.45, + "vlm_token_credit": -0.047524501010775565, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a medium close-up of a middle-aged craftsman with short, graying hair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the craftsman wearing a dusty, dark blue cotton work suit with sleeves rolled up to his elbows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the craftsman holding a steel chisel in his left hand and a wooden-handled hammer in his right hand?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera slowly panning forward while the craftsman strikes the chisel, sending white stone chips flying?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by a beam of bright sunlight creating a visible light column (Tyndall effect) through high windows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03839802369475365 + ], + "opposite_logprob_tokens": [ + -4.788397789001465 + ], + "logprob_reward": -0.03839802369475365, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.03839802369475365, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15801596641540527, + 0.18500567972660065, + 0.11808046698570251, + 0.058285366743803024, + 0.03860808163881302, + 0.07809527218341827, + 0.076789490878582, + 0.06023652106523514, + 0.11650251597166061, + 0.11038068681955338 + ], + "credit_map": [ + 0.01793082430958748, + 0.02727324888110161, + 0.017955971881747246, + 0.0195624940097332, + 0.01791028119623661, + 0.019698623567819595, + 0.019964952021837234, + 0.0177195705473423, + 0.021498143672943115, + 0.015109128318727016, + 0.06195152923464775, + 0.025586530566215515, + 0.012727257795631886, + 0.011014478281140327, + 0.010450062341988087, + 0.026668554171919823, + 0.01945742592215538, + 0.023740850389003754, + 0.007745594251900911, + 0.008513808250427246, + 0.021704940125346184, + 0.011470532044768333, + 0.016782300546765327, + 0.008665007539093494, + 0.008771547116339207, + 0.008808748796582222, + 0.005997630301862955, + 0.00770146120339632, + 0.004807800054550171, + 0.00706734461709857, + 0.00722171925008297, + 0.00790911354124546, + 0.005796925630420446, + 0.006781738251447678, + 0.005009623244404793, + 0.003972658421844244, + 0.005036910530179739, + 0.0042333705350756645, + 0.004353060852736235, + 0.0034237930085510015, + 0.009046072140336037, + 0.008078890852630138, + 0.009688110090792179, + 0.008882618509232998, + 0.006825312040746212, + 0.010498338378965855, + 0.012077252380549908, + 0.012998677790164948, + 0.009392819367349148, + 0.011881494894623756, + 0.008657507598400116, + 0.007910748943686485, + 0.01017831265926361, + 0.007547670975327492, + 0.011049497872591019, + 0.010171444155275822, + 0.008804364129900932, + 0.006067431997507811, + 0.009769548662006855, + 0.007674588821828365, + 0.00633912393823266, + 0.009816618636250496, + 0.006563855800777674, + 0.005200990941375494, + 0.010480000637471676, + 0.02686299756169319, + 0.010063454508781433, + 0.009180279448628426, + 0.03076384961605072, + 0.012895122170448303, + 0.009161027148365974, + 0.007095783017575741, + 0.011060488410294056, + 0.018154378980398178, + 0.01406447496265173, + 0.00812286976724863, + 0.010959245264530182, + 0.025804562494158745, + 0.010529148392379284, + 0.011685513891279697 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 2 + ], + "selected_reward": -0.03839802369475365 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a medium close-up of a middle-aged craftsman with short, graying hair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the craftsman wearing a dusty, dark blue cotton work suit with sleeves rolled up to his elbows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the craftsman holding a steel chisel in his left hand and a wooden-handled hammer in his right hand?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera slowly panning forward while the craftsman strikes the chisel, sending white stone chips flying?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by a beam of bright sunlight creating a visible light column (Tyndall effect) through high windows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.016095230355858803 + ], + "opposite_logprob_tokens": [ + -4.141095161437988 + ], + "logprob_reward": -0.016095230355858803, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.016095230355858803, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.16299869120121002, + 0.09021348506212234, + 0.1395035982131958, + 0.1976374238729477, + 0.06942988932132721, + 0.11748115718364716, + 0.055521849542856216, + 0.025853201746940613, + 0.063602976500988, + 0.07775768637657166 + ], + "credit_map": [ + 0.020105745643377304, + 0.021371493116021156, + 0.02021700330078602, + 0.017918718978762627, + 0.022721515968441963, + 0.0168325025588274, + 0.027511049062013626, + 0.01632065512239933, + 0.007407566532492638, + 0.007876217365264893, + 0.026759352535009384, + 0.012000322341918945, + 0.011478795669972897, + 0.007733958773314953, + 0.004181069787591696, + 0.012776201590895653, + 0.03426360338926315, + 0.010258419439196587, + 0.013205569237470627, + 0.01023668609559536, + 0.04485933110117912, + 0.012328529730439186, + 0.009057927876710892, + 0.005293539725244045, + 0.022985359653830528, + 0.022147739306092262, + 0.04146811366081238, + 0.05396490916609764, + 0.015072328969836235, + 0.017705073580145836, + 0.012255879119038582, + 0.012038018554449081, + 0.009517300873994827, + 0.020071590319275856, + 0.0067175161093473434, + 0.006672741379588842, + 0.008952047675848007, + 0.005805213935673237, + 0.0052279517985880375, + 0.006465529557317495, + 0.019793886691331863, + 0.019653405994176865, + 0.02702447585761547, + 0.01161358691751957, + 0.01398584246635437, + 0.010588991455733776, + 0.0066266851499676704, + 0.00819428265094757, + 0.007233465556055307, + 0.007228539790958166, + 0.009981558658182621, + 0.005257217679172754, + 0.007559352554380894, + 0.007577026728540659, + 0.005104095675051212, + 0.0055805956944823265, + 0.003045928431674838, + 0.0044057960622012615, + 0.003985763527452946, + 0.0029280746821314096, + 0.0024206952657550573, + 0.0048921480774879456, + 0.002111748792231083, + 0.0020630485378205776, + 0.006921261548995972, + 0.013844892382621765, + 0.007837088778614998, + 0.005778505466878414, + 0.010751728899776936, + 0.008473816327750683, + 0.00505557656288147, + 0.004940103739500046, + 0.005739586893469095, + 0.015506913885474205, + 0.010408150963485241, + 0.006339869927614927, + 0.008286463096737862, + 0.014915441162884235, + 0.007514325436204672, + 0.009046939201653004 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 0, + 2 + ], + "selected_reward": -0.016095230355858803 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a medium close-up of a middle-aged craftsman with short, graying hair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the craftsman wearing a dusty, dark blue cotton work suit with sleeves rolled up to his elbows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the craftsman holding a steel chisel in his left hand and a wooden-handled hammer in his right hand?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera slowly panning forward while the craftsman strikes the chisel, sending white stone chips flying?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by a beam of bright sunlight creating a visible light column (Tyndall effect) through high windows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04306694120168686 + ], + "opposite_logprob_tokens": [ + -3.16806697845459 + ], + "logprob_reward": -0.04306694120168686, + "yes_no_margin_reward": 3.125, + "token_credit_reward": -0.04306694120168686, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1464352309703827, + 0.1324772834777832, + 0.07296863198280334, + 0.1017838567495346, + 0.09929177910089493, + 0.1126064658164978, + 0.10237808525562286, + 0.07703574001789093, + 0.0631929412484169, + 0.09182994067668915 + ], + "credit_map": [ + 0.02071542479097843, + 0.027348555624485016, + 0.027041872963309288, + 0.019169719889760017, + 0.01351069938391447, + 0.011562883853912354, + 0.012921614572405815, + 0.014164465479552746, + 0.013525591231882572, + 0.01042156107723713, + 0.04029056802392006, + 0.02004103548824787, + 0.018577907234430313, + 0.011543822474777699, + 0.008721289224922657, + 0.009355509653687477, + 0.009804020635783672, + 0.0075228251516819, + 0.007543873507529497, + 0.006305905524641275, + 0.01418343000113964, + 0.012748428620398045, + 0.006316095590591431, + 0.008544051088392735, + 0.0157620869576931, + 0.016587775200605392, + 0.01640038564801216, + 0.01842721737921238, + 0.009602402336895466, + 0.007872640155255795, + 0.010274971835315228, + 0.006856372579932213, + 0.013606018386781216, + 0.013369865715503693, + 0.009774901904165745, + 0.01117789652198553, + 0.013852162286639214, + 0.01557078305631876, + 0.009761866182088852, + 0.012178285047411919, + 0.015302952378988266, + 0.014477252028882504, + 0.012788825668394566, + 0.012217252515256405, + 0.011665120720863342, + 0.01570233888924122, + 0.01599210873246193, + 0.014460615813732147, + 0.012712831608951092, + 0.009168650954961777, + 0.010501994751393795, + 0.006041770800948143, + 0.023791370913386345, + 0.008866517804563046, + 0.008127075619995594, + 0.02316787652671337, + 0.008576617576181889, + 0.013177945278584957, + 0.01020828541368246, + 0.0067318095825612545, + 0.0072220019064843655, + 0.013423029333353043, + 0.010477167554199696, + 0.0072188894264400005, + 0.006561409682035446, + 0.010682365857064724, + 0.008152597583830357, + 0.0069601270370185375, + 0.008227475918829441, + 0.009053738787770271, + 0.007741720415651798, + 0.0058135055005550385, + 0.008094578050076962, + 0.011640693992376328, + 0.012731178663671017, + 0.008157835341989994, + 0.008140865713357925, + 0.01841607503592968, + 0.011658587493002415, + 0.01299013290554285 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 5 + ], + "selected_reward": -0.04306694120168686 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a medium close-up of a middle-aged craftsman with short, graying hair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the craftsman wearing a dusty, dark blue cotton work suit with sleeves rolled up to his elbows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the craftsman holding a steel chisel in his left hand and a wooden-handled hammer in his right hand?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera slowly panning forward while the craftsman strikes the chisel, sending white stone chips flying?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by a beam of bright sunlight creating a visible light column (Tyndall effect) through high windows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07003426551818848 + ], + "opposite_logprob_tokens": [ + -2.6950342655181885 + ], + "logprob_reward": -0.07003426551818848, + "yes_no_margin_reward": 2.625, + "token_credit_reward": -0.07003426551818848, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.21256975829601288, + 0.12255001068115234, + 0.2039693146944046, + 0.10758305341005325, + 0.03722478449344635, + 0.05694019794464111, + 0.06202905625104904, + 0.054521337151527405, + 0.06096699461340904, + 0.08164546638727188 + ], + "credit_map": [ + 0.023375477641820908, + 0.037185944616794586, + 0.028045454993844032, + 0.022758040577173233, + 0.02485627681016922, + 0.02290366031229496, + 0.027941595762968063, + 0.025503309443593025, + 0.014637226238846779, + 0.010024397633969784, + 0.04221648722887039, + 0.017188439145684242, + 0.007464027032256126, + 0.007599906530231237, + 0.0050009009428322315, + 0.018418623134493828, + 0.034074846655130386, + 0.025332067161798477, + 0.018215226009488106, + 0.020744800567626953, + 0.051271162927150726, + 0.018372714519500732, + 0.023152660578489304, + 0.01280583068728447, + 0.012710039503872395, + 0.01688339188694954, + 0.018603889271616936, + 0.024604979902505875, + 0.00855363067239523, + 0.010328013449907303, + 0.007480763830244541, + 0.00841834768652916, + 0.005354529246687889, + 0.004355661105364561, + 0.0057535432279109955, + 0.003030309919267893, + 0.004775780253112316, + 0.005762551911175251, + 0.005226875655353069, + 0.0029655350372195244, + 0.006709625478833914, + 0.008418095298111439, + 0.008310838602483273, + 0.005843400489538908, + 0.004653028212487698, + 0.00647616945207119, + 0.0071021635085344315, + 0.009426874108612537, + 0.009862328879535198, + 0.007146945223212242, + 0.008443029597401619, + 0.0053968122228980064, + 0.011131231673061848, + 0.006943675689399242, + 0.004821618553251028, + 0.008283416740596294, + 0.008105560205876827, + 0.008216774091124535, + 0.010485244914889336, + 0.005267189349979162, + 0.006065306253731251, + 0.007050974294543266, + 0.005093418527394533, + 0.004236871376633644, + 0.0070165297947824, + 0.009499317966401577, + 0.008884035982191563, + 0.007576892618089914, + 0.00596057903021574, + 0.008250261656939983, + 0.008292163722217083, + 0.005487215239554644, + 0.009917800314724445, + 0.016533266752958298, + 0.010376893915235996, + 0.007477747276425362, + 0.008694915100932121, + 0.009970621205866337, + 0.009186326526105404, + 0.009487896226346493 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 1 + ], + "selected_reward": -0.07003426551818848 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a medium close-up of a middle-aged craftsman with short, graying hair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the craftsman wearing a dusty, dark blue cotton work suit with sleeves rolled up to his elbows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the craftsman holding a steel chisel in his left hand and a wooden-handled hammer in his right hand?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera slowly panning forward while the craftsman strikes the chisel, sending white stone chips flying?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by a beam of bright sunlight creating a visible light column (Tyndall effect) through high windows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07002804428339005 + ], + "opposite_logprob_tokens": [ + -2.695028066635132 + ], + "logprob_reward": -0.07002804428339005, + "yes_no_margin_reward": 2.625, + "token_credit_reward": -0.07002804428339005, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.4927542209625244, + 0.17588824033737183, + 0.11209104210138321, + 0.04414721950888634, + 0.026011843234300613, + 0.02409941516816616, + 0.043953344225883484, + 0.018620025366544724, + 0.023502444848418236, + 0.03893224522471428 + ], + "credit_map": [ + 0.046530961990356445, + 0.06323760747909546, + 0.0802803561091423, + 0.05342298746109009, + 0.07346728444099426, + 0.058294691145420074, + 0.06561686843633652, + 0.05190345644950867, + 0.019145777449011803, + 0.00872914306819439, + 0.06231100112199783, + 0.02500995621085167, + 0.0125100277364254, + 0.011217741295695305, + 0.009367110207676888, + 0.027597462758421898, + 0.02381778135895729, + 0.013606805354356766, + 0.009779613465070724, + 0.00814239401370287, + 0.03426145389676094, + 0.01063000038266182, + 0.00745190866291523, + 0.00440108310431242, + 0.004708601161837578, + 0.005253949202597141, + 0.009602207690477371, + 0.012068921700119972, + 0.003436771919950843, + 0.003456051927059889, + 0.0029167213942855597, + 0.002703986829146743, + 0.0036295917816460133, + 0.002525273710489273, + 0.0024475371465086937, + 0.0021624641958624125, + 0.004648727364838123, + 0.0038117272779345512, + 0.0023730495013296604, + 0.004413472954183817, + 0.004527581389993429, + 0.0028234024066478014, + 0.004086177796125412, + 0.0016698421677574515, + 0.0030131235253065825, + 0.0035823348443955183, + 0.002137198345735669, + 0.0022597536444664, + 0.004091397859156132, + 0.004718896001577377, + 0.004564644303172827, + 0.004388452507555485, + 0.00688876211643219, + 0.004046447109431028, + 0.003041426418349147, + 0.012213312089443207, + 0.002951285568997264, + 0.0028822089079767466, + 0.0030321392696350813, + 0.0014372896403074265, + 0.0017666735220700502, + 0.003036194248124957, + 0.002010981785133481, + 0.0015032518422231078, + 0.0019564058165997267, + 0.004079250153154135, + 0.002305289963260293, + 0.0015173675492405891, + 0.005474339704960585, + 0.003964858129620552, + 0.002176824724301696, + 0.0020281081087887287, + 0.003937585279345512, + 0.0030699793715029955, + 0.005246941931545734, + 0.002737038303166628, + 0.0036214955616742373, + 0.00761497113853693, + 0.004723973572254181, + 0.007980257272720337 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 2 + ], + "selected_reward": -0.07002804428339005 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "medium close-up shot, shot from a slightly upward angle, slowly pans forward, focusing on a middle-aged craftsman, rapt in a stone carving workshop imbued with a primitive atmosphere. Beneath his short, graying hair, beads of sweat beaded his wrinkled forehead. His eyes were fixed sharply on the towering, dark gray granite stele before him, his lips pursed in a look of intense concentration. He wore a dusty, dark blue cotton work suit, his sleeves rolled up to his elbows, revealing his muscular, veiny forearms. His left hand, rough and powerful, held a steel chisel firmly, its tip resting precisely on the pre-drawn strokes on the stele's surface. In his right hand, a wooden-handled hammer struck the end of the chisel with steady, rhythmic, and decisive strokes. Each strike is accompanied by a crisp clash, sending a burst of white stone chips flying and swirling through the air, illuminated by a beam of bright sunlight slanting in through the workshop's high windows, creating a dreamy, Tyndall-effect light column. The backdrop is the chaotic yet orderly workshop, with stones of various sizes and carving tools stacked against the wall. The overall scene is dominated by a calming color palette, dominated by the cool gray of the stone and the deep blue of the work clothes. The light imbues this focused and powerful creative scene with warmth and vitality, creating an almost sacred, tranquil and focused artistic atmosphere.", + "seed": 23384, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000027/group_1/index_8919_seed_23384_dp_2_rank_4_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 1, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.032999816536903384, + "vlm_logprob": -0.032999816536903384, + "vlm_yes_no_margin": 3.925, + "vlm_token_credit": -0.032999816536903384, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a medium close-up of a middle-aged craftsman with short, graying hair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the craftsman wearing a dusty, dark blue cotton work suit with sleeves rolled up to his elbows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the craftsman holding a steel chisel in his left hand and a wooden-handled hammer in his right hand?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera slowly panning forward while the craftsman strikes the chisel, sending white stone chips flying?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by a beam of bright sunlight creating a visible light column (Tyndall effect) through high windows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.021797023713588715 + ], + "opposite_logprob_tokens": [ + -5.646797180175781 + ], + "logprob_reward": -0.021797023713588715, + "yes_no_margin_reward": 5.625, + "token_credit_reward": -0.021797023713588715, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1693902462720871, + 0.05055110529065132, + 0.1648927479982376, + 0.13449884951114655, + 0.05470248684287071, + 0.04706487059593201, + 0.054700836539268494, + 0.07903217524290085, + 0.14987726509571075, + 0.09528938680887222 + ], + "credit_map": [ + 0.025900160893797874, + 0.018880443647503853, + 0.017236320301890373, + 0.02022445946931839, + 0.012775012291967869, + 0.02845185063779354, + 0.0225553996860981, + 0.023366598412394524, + 0.006690813694149256, + 0.008885725401341915, + 0.004969333298504353, + 0.006600049789994955, + 0.005090249236673117, + 0.00731321657076478, + 0.005093441344797611, + 0.005908277817070484, + 0.014703325927257538, + 0.013176877982914448, + 0.010571292601525784, + 0.04640187323093414, + 0.0095035620033741, + 0.04519375041127205, + 0.011299802921712399, + 0.014042260125279427, + 0.01170896366238594, + 0.009725631214678288, + 0.004159835167229176, + 0.010847152210772038, + 0.005043040029704571, + 0.036663927137851715, + 0.008205623365938663, + 0.04814467579126358, + 0.010110320523381233, + 0.008190031163394451, + 0.0053247734904289246, + 0.005737952888011932, + 0.007175002712756395, + 0.004076012875884771, + 0.005150599870830774, + 0.008937792852520943, + 0.008575725369155407, + 0.005987666081637144, + 0.007645426318049431, + 0.005048364866524935, + 0.00461809104308486, + 0.004862505942583084, + 0.005397670902311802, + 0.004929421469569206, + 0.009266400709748268, + 0.005672985687851906, + 0.006876027677208185, + 0.0059899939224123955, + 0.004898453131318092, + 0.006421780679374933, + 0.006878928747028112, + 0.008696266449987888, + 0.015505111776292324, + 0.01038119476288557, + 0.007462062872946262, + 0.008867206051945686, + 0.012902319431304932, + 0.005612441338598728, + 0.010077656246721745, + 0.008224177174270153, + 0.01969858631491661, + 0.013621533289551735, + 0.03936997428536415, + 0.020669350400567055, + 0.021336635574698448, + 0.017150649800896645, + 0.008507108315825462, + 0.009523431770503521, + 0.014866253361105919, + 0.014134674333035946, + 0.012647261843085289, + 0.01074256096035242, + 0.012199901044368744, + 0.008561681024730206, + 0.009046448394656181, + 0.013090616092085838 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 8 + ], + "selected_reward": -0.021797023713588715 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a medium close-up of a middle-aged craftsman with short, graying hair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the craftsman wearing a dusty, dark blue cotton work suit with sleeves rolled up to his elbows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the craftsman holding a steel chisel in his left hand and a wooden-handled hammer in his right hand?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera slowly panning forward while the craftsman strikes the chisel, sending white stone chips flying?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by a beam of bright sunlight creating a visible light column (Tyndall effect) through high windows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.009797343984246254 + ], + "opposite_logprob_tokens": [ + -4.6347975730896 + ], + "logprob_reward": -0.009797343984246254, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.009797343984246254, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14450186491012573, + 0.09396996349096298, + 0.2062748372554779, + 0.05870699882507324, + 0.04889073967933655, + 0.0773455798625946, + 0.051763251423835754, + 0.059154853224754333, + 0.1831483542919159, + 0.07624362409114838 + ], + "credit_map": [ + 0.015374939888715744, + 0.012368937022984028, + 0.02179885283112526, + 0.02077529765665531, + 0.012654215097427368, + 0.026567714288830757, + 0.01145602110773325, + 0.02350587025284767, + 0.012461739592254162, + 0.015466316603124142, + 0.01018694881349802, + 0.011390571482479572, + 0.004479945637285709, + 0.01698160171508789, + 0.007529931608587503, + 0.01547291036695242, + 0.01719667762517929, + 0.02514456957578659, + 0.01684258133172989, + 0.05729982629418373, + 0.00984025839716196, + 0.04174947738647461, + 0.018767939880490303, + 0.019433487206697464, + 0.013136639259755611, + 0.008454465307295322, + 0.007124364376068115, + 0.007288254797458649, + 0.00320422207005322, + 0.005898286122828722, + 0.00827969890087843, + 0.00532106589525938, + 0.007735579274594784, + 0.005704351235181093, + 0.004421698860824108, + 0.005247633438557386, + 0.005865138489753008, + 0.0034076799638569355, + 0.004765206016600132, + 0.011743446812033653, + 0.013108757324516773, + 0.009451130405068398, + 0.009332850575447083, + 0.008475431241095066, + 0.007383125834167004, + 0.006680317688733339, + 0.011269642040133476, + 0.011644315905869007, + 0.010325174778699875, + 0.007247254718095064, + 0.006374541204422712, + 0.005271147005259991, + 0.005941981915384531, + 0.007243742235004902, + 0.003628548001870513, + 0.00573086179792881, + 0.008808367885649204, + 0.007377687841653824, + 0.012972683645784855, + 0.008632230572402477, + 0.007390871644020081, + 0.003430252429097891, + 0.006047477014362812, + 0.0044952803291380405, + 0.028259772807359695, + 0.016150323674082756, + 0.044863056391477585, + 0.02496159076690674, + 0.022600287571549416, + 0.02143615484237671, + 0.010777102783322334, + 0.014100048691034317, + 0.012665598653256893, + 0.007844926789402962, + 0.009237246587872505, + 0.01148727536201477, + 0.013018086552619934, + 0.006003836635500193, + 0.007313224021345377, + 0.008673422038555145 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 8, + 0 + ], + "selected_reward": -0.009797343984246254 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a medium close-up of a middle-aged craftsman with short, graying hair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the craftsman wearing a dusty, dark blue cotton work suit with sleeves rolled up to his elbows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the craftsman holding a steel chisel in his left hand and a wooden-handled hammer in his right hand?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera slowly panning forward while the craftsman strikes the chisel, sending white stone chips flying?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by a beam of bright sunlight creating a visible light column (Tyndall effect) through high windows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04865000769495964 + ], + "opposite_logprob_tokens": [ + -3.048650026321411 + ], + "logprob_reward": -0.04865000769495964, + "yes_no_margin_reward": 3.0, + "token_credit_reward": -0.04865000769495964, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13583506643772125, + 0.08812431246042252, + 0.08403314650058746, + 0.09924673289060593, + 0.18738536536693573, + 0.129196435213089, + 0.1083158552646637, + 0.03570528328418732, + 0.09603272378444672, + 0.03612510487437248 + ], + "credit_map": [ + 0.017462344840168953, + 0.011580510064959526, + 0.013621573336422443, + 0.018354587256908417, + 0.009413047693669796, + 0.024902425706386566, + 0.01683550514280796, + 0.02366507053375244, + 0.017841607332229614, + 0.0153778912499547, + 0.007582190912216902, + 0.010738139972090721, + 0.00692448765039444, + 0.011559421196579933, + 0.006148823536932468, + 0.0119517482817173, + 0.008444252423942089, + 0.0074456012807786465, + 0.006755308248102665, + 0.016906946897506714, + 0.006288906093686819, + 0.023122260347008705, + 0.006079230457544327, + 0.008990638889372349, + 0.008565444499254227, + 0.006659437902271748, + 0.0068869320675730705, + 0.007846188731491566, + 0.0044823880307376385, + 0.026738544926047325, + 0.006726538296788931, + 0.03134125471115112, + 0.025621499866247177, + 0.018721409142017365, + 0.012279395014047623, + 0.01771252416074276, + 0.01037195511162281, + 0.010770479217171669, + 0.027099451050162315, + 0.06480862945318222, + 0.022007916122674942, + 0.018015453591942787, + 0.017051801085472107, + 0.015880919992923737, + 0.013424321077764034, + 0.009609226137399673, + 0.014046112075448036, + 0.0191606767475605, + 0.022918405011296272, + 0.014527494087815285, + 0.00997738353908062, + 0.012772511690855026, + 0.009758503176271915, + 0.009996455162763596, + 0.009153155609965324, + 0.01921194978058338, + 0.0053185876458883286, + 0.0034418287687003613, + 0.0049824705347418785, + 0.005328974686563015, + 0.0036789539735764265, + 0.002922729356214404, + 0.006676490884274244, + 0.003355245338752866, + 0.012007191777229309, + 0.007760717533528805, + 0.029287464916706085, + 0.013455754145979881, + 0.010392555966973305, + 0.011111577041447163, + 0.0050675892271101475, + 0.006949868984520435, + 0.006650526542216539, + 0.004838647786527872, + 0.004042597487568855, + 0.0037059024907648563, + 0.007756227161735296, + 0.0025235111825168133, + 0.0037817468401044607, + 0.0028259451501071453 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 0, + 5 + ], + "selected_reward": -0.04865000769495964 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a medium close-up of a middle-aged craftsman with short, graying hair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the craftsman wearing a dusty, dark blue cotton work suit with sleeves rolled up to his elbows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the craftsman holding a steel chisel in his left hand and a wooden-handled hammer in his right hand?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera slowly panning forward while the craftsman strikes the chisel, sending white stone chips flying?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by a beam of bright sunlight creating a visible light column (Tyndall effect) through high windows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.029805148020386696 + ], + "opposite_logprob_tokens": [ + -3.5298051834106445 + ], + "logprob_reward": -0.029805148020386696, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.029805148020386696, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12026689946651459, + 0.06879102438688278, + 0.08582840859889984, + 0.12098605930805206, + 0.06510928273200989, + 0.10033060610294342, + 0.1652587205171585, + 0.05065050348639488, + 0.15762262046337128, + 0.06515588611364365 + ], + "credit_map": [ + 0.015897687524557114, + 0.013343272730708122, + 0.026969602331519127, + 0.01465581078082323, + 0.009010104462504387, + 0.01612984947860241, + 0.008848069235682487, + 0.015412503853440285, + 0.005942357704043388, + 0.009786956012248993, + 0.007928375154733658, + 0.010235060937702656, + 0.006539673078805208, + 0.009996800683438778, + 0.0072167664766311646, + 0.011145034804940224, + 0.012665866874158382, + 0.00944632850587368, + 0.007345312274992466, + 0.010768307372927666, + 0.006829721853137016, + 0.01741650328040123, + 0.0060538966208696365, + 0.015302476473152637, + 0.01754455827176571, + 0.015249142423272133, + 0.011313069611787796, + 0.01602174900472164, + 0.006568786688148975, + 0.01704668067395687, + 0.009058084338903427, + 0.028183987364172935, + 0.009200716391205788, + 0.009985318407416344, + 0.006290149874985218, + 0.006867838092148304, + 0.0062746936455369, + 0.005028275772929192, + 0.005287852603942156, + 0.0161744337528944, + 0.016070589423179626, + 0.012531552463769913, + 0.013610509224236012, + 0.012535841204226017, + 0.010585760697722435, + 0.006804278120398521, + 0.011528773233294487, + 0.01666330359876156, + 0.028144197538495064, + 0.019403744488954544, + 0.03460054099559784, + 0.022715529426932335, + 0.01182476244866848, + 0.020897498354315758, + 0.009537668898701668, + 0.01813478395342827, + 0.009179696440696716, + 0.006164231337606907, + 0.007937293499708176, + 0.006665795110166073, + 0.007765216287225485, + 0.003164157271385193, + 0.005399975925683975, + 0.004374135285615921, + 0.01951596885919571, + 0.018389295786619186, + 0.019463220611214638, + 0.018466535955667496, + 0.025561654940247536, + 0.01211838610470295, + 0.022508317604660988, + 0.021599236875772476, + 0.008421196602284908, + 0.008761871606111526, + 0.008745639584958553, + 0.017504986375570297, + 0.006857531610876322, + 0.0031053770799189806, + 0.007779239676892757, + 0.0039800372906029224 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 8, + 3 + ], + "selected_reward": -0.029805148020386696 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a medium close-up of a middle-aged craftsman with short, graying hair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the craftsman wearing a dusty, dark blue cotton work suit with sleeves rolled up to his elbows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the craftsman holding a steel chisel in his left hand and a wooden-handled hammer in his right hand?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera slowly panning forward while the craftsman strikes the chisel, sending white stone chips flying?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by a beam of bright sunlight creating a visible light column (Tyndall effect) through high windows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0549495592713356 + ], + "opposite_logprob_tokens": [ + -2.9299495220184326 + ], + "logprob_reward": -0.0549495592713356, + "yes_no_margin_reward": 2.875, + "token_credit_reward": -0.0549495592713356, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2520645558834076, + 0.0766415223479271, + 0.09452439099550247, + 0.09029998630285263, + 0.04342637583613396, + 0.06855758279561996, + 0.136335551738739, + 0.09283487498760223, + 0.08275551348924637, + 0.06255961954593658 + ], + "credit_map": [ + 0.03210945054888725, + 0.0298356581479311, + 0.03063013032078743, + 0.035191237926483154, + 0.024228038266301155, + 0.049567341804504395, + 0.01971960999071598, + 0.03078310377895832, + 0.006485210731625557, + 0.00992339663207531, + 0.006937357131391764, + 0.009916423819959164, + 0.004151416942477226, + 0.016829170286655426, + 0.00913814827799797, + 0.013260407373309135, + 0.00875941477715969, + 0.010416547767817974, + 0.0059856511652469635, + 0.0263337641954422, + 0.005979025736451149, + 0.023524770513176918, + 0.006991715170443058, + 0.006533496547490358, + 0.012943973764777184, + 0.006959485821425915, + 0.004332303535193205, + 0.006781755480915308, + 0.006145435851067305, + 0.02272767201066017, + 0.007049188949167728, + 0.023360177874565125, + 0.005375987850129604, + 0.00464292336255312, + 0.003382051130756736, + 0.0035983193665742874, + 0.00624522939324379, + 0.0035908513236790895, + 0.006715705152601004, + 0.009875311516225338, + 0.012881266884505749, + 0.009622311219573021, + 0.009095350280404091, + 0.008441775105893612, + 0.0075009907595813274, + 0.005021252669394016, + 0.008759030140936375, + 0.007235611788928509, + 0.019592858850955963, + 0.0126542579382658, + 0.030876703560352325, + 0.01685069128870964, + 0.013757904060184956, + 0.0183070320636034, + 0.011212055571377277, + 0.013084049336612225, + 0.012251649983227253, + 0.013404217548668385, + 0.007130837067961693, + 0.009687685407698154, + 0.022344892844557762, + 0.007033292204141617, + 0.013601645827293396, + 0.007380657363682985, + 0.011786398477852345, + 0.00910724326968193, + 0.015817394480109215, + 0.00998540036380291, + 0.011119818314909935, + 0.00897966418415308, + 0.008401663042604923, + 0.007557927630841732, + 0.007209829054772854, + 0.008142692968249321, + 0.007380410563200712, + 0.011353121139109135, + 0.008233110420405865, + 0.005661559756845236, + 0.00665110582485795, + 0.00792778842151165 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 6, + 2 + ], + "selected_reward": -0.0549495592713356 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "medium close-up shot, shot from a slightly upward angle, slowly pans forward, focusing on a middle-aged craftsman, rapt in a stone carving workshop imbued with a primitive atmosphere. Beneath his short, graying hair, beads of sweat beaded his wrinkled forehead. His eyes were fixed sharply on the towering, dark gray granite stele before him, his lips pursed in a look of intense concentration. He wore a dusty, dark blue cotton work suit, his sleeves rolled up to his elbows, revealing his muscular, veiny forearms. His left hand, rough and powerful, held a steel chisel firmly, its tip resting precisely on the pre-drawn strokes on the stele's surface. In his right hand, a wooden-handled hammer struck the end of the chisel with steady, rhythmic, and decisive strokes. Each strike is accompanied by a crisp clash, sending a burst of white stone chips flying and swirling through the air, illuminated by a beam of bright sunlight slanting in through the workshop's high windows, creating a dreamy, Tyndall-effect light column. The backdrop is the chaotic yet orderly workshop, with stones of various sizes and carving tools stacked against the wall. The overall scene is dominated by a calming color palette, dominated by the cool gray of the stone and the deep blue of the work clothes. The light imbues this focused and powerful creative scene with warmth and vitality, creating an almost sacred, tranquil and focused artistic atmosphere.", + "seed": 23385, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000027/group_1/index_8919_seed_23385_dp_3_rank_6_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 1, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.09152305945754051, + "vlm_logprob": -0.09152305945754051, + "vlm_yes_no_margin": 2.975, + "vlm_token_credit": -0.09152305945754051, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a medium close-up of a middle-aged craftsman with short, graying hair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the craftsman wearing a dusty, dark blue cotton work suit with sleeves rolled up to his elbows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the craftsman holding a steel chisel in his left hand and a wooden-handled hammer in his right hand?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera slowly panning forward while the craftsman strikes the chisel, sending white stone chips flying?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by a beam of bright sunlight creating a visible light column (Tyndall effect) through high windows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02921486087143421 + ], + "opposite_logprob_tokens": [ + -4.779214859008789 + ], + "logprob_reward": -0.02921486087143421, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.02921486087143421, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.21870967745780945, + 0.09046828001737595, + 0.045849405229091644, + 0.025872668251395226, + 0.08468987792730331, + 0.1128758117556572, + 0.03546716645359993, + 0.2721591591835022, + 0.06830819696187973, + 0.04559971019625664 + ], + "credit_map": [ + 0.029246078804135323, + 0.04834863543510437, + 0.023731660097837448, + 0.010261736810207367, + 0.018314804881811142, + 0.03962886333465576, + 0.01315922662615776, + 0.03601868823170662, + 0.017740407958626747, + 0.02346106804907322, + 0.008915578946471214, + 0.005236071068793535, + 0.00859562773257494, + 0.009687499143183231, + 0.00480513833463192, + 0.012026895768940449, + 0.0067419204860925674, + 0.006380634382367134, + 0.0036700409837067127, + 0.004893957171589136, + 0.007834570482373238, + 0.0047458321787416935, + 0.006813996471464634, + 0.004768453538417816, + 0.004108679015189409, + 0.0032610774505883455, + 0.0026770050171762705, + 0.0026158830150961876, + 0.003985654562711716, + 0.0030188453383743763, + 0.0034029483795166016, + 0.0028025798965245485, + 0.025492331013083458, + 0.025857772678136826, + 0.00490025756880641, + 0.006363067775964737, + 0.00784742645919323, + 0.005112917628139257, + 0.0038324729539453983, + 0.00528364535421133, + 0.0115204444155097, + 0.012297559529542923, + 0.009327206760644913, + 0.020339829847216606, + 0.026538580656051636, + 0.008964172564446926, + 0.01187964715063572, + 0.012008393183350563, + 0.0062900385819375515, + 0.005752816796302795, + 0.003000412369146943, + 0.00478888675570488, + 0.006278422195464373, + 0.002883444307371974, + 0.002630476141348481, + 0.0038426739629358053, + 0.1081051379442215, + 0.09940096735954285, + 0.005953426938503981, + 0.008637921884655952, + 0.011873704381287098, + 0.010769103653728962, + 0.00802075956016779, + 0.01939813420176506, + 0.012645636685192585, + 0.010219039395451546, + 0.008395795710384846, + 0.0066328393295407295, + 0.006839422974735498, + 0.006319788750261068, + 0.005850495770573616, + 0.011405181139707565, + 0.005853669717907906, + 0.006935215089470148, + 0.005532838404178619, + 0.006630579009652138, + 0.005486528854817152, + 0.0053602843545377254, + 0.004720692988485098, + 0.005079904105514288 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 0, + 5 + ], + "selected_reward": -0.02921486087143421 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a medium close-up of a middle-aged craftsman with short, graying hair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the craftsman wearing a dusty, dark blue cotton work suit with sleeves rolled up to his elbows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the craftsman holding a steel chisel in his left hand and a wooden-handled hammer in his right hand?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera slowly panning forward while the craftsman strikes the chisel, sending white stone chips flying?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by a beam of bright sunlight creating a visible light column (Tyndall effect) through high windows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02331754006445408 + ], + "opposite_logprob_tokens": [ + -3.773317575454712 + ], + "logprob_reward": -0.02331754006445408, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.02331754006445408, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.41260334849357605, + 0.06789443641901016, + 0.027958989143371582, + 0.015681833028793335, + 0.03181460127234459, + 0.0380634069442749, + 0.017348555848002434, + 0.3059503436088562, + 0.05587988719344139, + 0.026804612949490547 + ], + "credit_map": [ + 0.14970482885837555, + 0.14285244047641754, + 0.020738815888762474, + 0.01456044614315033, + 0.03227399289608002, + 0.016615411266684532, + 0.011196858249604702, + 0.024660544469952583, + 0.006844025570899248, + 0.009394006803631783, + 0.008042819797992706, + 0.00392812630161643, + 0.007247796282172203, + 0.014783373102545738, + 0.0045805578120052814, + 0.013073726557195187, + 0.006030911114066839, + 0.005752819124609232, + 0.003173032309859991, + 0.0025219153612852097, + 0.0038230284117162228, + 0.0015843078726902604, + 0.002633903408423066, + 0.002439071424305439, + 0.002666155807673931, + 0.0019261575071141124, + 0.0016087586991488934, + 0.0015971759567037225, + 0.0021599056199193, + 0.002697133459150791, + 0.0012939405860379338, + 0.0017326041124761105, + 0.004524766001850367, + 0.00701456144452095, + 0.0019518572371453047, + 0.002521681599318981, + 0.005013708025217056, + 0.0028792130760848522, + 0.0025271927006542683, + 0.005381619092077017, + 0.008081518113613129, + 0.007451559416949749, + 0.004451982211321592, + 0.004762190394103527, + 0.00513495784252882, + 0.0020183399319648743, + 0.002725378843024373, + 0.003437478095293045, + 0.0021832003258168697, + 0.0026642330922186375, + 0.0013854345306754112, + 0.002019095467403531, + 0.003488450311124325, + 0.001554224407300353, + 0.002375772688537836, + 0.0016781453741714358, + 0.1245153471827507, + 0.11681763827800751, + 0.00576621200889349, + 0.007657660637050867, + 0.013837557286024094, + 0.011688263155519962, + 0.007438549771904945, + 0.018229098990559578, + 0.014070662669837475, + 0.009819243103265762, + 0.001895897788926959, + 0.00362352910451591, + 0.0049423654563724995, + 0.00545076746493578, + 0.004669071175158024, + 0.011408350430428982, + 0.003619798691943288, + 0.004041898995637894, + 0.002142867539077997, + 0.004418247379362583, + 0.004641438368707895, + 0.0020280254539102316, + 0.0027695270255208015, + 0.0031428078655153513 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 7, + 1 + ], + "selected_reward": -0.02331754006445408 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a medium close-up of a middle-aged craftsman with short, graying hair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the craftsman wearing a dusty, dark blue cotton work suit with sleeves rolled up to his elbows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the craftsman holding a steel chisel in his left hand and a wooden-handled hammer in his right hand?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera slowly panning forward while the craftsman strikes the chisel, sending white stone chips flying?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by a beam of bright sunlight creating a visible light column (Tyndall effect) through high windows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04306842386722565 + ], + "opposite_logprob_tokens": [ + -3.1680684089660645 + ], + "logprob_reward": -0.04306842386722565, + "yes_no_margin_reward": 3.125, + "token_credit_reward": -0.04306842386722565, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.431710809469223, + 0.12188751250505447, + 0.04712015762925148, + 0.03245859593153, + 0.079414501786232, + 0.06637634336948395, + 0.03003600239753723, + 0.11881155520677567, + 0.03216340392827988, + 0.040021125227212906 + ], + "credit_map": [ + 0.1308917999267578, + 0.13888874650001526, + 0.03079589456319809, + 0.021547526121139526, + 0.03969413414597511, + 0.021403564140200615, + 0.017579544335603714, + 0.03090960904955864, + 0.016298970207571983, + 0.013838176615536213, + 0.013446626253426075, + 0.011861010454595089, + 0.013321427628397942, + 0.026361312717199326, + 0.008137842640280724, + 0.0186221431940794, + 0.005419498775154352, + 0.004927415866404772, + 0.005816180258989334, + 0.005150424782186747, + 0.006585309281945229, + 0.008418506942689419, + 0.005939270369708538, + 0.004863550420850515, + 0.004244928248226643, + 0.003793899202719331, + 0.003092633094638586, + 0.0035277048591524363, + 0.005561087746173143, + 0.004786466248333454, + 0.00420777453109622, + 0.0032440992072224617, + 0.0073270536959171295, + 0.011747528798878193, + 0.007388146594166756, + 0.006317832041531801, + 0.012338612228631973, + 0.010468767024576664, + 0.010852154344320297, + 0.012974405661225319, + 0.011038411408662796, + 0.009706796146929264, + 0.007698134519159794, + 0.01013485249131918, + 0.010306081734597683, + 0.004717782139778137, + 0.0062365601770579815, + 0.006537726614624262, + 0.00351439556106925, + 0.0034217846114188433, + 0.0026342945639044046, + 0.003701150882989168, + 0.004356733523309231, + 0.0025318078696727753, + 0.006035735365003347, + 0.003840098390355706, + 0.03982674703001976, + 0.043700650334358215, + 0.00547556858509779, + 0.006651782896369696, + 0.008410220965743065, + 0.00324513902887702, + 0.0027674976736307144, + 0.008733944036066532, + 0.007534201256930828, + 0.005408405791968107, + 0.0024897668045014143, + 0.0029389280825853348, + 0.00316764903254807, + 0.0027651952113956213, + 0.002542724134400487, + 0.00531653081998229, + 0.006176453083753586, + 0.005926460027694702, + 0.003976545296609402, + 0.008171950466930866, + 0.0048272693529725075, + 0.002717304974794388, + 0.0038576978258788586, + 0.004367444198578596 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 7 + ], + "selected_reward": -0.04306842386722565 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a medium close-up of a middle-aged craftsman with short, graying hair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the craftsman wearing a dusty, dark blue cotton work suit with sleeves rolled up to his elbows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the craftsman holding a steel chisel in his left hand and a wooden-handled hammer in his right hand?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera slowly panning forward while the craftsman strikes the chisel, sending white stone chips flying?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by a beam of bright sunlight creating a visible light column (Tyndall effect) through high windows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.1604052633047104 + ], + "opposite_logprob_tokens": [ + -1.9104052782058716 + ], + "logprob_reward": -0.1604052633047104, + "yes_no_margin_reward": 1.75, + "token_credit_reward": -0.1604052633047104, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.4213733971118927, + 0.10607687383890152, + 0.053627822548151016, + 0.031439557671546936, + 0.09185495972633362, + 0.07331208139657974, + 0.023402014747262, + 0.14698171615600586, + 0.023372186347842216, + 0.028559310361742973 + ], + "credit_map": [ + 0.13280151784420013, + 0.14588330686092377, + 0.0247198436409235, + 0.014701695181429386, + 0.03154139965772629, + 0.022616568952798843, + 0.010981485247612, + 0.038127604871988297, + 0.020153600722551346, + 0.02693251520395279, + 0.011343003250658512, + 0.008627341128885746, + 0.009866615757346153, + 0.010413119569420815, + 0.005744027439504862, + 0.012996649369597435, + 0.014007974416017532, + 0.013305340893566608, + 0.004772871732711792, + 0.005439623724669218, + 0.0040984912775456905, + 0.0033501004800200462, + 0.004700548015534878, + 0.003952870145440102, + 0.004329700954258442, + 0.002950335154309869, + 0.004625665955245495, + 0.0038564023561775684, + 0.004308763891458511, + 0.003903557313606143, + 0.003957704640924931, + 0.00350742693990469, + 0.027390697970986366, + 0.02881491370499134, + 0.004290548153221607, + 0.006543426308780909, + 0.009234949015080929, + 0.005386266857385635, + 0.0038606140296906233, + 0.006333553232252598, + 0.008997743017971516, + 0.010875822976231575, + 0.009974246844649315, + 0.00893358699977398, + 0.013892114162445068, + 0.006064268760383129, + 0.007052349857985973, + 0.0075219557620584965, + 0.003039122559130192, + 0.0035788712557405233, + 0.002487603574991226, + 0.0027132246177643538, + 0.0039641582407057285, + 0.0017653562827035785, + 0.003433770965784788, + 0.0024199082981795073, + 0.05572017282247543, + 0.052960991859436035, + 0.0049086869694292545, + 0.005940951872617006, + 0.010786139406263828, + 0.004240616224706173, + 0.003834774484857917, + 0.008589399047195911, + 0.004438108764588833, + 0.004378274083137512, + 0.001829040120355785, + 0.002322189277037978, + 0.0029853833839297295, + 0.0014709826791658998, + 0.002174077555537224, + 0.003774130018427968, + 0.004696126561611891, + 0.005827742163091898, + 0.00289952103048563, + 0.004008288029581308, + 0.003555811708793044, + 0.0016213348135352135, + 0.0026384498924016953, + 0.003312035696581006 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 7, + 1 + ], + "selected_reward": -0.1604052633047104 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a medium close-up of a middle-aged craftsman with short, graying hair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the craftsman wearing a dusty, dark blue cotton work suit with sleeves rolled up to his elbows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the craftsman holding a steel chisel in his left hand and a wooden-handled hammer in his right hand?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera slowly panning forward while the craftsman strikes the chisel, sending white stone chips flying?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by a beam of bright sunlight creating a visible light column (Tyndall effect) through high windows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.20160920917987823 + ], + "opposite_logprob_tokens": [ + -1.7016092538833618 + ], + "logprob_reward": -0.20160920917987823, + "yes_no_margin_reward": 1.5, + "token_credit_reward": -0.20160920917987823, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2809250056743622, + 0.12107245624065399, + 0.0997738465666771, + 0.10535277426242828, + 0.10398143529891968, + 0.11672371625900269, + 0.02311844564974308, + 0.09261246770620346, + 0.030907059088349342, + 0.025532791391015053 + ], + "credit_map": [ + 0.06196258217096329, + 0.07037965953350067, + 0.039181020110845566, + 0.013364204205572605, + 0.024632753804326057, + 0.03181791305541992, + 0.010968412272632122, + 0.028618445619940758, + 0.0198748167604208, + 0.01748906448483467, + 0.01555539108812809, + 0.01088727731257677, + 0.009358658455312252, + 0.023099064826965332, + 0.006567748263478279, + 0.018240444362163544, + 0.032840143889188766, + 0.03141002729535103, + 0.008386867120862007, + 0.005010440479964018, + 0.007607039529830217, + 0.003310268046334386, + 0.005767263006418943, + 0.005441796965897083, + 0.01814947836101055, + 0.015142197720706463, + 0.014743635430932045, + 0.01213422417640686, + 0.014524011872708797, + 0.007210264913737774, + 0.011617784388363361, + 0.011831177398562431, + 0.023018214851617813, + 0.028813840821385384, + 0.007462610024958849, + 0.009419397450983524, + 0.01172370370477438, + 0.005007329396903515, + 0.006077820435166359, + 0.012458517216145992, + 0.026432543992996216, + 0.02225368842482567, + 0.012226492166519165, + 0.015036956407129765, + 0.012517263181507587, + 0.008562414906919003, + 0.007978662848472595, + 0.011715698055922985, + 0.00298620224930346, + 0.0034708024468272924, + 0.0011972516076639295, + 0.0027488344348967075, + 0.00378297152929008, + 0.0024620345793664455, + 0.0038555872160941362, + 0.0026147612370550632, + 0.03403798118233681, + 0.03156178072094917, + 0.0034681048709899187, + 0.00395815446972847, + 0.005383359268307686, + 0.004028416704386473, + 0.00337160169146955, + 0.0068030646070837975, + 0.004070719238370657, + 0.004784953780472279, + 0.0048818946816027164, + 0.004947689361870289, + 0.003561730030924082, + 0.0027024527080357075, + 0.00230292696505785, + 0.0036546909250319004, + 0.004634521435946226, + 0.004497155547142029, + 0.0029775486327707767, + 0.00361922150477767, + 0.0030378142837435007, + 0.0020215376280248165, + 0.001760415849275887, + 0.002984575927257538 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 5 + ], + "selected_reward": -0.20160920917987823 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "medium close-up shot, shot from a slightly upward angle, slowly pans forward, focusing on a middle-aged craftsman, rapt in a stone carving workshop imbued with a primitive atmosphere. Beneath his short, graying hair, beads of sweat beaded his wrinkled forehead. His eyes were fixed sharply on the towering, dark gray granite stele before him, his lips pursed in a look of intense concentration. He wore a dusty, dark blue cotton work suit, his sleeves rolled up to his elbows, revealing his muscular, veiny forearms. His left hand, rough and powerful, held a steel chisel firmly, its tip resting precisely on the pre-drawn strokes on the stele's surface. In his right hand, a wooden-handled hammer struck the end of the chisel with steady, rhythmic, and decisive strokes. Each strike is accompanied by a crisp clash, sending a burst of white stone chips flying and swirling through the air, illuminated by a beam of bright sunlight slanting in through the workshop's high windows, creating a dreamy, Tyndall-effect light column. The backdrop is the chaotic yet orderly workshop, with stones of various sizes and carving tools stacked against the wall. The overall scene is dominated by a calming color palette, dominated by the cool gray of the stone and the deep blue of the work clothes. The light imbues this focused and powerful creative scene with warmth and vitality, creating an almost sacred, tranquil and focused artistic atmosphere.", + "seed": 23386, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000027/group_1/index_8919_seed_23386_dp_3_rank_6_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 1, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.03706452250480652, + "vlm_logprob": -0.03706452250480652, + "vlm_yes_no_margin": 3.875, + "vlm_token_credit": -0.03706452250480652, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a medium close-up of a middle-aged craftsman with short, graying hair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the craftsman wearing a dusty, dark blue cotton work suit with sleeves rolled up to his elbows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the craftsman holding a steel chisel in his left hand and a wooden-handled hammer in his right hand?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera slowly panning forward while the craftsman strikes the chisel, sending white stone chips flying?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by a beam of bright sunlight creating a visible light column (Tyndall effect) through high windows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03571540489792824 + ], + "opposite_logprob_tokens": [ + -5.160715579986572 + ], + "logprob_reward": -0.03571540489792824, + "yes_no_margin_reward": 5.125, + "token_credit_reward": -0.03571540489792824, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.22738325595855713, + 0.10810848325490952, + 0.05503135174512863, + 0.049284033477306366, + 0.22950592637062073, + 0.06967239081859589, + 0.05035550147294998, + 0.06466005742549896, + 0.07857749611139297, + 0.06742148101329803 + ], + "credit_map": [ + 0.031922079622745514, + 0.02138671651482582, + 0.054929375648498535, + 0.014406965114176273, + 0.037724852561950684, + 0.022242717444896698, + 0.016472164541482925, + 0.028298400342464447, + 0.013151410967111588, + 0.010071469470858574, + 0.015259716659784317, + 0.009279521182179451, + 0.019460832700133324, + 0.01063341274857521, + 0.01623755320906639, + 0.01401456631720066, + 0.006991308648139238, + 0.008389131166040897, + 0.008876114152371883, + 0.004525807686150074, + 0.004248537588864565, + 0.0053269509226083755, + 0.00503979017958045, + 0.011633709073066711, + 0.006286533549427986, + 0.00667207594960928, + 0.006684530526399612, + 0.00428454764187336, + 0.0048329783603549, + 0.008249839767813683, + 0.005328083410859108, + 0.006945444270968437, + 0.020823314785957336, + 0.017767133191227913, + 0.022349432110786438, + 0.01439482532441616, + 0.03951876237988472, + 0.06054572016000748, + 0.024008609354496002, + 0.03009813092648983, + 0.007325011305510998, + 0.0063685462810099125, + 0.007085506338626146, + 0.005079540889710188, + 0.008071605116128922, + 0.0176127627491951, + 0.007292028516530991, + 0.010837391950190067, + 0.005617138929665089, + 0.011401227675378323, + 0.005573964677751064, + 0.004313907586038113, + 0.005246164742857218, + 0.007692399434745312, + 0.0057213488034904, + 0.0047893519513309, + 0.006347937975078821, + 0.007321349810808897, + 0.006521871313452721, + 0.004490335006266832, + 0.007722306996583939, + 0.01203086320310831, + 0.009213614277541637, + 0.011011773720383644, + 0.007901636883616447, + 0.011300637386739254, + 0.007296970579773188, + 0.005931007210165262, + 0.011924991384148598, + 0.011593286879360676, + 0.01497043389827013, + 0.007658532354980707, + 0.006534355226904154, + 0.010890433564782143, + 0.007176988758146763, + 0.004881090018898249, + 0.007780629210174084, + 0.008953072130680084, + 0.007969308644533157, + 0.013235606253147125 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 0, + 1 + ], + "selected_reward": -0.03571540489792824 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a medium close-up of a middle-aged craftsman with short, graying hair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the craftsman wearing a dusty, dark blue cotton work suit with sleeves rolled up to his elbows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the craftsman holding a steel chisel in his left hand and a wooden-handled hammer in his right hand?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera slowly panning forward while the craftsman strikes the chisel, sending white stone chips flying?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by a beam of bright sunlight creating a visible light column (Tyndall effect) through high windows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.012560781091451645 + ], + "opposite_logprob_tokens": [ + -4.387560844421387 + ], + "logprob_reward": -0.012560781091451645, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.012560781091451645, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15748175978660583, + 0.08195038884878159, + 0.11358947306871414, + 0.13811704516410828, + 0.12806087732315063, + 0.09364766627550125, + 0.05319927632808685, + 0.08419322967529297, + 0.06881943345069885, + 0.08094095438718796 + ], + "credit_map": [ + 0.026406852528452873, + 0.013547342270612717, + 0.032947178930044174, + 0.01546231284737587, + 0.019633052870631218, + 0.015236659906804562, + 0.01167230773717165, + 0.02257603406906128, + 0.009935315698385239, + 0.012597532942891121, + 0.011529605835676193, + 0.006484400946646929, + 0.009883127175271511, + 0.007907997816801071, + 0.011629540473222733, + 0.011982862837612629, + 0.016280999407172203, + 0.0234255101531744, + 0.017083529382944107, + 0.011380423791706562, + 0.012568741105496883, + 0.011129218153655529, + 0.010965141467750072, + 0.01075589656829834, + 0.015040705911815166, + 0.021404121071100235, + 0.019554877653717995, + 0.017199749127030373, + 0.012878605164587498, + 0.013242121785879135, + 0.016618521884083748, + 0.022178323939442635, + 0.016856925562024117, + 0.014010977931320667, + 0.01620069146156311, + 0.008658231236040592, + 0.016829675063490868, + 0.02606394700706005, + 0.013315624557435513, + 0.016124799847602844, + 0.011973107233643532, + 0.01643538288772106, + 0.009714341722428799, + 0.007322950754314661, + 0.009449522010982037, + 0.018566450104117393, + 0.008348251692950726, + 0.011837652884423733, + 0.006042425986379385, + 0.007343008648604155, + 0.006510909181088209, + 0.004846363794058561, + 0.00539167458191514, + 0.007745745126157999, + 0.009117107838392258, + 0.006202032323926687, + 0.009415007196366787, + 0.01994961127638817, + 0.008296126499772072, + 0.008874536491930485, + 0.008302020840346813, + 0.010196438990533352, + 0.010310683399438858, + 0.008848796598613262, + 0.005895302165299654, + 0.010022832080721855, + 0.005814520176500082, + 0.003673863597214222, + 0.012866348028182983, + 0.013525707647204399, + 0.010288640856742859, + 0.006732219830155373, + 0.007114602718502283, + 0.015344354324042797, + 0.00760940695181489, + 0.004787069279700518, + 0.009720837697386742, + 0.014300116337835789, + 0.008479924872517586, + 0.01358463242650032 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 4 + ], + "selected_reward": -0.012560781091451645 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a medium close-up of a middle-aged craftsman with short, graying hair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the craftsman wearing a dusty, dark blue cotton work suit with sleeves rolled up to his elbows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the craftsman holding a steel chisel in his left hand and a wooden-handled hammer in his right hand?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera slowly panning forward while the craftsman strikes the chisel, sending white stone chips flying?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by a beam of bright sunlight creating a visible light column (Tyndall effect) through high windows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.029802370816469193 + ], + "opposite_logprob_tokens": [ + -3.5298023223876953 + ], + "logprob_reward": -0.029802370816469193, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.029802370816469193, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.22575028240680695, + 0.11301221698522568, + 0.16281476616859436, + 0.07355081290006638, + 0.09644486010074615, + 0.08040618151426315, + 0.10683494806289673, + 0.0595615915954113, + 0.029626641422510147, + 0.05199771746993065 + ], + "credit_map": [ + 0.027761857956647873, + 0.025555817410349846, + 0.05400633066892624, + 0.014501928351819515, + 0.02453925646841526, + 0.017653798684477806, + 0.01936103031039238, + 0.04237023741006851, + 0.015202533453702927, + 0.012766912579536438, + 0.01760568656027317, + 0.008436977863311768, + 0.01278194971382618, + 0.010155824944376945, + 0.014351161196827888, + 0.021711163222789764, + 0.016096487641334534, + 0.019267449155449867, + 0.020495789125561714, + 0.010922105051577091, + 0.012708902359008789, + 0.025752121582627296, + 0.029345989227294922, + 0.028225913643836975, + 0.008797965943813324, + 0.007663573604077101, + 0.011208794079720974, + 0.006222727242857218, + 0.006777544971555471, + 0.010907679796218872, + 0.007735402323305607, + 0.014237119816243649, + 0.011501879431307316, + 0.012752157635986805, + 0.010981344617903233, + 0.009580567479133606, + 0.013144802302122116, + 0.01553888525813818, + 0.01037981640547514, + 0.01256539300084114, + 0.007677940186113119, + 0.009124964475631714, + 0.007714718114584684, + 0.00445513054728508, + 0.01043914444744587, + 0.021323813125491142, + 0.006085389293730259, + 0.013585076667368412, + 0.008098375052213669, + 0.010044570080935955, + 0.011224723421037197, + 0.00834405142813921, + 0.012108278460800648, + 0.0206915270537138, + 0.01630351133644581, + 0.02001989632844925, + 0.0058668977580964565, + 0.010264383628964424, + 0.005982823669910431, + 0.005074481479823589, + 0.00557416258379817, + 0.009038505144417286, + 0.008599583059549332, + 0.009160752408206463, + 0.002585852285847068, + 0.0048637958243489265, + 0.0022155146580189466, + 0.0024119585286825895, + 0.0061849565245211124, + 0.004046320915222168, + 0.004264827352017164, + 0.0030534095130860806, + 0.005467765033245087, + 0.00532197579741478, + 0.005673583131283522, + 0.0028626571875065565, + 0.006545607931911945, + 0.0101783387362957, + 0.0062258439138531685, + 0.00972194317728281 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 1 + ], + "selected_reward": -0.029802370816469193 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a medium close-up of a middle-aged craftsman with short, graying hair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the craftsman wearing a dusty, dark blue cotton work suit with sleeves rolled up to his elbows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the craftsman holding a steel chisel in his left hand and a wooden-handled hammer in his right hand?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera slowly panning forward while the craftsman strikes the chisel, sending white stone chips flying?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by a beam of bright sunlight creating a visible light column (Tyndall effect) through high windows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.08905909210443497 + ], + "opposite_logprob_tokens": [ + -2.4640591144561768 + ], + "logprob_reward": -0.08905909210443497, + "yes_no_margin_reward": 2.375, + "token_credit_reward": -0.08905909210443497, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2040300816297531, + 0.15180402994155884, + 0.11911021173000336, + 0.0819559320807457, + 0.07163497060537338, + 0.07486525923013687, + 0.05407322198152542, + 0.0586361289024353, + 0.11486881971359253, + 0.0690213218331337 + ], + "credit_map": [ + 0.033912546932697296, + 0.029468610882759094, + 0.0299922414124012, + 0.015154525637626648, + 0.02286149375140667, + 0.03062293492257595, + 0.012737592682242393, + 0.02928014099597931, + 0.024727605283260345, + 0.0233509112149477, + 0.021680399775505066, + 0.01744174025952816, + 0.01722799427807331, + 0.014659436419606209, + 0.01668585278093815, + 0.016030089929699898, + 0.015702467411756516, + 0.018873395398259163, + 0.017321202903985977, + 0.01151467114686966, + 0.008525733835995197, + 0.0157499760389328, + 0.017034152522683144, + 0.014388609677553177, + 0.008246143348515034, + 0.01458986010402441, + 0.01078941859304905, + 0.005225358065217733, + 0.008433775044977665, + 0.011322380974888802, + 0.008609910495579243, + 0.014739089645445347, + 0.009022869169712067, + 0.012104867957532406, + 0.00828795786947012, + 0.005400531925261021, + 0.010265274904668331, + 0.009666415862739086, + 0.007247129455208778, + 0.009639923460781574, + 0.005814907141029835, + 0.01040948461741209, + 0.007852484472095966, + 0.0049356697127223015, + 0.009533565491437912, + 0.01905437372624874, + 0.006639509461820126, + 0.010625267401337624, + 0.005339764058589935, + 0.006967251189053059, + 0.006312012206763029, + 0.003520856611430645, + 0.006088494788855314, + 0.009581436403095722, + 0.005429640877991915, + 0.010833765380084515, + 0.004886756185442209, + 0.0080262441188097, + 0.005496988072991371, + 0.004096189048141241, + 0.006541747599840164, + 0.013807757757604122, + 0.005919648800045252, + 0.009860791265964508, + 0.011156249791383743, + 0.012157034128904343, + 0.008812179788947105, + 0.0079293018206954, + 0.02263609692454338, + 0.021200524643063545, + 0.01728082075715065, + 0.01369661558419466, + 0.00561282504349947, + 0.008235854096710682, + 0.008013932965695858, + 0.004200718831270933, + 0.007931578904390335, + 0.011584466323256493, + 0.008169102482497692, + 0.015272846445441246 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 2 + ], + "selected_reward": -0.08905909210443497 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a medium close-up of a middle-aged craftsman with short, graying hair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the craftsman wearing a dusty, dark blue cotton work suit with sleeves rolled up to his elbows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the craftsman holding a steel chisel in his left hand and a wooden-handled hammer in his right hand?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera slowly panning forward while the craftsman strikes the chisel, sending white stone chips flying?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by a beam of bright sunlight creating a visible light column (Tyndall effect) through high windows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01818496361374855 + ], + "opposite_logprob_tokens": [ + -4.018185138702393 + ], + "logprob_reward": -0.01818496361374855, + "yes_no_margin_reward": 4.0, + "token_credit_reward": -0.01818496361374855, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11495263874530792, + 0.05858152359724045, + 0.03975381329655647, + 0.03276349976658821, + 0.035856977105140686, + 0.062040530145168304, + 0.05254748836159706, + 0.09851023554801941, + 0.37973591685295105, + 0.1252574324607849 + ], + "credit_map": [ + 0.013082199729979038, + 0.01477963849902153, + 0.016854271292686462, + 0.007897083647549152, + 0.021455571055412292, + 0.01301155611872673, + 0.011009793728590012, + 0.016862526535987854, + 0.005423325579613447, + 0.006016306113451719, + 0.010079311206936836, + 0.003989113960415125, + 0.007350138388574123, + 0.006123082246631384, + 0.011597190983593464, + 0.008003052324056625, + 0.0039020974654704332, + 0.005303150974214077, + 0.005629283841699362, + 0.003734160214662552, + 0.003926777746528387, + 0.006628096103668213, + 0.004018255043774843, + 0.0066119893454015255, + 0.002645796397700906, + 0.0037533175200223923, + 0.004344334360212088, + 0.002418405609205365, + 0.0034396746195852757, + 0.005575149320065975, + 0.004052192438393831, + 0.006534629967063665, + 0.003095268504694104, + 0.004037543199956417, + 0.003925061784684658, + 0.0026260025333613157, + 0.0052605583332479, + 0.008661869913339615, + 0.003133838064968586, + 0.005116830579936504, + 0.00636800704523921, + 0.006021510809659958, + 0.005894090514630079, + 0.0028647759463638067, + 0.00788592454046011, + 0.014205540530383587, + 0.007343333680182695, + 0.011457348242402077, + 0.003872321220114827, + 0.007337857037782669, + 0.005363958887755871, + 0.004603244364261627, + 0.006404820363968611, + 0.011082474142313004, + 0.0066572208888828754, + 0.007225590758025646, + 0.007716801017522812, + 0.005995126441121101, + 0.008806548081338406, + 0.005553870927542448, + 0.01329262088984251, + 0.024433057755231857, + 0.015169479884207249, + 0.017542731016874313, + 0.0355636328458786, + 0.035726919770240784, + 0.031641945242881775, + 0.024257533252239227, + 0.07121644914150238, + 0.07440315932035446, + 0.0704926997423172, + 0.03643357381224632, + 0.00981344934552908, + 0.014991605654358864, + 0.012456348165869713, + 0.007266158238053322, + 0.015318041667342186, + 0.025364801287651062, + 0.014381508342921734, + 0.0256655253469944 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 9, + 0 + ], + "selected_reward": -0.01818496361374855 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000027/group_11/index_tensor([5262, 5262])_group_11_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000027/group_11/index_tensor([5262, 5262])_group_11_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..f0e7b8624d04f463bc25cd4bb6fc4de70e58de39 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000027/group_11/index_tensor([5262, 5262])_group_11_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "In a perplexing, unbroken sequence, a low-angle tracking shot fluidly follows a lone figure as they perform impossible parkour across a rain-soaked, nighttime cityscape. The camera moves with them in dramatic slow motion, capturing their every gravity-defying movement. The figure is clad in a form-fitting, stylish black leather bodysuit that gleams beneath the perpetual drizzle, its texture subtly reflecting the chaotic city lights. A gleaming silver assault rifle is securely strapped to their back, its cold metallic shape capturing and shattering the vibrant neon lights of the surrounding skyscrapers. The scene becomes hyper-stylized as they traverse the dark, damp glass of towering buildings. They effortlessly navigate a shimmering stream of green digital code that drifts from the sky, the glowing characters parting around their bodies like digital curtains. Each raindrop is rendered with crystal clarity, splattering against the leather and walls in a suspended animation. The backdrop is a dizzying abyss of towering buildings and deep shadows, dotted with electric blue and fuchsia neon signs, whose reflections seep into the wet asphalt below, creating an atmosphere of cool, tense and surreal beauty.", + "seed": 6552, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000027/group_11/index_5262_seed_6552_dp_22_rank_44_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 11, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.018299678154289723, + "vlm_logprob": -0.018299678154289723, + "vlm_yes_no_margin": 4.1, + "vlm_token_credit": -0.018299678154289723, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a lone figure wearing a black leather bodysuit?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the figure performing parkour movements across a rain-soaked cityscape at night?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a silver assault rifle visible strapped to the figure's back?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the scene include a stream of green digital code drifting from the sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot a low-angle tracking shot moving with the figure?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.009556971490383148 + ], + "opposite_logprob_tokens": [ + -4.759556770324707 + ], + "logprob_reward": -0.009556971490383148, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.009556971490383148, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11088790744543076, + 0.08604670315980911, + 0.2703937590122223, + 0.07613882422447205, + 0.08838681131601334, + 0.04616758972406387, + 0.050256211310625076, + 0.11951510608196259, + 0.06061677262187004, + 0.09159038960933685 + ], + "credit_map": [ + 0.011182130314409733, + 0.013212219811975956, + 0.01233797799795866, + 0.008367529138922691, + 0.01065132301300764, + 0.022724714130163193, + 0.00905543752014637, + 0.023356568068265915, + 0.006210378371179104, + 0.009874025359749794, + 0.008996807970106602, + 0.00784152839332819, + 0.017123093828558922, + 0.009849938564002514, + 0.012091507203876972, + 0.014059425331652164, + 0.04480213671922684, + 0.018695225939154625, + 0.03951423615217209, + 0.029461929574608803, + 0.01673801802098751, + 0.04191366955637932, + 0.044432446360588074, + 0.03483610227704048, + 0.0100141866132617, + 0.005472883116453886, + 0.007558329962193966, + 0.006544054020196199, + 0.006854139734059572, + 0.011799291707575321, + 0.017911268398165703, + 0.009984668344259262, + 0.013765687122941017, + 0.008347664959728718, + 0.00979782734066248, + 0.006434716284275055, + 0.017748933285474777, + 0.013024590909481049, + 0.009335683658719063, + 0.00993170216679573, + 0.0065058330073952675, + 0.004711763467639685, + 0.0035890028811991215, + 0.004240487236529589, + 0.00799498986452818, + 0.006184859666973352, + 0.00837844330817461, + 0.004562211222946644, + 0.005106243304908276, + 0.006818403024226427, + 0.006728560663759708, + 0.004726755898445845, + 0.008406253531575203, + 0.006188679952174425, + 0.007571178488433361, + 0.004710133653134108, + 0.01378657203167677, + 0.008229798637330532, + 0.012504851445555687, + 0.010554471984505653, + 0.012039012275636196, + 0.017241720110177994, + 0.021234452724456787, + 0.02392422780394554, + 0.005862399935722351, + 0.010151530615985394, + 0.005977080203592777, + 0.005615586880594492, + 0.00841221958398819, + 0.011576475575566292, + 0.007027579937130213, + 0.005993898957967758, + 0.010145057924091816, + 0.010333257727324963, + 0.010555350221693516, + 0.01770598813891411, + 0.013786467723548412, + 0.004830949939787388, + 0.017217347398400307, + 0.007015967275947332 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 7, + 0 + ], + "selected_reward": -0.009556971490383148 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a lone figure wearing a black leather bodysuit?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the figure performing parkour movements across a rain-soaked cityscape at night?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a silver assault rifle visible strapped to the figure's back?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the scene include a stream of green digital code drifting from the sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot a low-angle tracking shot moving with the figure?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.026351768523454666 + ], + "opposite_logprob_tokens": [ + -3.6513516902923584 + ], + "logprob_reward": -0.026351768523454666, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.026351768523454666, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12851592898368835, + 0.0607023611664772, + 0.19959521293640137, + 0.08286509662866592, + 0.06326774507761002, + 0.10775531828403473, + 0.10866857320070267, + 0.12309049069881439, + 0.03551159054040909, + 0.09002765268087387 + ], + "credit_map": [ + 0.004625397268682718, + 0.008820499293506145, + 0.008294633589684963, + 0.006668596062809229, + 0.00876010861247778, + 0.04267284274101257, + 0.006831841077655554, + 0.041842006146907806, + 0.00979422964155674, + 0.007358771748840809, + 0.01009212527424097, + 0.00507572153583169, + 0.008762780576944351, + 0.006280497182160616, + 0.006602058652788401, + 0.006736178416758776, + 0.038464196026325226, + 0.01187372487038374, + 0.03325384482741356, + 0.019861532375216484, + 0.011153818108141422, + 0.029113642871379852, + 0.025464093312621117, + 0.03041035681962967, + 0.008767819032073021, + 0.0053618429228663445, + 0.009669640101492405, + 0.010881200432777405, + 0.007638079579919577, + 0.013518527150154114, + 0.016726845875382423, + 0.010301144793629646, + 0.005959286354482174, + 0.005564315244555473, + 0.006080072373151779, + 0.004342918284237385, + 0.016381917521357536, + 0.009021366015076637, + 0.008857439272105694, + 0.00706043466925621, + 0.0074135554023087025, + 0.007753845304250717, + 0.00565323093906045, + 0.006534759420901537, + 0.03212207928299904, + 0.015494825318455696, + 0.022398941218852997, + 0.01038409024477005, + 0.013960997574031353, + 0.012383089400827885, + 0.012706657871603966, + 0.0110039496794343, + 0.014802644960582256, + 0.014985316433012486, + 0.016900083050131798, + 0.01192583329975605, + 0.0124693363904953, + 0.014612206257879734, + 0.012393745593726635, + 0.01600007899105549, + 0.017159175127744675, + 0.01801234669983387, + 0.01866256073117256, + 0.013781046494841576, + 0.004678657278418541, + 0.00408054469153285, + 0.005884738638997078, + 0.007324427366256714, + 0.0030870891641825438, + 0.004033217206597328, + 0.0034789752680808306, + 0.002943937899544835, + 0.012079541571438313, + 0.011177318170666695, + 0.01340513676404953, + 0.012423132546246052, + 0.016009649261832237, + 0.006235158536583185, + 0.011067530140280724, + 0.007630196399986744 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 0, + 7 + ], + "selected_reward": -0.026351768523454666 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a lone figure wearing a black leather bodysuit?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the figure performing parkour movements across a rain-soaked cityscape at night?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a silver assault rifle visible strapped to the figure's back?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the scene include a stream of green digital code drifting from the sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot a low-angle tracking shot moving with the figure?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.018180398270487785 + ], + "opposite_logprob_tokens": [ + -4.0181803703308105 + ], + "logprob_reward": -0.018180398270487785, + "yes_no_margin_reward": 4.0, + "token_credit_reward": -0.018180398270487785, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1946203112602234, + 0.05893492326140404, + 0.1143583431839943, + 0.12206965684890747, + 0.11724008619785309, + 0.05917275696992874, + 0.05960264429450035, + 0.12100967764854431, + 0.05741852521896362, + 0.09557300060987473 + ], + "credit_map": [ + 0.015137993730604649, + 0.017300667241215706, + 0.016609661281108856, + 0.01782173477113247, + 0.01928767003118992, + 0.05013062804937363, + 0.017456388100981712, + 0.04087559133768082, + 0.005395839456468821, + 0.005830809473991394, + 0.007030651438981295, + 0.0057378727942705154, + 0.011354462243616581, + 0.006416879594326019, + 0.007825801149010658, + 0.009342612698674202, + 0.024182328954339027, + 0.0093489158898592, + 0.021959714591503143, + 0.008494311012327671, + 0.007634295150637627, + 0.015016545541584492, + 0.013436841778457165, + 0.014285396784543991, + 0.013524294830858707, + 0.00978781096637249, + 0.01184967439621687, + 0.02414281852543354, + 0.01551058143377304, + 0.014218899421393871, + 0.017930341884493828, + 0.015105248428881168, + 0.01514812558889389, + 0.012624816037714481, + 0.012269161641597748, + 0.010003515519201756, + 0.023064980283379555, + 0.018150122836232185, + 0.014417597092688084, + 0.01156176533550024, + 0.006581202615052462, + 0.006620481144636869, + 0.006278239656239748, + 0.0043829139322042465, + 0.011911280453205109, + 0.007706283126026392, + 0.010174985975027084, + 0.005517373792827129, + 0.005852423142641783, + 0.006300268229097128, + 0.009318141266703606, + 0.006190223153680563, + 0.00841866061091423, + 0.008379918523132801, + 0.008494396694004536, + 0.006648617796599865, + 0.01704470068216324, + 0.011418357491493225, + 0.0157217625528574, + 0.015800245106220245, + 0.016794353723526, + 0.017453063279390335, + 0.013266485184431076, + 0.013510718941688538, + 0.00616254610940814, + 0.008089940063655376, + 0.007138892076909542, + 0.006942541338503361, + 0.00886328425258398, + 0.009109408594667912, + 0.007581403944641352, + 0.0035305130295455456, + 0.01350274309515953, + 0.010261185467243195, + 0.01195712573826313, + 0.012463867664337158, + 0.012541857548058033, + 0.006301152054220438, + 0.016792908310890198, + 0.011752166785299778 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 7 + ], + "selected_reward": -0.018180398270487785 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a lone figure wearing a black leather bodysuit?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the figure performing parkour movements across a rain-soaked cityscape at night?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a silver assault rifle visible strapped to the figure's back?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the scene include a stream of green digital code drifting from the sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot a low-angle tracking shot moving with the figure?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.026339344680309296 + ], + "opposite_logprob_tokens": [ + -3.651339292526245 + ], + "logprob_reward": -0.026339344680309296, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.026339344680309296, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1330820918083191, + 0.0372023768723011, + 0.1189972385764122, + 0.05118295177817345, + 0.2639939486980438, + 0.044798169285058975, + 0.0776318907737732, + 0.1221947968006134, + 0.08401626348495483, + 0.06690028309822083 + ], + "credit_map": [ + 0.008287143893539906, + 0.008989761583507061, + 0.009060356765985489, + 0.006569806020706892, + 0.008615692146122456, + 0.04363603517413139, + 0.007367101032286882, + 0.040556181222200394, + 0.0038799045141786337, + 0.0033768368884921074, + 0.006021292880177498, + 0.003744544694200158, + 0.005491976626217365, + 0.0043591707944869995, + 0.004659464582800865, + 0.005669183097779751, + 0.02286774106323719, + 0.012195936404168606, + 0.02009476348757744, + 0.00823637843132019, + 0.008785856887698174, + 0.018631793558597565, + 0.01384055893868208, + 0.014344201423227787, + 0.007466049399226904, + 0.0031947498209774494, + 0.00566808320581913, + 0.00517607294023037, + 0.0034343921579420567, + 0.008032050915062428, + 0.006093336269259453, + 0.012118211947381496, + 0.03541547805070877, + 0.02390148490667343, + 0.051705144345760345, + 0.02644619718194008, + 0.019557083025574684, + 0.03468170762062073, + 0.0334913395345211, + 0.03879550099372864, + 0.00746435672044754, + 0.004847277887165546, + 0.005158113315701485, + 0.0031277230009436607, + 0.0065195076167583466, + 0.00597058329731226, + 0.006792386993765831, + 0.004918216727674007, + 0.010126633569598198, + 0.01080012135207653, + 0.009577655233442783, + 0.007061985787004232, + 0.011886403895914555, + 0.01123763807117939, + 0.010640029795467854, + 0.006301417015492916, + 0.011871826834976673, + 0.01318745594471693, + 0.01604343205690384, + 0.01661892794072628, + 0.018326206132769585, + 0.017389947548508644, + 0.016659224405884743, + 0.012097769416868687, + 0.007096915505826473, + 0.012344068847596645, + 0.011371240951120853, + 0.01270380336791277, + 0.009205572307109833, + 0.011720924638211727, + 0.012133792974054813, + 0.007439938373863697, + 0.00961326528340578, + 0.0071938433684408665, + 0.011277793906629086, + 0.009300767444074154, + 0.009078298695385456, + 0.004859148524701595, + 0.007450526114553213, + 0.008126633241772652 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 0, + 7 + ], + "selected_reward": -0.026339344680309296 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a lone figure wearing a black leather bodysuit?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the figure performing parkour movements across a rain-soaked cityscape at night?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a silver assault rifle visible strapped to the figure's back?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the scene include a stream of green digital code drifting from the sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot a low-angle tracking shot moving with the figure?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.011069907806813717 + ], + "opposite_logprob_tokens": [ + -4.5110697746276855 + ], + "logprob_reward": -0.011069907806813717, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.011069907806813717, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.23522965610027313, + 0.0957530066370964, + 0.16787268221378326, + 0.06214887276291847, + 0.0675906240940094, + 0.06712064146995544, + 0.10538483411073685, + 0.10794281959533691, + 0.04753364995121956, + 0.04342322796583176 + ], + "credit_map": [ + 0.01021083164960146, + 0.011701730079948902, + 0.01192265935242176, + 0.01194848958402872, + 0.011118460446596146, + 0.084866002202034, + 0.010165964253246784, + 0.08329553157091141, + 0.017010217532515526, + 0.008946542628109455, + 0.015878483653068542, + 0.011915401555597782, + 0.00842981692403555, + 0.010761355049908161, + 0.011072961613535881, + 0.01173822209239006, + 0.03578169643878937, + 0.011047406122088432, + 0.03187679871916771, + 0.013528628274798393, + 0.008363287895917892, + 0.025671744719147682, + 0.018386781215667725, + 0.02321634069085121, + 0.00797333288937807, + 0.005669098813086748, + 0.007882347330451012, + 0.007773975376039743, + 0.007809733971953392, + 0.00804674532264471, + 0.008531980216503143, + 0.008461660705506802, + 0.007998128421604633, + 0.00865444540977478, + 0.009387331083416939, + 0.005737473256886005, + 0.012596736662089825, + 0.007934419438242912, + 0.007084468845278025, + 0.008197621442377567, + 0.008609045296907425, + 0.006704420316964388, + 0.005439695902168751, + 0.004548172000795603, + 0.011938987299799919, + 0.009300464764237404, + 0.012188645079731941, + 0.008391208946704865, + 0.009622816927731037, + 0.013671498745679855, + 0.010860027745366096, + 0.010418792255222797, + 0.01890607178211212, + 0.014287950471043587, + 0.017599884420633316, + 0.010017786175012589, + 0.01083704549819231, + 0.012495498172938824, + 0.014780032448470592, + 0.01409049704670906, + 0.017577748745679855, + 0.013900156132876873, + 0.012239730916917324, + 0.012022111564874649, + 0.0061073205433785915, + 0.011321639642119408, + 0.0058927470818161964, + 0.004614996258169413, + 0.005276600364595652, + 0.006612107623368502, + 0.004129140172153711, + 0.0035790940746665, + 0.004814125131815672, + 0.004986217711120844, + 0.007086870726197958, + 0.006163138430565596, + 0.0056062922812998295, + 0.003667424200102687, + 0.005534703377634287, + 0.005564453080296516 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 7 + ], + "selected_reward": -0.011069907806813717 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "In a perplexing, unbroken sequence, a low-angle tracking shot fluidly follows a lone figure as they perform impossible parkour across a rain-soaked, nighttime cityscape. The camera moves with them in dramatic slow motion, capturing their every gravity-defying movement. The figure is clad in a form-fitting, stylish black leather bodysuit that gleams beneath the perpetual drizzle, its texture subtly reflecting the chaotic city lights. A gleaming silver assault rifle is securely strapped to their back, its cold metallic shape capturing and shattering the vibrant neon lights of the surrounding skyscrapers. The scene becomes hyper-stylized as they traverse the dark, damp glass of towering buildings. They effortlessly navigate a shimmering stream of green digital code that drifts from the sky, the glowing characters parting around their bodies like digital curtains. Each raindrop is rendered with crystal clarity, splattering against the leather and walls in a suspended animation. The backdrop is a dizzying abyss of towering buildings and deep shadows, dotted with electric blue and fuchsia neon signs, whose reflections seep into the wet asphalt below, creating an atmosphere of cool, tense and surreal beauty.", + "seed": 6553, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000027/group_11/index_5262_seed_6553_dp_22_rank_44_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 11, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.023342438228428363, + "vlm_logprob": -0.023342438228428363, + "vlm_yes_no_margin": 3.974999952316284, + "vlm_token_credit": -0.023342438228428363, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a lone figure wearing a black leather bodysuit?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the figure performing parkour movements across a rain-soaked cityscape at night?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a silver assault rifle visible strapped to the figure's back?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the scene include a stream of green digital code drifting from the sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot a low-angle tracking shot moving with the figure?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.018902992829680443 + ], + "opposite_logprob_tokens": [ + -4.018902778625488 + ], + "logprob_reward": -0.018902992829680443, + "yes_no_margin_reward": 3.999999761581421, + "token_credit_reward": -0.018902992829680443, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08312429487705231, + 0.04968307539820671, + 0.09774917364120483, + 0.15936116874217987, + 0.05649309232831001, + 0.06885504722595215, + 0.07359008491039276, + 0.27460750937461853, + 0.1188545674085617, + 0.017682000994682312 + ], + "credit_map": [ + 0.0058337487280368805, + 0.006443011574447155, + 0.007158656604588032, + 0.008591121062636375, + 0.01822306588292122, + 0.012529143132269382, + 0.0139425965026021, + 0.010402948595583439, + 0.006489391438663006, + 0.005480882711708546, + 0.0050454651936888695, + 0.0048838951624929905, + 0.004655116703361273, + 0.0077449968084692955, + 0.00726680364459753, + 0.008116521872580051, + 0.002791758393868804, + 0.0036412933841347694, + 0.004728787578642368, + 0.005907988175749779, + 0.004285457078367472, + 0.032868415117263794, + 0.006818436086177826, + 0.036707036197185516, + 0.01165852602571249, + 0.021267596632242203, + 0.01436225138604641, + 0.03342742472887039, + 0.014000261202454567, + 0.018852556124329567, + 0.020176533609628677, + 0.02561601996421814, + 0.003738740226253867, + 0.0069258990697562695, + 0.010979407466948032, + 0.006156021263450384, + 0.007347573060542345, + 0.011792279779911041, + 0.00556213641539216, + 0.003991035744547844, + 0.00966493971645832, + 0.009746505878865719, + 0.008630504831671715, + 0.011769982986152172, + 0.007243632338941097, + 0.010517092421650887, + 0.00584465591236949, + 0.005437733605504036, + 0.008352679200470448, + 0.01408641692250967, + 0.010454915463924408, + 0.014103262685239315, + 0.0069800689816474915, + 0.004464346915483475, + 0.008580468595027924, + 0.006567923352122307, + 0.019222691655158997, + 0.04178369790315628, + 0.021257193759083748, + 0.04891414940357208, + 0.045187998563051224, + 0.03928770497441292, + 0.045729734003543854, + 0.013224350288510323, + 0.011950146406888962, + 0.009362692944705486, + 0.01682317443192005, + 0.025345658883452415, + 0.009690509177744389, + 0.014057178981602192, + 0.015442965552210808, + 0.01618223823606968, + 0.002679766621440649, + 0.002402413636445999, + 0.001934835105203092, + 0.001273688510991633, + 0.001924139098264277, + 0.0029515493661165237, + 0.0021813518833369017, + 0.002334256423637271 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 3, + 8 + ], + "selected_reward": -0.018902992829680443 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a lone figure wearing a black leather bodysuit?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the figure performing parkour movements across a rain-soaked cityscape at night?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a silver assault rifle visible strapped to the figure's back?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the scene include a stream of green digital code drifting from the sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot a low-angle tracking shot moving with the figure?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.054968513548374176 + ], + "opposite_logprob_tokens": [ + -2.9299685955047607 + ], + "logprob_reward": -0.054968513548374176, + "yes_no_margin_reward": 2.875, + "token_credit_reward": -0.054968513548374176, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12952309846878052, + 0.08225822448730469, + 0.1130865141749382, + 0.05098320171236992, + 0.04860805720090866, + 0.041271232068538666, + 0.06450650095939636, + 0.3394605815410614, + 0.07448147982358932, + 0.05582110211253166 + ], + "credit_map": [ + 0.011499582789838314, + 0.011393554508686066, + 0.010376250371336937, + 0.011267535388469696, + 0.02761467918753624, + 0.019295401871204376, + 0.02079181745648384, + 0.017284274101257324, + 0.0077310469932854176, + 0.007358516566455364, + 0.00714923394843936, + 0.006372267846018076, + 0.008513596840202808, + 0.018085097894072533, + 0.008204740472137928, + 0.018843725323677063, + 0.004992325324565172, + 0.006880434229969978, + 0.008950748480856419, + 0.009482751600444317, + 0.005934709683060646, + 0.0320112369954586, + 0.010503526777029037, + 0.03433077409863472, + 0.0033845705911517143, + 0.004031649325042963, + 0.0049098641611635685, + 0.007454647682607174, + 0.006517949979752302, + 0.007442955859005451, + 0.008395001292228699, + 0.008846564218401909, + 0.0037485698703676462, + 0.006663450039923191, + 0.008920446038246155, + 0.004643539898097515, + 0.006721236743032932, + 0.01040650811046362, + 0.004202309995889664, + 0.003301996737718582, + 0.007195803336799145, + 0.004620560444891453, + 0.004585434217005968, + 0.005557466764003038, + 0.005337638780474663, + 0.004890152253210545, + 0.00376688688993454, + 0.005317290313541889, + 0.006380711682140827, + 0.011722847819328308, + 0.009401533752679825, + 0.011525347828865051, + 0.007072643376886845, + 0.004656385164707899, + 0.00778975710272789, + 0.005957276094704866, + 0.02279749885201454, + 0.05161919817328453, + 0.027664517983794212, + 0.058170903474092484, + 0.05466359108686447, + 0.04768475890159607, + 0.05423508584499359, + 0.022625038400292397, + 0.010706550441682339, + 0.007559886667877436, + 0.01009015366435051, + 0.007060339208692312, + 0.011033155955374241, + 0.009949463419616222, + 0.009561416693031788, + 0.008520510047674179, + 0.010039692744612694, + 0.009594254195690155, + 0.0055668894201517105, + 0.0044967238791286945, + 0.006124521140009165, + 0.00997563824057579, + 0.006060961168259382, + 0.0039624241180717945 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 0, + 2 + ], + "selected_reward": -0.054968513548374176 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a lone figure wearing a black leather bodysuit?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the figure performing parkour movements across a rain-soaked cityscape at night?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a silver assault rifle visible strapped to the figure's back?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the scene include a stream of green digital code drifting from the sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot a low-angle tracking shot moving with the figure?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.009780817665159702 + ], + "opposite_logprob_tokens": [ + -4.6347808837890625 + ], + "logprob_reward": -0.009780817665159702, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.009780817665159702, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09285843372344971, + 0.1154792308807373, + 0.06747511029243469, + 0.07605622708797455, + 0.09057013690471649, + 0.21137189865112305, + 0.06014155223965645, + 0.13273486495018005, + 0.06926658749580383, + 0.0840458944439888 + ], + "credit_map": [ + 0.014394021593034267, + 0.005893209483474493, + 0.011254139244556427, + 0.009232813492417336, + 0.010018392466008663, + 0.011645453982055187, + 0.018246246501803398, + 0.012174158357083797, + 0.012950925156474113, + 0.011964495293796062, + 0.0117509039118886, + 0.011868199333548546, + 0.015501061454415321, + 0.01635425165295601, + 0.0178427305072546, + 0.017246665433049202, + 0.0027052590157836676, + 0.0032599796541035175, + 0.002773360814899206, + 0.003211874980479479, + 0.004948899615556002, + 0.02213502675294876, + 0.005496020894497633, + 0.02294468879699707, + 0.00522918626666069, + 0.008309343829751015, + 0.008131672628223896, + 0.011549822054803371, + 0.00987988617271185, + 0.011039943434298038, + 0.009176959283649921, + 0.012739408761262894, + 0.005753884557634592, + 0.020552409812808037, + 0.011049652472138405, + 0.006711218506097794, + 0.015433264896273613, + 0.017176683992147446, + 0.007254085503518581, + 0.00663893623277545, + 0.03706106171011925, + 0.026886502280831337, + 0.025379018858075142, + 0.03404251113533974, + 0.022846287116408348, + 0.020306942984461784, + 0.018808802589774132, + 0.026040783151984215, + 0.007961882278323174, + 0.012161452323198318, + 0.008918426930904388, + 0.009732712991535664, + 0.005155618768185377, + 0.004238754510879517, + 0.005722898058593273, + 0.006249807775020599, + 0.009525363333523273, + 0.019269825890660286, + 0.010821121744811535, + 0.024368125945329666, + 0.01935723051428795, + 0.01996903494000435, + 0.022302960976958275, + 0.007121194154024124, + 0.006795338820666075, + 0.006527986377477646, + 0.010299588553607464, + 0.013113238848745823, + 0.0070390342734754086, + 0.007001066114753485, + 0.00917741097509861, + 0.009312925860285759, + 0.014255916699767113, + 0.014845972880721092, + 0.008942347019910812, + 0.014050337485969067, + 0.006218107882887125, + 0.009169056080281734, + 0.008435127325356007, + 0.008129036985337734 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 7, + 1 + ], + "selected_reward": -0.009780817665159702 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a lone figure wearing a black leather bodysuit?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the figure performing parkour movements across a rain-soaked cityscape at night?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a silver assault rifle visible strapped to the figure's back?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the scene include a stream of green digital code drifting from the sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot a low-angle tracking shot moving with the figure?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.023283299058675766 + ], + "opposite_logprob_tokens": [ + -3.7732832431793213 + ], + "logprob_reward": -0.023283299058675766, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.023283299058675766, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06066657230257988, + 0.04608536139130592, + 0.1048417016863823, + 0.064690962433815, + 0.05004182457923889, + 0.031355440616607666, + 0.049768365919589996, + 0.05459625646471977, + 0.05170338228344917, + 0.4862501919269562 + ], + "credit_map": [ + 0.00583959836512804, + 0.004720137920230627, + 0.008281980641186237, + 0.007045015227049589, + 0.008752280846238136, + 0.00982052180916071, + 0.007731778081506491, + 0.00847526267170906, + 0.004945697262883186, + 0.004613526631146669, + 0.006050020921975374, + 0.00440159672871232, + 0.005480397026985884, + 0.006823804695159197, + 0.007796403951942921, + 0.0059739165008068085, + 0.0041238837875425816, + 0.0057832542806863785, + 0.005773063283413649, + 0.005377042572945356, + 0.005264751613140106, + 0.03310447558760643, + 0.008729272522032261, + 0.03668596222996712, + 0.0039022001437842846, + 0.006208764389157295, + 0.005019589327275753, + 0.010139762423932552, + 0.008259305730462074, + 0.01244308054447174, + 0.007040542084723711, + 0.011677718721330166, + 0.0037045348435640335, + 0.007200290914624929, + 0.008102765306830406, + 0.004929486662149429, + 0.006539131049066782, + 0.011991809122264385, + 0.004282514564692974, + 0.0032912935130298138, + 0.005481458269059658, + 0.002853800542652607, + 0.004613037686794996, + 0.0040631890296936035, + 0.003738460596650839, + 0.005229187663644552, + 0.0022784462198615074, + 0.00309785851277411, + 0.00907097477465868, + 0.013150466606020927, + 0.00865841656923294, + 0.006442044861614704, + 0.003989900927990675, + 0.0026049031876027584, + 0.0032003798987716436, + 0.002651280490681529, + 0.005535653326660395, + 0.00801704078912735, + 0.0059815370477736, + 0.00573732191696763, + 0.009260234422981739, + 0.008752821013331413, + 0.00821249932050705, + 0.003099146531894803, + 0.0075260912999510765, + 0.0065771727822721004, + 0.006407764740288258, + 0.00682388711720705, + 0.005618577357381582, + 0.006002085283398628, + 0.006545730400830507, + 0.0062020705081522465, + 0.07546442002058029, + 0.07844867557287216, + 0.055609896779060364, + 0.07971403002738953, + 0.041864681988954544, + 0.055313486605882645, + 0.05329539254307747, + 0.046539586037397385 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 2, + 3 + ], + "selected_reward": -0.023283299058675766 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a lone figure wearing a black leather bodysuit?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the figure performing parkour movements across a rain-soaked cityscape at night?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a silver assault rifle visible strapped to the figure's back?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the scene include a stream of green digital code drifting from the sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot a low-angle tracking shot moving with the figure?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.009776568040251732 + ], + "opposite_logprob_tokens": [ + -4.634776592254639 + ], + "logprob_reward": -0.009776568040251732, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.009776568040251732, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06203721463680267, + 0.05432684347033501, + 0.2078961580991745, + 0.08072852343320847, + 0.20433320105075836, + 0.0458466000854969, + 0.03695835545659065, + 0.14848895370960236, + 0.10534940659999847, + 0.05403480306267738 + ], + "credit_map": [ + 0.004392432048916817, + 0.005814951378852129, + 0.006445886101573706, + 0.008355275727808475, + 0.011915775015950203, + 0.008647597394883633, + 0.00990983471274376, + 0.006555458530783653, + 0.005446933209896088, + 0.004961170721799135, + 0.009033686481416225, + 0.005663442425429821, + 0.005603953264653683, + 0.007926207035779953, + 0.007634528446942568, + 0.008056911639869213, + 0.006127441767603159, + 0.006206539459526539, + 0.008925667963922024, + 0.011332719586789608, + 0.010609009303152561, + 0.07223118841648102, + 0.013097002170979977, + 0.07936657965183258, + 0.006561096757650375, + 0.008833077736198902, + 0.008271799422800541, + 0.01145890261977911, + 0.00977279245853424, + 0.015262848697602749, + 0.007059368770569563, + 0.013508628122508526, + 0.015036055818200111, + 0.030584139749407768, + 0.034211792051792145, + 0.014081589877605438, + 0.029490290209650993, + 0.05166985094547272, + 0.013478085398674011, + 0.015781383961439133, + 0.006510313134640455, + 0.006082190200686455, + 0.005358087830245495, + 0.008305463939905167, + 0.005161297507584095, + 0.005207266192883253, + 0.0043767415918409824, + 0.004845235031098127, + 0.005506966728717089, + 0.00841075461357832, + 0.005729868542402983, + 0.006006836891174316, + 0.00256216898560524, + 0.002611577045172453, + 0.0036754498723894358, + 0.0024547285865992308, + 0.011967896483838558, + 0.025183677673339844, + 0.011962736956775188, + 0.023800697177648544, + 0.02696184627711773, + 0.020245863124728203, + 0.0217157993465662, + 0.006650421302765608, + 0.00957188569009304, + 0.009546210058033466, + 0.01508693490177393, + 0.0207620058208704, + 0.009183622896671295, + 0.011495515704154968, + 0.015153431333601475, + 0.014549797400832176, + 0.0059004370123147964, + 0.008690510876476765, + 0.005653820000588894, + 0.009874951094388962, + 0.004331517964601517, + 0.0057142749428749084, + 0.007536540273576975, + 0.006332745775580406 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 4, + 7 + ], + "selected_reward": -0.009776568040251732 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "In a perplexing, unbroken sequence, a low-angle tracking shot fluidly follows a lone figure as they perform impossible parkour across a rain-soaked, nighttime cityscape. The camera moves with them in dramatic slow motion, capturing their every gravity-defying movement. The figure is clad in a form-fitting, stylish black leather bodysuit that gleams beneath the perpetual drizzle, its texture subtly reflecting the chaotic city lights. A gleaming silver assault rifle is securely strapped to their back, its cold metallic shape capturing and shattering the vibrant neon lights of the surrounding skyscrapers. The scene becomes hyper-stylized as they traverse the dark, damp glass of towering buildings. They effortlessly navigate a shimmering stream of green digital code that drifts from the sky, the glowing characters parting around their bodies like digital curtains. Each raindrop is rendered with crystal clarity, splattering against the leather and walls in a suspended animation. The backdrop is a dizzying abyss of towering buildings and deep shadows, dotted with electric blue and fuchsia neon signs, whose reflections seep into the wet asphalt below, creating an atmosphere of cool, tense and surreal beauty.", + "seed": 6554, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000027/group_11/index_5262_seed_6554_dp_23_rank_46_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 11, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.028173770010471343, + "vlm_logprob": -0.028173770010471343, + "vlm_yes_no_margin": 3.75, + "vlm_token_credit": -0.028173770010471343, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a lone figure wearing a black leather bodysuit?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the figure performing parkour movements across a rain-soaked cityscape at night?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a silver assault rifle visible strapped to the figure's back?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the scene include a stream of green digital code drifting from the sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot a low-angle tracking shot moving with the figure?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.034130919724702835 + ], + "opposite_logprob_tokens": [ + -3.409130811691284 + ], + "logprob_reward": -0.034130919724702835, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.034130919724702835, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15753795206546783, + 0.09443924576044083, + 0.17306426167488098, + 0.049708157777786255, + 0.07705099880695343, + 0.0340433344244957, + 0.10229858756065369, + 0.04330049455165863, + 0.08285991102457047, + 0.18569707870483398 + ], + "credit_map": [ + 0.010821756906807423, + 0.012919269502162933, + 0.008440449833869934, + 0.024367665871977806, + 0.011245088651776314, + 0.007829836569726467, + 0.033931147307157516, + 0.047982748597860336, + 0.010045481845736504, + 0.014653278514742851, + 0.008745484054088593, + 0.010012899525463581, + 0.011674067005515099, + 0.013818302191793919, + 0.011953616514801979, + 0.013536115176975727, + 0.011631092056632042, + 0.028204116970300674, + 0.012123685330152512, + 0.019411712884902954, + 0.03115191124379635, + 0.025941791012883186, + 0.022690484300255775, + 0.02190946415066719, + 0.00409731175750494, + 0.005942598916590214, + 0.0036635729484260082, + 0.007183598820120096, + 0.00745417969301343, + 0.006938031408935785, + 0.00799699779599905, + 0.0064318678341805935, + 0.006188794504851103, + 0.005614956840872765, + 0.0042456672526896, + 0.004870267119258642, + 0.00787444319576025, + 0.020432131364941597, + 0.01233239658176899, + 0.015492342412471771, + 0.003235150594264269, + 0.0036023035645484924, + 0.003285292536020279, + 0.005127103999257088, + 0.003204864216968417, + 0.005050105042755604, + 0.005317229311913252, + 0.005221287254244089, + 0.00834773387759924, + 0.00991109013557434, + 0.007923828437924385, + 0.015059420838952065, + 0.01657242141664028, + 0.010889622382819653, + 0.021068137139081955, + 0.012526331469416618, + 0.004774684552103281, + 0.004126565996557474, + 0.0031201110687106848, + 0.003978679422289133, + 0.006777688395231962, + 0.006364875473082066, + 0.0057962811551988125, + 0.00836160872131586, + 0.006279420107603073, + 0.007212809287011623, + 0.008411429822444916, + 0.012426914647221565, + 0.009509297087788582, + 0.011457264423370361, + 0.012433628551661968, + 0.015129143372178078, + 0.043659113347530365, + 0.02049759216606617, + 0.016197171062231064, + 0.010583637282252312, + 0.011058485135436058, + 0.050403792411088943, + 0.018085183575749397, + 0.01521210465580225 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 2, + 0 + ], + "selected_reward": -0.034130919724702835 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a lone figure wearing a black leather bodysuit?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the figure performing parkour movements across a rain-soaked cityscape at night?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a silver assault rifle visible strapped to the figure's back?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the scene include a stream of green digital code drifting from the sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot a low-angle tracking shot moving with the figure?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.06207158789038658 + ], + "opposite_logprob_tokens": [ + -2.8120715618133545 + ], + "logprob_reward": -0.06207158789038658, + "yes_no_margin_reward": 2.75, + "token_credit_reward": -0.06207158789038658, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1548161655664444, + 0.04538877308368683, + 0.21871142089366913, + 0.03889698162674904, + 0.22468936443328857, + 0.024447815492749214, + 0.04414540156722069, + 0.014948050491511822, + 0.03062024712562561, + 0.20333576202392578 + ], + "credit_map": [ + 0.01308670174330473, + 0.011481541208922863, + 0.009198215790092945, + 0.021158665418624878, + 0.010529657825827599, + 0.01148707140237093, + 0.04476284235715866, + 0.03311144933104515, + 0.005324481055140495, + 0.00437987269833684, + 0.004181456286460161, + 0.007204830180853605, + 0.005173333454877138, + 0.006622706539928913, + 0.006259145215153694, + 0.006242944858968258, + 0.015516837127506733, + 0.0357208214700222, + 0.015136457979679108, + 0.02130766585469246, + 0.04067039117217064, + 0.03503533825278282, + 0.027019359171390533, + 0.02830454520881176, + 0.0036045711021870375, + 0.003948088735342026, + 0.003859982592985034, + 0.005371356848627329, + 0.005432319827377796, + 0.006115846801549196, + 0.005122579168528318, + 0.005442232359200716, + 0.017302606254816055, + 0.014426715672016144, + 0.009902125224471092, + 0.016356950625777245, + 0.02204432524740696, + 0.07400022447109222, + 0.02592451311647892, + 0.044731881469488144, + 0.0018005709862336516, + 0.0019429120002314448, + 0.002832841593772173, + 0.004306760150939226, + 0.0022380799055099487, + 0.0035640685819089413, + 0.0039176857098937035, + 0.0038448970299214125, + 0.0035755273420363665, + 0.004216659814119339, + 0.0049612196162343025, + 0.008976324461400509, + 0.005501878447830677, + 0.005129171535372734, + 0.00640614191070199, + 0.005378470290452242, + 0.0014948907773941755, + 0.0012099712621420622, + 0.0007574206683784723, + 0.0012179812183603644, + 0.002218829933553934, + 0.0026374775916337967, + 0.002021371154114604, + 0.0033901066053658724, + 0.0026070622261613607, + 0.0034385237377136946, + 0.0027471508365124464, + 0.003519979305565357, + 0.0037846704944968224, + 0.004908611532300711, + 0.005071183666586876, + 0.004543064162135124, + 0.05263713374733925, + 0.016743384301662445, + 0.02072015590965748, + 0.006776283495128155, + 0.0076454742811620235, + 0.06218784675002098, + 0.02247484400868416, + 0.01415063813328743 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 2, + 9 + ], + "selected_reward": -0.06207158789038658 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a lone figure wearing a black leather bodysuit?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the figure performing parkour movements across a rain-soaked cityscape at night?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a silver assault rifle visible strapped to the figure's back?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the scene include a stream of green digital code drifting from the sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot a low-angle tracking shot moving with the figure?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.016066139563918114 + ], + "opposite_logprob_tokens": [ + -4.141066074371338 + ], + "logprob_reward": -0.016066139563918114, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.016066139563918114, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.4690774083137512, + 0.05190051719546318, + 0.0981810912489891, + 0.030733391642570496, + 0.06668408215045929, + 0.056669048964977264, + 0.04025226831436157, + 0.05889012664556503, + 0.060176119208335876, + 0.06743592768907547 + ], + "credit_map": [ + 0.052032459527254105, + 0.03324724733829498, + 0.025639817118644714, + 0.05533502250909805, + 0.05326827988028526, + 0.032112378627061844, + 0.16638684272766113, + 0.05105536803603172, + 0.006935677956789732, + 0.006204633973538876, + 0.005597375333309174, + 0.006771290674805641, + 0.006613819859921932, + 0.007351494859904051, + 0.006810061167925596, + 0.005616168491542339, + 0.007784613408148289, + 0.016389671713113785, + 0.007774068973958492, + 0.011318733915686607, + 0.013276750221848488, + 0.014180252328515053, + 0.015213873237371445, + 0.012243121862411499, + 0.002955602016299963, + 0.0027867979370057583, + 0.002297869184985757, + 0.004982380196452141, + 0.0044556413777172565, + 0.004187134560197592, + 0.00511572603136301, + 0.003952242899686098, + 0.006149215158075094, + 0.00799979642033577, + 0.006319515872746706, + 0.006132204085588455, + 0.008060862310230732, + 0.013140900991857052, + 0.007904603146016598, + 0.010976983234286308, + 0.003760643070563674, + 0.004629202652722597, + 0.0059953308664262295, + 0.009908465668559074, + 0.005194846075028181, + 0.007984639145433903, + 0.01022826787084341, + 0.008967652916908264, + 0.0029291382525116205, + 0.0038531972095370293, + 0.0035920829977840185, + 0.005928268190473318, + 0.005800256505608559, + 0.00412259204313159, + 0.008793654851615429, + 0.005233080126345158, + 0.00503139803186059, + 0.005606483668088913, + 0.002874495927244425, + 0.006773712579160929, + 0.00736693711951375, + 0.010568642988801003, + 0.008396891877055168, + 0.012271564453840256, + 0.0036587605718523264, + 0.007212566211819649, + 0.005050957668572664, + 0.00985781755298376, + 0.0072212521918118, + 0.009250354021787643, + 0.00817171297967434, + 0.009752694517374039, + 0.011632806621491909, + 0.011657578870654106, + 0.007302031386643648, + 0.00753122940659523, + 0.006427962798625231, + 0.011509907431900501, + 0.004784930031746626, + 0.0065894825384020805 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 9 + ], + "selected_reward": -0.016066139563918114 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a lone figure wearing a black leather bodysuit?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the figure performing parkour movements across a rain-soaked cityscape at night?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a silver assault rifle visible strapped to the figure's back?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the scene include a stream of green digital code drifting from the sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot a low-angle tracking shot moving with the figure?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.016064731404185295 + ], + "opposite_logprob_tokens": [ + -4.141064643859863 + ], + "logprob_reward": -0.016064731404185295, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.016064731404185295, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3685675263404846, + 0.06802250444889069, + 0.09321565181016922, + 0.044815193861722946, + 0.05566169694066048, + 0.022373488172888756, + 0.016094474121928215, + 0.016498208045959473, + 0.02978924661874771, + 0.2849620580673218 + ], + "credit_map": [ + 0.04742933064699173, + 0.030439836904406548, + 0.024125222116708755, + 0.035870570689439774, + 0.04761757329106331, + 0.0333046056330204, + 0.11423910409212112, + 0.03554126247763634, + 0.005682757124304771, + 0.007696939632296562, + 0.006496403366327286, + 0.0062544033862650394, + 0.0072851888835430145, + 0.011337361298501492, + 0.009843593463301659, + 0.013425848446786404, + 0.006905955262482166, + 0.013577459380030632, + 0.00817814189940691, + 0.011521662585437298, + 0.01611650362610817, + 0.013431623578071594, + 0.011706927791237831, + 0.011777365580201149, + 0.004051125142723322, + 0.00477567920461297, + 0.0040373788215219975, + 0.006640303414314985, + 0.004336833488196135, + 0.007478635758161545, + 0.006205590907484293, + 0.007289647590368986, + 0.00590334739536047, + 0.006421581842005253, + 0.006255557760596275, + 0.0056022475473582745, + 0.00858961883932352, + 0.009463286958634853, + 0.006635380908846855, + 0.006790670100599527, + 0.002195890061557293, + 0.0023058129008859396, + 0.002825565403327346, + 0.00323444209061563, + 0.0017331212293356657, + 0.0036929266061633825, + 0.0033942293375730515, + 0.0029914970509707928, + 0.0016185161657631397, + 0.001546265440993011, + 0.002272211015224457, + 0.0016823740443214774, + 0.00198866194114089, + 0.0022561438381671906, + 0.002768254606053233, + 0.0019620463717728853, + 0.001629662117920816, + 0.0014900914393365383, + 0.0014662602916359901, + 0.002416397212073207, + 0.0019550512079149485, + 0.002586320275440812, + 0.002020089188590646, + 0.002934335032477975, + 0.002295480575412512, + 0.002484894124791026, + 0.0022607867140322924, + 0.0038384203799068928, + 0.002828910481184721, + 0.004622193519026041, + 0.005006922874599695, + 0.006451636552810669, + 0.057770222425460815, + 0.0400671623647213, + 0.028296692296862602, + 0.022408710792660713, + 0.015215937048196793, + 0.07537510991096497, + 0.026002300903201103, + 0.019825896248221397 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 2 + ], + "selected_reward": -0.016064731404185295 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a lone figure wearing a black leather bodysuit?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the figure performing parkour movements across a rain-soaked cityscape at night?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a silver assault rifle visible strapped to the figure's back?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the scene include a stream of green digital code drifting from the sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot a low-angle tracking shot moving with the figure?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.012535471469163895 + ], + "opposite_logprob_tokens": [ + -4.387535572052002 + ], + "logprob_reward": -0.012535471469163895, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.012535471469163895, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14786086976528168, + 0.06884662061929703, + 0.12206675112247467, + 0.1363966464996338, + 0.17287717759609222, + 0.0408303439617157, + 0.04407350346446037, + 0.027719411998987198, + 0.021045979112386703, + 0.21828265488147736 + ], + "credit_map": [ + 0.014385733753442764, + 0.020819716155529022, + 0.009940161369740963, + 0.03104991279542446, + 0.013130811043083668, + 0.008361033163964748, + 0.021676786243915558, + 0.028496718034148216, + 0.006504044868052006, + 0.007449879311025143, + 0.007028832565993071, + 0.009093502536416054, + 0.008152260445058346, + 0.009732196107506752, + 0.008520688861608505, + 0.01236521452665329, + 0.010538021102547646, + 0.01989048719406128, + 0.008461805060505867, + 0.011892571114003658, + 0.019366363063454628, + 0.020311005413532257, + 0.01666586473584175, + 0.014940638095140457, + 0.009065159596502781, + 0.014926514588296413, + 0.0137372687458992, + 0.017742786556482315, + 0.016381215304136276, + 0.022547468543052673, + 0.022803686559200287, + 0.019192561507225037, + 0.012808021157979965, + 0.010973915457725525, + 0.008471082895994186, + 0.012072351761162281, + 0.016830895096063614, + 0.05906955152750015, + 0.017094776034355164, + 0.035556577146053314, + 0.0040077646262943745, + 0.0032007016707211733, + 0.003776520024985075, + 0.006608623079955578, + 0.0035733170807361603, + 0.006863859947770834, + 0.005890479777008295, + 0.006909077055752277, + 0.003923354670405388, + 0.004164889920502901, + 0.004894992336630821, + 0.008963457308709621, + 0.005878279451280832, + 0.004559892229735851, + 0.006343798246234655, + 0.005344841163605452, + 0.0024256776086986065, + 0.003456014208495617, + 0.0017001335509121418, + 0.0032204247545450926, + 0.004064269829541445, + 0.004046535585075617, + 0.003026608843356371, + 0.005779747385531664, + 0.0016544930404052138, + 0.0018337202491238713, + 0.0016663846326991916, + 0.0031381435692310333, + 0.0030765621922910213, + 0.0036521167494356632, + 0.0031187173444777727, + 0.0029058409854769707, + 0.054428163915872574, + 0.017887093126773834, + 0.020976122468709946, + 0.010181084275245667, + 0.01008754875510931, + 0.06535691767930984, + 0.023241214454174042, + 0.016124488785862923 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 4, + 0 + ], + "selected_reward": -0.012535471469163895 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "In a perplexing, unbroken sequence, a low-angle tracking shot fluidly follows a lone figure as they perform impossible parkour across a rain-soaked, nighttime cityscape. The camera moves with them in dramatic slow motion, capturing their every gravity-defying movement. The figure is clad in a form-fitting, stylish black leather bodysuit that gleams beneath the perpetual drizzle, its texture subtly reflecting the chaotic city lights. A gleaming silver assault rifle is securely strapped to their back, its cold metallic shape capturing and shattering the vibrant neon lights of the surrounding skyscrapers. The scene becomes hyper-stylized as they traverse the dark, damp glass of towering buildings. They effortlessly navigate a shimmering stream of green digital code that drifts from the sky, the glowing characters parting around their bodies like digital curtains. Each raindrop is rendered with crystal clarity, splattering against the leather and walls in a suspended animation. The backdrop is a dizzying abyss of towering buildings and deep shadows, dotted with electric blue and fuchsia neon signs, whose reflections seep into the wet asphalt below, creating an atmosphere of cool, tense and surreal beauty.", + "seed": 6555, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000027/group_11/index_5262_seed_6555_dp_23_rank_46_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 11, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.033335340488702056, + "vlm_logprob": -0.033335340488702056, + "vlm_yes_no_margin": 3.975, + "vlm_token_credit": -0.033335340488702056, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a lone figure wearing a black leather bodysuit?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the figure performing parkour movements across a rain-soaked cityscape at night?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a silver assault rifle visible strapped to the figure's back?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the scene include a stream of green digital code drifting from the sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot a low-angle tracking shot moving with the figure?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.013720369897782803 + ], + "opposite_logprob_tokens": [ + -4.388720512390137 + ], + "logprob_reward": -0.013720369897782803, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.013720369897782803, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1448037028312683, + 0.22256551682949066, + 0.07160165905952454, + 0.0409412607550621, + 0.07695133984088898, + 0.10368815064430237, + 0.040004342794418335, + 0.21125428378582, + 0.04063381627202034, + 0.04755593091249466 + ], + "credit_map": [ + 0.010117580182850361, + 0.012914787046611309, + 0.018946798518300056, + 0.014403989538550377, + 0.013224456459283829, + 0.02657676301896572, + 0.015356001444160938, + 0.03326333314180374, + 0.01882653310894966, + 0.018205443397164345, + 0.025661056861281395, + 0.04071372374892235, + 0.015706855803728104, + 0.052132513374090195, + 0.026108060032129288, + 0.025211328640580177, + 0.007343723438680172, + 0.006728101056069136, + 0.008199513889849186, + 0.012695875950157642, + 0.007173694204539061, + 0.012876048684120178, + 0.010163281112909317, + 0.006421422120183706, + 0.004502370022237301, + 0.004778560716658831, + 0.00530075840651989, + 0.0043191914446651936, + 0.0062960474751889706, + 0.00571461534127593, + 0.004823924973607063, + 0.0052057895809412, + 0.008619247935712337, + 0.007102294359356165, + 0.005552058573812246, + 0.007667135912925005, + 0.008598408661782742, + 0.005851656198501587, + 0.013576998375356197, + 0.019983535632491112, + 0.017625730484724045, + 0.00955683272331953, + 0.008355018682777882, + 0.010900444351136684, + 0.008939891122281551, + 0.015196867287158966, + 0.017129655927419662, + 0.015983710065484047, + 0.011718375608325005, + 0.004717300646007061, + 0.004319803789258003, + 0.003989761229604483, + 0.008491503074765205, + 0.0024916569236665964, + 0.002369292313233018, + 0.0019066515378654003, + 0.012219843454658985, + 0.011863486841320992, + 0.0661020427942276, + 0.020929371938109398, + 0.031637612730264664, + 0.03914759308099747, + 0.01827128790318966, + 0.011083041317760944, + 0.0038921572268009186, + 0.006020036991685629, + 0.004533103667199612, + 0.0047464147210121155, + 0.005090344697237015, + 0.0052606165409088135, + 0.005310350563377142, + 0.005780791863799095, + 0.0037399313878268003, + 0.007661701645702124, + 0.002950782887637615, + 0.0029111660551279783, + 0.009980308823287487, + 0.006596070248633623, + 0.0050793797709047794, + 0.008636588230729103 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 7, + 0 + ], + "selected_reward": -0.013720369897782803 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a lone figure wearing a black leather bodysuit?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the figure performing parkour movements across a rain-soaked cityscape at night?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a silver assault rifle visible strapped to the figure's back?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the scene include a stream of green digital code drifting from the sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot a low-angle tracking shot moving with the figure?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.11294278502464294 + ], + "opposite_logprob_tokens": [ + -2.237942695617676 + ], + "logprob_reward": -0.11294278502464294, + "yes_no_margin_reward": 2.125, + "token_credit_reward": -0.11294278502464294, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09702125191688538, + 0.08960063755512238, + 0.1307489424943924, + 0.06876640021800995, + 0.05691320821642876, + 0.06764120608568192, + 0.127324178814888, + 0.2294982224702835, + 0.07822179794311523, + 0.05426419898867607 + ], + "credit_map": [ + 0.008817978203296661, + 0.015331510454416275, + 0.01819596439599991, + 0.010783426463603973, + 0.00717157032340765, + 0.01301446370780468, + 0.01193331740796566, + 0.01177301350980997, + 0.009710779413580894, + 0.014369996264576912, + 0.012480027042329311, + 0.009669672697782516, + 0.008268123492598534, + 0.01577604189515114, + 0.009184149093925953, + 0.010141837410628796, + 0.010496499948203564, + 0.014000634662806988, + 0.015328757464885712, + 0.023764682933688164, + 0.011359180323779583, + 0.01882927678525448, + 0.02511707879602909, + 0.011852825991809368, + 0.00670705595985055, + 0.008628277108073235, + 0.00894236471503973, + 0.00978960283100605, + 0.010272019542753696, + 0.007540158927440643, + 0.008479265496134758, + 0.008407660759985447, + 0.007877401076257229, + 0.005360430106520653, + 0.0054102204740047455, + 0.005486908368766308, + 0.007344644982367754, + 0.009913223795592785, + 0.006152007728815079, + 0.009368374943733215, + 0.009677937254309654, + 0.005504829343408346, + 0.006653275806456804, + 0.00691747572273016, + 0.008233917877078056, + 0.008555998094379902, + 0.012673888355493546, + 0.009423883631825447, + 0.0392722524702549, + 0.015236726962029934, + 0.011867051012814045, + 0.008318755775690079, + 0.02474098838865757, + 0.00691530155017972, + 0.010608868673443794, + 0.010364232584834099, + 0.019207242876291275, + 0.014628217555582523, + 0.06235537305474281, + 0.014900067821145058, + 0.019252261146903038, + 0.06558191776275635, + 0.01759597659111023, + 0.015977157279849052, + 0.008069561794400215, + 0.01268947683274746, + 0.009223967790603638, + 0.007041086442768574, + 0.010655073449015617, + 0.01057444978505373, + 0.008358823135495186, + 0.011609361506998539, + 0.003861297620460391, + 0.005338974297046661, + 0.003374256892129779, + 0.004688142333179712, + 0.010702904313802719, + 0.008014139719307423, + 0.008235204964876175, + 0.010049277916550636 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 2, + 6 + ], + "selected_reward": -0.11294278502464294 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a lone figure wearing a black leather bodysuit?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the figure performing parkour movements across a rain-soaked cityscape at night?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a silver assault rifle visible strapped to the figure's back?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the scene include a stream of green digital code drifting from the sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot a low-angle tracking shot moving with the figure?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.020573899149894714 + ], + "opposite_logprob_tokens": [ + -3.895573854446411 + ], + "logprob_reward": -0.020573899149894714, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.020573899149894714, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12960930168628693, + 0.2952941358089447, + 0.06488994508981705, + 0.15975281596183777, + 0.050947416573762894, + 0.0747239738702774, + 0.040619995445013046, + 0.09646634757518768, + 0.04410584270954132, + 0.043590325862169266 + ], + "credit_map": [ + 0.011844920925796032, + 0.014159112237393856, + 0.02233501523733139, + 0.013374707661569118, + 0.010559926740825176, + 0.018933281302452087, + 0.01459831465035677, + 0.02380400151014328, + 0.021834049373865128, + 0.02907251939177513, + 0.032255619764328, + 0.05324140936136246, + 0.021063195541501045, + 0.06598690897226334, + 0.03943178430199623, + 0.03240860253572464, + 0.0058115762658417225, + 0.006690838374197483, + 0.005678632762283087, + 0.01340460404753685, + 0.00785097386687994, + 0.009820638224482536, + 0.009525683708488941, + 0.006106994580477476, + 0.015987057238817215, + 0.021014029160141945, + 0.014477725140750408, + 0.017118267714977264, + 0.03016992285847664, + 0.021928012371063232, + 0.02287663333117962, + 0.01618115045130253, + 0.0049733687192201614, + 0.005980359390377998, + 0.004604721907526255, + 0.006276522763073444, + 0.007357163354754448, + 0.00663949828594923, + 0.007724490016698837, + 0.007391286548227072, + 0.01498344074934721, + 0.006819293834269047, + 0.008676917292177677, + 0.009045954793691635, + 0.006898034829646349, + 0.009092840366065502, + 0.008993151597678661, + 0.010214337147772312, + 0.01047681737691164, + 0.007348977029323578, + 0.004271466284990311, + 0.003593382891267538, + 0.004011726472526789, + 0.002125681145116687, + 0.0039903391152620316, + 0.004801599774509668, + 0.006149960216134787, + 0.00931897945702076, + 0.01638946495950222, + 0.014397794380784035, + 0.016682066023349762, + 0.012487166561186314, + 0.01549550425261259, + 0.005545399617403746, + 0.0055021014995872974, + 0.0038956215139478445, + 0.004697548691183329, + 0.005592017434537411, + 0.005601523909717798, + 0.007324159145355225, + 0.005780211184173822, + 0.005712655372917652, + 0.004178005736321211, + 0.006713598035275936, + 0.0034503780771046877, + 0.004123750142753124, + 0.0057257129810750484, + 0.006181051023304462, + 0.0054177334532141685, + 0.007800091989338398 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 3, + 0 + ], + "selected_reward": -0.020573899149894714 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a lone figure wearing a black leather bodysuit?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the figure performing parkour movements across a rain-soaked cityscape at night?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a silver assault rifle visible strapped to the figure's back?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the scene include a stream of green digital code drifting from the sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot a low-angle tracking shot moving with the figure?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005248814355581999 + ], + "opposite_logprob_tokens": [ + -5.2552490234375 + ], + "logprob_reward": -0.005248814355581999, + "yes_no_margin_reward": 5.25, + "token_credit_reward": -0.005248814355581999, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.03806425631046295, + 0.04314269497990608, + 0.04370832443237305, + 0.04436219111084938, + 0.05217466130852699, + 0.025628702715039253, + 0.09226439148187637, + 0.5194190740585327, + 0.09714917838573456, + 0.04408654943108559 + ], + "credit_map": [ + 0.005467210430651903, + 0.00506657175719738, + 0.006852257065474987, + 0.004461399279534817, + 0.0031745757441967726, + 0.004708674270659685, + 0.004158791154623032, + 0.00417478010058403, + 0.004629816394299269, + 0.004564028698951006, + 0.004847337957471609, + 0.007045023143291473, + 0.003316151211038232, + 0.006609585136175156, + 0.006198370363563299, + 0.005932377185672522, + 0.0029985581059008837, + 0.004241719841957092, + 0.004584773909300566, + 0.008602323941886425, + 0.0031775503885000944, + 0.00514237629249692, + 0.010785914026200771, + 0.004175110720098019, + 0.0038524481933563948, + 0.006650240160524845, + 0.0032888895366340876, + 0.004481098148971796, + 0.012369690462946892, + 0.005143152549862862, + 0.005020328797399998, + 0.0035563441924750805, + 0.005023630801588297, + 0.0049570403061807156, + 0.004609199240803719, + 0.006048151757568121, + 0.006082132458686829, + 0.009955091401934624, + 0.007301495410501957, + 0.008197920396924019, + 0.0022300032433122396, + 0.0037077756132930517, + 0.0037931078113615513, + 0.0029279731679707766, + 0.0030056217219680548, + 0.0028244282584637403, + 0.003887247759848833, + 0.0032525453716516495, + 0.034440599381923676, + 0.01563631370663643, + 0.007489720359444618, + 0.003655662527307868, + 0.010463244281709194, + 0.004521544557064772, + 0.006505066063255072, + 0.009552234783768654, + 0.02908419445157051, + 0.02314593829214573, + 0.19756601750850677, + 0.02320675179362297, + 0.03039792738854885, + 0.17274750769138336, + 0.025264892727136612, + 0.0180058516561985, + 0.009139989502727985, + 0.016173118725419044, + 0.013498920947313309, + 0.011979890055954456, + 0.010436812415719032, + 0.012484224513173103, + 0.009398712776601315, + 0.014037510380148888, + 0.0041215927340090275, + 0.008903667330741882, + 0.004104632418602705, + 0.00358275743201375, + 0.006828391924500465, + 0.0055010668002069, + 0.005106030963361263, + 0.005938410758972168 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 8, + 6 + ], + "selected_reward": -0.005248814355581999 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a lone figure wearing a black leather bodysuit?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the figure performing parkour movements across a rain-soaked cityscape at night?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a silver assault rifle visible strapped to the figure's back?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the scene include a stream of green digital code drifting from the sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot a low-angle tracking shot moving with the figure?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.014190834015607834 + ], + "opposite_logprob_tokens": [ + -4.264190673828125 + ], + "logprob_reward": -0.014190834015607834, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.014190834015607834, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1770070642232895, + 0.08148052543401718, + 0.19581151008605957, + 0.1201729029417038, + 0.1332240253686905, + 0.03544120490550995, + 0.05127819627523422, + 0.0918324887752533, + 0.03625413402915001, + 0.07749789953231812 + ], + "credit_map": [ + 0.01769370585680008, + 0.017044980078935623, + 0.03764648362994194, + 0.017483776435256004, + 0.013872336596250534, + 0.03355967253446579, + 0.020593544468283653, + 0.01911257952451706, + 0.01012470480054617, + 0.008217578753829002, + 0.012001259252429008, + 0.010662103071808815, + 0.006113738287240267, + 0.012113404460251331, + 0.009277710691094398, + 0.012970034033060074, + 0.014935356564819813, + 0.017011502757668495, + 0.02098877541720867, + 0.04416102170944214, + 0.017837082967162132, + 0.036827169358730316, + 0.026109803467988968, + 0.017940834164619446, + 0.00987258180975914, + 0.013750637881457806, + 0.014009161852300167, + 0.015579665079712868, + 0.016052406281232834, + 0.021977540105581284, + 0.013413289561867714, + 0.015517630614340305, + 0.015151442028582096, + 0.014036726206541061, + 0.014455159194767475, + 0.014798549935221672, + 0.017501162365078926, + 0.010238057002425194, + 0.022779513150453568, + 0.024263424798846245, + 0.0065288967452943325, + 0.0037469894159585238, + 0.0032138600945472717, + 0.0031341270077973604, + 0.0033680968917906284, + 0.004692217335104942, + 0.005833305884152651, + 0.004923713859170675, + 0.012279360555112362, + 0.007116793189197779, + 0.0064616999588906765, + 0.007089199032634497, + 0.0057755205780267715, + 0.0032385645899921656, + 0.005091110244393349, + 0.0042259530164301395, + 0.005770621355623007, + 0.004701036028563976, + 0.029201528057456017, + 0.007899937219917774, + 0.005763500463217497, + 0.02488860860466957, + 0.006908773444592953, + 0.006698483135551214, + 0.0032896441407501698, + 0.005457838997244835, + 0.0034874475095421076, + 0.0035486205015331507, + 0.004824781324714422, + 0.004450903739780188, + 0.005066289566457272, + 0.006128611508756876, + 0.006105302833020687, + 0.006210213992744684, + 0.0022442005574703217, + 0.0037074091378599405, + 0.020708605647087097, + 0.013419696129858494, + 0.01114554051309824, + 0.013956944458186626 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 0, + 4 + ], + "selected_reward": -0.014190834015607834 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000027/group_12/index_tensor([3791, 3791])_group_12_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000027/group_12/index_tensor([3791, 3791])_group_12_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..4116a078d32d8ca357a8c4b03c25667d88905038 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000027/group_12/index_tensor([3791, 3791])_group_12_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "shot shot from a level perspective, the camera positioned in the center of a narrow, deep alley, begins a steady, slow horizontal rotation. As the camera pans, the mottled, high walls on either side slowly flow across the frame. The walls are made of damp, dark masonry, large swathes of plaster peeling away, revealing the rough texture beneath. Dark green moss creeps through corners and cracks, and tangled strands of electrical wire dangle like black vines. Underfoot, the uneven, slippery cobblestones lie wet, with a few puddles reflecting the lone light source in the distance. At the end of the alley, a lone street lamp casts a faint, warm yellow glow, piercing the dense darkness, casting long, twisting shadows on the ground and bringing a glimmer of light to the claustrophobic space. The entire rotation creates an atmosphere of exploration, suspense, and loneliness, as if uncovering the secrets of a forgotten corner of the city.", + "seed": 20497, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000027/group_12/index_3791_seed_20497_dp_24_rank_48_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 12, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.0517170125618577, + "vlm_logprob": -0.0517170125618577, + "vlm_yes_no_margin": 3.975, + "vlm_token_credit": -0.0517170125618577, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned in the center of a narrow, deep alley?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera perform a steady, slow horizontal rotation?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the walls made of damp, dark masonry with peeling plaster revealing rough texture?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is there a lone street lamp at the end of the alley casting a faint, warm yellow glow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are the cobblestones underfoot uneven, slippery, and wet with puddles reflecting light?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.029960302636027336 + ], + "opposite_logprob_tokens": [ + -5.6549601554870605 + ], + "logprob_reward": -0.029960302636027336, + "yes_no_margin_reward": 5.625, + "token_credit_reward": -0.029960302636027336, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1589839905500412, + 0.06655523180961609, + 0.048760637640953064, + 0.09249551594257355, + 0.13296738266944885, + 0.08377843350172043, + 0.0725933313369751, + 0.1413286179304123, + 0.07941122353076935, + 0.12312562763690948 + ], + "credit_map": [ + 0.01149232778698206, + 0.01648341491818428, + 0.024498794227838516, + 0.02657332271337509, + 0.012596847489476204, + 0.021712005138397217, + 0.0210464708507061, + 0.024580806493759155, + 0.0060446313582360744, + 0.009680651128292084, + 0.008714079856872559, + 0.010957321152091026, + 0.004832116886973381, + 0.0066618528217077255, + 0.009893205016851425, + 0.009771374985575676, + 0.004973732866346836, + 0.00535307452082634, + 0.004909476265311241, + 0.0038073512259870768, + 0.0043558948673307896, + 0.01349467784166336, + 0.004661201033741236, + 0.007205228321254253, + 0.011808021925389767, + 0.011722509749233723, + 0.013017688877880573, + 0.01097403559833765, + 0.009411393664777279, + 0.01187688484787941, + 0.012073059566318989, + 0.01161192636936903, + 0.014890456572175026, + 0.017519688233733177, + 0.02189120091497898, + 0.012903331778943539, + 0.014234158210456371, + 0.01881878264248371, + 0.01127881184220314, + 0.021430950611829758, + 0.008183484897017479, + 0.010729086585342884, + 0.011598685756325722, + 0.010920642875134945, + 0.00905162375420332, + 0.014033142477273941, + 0.0071945954114198685, + 0.012067173607647419, + 0.0072688087821006775, + 0.005191021133214235, + 0.012250949628651142, + 0.010341011919081211, + 0.00766284903511405, + 0.01034562662243843, + 0.009612890891730785, + 0.009920173324644566, + 0.018015995621681213, + 0.012157469056546688, + 0.025135746225714684, + 0.016274085268378258, + 0.015706947073340416, + 0.019203245639801025, + 0.021843375638127327, + 0.012991750612854958, + 0.0057232994586229324, + 0.004236155655235052, + 0.01802019774913788, + 0.01091604121029377, + 0.010077478364109993, + 0.005972269456833601, + 0.011444569565355778, + 0.013021205551922321, + 0.007953423075377941, + 0.006416281685233116, + 0.029832037165760994, + 0.012921616435050964, + 0.010918020270764828, + 0.010730080306529999, + 0.013253926299512386, + 0.03110024333000183 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 7, + 4 + ], + "selected_reward": -0.029960302636027336 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned in the center of a narrow, deep alley?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera perform a steady, slow horizontal rotation?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the walls made of damp, dark masonry with peeling plaster revealing rough texture?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is there a lone street lamp at the end of the alley casting a faint, warm yellow glow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are the cobblestones underfoot uneven, slippery, and wet with puddles reflecting light?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.14283248782157898 + ], + "opposite_logprob_tokens": [ + -2.0178325176239014 + ], + "logprob_reward": -0.14283248782157898, + "yes_no_margin_reward": 1.875, + "token_credit_reward": -0.14283248782157898, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14387379586696625, + 0.07046432048082352, + 0.03407049924135208, + 0.14933894574642181, + 0.13798364996910095, + 0.16368326544761658, + 0.07842375338077545, + 0.06837582588195801, + 0.08066964894533157, + 0.07311631739139557 + ], + "credit_map": [ + 0.01228402554988861, + 0.01796307973563671, + 0.01982085406780243, + 0.022195206955075264, + 0.010915342718362808, + 0.013060026802122593, + 0.022388437762856483, + 0.02524680644273758, + 0.0050276839174330235, + 0.007729048375040293, + 0.010765732266008854, + 0.012790440581738949, + 0.00460432656109333, + 0.008977700024843216, + 0.010708488523960114, + 0.009860901162028313, + 0.003466175403445959, + 0.0049918885342776775, + 0.004758304916322231, + 0.0038373100105673075, + 0.0036758524365723133, + 0.005065875127911568, + 0.0036637233570218086, + 0.00461136968806386, + 0.01748250238597393, + 0.018268657848238945, + 0.023155704140663147, + 0.01704252138733864, + 0.019228404387831688, + 0.019982216879725456, + 0.014754585921764374, + 0.019424356520175934, + 0.015390003100037575, + 0.01566144824028015, + 0.022069254890084267, + 0.014547913335263729, + 0.014319872483611107, + 0.0210382342338562, + 0.014671484008431435, + 0.02028544619679451, + 0.019783541560173035, + 0.009941273368895054, + 0.02353748306632042, + 0.016743946820497513, + 0.01670878939330578, + 0.023169122636318207, + 0.020698361098766327, + 0.03310074284672737, + 0.01026905421167612, + 0.005216754041612148, + 0.012968881987035275, + 0.00967995822429657, + 0.008050106465816498, + 0.009978536516427994, + 0.010543562471866608, + 0.011716893874108791, + 0.009806368499994278, + 0.00452288705855608, + 0.009750930592417717, + 0.010702010244131088, + 0.008950283750891685, + 0.011502798646688461, + 0.007726560812443495, + 0.005413981154561043, + 0.005221084225922823, + 0.007183869369328022, + 0.01151453610509634, + 0.016823329031467438, + 0.007877925410866737, + 0.007134636864066124, + 0.017675722017884254, + 0.007238544523715973, + 0.004691451322287321, + 0.003653915598988533, + 0.017599692568182945, + 0.00879056565463543, + 0.0058660972863435745, + 0.005906505044549704, + 0.009204854257404804, + 0.017403239384293556 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 3, + 0 + ], + "selected_reward": -0.14283248782157898 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned in the center of a narrow, deep alley?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera perform a steady, slow horizontal rotation?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the walls made of damp, dark masonry with peeling plaster revealing rough texture?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is there a lone street lamp at the end of the alley casting a faint, warm yellow glow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are the cobblestones underfoot uneven, slippery, and wet with puddles reflecting light?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.00979462917894125 + ], + "opposite_logprob_tokens": [ + -4.63479471206665 + ], + "logprob_reward": -0.00979462917894125, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.00979462917894125, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11439480632543564, + 0.07861391454935074, + 0.10800173878669739, + 0.11531028151512146, + 0.09928535670042038, + 0.05355582386255264, + 0.070051409304142, + 0.19572031497955322, + 0.04520030692219734, + 0.11986605823040009 + ], + "credit_map": [ + 0.011269600130617619, + 0.013475311920046806, + 0.018099727109074593, + 0.01637629233300686, + 0.007609354332089424, + 0.013621287420392036, + 0.01200010348111391, + 0.02194312959909439, + 0.0059615494683384895, + 0.01085511315613985, + 0.0105056744068861, + 0.014170118607580662, + 0.005645750090479851, + 0.010811789892613888, + 0.009047071449458599, + 0.01161684188991785, + 0.017105743288993835, + 0.009081178344786167, + 0.0242738276720047, + 0.012449316680431366, + 0.007107917219400406, + 0.012501289136707783, + 0.011225700378417969, + 0.01425675954669714, + 0.00964414607733488, + 0.012582169845700264, + 0.015252809040248394, + 0.015290903858840466, + 0.012523382902145386, + 0.015651868656277657, + 0.01705041341483593, + 0.017314575612545013, + 0.009654400870203972, + 0.012129060924053192, + 0.015090379863977432, + 0.012322808615863323, + 0.009977578185498714, + 0.013193409889936447, + 0.010156841948628426, + 0.016760867089033127, + 0.006100155878812075, + 0.007399373687803745, + 0.006313970312476158, + 0.008037851192057133, + 0.00686620594933629, + 0.007177154067903757, + 0.00560259073972702, + 0.00605852622538805, + 0.008899313397705555, + 0.005782769527286291, + 0.010301461443305016, + 0.008005904965102673, + 0.008834070526063442, + 0.009847980923950672, + 0.011310882866382599, + 0.007069027982652187, + 0.02470180206000805, + 0.010600490495562553, + 0.031216947361826897, + 0.02582535147666931, + 0.02230343595147133, + 0.028199225664138794, + 0.03058132529258728, + 0.022291742265224457, + 0.003237854689359665, + 0.003276645205914974, + 0.008632835932075977, + 0.007995151914656162, + 0.004211218096315861, + 0.0039823018014431, + 0.008531426079571247, + 0.005332875531166792, + 0.005830397829413414, + 0.00371731398627162, + 0.038722362369298935, + 0.00679625291377306, + 0.009585104882717133, + 0.008458338677883148, + 0.006866231095045805, + 0.03989005088806152 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 9, + 3 + ], + "selected_reward": -0.00979462917894125 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned in the center of a narrow, deep alley?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera perform a steady, slow horizontal rotation?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the walls made of damp, dark masonry with peeling plaster revealing rough texture?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is there a lone street lamp at the end of the alley casting a faint, warm yellow glow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are the cobblestones underfoot uneven, slippery, and wet with puddles reflecting light?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005950592458248138 + ], + "opposite_logprob_tokens": [ + -5.130950450897217 + ], + "logprob_reward": -0.005950592458248138, + "yes_no_margin_reward": 5.125, + "token_credit_reward": -0.005950592458248138, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11111842095851898, + 0.1397743970155716, + 0.20418086647987366, + 0.09900270402431488, + 0.05094823241233826, + 0.055071841925382614, + 0.07167854905128479, + 0.12174804508686066, + 0.030484052374958992, + 0.11599289625883102 + ], + "credit_map": [ + 0.007228484842926264, + 0.021055687218904495, + 0.013234805315732956, + 0.018744472414255142, + 0.008382769301533699, + 0.012743495404720306, + 0.01408405415713787, + 0.01564464531838894, + 0.015259024687111378, + 0.017972303554415703, + 0.02202625945210457, + 0.021903961896896362, + 0.010267903096973896, + 0.01971088908612728, + 0.01330339815467596, + 0.01933065429329872, + 0.01528975646942854, + 0.026832573115825653, + 0.02116960659623146, + 0.025674838572740555, + 0.016090793535113335, + 0.04987257346510887, + 0.017916634678840637, + 0.03133409842848778, + 0.011527236551046371, + 0.010215302929282188, + 0.018187416717410088, + 0.010068275034427643, + 0.00700623681768775, + 0.0170854814350605, + 0.009624183177947998, + 0.015288562513887882, + 0.005942677613347769, + 0.004775318782776594, + 0.006985231768339872, + 0.005741622764617205, + 0.007298686541616917, + 0.006260102614760399, + 0.006013454869389534, + 0.007931139320135117, + 0.006632233038544655, + 0.006889615673571825, + 0.0070802755653858185, + 0.005605973303318024, + 0.005648494698107243, + 0.00722750136628747, + 0.005569600500166416, + 0.010418147780001163, + 0.008392179384827614, + 0.005026654805988073, + 0.009771515615284443, + 0.010625416412949562, + 0.008031620644032955, + 0.012071734294295311, + 0.010113054886460304, + 0.007646370679140091, + 0.013858508318662643, + 0.00940980389714241, + 0.01761098951101303, + 0.016137439757585526, + 0.013972533866763115, + 0.01993934065103531, + 0.01661544479429722, + 0.014203990809619427, + 0.0024176486767828465, + 0.0018301383825019002, + 0.006103404797613621, + 0.004253132734447718, + 0.004338248632848263, + 0.0023946904111653566, + 0.00492096645757556, + 0.004225822165608406, + 0.008105236105620861, + 0.006558624561876059, + 0.030005773529410362, + 0.007583734579384327, + 0.014675911515951157, + 0.010582739487290382, + 0.009466235525906086, + 0.029014647006988525 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 1, + 7 + ], + "selected_reward": -0.005950592458248138 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned in the center of a narrow, deep alley?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera perform a steady, slow horizontal rotation?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the walls made of damp, dark masonry with peeling plaster revealing rough texture?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is there a lone street lamp at the end of the alley casting a faint, warm yellow glow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are the cobblestones underfoot uneven, slippery, and wet with puddles reflecting light?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0700470507144928 + ], + "opposite_logprob_tokens": [ + -2.69504714012146 + ], + "logprob_reward": -0.0700470507144928, + "yes_no_margin_reward": 2.625, + "token_credit_reward": -0.0700470507144928, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.16574521362781525, + 0.08081123232841492, + 0.07955087721347809, + 0.13719390332698822, + 0.07229720056056976, + 0.026470432057976723, + 0.06011718139052391, + 0.26868075132369995, + 0.048580434173345566, + 0.060552723705768585 + ], + "credit_map": [ + 0.011040027253329754, + 0.021718420088291168, + 0.022441362962126732, + 0.027810053899884224, + 0.014841340482234955, + 0.01669108308851719, + 0.024652816355228424, + 0.026550104841589928, + 0.006193391513079405, + 0.009769678115844727, + 0.010174664668738842, + 0.016438746824860573, + 0.005022663157433271, + 0.009105704724788666, + 0.012735738418996334, + 0.011370641179382801, + 0.010859301313757896, + 0.009414253756403923, + 0.009592083282768726, + 0.007283332291990519, + 0.0066432952880859375, + 0.015629906207323074, + 0.007790970616042614, + 0.012337728403508663, + 0.011666006408631802, + 0.014294915832579136, + 0.02103988453745842, + 0.017798569053411484, + 0.016517356038093567, + 0.0208884309977293, + 0.01811850257217884, + 0.01687024161219597, + 0.008203493431210518, + 0.008623740635812283, + 0.010370834730565548, + 0.009367936290800571, + 0.007176840212196112, + 0.008755658753216267, + 0.007276165299117565, + 0.012522532604634762, + 0.0025993306189775467, + 0.002297693397849798, + 0.003380819456651807, + 0.0030249219853430986, + 0.003984135575592518, + 0.0045370436273515224, + 0.0037517102900892496, + 0.0028947775717824697, + 0.00587515439838171, + 0.00517024053260684, + 0.00947718508541584, + 0.009422102943062782, + 0.007352507207542658, + 0.006216447334736586, + 0.010618477128446102, + 0.00598506536334753, + 0.03307685628533363, + 0.016955778002738953, + 0.04343394562602043, + 0.03352504223585129, + 0.030865248292684555, + 0.04098043590784073, + 0.037268057465553284, + 0.03257540613412857, + 0.003954279236495495, + 0.0033420405816286802, + 0.0067549641244113445, + 0.011304751969873905, + 0.005073862615972757, + 0.0038586498703807592, + 0.009509244002401829, + 0.004782639443874359, + 0.003647375153377652, + 0.0029355462174862623, + 0.01711747795343399, + 0.005750420968979597, + 0.004697501193732023, + 0.004688948392868042, + 0.0051330626010894775, + 0.016582390293478966 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 0, + 3 + ], + "selected_reward": -0.0700470507144928 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "shot shot from a level perspective, the camera positioned in the center of a narrow, deep alley, begins a steady, slow horizontal rotation. As the camera pans, the mottled, high walls on either side slowly flow across the frame. The walls are made of damp, dark masonry, large swathes of plaster peeling away, revealing the rough texture beneath. Dark green moss creeps through corners and cracks, and tangled strands of electrical wire dangle like black vines. Underfoot, the uneven, slippery cobblestones lie wet, with a few puddles reflecting the lone light source in the distance. At the end of the alley, a lone street lamp casts a faint, warm yellow glow, piercing the dense darkness, casting long, twisting shadows on the ground and bringing a glimmer of light to the claustrophobic space. The entire rotation creates an atmosphere of exploration, suspense, and loneliness, as if uncovering the secrets of a forgotten corner of the city.", + "seed": 20498, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000027/group_12/index_3791_seed_20498_dp_24_rank_48_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 12, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.0457074074074626, + "vlm_logprob": -0.0457074074074626, + "vlm_yes_no_margin": 3.6, + "vlm_token_credit": -0.0457074074074626, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned in the center of a narrow, deep alley?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera perform a steady, slow horizontal rotation?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the walls made of damp, dark masonry with peeling plaster revealing rough texture?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is there a lone street lamp at the end of the alley casting a faint, warm yellow glow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are the cobblestones underfoot uneven, slippery, and wet with puddles reflecting light?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.042329609394073486 + ], + "opposite_logprob_tokens": [ + -4.167329788208008 + ], + "logprob_reward": -0.042329609394073486, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.042329609394073486, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13162526488304138, + 0.03635699674487114, + 0.04786933213472366, + 0.06004273518919945, + 0.08479851484298706, + 0.05935947969555855, + 0.0779719203710556, + 0.045326169580221176, + 0.3425271809101105, + 0.11412238329648972 + ], + "credit_map": [ + 0.010990181006491184, + 0.01824389025568962, + 0.012510336935520172, + 0.01836397871375084, + 0.009592707268893719, + 0.021459897980093956, + 0.017141470685601234, + 0.02332279086112976, + 0.0028210419695824385, + 0.0033719148486852646, + 0.004525997210294008, + 0.00414921622723341, + 0.0030496991239488125, + 0.005434730555862188, + 0.0070772296749055386, + 0.005927166435867548, + 0.004054599441587925, + 0.0036205414216965437, + 0.006742245517671108, + 0.0081906383857131, + 0.004154144320636988, + 0.007358040660619736, + 0.007284570951014757, + 0.006464552134275436, + 0.005800897721201181, + 0.01534079760313034, + 0.004803038667887449, + 0.009255164302885532, + 0.006594613194465637, + 0.004668017849326134, + 0.006852220743894577, + 0.006727979984134436, + 0.006192095577716827, + 0.009340494871139526, + 0.012798481620848179, + 0.01551695354282856, + 0.00913228839635849, + 0.007701514288783073, + 0.01409914344549179, + 0.010017543099820614, + 0.00625956803560257, + 0.009437873028218746, + 0.007029601838439703, + 0.011987368576228619, + 0.006654005963355303, + 0.006067429203540087, + 0.005797666497528553, + 0.006125964689999819, + 0.008062679320573807, + 0.008603098802268505, + 0.019060062244534492, + 0.011057941243052483, + 0.005341910757124424, + 0.009024698287248611, + 0.011148013174533844, + 0.005673512350767851, + 0.0036541512235999107, + 0.0033555657137185335, + 0.007655565161257982, + 0.006446688435971737, + 0.00562499463558197, + 0.007761670276522636, + 0.006128391716629267, + 0.004699144046753645, + 0.017573731020092964, + 0.012139933183789253, + 0.1137499287724495, + 0.020527727901935577, + 0.01763126254081726, + 0.03395812585949898, + 0.024909108877182007, + 0.10203736275434494, + 0.008573194965720177, + 0.006079657468944788, + 0.02782447449862957, + 0.013784550130367279, + 0.0065492200665175915, + 0.010710838250815868, + 0.013388918712735176, + 0.02721153385937214 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 0, + 9 + ], + "selected_reward": -0.042329609394073486 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned in the center of a narrow, deep alley?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera perform a steady, slow horizontal rotation?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the walls made of damp, dark masonry with peeling plaster revealing rough texture?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is there a lone street lamp at the end of the alley casting a faint, warm yellow glow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are the cobblestones underfoot uneven, slippery, and wet with puddles reflecting light?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.08906880021095276 + ], + "opposite_logprob_tokens": [ + -2.46406888961792 + ], + "logprob_reward": -0.08906880021095276, + "yes_no_margin_reward": 2.375, + "token_credit_reward": -0.08906880021095276, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13693135976791382, + 0.053328514099121094, + 0.04024406149983406, + 0.07616686075925827, + 0.11034010350704193, + 0.05454690009355545, + 0.11634338647127151, + 0.044680848717689514, + 0.17965996265411377, + 0.18775799870491028 + ], + "credit_map": [ + 0.011919042095541954, + 0.024098578840494156, + 0.01607520505785942, + 0.01221555843949318, + 0.010361908935010433, + 0.022367767989635468, + 0.021123385056853294, + 0.018769914284348488, + 0.003880177391692996, + 0.0064587946981191635, + 0.006577917840331793, + 0.005881438963115215, + 0.004884267691522837, + 0.008539652451872826, + 0.008505667559802532, + 0.008600602857768536, + 0.004503889009356499, + 0.0034138772170990705, + 0.0057945516891777515, + 0.007601010613143444, + 0.0039521618746221066, + 0.005213565658777952, + 0.004470564890652895, + 0.0052944389171898365, + 0.006865416653454304, + 0.012156831100583076, + 0.01158444955945015, + 0.011150448583066463, + 0.007636775728315115, + 0.009671388193964958, + 0.009025707840919495, + 0.008075844496488571, + 0.009969234466552734, + 0.011336993426084518, + 0.01821807026863098, + 0.01443918701261282, + 0.014395319856703281, + 0.011349152773618698, + 0.015713205561041832, + 0.014918935485184193, + 0.005777730140835047, + 0.008337479084730148, + 0.007821234874427319, + 0.008025850169360638, + 0.008411052636802197, + 0.004894303157925606, + 0.005393593572080135, + 0.005885652732104063, + 0.011332817375659943, + 0.013302866369485855, + 0.020756904035806656, + 0.012907414697110653, + 0.011152004823088646, + 0.01898796856403351, + 0.014722988940775394, + 0.01318042166531086, + 0.003107862314209342, + 0.003369647543877363, + 0.010082470253109932, + 0.006506310775876045, + 0.0037499042227864265, + 0.007150116842240095, + 0.005063417833298445, + 0.005651117768138647, + 0.009780443273484707, + 0.006982775405049324, + 0.06151800975203514, + 0.009912153705954552, + 0.007684104144573212, + 0.017882488667964935, + 0.011989830061793327, + 0.05391016602516174, + 0.015846241265535355, + 0.015984177589416504, + 0.030550885945558548, + 0.03986877575516701, + 0.011160976253449917, + 0.01664407178759575, + 0.028488291427493095, + 0.02921457774937153 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 8, + 0 + ], + "selected_reward": -0.08906880021095276 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned in the center of a narrow, deep alley?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera perform a steady, slow horizontal rotation?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the walls made of damp, dark masonry with peeling plaster revealing rough texture?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is there a lone street lamp at the end of the alley casting a faint, warm yellow glow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are the cobblestones underfoot uneven, slippery, and wet with puddles reflecting light?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005260198377072811 + ], + "opposite_logprob_tokens": [ + -5.255259990692139 + ], + "logprob_reward": -0.005260198377072811, + "yes_no_margin_reward": 5.25, + "token_credit_reward": -0.005260198377072811, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.05458449199795723, + 0.02299126610159874, + 0.026561280712485313, + 0.0813383087515831, + 0.07839687913656235, + 0.022472631186246872, + 0.03214128687977791, + 0.03074866719543934, + 0.616731584072113, + 0.03403370827436447 + ], + "credit_map": [ + 0.005191595759242773, + 0.004061264451593161, + 0.006223930045962334, + 0.004305538721382618, + 0.0031401682645082474, + 0.014649714343249798, + 0.0044441744685173035, + 0.012568099424242973, + 0.0016566134290769696, + 0.003026873106136918, + 0.002351165981963277, + 0.0026133430656045675, + 0.0025775106623768806, + 0.0031147419940680265, + 0.0038559904787689447, + 0.003795025171712041, + 0.00403590640053153, + 0.0030296791810542345, + 0.0031973812729120255, + 0.0029656027909368277, + 0.002910470589995384, + 0.003520264057442546, + 0.002922322601079941, + 0.003979649860411882, + 0.008310494013130665, + 0.015389478765428066, + 0.011920087039470673, + 0.010157514363527298, + 0.006567027419805527, + 0.00667837867513299, + 0.011789029464125633, + 0.010526294820010662, + 0.006282626185566187, + 0.008210284635424614, + 0.014138012193143368, + 0.012101502157747746, + 0.008909907191991806, + 0.0074546621181070805, + 0.011743622832000256, + 0.009556254372000694, + 0.001710590673610568, + 0.0029818976763635874, + 0.003486757865175605, + 0.003848021849989891, + 0.003211963688954711, + 0.0027823634445667267, + 0.002623680979013443, + 0.0018273534951731563, + 0.004105949308723211, + 0.002485863398760557, + 0.007158215157687664, + 0.003458054969087243, + 0.0033481845166534185, + 0.004051658790558577, + 0.004295375198125839, + 0.0032379815820604563, + 0.0022943217772990465, + 0.002240401227027178, + 0.005932101048529148, + 0.004180791787803173, + 0.0036504112649708986, + 0.005107228644192219, + 0.0037020707968622446, + 0.003641339484602213, + 0.03497016802430153, + 0.015161451883614063, + 0.22922880947589874, + 0.029893286526203156, + 0.017339518293738365, + 0.0593358613550663, + 0.023724732920527458, + 0.20707769691944122, + 0.0024086814373731613, + 0.0019873762503266335, + 0.008494023233652115, + 0.003181891515851021, + 0.0024398311506956816, + 0.00294332392513752, + 0.0033664254005998373, + 0.009212150238454342 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 3, + 4 + ], + "selected_reward": -0.005260198377072811 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned in the center of a narrow, deep alley?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera perform a steady, slow horizontal rotation?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the walls made of damp, dark masonry with peeling plaster revealing rough texture?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is there a lone street lamp at the end of the alley casting a faint, warm yellow glow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are the cobblestones underfoot uneven, slippery, and wet with puddles reflecting light?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.029802024364471436 + ], + "opposite_logprob_tokens": [ + -3.529802083969116 + ], + "logprob_reward": -0.029802024364471436, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.029802024364471436, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11044835299253464, + 0.03543667495250702, + 0.041614267975091934, + 0.04933781549334526, + 0.041283849626779556, + 0.07307641953229904, + 0.19327670335769653, + 0.073540598154068, + 0.3325459659099579, + 0.049439359456300735 + ], + "credit_map": [ + 0.00873017031699419, + 0.0139596713706851, + 0.014348786324262619, + 0.017118871212005615, + 0.010650763288140297, + 0.015447060577571392, + 0.020516034215688705, + 0.009677005000412464, + 0.002178475959226489, + 0.002991661662235856, + 0.005372316110879183, + 0.003935926593840122, + 0.002352943643927574, + 0.006637747399508953, + 0.007013609167188406, + 0.004953994881361723, + 0.009990716353058815, + 0.0032884085085242987, + 0.004574073012918234, + 0.00528732780367136, + 0.0042675998993217945, + 0.00490991398692131, + 0.004573300946503878, + 0.0047229272313416, + 0.003953327890485525, + 0.006429014727473259, + 0.005365556105971336, + 0.009618032723665237, + 0.005870935972779989, + 0.0059436894953250885, + 0.006506401114165783, + 0.005650858860462904, + 0.0054723541252315044, + 0.004266954958438873, + 0.005047916434705257, + 0.005060194991528988, + 0.0038614983204752207, + 0.004371647257357836, + 0.005825292319059372, + 0.007377991918474436, + 0.007169700227677822, + 0.009783994406461716, + 0.01304844580590725, + 0.011610801331698895, + 0.013629447668790817, + 0.006304149515926838, + 0.007157078478485346, + 0.004372805822640657, + 0.022989602759480476, + 0.019979624077677727, + 0.04452241584658623, + 0.025799786671996117, + 0.018117984756827354, + 0.019565755501389503, + 0.026082491502165794, + 0.016219034790992737, + 0.005233163945376873, + 0.003923410549759865, + 0.0187278650701046, + 0.010252108797430992, + 0.004943318665027618, + 0.012845891527831554, + 0.009450383484363556, + 0.008164456114172935, + 0.01803191378712654, + 0.010470885783433914, + 0.11679645627737045, + 0.019025355577468872, + 0.01210065372288227, + 0.03186000883579254, + 0.017882339656352997, + 0.10637833178043365, + 0.004289218224585056, + 0.005473271943628788, + 0.006451635621488094, + 0.009408464655280113, + 0.004071322735399008, + 0.005600473843514919, + 0.007317017763853073, + 0.006827954668551683 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 6, + 0 + ], + "selected_reward": -0.029802024364471436 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned in the center of a narrow, deep alley?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera perform a steady, slow horizontal rotation?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the walls made of damp, dark masonry with peeling plaster revealing rough texture?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is there a lone street lamp at the end of the alley casting a faint, warm yellow glow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are the cobblestones underfoot uneven, slippery, and wet with puddles reflecting light?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.06207640469074249 + ], + "opposite_logprob_tokens": [ + -2.8120763301849365 + ], + "logprob_reward": -0.06207640469074249, + "yes_no_margin_reward": 2.75, + "token_credit_reward": -0.06207640469074249, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.019797440618276596, + 0.010931326076388359, + 0.011361724697053432, + 0.02242087945342064, + 0.026233568787574768, + 0.013961253687739372, + 0.04933292418718338, + 0.03414071351289749, + 0.7956698536872864, + 0.016150254756212234 + ], + "credit_map": [ + 0.001467201393097639, + 0.002043170388787985, + 0.0018883576849475503, + 0.0019114279421046376, + 0.001157763646915555, + 0.004919206257909536, + 0.002056130673736334, + 0.004354183096438646, + 0.0008998740231618285, + 0.0008600454893894494, + 0.00114146969281137, + 0.0010360941523686051, + 0.001366969896480441, + 0.0018268170533701777, + 0.001783382031135261, + 0.002016673097386956, + 0.0016125538386404514, + 0.0009807568276301026, + 0.0015325406566262245, + 0.001636603381484747, + 0.0012614381266757846, + 0.001449046772904694, + 0.0013164576375856996, + 0.0015723276883363724, + 0.001967456191778183, + 0.004001005552709103, + 0.0030127177014946938, + 0.0032255405094474554, + 0.002429394982755184, + 0.0027452425565570593, + 0.0025326090399175882, + 0.0025069129187613726, + 0.002186637604609132, + 0.003290181513875723, + 0.003990092780441046, + 0.0047807954251766205, + 0.0030011499766260386, + 0.002467632992193103, + 0.003574542934074998, + 0.0029425364919006824, + 0.0013215809594839811, + 0.0015144981443881989, + 0.0023635425604879856, + 0.0021359813399612904, + 0.0018279022770002484, + 0.00174427789170295, + 0.0016064689261838794, + 0.0014470013557001948, + 0.005760821979492903, + 0.00437890412285924, + 0.01406923495233059, + 0.0046316711232066154, + 0.0039437562227249146, + 0.006818740162998438, + 0.004414346069097519, + 0.005315448623150587, + 0.002530478872358799, + 0.0026887571439146996, + 0.006610754411667585, + 0.004368385300040245, + 0.004500189330428839, + 0.00539590697735548, + 0.003989698365330696, + 0.004056544974446297, + 0.04555550962686539, + 0.020296471193432808, + 0.29381033778190613, + 0.038888901472091675, + 0.023294610902667046, + 0.07705020904541016, + 0.03281603381037712, + 0.26395776867866516, + 0.0014209112850949168, + 0.0012242961674928665, + 0.002105721738189459, + 0.0030528302304446697, + 0.0011534967925399542, + 0.002109949477016926, + 0.0033121458254754543, + 0.001770903472788632 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 6, + 7 + ], + "selected_reward": -0.06207640469074249 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "shot shot from a level perspective, the camera positioned in the center of a narrow, deep alley, begins a steady, slow horizontal rotation. As the camera pans, the mottled, high walls on either side slowly flow across the frame. The walls are made of damp, dark masonry, large swathes of plaster peeling away, revealing the rough texture beneath. Dark green moss creeps through corners and cracks, and tangled strands of electrical wire dangle like black vines. Underfoot, the uneven, slippery cobblestones lie wet, with a few puddles reflecting the lone light source in the distance. At the end of the alley, a lone street lamp casts a faint, warm yellow glow, piercing the dense darkness, casting long, twisting shadows on the ground and bringing a glimmer of light to the claustrophobic space. The entire rotation creates an atmosphere of exploration, suspense, and loneliness, as if uncovering the secrets of a forgotten corner of the city.", + "seed": 20499, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000027/group_12/index_3791_seed_20499_dp_25_rank_50_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 12, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.029221044667065144, + "vlm_logprob": -0.029221044667065144, + "vlm_yes_no_margin": 4.1, + "vlm_token_credit": -0.029221044667065144, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned in the center of a narrow, deep alley?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera perform a steady, slow horizontal rotation?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the walls made of damp, dark masonry with peeling plaster revealing rough texture?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is there a lone street lamp at the end of the alley casting a faint, warm yellow glow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are the cobblestones underfoot uneven, slippery, and wet with puddles reflecting light?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.046475689858198166 + ], + "opposite_logprob_tokens": [ + -3.796475648880005 + ], + "logprob_reward": -0.046475689858198166, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.046475689858198166, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.16979318857192993, + 0.24490025639533997, + 0.07237958908081055, + 0.13453955948352814, + 0.036848362535238266, + 0.04288722574710846, + 0.0423424169421196, + 0.09415342658758163, + 0.07110291719436646, + 0.09105303138494492 + ], + "credit_map": [ + 0.010423746891319752, + 0.017915476113557816, + 0.020567385479807854, + 0.030226824805140495, + 0.01600133627653122, + 0.04066503047943115, + 0.016921820119023323, + 0.017071586102247238, + 0.01091872900724411, + 0.020248446613550186, + 0.014246954582631588, + 0.016731515526771545, + 0.037722039967775345, + 0.07735110074281693, + 0.02123316004872322, + 0.04644833132624626, + 0.006153552792966366, + 0.005433241371065378, + 0.0063260216265916824, + 0.008960914798080921, + 0.011631038039922714, + 0.011849475093185902, + 0.009491383098065853, + 0.01253396924585104, + 0.012549084611237049, + 0.010772706009447575, + 0.020080657675862312, + 0.015733681619167328, + 0.0188482366502285, + 0.015061670914292336, + 0.02459471859037876, + 0.016898823902010918, + 0.004824196919798851, + 0.003377459943294525, + 0.004843249451369047, + 0.003960530739277601, + 0.004218362737447023, + 0.0043424381874501705, + 0.005283655598759651, + 0.005998469889163971, + 0.0037523629143834114, + 0.006261390168219805, + 0.006202477030456066, + 0.004580979701131582, + 0.006542380899190903, + 0.005367367994040251, + 0.005056450143456459, + 0.00512382248416543, + 0.004916582256555557, + 0.004461744800209999, + 0.005568966269493103, + 0.003691716119647026, + 0.005322218872606754, + 0.005211832467466593, + 0.00667177326977253, + 0.00649758568033576, + 0.008393397554755211, + 0.005753615405410528, + 0.019263366237282753, + 0.007637795060873032, + 0.009635659866034985, + 0.010041484609246254, + 0.017104104161262512, + 0.016324007883667946, + 0.009203454479575157, + 0.009036717936396599, + 0.011148268356919289, + 0.007849695160984993, + 0.009026623331010342, + 0.006183980032801628, + 0.009890519082546234, + 0.008763672783970833, + 0.008932000957429409, + 0.012718762271106243, + 0.010967044159770012, + 0.01516351755708456, + 0.008915885351598263, + 0.010279295034706593, + 0.011736039072275162, + 0.012340500019490719 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 3 + ], + "selected_reward": -0.046475689858198166 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned in the center of a narrow, deep alley?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera perform a steady, slow horizontal rotation?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the walls made of damp, dark masonry with peeling plaster revealing rough texture?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is there a lone street lamp at the end of the alley casting a faint, warm yellow glow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are the cobblestones underfoot uneven, slippery, and wet with puddles reflecting light?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07005304843187332 + ], + "opposite_logprob_tokens": [ + -2.6950531005859375 + ], + "logprob_reward": -0.07005304843187332, + "yes_no_margin_reward": 2.625, + "token_credit_reward": -0.07005304843187332, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.16475024819374084, + 0.1910819411277771, + 0.09659609198570251, + 0.06626468151807785, + 0.07963483035564423, + 0.08527811616659164, + 0.04027628153562546, + 0.04499324783682823, + 0.10927480459213257, + 0.12184978276491165 + ], + "credit_map": [ + 0.014963622204959393, + 0.01725519262254238, + 0.022440040484070778, + 0.024626825004816055, + 0.019282745197415352, + 0.02453656494617462, + 0.013058504089713097, + 0.02858675643801689, + 0.010188231244683266, + 0.014438722282648087, + 0.01750854402780533, + 0.013206186704337597, + 0.02736975997686386, + 0.04446985945105553, + 0.03277070075273514, + 0.03112993761897087, + 0.009085164405405521, + 0.012688963674008846, + 0.00980288814753294, + 0.018377603963017464, + 0.013530422002077103, + 0.013022406026721, + 0.011205045506358147, + 0.00888359546661377, + 0.008027437143027782, + 0.007359397131949663, + 0.009663207456469536, + 0.007374683860689402, + 0.00823906995356083, + 0.006232419982552528, + 0.011656868271529675, + 0.007711595855653286, + 0.0060394820757210255, + 0.0066252523101866245, + 0.010726097039878368, + 0.011619895696640015, + 0.009118433110415936, + 0.007505429442971945, + 0.013990902341902256, + 0.014009341597557068, + 0.00705978088080883, + 0.014508096501231194, + 0.006708886008709669, + 0.01228417083621025, + 0.015076016075909138, + 0.007198070175945759, + 0.011299988254904747, + 0.011143106035888195, + 0.004629587754607201, + 0.004308085888624191, + 0.004330675583332777, + 0.004424276761710644, + 0.00433498527854681, + 0.004902200307697058, + 0.006112113129347563, + 0.007234358694404364, + 0.004488172009587288, + 0.005254371557384729, + 0.006681140046566725, + 0.005748071242123842, + 0.0058344039134681225, + 0.004710712004452944, + 0.007727646268904209, + 0.004548731260001659, + 0.010986840352416039, + 0.0117306774482131, + 0.015136892907321453, + 0.019652068614959717, + 0.010881385765969753, + 0.010571558959782124, + 0.015716876834630966, + 0.01459850650280714, + 0.011275616474449635, + 0.01824507862329483, + 0.01824161596596241, + 0.01491039339452982, + 0.010747219435870647, + 0.014358761720359325, + 0.01472693681716919, + 0.019344154745340347 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 9 + ], + "selected_reward": -0.07005304843187332 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned in the center of a narrow, deep alley?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera perform a steady, slow horizontal rotation?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the walls made of damp, dark masonry with peeling plaster revealing rough texture?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is there a lone street lamp at the end of the alley casting a faint, warm yellow glow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are the cobblestones underfoot uneven, slippery, and wet with puddles reflecting light?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0067512271925807 + ], + "opposite_logprob_tokens": [ + -5.00675106048584 + ], + "logprob_reward": -0.0067512271925807, + "yes_no_margin_reward": 5.0, + "token_credit_reward": -0.0067512271925807, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12330297380685806, + 0.10552294552326202, + 0.08159826695919037, + 0.10025572776794434, + 0.053215041756629944, + 0.09217134118080139, + 0.07046770304441452, + 0.09180345386266708, + 0.14778147637844086, + 0.13388103246688843 + ], + "credit_map": [ + 0.009339986369013786, + 0.011955179274082184, + 0.013330000452697277, + 0.02671639248728752, + 0.016714761033654213, + 0.017934536561369896, + 0.011864669620990753, + 0.015447452664375305, + 0.0062707834877073765, + 0.010418192483484745, + 0.014729773625731468, + 0.011133354157209396, + 0.013243832625448704, + 0.013878129422664642, + 0.019711069762706757, + 0.01613781414926052, + 0.007020177785307169, + 0.007132755126804113, + 0.006718624383211136, + 0.01020796224474907, + 0.013592507690191269, + 0.015594889409840107, + 0.008928956463932991, + 0.012402393855154514, + 0.010149234905838966, + 0.007783157750964165, + 0.012096717022359371, + 0.013705838471651077, + 0.015288189984858036, + 0.010331913828849792, + 0.019861266016960144, + 0.011039410717785358, + 0.005887338891625404, + 0.005225725471973419, + 0.008260956034064293, + 0.0062514664605259895, + 0.0052920151501894, + 0.00661898497492075, + 0.007228450383991003, + 0.008450103923678398, + 0.010238678194582462, + 0.012736205011606216, + 0.014021562412381172, + 0.011720489710569382, + 0.012796655297279358, + 0.010830413550138474, + 0.0095425546169281, + 0.010284778662025928, + 0.005930852610617876, + 0.01010600384324789, + 0.006558137480169535, + 0.007708916440606117, + 0.008973486721515656, + 0.006918986327946186, + 0.013734783977270126, + 0.010536539368331432, + 0.0072290473617613316, + 0.007102873641997576, + 0.020923960953950882, + 0.0074396440759301186, + 0.009127899073064327, + 0.009802678599953651, + 0.012670349329710007, + 0.017506998032331467, + 0.01623895950615406, + 0.017072424292564392, + 0.020382920280098915, + 0.02538815326988697, + 0.016823476180434227, + 0.015181850641965866, + 0.023509757593274117, + 0.013183943927288055, + 0.0107749467715621, + 0.01962425373494625, + 0.015099091455340385, + 0.024969229474663734, + 0.016710123047232628, + 0.015911994501948357, + 0.016255255788564682, + 0.014536149799823761 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 9, + 0 + ], + "selected_reward": -0.0067512271925807 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned in the center of a narrow, deep alley?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera perform a steady, slow horizontal rotation?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the walls made of damp, dark masonry with peeling plaster revealing rough texture?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is there a lone street lamp at the end of the alley casting a faint, warm yellow glow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are the cobblestones underfoot uneven, slippery, and wet with puddles reflecting light?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.006741991266608238 + ], + "opposite_logprob_tokens": [ + -5.006742000579834 + ], + "logprob_reward": -0.006741991266608238, + "yes_no_margin_reward": 5.0, + "token_credit_reward": -0.006741991266608238, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.19244760274887085, + 0.10354200005531311, + 0.047922566533088684, + 0.051747169345617294, + 0.046569496393203735, + 0.08727089315652847, + 0.08448149263858795, + 0.15053479373455048, + 0.18359965085983276, + 0.05188438296318054 + ], + "credit_map": [ + 0.015549951232969761, + 0.021493496373295784, + 0.027277270331978798, + 0.03473188728094101, + 0.02177422307431698, + 0.026538416743278503, + 0.01788162812590599, + 0.027200734242796898, + 0.0058868457563221455, + 0.007747096009552479, + 0.015391421504318714, + 0.008433124050498009, + 0.01294118445366621, + 0.022169509902596474, + 0.010990823619067669, + 0.019981997087597847, + 0.0047177504748106, + 0.005518267396837473, + 0.004519524984061718, + 0.007258332800120115, + 0.008808612823486328, + 0.005956249311566353, + 0.005929260980337858, + 0.005214564967900515, + 0.005890242755413055, + 0.006600288674235344, + 0.007166743278503418, + 0.005554869305342436, + 0.00652167946100235, + 0.005623962264508009, + 0.008338467217981815, + 0.006050918251276016, + 0.006582982838153839, + 0.004692475777119398, + 0.006612597499042749, + 0.006304614711552858, + 0.003896167501807213, + 0.004375799093395472, + 0.005127090960741043, + 0.008977766148746014, + 0.011480847373604774, + 0.013771237805485725, + 0.007160361856222153, + 0.010550265200436115, + 0.0126598309725523, + 0.011662003584206104, + 0.009780283086001873, + 0.010206056758761406, + 0.010066283866763115, + 0.010293661616742611, + 0.010623123496770859, + 0.012265553697943687, + 0.00840043742209673, + 0.010502435266971588, + 0.010948334820568562, + 0.011381663382053375, + 0.012299671769142151, + 0.01388054434210062, + 0.03259332850575447, + 0.015002149157226086, + 0.01550471130758524, + 0.011703677475452423, + 0.02117035537958145, + 0.028380364179611206, + 0.021598173305392265, + 0.021522192284464836, + 0.026428375393152237, + 0.032292816787958145, + 0.019475100561976433, + 0.016597233712673187, + 0.02311660163104534, + 0.02256914973258972, + 0.0049452390521764755, + 0.0077752359211444855, + 0.004632876254618168, + 0.005708970129489899, + 0.0045595853589475155, + 0.005821205209940672, + 0.006115468218922615, + 0.012325801886618137 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 7 + ], + "selected_reward": -0.006741991266608238 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned in the center of a narrow, deep alley?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera perform a steady, slow horizontal rotation?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the walls made of damp, dark masonry with peeling plaster revealing rough texture?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is there a lone street lamp at the end of the alley casting a faint, warm yellow glow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are the cobblestones underfoot uneven, slippery, and wet with puddles reflecting light?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.016083266586065292 + ], + "opposite_logprob_tokens": [ + -4.141083240509033 + ], + "logprob_reward": -0.016083266586065292, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.016083266586065292, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.05965492129325867, + 0.0400250107049942, + 0.05477122962474823, + 0.038756534457206726, + 0.06055095046758652, + 0.08746852725744247, + 0.06409244984388351, + 0.07894378900527954, + 0.41203317046165466, + 0.10370337963104248 + ], + "credit_map": [ + 0.003977588843554258, + 0.006655693054199219, + 0.006791330873966217, + 0.01164229679852724, + 0.006055224686861038, + 0.010656008496880531, + 0.005865420680493116, + 0.008011355064809322, + 0.0027101163286715746, + 0.0038572156336158514, + 0.00615426991134882, + 0.004047391004860401, + 0.005459287203848362, + 0.005403383634984493, + 0.006361500825732946, + 0.006031842902302742, + 0.004662846215069294, + 0.005051733925938606, + 0.004030895885080099, + 0.007040227763354778, + 0.008410360664129257, + 0.009760525077581406, + 0.008002412505447865, + 0.007812232244759798, + 0.0040253568440675735, + 0.003682827576994896, + 0.004988709464669228, + 0.005611599422991276, + 0.004578451160341501, + 0.005135578569024801, + 0.00637521967291832, + 0.004358790349215269, + 0.00618396420031786, + 0.00685516931116581, + 0.007953465916216373, + 0.007759236264973879, + 0.007690642960369587, + 0.006634584628045559, + 0.008176170289516449, + 0.00929771363735199, + 0.009883806109428406, + 0.015903452411293983, + 0.011275559663772583, + 0.013003109954297543, + 0.011361394077539444, + 0.008159901946783066, + 0.007814728654921055, + 0.010066568851470947, + 0.006559662986546755, + 0.008317266590893269, + 0.00903171207755804, + 0.010117604397237301, + 0.006850747857242823, + 0.006948359310626984, + 0.008073962293565273, + 0.00819313246756792, + 0.005567393731325865, + 0.004932194482535124, + 0.02055300585925579, + 0.005997535306960344, + 0.006631563417613506, + 0.00671734381467104, + 0.009591735899448395, + 0.01895301602780819, + 0.04545774310827255, + 0.053759749978780746, + 0.05396192520856857, + 0.06946281343698502, + 0.0444527305662632, + 0.04350179061293602, + 0.06370314955711365, + 0.03773326799273491, + 0.00858883373439312, + 0.012439883314073086, + 0.013480978086590767, + 0.019913174211978912, + 0.013450033031404018, + 0.008695525117218494, + 0.015299570746719837, + 0.011835383251309395 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 9, + 5 + ], + "selected_reward": -0.016083266586065292 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "shot shot from a level perspective, the camera positioned in the center of a narrow, deep alley, begins a steady, slow horizontal rotation. As the camera pans, the mottled, high walls on either side slowly flow across the frame. The walls are made of damp, dark masonry, large swathes of plaster peeling away, revealing the rough texture beneath. Dark green moss creeps through corners and cracks, and tangled strands of electrical wire dangle like black vines. Underfoot, the uneven, slippery cobblestones lie wet, with a few puddles reflecting the lone light source in the distance. At the end of the alley, a lone street lamp casts a faint, warm yellow glow, piercing the dense darkness, casting long, twisting shadows on the ground and bringing a glimmer of light to the claustrophobic space. The entire rotation creates an atmosphere of exploration, suspense, and loneliness, as if uncovering the secrets of a forgotten corner of the city.", + "seed": 20500, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000027/group_12/index_3791_seed_20500_dp_25_rank_50_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 12, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.045145031064748764, + "vlm_logprob": -0.045145031064748764, + "vlm_yes_no_margin": 3.5, + "vlm_token_credit": -0.045145031064748764, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned in the center of a narrow, deep alley?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera perform a steady, slow horizontal rotation?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the walls made of damp, dark masonry with peeling plaster revealing rough texture?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is there a lone street lamp at the end of the alley casting a faint, warm yellow glow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are the cobblestones underfoot uneven, slippery, and wet with puddles reflecting light?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.06346724182367325 + ], + "opposite_logprob_tokens": [ + -3.938467264175415 + ], + "logprob_reward": -0.06346724182367325, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.06346724182367325, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.40551242232322693, + 0.03060397505760193, + 0.02828432247042656, + 0.028951384127140045, + 0.0702594444155693, + 0.0574595145881176, + 0.14116238057613373, + 0.06962549686431885, + 0.08914650231599808, + 0.07899460941553116 + ], + "credit_map": [ + 0.03685358166694641, + 0.042376622557640076, + 0.02121548540890217, + 0.030972547829151154, + 0.04160422831773758, + 0.07131186127662659, + 0.0399673730134964, + 0.12121067941188812, + 0.0028473306447267532, + 0.0034512344282120466, + 0.0035157976672053337, + 0.003479013219475746, + 0.005267843138426542, + 0.004177295137196779, + 0.004074683412909508, + 0.0037907774094492197, + 0.0025944856461137533, + 0.002657434204593301, + 0.00422003073617816, + 0.0045278905890882015, + 0.003737952094525099, + 0.0027902889996767044, + 0.004132627043873072, + 0.0036236138548702, + 0.003551788628101349, + 0.0033199875615537167, + 0.0033210788387805223, + 0.0037791398353874683, + 0.003985781688243151, + 0.004073204938322306, + 0.0038582708220928907, + 0.0030621313489973545, + 0.0067160651087760925, + 0.011838718317449093, + 0.006892570294439793, + 0.008797976188361645, + 0.009662032127380371, + 0.008272955194115639, + 0.009074116125702858, + 0.009005009196698666, + 0.002607170958071947, + 0.006523302290588617, + 0.002458008239045739, + 0.0032256257254630327, + 0.004066103138029575, + 0.013682324439287186, + 0.00489445636048913, + 0.020002521574497223, + 0.015609239228069782, + 0.01273538451641798, + 0.019958261400461197, + 0.013461238704621792, + 0.017059845849871635, + 0.017149290069937706, + 0.018256014212965965, + 0.026933109387755394, + 0.009650422260165215, + 0.0053263320587575436, + 0.0086058359593153, + 0.005736477207392454, + 0.010523288510739803, + 0.00899267103523016, + 0.012158098630607128, + 0.00863236840814352, + 0.011320183053612709, + 0.005461147986352444, + 0.01702958159148693, + 0.009244509972631931, + 0.010203531943261623, + 0.009188935160636902, + 0.009062763303518295, + 0.01763584278523922, + 0.006735347677022219, + 0.007065347861498594, + 0.00854030903428793, + 0.010031400248408318, + 0.007609952706843615, + 0.010414603166282177, + 0.01573800854384899, + 0.01285963598638773 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 6, + 8 + ], + "selected_reward": -0.06346724182367325 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned in the center of a narrow, deep alley?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera perform a steady, slow horizontal rotation?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the walls made of damp, dark masonry with peeling plaster revealing rough texture?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is there a lone street lamp at the end of the alley casting a faint, warm yellow glow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are the cobblestones underfoot uneven, slippery, and wet with puddles reflecting light?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04868759214878082 + ], + "opposite_logprob_tokens": [ + -3.048687696456909 + ], + "logprob_reward": -0.04868759214878082, + "yes_no_margin_reward": 3.0, + "token_credit_reward": -0.04868759214878082, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.23216526210308075, + 0.06159484386444092, + 0.02444746531546116, + 0.03952110558748245, + 0.06609411537647247, + 0.1428215056657791, + 0.0969533920288086, + 0.1284395456314087, + 0.14944495260715485, + 0.058517780154943466 + ], + "credit_map": [ + 0.021659623831510544, + 0.017255226150155067, + 0.027278222143650055, + 0.018933258950710297, + 0.033362433314323425, + 0.04584839195013046, + 0.02022772841155529, + 0.0476003997027874, + 0.0037053744308650494, + 0.00643726484850049, + 0.004977991338819265, + 0.011677228845655918, + 0.013556215912103653, + 0.0070411041378974915, + 0.009180862456560135, + 0.0050188107416033745, + 0.001940595917403698, + 0.002387505490332842, + 0.0028335789684206247, + 0.0028010262176394463, + 0.0045971558429300785, + 0.003308648942038417, + 0.0037222709506750107, + 0.002856683684512973, + 0.004043424502015114, + 0.0051547992043197155, + 0.004457966890186071, + 0.006376737728714943, + 0.0028997764457017183, + 0.006953496020287275, + 0.0060144271701574326, + 0.003620481351390481, + 0.005560167133808136, + 0.007261370774358511, + 0.006542789749801159, + 0.011789425276219845, + 0.009487934410572052, + 0.007325343322008848, + 0.00990393664687872, + 0.008223156444728374, + 0.00576715636998415, + 0.01764577068388462, + 0.004054354038089514, + 0.006841436959803104, + 0.007225614972412586, + 0.03762786462903023, + 0.008765878155827522, + 0.05489342287182808, + 0.008401228114962578, + 0.011796576902270317, + 0.00993768684566021, + 0.010651415213942528, + 0.012883077375590801, + 0.00959726981818676, + 0.011901776306331158, + 0.021784374490380287, + 0.022192971780896187, + 0.009795339778065681, + 0.014379863627254963, + 0.014484190382063389, + 0.014149859547615051, + 0.014254352077841759, + 0.02157650515437126, + 0.017606468871235847, + 0.016229495406150818, + 0.0093877287581563, + 0.026444485411047935, + 0.0227230042219162, + 0.015529530122876167, + 0.01493180450052023, + 0.017306093126535416, + 0.026892827823758125, + 0.004405141342431307, + 0.004543048329651356, + 0.007989302277565002, + 0.004945908207446337, + 0.00478410255163908, + 0.015645040199160576, + 0.005083806347101927, + 0.011121434159576893 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 5 + ], + "selected_reward": -0.04868759214878082 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned in the center of a narrow, deep alley?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera perform a steady, slow horizontal rotation?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the walls made of damp, dark masonry with peeling plaster revealing rough texture?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is there a lone street lamp at the end of the alley casting a faint, warm yellow glow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are the cobblestones underfoot uneven, slippery, and wet with puddles reflecting light?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.009798642247915268 + ], + "opposite_logprob_tokens": [ + -4.634798526763916 + ], + "logprob_reward": -0.009798642247915268, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.009798642247915268, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.4847680926322937, + 0.032679133117198944, + 0.015963872894644737, + 0.0276014506816864, + 0.03233533352613449, + 0.13544128835201263, + 0.047234829515218735, + 0.03163822740316391, + 0.15876968204975128, + 0.033568043261766434 + ], + "credit_map": [ + 0.05733129009604454, + 0.029724452644586563, + 0.055123843252658844, + 0.039038579910993576, + 0.036664001643657684, + 0.09457375109195709, + 0.03214014321565628, + 0.14017203450202942, + 0.0024426488671451807, + 0.002657809527590871, + 0.002200300106778741, + 0.0037068098317831755, + 0.004931071773171425, + 0.007395921740680933, + 0.004662978462874889, + 0.004681591875851154, + 0.0012218948686495423, + 0.001275216112844646, + 0.0016859661554917693, + 0.002337744692340493, + 0.003395789535716176, + 0.0017842627130448818, + 0.00284934020601213, + 0.0014136588433757424, + 0.002283581532537937, + 0.0022731011267751455, + 0.0034531967248767614, + 0.004007949959486723, + 0.002941970946267247, + 0.005076580680906773, + 0.005154244136065245, + 0.0024108244106173515, + 0.0025444829370826483, + 0.005278684198856354, + 0.003110104240477085, + 0.004583515226840973, + 0.006267343182116747, + 0.0034187245182693005, + 0.004773967899382114, + 0.0023585082963109016, + 0.006220775656402111, + 0.014821815304458141, + 0.0031216407660394907, + 0.0062184822745621204, + 0.00464349240064621, + 0.038086362183094025, + 0.007660922594368458, + 0.05466780439019203, + 0.004201381467282772, + 0.005494000855833292, + 0.004733029752969742, + 0.0065152388997375965, + 0.00513494573533535, + 0.003528566099703312, + 0.008306336589157581, + 0.009321331977844238, + 0.0030438383109867573, + 0.002182858530431986, + 0.0032326539512723684, + 0.005486674141138792, + 0.004274438135325909, + 0.002326610032469034, + 0.004902348853647709, + 0.006188804749399424, + 0.013330516405403614, + 0.00874392595142126, + 0.01927432045340538, + 0.03163600340485573, + 0.015339785255491734, + 0.016125261783599854, + 0.020195554941892624, + 0.03412432223558426, + 0.00293960259296, + 0.0034778511617332697, + 0.004793654195964336, + 0.005227444227784872, + 0.0030056918039917946, + 0.0047410097904503345, + 0.005624753423035145, + 0.0037580353673547506 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 5 + ], + "selected_reward": -0.009798642247915268 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned in the center of a narrow, deep alley?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera perform a steady, slow horizontal rotation?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the walls made of damp, dark masonry with peeling plaster revealing rough texture?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is there a lone street lamp at the end of the alley casting a faint, warm yellow glow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are the cobblestones underfoot uneven, slippery, and wet with puddles reflecting light?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.033706627786159515 + ], + "opposite_logprob_tokens": [ + -3.4087066650390625 + ], + "logprob_reward": -0.033706627786159515, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.033706627786159515, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.6237767934799194, + 0.026308082044124603, + 0.04226835444569588, + 0.06358376890420914, + 0.0652633085846901, + 0.05515320226550102, + 0.016050036996603012, + 0.026637036353349686, + 0.04933728650212288, + 0.03162210062146187 + ], + "credit_map": [ + 0.080312080681324, + 0.04774918407201767, + 0.09398054331541061, + 0.06270603090524673, + 0.045269571244716644, + 0.10490120947360992, + 0.040323738008737564, + 0.14853449165821075, + 0.0015469540376216173, + 0.0022844935301691294, + 0.0019185664132237434, + 0.00419845525175333, + 0.0037926670629531145, + 0.004701565485447645, + 0.004045685287564993, + 0.003819699864834547, + 0.0034505054354667664, + 0.0031684120185673237, + 0.005128065589815378, + 0.0057274894788861275, + 0.006606054957956076, + 0.005240674130618572, + 0.007965395227074623, + 0.004981759935617447, + 0.008649297058582306, + 0.010636338964104652, + 0.00845851469784975, + 0.010295825079083443, + 0.006105225533246994, + 0.007602371275424957, + 0.006375163793563843, + 0.0054610371589660645, + 0.004926237277686596, + 0.006518278270959854, + 0.005880571436136961, + 0.010445346124470234, + 0.008634300902485847, + 0.007033699657768011, + 0.014162036590278149, + 0.00766284205019474, + 0.002247747965157032, + 0.007695167325437069, + 0.001719941501505673, + 0.003263553138822317, + 0.0036379036027938128, + 0.012248058803379536, + 0.004966016858816147, + 0.019374819472432137, + 0.0011147080222144723, + 0.0024507613852620125, + 0.0021429823245853186, + 0.001826688414439559, + 0.0020239518489688635, + 0.001582732074894011, + 0.0024514400865882635, + 0.0024567756336182356, + 0.005042728036642075, + 0.001821280224248767, + 0.0026551701594144106, + 0.002317111473530531, + 0.003896369831636548, + 0.003443268360570073, + 0.0034660573583096266, + 0.003995056264102459, + 0.006026851944625378, + 0.003503039013594389, + 0.007457863073796034, + 0.007760199252516031, + 0.005591102410107851, + 0.005638290196657181, + 0.0064005302265286446, + 0.006959415040910244, + 0.003967202268540859, + 0.0030965779442340136, + 0.0054303426295518875, + 0.004448457155376673, + 0.0030595557764172554, + 0.0038080390077084303, + 0.003965806681662798, + 0.0038461233489215374 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 3 + ], + "selected_reward": -0.033706627786159515 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned in the center of a narrow, deep alley?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera perform a steady, slow horizontal rotation?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the walls made of damp, dark masonry with peeling plaster revealing rough texture?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is there a lone street lamp at the end of the alley casting a faint, warm yellow glow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are the cobblestones underfoot uneven, slippery, and wet with puddles reflecting light?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07006505131721497 + ], + "opposite_logprob_tokens": [ + -2.6950650215148926 + ], + "logprob_reward": -0.07006505131721497, + "yes_no_margin_reward": 2.625, + "token_credit_reward": -0.07006505131721497, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2550172805786133, + 0.05503956973552704, + 0.03254464268684387, + 0.03202606365084648, + 0.036648549139499664, + 0.1402783989906311, + 0.03976975381374359, + 0.10219403356313705, + 0.26081427931785583, + 0.0456673800945282 + ], + "credit_map": [ + 0.02181435190141201, + 0.018917735666036606, + 0.011707545258104801, + 0.01782945729792118, + 0.036428969353437424, + 0.04260333999991417, + 0.030344408005475998, + 0.07537149637937546, + 0.004179089795798063, + 0.005630868952721357, + 0.0039060271810740232, + 0.005214247852563858, + 0.012049280107021332, + 0.011472069658339024, + 0.005769491661339998, + 0.006818499881774187, + 0.002403444144874811, + 0.002861633198335767, + 0.0036187798250466585, + 0.005183011293411255, + 0.006326624657958746, + 0.0034953532740473747, + 0.005607422441244125, + 0.00304837292060256, + 0.00384603813290596, + 0.0034075002186000347, + 0.004116551019251347, + 0.0038307332433760166, + 0.005332987755537033, + 0.0047031608410179615, + 0.003731529228389263, + 0.00305756414309144, + 0.003064699238166213, + 0.006291876547038555, + 0.0038746551144868135, + 0.006470533553510904, + 0.004253295715898275, + 0.0029791740234941244, + 0.006204742006957531, + 0.0035095757339149714, + 0.007463571149855852, + 0.014128985814750195, + 0.0039815339259803295, + 0.007005186751484871, + 0.006780186668038368, + 0.036808982491493225, + 0.011700144968926907, + 0.052409812808036804, + 0.002231236081570387, + 0.005101100541651249, + 0.0034898780286312103, + 0.00799888838082552, + 0.0037996149621903896, + 0.0026071665342897177, + 0.00888729840517044, + 0.0056545729748904705, + 0.01923763006925583, + 0.006544421426951885, + 0.012318465858697891, + 0.015168131329119205, + 0.010290466248989105, + 0.009471074678003788, + 0.011957963928580284, + 0.01720587909221649, + 0.019176678732037544, + 0.0162031427025795, + 0.022798342630267143, + 0.06345559656620026, + 0.025323156267404556, + 0.029587170109152794, + 0.03630969673395157, + 0.04796052724123001, + 0.003212948329746723, + 0.002930928487330675, + 0.0038101852405816317, + 0.005547196138650179, + 0.00431315042078495, + 0.009031437337398529, + 0.00955459475517273, + 0.00726694380864501 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 0, + 5 + ], + "selected_reward": -0.07006505131721497 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000027/group_15/index_tensor([921, 921])_group_15_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000027/group_15/index_tensor([921, 921])_group_15_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..cff4181fd2135daf95cf72cb58ca6ace740164c3 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000027/group_15/index_tensor([921, 921])_group_15_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "fluid drone shot, captured in Annie Leibovitz's signature dramatic style, slowly and gracefully ascends around Tokyo Tower. High-contrast lighting sculpts the tower's iconic international orange-red and white steel structure against a deep, midnight-blue velvety backdrop, each beam and support clearly visible. Thousands of warm golden lights illuminate its intricate truss structure, casting deep, layered shadows, while the signal lights at the spire rhythmically flicker a solitary, proud red. As the camera swirls, it takes in the vast expanse of Tokyo's cityscape below. The lights of countless buildings converge into a dazzling sea of light, while the shimmering streets pulsate like luminous arteries through the city's fabric. The rich, saturated tones and dramatic composition of light and shadow create an atmosphere both majestic and tinged with loneliness, as if capturing a powerful and emotional portrait of this iconic giant of a city that never sleeps.", + "seed": 4156, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000027/group_15/index_921_seed_4156_dp_30_rank_60_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 15, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.026308690570294856, + "vlm_logprob": -0.026308690570294856, + "vlm_yes_no_margin": 3.95, + "vlm_token_credit": -0.026308690570294856, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject Tokyo Tower visible in the video?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the tower's structure colored in international orange-red and white steel?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera slowly and gracefully ascend around the tower?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are thousands of warm golden lights illuminating the tower's truss structure?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera swirl to reveal the cityscape below with a sea of building lights?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02697647176682949 + ], + "opposite_logprob_tokens": [ + -3.6519765853881836 + ], + "logprob_reward": -0.02697647176682949, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.02697647176682949, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09333708882331848, + 0.07141346484422684, + 0.11838576197624207, + 0.0895223394036293, + 0.02624879963696003, + 0.19305568933486938, + 0.0702424868941307, + 0.028762247413396835, + 0.2497876137495041, + 0.0592445507645607 + ], + "credit_map": [ + 0.009273175150156021, + 0.007219679653644562, + 0.0146177364513278, + 0.014855548739433289, + 0.014834161847829819, + 0.010706890374422073, + 0.012169871479272842, + 0.009660028852522373, + 0.009591538459062576, + 0.006444713566452265, + 0.01534925028681755, + 0.006499255541712046, + 0.008314842358231544, + 0.006528856232762337, + 0.009019256569445133, + 0.009665749035775661, + 0.01652444154024124, + 0.010079648345708847, + 0.018357839435338974, + 0.009440198540687561, + 0.013242529705166817, + 0.01476648822426796, + 0.021809270605444908, + 0.01416533812880516, + 0.022183844819664955, + 0.005607318133115768, + 0.024547500535845757, + 0.0055861095897853374, + 0.007374856621026993, + 0.007925924845039845, + 0.009049560874700546, + 0.007247226778417826, + 0.003019646741449833, + 0.002759101102128625, + 0.004137001466006041, + 0.003473754972219467, + 0.003601278644055128, + 0.003456261707469821, + 0.003232819028198719, + 0.002568934578448534, + 0.021892501041293144, + 0.02705538272857666, + 0.029411878436803818, + 0.014771133661270142, + 0.021404802799224854, + 0.0351998507976532, + 0.027471154928207397, + 0.015848975628614426, + 0.013629910536110401, + 0.00647337269037962, + 0.014241856522858143, + 0.0056143999099731445, + 0.008254960179328918, + 0.006591075100004673, + 0.008499469608068466, + 0.006937442813068628, + 0.004822587128728628, + 0.0046317060478031635, + 0.005040880758315325, + 0.0020634671673178673, + 0.002972646849229932, + 0.002987936604768038, + 0.0029272560495883226, + 0.003315765643492341, + 0.029867129400372505, + 0.04233561083674431, + 0.04041992500424385, + 0.02446610853075981, + 0.025177402421832085, + 0.033022962510585785, + 0.03392252326011658, + 0.020575959235429764, + 0.009469849988818169, + 0.006220633629709482, + 0.010297972708940506, + 0.002820407273247838, + 0.004995077382773161, + 0.014543299563229084, + 0.004703035578131676, + 0.006194274872541428 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 5, + 2 + ], + "selected_reward": -0.02697647176682949 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject Tokyo Tower visible in the video?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the tower's structure colored in international orange-red and white steel?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera slowly and gracefully ascend around the tower?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are thousands of warm golden lights illuminating the tower's truss structure?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera swirl to reveal the cityscape below with a sea of building lights?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04866976663470268 + ], + "opposite_logprob_tokens": [ + -3.0486698150634766 + ], + "logprob_reward": -0.04866976663470268, + "yes_no_margin_reward": 3.0, + "token_credit_reward": -0.04866976663470268, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13366080820560455, + 0.0553593710064888, + 0.08607479929924011, + 0.031961262226104736, + 0.04031366482377052, + 0.1044282615184784, + 0.08803456276655197, + 0.048796068876981735, + 0.37360459566116333, + 0.037766605615615845 + ], + "credit_map": [ + 0.015140182338654995, + 0.008600852452218533, + 0.024883732199668884, + 0.01866834983229637, + 0.019672829657793045, + 0.017659243196249008, + 0.018971167504787445, + 0.01006444450467825, + 0.00745823048055172, + 0.0034355339594185352, + 0.010959895327687263, + 0.004427394364029169, + 0.00600980781018734, + 0.00879048090428114, + 0.006077344063669443, + 0.008200686424970627, + 0.012247573584318161, + 0.009325793012976646, + 0.013866513967514038, + 0.00680546322837472, + 0.009009572677314281, + 0.01091373898088932, + 0.012384386733174324, + 0.011521763168275356, + 0.00531330332159996, + 0.002601012121886015, + 0.005454903468489647, + 0.0027526570484042168, + 0.004327327013015747, + 0.004145713523030281, + 0.0040017394348979, + 0.0033646044321358204, + 0.006223974283784628, + 0.0032763180788606405, + 0.006904446519911289, + 0.004043698310852051, + 0.005189604125916958, + 0.004955770913511515, + 0.004890676587820053, + 0.004829170648008585, + 0.012821401469409466, + 0.01355294231325388, + 0.019866585731506348, + 0.009176339954137802, + 0.009001499973237514, + 0.018568115308880806, + 0.0112199317663908, + 0.010221455246210098, + 0.013106236234307289, + 0.018490726128220558, + 0.014941293746232986, + 0.006668440531939268, + 0.010376068763434887, + 0.010212100110948086, + 0.00842036958783865, + 0.005819322541356087, + 0.004126339219510555, + 0.004586680792272091, + 0.0073684211820364, + 0.003645502496510744, + 0.006283260881900787, + 0.008546244353055954, + 0.007949896156787872, + 0.006289723329246044, + 0.04963351786136627, + 0.042788513004779816, + 0.06094246357679367, + 0.036538541316986084, + 0.046056054532527924, + 0.042384661734104156, + 0.046167612075805664, + 0.049093205481767654, + 0.004619266837835312, + 0.004859842825680971, + 0.006431231275200844, + 0.002432436915114522, + 0.006549594458192587, + 0.004592441953718662, + 0.0042698318138718605, + 0.004011959303170443 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 0, + 5 + ], + "selected_reward": -0.04866976663470268 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject Tokyo Tower visible in the video?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the tower's structure colored in international orange-red and white steel?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera slowly and gracefully ascend around the tower?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are thousands of warm golden lights illuminating the tower's truss structure?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera swirl to reveal the cityscape below with a sea of building lights?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02327992208302021 + ], + "opposite_logprob_tokens": [ + -3.773279905319214 + ], + "logprob_reward": -0.02327992208302021, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.02327992208302021, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08572714030742645, + 0.0641326829791069, + 0.1385568082332611, + 0.31556522846221924, + 0.04467451944947243, + 0.08628728240728378, + 0.06056783348321915, + 0.026169147342443466, + 0.15271331369876862, + 0.0256060641258955 + ], + "credit_map": [ + 0.009004568681120872, + 0.0069922166876494884, + 0.015145753510296345, + 0.010132666677236557, + 0.013710242696106434, + 0.010493879206478596, + 0.01097527053207159, + 0.009272539988160133, + 0.01035738829523325, + 0.008159313350915909, + 0.011922175996005535, + 0.005642616655677557, + 0.0071367365308105946, + 0.006134495139122009, + 0.007623903453350067, + 0.007156059145927429, + 0.015334219671785831, + 0.010232998989522457, + 0.023685552179813385, + 0.011819533072412014, + 0.013716396875679493, + 0.015148306265473366, + 0.03134947642683983, + 0.017270324751734734, + 0.08121334761381149, + 0.019847901538014412, + 0.09627225995063782, + 0.021923953667283058, + 0.021171648055315018, + 0.025933539494872093, + 0.028872735798358917, + 0.02032986655831337, + 0.007011388894170523, + 0.002922247862443328, + 0.007715194020420313, + 0.003873523324728012, + 0.006948254536837339, + 0.005066380370408297, + 0.005163609050214291, + 0.0059739225544035435, + 0.009756568819284439, + 0.008599963039159775, + 0.014961175620555878, + 0.006791071966290474, + 0.008962631225585938, + 0.012587813660502434, + 0.013236043974757195, + 0.011392015963792801, + 0.011437546461820602, + 0.005013037007302046, + 0.012223170138895512, + 0.004450553562492132, + 0.007748298812657595, + 0.006142801139503717, + 0.007903296500444412, + 0.005649129394441843, + 0.004828040022403002, + 0.003483953420072794, + 0.004799296613782644, + 0.0017212878447026014, + 0.0032295817509293556, + 0.0032398307230323553, + 0.002811994170770049, + 0.002055160468444228, + 0.02014332450926304, + 0.01887389086186886, + 0.024261269718408585, + 0.014724929817020893, + 0.016527537256479263, + 0.01886991411447525, + 0.01812734082341194, + 0.021185114979743958, + 0.004662589635699987, + 0.002647270215675235, + 0.004151824861764908, + 0.0019068336114287376, + 0.0031433566473424435, + 0.0039281584322452545, + 0.0021259719505906105, + 0.0030400604009628296 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 8, + 2 + ], + "selected_reward": -0.02327992208302021 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject Tokyo Tower visible in the video?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the tower's structure colored in international orange-red and white steel?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera slowly and gracefully ascend around the tower?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are thousands of warm golden lights illuminating the tower's truss structure?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera swirl to reveal the cityscape below with a sea of building lights?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.002817234955728054 + ], + "opposite_logprob_tokens": [ + -5.877817153930664 + ], + "logprob_reward": -0.002817234955728054, + "yes_no_margin_reward": 5.875, + "token_credit_reward": -0.002817234955728054, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15779870748519897, + 0.08037467300891876, + 0.04020391404628754, + 0.09718440473079681, + 0.06078268215060234, + 0.12085877358913422, + 0.04851091280579567, + 0.05918193235993385, + 0.2828483581542969, + 0.05225569009780884 + ], + "credit_map": [ + 0.013821487314999104, + 0.014748099260032177, + 0.032524581998586655, + 0.01641257107257843, + 0.03083513304591179, + 0.012736830860376358, + 0.02224758267402649, + 0.014472408220171928, + 0.010733223520219326, + 0.005780222825706005, + 0.015130735002458096, + 0.006612008903175592, + 0.009828357957303524, + 0.008585065603256226, + 0.011822237633168697, + 0.011882811784744263, + 0.005779135972261429, + 0.0037009743973612785, + 0.006197597831487656, + 0.004069135989993811, + 0.005177836399525404, + 0.0047876061871647835, + 0.0061223432421684265, + 0.004369279369711876, + 0.02688099443912506, + 0.005538548342883587, + 0.03168608993291855, + 0.005928772035986185, + 0.006372103467583656, + 0.0077566122636199, + 0.006869768723845482, + 0.006151498761028051, + 0.007488357834517956, + 0.00372729217633605, + 0.011330455541610718, + 0.004931267350912094, + 0.008869287557899952, + 0.008920673280954361, + 0.009491348639130592, + 0.0060239932499825954, + 0.012978415936231613, + 0.014974541962146759, + 0.020720988512039185, + 0.010507762432098389, + 0.015200557187199593, + 0.020275499671697617, + 0.016066212207078934, + 0.01013477984815836, + 0.0072556305676698685, + 0.0075705209746956825, + 0.008841889910399914, + 0.003739375388249755, + 0.006742226891219616, + 0.005731296259909868, + 0.0050185020081698895, + 0.0036114652175456285, + 0.007578786462545395, + 0.006415874231606722, + 0.009106124751269817, + 0.005324383731931448, + 0.007138337008655071, + 0.008512777276337147, + 0.008141888305544853, + 0.006963752675801516, + 0.03419552370905876, + 0.04060390964150429, + 0.044858649373054504, + 0.027906248345971107, + 0.030341481789946556, + 0.03795156627893448, + 0.02855515107512474, + 0.03843582421541214, + 0.0066972034983336926, + 0.0051686582155525684, + 0.008519738912582397, + 0.003169752424582839, + 0.005052783992141485, + 0.012374836951494217, + 0.004846095107495785, + 0.006426610052585602 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 0, + 5 + ], + "selected_reward": -0.002817234955728054 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject Tokyo Tower visible in the video?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the tower's structure colored in international orange-red and white steel?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera slowly and gracefully ascend around the tower?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are thousands of warm golden lights illuminating the tower's truss structure?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera swirl to reveal the cityscape below with a sea of building lights?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.029800057411193848 + ], + "opposite_logprob_tokens": [ + -3.5297999382019043 + ], + "logprob_reward": -0.029800057411193848, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.029800057411193848, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1045486107468605, + 0.283516526222229, + 0.042585209012031555, + 0.1452372819185257, + 0.03760474920272827, + 0.08078327775001526, + 0.04433605819940567, + 0.05502433702349663, + 0.12738892436027527, + 0.07897504419088364 + ], + "credit_map": [ + 0.012140586972236633, + 0.009541292674839497, + 0.026058515533804893, + 0.00943834614008665, + 0.012990470975637436, + 0.010559719987213612, + 0.015008619986474514, + 0.008811059407889843, + 0.040897760540246964, + 0.02783953957259655, + 0.07892130315303802, + 0.018505094572901726, + 0.022064177319407463, + 0.017481600865721703, + 0.04327230527997017, + 0.034534744918346405, + 0.0064210607670247555, + 0.004744181875139475, + 0.007877646945416927, + 0.003091607242822647, + 0.003478973638266325, + 0.004728415049612522, + 0.007349363062530756, + 0.004893959034234285, + 0.04902021214365959, + 0.007041269913315773, + 0.055267393589019775, + 0.007614923641085625, + 0.006555724889039993, + 0.007035217247903347, + 0.006931221578270197, + 0.005771304480731487, + 0.00861378200352192, + 0.0032731059473007917, + 0.007509823888540268, + 0.0030983209144324064, + 0.004112935625016689, + 0.004071184899657965, + 0.0032063061371445656, + 0.003719288157299161, + 0.011302025988698006, + 0.005366142373532057, + 0.019983870908617973, + 0.005313561763614416, + 0.012700733728706837, + 0.007718109525740147, + 0.011200717650353909, + 0.007198116276413202, + 0.006455558817833662, + 0.009016088210046291, + 0.006641353480517864, + 0.003888757899403572, + 0.0049008834175765514, + 0.0062565128318965435, + 0.003722100518643856, + 0.0034548037219792604, + 0.007858500815927982, + 0.00603509321808815, + 0.009245459921658039, + 0.0046494570560753345, + 0.006233582738786936, + 0.008535043336451054, + 0.007360766641795635, + 0.005106430966407061, + 0.013225940056145191, + 0.01664719544351101, + 0.019633818417787552, + 0.012856225483119488, + 0.013130980543792248, + 0.018327713012695312, + 0.01567799411714077, + 0.017889056354761124, + 0.014465587213635445, + 0.006725935265421867, + 0.014748563058674335, + 0.003859409596771002, + 0.00607330072671175, + 0.021119344979524612, + 0.003358920104801655, + 0.008623991161584854 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 3, + 8 + ], + "selected_reward": -0.029800057411193848 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "fluid drone shot, captured in Annie Leibovitz's signature dramatic style, slowly and gracefully ascends around Tokyo Tower. High-contrast lighting sculpts the tower's iconic international orange-red and white steel structure against a deep, midnight-blue velvety backdrop, each beam and support clearly visible. Thousands of warm golden lights illuminate its intricate truss structure, casting deep, layered shadows, while the signal lights at the spire rhythmically flicker a solitary, proud red. As the camera swirls, it takes in the vast expanse of Tokyo's cityscape below. The lights of countless buildings converge into a dazzling sea of light, while the shimmering streets pulsate like luminous arteries through the city's fabric. The rich, saturated tones and dramatic composition of light and shadow create an atmosphere both majestic and tinged with loneliness, as if capturing a powerful and emotional portrait of this iconic giant of a city that never sleeps.", + "seed": 4157, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000027/group_15/index_921_seed_4157_dp_30_rank_60_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 15, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.016225069854408504, + "vlm_logprob": -0.016225069854408504, + "vlm_yes_no_margin": 4.7, + "vlm_token_credit": -0.016225069854408504, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject Tokyo Tower visible in the video?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the tower's structure colored in international orange-red and white steel?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera slowly and gracefully ascend around the tower?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are thousands of warm golden lights illuminating the tower's truss structure?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera swirl to reveal the cityscape below with a sea of building lights?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0017726197838783264 + ], + "opposite_logprob_tokens": [ + -6.626772403717041 + ], + "logprob_reward": -0.0017726197838783264, + "yes_no_margin_reward": 6.625, + "token_credit_reward": -0.0017726197838783264, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.19094491004943848, + 0.06300556659698486, + 0.08051608502864838, + 0.13839562237262726, + 0.05132841318845749, + 0.12158938497304916, + 0.06928956508636475, + 0.057044804096221924, + 0.19322988390922546, + 0.03465580940246582 + ], + "credit_map": [ + 0.02288963831961155, + 0.0206612516194582, + 0.028531381860375404, + 0.02603963017463684, + 0.02391475811600685, + 0.025184256955981255, + 0.02338566444814205, + 0.020338337868452072, + 0.011049504391849041, + 0.007671283092349768, + 0.007963484153151512, + 0.007688469253480434, + 0.006216126494109631, + 0.007713136728852987, + 0.005408237688243389, + 0.009295322932302952, + 0.010051093064248562, + 0.010079214349389076, + 0.011823042295873165, + 0.008359143510460854, + 0.010163357481360435, + 0.01256799977272749, + 0.008990395814180374, + 0.00848183874040842, + 0.013638905249536037, + 0.027548495680093765, + 0.02180618792772293, + 0.01768467389047146, + 0.019755948334932327, + 0.013750969432294369, + 0.009968960657715797, + 0.014241483062505722, + 0.008388848975300789, + 0.004023153334856033, + 0.00742329889908433, + 0.0048185852356255054, + 0.002886996604502201, + 0.009720449335873127, + 0.0040064724162220955, + 0.010060609318315983, + 0.01569075509905815, + 0.009485551156103611, + 0.015704108402132988, + 0.016093077138066292, + 0.010463738813996315, + 0.028785504400730133, + 0.009640352800488472, + 0.015726299956440926, + 0.009056208655238152, + 0.0091194873675704, + 0.007520632352679968, + 0.005837880540639162, + 0.005674479063600302, + 0.014203546568751335, + 0.006008842960000038, + 0.011868493631482124, + 0.008668526075780392, + 0.0068581243976950645, + 0.004935594275593758, + 0.0075943199917674065, + 0.005712611600756645, + 0.008676685392856598, + 0.006106714252382517, + 0.00849222857505083, + 0.021165497601032257, + 0.016941986978054047, + 0.020096248015761375, + 0.03890667110681534, + 0.02013077214360237, + 0.019123191013932228, + 0.03435579314827919, + 0.02250971831381321, + 0.004831539001315832, + 0.0033544511534273624, + 0.005837026983499527, + 0.003974335268139839, + 0.003611647989600897, + 0.003717212937772274, + 0.003137753112241626, + 0.006191844120621681 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 0, + 3 + ], + "selected_reward": -0.0017726197838783264 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject Tokyo Tower visible in the video?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the tower's structure colored in international orange-red and white steel?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera slowly and gracefully ascend around the tower?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are thousands of warm golden lights illuminating the tower's truss structure?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera swirl to reveal the cityscape below with a sea of building lights?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.020591065287590027 + ], + "opposite_logprob_tokens": [ + -3.8955910205841064 + ], + "logprob_reward": -0.020591065287590027, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.020591065287590027, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2446264922618866, + 0.07825800031423569, + 0.06852125376462936, + 0.10660690069198608, + 0.13242065906524658, + 0.08357511460781097, + 0.07033956050872803, + 0.14288264513015747, + 0.03954816237092018, + 0.033221203833818436 + ], + "credit_map": [ + 0.029828885570168495, + 0.04481339082121849, + 0.03111046925187111, + 0.027170974761247635, + 0.028066905215382576, + 0.029939526692032814, + 0.026441168040037155, + 0.027255170047283173, + 0.01357288844883442, + 0.006771480198949575, + 0.010996261611580849, + 0.013125172816216946, + 0.009160002693533897, + 0.007680700626224279, + 0.00882258452475071, + 0.008128909394145012, + 0.008339564315974712, + 0.008348979987204075, + 0.009964028373360634, + 0.00708305649459362, + 0.008966307155787945, + 0.011158603243529797, + 0.007530688773840666, + 0.007130028679966927, + 0.014265728183090687, + 0.016004502773284912, + 0.01599039137363434, + 0.009811061434447765, + 0.010270403698086739, + 0.018164783716201782, + 0.010429920628666878, + 0.011670110747218132, + 0.01741758920252323, + 0.0182652547955513, + 0.017468269914388657, + 0.01907404139637947, + 0.011383002623915672, + 0.016750378534197807, + 0.009228533133864403, + 0.02283359132707119, + 0.011083198711276054, + 0.014386872760951519, + 0.011101080104708672, + 0.008242647163569927, + 0.008473253808915615, + 0.01229157205671072, + 0.0069427089765667915, + 0.011053774505853653, + 0.008367953822016716, + 0.008869587443768978, + 0.008683806285262108, + 0.006421881727874279, + 0.006196439266204834, + 0.013056065887212753, + 0.006954140029847622, + 0.011789686977863312, + 0.021111290901899338, + 0.01616768166422844, + 0.015605541877448559, + 0.019496595486998558, + 0.014472405426204205, + 0.018915867432951927, + 0.017095163464546204, + 0.02001810073852539, + 0.005931174382567406, + 0.0031028706580400467, + 0.005802315194159746, + 0.00447955634444952, + 0.0037216825876384974, + 0.00605761120095849, + 0.004724881611764431, + 0.005728071555495262, + 0.003910400439053774, + 0.003844617633149028, + 0.005831180606037378, + 0.003763860324397683, + 0.003209077287465334, + 0.0032120496034622192, + 0.0035566966980695724, + 0.0058933221735060215 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 7, + 4 + ], + "selected_reward": -0.020591065287590027 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject Tokyo Tower visible in the video?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the tower's structure colored in international orange-red and white steel?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera slowly and gracefully ascend around the tower?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are thousands of warm golden lights illuminating the tower's truss structure?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera swirl to reveal the cityscape below with a sea of building lights?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0430525541305542 + ], + "opposite_logprob_tokens": [ + -3.1680526733398438 + ], + "logprob_reward": -0.0430525541305542, + "yes_no_margin_reward": 3.125, + "token_credit_reward": -0.0430525541305542, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1334874927997589, + 0.054826319217681885, + 0.038098372519016266, + 0.11642418801784515, + 0.06427481770515442, + 0.07873311638832092, + 0.06833097338676453, + 0.03533080592751503, + 0.24296100437641144, + 0.16753284633159637 + ], + "credit_map": [ + 0.0162278413772583, + 0.029855169355869293, + 0.021605664864182472, + 0.01177369337528944, + 0.017832007259130478, + 0.013161401264369488, + 0.010522625409066677, + 0.01250910572707653, + 0.008423181250691414, + 0.0077604628168046474, + 0.007898454554378986, + 0.008073577657341957, + 0.00499480776488781, + 0.006109227426350117, + 0.0038953276816755533, + 0.007671283558011055, + 0.005534149706363678, + 0.004226335324347019, + 0.005606490187346935, + 0.00392214534804225, + 0.004404646810144186, + 0.0055564227513968945, + 0.0029443625826388597, + 0.005903826095163822, + 0.014573797583580017, + 0.01876063458621502, + 0.017659561708569527, + 0.011668535880744457, + 0.013290963135659695, + 0.020194046199321747, + 0.008082161657512188, + 0.012194490991532803, + 0.008407073095440865, + 0.006575966253876686, + 0.008591822348535061, + 0.006815431173890829, + 0.0049551487900316715, + 0.012562980875372887, + 0.0040795146487653255, + 0.012286889366805553, + 0.007575008552521467, + 0.009897390380501747, + 0.009373078122735023, + 0.011816051788628101, + 0.009529110044240952, + 0.012938705272972584, + 0.007740122731775045, + 0.009863647632300854, + 0.008178015239536762, + 0.006371663883328438, + 0.007920361123979092, + 0.003900824813172221, + 0.004107796587049961, + 0.016514571383595467, + 0.004937891848385334, + 0.01639985479414463, + 0.0050436630845069885, + 0.004280091729015112, + 0.003800678299739957, + 0.005516262259334326, + 0.0028118258342146873, + 0.00364675116725266, + 0.005571440793573856, + 0.004660093691200018, + 0.026746630668640137, + 0.02214634232223034, + 0.0265270434319973, + 0.047934260219335556, + 0.026315711438655853, + 0.02329329214990139, + 0.041517842561006546, + 0.028479905799031258, + 0.018566949293017387, + 0.021811552345752716, + 0.01721375808119774, + 0.019073696807026863, + 0.03126465529203415, + 0.018749771639704704, + 0.01767813041806221, + 0.023174341768026352 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 9, + 0 + ], + "selected_reward": -0.0430525541305542 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject Tokyo Tower visible in the video?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the tower's structure colored in international orange-red and white steel?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera slowly and gracefully ascend around the tower?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are thousands of warm golden lights illuminating the tower's truss structure?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera swirl to reveal the cityscape below with a sea of building lights?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004636608529835939 + ], + "opposite_logprob_tokens": [ + -5.379636764526367 + ], + "logprob_reward": -0.004636608529835939, + "yes_no_margin_reward": 5.375, + "token_credit_reward": -0.004636608529835939, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15552490949630737, + 0.1056860163807869, + 0.14731423556804657, + 0.06992913782596588, + 0.20991559326648712, + 0.06783030927181244, + 0.06590693444013596, + 0.07440685480833054, + 0.05513496696949005, + 0.04835100844502449 + ], + "credit_map": [ + 0.025283580645918846, + 0.025534609332680702, + 0.021539589390158653, + 0.014356594532728195, + 0.021397892385721207, + 0.01739475317299366, + 0.012975490652024746, + 0.017042404040694237, + 0.017748933285474777, + 0.0117800896987319, + 0.012586387805640697, + 0.013284420594573021, + 0.013917505741119385, + 0.010541587136685848, + 0.012877071276307106, + 0.012950032018125057, + 0.018352724611759186, + 0.01806984655559063, + 0.020380813628435135, + 0.022796904668211937, + 0.02017849311232567, + 0.018349003046751022, + 0.016322551295161247, + 0.01286389958113432, + 0.009150794707238674, + 0.00820318702608347, + 0.010869717225432396, + 0.0070817843079566956, + 0.006419834680855274, + 0.01295082550495863, + 0.005454570055007935, + 0.009798428043723106, + 0.025166204199194908, + 0.029705680906772614, + 0.02733922377228737, + 0.03658798336982727, + 0.019018571823835373, + 0.023742884397506714, + 0.01743139885365963, + 0.030923636630177498, + 0.009524475783109665, + 0.007174134254455566, + 0.009924829006195068, + 0.009235658682882786, + 0.005138969514518976, + 0.013438163325190544, + 0.004908052273094654, + 0.008486025035381317, + 0.008829248137772083, + 0.006582476664334536, + 0.006888723466545343, + 0.005297393538057804, + 0.005011224653571844, + 0.01439157035201788, + 0.006042297929525375, + 0.012864002026617527, + 0.012478958815336227, + 0.0071612331084907055, + 0.009241331368684769, + 0.009797844104468822, + 0.005455418955534697, + 0.009607195854187012, + 0.008070535026490688, + 0.012594346888363361, + 0.007867214269936085, + 0.005787684582173824, + 0.007257118355482817, + 0.007642228621989489, + 0.0055033545941114426, + 0.00685541657730937, + 0.007174426689743996, + 0.0070475246757268906, + 0.0063122911378741264, + 0.005492310505360365, + 0.0061058104038238525, + 0.005005023442208767, + 0.007731284946203232, + 0.005271026398986578, + 0.004763384349644184, + 0.007669881451874971 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 0, + 2 + ], + "selected_reward": -0.004636608529835939 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject Tokyo Tower visible in the video?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the tower's structure colored in international orange-red and white steel?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera slowly and gracefully ascend around the tower?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are thousands of warm golden lights illuminating the tower's truss structure?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera swirl to reveal the cityscape below with a sea of building lights?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.011072501540184021 + ], + "opposite_logprob_tokens": [ + -4.511072635650635 + ], + "logprob_reward": -0.011072501540184021, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.011072501540184021, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14416630566120148, + 0.04866141080856323, + 0.040336042642593384, + 0.07023400068283081, + 0.14640125632286072, + 0.21683479845523834, + 0.09771468490362167, + 0.039476025849580765, + 0.14455285668373108, + 0.051622603088617325 + ], + "credit_map": [ + 0.03328684717416763, + 0.014169318601489067, + 0.021543679758906364, + 0.01428918819874525, + 0.012960150837898254, + 0.01951272413134575, + 0.00828468892723322, + 0.020119700580835342, + 0.0064787534065544605, + 0.006162870209664106, + 0.00744041008874774, + 0.005997856613248587, + 0.005388396326452494, + 0.005216227378696203, + 0.004296195227652788, + 0.007680702023208141, + 0.00565155828371644, + 0.005178206600248814, + 0.005521026905626059, + 0.004634246230125427, + 0.00551024591550231, + 0.004337141755968332, + 0.0046882107853889465, + 0.004815409425646067, + 0.007362349424511194, + 0.013579290360212326, + 0.010896467603743076, + 0.009341123513877392, + 0.00911431573331356, + 0.007922305725514889, + 0.005660264287143946, + 0.006357889156788588, + 0.020164605230093002, + 0.01733526401221752, + 0.018182197585701942, + 0.0162757970392704, + 0.014874085783958435, + 0.025723330676555634, + 0.01105062011629343, + 0.022795354947447777, + 0.027909090742468834, + 0.018271679058670998, + 0.02389201521873474, + 0.027571892365813255, + 0.024229705333709717, + 0.04962840676307678, + 0.01921694166958332, + 0.02611505426466465, + 0.009697937406599522, + 0.005415262654423714, + 0.009547295048832893, + 0.004357919562608004, + 0.004132836125791073, + 0.03031335212290287, + 0.005134499166160822, + 0.0291155818849802, + 0.00530785508453846, + 0.0047443462535738945, + 0.0051485165022313595, + 0.005649711936712265, + 0.0035967035219073296, + 0.003512244438752532, + 0.006398405414074659, + 0.005118242930620909, + 0.01853322796523571, + 0.012648181989789009, + 0.014840764924883842, + 0.025828517973423004, + 0.01635402999818325, + 0.014852489344775677, + 0.026175707578659058, + 0.01531993132084608, + 0.007655187975615263, + 0.00559124955907464, + 0.007781105116009712, + 0.005732119083404541, + 0.006360007915645838, + 0.006351083051413298, + 0.004649532958865166, + 0.007502319291234016 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 4, + 8 + ], + "selected_reward": -0.011072501540184021 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "fluid drone shot, captured in Annie Leibovitz's signature dramatic style, slowly and gracefully ascends around Tokyo Tower. High-contrast lighting sculpts the tower's iconic international orange-red and white steel structure against a deep, midnight-blue velvety backdrop, each beam and support clearly visible. Thousands of warm golden lights illuminate its intricate truss structure, casting deep, layered shadows, while the signal lights at the spire rhythmically flicker a solitary, proud red. As the camera swirls, it takes in the vast expanse of Tokyo's cityscape below. The lights of countless buildings converge into a dazzling sea of light, while the shimmering streets pulsate like luminous arteries through the city's fabric. The rich, saturated tones and dramatic composition of light and shadow create an atmosphere both majestic and tinged with loneliness, as if capturing a powerful and emotional portrait of this iconic giant of a city that never sleeps.", + "seed": 4158, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000027/group_15/index_921_seed_4158_dp_31_rank_62_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 15, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.016513931518420576, + "vlm_logprob": -0.016513931518420576, + "vlm_yes_no_margin": 5.0, + "vlm_token_credit": -0.016513931518420576, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject Tokyo Tower visible in the video?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the tower's structure colored in international orange-red and white steel?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera slowly and gracefully ascend around the tower?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are thousands of warm golden lights illuminating the tower's truss structure?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera swirl to reveal the cityscape below with a sea of building lights?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.003908852580934763 + ], + "opposite_logprob_tokens": [ + -5.753908634185791 + ], + "logprob_reward": -0.003908852580934763, + "yes_no_margin_reward": 5.75, + "token_credit_reward": -0.003908852580934763, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3027813732624054, + 0.09948521852493286, + 0.09204143285751343, + 0.12145522236824036, + 0.18405595421791077, + 0.07105141133069992, + 0.03799743577837944, + 0.03232048079371452, + 0.01882997341454029, + 0.03998147323727608 + ], + "credit_map": [ + 0.0365593247115612, + 0.011503956280648708, + 0.11386080086231232, + 0.012783556245267391, + 0.011045333929359913, + 0.05982217192649841, + 0.013226509094238281, + 0.043979767709970474, + 0.009490440599620342, + 0.009405720978975296, + 0.01118743047118187, + 0.006826634984463453, + 0.004701496567577124, + 0.025037195533514023, + 0.00658407574519515, + 0.026252225041389465, + 0.010428069159388542, + 0.009921836666762829, + 0.014037339016795158, + 0.011394243687391281, + 0.004459561314433813, + 0.013541696593165398, + 0.00897257961332798, + 0.019286103546619415, + 0.016069693490862846, + 0.014690164476633072, + 0.025278691202402115, + 0.012715756893157959, + 0.00889835599809885, + 0.01728139817714691, + 0.008832814171910286, + 0.01768835447728634, + 0.026420824229717255, + 0.02700251154601574, + 0.02119128778576851, + 0.027788586914539337, + 0.016450820490717888, + 0.021636441349983215, + 0.016227537766098976, + 0.0273379348218441, + 0.012700062245130539, + 0.012020135298371315, + 0.008070479147136211, + 0.006161003839224577, + 0.005820176098495722, + 0.009154471568763256, + 0.005832702852785587, + 0.011292383074760437, + 0.003576527815312147, + 0.004417459014803171, + 0.0028284781146794558, + 0.005407118238508701, + 0.004498571157455444, + 0.008556249551475048, + 0.0032834908924996853, + 0.005429539363831282, + 0.0031485860235989094, + 0.0022603224497288465, + 0.005387323908507824, + 0.00319692469201982, + 0.003414356615394354, + 0.006169568747282028, + 0.00438264012336731, + 0.00436075683683157, + 0.0017364001832902431, + 0.0012110101524740458, + 0.0022481027990579605, + 0.00188432855065912, + 0.0015916115371510386, + 0.0046764034777879715, + 0.0027335474733263254, + 0.002748567843809724, + 0.0060937595553696156, + 0.004142914433032274, + 0.003950067330151796, + 0.004541375674307346, + 0.007499690167605877, + 0.004709922708570957, + 0.004784977063536644, + 0.004258765839040279 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 3 + ], + "selected_reward": -0.003908852580934763 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject Tokyo Tower visible in the video?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the tower's structure colored in international orange-red and white steel?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera slowly and gracefully ascend around the tower?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are thousands of warm golden lights illuminating the tower's truss structure?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera swirl to reveal the cityscape below with a sea of building lights?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04867033287882805 + ], + "opposite_logprob_tokens": [ + -3.0486702919006348 + ], + "logprob_reward": -0.04867033287882805, + "yes_no_margin_reward": 3.0, + "token_credit_reward": -0.04867033287882805, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15256255865097046, + 0.2743736803531647, + 0.12804760038852692, + 0.12763987481594086, + 0.15617872774600983, + 0.07224307209253311, + 0.015238801948726177, + 0.021369105204939842, + 0.015384153462946415, + 0.036962445825338364 + ], + "credit_map": [ + 0.015279977582395077, + 0.008201486431062222, + 0.03498539328575134, + 0.00842667929828167, + 0.009378761984407902, + 0.019803058356046677, + 0.00843262579292059, + 0.04805457964539528, + 0.018308574333786964, + 0.018547622486948967, + 0.05201336368918419, + 0.013098626397550106, + 0.008141146041452885, + 0.0595463328063488, + 0.010804149322211742, + 0.09391385316848755, + 0.013830103911459446, + 0.01660737954080105, + 0.017627298831939697, + 0.016153259202837944, + 0.007534828968346119, + 0.018089523538947105, + 0.013017958961427212, + 0.025187253952026367, + 0.018243219703435898, + 0.016733745113015175, + 0.025010868906974792, + 0.011915285140275955, + 0.013036109507083893, + 0.01601751707494259, + 0.011127362959086895, + 0.015555759891867638, + 0.037407372146844864, + 0.01948503777384758, + 0.019224433228373528, + 0.014977678656578064, + 0.01389350276440382, + 0.02393658086657524, + 0.011744276620447636, + 0.01550985872745514, + 0.008532445877790451, + 0.01884494349360466, + 0.0067036887630820274, + 0.005669482983648777, + 0.007682379335165024, + 0.007223661057651043, + 0.0071943397633731365, + 0.010392123833298683, + 0.0024946746416389942, + 0.0016943729715421796, + 0.0015997927403077483, + 0.0021889214403927326, + 0.0014688019873574376, + 0.002109343418851495, + 0.001720234751701355, + 0.0019626584835350513, + 0.002148826140910387, + 0.0018173796124756336, + 0.002486636396497488, + 0.002070477930828929, + 0.002551598474383354, + 0.0037915739230811596, + 0.0031381987500935793, + 0.003364414907991886, + 0.0013211965560913086, + 0.0014904250856488943, + 0.001876671682111919, + 0.0015458723064512014, + 0.001737741637043655, + 0.0026955658104270697, + 0.001643131603486836, + 0.003073547501116991, + 0.006892424542456865, + 0.004448224790394306, + 0.003502788720652461, + 0.004261703230440617, + 0.005706888623535633, + 0.004836094565689564, + 0.003885835176333785, + 0.0034284868743270636 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 4, + 0 + ], + "selected_reward": -0.04867033287882805 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject Tokyo Tower visible in the video?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the tower's structure colored in international orange-red and white steel?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera slowly and gracefully ascend around the tower?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are thousands of warm golden lights illuminating the tower's truss structure?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera swirl to reveal the cityscape below with a sea of building lights?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02633957751095295 + ], + "opposite_logprob_tokens": [ + -3.651339530944824 + ], + "logprob_reward": -0.02633957751095295, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.02633957751095295, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13137716054916382, + 0.10259917378425598, + 0.16556064784526825, + 0.0871414989233017, + 0.22451402246952057, + 0.13921228051185608, + 0.056722771376371384, + 0.020752979442477226, + 0.031079676002264023, + 0.041039738804101944 + ], + "credit_map": [ + 0.012027263641357422, + 0.009948273189365864, + 0.02192535810172558, + 0.007428835611790419, + 0.012491658329963684, + 0.02060098387300968, + 0.007422565016895533, + 0.03953223675489426, + 0.008288474753499031, + 0.006573524326086044, + 0.026197578758001328, + 0.006364605389535427, + 0.003863218706101179, + 0.03328009322285652, + 0.004822643008083105, + 0.013209043070673943, + 0.020814374089241028, + 0.021111220121383667, + 0.02356688492000103, + 0.02518385462462902, + 0.012298421934247017, + 0.025521378964185715, + 0.01433311216533184, + 0.02273140661418438, + 0.010471705347299576, + 0.014680112712085247, + 0.010829747654497623, + 0.009048664942383766, + 0.009896191768348217, + 0.0104981679469347, + 0.009536859579384327, + 0.012180056422948837, + 0.03881160914897919, + 0.026486096903681755, + 0.025494927540421486, + 0.033469583839178085, + 0.023523995652794838, + 0.02859688550233841, + 0.020416200160980225, + 0.027714740484952927, + 0.017345314845442772, + 0.018820330500602722, + 0.012160512618720531, + 0.010863350704312325, + 0.012503400444984436, + 0.025649406015872955, + 0.011941313743591309, + 0.029928652569651604, + 0.006777892354875803, + 0.005603833124041557, + 0.004760308191180229, + 0.005918364506214857, + 0.006120257545262575, + 0.013671108521521091, + 0.004621519241482019, + 0.009249492548406124, + 0.0014490106841549277, + 0.0012016292894259095, + 0.0035420842468738556, + 0.001898077200166881, + 0.002062611747533083, + 0.00527331093326211, + 0.0020394178573042154, + 0.003286838997155428, + 0.0041707176715135574, + 0.0018360811518505216, + 0.004832794424146414, + 0.003975271712988615, + 0.0029126314911991358, + 0.005016247741878033, + 0.0033733630552887917, + 0.0049625723622739315, + 0.0044688573107123375, + 0.003274668473750353, + 0.004457779694348574, + 0.005306177772581577, + 0.005615083500742912, + 0.006605185102671385, + 0.0035938925575464964, + 0.007718102540820837 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 2, + 5 + ], + "selected_reward": -0.02633957751095295 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject Tokyo Tower visible in the video?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the tower's structure colored in international orange-red and white steel?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera slowly and gracefully ascend around the tower?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are thousands of warm golden lights illuminating the tower's truss structure?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera swirl to reveal the cityscape below with a sea of building lights?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0019378946162760258 + ], + "opposite_logprob_tokens": [ + -6.2519378662109375 + ], + "logprob_reward": -0.0019378946162760258, + "yes_no_margin_reward": 6.25, + "token_credit_reward": -0.0019378946162760258, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.21119864284992218, + 0.1808166354894638, + 0.12755133211612701, + 0.0887434333562851, + 0.14122173190116882, + 0.11176926642656326, + 0.02479417249560356, + 0.018165798857808113, + 0.02995750494301319, + 0.06578144431114197 + ], + "credit_map": [ + 0.02025829814374447, + 0.005843274295330048, + 0.07258594036102295, + 0.0066678691655397415, + 0.005620892625302076, + 0.018656164407730103, + 0.006359785329550505, + 0.07520643621683121, + 0.010308117605745792, + 0.004334187600761652, + 0.01155293919146061, + 0.0037826949264854193, + 0.0032403001096099615, + 0.0758804976940155, + 0.00417136587202549, + 0.067546546459198, + 0.015094107948243618, + 0.010777401737868786, + 0.018255600705742836, + 0.022310663014650345, + 0.006546645425260067, + 0.018879735842347145, + 0.01535699050873518, + 0.020330194383859634, + 0.010813453234732151, + 0.009071210399270058, + 0.018812980502843857, + 0.010515532456338406, + 0.00848847534507513, + 0.010964451357722282, + 0.01016607042402029, + 0.009911262430250645, + 0.026840656995773315, + 0.018712958320975304, + 0.0169659573584795, + 0.01958351582288742, + 0.01215349417179823, + 0.01872961036860943, + 0.010292296297848225, + 0.017943238839507103, + 0.01650998555123806, + 0.026472648605704308, + 0.010233297944068909, + 0.010879551991820335, + 0.01061408594250679, + 0.016538821160793304, + 0.010108095593750477, + 0.010412781499326229, + 0.0025121343787759542, + 0.0022776771802455187, + 0.0027680506464093924, + 0.0023049574811011553, + 0.003092970000579953, + 0.005170092452317476, + 0.003115519881248474, + 0.003552772104740143, + 0.001600953284651041, + 0.0012391182826831937, + 0.002778230234980583, + 0.0018726527923718095, + 0.0016517434269189835, + 0.004735102411359549, + 0.0014582929434254766, + 0.0028297046665102243, + 0.0016457000747323036, + 0.0013066537212580442, + 0.004103959538042545, + 0.0036989301443099976, + 0.002665000967681408, + 0.009258339181542397, + 0.0022339385468512774, + 0.005044983699917793, + 0.00702640088275075, + 0.004577194806188345, + 0.00775102199986577, + 0.007021329831331968, + 0.012177936732769012, + 0.01001481432467699, + 0.0057233963161706924, + 0.01148934569209814 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 4 + ], + "selected_reward": -0.0019378946162760258 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject Tokyo Tower visible in the video?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the tower's structure colored in international orange-red and white steel?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera slowly and gracefully ascend around the tower?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are thousands of warm golden lights illuminating the tower's truss structure?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera swirl to reveal the cityscape below with a sea of building lights?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0017130000051110983 + ], + "opposite_logprob_tokens": [ + -6.376712799072266 + ], + "logprob_reward": -0.0017130000051110983, + "yes_no_margin_reward": 6.375, + "token_credit_reward": -0.0017130000051110983, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11078307032585144, + 0.09025391191244125, + 0.07429230958223343, + 0.1124480739235878, + 0.18810202181339264, + 0.215153768658638, + 0.06194575875997543, + 0.04486031457781792, + 0.04732341691851616, + 0.05483737960457802 + ], + "credit_map": [ + 0.015378531068563461, + 0.006789324339479208, + 0.03121660277247429, + 0.005809735041111708, + 0.006447942461818457, + 0.010827588848769665, + 0.005119497422128916, + 0.02919384278357029, + 0.011202161200344563, + 0.006625154055655003, + 0.010848091915249825, + 0.004354648757725954, + 0.003037302289158106, + 0.025430044159293175, + 0.0049134306609630585, + 0.02384306490421295, + 0.011159764602780342, + 0.004804553464055061, + 0.010831593535840511, + 0.007429424673318863, + 0.0031539564952254295, + 0.011330892331898212, + 0.0058572483249008656, + 0.01972486451268196, + 0.013954009860754013, + 0.015248941257596016, + 0.022683681920170784, + 0.007506511639803648, + 0.010680250823497772, + 0.01625712402164936, + 0.009384437464177608, + 0.01673310250043869, + 0.03186553344130516, + 0.029634486883878708, + 0.020200999453663826, + 0.03286007046699524, + 0.0170208141207695, + 0.023136558011174202, + 0.013572243973612785, + 0.019811304286122322, + 0.02427760139107704, + 0.04153130203485489, + 0.02045617625117302, + 0.013450106605887413, + 0.01741190068423748, + 0.0370306633412838, + 0.017192305997014046, + 0.043803710490465164, + 0.007100986782461405, + 0.00569445313885808, + 0.005501000676304102, + 0.005955556407570839, + 0.005427794996649027, + 0.014340049587190151, + 0.0047515076585114, + 0.013174405321478844, + 0.00259567704051733, + 0.002058563753962517, + 0.005721804685890675, + 0.003665495663881302, + 0.005792274139821529, + 0.015123485587537289, + 0.002917804289609194, + 0.006985203363001347, + 0.004606991074979305, + 0.00292101688683033, + 0.010740109719336033, + 0.00442010723054409, + 0.0037200236693024635, + 0.008966161869466305, + 0.0047585065476596355, + 0.007190499920397997, + 0.010874244384467602, + 0.005466946400702, + 0.005482370033860207, + 0.00688139209523797, + 0.0065838941372931, + 0.006483475212007761, + 0.006172938272356987, + 0.006892113480716944 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 4, + 3 + ], + "selected_reward": -0.0017130000051110983 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "fluid drone shot, captured in Annie Leibovitz's signature dramatic style, slowly and gracefully ascends around Tokyo Tower. High-contrast lighting sculpts the tower's iconic international orange-red and white steel structure against a deep, midnight-blue velvety backdrop, each beam and support clearly visible. Thousands of warm golden lights illuminate its intricate truss structure, casting deep, layered shadows, while the signal lights at the spire rhythmically flicker a solitary, proud red. As the camera swirls, it takes in the vast expanse of Tokyo's cityscape below. The lights of countless buildings converge into a dazzling sea of light, while the shimmering streets pulsate like luminous arteries through the city's fabric. The rich, saturated tones and dramatic composition of light and shadow create an atmosphere both majestic and tinged with loneliness, as if capturing a powerful and emotional portrait of this iconic giant of a city that never sleeps.", + "seed": 4159, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000027/group_15/index_921_seed_4159_dp_31_rank_62_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 15, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.0607784176710993, + "vlm_logprob": -0.0607784176710993, + "vlm_yes_no_margin": 3.95, + "vlm_token_credit": -0.0607784176710993, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject Tokyo Tower visible in the video?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the tower's structure colored in international orange-red and white steel?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera slowly and gracefully ascend around the tower?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are thousands of warm golden lights illuminating the tower's truss structure?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera swirl to reveal the cityscape below with a sea of building lights?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.22678548097610474 + ], + "opposite_logprob_tokens": [ + -1.60178542137146 + ], + "logprob_reward": -0.22678548097610474, + "yes_no_margin_reward": 1.375, + "token_credit_reward": -0.22678548097610474, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.4653700590133667, + 0.07676471769809723, + 0.16132420301437378, + 0.028163274750113487, + 0.03999504819512367, + 0.06179530546069145, + 0.04138291999697685, + 0.028139665722846985, + 0.057708531618118286, + 0.039356306195259094 + ], + "credit_map": [ + 0.052657272666692734, + 0.05106061324477196, + 0.07592511177062988, + 0.04519268870353699, + 0.07768969982862473, + 0.06172795966267586, + 0.05484766513109207, + 0.04626903682947159, + 0.010097942315042019, + 0.007407444063574076, + 0.011945150792598724, + 0.010187980718910694, + 0.009370527230203152, + 0.008352602832019329, + 0.009759398177266121, + 0.009643662720918655, + 0.02234754152595997, + 0.018357766792178154, + 0.025788813829421997, + 0.019550085067749023, + 0.016553519293665886, + 0.016840536147356033, + 0.02428269572556019, + 0.01760323904454708, + 0.0032149204052984715, + 0.004352578427642584, + 0.004817185923457146, + 0.0023960527032613754, + 0.004036816768348217, + 0.003976134583353996, + 0.0034292880445718765, + 0.0019402968464419246, + 0.0059249307960271835, + 0.0031588550191372633, + 0.008975738659501076, + 0.004704951774328947, + 0.006145052146166563, + 0.002992370631545782, + 0.00564334262162447, + 0.002449804451316595, + 0.007202967535704374, + 0.004889713600277901, + 0.009570101276040077, + 0.009246142581105232, + 0.0083612697198987, + 0.004922604653984308, + 0.0077520376071333885, + 0.009850463829934597, + 0.007276095915585756, + 0.0051442296244204044, + 0.01208674255758524, + 0.0036142184399068356, + 0.0034397223498672247, + 0.00359158287756145, + 0.0031803715974092484, + 0.003049951046705246, + 0.0035349796526134014, + 0.005222869571298361, + 0.003569039748981595, + 0.0019347647903487086, + 0.0027716385666280985, + 0.003772565396502614, + 0.005287121050059795, + 0.0020466861315071583, + 0.00879422202706337, + 0.006593877915292978, + 0.0069139921106398106, + 0.005170801654458046, + 0.0053672282956540585, + 0.007379215210676193, + 0.010914080776274204, + 0.006575106177479029, + 0.004602893255650997, + 0.008466136641800404, + 0.0031394269317388535, + 0.0023120855912566185, + 0.006455313414335251, + 0.003706886200234294, + 0.00877920538187027, + 0.0018943516770377755 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 1 + ], + "selected_reward": -0.22678548097610474 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject Tokyo Tower visible in the video?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the tower's structure colored in international orange-red and white steel?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera slowly and gracefully ascend around the tower?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are thousands of warm golden lights illuminating the tower's truss structure?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera swirl to reveal the cityscape below with a sea of building lights?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.054996270686388016 + ], + "opposite_logprob_tokens": [ + -2.9299962520599365 + ], + "logprob_reward": -0.054996270686388016, + "yes_no_margin_reward": 2.875, + "token_credit_reward": -0.054996270686388016, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.23132051527500153, + 0.11672090739011765, + 0.09486424177885056, + 0.050785306841135025, + 0.04028802365064621, + 0.13907577097415924, + 0.027295907959342003, + 0.06371660530567169, + 0.18349897861480713, + 0.05243375524878502 + ], + "credit_map": [ + 0.04088786616921425, + 0.02094889245927334, + 0.04725725203752518, + 0.019351424649357796, + 0.03054674342274666, + 0.021439827978610992, + 0.02960643731057644, + 0.021282074972987175, + 0.01572335511445999, + 0.012996888719499111, + 0.017793256789445877, + 0.011036965996026993, + 0.015843121334910393, + 0.011789066717028618, + 0.015109769068658352, + 0.016428491100668907, + 0.011268292553722858, + 0.011194033548235893, + 0.01460753008723259, + 0.00817583967000246, + 0.014220298267900944, + 0.013348965905606747, + 0.01129798125475645, + 0.010751297697424889, + 0.007655688561499119, + 0.0050463005900382996, + 0.009089608676731586, + 0.007658065762370825, + 0.00514585617929697, + 0.005758003797382116, + 0.0060502914711833, + 0.004381493665277958, + 0.00522274523973465, + 0.004495608154684305, + 0.005533741787075996, + 0.004964796826243401, + 0.00660276273265481, + 0.003877174574881792, + 0.005910858046263456, + 0.0036803395487368107, + 0.019945643842220306, + 0.024067608639597893, + 0.01793615147471428, + 0.010301893576979637, + 0.032454561442136765, + 0.01524188183248043, + 0.011641595512628555, + 0.007486438378691673, + 0.004595139995217323, + 0.003602034645155072, + 0.008187799714505672, + 0.0019041498890146613, + 0.002012006239965558, + 0.002406161045655608, + 0.00244913506321609, + 0.002139480784535408, + 0.008864806033670902, + 0.01606561988592148, + 0.006548665929585695, + 0.004187757149338722, + 0.007593739777803421, + 0.008414898999035358, + 0.008059064857661724, + 0.003982050810009241, + 0.02442934736609459, + 0.02022210694849491, + 0.017993956804275513, + 0.015527966432273388, + 0.01918412558734417, + 0.023917974904179573, + 0.03307344764471054, + 0.029150044545531273, + 0.005539994686841965, + 0.008974495343863964, + 0.004669482819736004, + 0.003514990210533142, + 0.009383238852024078, + 0.004794498439878225, + 0.012311960570514202, + 0.0032450950238853693 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 5 + ], + "selected_reward": -0.054996270686388016 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject Tokyo Tower visible in the video?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the tower's structure colored in international orange-red and white steel?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera slowly and gracefully ascend around the tower?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are thousands of warm golden lights illuminating the tower's truss structure?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera swirl to reveal the cityscape below with a sea of building lights?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.00762906763702631 + ], + "opposite_logprob_tokens": [ + -4.882628917694092 + ], + "logprob_reward": -0.00762906763702631, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.00762906763702631, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.5114255547523499, + 0.060349129140377045, + 0.09052442014217377, + 0.03436971828341484, + 0.049153611063957214, + 0.10891997814178467, + 0.040010370314121246, + 0.028522301465272903, + 0.04230091720819473, + 0.034423962235450745 + ], + "credit_map": [ + 0.17095737159252167, + 0.022360021248459816, + 0.17519864439964294, + 0.025107450783252716, + 0.03849288076162338, + 0.028655508533120155, + 0.02889840677380562, + 0.02175532467663288, + 0.008793172426521778, + 0.006194720976054668, + 0.009693547151982784, + 0.0062652998603880405, + 0.005876472685486078, + 0.0065058209002017975, + 0.008724112063646317, + 0.008295981213450432, + 0.015523682348430157, + 0.01109866052865982, + 0.01607416942715645, + 0.005808901973068714, + 0.011299576610326767, + 0.007883426733314991, + 0.014793474227190018, + 0.008042527362704277, + 0.006072196178138256, + 0.0038984743878245354, + 0.005415549501776695, + 0.003714929101988673, + 0.0030470637138932943, + 0.003526334883645177, + 0.0054867020808160305, + 0.0032084686681628227, + 0.005983256734907627, + 0.007733217440545559, + 0.005715149454772472, + 0.005850683897733688, + 0.006753298453986645, + 0.005371340084820986, + 0.008220222778618336, + 0.0035264422185719013, + 0.017412273213267326, + 0.019023308530449867, + 0.0157330222427845, + 0.008149757981300354, + 0.022130373865365982, + 0.009767178446054459, + 0.010352659970521927, + 0.006351406220346689, + 0.005520917009562254, + 0.0046294028870761395, + 0.007650827523320913, + 0.003116212785243988, + 0.006176222115755081, + 0.004788573831319809, + 0.004069601185619831, + 0.004058613441884518, + 0.003983525093644857, + 0.006004880648106337, + 0.002867072820663452, + 0.0019748136401176453, + 0.003840148216113448, + 0.0032690565567463636, + 0.004494866821914911, + 0.0020879379007965326, + 0.005598805844783783, + 0.005694243125617504, + 0.005501273088157177, + 0.004180336836725473, + 0.004317756742238998, + 0.00471549341455102, + 0.0074114021845161915, + 0.00488160690292716, + 0.003971559461206198, + 0.00619878014549613, + 0.003977361600846052, + 0.0021991615649312735, + 0.005433659069240093, + 0.0035968192387372255, + 0.006331305485218763, + 0.0027153135742992163 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 5, + 2 + ], + "selected_reward": -0.00762906763702631 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject Tokyo Tower visible in the video?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the tower's structure colored in international orange-red and white steel?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera slowly and gracefully ascend around the tower?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are thousands of warm golden lights illuminating the tower's truss structure?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera swirl to reveal the cityscape below with a sea of building lights?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0019394413102418184 + ], + "opposite_logprob_tokens": [ + -6.251939296722412 + ], + "logprob_reward": -0.0019394413102418184, + "yes_no_margin_reward": 6.25, + "token_credit_reward": -0.0019394413102418184, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.4485478103160858, + 0.0849764496088028, + 0.14714807271957397, + 0.023625187575817108, + 0.04272773116827011, + 0.047754764556884766, + 0.045355621725320816, + 0.03230234235525131, + 0.09104082733392715, + 0.03652117773890495 + ], + "credit_map": [ + 0.17038559913635254, + 0.013994495384395123, + 0.174515038728714, + 0.01767583191394806, + 0.029465973377227783, + 0.01081341877579689, + 0.019431402906775475, + 0.012266064062714577, + 0.010618310421705246, + 0.009510406292974949, + 0.013419664464890957, + 0.008571715094149113, + 0.010618399828672409, + 0.008945142850279808, + 0.012216849252581596, + 0.011075966991484165, + 0.01758400909602642, + 0.010554766282439232, + 0.023231755942106247, + 0.012808702886104584, + 0.01816357858479023, + 0.01791500486433506, + 0.03136950731277466, + 0.015520752407610416, + 0.0028710616752505302, + 0.002600736916065216, + 0.0036333445459604263, + 0.0037645648699253798, + 0.002125590341165662, + 0.0028797052800655365, + 0.002433467423543334, + 0.0033167190849781036, + 0.006937770638614893, + 0.004526892676949501, + 0.008770340122282505, + 0.0033903084695339203, + 0.00584745965898037, + 0.0038325439672917128, + 0.006085289642214775, + 0.003337128786370158, + 0.007640640251338482, + 0.008648591116070747, + 0.008406230248510838, + 0.003441260429099202, + 0.006711839698255062, + 0.003515789518132806, + 0.005100842099636793, + 0.004289576318114996, + 0.007741694804280996, + 0.00569193996489048, + 0.013555175624787807, + 0.0037072107661515474, + 0.003557016374543309, + 0.0033665515948086977, + 0.0031702311243861914, + 0.004565801937133074, + 0.0044954014010727406, + 0.005774862598627806, + 0.003463060362264514, + 0.002223749877884984, + 0.004299636464565992, + 0.004904765170067549, + 0.004622369073331356, + 0.0025184997357428074, + 0.01098052877932787, + 0.010404369793832302, + 0.009802460670471191, + 0.01031805481761694, + 0.00707605667412281, + 0.009924503974616528, + 0.01667356863617897, + 0.015861287713050842, + 0.004100664518773556, + 0.00458889314904809, + 0.0033489710185676813, + 0.002604949288070202, + 0.0039986502379179, + 0.0036101487930864096, + 0.011813552118837833, + 0.0024553497787564993 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 8 + ], + "selected_reward": -0.0019394413102418184 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject Tokyo Tower visible in the video?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the tower's structure colored in international orange-red and white steel?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera slowly and gracefully ascend around the tower?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are thousands of warm golden lights illuminating the tower's truss structure?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera swirl to reveal the cityscape below with a sea of building lights?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.012541827745735645 + ], + "opposite_logprob_tokens": [ + -4.387541770935059 + ], + "logprob_reward": -0.012541827745735645, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.012541827745735645, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1720404475927353, + 0.032513901591300964, + 0.07600966840982437, + 0.030696094036102295, + 0.04706139117479324, + 0.35882440209388733, + 0.2038385272026062, + 0.02465960942208767, + 0.03609085455536842, + 0.01826510950922966 + ], + "credit_map": [ + 0.05148111656308174, + 0.01070266030728817, + 0.05251814052462578, + 0.010939901694655418, + 0.013979953713715076, + 0.01262096967548132, + 0.009918960742652416, + 0.009878735058009624, + 0.005982760339975357, + 0.0032076274510473013, + 0.005017091054469347, + 0.0028421434108167887, + 0.003774339333176613, + 0.0037378857377916574, + 0.003938200883567333, + 0.004013854078948498, + 0.012110810726881027, + 0.006156210787594318, + 0.012804037891328335, + 0.008791609667241573, + 0.007507431786507368, + 0.00855162087827921, + 0.011915607377886772, + 0.008172334171831608, + 0.0062309917993843555, + 0.0031497320160269737, + 0.006880817003548145, + 0.002550677862018347, + 0.0026111463084816933, + 0.004224626813083887, + 0.0027122467290610075, + 0.002335853176191449, + 0.007815881632268429, + 0.005247669294476509, + 0.01219114102423191, + 0.005941839423030615, + 0.003970058169215918, + 0.0033570658415555954, + 0.005966791417449713, + 0.0025709406472742558, + 0.057240650057792664, + 0.05960051342844963, + 0.06385902315378189, + 0.03266901895403862, + 0.06161247193813324, + 0.03169642761349678, + 0.027250170707702637, + 0.024896100163459778, + 0.03801848366856575, + 0.020348260179162025, + 0.07207253575325012, + 0.01668928936123848, + 0.011647101491689682, + 0.013064037077128887, + 0.013196518644690514, + 0.01880229450762272, + 0.0036394880153238773, + 0.004608022980391979, + 0.00326526234857738, + 0.002157611772418022, + 0.003136861138045788, + 0.002876098733395338, + 0.0030682827346026897, + 0.0019079801859334111, + 0.004368025343865156, + 0.007077877875417471, + 0.003678695298731327, + 0.0034714641515165567, + 0.005117469932883978, + 0.003733636811375618, + 0.004625345580279827, + 0.004018341191112995, + 0.0024741438683122396, + 0.0028774074744433165, + 0.0016146383713930845, + 0.0015233388403430581, + 0.0021359017118811607, + 0.002538618165999651, + 0.0035999238025397062, + 0.0015011350624263287 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 6, + 0 + ], + "selected_reward": -0.012541827745735645 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000027/group_2/index_tensor([6487, 6487])_group_2_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000027/group_2/index_tensor([6487, 6487])_group_2_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..f3e1032e79bb435cd9752ea7c51d897caba257c8 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000027/group_2/index_tensor([6487, 6487])_group_2_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "striking close-up closely tracks the head of an anthropomorphic scarab beetle, caught in the middle of an epic confrontation. The behemoth's entire body is clad in iridescent metallic armor shimmering in emerald, deep purple, and bronze, shimmering brilliantly in the dim light. The camera zooms in on the complex, imposing structure of its head, its compound eyes ablaze with a fiery orange glow, betraying a relentless will to fight. At that moment, its disproportionately large, weaponized forearm, like a battering ram, slams forward, colliding violently with the weapon of an equally massive opponent at the edge of the frame. The moment of impact sends sparks of dazzling cyan-blue energy flying like fireworks, the force of the impact shaking the camera. The cascading sparks become the primary light source, carving sharp highlights and deep shadows into the smooth, curved surface of the armor, perfectly illustrating the intense contrast between light and dark. The entire picture presents a unique \"Battle of Two Cities\" animation style. Each frame seems to be a carefully drawn tough oil painting. The character's contours are rough and clear, and there are clear transitions between color blocks, creating a tense, violent and powerful atmosphere of epic confrontation.", + "seed": 8818, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000027/group_2/index_6487_seed_8818_dp_4_rank_8_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 2, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.029096917249262333, + "vlm_logprob": -0.029096917249262333, + "vlm_yes_no_margin": 4.025, + "vlm_token_credit": -0.029096917249262333, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an anthropomorphic scarab beetle wearing iridescent metallic armor in emerald, deep purple, and bronze?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the beetle's compound eyes glow with a fiery orange light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the beetle slam its large weaponized forearm forward to collide with an opponent's weapon at the edge of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do sparks of cyan-blue energy fly from the impact point like fireworks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the visual style consistent with a 'Battle of Two Cities' animation featuring rough, clear contours and distinct color blocks like a tough oil painting?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005448373034596443 + ], + "opposite_logprob_tokens": [ + -5.505448341369629 + ], + "logprob_reward": -0.005448373034596443, + "yes_no_margin_reward": 5.5, + "token_credit_reward": -0.005448373034596443, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.16325916349887848, + 0.11178553104400635, + 0.10582852363586426, + 0.08634841442108154, + 0.05362195149064064, + 0.11934656649827957, + 0.0747210904955864, + 0.09192072600126266, + 0.05273250490427017, + 0.14043553173542023 + ], + "credit_map": [ + 0.01868940144777298, + 0.015394915826618671, + 0.017882665619254112, + 0.022326035425066948, + 0.01749780774116516, + 0.017913537099957466, + 0.035750433802604675, + 0.01780436560511589, + 0.0166596919298172, + 0.021561717614531517, + 0.006806906312704086, + 0.011691564694046974, + 0.02358913980424404, + 0.013483482412993908, + 0.007995447143912315, + 0.009997574612498283, + 0.009734898805618286, + 0.010143578052520752, + 0.008029278367757797, + 0.008629702031612396, + 0.009231393225491047, + 0.03388514369726181, + 0.016324728727340698, + 0.009849801659584045, + 0.008249204605817795, + 0.008527996949851513, + 0.012342705391347408, + 0.020528601482510567, + 0.00857282429933548, + 0.0066645448096096516, + 0.011575279757380486, + 0.009887256659567356, + 0.007149711716920137, + 0.006384414620697498, + 0.007421588059514761, + 0.0055507151409983635, + 0.00548073835670948, + 0.004980641882866621, + 0.007984590716660023, + 0.00866954680532217, + 0.018950797617435455, + 0.009869090281426907, + 0.01631069742143154, + 0.008347627706825733, + 0.027763385325670242, + 0.012413723394274712, + 0.009318066760897636, + 0.016373181715607643, + 0.006091061048209667, + 0.01521995384246111, + 0.010575946420431137, + 0.02080601267516613, + 0.005373332183808088, + 0.004344069864600897, + 0.007178802974522114, + 0.00513191195204854, + 0.008512618951499462, + 0.01244623214006424, + 0.009110070765018463, + 0.015363446436822414, + 0.009815272875130177, + 0.012737036682665348, + 0.01639683172106743, + 0.007539208512753248, + 0.005651287734508514, + 0.005322640761733055, + 0.005595516413450241, + 0.008425570093095303, + 0.00761870201677084, + 0.007490808144211769, + 0.0066664498299360275, + 0.005961528513580561, + 0.00703114690259099, + 0.00600280100479722, + 0.00548733863979578, + 0.00875503197312355, + 0.010419943369925022, + 0.01030516717582941, + 0.045209866017103195, + 0.04722423106431961 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 5 + ], + "selected_reward": -0.005448373034596443 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an anthropomorphic scarab beetle wearing iridescent metallic armor in emerald, deep purple, and bronze?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the beetle's compound eyes glow with a fiery orange light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the beetle slam its large weaponized forearm forward to collide with an opponent's weapon at the edge of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do sparks of cyan-blue energy fly from the impact point like fireworks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the visual style consistent with a 'Battle of Two Cities' animation featuring rough, clear contours and distinct color blocks like a tough oil painting?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04865682125091553 + ], + "opposite_logprob_tokens": [ + -3.048656940460205 + ], + "logprob_reward": -0.04865682125091553, + "yes_no_margin_reward": 3.0, + "token_credit_reward": -0.04865682125091553, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.34738031029701233, + 0.044230762869119644, + 0.13951021432876587, + 0.05101356655359268, + 0.06156828999519348, + 0.0628771111369133, + 0.02703256905078888, + 0.03694835305213928, + 0.009682265110313892, + 0.2197565734386444 + ], + "credit_map": [ + 0.03862495347857475, + 0.04265095293521881, + 0.03361432999372482, + 0.033937469124794006, + 0.052921194583177567, + 0.04131057485938072, + 0.06725180894136429, + 0.03706903010606766, + 0.005231441464275122, + 0.007388931233435869, + 0.00367095903493464, + 0.005199396517127752, + 0.0075792851857841015, + 0.0047993226908147335, + 0.006890362594276667, + 0.0034710653126239777, + 0.01167018711566925, + 0.014186208136379719, + 0.009094931185245514, + 0.011052444577217102, + 0.010217476636171341, + 0.054447147995233536, + 0.015400523319840431, + 0.013441295363008976, + 0.004653092473745346, + 0.008507396094501019, + 0.004593813791871071, + 0.0065619912929832935, + 0.006182522047311068, + 0.008363251574337482, + 0.00553111732006073, + 0.006620380096137524, + 0.0029561284463852644, + 0.004015408921986818, + 0.005284818820655346, + 0.006225088611245155, + 0.007850409485399723, + 0.0069313058629632, + 0.019483285024762154, + 0.008821841329336166, + 0.008860434405505657, + 0.0052587976679205894, + 0.005624864716082811, + 0.005039990413933992, + 0.01569851115345955, + 0.01001342386007309, + 0.003917982801795006, + 0.008463109843432903, + 0.0037517109885811806, + 0.007302936166524887, + 0.003726864466443658, + 0.0030365376733243465, + 0.0026653714012354612, + 0.0022157144267112017, + 0.0024064970202744007, + 0.0019269352778792381, + 0.0035891220904886723, + 0.00862093921750784, + 0.0033072035294026136, + 0.004990759771317244, + 0.003471149131655693, + 0.004431452602148056, + 0.005038411356508732, + 0.0034993155859410763, + 0.001018164330162108, + 0.001337811117991805, + 0.0011555522214621305, + 0.001338373520411551, + 0.0012946235947310925, + 0.0012252526357769966, + 0.001200550701469183, + 0.0011119364062324166, + 0.007906796410679817, + 0.004717699252068996, + 0.0027749331202358007, + 0.01203212607651949, + 0.008449737913906574, + 0.008547313511371613, + 0.09126518666744232, + 0.08406277745962143 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 2 + ], + "selected_reward": -0.04865682125091553 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an anthropomorphic scarab beetle wearing iridescent metallic armor in emerald, deep purple, and bronze?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the beetle's compound eyes glow with a fiery orange light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the beetle slam its large weaponized forearm forward to collide with an opponent's weapon at the edge of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do sparks of cyan-blue energy fly from the impact point like fireworks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the visual style consistent with a 'Battle of Two Cities' animation featuring rough, clear contours and distinct color blocks like a tough oil painting?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03810028359293938 + ], + "opposite_logprob_tokens": [ + -3.288100242614746 + ], + "logprob_reward": -0.03810028359293938, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.03810028359293938, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12811033427715302, + 0.08799882233142853, + 0.2378712147474289, + 0.051675740629434586, + 0.06436886638402939, + 0.12018139660358429, + 0.11548254638910294, + 0.08330778032541275, + 0.049949515610933304, + 0.0610538087785244 + ], + "credit_map": [ + 0.015550365671515465, + 0.015988433733582497, + 0.014759711921215057, + 0.01314691174775362, + 0.012743611820042133, + 0.015559492632746696, + 0.021880151703953743, + 0.018481643870472908, + 0.013695984147489071, + 0.01289079338312149, + 0.009937705472111702, + 0.008818186819553375, + 0.01218857616186142, + 0.010846471413969994, + 0.009634476155042648, + 0.009986626915633678, + 0.022029737010598183, + 0.01984386332333088, + 0.014612813480198383, + 0.020204003900289536, + 0.020527882501482964, + 0.0846359059214592, + 0.03510267287492752, + 0.02091434970498085, + 0.005296036135405302, + 0.004984419327229261, + 0.00877787172794342, + 0.006788584403693676, + 0.0055590127594769, + 0.005538595374673605, + 0.0069751194678246975, + 0.007756102364510298, + 0.004747891332954168, + 0.004892515949904919, + 0.01101913396269083, + 0.007931840606033802, + 0.009757951833307743, + 0.005559930577874184, + 0.01288637612015009, + 0.007573230657726526, + 0.015564228408038616, + 0.01570712961256504, + 0.013697405345737934, + 0.011568360961973667, + 0.017465602606534958, + 0.018337825313210487, + 0.009899161756038666, + 0.017941690981388092, + 0.012373770587146282, + 0.022982299327850342, + 0.012930594384670258, + 0.016999393701553345, + 0.011539649218320847, + 0.01173421274870634, + 0.018666386604309082, + 0.008256236091256142, + 0.009086078964173794, + 0.011476882733404636, + 0.010273457504808903, + 0.012657792307436466, + 0.008632264100015163, + 0.011719577945768833, + 0.00906710047274828, + 0.010394630022346973, + 0.005558033939450979, + 0.00652556074783206, + 0.0061644562520086765, + 0.006224177777767181, + 0.005675280466675758, + 0.006385161075741053, + 0.006465993355959654, + 0.006950848735868931, + 0.004895050078630447, + 0.007050869055092335, + 0.004886267706751823, + 0.009546556510031223, + 0.004445259924978018, + 0.005713068880140781, + 0.013335960917174816, + 0.011180773377418518 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 0, + 5 + ], + "selected_reward": -0.03810028359293938 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an anthropomorphic scarab beetle wearing iridescent metallic armor in emerald, deep purple, and bronze?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the beetle's compound eyes glow with a fiery orange light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the beetle slam its large weaponized forearm forward to collide with an opponent's weapon at the edge of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do sparks of cyan-blue energy fly from the impact point like fireworks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the visual style consistent with a 'Battle of Two Cities' animation featuring rough, clear contours and distinct color blocks like a tough oil painting?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004633641801774502 + ], + "opposite_logprob_tokens": [ + -5.37963342666626 + ], + "logprob_reward": -0.004633641801774502, + "yes_no_margin_reward": 5.375, + "token_credit_reward": -0.004633641801774502, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14093448221683502, + 0.10832224041223526, + 0.08563172072172165, + 0.04414054751396179, + 0.06960806995630264, + 0.04186658561229706, + 0.13254213333129883, + 0.1653001308441162, + 0.02074383944272995, + 0.19091013073921204 + ], + "credit_map": [ + 0.012423356994986534, + 0.015407164581120014, + 0.010411138646304607, + 0.013609727844595909, + 0.018598025664687157, + 0.019035784527659416, + 0.03356613963842392, + 0.017883162945508957, + 0.012187950313091278, + 0.0199651587754488, + 0.007518543396145105, + 0.009131048806011677, + 0.028299804776906967, + 0.01428740844130516, + 0.007644160185009241, + 0.009288168512284756, + 0.006495628040283918, + 0.007002942264080048, + 0.004815327003598213, + 0.009965566918253899, + 0.0067378077656030655, + 0.03149585425853729, + 0.010409211739897728, + 0.008709391579031944, + 0.00539206899702549, + 0.004592548590153456, + 0.005673276726156473, + 0.007272142916917801, + 0.006029116455465555, + 0.0038623905275017023, + 0.005780154373496771, + 0.005538849625736475, + 0.00454636150971055, + 0.00917647872120142, + 0.012500149197876453, + 0.006979598663747311, + 0.006001794245094061, + 0.004978614393621683, + 0.009696848690509796, + 0.01572822593152523, + 0.004646266344934702, + 0.0057791671715676785, + 0.005945980548858643, + 0.004791934508830309, + 0.004652765113860369, + 0.005412188358604908, + 0.004059203900396824, + 0.006579083856195211, + 0.01672803983092308, + 0.027172762900590897, + 0.01603681780397892, + 0.021871943026781082, + 0.01377762109041214, + 0.011547683738172054, + 0.017913687974214554, + 0.007493600249290466, + 0.013541970402002335, + 0.02114146761596203, + 0.018703732639551163, + 0.021286267787218094, + 0.023120123893022537, + 0.026534948498010635, + 0.023823540657758713, + 0.017148099839687347, + 0.002076040953397751, + 0.0024248063564300537, + 0.0023320044856518507, + 0.00322256563231349, + 0.002422695280984044, + 0.002646499313414097, + 0.002758750459179282, + 0.0028604781255126, + 0.0072966753505170345, + 0.004088772460818291, + 0.003011730033904314, + 0.010316594503819942, + 0.007585127837955952, + 0.00728309853002429, + 0.07880212366580963, + 0.07252601534128189 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 7, + 0 + ], + "selected_reward": -0.004633641801774502 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an anthropomorphic scarab beetle wearing iridescent metallic armor in emerald, deep purple, and bronze?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the beetle's compound eyes glow with a fiery orange light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the beetle slam its large weaponized forearm forward to collide with an opponent's weapon at the edge of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do sparks of cyan-blue energy fly from the impact point like fireworks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the visual style consistent with a 'Battle of Two Cities' animation featuring rough, clear contours and distinct color blocks like a tough oil painting?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.048645466566085815 + ], + "opposite_logprob_tokens": [ + -3.048645496368408 + ], + "logprob_reward": -0.048645466566085815, + "yes_no_margin_reward": 3.0, + "token_credit_reward": -0.048645466566085815, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0951530784368515, + 0.10496705025434494, + 0.20762786269187927, + 0.07999458909034729, + 0.059561025351285934, + 0.057498496025800705, + 0.04799694940447807, + 0.08623725175857544, + 0.028887176886200905, + 0.23207654058933258 + ], + "credit_map": [ + 0.006232123356312513, + 0.008208444342017174, + 0.007932028733193874, + 0.00848213117569685, + 0.01084187626838684, + 0.011889061890542507, + 0.02939673140645027, + 0.012170687317848206, + 0.011727062053978443, + 0.014203192666172981, + 0.0069311331026256084, + 0.010558768175542355, + 0.024511460214853287, + 0.012430255301296711, + 0.01587553694844246, + 0.008729647845029831, + 0.016956361010670662, + 0.01469458732753992, + 0.01033763773739338, + 0.014657380990684032, + 0.01557388436049223, + 0.08723626285791397, + 0.027275962755084038, + 0.020895792171359062, + 0.009636945091187954, + 0.011487999930977821, + 0.010720609687268734, + 0.007152852602303028, + 0.008943487890064716, + 0.010071998462080956, + 0.01015259325504303, + 0.011828097514808178, + 0.002908129943534732, + 0.004597277846187353, + 0.006609347183257341, + 0.007104544434696436, + 0.007849437184631824, + 0.007542796898633242, + 0.017235832288861275, + 0.005713661201298237, + 0.011102810502052307, + 0.005376540124416351, + 0.006287633907049894, + 0.0038758893497288227, + 0.01410065870732069, + 0.005334123969078064, + 0.004445978906005621, + 0.0069748591631650925, + 0.004362510517239571, + 0.007988492958247662, + 0.006194767076522112, + 0.01010507345199585, + 0.005244243890047073, + 0.0032020972575992346, + 0.0057630096562206745, + 0.005136755760759115, + 0.009723236784338951, + 0.012648883275687695, + 0.008081779815256596, + 0.011876974254846573, + 0.01030727755278349, + 0.01203260850161314, + 0.011727401986718178, + 0.009839087724685669, + 0.0038483983371406794, + 0.003599383169785142, + 0.0034575711470097303, + 0.0030984333716332912, + 0.0032976677175611258, + 0.004117078147828579, + 0.003202731255441904, + 0.0042659142054617405, + 0.00838211178779602, + 0.004324383568018675, + 0.003198888385668397, + 0.012002871371805668, + 0.009107139892876148, + 0.008832993917167187, + 0.09721244871616364, + 0.08901571482419968 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 2, + 1 + ], + "selected_reward": -0.048645466566085815 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "striking close-up closely tracks the head of an anthropomorphic scarab beetle, caught in the middle of an epic confrontation. The behemoth's entire body is clad in iridescent metallic armor shimmering in emerald, deep purple, and bronze, shimmering brilliantly in the dim light. The camera zooms in on the complex, imposing structure of its head, its compound eyes ablaze with a fiery orange glow, betraying a relentless will to fight. At that moment, its disproportionately large, weaponized forearm, like a battering ram, slams forward, colliding violently with the weapon of an equally massive opponent at the edge of the frame. The moment of impact sends sparks of dazzling cyan-blue energy flying like fireworks, the force of the impact shaking the camera. The cascading sparks become the primary light source, carving sharp highlights and deep shadows into the smooth, curved surface of the armor, perfectly illustrating the intense contrast between light and dark. The entire picture presents a unique \"Battle of Two Cities\" animation style. Each frame seems to be a carefully drawn tough oil painting. The character's contours are rough and clear, and there are clear transitions between color blocks, creating a tense, violent and powerful atmosphere of epic confrontation.", + "seed": 8819, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000027/group_2/index_6487_seed_8819_dp_4_rank_8_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 2, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.03524780757725239, + "vlm_logprob": -0.03524780757725239, + "vlm_yes_no_margin": 3.875, + "vlm_token_credit": -0.03524780757725239, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an anthropomorphic scarab beetle wearing iridescent metallic armor in emerald, deep purple, and bronze?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the beetle's compound eyes glow with a fiery orange light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the beetle slam its large weaponized forearm forward to collide with an opponent's weapon at the edge of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do sparks of cyan-blue energy fly from the impact point like fireworks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the visual style consistent with a 'Battle of Two Cities' animation featuring rough, clear contours and distinct color blocks like a tough oil painting?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.009458491578698158 + ], + "opposite_logprob_tokens": [ + -4.759458541870117 + ], + "logprob_reward": -0.009458491578698158, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.009458491578698158, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15125006437301636, + 0.055069684982299805, + 0.04128839075565338, + 0.05008295178413391, + 0.16426055133342743, + 0.19834540784358978, + 0.07500394433736801, + 0.10471726208925247, + 0.10219975560903549, + 0.057781971991062164 + ], + "credit_map": [ + 0.018047917634248734, + 0.02206839807331562, + 0.02684297412633896, + 0.024456270039081573, + 0.013896113261580467, + 0.016400925815105438, + 0.01691245660185814, + 0.012625009752810001, + 0.005189416464418173, + 0.00600152974948287, + 0.004382438492029905, + 0.0061807893216609955, + 0.011650161817669868, + 0.005996068473905325, + 0.009486095979809761, + 0.006183183286339045, + 0.0034006682690232992, + 0.0034768718760460615, + 0.0035147671587765217, + 0.004692597780376673, + 0.006651011295616627, + 0.007047178223729134, + 0.006090768147259951, + 0.006414526607841253, + 0.0053824144415557384, + 0.003420555964112282, + 0.00913943536579609, + 0.006034386344254017, + 0.005417069420218468, + 0.006210301537066698, + 0.007527845911681652, + 0.006950941868126392, + 0.013153654523193836, + 0.014159547165036201, + 0.0226434338837862, + 0.01730269566178322, + 0.019133007153868675, + 0.031198961660265923, + 0.02481316775083542, + 0.021856077015399933, + 0.00889438483864069, + 0.009712735190987587, + 0.014232181943953037, + 0.01696787029504776, + 0.019467968493700027, + 0.06381484121084213, + 0.03285369649529457, + 0.03240171819925308, + 0.006200550124049187, + 0.008742091245949268, + 0.008015166968107224, + 0.012069344520568848, + 0.006351503077894449, + 0.010334562510251999, + 0.014172228053212166, + 0.009118504822254181, + 0.006762355100363493, + 0.009013193659484386, + 0.014481313526630402, + 0.02034207247197628, + 0.00936668086796999, + 0.007745692040771246, + 0.02510710619390011, + 0.011898848228156567, + 0.009447237476706505, + 0.018138641491532326, + 0.014801020734012127, + 0.0157551821321249, + 0.008293939754366875, + 0.010049420408904552, + 0.014060222543776035, + 0.011654086410999298, + 0.004439837299287319, + 0.006797424051910639, + 0.0036465306766331196, + 0.00922754593193531, + 0.006089397706091404, + 0.009323535487055779, + 0.00947490707039833, + 0.008782794699072838 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 4, + 0 + ], + "selected_reward": -0.009458491578698158 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an anthropomorphic scarab beetle wearing iridescent metallic armor in emerald, deep purple, and bronze?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the beetle's compound eyes glow with a fiery orange light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the beetle slam its large weaponized forearm forward to collide with an opponent's weapon at the edge of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do sparks of cyan-blue energy fly from the impact point like fireworks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the visual style consistent with a 'Battle of Two Cities' animation featuring rough, clear contours and distinct color blocks like a tough oil painting?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.011078868061304092 + ], + "opposite_logprob_tokens": [ + -4.511078834533691 + ], + "logprob_reward": -0.011078868061304092, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.011078868061304092, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0425071157515049, + 0.04179447889328003, + 0.05113952234387398, + 0.07445550709962845, + 0.058332979679107666, + 0.16368578374385834, + 0.05469822511076927, + 0.08256122469902039, + 0.3431773781776428, + 0.08764781057834625 + ], + "credit_map": [ + 0.005239557009190321, + 0.005643963813781738, + 0.00503602996468544, + 0.005240409169346094, + 0.00611545704305172, + 0.005520860198885202, + 0.0066163018345832825, + 0.0030945318285375834, + 0.003930381499230862, + 0.005908441264182329, + 0.004693685099482536, + 0.004369065165519714, + 0.0067876894026994705, + 0.005084762815386057, + 0.007972467690706253, + 0.0030479850247502327, + 0.004562526009976864, + 0.004567818250507116, + 0.0031303665600717068, + 0.005958305671811104, + 0.011863903142511845, + 0.006654446944594383, + 0.00786281656473875, + 0.006539337802678347, + 0.0059142811223864555, + 0.008134908974170685, + 0.008216563612222672, + 0.012486548162996769, + 0.008286983706057072, + 0.009535607881844044, + 0.01081831380724907, + 0.01106230542063713, + 0.003133204998448491, + 0.004371838644146919, + 0.004738435614854097, + 0.006383286323398352, + 0.011496656574308872, + 0.00614493852481246, + 0.014529012143611908, + 0.007535604294389486, + 0.0071536158211529255, + 0.008445034734904766, + 0.008535867556929588, + 0.01577131636440754, + 0.024104027077555656, + 0.037688300013542175, + 0.037438374012708664, + 0.024549230933189392, + 0.0048225438222289085, + 0.0040024276822805405, + 0.0061599682085216045, + 0.00921502523124218, + 0.004237078595906496, + 0.010045547969639301, + 0.008749398402869701, + 0.007466231007128954, + 0.007486519869416952, + 0.006438519340008497, + 0.013041086494922638, + 0.016336046159267426, + 0.009674947708845139, + 0.0073582567274570465, + 0.012139304541051388, + 0.010086538270115852, + 0.031148148700594902, + 0.06827246397733688, + 0.04739486426115036, + 0.05319121479988098, + 0.03978510946035385, + 0.02355305664241314, + 0.04232928529381752, + 0.03750322014093399, + 0.00775717431679368, + 0.011291204951703548, + 0.0065487758256495, + 0.020987864583730698, + 0.004654345568269491, + 0.008512698113918304, + 0.007531540002673864, + 0.020364202558994293 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 5, + 9 + ], + "selected_reward": -0.011078868061304092 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an anthropomorphic scarab beetle wearing iridescent metallic armor in emerald, deep purple, and bronze?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the beetle's compound eyes glow with a fiery orange light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the beetle slam its large weaponized forearm forward to collide with an opponent's weapon at the edge of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do sparks of cyan-blue energy fly from the impact point like fireworks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the visual style consistent with a 'Battle of Two Cities' animation featuring rough, clear contours and distinct color blocks like a tough oil painting?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0789700523018837 + ], + "opposite_logprob_tokens": [ + -2.578969955444336 + ], + "logprob_reward": -0.0789700523018837, + "yes_no_margin_reward": 2.5, + "token_credit_reward": -0.0789700523018837, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14395740628242493, + 0.07503970712423325, + 0.06190945953130722, + 0.1251782327890396, + 0.05335654318332672, + 0.20051221549510956, + 0.05206331983208656, + 0.05987987667322159, + 0.08251480758190155, + 0.14558838307857513 + ], + "credit_map": [ + 0.01729409396648407, + 0.018188636749982834, + 0.0254533588886261, + 0.020654745399951935, + 0.017097540199756622, + 0.01570378802716732, + 0.015740953385829926, + 0.013824275694787502, + 0.007556351367384195, + 0.007001269608736038, + 0.006423195358365774, + 0.006633677985519171, + 0.012698626145720482, + 0.008308680728077888, + 0.018449857831001282, + 0.007968046702444553, + 0.005873446818441153, + 0.006695651914924383, + 0.004763872362673283, + 0.006638083606958389, + 0.010608257725834846, + 0.009848921559751034, + 0.00944460742175579, + 0.008036619052290916, + 0.011293710209429264, + 0.011455243453383446, + 0.0138861034065485, + 0.020407192409038544, + 0.013409643433988094, + 0.013464066199958324, + 0.021953102201223373, + 0.01930917054414749, + 0.004486505873501301, + 0.004742263816297054, + 0.0075775072909891605, + 0.0068824756890535355, + 0.007071156054735184, + 0.00867792870849371, + 0.007376713212579489, + 0.0065419916063547134, + 0.01016903854906559, + 0.0117388516664505, + 0.014476281590759754, + 0.019661782309412956, + 0.030812855809926987, + 0.043955761939287186, + 0.040470436215400696, + 0.029227199032902718, + 0.005464320536702871, + 0.0064775655046105385, + 0.0066692279651761055, + 0.005693391431123018, + 0.004622458014637232, + 0.007656202185899019, + 0.00746163772419095, + 0.008018516935408115, + 0.005103955511003733, + 0.004303405527025461, + 0.009709715843200684, + 0.010007314383983612, + 0.006941511761397123, + 0.003984055016189814, + 0.010158161632716656, + 0.009671754203736782, + 0.00911264680325985, + 0.018915731459856033, + 0.008570428006350994, + 0.01296892762184143, + 0.00798400305211544, + 0.0072425659745931625, + 0.008476067334413528, + 0.00924443919211626, + 0.00871109589934349, + 0.013460281305015087, + 0.008663552813231945, + 0.037917789071798325, + 0.009674781933426857, + 0.013383157551288605, + 0.017703166231513023, + 0.03607455641031265 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 9, + 0 + ], + "selected_reward": -0.0789700523018837 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an anthropomorphic scarab beetle wearing iridescent metallic armor in emerald, deep purple, and bronze?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the beetle's compound eyes glow with a fiery orange light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the beetle slam its large weaponized forearm forward to collide with an opponent's weapon at the edge of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do sparks of cyan-blue energy fly from the impact point like fireworks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the visual style consistent with a 'Battle of Two Cities' animation featuring rough, clear contours and distinct color blocks like a tough oil painting?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.006730150431394577 + ], + "opposite_logprob_tokens": [ + -5.006730079650879 + ], + "logprob_reward": -0.006730150431394577, + "yes_no_margin_reward": 5.0, + "token_credit_reward": -0.006730150431394577, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2941609025001526, + 0.029323657974600792, + 0.07500984519720078, + 0.03643875941634178, + 0.04016687721014023, + 0.2056790441274643, + 0.0515856072306633, + 0.052342820912599564, + 0.0807080790400505, + 0.13458438217639923 + ], + "credit_map": [ + 0.03536185249686241, + 0.0349968858063221, + 0.04351499304175377, + 0.03041672334074974, + 0.0486585833132267, + 0.03245891258120537, + 0.042786821722984314, + 0.025966165587306023, + 0.0032849940471351147, + 0.0033900446724146605, + 0.003031392814591527, + 0.003211738308891654, + 0.0041610863991081715, + 0.0034441847819834948, + 0.004848932847380638, + 0.003951287362724543, + 0.006283248774707317, + 0.007585851475596428, + 0.005401494912803173, + 0.007455937098711729, + 0.016583988443017006, + 0.008493592962622643, + 0.011819915845990181, + 0.011385820806026459, + 0.0030926887411624193, + 0.0034980601631104946, + 0.0051692393608391285, + 0.004565069451928139, + 0.004772379528731108, + 0.0046243262477219105, + 0.005097493529319763, + 0.005619505420327187, + 0.0031533269211649895, + 0.0029485782142728567, + 0.004739093594253063, + 0.004147335421293974, + 0.006116827484220266, + 0.006745713762938976, + 0.006903498433530331, + 0.005412509199231863, + 0.009463486261665821, + 0.010391147807240486, + 0.010083870962262154, + 0.017573287710547447, + 0.030265023931860924, + 0.0519593320786953, + 0.044971588999032974, + 0.030971312895417213, + 0.004069304093718529, + 0.004133972339332104, + 0.005551887210458517, + 0.008257070556282997, + 0.003676413558423519, + 0.008766775950789452, + 0.008341719396412373, + 0.00878846738487482, + 0.0028971207793802023, + 0.0043542529456317425, + 0.012802963145077229, + 0.009784881956875324, + 0.004204325377941132, + 0.004003806971013546, + 0.007697728928178549, + 0.0065977443009614944, + 0.009314256720244884, + 0.020106367766857147, + 0.007592992391437292, + 0.012928830459713936, + 0.006539215799421072, + 0.008606935851275921, + 0.0073647997342050076, + 0.008254687301814556, + 0.007571497466415167, + 0.012924667447805405, + 0.0071482062339782715, + 0.04024975746870041, + 0.0063448501750826836, + 0.012347649782896042, + 0.011233517900109291, + 0.0367642380297184 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 5, + 9 + ], + "selected_reward": -0.006730150431394577 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an anthropomorphic scarab beetle wearing iridescent metallic armor in emerald, deep purple, and bronze?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the beetle's compound eyes glow with a fiery orange light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the beetle slam its large weaponized forearm forward to collide with an opponent's weapon at the edge of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do sparks of cyan-blue energy fly from the impact point like fireworks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the visual style consistent with a 'Battle of Two Cities' animation featuring rough, clear contours and distinct color blocks like a tough oil painting?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07000147551298141 + ], + "opposite_logprob_tokens": [ + -2.6950013637542725 + ], + "logprob_reward": -0.07000147551298141, + "yes_no_margin_reward": 2.625, + "token_credit_reward": -0.07000147551298141, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11173085123300552, + 0.05486246570944786, + 0.028695430606603622, + 0.059136491268873215, + 0.3680918514728546, + 0.20870298147201538, + 0.03483986482024193, + 0.045914456248283386, + 0.044463060796260834, + 0.04356257990002632 + ], + "credit_map": [ + 0.011713911779224873, + 0.013403068296611309, + 0.02217068150639534, + 0.014362329617142677, + 0.011469916440546513, + 0.013262063264846802, + 0.014409283176064491, + 0.010939596220850945, + 0.005930103827267885, + 0.008538381196558475, + 0.005691254045814276, + 0.007207351736724377, + 0.008003244176506996, + 0.0059948572888970375, + 0.007973292842507362, + 0.005523980595171452, + 0.002802102593705058, + 0.0032072621397674084, + 0.0029257473070174456, + 0.0026190057396888733, + 0.005121913738548756, + 0.004357482306659222, + 0.003914814442396164, + 0.0037471018731594086, + 0.0047961752861738205, + 0.005642799194902182, + 0.00977229606360197, + 0.008878007531166077, + 0.0058805085718631744, + 0.006808631122112274, + 0.00899870041757822, + 0.008359374478459358, + 0.021965427324175835, + 0.03121926449239254, + 0.026189105585217476, + 0.03319060057401657, + 0.08350907266139984, + 0.04153473302721977, + 0.08352990448474884, + 0.04695374146103859, + 0.0097250547260046, + 0.011185435578227043, + 0.010649474337697029, + 0.018707619979977608, + 0.030289625748991966, + 0.05096908286213875, + 0.045404259115457535, + 0.0317724347114563, + 0.0031241588294506073, + 0.0024145706556737423, + 0.0038835008163005114, + 0.004217751789838076, + 0.0037483754567801952, + 0.006720943376421928, + 0.006274879444390535, + 0.004455687012523413, + 0.0029171546921133995, + 0.0029517400544136763, + 0.004759952891618013, + 0.007783217821270227, + 0.005169345065951347, + 0.004510453436523676, + 0.010392026975750923, + 0.007430569268763065, + 0.003773235250264406, + 0.007028461899608374, + 0.004854200873523951, + 0.006412188522517681, + 0.005704402457922697, + 0.0041671860963106155, + 0.007660271599888802, + 0.004863115958869457, + 0.001950100064277649, + 0.003847443498671055, + 0.0024537728168070316, + 0.012827551923692226, + 0.002484292257577181, + 0.0037137512117624283, + 0.004232431761920452, + 0.012053238227963448 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 5, + 0 + ], + "selected_reward": -0.07000147551298141 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "striking close-up closely tracks the head of an anthropomorphic scarab beetle, caught in the middle of an epic confrontation. The behemoth's entire body is clad in iridescent metallic armor shimmering in emerald, deep purple, and bronze, shimmering brilliantly in the dim light. The camera zooms in on the complex, imposing structure of its head, its compound eyes ablaze with a fiery orange glow, betraying a relentless will to fight. At that moment, its disproportionately large, weaponized forearm, like a battering ram, slams forward, colliding violently with the weapon of an equally massive opponent at the edge of the frame. The moment of impact sends sparks of dazzling cyan-blue energy flying like fireworks, the force of the impact shaking the camera. The cascading sparks become the primary light source, carving sharp highlights and deep shadows into the smooth, curved surface of the armor, perfectly illustrating the intense contrast between light and dark. The entire picture presents a unique \"Battle of Two Cities\" animation style. Each frame seems to be a carefully drawn tough oil painting. The character's contours are rough and clear, and there are clear transitions between color blocks, creating a tense, violent and powerful atmosphere of epic confrontation.", + "seed": 8820, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000027/group_2/index_6487_seed_8820_dp_5_rank_10_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 2, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.03417492583394051, + "vlm_logprob": -0.03417492583394051, + "vlm_yes_no_margin": 4.15, + "vlm_token_credit": -0.03417492583394051, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an anthropomorphic scarab beetle wearing iridescent metallic armor in emerald, deep purple, and bronze?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the beetle's compound eyes glow with a fiery orange light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the beetle slam its large weaponized forearm forward to collide with an opponent's weapon at the edge of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do sparks of cyan-blue energy fly from the impact point like fireworks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the visual style consistent with a 'Battle of Two Cities' animation featuring rough, clear contours and distinct color blocks like a tough oil painting?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.003244614228606224 + ], + "opposite_logprob_tokens": [ + -6.128244400024414 + ], + "logprob_reward": -0.003244614228606224, + "yes_no_margin_reward": 6.125, + "token_credit_reward": -0.003244614228606224, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2258467972278595, + 0.23601795732975006, + 0.06542688608169556, + 0.06784958392381668, + 0.05437256023287773, + 0.07256180793046951, + 0.08966248482465744, + 0.050996363162994385, + 0.09155118465423584, + 0.04571440443396568 + ], + "credit_map": [ + 0.02868834137916565, + 0.0191081240773201, + 0.043601591140031815, + 0.02461511641740799, + 0.025403665378689766, + 0.030682174488902092, + 0.029292963445186615, + 0.02445482648909092, + 0.026480868458747864, + 0.029968062415719032, + 0.032926250249147415, + 0.02494640462100506, + 0.029664400964975357, + 0.028771797195076942, + 0.03129725158214569, + 0.031962912529706955, + 0.00555189186707139, + 0.007520510349422693, + 0.011732612736523151, + 0.00932331196963787, + 0.005244341678917408, + 0.00856045726686716, + 0.00800715945661068, + 0.009486597031354904, + 0.00959356315433979, + 0.00820200052112341, + 0.007223205640912056, + 0.008754817768931389, + 0.006954526528716087, + 0.00610585231333971, + 0.01046662125736475, + 0.010548985563218594, + 0.00476039107888937, + 0.006683229934424162, + 0.004520414397120476, + 0.00926205050200224, + 0.006329014431685209, + 0.009183255024254322, + 0.008824968710541725, + 0.004809231497347355, + 0.010175458155572414, + 0.006102592218667269, + 0.009859003126621246, + 0.009473350830376148, + 0.006229153834283352, + 0.009410824626684189, + 0.011945358477532864, + 0.009366062469780445, + 0.011881736107170582, + 0.008377944119274616, + 0.009572179056704044, + 0.009678819216787815, + 0.011339548043906689, + 0.008764369413256645, + 0.017545858398079872, + 0.01250202115625143, + 0.005460931919515133, + 0.0051219165325164795, + 0.005429146345704794, + 0.005447788164019585, + 0.00819283165037632, + 0.005830384790897369, + 0.008671381510794163, + 0.006841978523880243, + 0.008330060169100761, + 0.012256541289389133, + 0.007560047786682844, + 0.01631186529994011, + 0.015396198257803917, + 0.009528513997793198, + 0.0111553855240345, + 0.011012562550604343, + 0.006712096743285656, + 0.004700955469161272, + 0.007754128891974688, + 0.007252399809658527, + 0.0039479099214077, + 0.005582904443144798, + 0.0043975249864161015, + 0.005366484634578228 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 8 + ], + "selected_reward": -0.003244614228606224 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an anthropomorphic scarab beetle wearing iridescent metallic armor in emerald, deep purple, and bronze?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the beetle's compound eyes glow with a fiery orange light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the beetle slam its large weaponized forearm forward to collide with an opponent's weapon at the edge of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do sparks of cyan-blue energy fly from the impact point like fireworks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the visual style consistent with a 'Battle of Two Cities' animation featuring rough, clear contours and distinct color blocks like a tough oil painting?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.007632380351424217 + ], + "opposite_logprob_tokens": [ + -4.882632255554199 + ], + "logprob_reward": -0.007632380351424217, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.007632380351424217, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.037098150700330734, + 0.06734488904476166, + 0.05838621407747269, + 0.01795192062854767, + 0.09714926779270172, + 0.019086578860878944, + 0.061960309743881226, + 0.04606618359684944, + 0.5378232002258301, + 0.0571332685649395 + ], + "credit_map": [ + 0.005618095397949219, + 0.002802178729325533, + 0.005394377280026674, + 0.0030338349752128124, + 0.0037863715551793575, + 0.006181268487125635, + 0.0035800044424831867, + 0.0067020198330283165, + 0.007628146559000015, + 0.008371859788894653, + 0.009022574871778488, + 0.008096862584352493, + 0.008400819264352322, + 0.008010385558009148, + 0.00891024898737669, + 0.008903988637030125, + 0.0054395911283791065, + 0.005625670310109854, + 0.009416370652616024, + 0.008905340917408466, + 0.005559063050895929, + 0.007328368257731199, + 0.008264911361038685, + 0.007846896536648273, + 0.002113214461132884, + 0.001903610653243959, + 0.001929539255797863, + 0.0021099341101944447, + 0.0016281602438539267, + 0.0018925630720332265, + 0.0032657620031386614, + 0.003109137061983347, + 0.00520208477973938, + 0.0070563191547989845, + 0.004586867988109589, + 0.01297455932945013, + 0.007184076122939587, + 0.024440452456474304, + 0.008645210415124893, + 0.02705969102680683, + 0.0017925577703863382, + 0.001797839067876339, + 0.002135796705260873, + 0.0025393248070031404, + 0.0023636987898498774, + 0.0023347726091742516, + 0.003536549862474203, + 0.0025860387831926346, + 0.0073729585856199265, + 0.005305697675794363, + 0.007060161326080561, + 0.005590490065515041, + 0.006655833683907986, + 0.006573112681508064, + 0.013723026029765606, + 0.009679033420979977, + 0.005859509576112032, + 0.003907571546733379, + 0.00575953209772706, + 0.005037236958742142, + 0.005353291518986225, + 0.009223644621670246, + 0.003988737240433693, + 0.0069366637617349625, + 0.060186974704265594, + 0.060470081865787506, + 0.040460146963596344, + 0.05735301598906517, + 0.11993445456027985, + 0.07353817671537399, + 0.05645820498466492, + 0.0694221630692482, + 0.005675798282027245, + 0.008295707404613495, + 0.006867567542940378, + 0.009837718680500984, + 0.0076405759900808334, + 0.005623129196465015, + 0.005356383044272661, + 0.007836390286684036 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 4, + 1 + ], + "selected_reward": -0.007632380351424217 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an anthropomorphic scarab beetle wearing iridescent metallic armor in emerald, deep purple, and bronze?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the beetle's compound eyes glow with a fiery orange light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the beetle slam its large weaponized forearm forward to collide with an opponent's weapon at the edge of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do sparks of cyan-blue energy fly from the impact point like fireworks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the visual style consistent with a 'Battle of Two Cities' animation featuring rough, clear contours and distinct color blocks like a tough oil painting?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.048646945506334305 + ], + "opposite_logprob_tokens": [ + -3.048646926879883 + ], + "logprob_reward": -0.048646945506334305, + "yes_no_margin_reward": 3.0, + "token_credit_reward": -0.048646945506334305, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1500016152858734, + 0.12591895461082458, + 0.05272491276264191, + 0.04984443634748459, + 0.22068409621715546, + 0.07543102651834488, + 0.12085743248462677, + 0.09541959315538406, + 0.04449920728802681, + 0.06461873650550842 + ], + "credit_map": [ + 0.016664372757077217, + 0.011082123965024948, + 0.03255477920174599, + 0.018255501985549927, + 0.01771152764558792, + 0.022773046046495438, + 0.015913955867290497, + 0.0150463180616498, + 0.011489673517644405, + 0.01929483562707901, + 0.017173610627651215, + 0.014017042703926563, + 0.018234295770525932, + 0.014509649947285652, + 0.015100035816431046, + 0.016099808737635612, + 0.004186022561043501, + 0.007074200082570314, + 0.007143001537770033, + 0.005866081453859806, + 0.007924596779048443, + 0.006512314081192017, + 0.0073881386779248714, + 0.006630556657910347, + 0.0065296608954668045, + 0.007067249622195959, + 0.004494252149015665, + 0.007696500979363918, + 0.004513856954872608, + 0.004568702541291714, + 0.007977006025612354, + 0.0069972011260688305, + 0.01176564209163189, + 0.011477761901915073, + 0.007864153012633324, + 0.04454052448272705, + 0.011356559582054615, + 0.04961510747671127, + 0.012611550278961658, + 0.0714527890086174, + 0.00826855655759573, + 0.006766856648027897, + 0.011358078569173813, + 0.009087314829230309, + 0.00863964669406414, + 0.010240592993795872, + 0.012351145967841148, + 0.008718837052583694, + 0.01377992145717144, + 0.010468356311321259, + 0.01432922575622797, + 0.012976536527276039, + 0.013866248540580273, + 0.012558555230498314, + 0.023294739425182343, + 0.019583845511078835, + 0.009220365434885025, + 0.010188865475356579, + 0.010871617123484612, + 0.009551254101097584, + 0.0158805251121521, + 0.01644260063767433, + 0.010917112231254578, + 0.012347248382866383, + 0.00544000742956996, + 0.005619187373667955, + 0.003983347211033106, + 0.0058977860026061535, + 0.006519061513245106, + 0.004925957415252924, + 0.00413539120927453, + 0.0079784682020545, + 0.006648596376180649, + 0.012865977361798286, + 0.006884036120027304, + 0.008829042315483093, + 0.007953879423439503, + 0.005072716623544693, + 0.007232316769659519, + 0.009132174775004387 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 0, + 1 + ], + "selected_reward": -0.048646945506334305 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an anthropomorphic scarab beetle wearing iridescent metallic armor in emerald, deep purple, and bronze?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the beetle's compound eyes glow with a fiery orange light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the beetle slam its large weaponized forearm forward to collide with an opponent's weapon at the edge of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do sparks of cyan-blue energy fly from the impact point like fireworks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the visual style consistent with a 'Battle of Two Cities' animation featuring rough, clear contours and distinct color blocks like a tough oil painting?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.011067431420087814 + ], + "opposite_logprob_tokens": [ + -4.5110673904418945 + ], + "logprob_reward": -0.011067431420087814, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.011067431420087814, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1307576298713684, + 0.08868353813886642, + 0.036131974309682846, + 0.03209061175584793, + 0.16537423431873322, + 0.1574045568704605, + 0.026556096971035004, + 0.16127924621105194, + 0.1655871868133545, + 0.03613491728901863 + ], + "credit_map": [ + 0.015546310693025589, + 0.010029199533164501, + 0.0313270166516304, + 0.01406357903033495, + 0.01300142239779234, + 0.019515732303261757, + 0.01144348829984665, + 0.015830881893634796, + 0.008700753562152386, + 0.011204995214939117, + 0.01319365855306387, + 0.011240986175835133, + 0.01257309690117836, + 0.010111751966178417, + 0.010253951884806156, + 0.011404342018067837, + 0.0031750998459756374, + 0.005455970764160156, + 0.004568233620375395, + 0.004744201432913542, + 0.004433904308825731, + 0.0034626151900738478, + 0.005346683785319328, + 0.004945266991853714, + 0.0047111171297729015, + 0.0028743885923177004, + 0.002770835068076849, + 0.004378030076622963, + 0.002973613329231739, + 0.003538084216415882, + 0.004779969807714224, + 0.0060645719058811665, + 0.008473346941173077, + 0.00652959244325757, + 0.005718432832509279, + 0.011911987327039242, + 0.008976778015494347, + 0.06675819307565689, + 0.007751817349344492, + 0.049254074692726135, + 0.01600533537566662, + 0.009092846885323524, + 0.02470831386744976, + 0.015362190082669258, + 0.019824648275971413, + 0.037445176392793655, + 0.021569963544607162, + 0.013396080583333969, + 0.002133610425516963, + 0.002661851467564702, + 0.0031508563552051783, + 0.0031104767695069313, + 0.002471198793500662, + 0.004566652700304985, + 0.0037629057187587023, + 0.004698545206338167, + 0.015516480430960655, + 0.01514557097107172, + 0.01855543814599514, + 0.015732184052467346, + 0.02467571571469307, + 0.033354222774505615, + 0.016227848827838898, + 0.022071784362196922, + 0.017543895170092583, + 0.018519503995776176, + 0.01626831665635109, + 0.01637372374534607, + 0.03335265442728996, + 0.02240181714296341, + 0.01818266697227955, + 0.02294461987912655, + 0.004223792813718319, + 0.006434669718146324, + 0.004101991653442383, + 0.0051140934228897095, + 0.0045863790437579155, + 0.003044064389541745, + 0.0036643273197114468, + 0.00496559776365757 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 4, + 7 + ], + "selected_reward": -0.011067431420087814 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an anthropomorphic scarab beetle wearing iridescent metallic armor in emerald, deep purple, and bronze?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the beetle's compound eyes glow with a fiery orange light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the beetle slam its large weaponized forearm forward to collide with an opponent's weapon at the edge of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do sparks of cyan-blue energy fly from the impact point like fireworks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the visual style consistent with a 'Battle of Two Cities' animation featuring rough, clear contours and distinct color blocks like a tough oil painting?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.10028325766324997 + ], + "opposite_logprob_tokens": [ + -2.350283145904541 + ], + "logprob_reward": -0.10028325766324997, + "yes_no_margin_reward": 2.25, + "token_credit_reward": -0.10028325766324997, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.24769166111946106, + 0.11812272667884827, + 0.06299459934234619, + 0.03222208842635155, + 0.14703483879566193, + 0.06960570067167282, + 0.092603899538517, + 0.10331089049577713, + 0.08882374316453934, + 0.03758985549211502 + ], + "credit_map": [ + 0.036924343556165695, + 0.022267339751124382, + 0.04524616897106171, + 0.0334148108959198, + 0.02280089631676674, + 0.03344105929136276, + 0.02876594290137291, + 0.02483108639717102, + 0.014427543617784977, + 0.011717387475073338, + 0.013435249216854572, + 0.015024542808532715, + 0.016381412744522095, + 0.015727585181593895, + 0.014887149445712566, + 0.016521846875548363, + 0.0071002645418047905, + 0.005645331460982561, + 0.009198883548378944, + 0.00831769872456789, + 0.009093460626900196, + 0.006299068219959736, + 0.008098314516246319, + 0.009241575375199318, + 0.0030059253331273794, + 0.0029249629005789757, + 0.003125311341136694, + 0.004747492726892233, + 0.0034150201827287674, + 0.003353091888129711, + 0.006245167460292578, + 0.005405114497989416, + 0.007664924953132868, + 0.01058839913457632, + 0.005921886768192053, + 0.021758077666163445, + 0.01746486872434616, + 0.03497527539730072, + 0.009550821967422962, + 0.039110586047172546, + 0.0062886793166399, + 0.007479475345462561, + 0.008683734573423862, + 0.008859669789671898, + 0.007569958455860615, + 0.01139176543802023, + 0.011406409554183483, + 0.00792600866407156, + 0.013011929579079151, + 0.008254188112914562, + 0.010911538265645504, + 0.008739539422094822, + 0.009934473782777786, + 0.009281623177230358, + 0.01810060255229473, + 0.014369999058544636, + 0.015927456319332123, + 0.009973702020943165, + 0.012236885726451874, + 0.012055720202624798, + 0.012594256550073624, + 0.011510924436151981, + 0.01156072597950697, + 0.017451217398047447, + 0.007770652417093515, + 0.011068297550082207, + 0.00780149269849062, + 0.012224625796079636, + 0.015974415466189384, + 0.012745539657771587, + 0.010697433724999428, + 0.010541283525526524, + 0.005322740878909826, + 0.004723795689642429, + 0.006352897733449936, + 0.003731211181730032, + 0.003476859303191304, + 0.003840747755020857, + 0.0039565772749483585, + 0.0061850231140851974 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 1 + ], + "selected_reward": -0.10028325766324997 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "striking close-up closely tracks the head of an anthropomorphic scarab beetle, caught in the middle of an epic confrontation. The behemoth's entire body is clad in iridescent metallic armor shimmering in emerald, deep purple, and bronze, shimmering brilliantly in the dim light. The camera zooms in on the complex, imposing structure of its head, its compound eyes ablaze with a fiery orange glow, betraying a relentless will to fight. At that moment, its disproportionately large, weaponized forearm, like a battering ram, slams forward, colliding violently with the weapon of an equally massive opponent at the edge of the frame. The moment of impact sends sparks of dazzling cyan-blue energy flying like fireworks, the force of the impact shaking the camera. The cascading sparks become the primary light source, carving sharp highlights and deep shadows into the smooth, curved surface of the armor, perfectly illustrating the intense contrast between light and dark. The entire picture presents a unique \"Battle of Two Cities\" animation style. Each frame seems to be a carefully drawn tough oil painting. The character's contours are rough and clear, and there are clear transitions between color blocks, creating a tense, violent and powerful atmosphere of epic confrontation.", + "seed": 8821, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000027/group_2/index_6487_seed_8821_dp_5_rank_10_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 2, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.029536423133686186, + "vlm_logprob": -0.029536423133686186, + "vlm_yes_no_margin": 4.375, + "vlm_token_credit": -0.029536423133686186, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an anthropomorphic scarab beetle wearing iridescent metallic armor in emerald, deep purple, and bronze?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the beetle's compound eyes glow with a fiery orange light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the beetle slam its large weaponized forearm forward to collide with an opponent's weapon at the edge of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do sparks of cyan-blue energy fly from the impact point like fireworks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the visual style consistent with a 'Battle of Two Cities' animation featuring rough, clear contours and distinct color blocks like a tough oil painting?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.00513710081577301 + ], + "opposite_logprob_tokens": [ + -5.630136966705322 + ], + "logprob_reward": -0.00513710081577301, + "yes_no_margin_reward": 5.625, + "token_credit_reward": -0.00513710081577301, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13920368254184723, + 0.19690243899822235, + 0.11344441026449203, + 0.0272386334836483, + 0.08132903277873993, + 0.16008399426937103, + 0.06796051561832428, + 0.02959839068353176, + 0.10763213038444519, + 0.07660674303770065 + ], + "credit_map": [ + 0.007502575870603323, + 0.011583457700908184, + 0.01287140417844057, + 0.01519576832652092, + 0.03175702691078186, + 0.014186494052410126, + 0.025317717343568802, + 0.0207892507314682, + 0.010912680998444557, + 0.012618109583854675, + 0.00970104243606329, + 0.05620996654033661, + 0.008924189954996109, + 0.016108425334095955, + 0.01552356593310833, + 0.0669044554233551, + 0.033732857555150986, + 0.007644678000360727, + 0.010484644211828709, + 0.009014995768666267, + 0.025589853525161743, + 0.007628175895661116, + 0.0076142773032188416, + 0.011734922416508198, + 0.0029566658195108175, + 0.0030096457339823246, + 0.002274496015161276, + 0.002596362726762891, + 0.004669186659157276, + 0.003315580077469349, + 0.0036858029197901487, + 0.004730895161628723, + 0.006847935263067484, + 0.007850544527173042, + 0.005070943385362625, + 0.013952834531664848, + 0.013161216862499714, + 0.01095523126423359, + 0.010973907075822353, + 0.012516414746642113, + 0.00878251250833273, + 0.009854039177298546, + 0.03007921390235424, + 0.01598476618528366, + 0.03697040677070618, + 0.01773551106452942, + 0.027768446132540703, + 0.01290910691022873, + 0.004919484257698059, + 0.006676298100501299, + 0.004380778409540653, + 0.004644123371690512, + 0.01802828535437584, + 0.011486549861729145, + 0.00909670814871788, + 0.00872829183936119, + 0.003401099471375346, + 0.0036379641387611628, + 0.0022970084100961685, + 0.003467082977294922, + 0.005418215412646532, + 0.0036736945621669292, + 0.004512831103056669, + 0.0031904939096421003, + 0.00989672914147377, + 0.008308269083499908, + 0.009305574931204319, + 0.01652582734823227, + 0.009673438966274261, + 0.011265677399933338, + 0.029493452981114388, + 0.013163154013454914, + 0.004679036792367697, + 0.0045771426521241665, + 0.008179949596524239, + 0.012704354710876942, + 0.004972042981535196, + 0.013188476674258709, + 0.0073762331157922745, + 0.020929507911205292 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 5, + 0 + ], + "selected_reward": -0.00513710081577301 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an anthropomorphic scarab beetle wearing iridescent metallic armor in emerald, deep purple, and bronze?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the beetle's compound eyes glow with a fiery orange light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the beetle slam its large weaponized forearm forward to collide with an opponent's weapon at the edge of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do sparks of cyan-blue energy fly from the impact point like fireworks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the visual style consistent with a 'Battle of Two Cities' animation featuring rough, clear contours and distinct color blocks like a tough oil painting?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0430624857544899 + ], + "opposite_logprob_tokens": [ + -3.168062448501587 + ], + "logprob_reward": -0.0430624857544899, + "yes_no_margin_reward": 3.125, + "token_credit_reward": -0.0430624857544899, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07190478593111038, + 0.321563184261322, + 0.3090435564517975, + 0.01738554984331131, + 0.027693219482898712, + 0.024336136877536774, + 0.023072129115462303, + 0.03240189328789711, + 0.12203069031238556, + 0.050568852573633194 + ], + "credit_map": [ + 0.007612853776663542, + 0.015556585974991322, + 0.007992767728865147, + 0.006948857568204403, + 0.007583186961710453, + 0.008388148620724678, + 0.008967459201812744, + 0.008854929357767105, + 0.01782054826617241, + 0.018327374011278152, + 0.012060750275850296, + 0.09770092368125916, + 0.010438717901706696, + 0.02307816781103611, + 0.02481376938521862, + 0.1173229068517685, + 0.04965856671333313, + 0.01424878928810358, + 0.01250279787927866, + 0.025727668777108192, + 0.14071717858314514, + 0.02373412437736988, + 0.029481850564479828, + 0.012972569093108177, + 0.0020787303801625967, + 0.0028388346545398235, + 0.001980692381039262, + 0.001998533960431814, + 0.0018659878987818956, + 0.0020471063908189535, + 0.002503786003217101, + 0.0020718767773360014, + 0.0025706251617521048, + 0.003362522693350911, + 0.0026298719458281994, + 0.003100419882684946, + 0.004022885579615831, + 0.0039847008883953094, + 0.0034067435190081596, + 0.004615450277924538, + 0.001405168790370226, + 0.001403798465617001, + 0.004407496657222509, + 0.0025066076777875423, + 0.005408470518887043, + 0.002732605440542102, + 0.004340366926044226, + 0.002131623448804021, + 0.001553651294670999, + 0.0025164319667965174, + 0.0016532761510461569, + 0.0012308877194300294, + 0.006781077012419701, + 0.00447625620290637, + 0.0028826792258769274, + 0.0019778695423156023, + 0.0038427752442657948, + 0.0030449454206973314, + 0.0016713823424652219, + 0.0033365879207849503, + 0.006797170266509056, + 0.006469190586358309, + 0.004366532899439335, + 0.0028733094222843647, + 0.00736515736207366, + 0.0074447151273489, + 0.0054255700670182705, + 0.018053928390145302, + 0.011118171736598015, + 0.015340653248131275, + 0.03924870118498802, + 0.018033800646662712, + 0.00458160974085331, + 0.005322678945958614, + 0.002953198039904237, + 0.006738121621310711, + 0.007447931915521622, + 0.008035178296267986, + 0.009296764619648457, + 0.006193371489644051 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 2, + 8 + ], + "selected_reward": -0.0430624857544899 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an anthropomorphic scarab beetle wearing iridescent metallic armor in emerald, deep purple, and bronze?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the beetle's compound eyes glow with a fiery orange light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the beetle slam its large weaponized forearm forward to collide with an opponent's weapon at the edge of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do sparks of cyan-blue energy fly from the impact point like fireworks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the visual style consistent with a 'Battle of Two Cities' animation featuring rough, clear contours and distinct color blocks like a tough oil painting?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.007638413459062576 + ], + "opposite_logprob_tokens": [ + -4.882638454437256 + ], + "logprob_reward": -0.007638413459062576, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.007638413459062576, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.03470304608345032, + 0.059229619801044464, + 0.6468119621276855, + 0.032780807465314865, + 0.028330838307738304, + 0.03419617563486099, + 0.043431032449007034, + 0.022186115384101868, + 0.047152549028396606, + 0.05117785558104515 + ], + "credit_map": [ + 0.0031529844272881746, + 0.004129615146666765, + 0.0039207893423736095, + 0.00481435377150774, + 0.00511708902195096, + 0.004811384249478579, + 0.0048900628462433815, + 0.003866765880957246, + 0.003239648649469018, + 0.0046398574486374855, + 0.0026393812149763107, + 0.016817843541502953, + 0.0029307531658560038, + 0.003753929864615202, + 0.005943941883742809, + 0.019264256581664085, + 0.15198379755020142, + 0.03387966752052307, + 0.035406384617090225, + 0.043450091034173965, + 0.2655465304851532, + 0.0449451319873333, + 0.048376794904470444, + 0.023223528638482094, + 0.0035146893933415413, + 0.004042045678943396, + 0.004530242178589106, + 0.0030159440357238054, + 0.005043452605605125, + 0.0036902311258018017, + 0.0042118169367313385, + 0.004732381086796522, + 0.0031291109044104815, + 0.0030353334732353687, + 0.0020575777161866426, + 0.004767293576151133, + 0.0035489408764988184, + 0.0027640312910079956, + 0.0048812031745910645, + 0.004147345200181007, + 0.0021617000456899405, + 0.002529693068936467, + 0.0048004076816141605, + 0.00358243053779006, + 0.005753024015575647, + 0.003098751651123166, + 0.006722001358866692, + 0.005548166576772928, + 0.0037708876188844442, + 0.005536955781280994, + 0.0023161687422543764, + 0.0026457367930561304, + 0.014647641219198704, + 0.00806494615972042, + 0.0033296067267656326, + 0.0031190840527415276, + 0.0022981828078627586, + 0.0019741677679121494, + 0.0019426833605393767, + 0.0021988442167639732, + 0.004514225292950869, + 0.0036884373985230923, + 0.0030795244965702295, + 0.0024900471325963736, + 0.004171415232121944, + 0.003905189223587513, + 0.0034845806658267975, + 0.007477802224457264, + 0.0065391045063734055, + 0.005811093375086784, + 0.011293414980173111, + 0.0044699423015117645, + 0.002849293639883399, + 0.002880654064938426, + 0.005987848620861769, + 0.005711570382118225, + 0.005077317822724581, + 0.012435927987098694, + 0.006634256336838007, + 0.009600977413356304 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 1, + 9 + ], + "selected_reward": -0.007638413459062576 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an anthropomorphic scarab beetle wearing iridescent metallic armor in emerald, deep purple, and bronze?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the beetle's compound eyes glow with a fiery orange light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the beetle slam its large weaponized forearm forward to collide with an opponent's weapon at the edge of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do sparks of cyan-blue energy fly from the impact point like fireworks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the visual style consistent with a 'Battle of Two Cities' animation featuring rough, clear contours and distinct color blocks like a tough oil painting?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.002815451705828309 + ], + "opposite_logprob_tokens": [ + -5.877815246582031 + ], + "logprob_reward": -0.002815451705828309, + "yes_no_margin_reward": 5.875, + "token_credit_reward": -0.002815451705828309, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08840448409318924, + 0.09995976835489273, + 0.06560621410608292, + 0.02452983148396015, + 0.02390940859913826, + 0.046453557908535004, + 0.19175733625888824, + 0.05617731437087059, + 0.11987856775522232, + 0.2833234965801239 + ], + "credit_map": [ + 0.007338980678468943, + 0.009518414735794067, + 0.010728019289672375, + 0.010567694902420044, + 0.0162457674741745, + 0.012136655859649181, + 0.013461618684232235, + 0.008407334797084332, + 0.0064792088232934475, + 0.008448958396911621, + 0.005211871117353439, + 0.028112336993217468, + 0.004778925329446793, + 0.006547724828124046, + 0.008512704633176327, + 0.03186803683638573, + 0.014614686369895935, + 0.004031922202557325, + 0.003901482094079256, + 0.004863944370299578, + 0.024690555408596992, + 0.006244182586669922, + 0.0039981636218726635, + 0.003261271631345153, + 0.0035761550534516573, + 0.002813157858327031, + 0.0023293981794267893, + 0.002240248955786228, + 0.0029549631290137768, + 0.0040972610004246235, + 0.0026968319434672594, + 0.0038218151312321424, + 0.0022313932422548532, + 0.0027449121698737144, + 0.00294112809933722, + 0.0028486831579357386, + 0.0034427945502102375, + 0.0025224583223462105, + 0.0036545332986861467, + 0.0035235052928328514, + 0.0034865604247897863, + 0.0037277345545589924, + 0.006429634988307953, + 0.0037392820231616497, + 0.011068794876337051, + 0.005836932919919491, + 0.007380833849310875, + 0.004783788695931435, + 0.016261320561170578, + 0.022775255143642426, + 0.014855432324111462, + 0.011985257267951965, + 0.06263329833745956, + 0.03597640618681908, + 0.015438861213624477, + 0.011831511743366718, + 0.005066444166004658, + 0.004346547648310661, + 0.0034826442133635283, + 0.005171247757971287, + 0.012836492620408535, + 0.011037573218345642, + 0.008504076860845089, + 0.005732288584113121, + 0.006133738439530134, + 0.00694491621106863, + 0.007086589466780424, + 0.01618390530347824, + 0.012962305918335915, + 0.010615252889692783, + 0.03938717022538185, + 0.020564695820212364, + 0.013886723667383194, + 0.01791556179523468, + 0.021270453929901123, + 0.06367219239473343, + 0.03205294907093048, + 0.07230723649263382, + 0.029063737019896507, + 0.033154651522636414 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 6, + 8 + ], + "selected_reward": -0.002815451705828309 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an anthropomorphic scarab beetle wearing iridescent metallic armor in emerald, deep purple, and bronze?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the beetle's compound eyes glow with a fiery orange light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the beetle slam its large weaponized forearm forward to collide with an opponent's weapon at the edge of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do sparks of cyan-blue energy fly from the impact point like fireworks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the visual style consistent with a 'Battle of Two Cities' animation featuring rough, clear contours and distinct color blocks like a tough oil painting?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.08902866393327713 + ], + "opposite_logprob_tokens": [ + -2.4640285968780518 + ], + "logprob_reward": -0.08902866393327713, + "yes_no_margin_reward": 2.375, + "token_credit_reward": -0.08902866393327713, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09488115459680557, + 0.14018309116363525, + 0.14694367349147797, + 0.019499894231557846, + 0.09541782736778259, + 0.2142077088356018, + 0.0498843789100647, + 0.04618438333272934, + 0.07608446478843689, + 0.11671342700719833 + ], + "credit_map": [ + 0.008560708723962307, + 0.008953980170190334, + 0.009047042578458786, + 0.011090526357293129, + 0.01505095325410366, + 0.012369281612336636, + 0.014123212546110153, + 0.01568545401096344, + 0.006903435569256544, + 0.008368972688913345, + 0.005981756839901209, + 0.041873861104249954, + 0.005722813308238983, + 0.01041446439921856, + 0.011187420226633549, + 0.049730364233255386, + 0.036710336804389954, + 0.008524865843355656, + 0.009075277484953403, + 0.010326122865080833, + 0.053510844707489014, + 0.012006944976747036, + 0.010369840078055859, + 0.00641945656388998, + 0.002287183189764619, + 0.0034773426596075296, + 0.0023592114448547363, + 0.001786713139154017, + 0.002879922278225422, + 0.0021090928930789232, + 0.0021510503720492125, + 0.002449378604069352, + 0.007061422802507877, + 0.008334698155522346, + 0.006020896136760712, + 0.014338352717459202, + 0.013548996299505234, + 0.014716017059981823, + 0.014846530742943287, + 0.016550906002521515, + 0.011940901167690754, + 0.012257746420800686, + 0.04007686302065849, + 0.02035842277109623, + 0.050256915390491486, + 0.024231817573308945, + 0.03777031973004341, + 0.017314724624156952, + 0.003504083724692464, + 0.0063692787662148476, + 0.0049621593207120895, + 0.005008500535041094, + 0.010026847012341022, + 0.007982642389833927, + 0.006331453565508127, + 0.005699415225535631, + 0.005052174441516399, + 0.003219882259145379, + 0.003107459517195821, + 0.004489765968173742, + 0.009511190466582775, + 0.010786537081003189, + 0.006382958963513374, + 0.0036344134714454412, + 0.005743357818573713, + 0.004895850084722042, + 0.00433060247451067, + 0.00833356287330389, + 0.010603907518088818, + 0.01770879328250885, + 0.01578962802886963, + 0.008678759448230267, + 0.006964236497879028, + 0.006610000040382147, + 0.012503070756793022, + 0.025387685745954514, + 0.010720501653850079, + 0.016401439905166626, + 0.01168480608612299, + 0.026441680267453194 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 2, + 1 + ], + "selected_reward": -0.08902866393327713 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000027/group_5/index_tensor([4364, 4364])_group_5_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000027/group_5/index_tensor([4364, 4364])_group_5_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..68b6e97d3ef1e5f8f0aae1bf62094925ff44eae5 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000027/group_5/index_tensor([4364, 4364])_group_5_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "striking low-angle close-up shot follows, the camera panning rapidly in an arcing trajectory to capture a futuristically designed heavy main battle tank. Its angular armored hull, constructed of a dark gray, scratched composite material, is covered in a complex web of pipes and radiator grilles. Under a gloomy sky, icy raindrops fall, drenching the concrete plaza of an abandoned industrial area. With astonishing agility belying its size, the behemoth performs a fierce, wide-angle power drift, its engine roaring deafeningly and its exhaust spewing searing streams of air. Its wide, muddy metal tracks spin frantically on the slippery concrete, grinding violently against the ground, sending forth a trail of dazzling orange-red sparks that briefly illuminate the metal details and splashing water on the tank's side. The violent turning of the tracks splashed the accumulated water on the ground violently in all directions, forming a spectacular water curtain. The whole scene was ignited by hot sparks in the cold blue-gray tones, full of primitive power, tense dynamic aesthetics and the cool atmosphere of heavy industry.", + "seed": 12850, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000027/group_5/index_4364_seed_12850_dp_10_rank_20_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 5, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.028340984508395194, + "vlm_logprob": -0.028340984508395194, + "vlm_yes_no_margin": 3.975, + "vlm_token_credit": -0.028340984508395194, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a futuristically designed heavy battle tank?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the tank's hull dark gray, scratched, and covered in pipes and radiator grilles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the tank performing a wide-angle power drift with spinning muddy tracks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are there dazzling orange-red sparks and a water curtain created by the tank's movement?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera pan rapidly in an arcing trajectory while maintaining a low-angle close-up?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01571156643331051 + ], + "opposite_logprob_tokens": [ + -5.140711784362793 + ], + "logprob_reward": -0.01571156643331051, + "yes_no_margin_reward": 5.125, + "token_credit_reward": -0.01571156643331051, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.5077487826347351, + 0.03914528712630272, + 0.06012481078505516, + 0.029524311423301697, + 0.05773788318037987, + 0.03509575501084328, + 0.01599692367017269, + 0.07303576171398163, + 0.09150690585374832, + 0.09008361399173737 + ], + "credit_map": [ + 0.03971843421459198, + 0.01512511819601059, + 0.2182493656873703, + 0.026399943977594376, + 0.014486948028206825, + 0.12715765833854675, + 0.03439135476946831, + 0.03221990540623665, + 0.004179744981229305, + 0.005938314367085695, + 0.0038710259832441807, + 0.006314116530120373, + 0.0050542885437607765, + 0.00590519467368722, + 0.003757662605494261, + 0.0041249413043260574, + 0.0070052058435976505, + 0.006058718077838421, + 0.006852165795862675, + 0.013202572241425514, + 0.006953221280127764, + 0.006210885010659695, + 0.007840747013688087, + 0.006001296453177929, + 0.0028825297486037016, + 0.004840717650949955, + 0.006670153234153986, + 0.0034258223604410887, + 0.0029096798971295357, + 0.0026494639459997416, + 0.003394583472982049, + 0.002751363441348076, + 0.007599896751344204, + 0.007192554417997599, + 0.008713298477232456, + 0.006753318011760712, + 0.007355372421443462, + 0.006234358996152878, + 0.008802271448075771, + 0.005086812656372786, + 0.002652688417583704, + 0.004146994557231665, + 0.004740232601761818, + 0.004004825372248888, + 0.004197464790195227, + 0.0031706818845123053, + 0.005080164410173893, + 0.0071027022786438465, + 0.0018097320571541786, + 0.0018715762998908758, + 0.002215222455561161, + 0.0023420234210789204, + 0.002347062574699521, + 0.0014676636783406138, + 0.002359953476116061, + 0.0015836884267628193, + 0.009746847674250603, + 0.006041435524821281, + 0.014642415568232536, + 0.01091824471950531, + 0.005955024156719446, + 0.008609910495579243, + 0.008573687635362148, + 0.008548193611204624, + 0.008133796975016594, + 0.005541679449379444, + 0.007664571516215801, + 0.005405189003795385, + 0.007544761057943106, + 0.02237655222415924, + 0.009037504903972149, + 0.025802846997976303, + 0.01785949431359768, + 0.007544673979282379, + 0.011067471466958523, + 0.009828430600464344, + 0.00762965576723218, + 0.00724969245493412, + 0.009078864939510822, + 0.01982533559203148 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 9 + ], + "selected_reward": -0.01571156643331051 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a futuristically designed heavy battle tank?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the tank's hull dark gray, scratched, and covered in pipes and radiator grilles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the tank performing a wide-angle power drift with spinning muddy tracks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are there dazzling orange-red sparks and a water curtain created by the tank's movement?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera pan rapidly in an arcing trajectory while maintaining a low-angle close-up?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.009799586609005928 + ], + "opposite_logprob_tokens": [ + -4.634799480438232 + ], + "logprob_reward": -0.009799586609005928, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.009799586609005928, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.6210834980010986, + 0.03579404205083847, + 0.036880794912576675, + 0.024392955005168915, + 0.029795236885547638, + 0.08125365525484085, + 0.0439077690243721, + 0.0430120974779129, + 0.04350016266107559, + 0.04037977010011673 + ], + "credit_map": [ + 0.04800673946738243, + 0.018010715022683144, + 0.2641996741294861, + 0.03239433839917183, + 0.01772441901266575, + 0.16157768666744232, + 0.04120645672082901, + 0.03796343132853508, + 0.004291972145438194, + 0.005457745399326086, + 0.003315286012366414, + 0.00546986423432827, + 0.004617844708263874, + 0.005996570456773043, + 0.0035512694157660007, + 0.0030934857204556465, + 0.0043311892077326775, + 0.0036133681423962116, + 0.005312174092978239, + 0.007508574984967709, + 0.0033355639316141605, + 0.004059319384396076, + 0.004590711556375027, + 0.004129891283810139, + 0.0020644045434892178, + 0.0025565281976014376, + 0.004066480789333582, + 0.002740521216765046, + 0.003058492438867688, + 0.002523597562685609, + 0.00482055963948369, + 0.002562368754297495, + 0.0029443218372762203, + 0.004120897967368364, + 0.004684532061219215, + 0.002747138263657689, + 0.0037651127204298973, + 0.002525900024920702, + 0.005614911206066608, + 0.003392420941963792, + 0.006558999419212341, + 0.008796410635113716, + 0.011098731309175491, + 0.00902103167027235, + 0.008523953147232533, + 0.0076348911970853806, + 0.01173324789851904, + 0.017886390909552574, + 0.00422271154820919, + 0.004749549552798271, + 0.007148500997573137, + 0.005162314046174288, + 0.005520830862224102, + 0.0040706004947423935, + 0.006458663381636143, + 0.006574599072337151, + 0.00785095989704132, + 0.0032355026341974735, + 0.008539723232388496, + 0.005341893061995506, + 0.00248412205837667, + 0.004883908666670322, + 0.004029626492410898, + 0.006646359339356422, + 0.0034744050353765488, + 0.003098106011748314, + 0.0034765375312417746, + 0.0027949251234531403, + 0.004237284418195486, + 0.010406291112303734, + 0.004347568843513727, + 0.011665042489767075, + 0.006802170071750879, + 0.0035236477851867676, + 0.005768831353634596, + 0.004550632555037737, + 0.0035338609013706446, + 0.0029106519650667906, + 0.004649733193218708, + 0.008640239015221596 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 5, + 6 + ], + "selected_reward": -0.009799586609005928 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a futuristically designed heavy battle tank?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the tank's hull dark gray, scratched, and covered in pipes and radiator grilles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the tank performing a wide-angle power drift with spinning muddy tracks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are there dazzling orange-red sparks and a water curtain created by the tank's movement?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera pan rapidly in an arcing trajectory while maintaining a low-angle close-up?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.06204032897949219 + ], + "opposite_logprob_tokens": [ + -2.812040328979492 + ], + "logprob_reward": -0.06204032897949219, + "yes_no_margin_reward": 2.75, + "token_credit_reward": -0.06204032897949219, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.24361376464366913, + 0.04186863452196121, + 0.09370767325162888, + 0.11361905187368393, + 0.06146341562271118, + 0.170290008187294, + 0.055911559611558914, + 0.03759446740150452, + 0.14581839740276337, + 0.03611307591199875 + ], + "credit_map": [ + 0.021653445437550545, + 0.01242374163120985, + 0.07759997993707657, + 0.013776267878711224, + 0.010686507448554039, + 0.07023110240697861, + 0.014319062232971191, + 0.022923683747649193, + 0.007160452660173178, + 0.004216569475829601, + 0.005279709119349718, + 0.005504989065229893, + 0.00421446468681097, + 0.006626012269407511, + 0.004935800097882748, + 0.00393063435330987, + 0.011161142028868198, + 0.010295761749148369, + 0.013678777031600475, + 0.017420319840312004, + 0.009726043790578842, + 0.011577129364013672, + 0.010886046104133129, + 0.008962449617683887, + 0.011045447550714016, + 0.013316336087882519, + 0.016742927953600883, + 0.013003824278712273, + 0.014292696490883827, + 0.011523261666297913, + 0.019183505326509476, + 0.014511043205857277, + 0.005790120922029018, + 0.00695051159709692, + 0.01392411720007658, + 0.004144671838730574, + 0.00855595525354147, + 0.005343791097402573, + 0.007880213670432568, + 0.008874036371707916, + 0.012934320606291294, + 0.018694186583161354, + 0.0260938573628664, + 0.017764747142791748, + 0.015493453480303288, + 0.01606527343392372, + 0.0245276540517807, + 0.03871651738882065, + 0.005882388446480036, + 0.008644280955195427, + 0.00928202923387289, + 0.005144012160599232, + 0.009315120987594128, + 0.00516866147518158, + 0.0073465765453875065, + 0.005128488410264254, + 0.004756752867251635, + 0.0026388182304799557, + 0.009131411090493202, + 0.004816100932657719, + 0.002970177913084626, + 0.004942498169839382, + 0.004411792848259211, + 0.003926912322640419, + 0.011147001758217812, + 0.00917133316397667, + 0.009663284756243229, + 0.00985647737979889, + 0.01093864906579256, + 0.04012111946940422, + 0.014611832797527313, + 0.040308691561222076, + 0.006243211217224598, + 0.002783330390229821, + 0.003700203262269497, + 0.004003901034593582, + 0.003361497074365616, + 0.003483780426904559, + 0.005122192203998566, + 0.007414958439767361 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 5, + 8 + ], + "selected_reward": -0.06204032897949219 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a futuristically designed heavy battle tank?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the tank's hull dark gray, scratched, and covered in pipes and radiator grilles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the tank performing a wide-angle power drift with spinning muddy tracks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are there dazzling orange-red sparks and a water curtain created by the tank's movement?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera pan rapidly in an arcing trajectory while maintaining a low-angle close-up?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.016065435484051704 + ], + "opposite_logprob_tokens": [ + -4.14106559753418 + ], + "logprob_reward": -0.016065435484051704, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.016065435484051704, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3668402433395386, + 0.057031307369470596, + 0.1478077620267868, + 0.05483649671077728, + 0.038102127611637115, + 0.11999278515577316, + 0.02825416810810566, + 0.04485980048775673, + 0.06419014185667038, + 0.07808520644903183 + ], + "credit_map": [ + 0.058986663818359375, + 0.018347762525081635, + 0.08390732854604721, + 0.019746946170926094, + 0.01392302569001913, + 0.13413389027118683, + 0.014730572700500488, + 0.023064054548740387, + 0.006553858518600464, + 0.007136705331504345, + 0.005635484587401152, + 0.010316358879208565, + 0.0067448741756379604, + 0.00769613916054368, + 0.005696855951100588, + 0.007251034956425428, + 0.0158785879611969, + 0.019745392724871635, + 0.019803984090685844, + 0.0343918576836586, + 0.013706480152904987, + 0.012680914252996445, + 0.01862168125808239, + 0.01297886110842228, + 0.005266742315143347, + 0.004251664038747549, + 0.00948782917112112, + 0.005452783312648535, + 0.007250133901834488, + 0.007345043588429689, + 0.009955260902643204, + 0.005827039480209351, + 0.0043178447522223, + 0.003355897730216384, + 0.004765584133565426, + 0.003374668536707759, + 0.009581212885677814, + 0.00378506095148623, + 0.004994107410311699, + 0.003927748184651136, + 0.009322001598775387, + 0.012799473479390144, + 0.016787009313702583, + 0.01379240769892931, + 0.012469925917685032, + 0.011034807190299034, + 0.016572806984186172, + 0.02721434272825718, + 0.002839256776496768, + 0.004056892823427916, + 0.004716396331787109, + 0.002790563041344285, + 0.004077007994055748, + 0.0029217535629868507, + 0.002892795018851757, + 0.0039595006965100765, + 0.0077927568927407265, + 0.0032717110589146614, + 0.00838495884090662, + 0.005558430682867765, + 0.0036075455136597157, + 0.00537292892113328, + 0.004526599310338497, + 0.006344867870211601, + 0.004778461065143347, + 0.004175915382802486, + 0.005662726238369942, + 0.005212611053138971, + 0.0048681036569178104, + 0.015271613374352455, + 0.008182343095541, + 0.01603836938738823, + 0.012571079656481743, + 0.006008388940244913, + 0.011564773507416248, + 0.008594011887907982, + 0.006711521651595831, + 0.005571297835558653, + 0.009094813838601112, + 0.017969323322176933 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 5 + ], + "selected_reward": -0.016065435484051704 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a futuristically designed heavy battle tank?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the tank's hull dark gray, scratched, and covered in pipes and radiator grilles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the tank performing a wide-angle power drift with spinning muddy tracks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are there dazzling orange-red sparks and a water curtain created by the tank's movement?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera pan rapidly in an arcing trajectory while maintaining a low-angle close-up?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.038088005036115646 + ], + "opposite_logprob_tokens": [ + -3.288088083267212 + ], + "logprob_reward": -0.038088005036115646, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.038088005036115646, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.48230117559432983, + 0.024084558710455894, + 0.03507930412888527, + 0.10855548828840256, + 0.02722856216132641, + 0.02633056230843067, + 0.033809904009103775, + 0.030710630118846893, + 0.16658395528793335, + 0.06531590223312378 + ], + "credit_map": [ + 0.03863804042339325, + 0.01514695305377245, + 0.20328234136104584, + 0.024243220686912537, + 0.014762124046683311, + 0.1267397701740265, + 0.030002575367689133, + 0.02948613464832306, + 0.003116541774943471, + 0.00337249250151217, + 0.0024886331520974636, + 0.003390881000086665, + 0.0031035973224788904, + 0.0034099335316568613, + 0.0022349690552800894, + 0.002967508975416422, + 0.0031149135902523994, + 0.004140135832130909, + 0.003946913871914148, + 0.007044252473860979, + 0.004953792318701744, + 0.004666931927204132, + 0.004117738921195269, + 0.003094625659286976, + 0.01067697536200285, + 0.011652332730591297, + 0.02400904707610607, + 0.019432175904512405, + 0.00959813129156828, + 0.012108184397220612, + 0.012178102508187294, + 0.008900536224246025, + 0.003237662138417363, + 0.0030495724640786648, + 0.004438699223101139, + 0.003044505836442113, + 0.00482021551579237, + 0.0026262847241014242, + 0.0030748844146728516, + 0.002936737146228552, + 0.002371052745729685, + 0.002533737337216735, + 0.004086161497980356, + 0.0035067410208284855, + 0.0026439055800437927, + 0.0025052777491509914, + 0.004819526802748442, + 0.0038641593419015408, + 0.0033550274092704058, + 0.004581680055707693, + 0.005347609519958496, + 0.004924929700791836, + 0.004855564795434475, + 0.003232749179005623, + 0.003465852700173855, + 0.004046493209898472, + 0.005340809468179941, + 0.0039206561632454395, + 0.004354577511548996, + 0.003128685988485813, + 0.004114728886634111, + 0.004237593617290258, + 0.003277033567428589, + 0.002336547477170825, + 0.015159262344241142, + 0.009582988917827606, + 0.01675526611506939, + 0.011674754321575165, + 0.013215754181146622, + 0.040269773453474045, + 0.016733072698116302, + 0.04319308325648308, + 0.011775690130889416, + 0.005603174678981304, + 0.008205250836908817, + 0.006053163670003414, + 0.005766781512647867, + 0.005198781378567219, + 0.005059546325355768, + 0.017653513699769974 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 3 + ], + "selected_reward": -0.038088005036115646 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "striking low-angle close-up shot follows, the camera panning rapidly in an arcing trajectory to capture a futuristically designed heavy main battle tank. Its angular armored hull, constructed of a dark gray, scratched composite material, is covered in a complex web of pipes and radiator grilles. Under a gloomy sky, icy raindrops fall, drenching the concrete plaza of an abandoned industrial area. With astonishing agility belying its size, the behemoth performs a fierce, wide-angle power drift, its engine roaring deafeningly and its exhaust spewing searing streams of air. Its wide, muddy metal tracks spin frantically on the slippery concrete, grinding violently against the ground, sending forth a trail of dazzling orange-red sparks that briefly illuminate the metal details and splashing water on the tank's side. The violent turning of the tracks splashed the accumulated water on the ground violently in all directions, forming a spectacular water curtain. The whole scene was ignited by hot sparks in the cold blue-gray tones, full of primitive power, tense dynamic aesthetics and the cool atmosphere of heavy industry.", + "seed": 12851, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000027/group_5/index_4364_seed_12851_dp_10_rank_20_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 5, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.03554443418979645, + "vlm_logprob": -0.03554443418979645, + "vlm_yes_no_margin": 3.725, + "vlm_token_credit": -0.03554443418979645, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a futuristically designed heavy battle tank?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the tank's hull dark gray, scratched, and covered in pipes and radiator grilles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the tank performing a wide-angle power drift with spinning muddy tracks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are there dazzling orange-red sparks and a water curtain created by the tank's movement?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera pan rapidly in an arcing trajectory while maintaining a low-angle close-up?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.047749847173690796 + ], + "opposite_logprob_tokens": [ + -3.5477497577667236 + ], + "logprob_reward": -0.047749847173690796, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.047749847173690796, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11417228728532791, + 0.07442063093185425, + 0.11903426796197891, + 0.12894105911254883, + 0.060245342552661896, + 0.09679252654314041, + 0.2550027370452881, + 0.060603708028793335, + 0.04570496454834938, + 0.045082468539476395 + ], + "credit_map": [ + 0.019656378775835037, + 0.010431675240397453, + 0.01812228187918663, + 0.01054275594651699, + 0.013693499378859997, + 0.01441942062228918, + 0.015122837387025356, + 0.012183436192572117, + 0.011039244942367077, + 0.00578675651922822, + 0.00947052426636219, + 0.010558414272964, + 0.009339454583823681, + 0.009357312694191933, + 0.010247014462947845, + 0.00862190779298544, + 0.018726550042629242, + 0.014363802038133144, + 0.014422062784433365, + 0.013436851091682911, + 0.010664939880371094, + 0.01477473322302103, + 0.014632700011134148, + 0.018012627959251404, + 0.01548992469906807, + 0.0117746377363801, + 0.026527902111411095, + 0.017648570239543915, + 0.01363108865916729, + 0.015295546501874924, + 0.016826579347252846, + 0.011746812611818314, + 0.008281927555799484, + 0.005370600614696741, + 0.006799985654652119, + 0.008668670430779457, + 0.004768643993884325, + 0.006400169339030981, + 0.011445782147347927, + 0.008509564213454723, + 0.007869292050600052, + 0.008515963330864906, + 0.01063550915569067, + 0.017685454338788986, + 0.010458998382091522, + 0.015469444915652275, + 0.015761855989694595, + 0.010396001860499382, + 0.018584899604320526, + 0.01673867553472519, + 0.02113483101129532, + 0.054454125463962555, + 0.019326554611325264, + 0.018521752208471298, + 0.08120695501565933, + 0.025034938007593155, + 0.0063623846508562565, + 0.009194271638989449, + 0.006970774382352829, + 0.007550403475761414, + 0.010028479620814323, + 0.006027222145348787, + 0.009172427468001842, + 0.005297742784023285, + 0.004273878410458565, + 0.006681026425212622, + 0.006716324482113123, + 0.007357886526733637, + 0.007444830611348152, + 0.003578456584364176, + 0.005123674403876066, + 0.004528885707259178, + 0.005435710772871971, + 0.003762810258194804, + 0.007162416819483042, + 0.0058501181192696095, + 0.005948283709585667, + 0.005902665201574564, + 0.005323568359017372, + 0.005696895532310009 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 3, + 2 + ], + "selected_reward": -0.047749847173690796 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a futuristically designed heavy battle tank?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the tank's hull dark gray, scratched, and covered in pipes and radiator grilles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the tank performing a wide-angle power drift with spinning muddy tracks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are there dazzling orange-red sparks and a water curtain created by the tank's movement?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera pan rapidly in an arcing trajectory while maintaining a low-angle close-up?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.012567373923957348 + ], + "opposite_logprob_tokens": [ + -4.387567520141602 + ], + "logprob_reward": -0.012567373923957348, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.012567373923957348, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11886826157569885, + 0.0687766894698143, + 0.07110188901424408, + 0.18184246122837067, + 0.13355286419391632, + 0.06875664740800858, + 0.11093202978372574, + 0.08054107427597046, + 0.08347982913255692, + 0.08214820176362991 + ], + "credit_map": [ + 0.02102058380842209, + 0.011408577673137188, + 0.01912679895758629, + 0.009234365075826645, + 0.017104318365454674, + 0.018890026956796646, + 0.011890089139342308, + 0.010193507187068462, + 0.00812976062297821, + 0.0065741525031626225, + 0.007971060462296009, + 0.008955870755016804, + 0.014127122238278389, + 0.009194358251988888, + 0.008389778435230255, + 0.005434584803879261, + 0.009516377933323383, + 0.007672778330743313, + 0.009758371859788895, + 0.011077851988375187, + 0.005612980108708143, + 0.008514792658388615, + 0.011305547319352627, + 0.007643192540854216, + 0.027327438816428185, + 0.016790252178907394, + 0.03205365315079689, + 0.02107211761176586, + 0.023229220882058144, + 0.024981442838907242, + 0.020934615284204483, + 0.015453719533979893, + 0.018155818805098534, + 0.013612327165901661, + 0.015946703031659126, + 0.019730620086193085, + 0.014071824960410595, + 0.01507310476154089, + 0.022257279604673386, + 0.014705182984471321, + 0.006182931829243898, + 0.0060929493047297, + 0.00696309469640255, + 0.007928308099508286, + 0.009478981606662273, + 0.008701981045305729, + 0.014540466479957104, + 0.008867934346199036, + 0.008008470758795738, + 0.010087404400110245, + 0.0072959838435053825, + 0.022795207798480988, + 0.008442713879048824, + 0.007987047545611858, + 0.03395707160234451, + 0.012358129024505615, + 0.008114730939269066, + 0.015719791874289513, + 0.00815799180418253, + 0.007158744614571333, + 0.011647257953882217, + 0.005731415003538132, + 0.01585555262863636, + 0.008155591785907745, + 0.00781845860183239, + 0.014158829115331173, + 0.012085359543561935, + 0.01198958046734333, + 0.01369190588593483, + 0.007853325456380844, + 0.009244760498404503, + 0.006637609098106623, + 0.012455550022423267, + 0.004976294469088316, + 0.01810780167579651, + 0.0108836330473423, + 0.009625492617487907, + 0.009598377160727978, + 0.007582519203424454, + 0.008918536826968193 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 4, + 0 + ], + "selected_reward": -0.012567373923957348 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a futuristically designed heavy battle tank?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the tank's hull dark gray, scratched, and covered in pipes and radiator grilles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the tank performing a wide-angle power drift with spinning muddy tracks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are there dazzling orange-red sparks and a water curtain created by the tank's movement?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera pan rapidly in an arcing trajectory while maintaining a low-angle close-up?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.029796818271279335 + ], + "opposite_logprob_tokens": [ + -3.529796838760376 + ], + "logprob_reward": -0.029796818271279335, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.029796818271279335, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1300050914287567, + 0.0907624363899231, + 0.1586119532585144, + 0.12918251752853394, + 0.04593472555279732, + 0.06526553630828857, + 0.07809760421514511, + 0.027729518711566925, + 0.15560728311538696, + 0.11880332976579666 + ], + "credit_map": [ + 0.014957736246287823, + 0.010174592025578022, + 0.022180572152137756, + 0.014398871921002865, + 0.013930867426097393, + 0.017569107934832573, + 0.02078082785010338, + 0.016012515872716904, + 0.013609293848276138, + 0.005541813559830189, + 0.01200732123106718, + 0.008009694516658783, + 0.009365142323076725, + 0.014355986379086971, + 0.008015993051230907, + 0.019857190549373627, + 0.02540927194058895, + 0.0221992339938879, + 0.01968109793961048, + 0.017712701112031937, + 0.016684746369719505, + 0.019777923822402954, + 0.014382903464138508, + 0.022764066234230995, + 0.031216030940413475, + 0.009951720014214516, + 0.021029246971011162, + 0.012953432276844978, + 0.013235263526439667, + 0.021287940442562103, + 0.011326023377478123, + 0.008182861842215061, + 0.008547871373593807, + 0.003950474783778191, + 0.005784616339951754, + 0.005991630721837282, + 0.0040024821646511555, + 0.005351996514946222, + 0.007395868189632893, + 0.0049097854644060135, + 0.006354299373924732, + 0.0077490052208304405, + 0.007312547415494919, + 0.007343145087361336, + 0.009850171394646168, + 0.009048533625900745, + 0.008958357386291027, + 0.008649471215903759, + 0.009717184118926525, + 0.009648610837757587, + 0.004932385869324207, + 0.014253247529268265, + 0.005085914861410856, + 0.005627429112792015, + 0.020386599004268646, + 0.008446238934993744, + 0.002596575766801834, + 0.0035952716134488583, + 0.003191092750057578, + 0.004531617742031813, + 0.0033038079272955656, + 0.0023413007147610188, + 0.005354991648346186, + 0.0028148621786385775, + 0.01240259874612093, + 0.025496812537312508, + 0.015320515260100365, + 0.02462095580995083, + 0.028908247128129005, + 0.013745960779488087, + 0.021602172404527664, + 0.013510007411241531, + 0.019248880445957184, + 0.007155226543545723, + 0.03132680058479309, + 0.019002387300133705, + 0.009634427726268768, + 0.010214987210929394, + 0.008625458925962448, + 0.013595163822174072 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 8, + 0 + ], + "selected_reward": -0.029796818271279335 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a futuristically designed heavy battle tank?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the tank's hull dark gray, scratched, and covered in pipes and radiator grilles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the tank performing a wide-angle power drift with spinning muddy tracks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are there dazzling orange-red sparks and a water curtain created by the tank's movement?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera pan rapidly in an arcing trajectory while maintaining a low-angle close-up?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.008636976592242718 + ], + "opposite_logprob_tokens": [ + -4.758636951446533 + ], + "logprob_reward": -0.008636976592242718, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.008636976592242718, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11006083339452744, + 0.13509823381900787, + 0.12312910705804825, + 0.07384730875492096, + 0.04872053116559982, + 0.05668098106980324, + 0.04382864385843277, + 0.09348125755786896, + 0.1405395120382309, + 0.1746136099100113 + ], + "credit_map": [ + 0.017847761511802673, + 0.010726258158683777, + 0.01599367894232273, + 0.010358791798353195, + 0.01265714131295681, + 0.014839038252830505, + 0.013927814550697803, + 0.013710344210267067, + 0.02009240724146366, + 0.013436338864266872, + 0.022041387856006622, + 0.011550870724022388, + 0.015023654326796532, + 0.02346888929605484, + 0.009115193039178848, + 0.02036947011947632, + 0.019642487168312073, + 0.01752270758152008, + 0.012879214249551296, + 0.013467706739902496, + 0.013726923614740372, + 0.014992601238191128, + 0.011674506589770317, + 0.019222941249608994, + 0.0123489024117589, + 0.006543157156556845, + 0.012762555852532387, + 0.00789426825940609, + 0.009145972318947315, + 0.009586255997419357, + 0.008136267773807049, + 0.007429920602589846, + 0.008138492703437805, + 0.0033768557477742434, + 0.006300087086856365, + 0.008256720378994942, + 0.0036182915791869164, + 0.006711027584969997, + 0.006645638961344957, + 0.0056734089739620686, + 0.005320258438587189, + 0.0056299190036952496, + 0.005913001950830221, + 0.005965518299490213, + 0.0084013557061553, + 0.006982767954468727, + 0.01138089969754219, + 0.00708725256845355, + 0.007275766227394342, + 0.008994096890091896, + 0.0028503674548119307, + 0.006585817784070969, + 0.0035500514786690474, + 0.0033191731199622154, + 0.0073309773579239845, + 0.003922391217201948, + 0.008541123010218143, + 0.010935894213616848, + 0.009715341962873936, + 0.00876742135733366, + 0.012225324288010597, + 0.007544247433543205, + 0.023370977491140366, + 0.012380917556583881, + 0.00983665231615305, + 0.017706822603940964, + 0.023976866155862808, + 0.02537534199655056, + 0.014703618362545967, + 0.00889477226883173, + 0.0237993523478508, + 0.016246074810624123, + 0.03190112113952637, + 0.009368461556732655, + 0.05057073011994362, + 0.02263331413269043, + 0.016626060009002686, + 0.020370405167341232, + 0.010511528700590134, + 0.012631970457732677 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 8, + 1 + ], + "selected_reward": -0.008636976592242718 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a futuristically designed heavy battle tank?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the tank's hull dark gray, scratched, and covered in pipes and radiator grilles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the tank performing a wide-angle power drift with spinning muddy tracks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are there dazzling orange-red sparks and a water curtain created by the tank's movement?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera pan rapidly in an arcing trajectory while maintaining a low-angle close-up?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07897115498781204 + ], + "opposite_logprob_tokens": [ + -2.5789711475372314 + ], + "logprob_reward": -0.07897115498781204, + "yes_no_margin_reward": 2.5, + "token_credit_reward": -0.07897115498781204, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10138202458620071, + 0.06770111620426178, + 0.20291346311569214, + 0.1451898068189621, + 0.07457222789525986, + 0.08739940077066422, + 0.10250923782587051, + 0.04321549832820892, + 0.06357619166374207, + 0.11154109239578247 + ], + "credit_map": [ + 0.014876860193908215, + 0.01144885178655386, + 0.015572620555758476, + 0.008709040470421314, + 0.013773496262729168, + 0.01210158783942461, + 0.009693424217402935, + 0.015206124633550644, + 0.009168149903416634, + 0.004863044712692499, + 0.009929212741553783, + 0.008421765640377998, + 0.008677701465785503, + 0.010181627236306667, + 0.007283832877874374, + 0.00917577650398016, + 0.03398628532886505, + 0.026439206674695015, + 0.023103801533579826, + 0.02109728753566742, + 0.021836258471012115, + 0.025836503133177757, + 0.019158581271767616, + 0.03145552799105644, + 0.018566738814115524, + 0.012379954569041729, + 0.028844255954027176, + 0.018416952341794968, + 0.014752847142517567, + 0.01595987193286419, + 0.016727115958929062, + 0.019542064517736435, + 0.01808515004813671, + 0.007920031435787678, + 0.007246932480484247, + 0.008189467713236809, + 0.006146266125142574, + 0.011074328795075417, + 0.009147904813289642, + 0.006762143690139055, + 0.009024995379149914, + 0.00841658003628254, + 0.008139806799590588, + 0.00988262053579092, + 0.010594991967082024, + 0.012064861133694649, + 0.017446091398596764, + 0.011829443275928497, + 0.011869394220411777, + 0.006039940752089024, + 0.010832054540514946, + 0.020289156585931778, + 0.00720458896830678, + 0.007200849242508411, + 0.028381839394569397, + 0.010691394098103046, + 0.003169316565617919, + 0.006144088227301836, + 0.005309290252625942, + 0.005066287238150835, + 0.006695464253425598, + 0.0036474294029176235, + 0.007086688652634621, + 0.006096928380429745, + 0.006342690903693438, + 0.009671651758253574, + 0.009587813168764114, + 0.007747544441372156, + 0.008096379227936268, + 0.0062031871639192104, + 0.009687020443379879, + 0.0062398966401815414, + 0.014789371751248837, + 0.012343292124569416, + 0.014430811628699303, + 0.011067752726376057, + 0.017917195335030556, + 0.01662185601890087, + 0.013339116238057613, + 0.011031696572899818 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 3, + 9 + ], + "selected_reward": -0.07897115498781204 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "striking low-angle close-up shot follows, the camera panning rapidly in an arcing trajectory to capture a futuristically designed heavy main battle tank. Its angular armored hull, constructed of a dark gray, scratched composite material, is covered in a complex web of pipes and radiator grilles. Under a gloomy sky, icy raindrops fall, drenching the concrete plaza of an abandoned industrial area. With astonishing agility belying its size, the behemoth performs a fierce, wide-angle power drift, its engine roaring deafeningly and its exhaust spewing searing streams of air. Its wide, muddy metal tracks spin frantically on the slippery concrete, grinding violently against the ground, sending forth a trail of dazzling orange-red sparks that briefly illuminate the metal details and splashing water on the tank's side. The violent turning of the tracks splashed the accumulated water on the ground violently in all directions, forming a spectacular water curtain. The whole scene was ignited by hot sparks in the cold blue-gray tones, full of primitive power, tense dynamic aesthetics and the cool atmosphere of heavy industry.", + "seed": 12852, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000027/group_5/index_4364_seed_12852_dp_11_rank_22_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 5, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.04653307665139437, + "vlm_logprob": -0.04653307665139437, + "vlm_yes_no_margin": 3.75, + "vlm_token_credit": -0.04653307665139437, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a futuristically designed heavy battle tank?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the tank's hull dark gray, scratched, and covered in pipes and radiator grilles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the tank performing a wide-angle power drift with spinning muddy tracks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are there dazzling orange-red sparks and a water curtain created by the tank's movement?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera pan rapidly in an arcing trajectory while maintaining a low-angle close-up?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.016506772488355637 + ], + "opposite_logprob_tokens": [ + -5.016506671905518 + ], + "logprob_reward": -0.016506772488355637, + "yes_no_margin_reward": 5.0, + "token_credit_reward": -0.016506772488355637, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07710672169923782, + 0.15002508461475372, + 0.034728724509477615, + 0.08117339015007019, + 0.08411171287298203, + 0.024744411930441856, + 0.04554824158549309, + 0.031094990670681, + 0.326350599527359, + 0.1451161652803421 + ], + "credit_map": [ + 0.009548916481435299, + 0.006567855831235647, + 0.008653157390654087, + 0.008557798340916634, + 0.007826466113328934, + 0.008519644849002361, + 0.013053282164037228, + 0.01437959261238575, + 0.01775955967605114, + 0.017570381984114647, + 0.015403132885694504, + 0.02035573124885559, + 0.018189355731010437, + 0.015946120023727417, + 0.021438201889395714, + 0.023362591862678528, + 0.0037157286424189806, + 0.0036169455852359533, + 0.005800618324428797, + 0.006831648293882608, + 0.0033061702270060778, + 0.0031741370912641287, + 0.004014765843749046, + 0.004268712364137173, + 0.009316259995102882, + 0.006783613003790379, + 0.019335001707077026, + 0.01533455029129982, + 0.006177179049700499, + 0.008038859814405441, + 0.007836775854229927, + 0.008351147174835205, + 0.0238531231880188, + 0.0032452710438519716, + 0.029979059472680092, + 0.007908875122666359, + 0.0027121540624648333, + 0.00589052215218544, + 0.0037716596852988005, + 0.0067510525695979595, + 0.0031879087910056114, + 0.0025374095421284437, + 0.005067283753305674, + 0.0022007636725902557, + 0.0025352779775857925, + 0.0026569021865725517, + 0.0028106658719480038, + 0.003748200135305524, + 0.004607692826539278, + 0.005260631442070007, + 0.006979572121053934, + 0.00919169932603836, + 0.0053343577310442924, + 0.003841969883069396, + 0.0053110248409211636, + 0.0050212969072163105, + 0.003970061428844929, + 0.003606091719120741, + 0.003214067779481411, + 0.0032942225225269794, + 0.005220544990152121, + 0.0030169147066771984, + 0.003981728572398424, + 0.004791357554495335, + 0.0427335649728775, + 0.03288189694285393, + 0.04396570846438408, + 0.06583478301763535, + 0.02210775390267372, + 0.06378817558288574, + 0.030413324013352394, + 0.02462538704276085, + 0.029408637434244156, + 0.012907393276691437, + 0.03328188881278038, + 0.020182112231850624, + 0.011056061834096909, + 0.012119720689952374, + 0.014017271809279919, + 0.012143081985414028 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 1, + 9 + ], + "selected_reward": -0.016506772488355637 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a futuristically designed heavy battle tank?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the tank's hull dark gray, scratched, and covered in pipes and radiator grilles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the tank performing a wide-angle power drift with spinning muddy tracks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are there dazzling orange-red sparks and a water curtain created by the tank's movement?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera pan rapidly in an arcing trajectory while maintaining a low-angle close-up?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.008663213811814785 + ], + "opposite_logprob_tokens": [ + -4.758663177490234 + ], + "logprob_reward": -0.008663213811814785, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.008663213811814785, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11727457493543625, + 0.08027907460927963, + 0.06094279885292053, + 0.14626453816890717, + 0.14702492952346802, + 0.05912057310342789, + 0.08170685917139053, + 0.05938344821333885, + 0.1960146725177765, + 0.05198849365115166 + ], + "credit_map": [ + 0.016192782670259476, + 0.008624977432191372, + 0.012381521984934807, + 0.014591656625270844, + 0.012602560222148895, + 0.018566712737083435, + 0.016609588637948036, + 0.017704784870147705, + 0.009808914735913277, + 0.012852762825787067, + 0.008643409237265587, + 0.010075111873447895, + 0.010246996767818928, + 0.008400720544159412, + 0.008972583338618279, + 0.01127858180552721, + 0.006756303366273642, + 0.005905026104301214, + 0.00894518569111824, + 0.009821601212024689, + 0.008126052096486092, + 0.005277618765830994, + 0.008938800543546677, + 0.007172215264290571, + 0.01956745609641075, + 0.013107341714203358, + 0.03785443678498268, + 0.026148980483412743, + 0.012465806677937508, + 0.00973541010171175, + 0.014121705666184425, + 0.013263416476547718, + 0.037987906485795975, + 0.008219505660235882, + 0.04219714552164078, + 0.014295641332864761, + 0.008804542012512684, + 0.017741335555911064, + 0.008056016638875008, + 0.009722842834889889, + 0.005091413855552673, + 0.006296437233686447, + 0.013968721963465214, + 0.005219480488449335, + 0.005479644984006882, + 0.00507062952965498, + 0.008074582554399967, + 0.009919668547809124, + 0.007487783208489418, + 0.010810387320816517, + 0.00908600352704525, + 0.009602753445506096, + 0.01281838770955801, + 0.009367059916257858, + 0.009893693961203098, + 0.012640795670449734, + 0.006025445181876421, + 0.004158912692219019, + 0.005867969710379839, + 0.004412151873111725, + 0.007855533622205257, + 0.007559735793620348, + 0.011172440834343433, + 0.01233125850558281, + 0.029964929446578026, + 0.02342664822936058, + 0.03447750210762024, + 0.024807734414935112, + 0.01715586520731449, + 0.031640440225601196, + 0.01922924816608429, + 0.01531233824789524, + 0.00578257767483592, + 0.006032917182892561, + 0.005366342607885599, + 0.010380000807344913, + 0.005148511379957199, + 0.007817748934030533, + 0.00592257222160697, + 0.005537833087146282 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 4, + 3 + ], + "selected_reward": -0.008663213811814785 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a futuristically designed heavy battle tank?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the tank's hull dark gray, scratched, and covered in pipes and radiator grilles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the tank performing a wide-angle power drift with spinning muddy tracks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are there dazzling orange-red sparks and a water curtain created by the tank's movement?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera pan rapidly in an arcing trajectory while maintaining a low-angle close-up?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.14278382062911987 + ], + "opposite_logprob_tokens": [ + -2.0177838802337646 + ], + "logprob_reward": -0.14278382062911987, + "yes_no_margin_reward": 1.875, + "token_credit_reward": -0.14278382062911987, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0810706615447998, + 0.11214066296815872, + 0.021286221221089363, + 0.18806329369544983, + 0.07426992058753967, + 0.05492445081472397, + 0.10739298909902573, + 0.10699883848428726, + 0.17163287103176117, + 0.08222012221813202 + ], + "credit_map": [ + 0.010116135701537132, + 0.007931391708552837, + 0.008874047547578812, + 0.00949043594300747, + 0.007943596690893173, + 0.008257865905761719, + 0.014266591519117355, + 0.014190598390996456, + 0.012207049876451492, + 0.011162301525473595, + 0.011189318262040615, + 0.018966633826494217, + 0.013701840303838253, + 0.014578514732420444, + 0.01659430004656315, + 0.013740697875618935, + 0.002499069320037961, + 0.0028496833983808756, + 0.003451061900705099, + 0.003277462674304843, + 0.002295116428285837, + 0.0018532670801505446, + 0.0025849088560789824, + 0.002475649118423462, + 0.0286344513297081, + 0.017914757132530212, + 0.054385874420404434, + 0.02670877054333687, + 0.012582614086568356, + 0.011150414124131203, + 0.01860860176384449, + 0.018077794462442398, + 0.01207293476909399, + 0.0071141524240374565, + 0.008234321139752865, + 0.010523851029574871, + 0.008100494742393494, + 0.0150781674310565, + 0.006545951124280691, + 0.006600040942430496, + 0.005750754848122597, + 0.007988645695149899, + 0.009435484185814857, + 0.00548205804079771, + 0.005090630613267422, + 0.004627120681107044, + 0.007318780291825533, + 0.00923097226768732, + 0.007988393306732178, + 0.01210961863398552, + 0.013817545026540756, + 0.016015728935599327, + 0.012055200524628162, + 0.010948210023343563, + 0.01497780904173851, + 0.019480472430586815, + 0.008028802461922169, + 0.006953561678528786, + 0.011390064842998981, + 0.009417115710675716, + 0.00999100785702467, + 0.013376270420849323, + 0.02340610697865486, + 0.024435900151729584, + 0.020420050248503685, + 0.022828184068202972, + 0.02422911673784256, + 0.028386831283569336, + 0.01501529198139906, + 0.026541471481323242, + 0.017545191571116447, + 0.016666704788804054, + 0.013615109957754612, + 0.007076551206409931, + 0.01604151912033558, + 0.010804391466081142, + 0.007761557586491108, + 0.009500040672719479, + 0.007491342257708311, + 0.009929611347615719 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 8, + 1 + ], + "selected_reward": -0.14278382062911987 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a futuristically designed heavy battle tank?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the tank's hull dark gray, scratched, and covered in pipes and radiator grilles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the tank performing a wide-angle power drift with spinning muddy tracks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are there dazzling orange-red sparks and a water curtain created by the tank's movement?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera pan rapidly in an arcing trajectory while maintaining a low-angle close-up?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0160652007907629 + ], + "opposite_logprob_tokens": [ + -4.1410651206970215 + ], + "logprob_reward": -0.0160652007907629, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.0160652007907629, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.053073931485414505, + 0.060061559081077576, + 0.04727809876203537, + 0.06758255511522293, + 0.2058396190404892, + 0.041293371468782425, + 0.06010660529136658, + 0.0993151143193245, + 0.3207724690437317, + 0.04467667639255524 + ], + "credit_map": [ + 0.007772224955260754, + 0.0052182930521667, + 0.00495067099109292, + 0.005334861110895872, + 0.004851959180086851, + 0.0033418438397347927, + 0.010244199074804783, + 0.01135987974703312, + 0.007024598773568869, + 0.006996197160333395, + 0.0065459017641842365, + 0.0066184853203594685, + 0.008780558593571186, + 0.008680745027959347, + 0.007067733909934759, + 0.008347337134182453, + 0.004663689061999321, + 0.004430484026670456, + 0.006950849667191505, + 0.009979832917451859, + 0.004365239292383194, + 0.0032915996853262186, + 0.006651405710726976, + 0.006944999098777771, + 0.009927053935825825, + 0.006148055195808411, + 0.019040239974856377, + 0.011399691924452782, + 0.005111479200422764, + 0.004841207526624203, + 0.005250837653875351, + 0.005863991100341082, + 0.05624458193778992, + 0.010991349816322327, + 0.061410509049892426, + 0.017241543158888817, + 0.011328290216624737, + 0.024378318339586258, + 0.011867629364132881, + 0.01237740833312273, + 0.004604497458785772, + 0.00364427687600255, + 0.010043637827038765, + 0.004881484899669886, + 0.00424384418874979, + 0.0035730144008994102, + 0.005378203932195902, + 0.004924414679408073, + 0.0034111677668988705, + 0.005637281108647585, + 0.007262377068400383, + 0.006909061223268509, + 0.009702272713184357, + 0.007204852998256683, + 0.007409602403640747, + 0.012569987215101719, + 0.00782844889909029, + 0.005869387648999691, + 0.011277184821665287, + 0.008938707411289215, + 0.012948818504810333, + 0.01163710467517376, + 0.019221369177103043, + 0.02159409038722515, + 0.03540128469467163, + 0.05142008513212204, + 0.0510529950261116, + 0.03901524841785431, + 0.036613062024116516, + 0.031491801142692566, + 0.03751913458108902, + 0.038258858025074005, + 0.009556896984577179, + 0.003386532887816429, + 0.011525638401508331, + 0.004670721013098955, + 0.0036282408982515335, + 0.0042268973775208, + 0.00396905979141593, + 0.0037126883398741484 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 4, + 7 + ], + "selected_reward": -0.0160652007907629 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a futuristically designed heavy battle tank?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the tank's hull dark gray, scratched, and covered in pipes and radiator grilles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the tank performing a wide-angle power drift with spinning muddy tracks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are there dazzling orange-red sparks and a water curtain created by the tank's movement?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera pan rapidly in an arcing trajectory while maintaining a low-angle close-up?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04864637553691864 + ], + "opposite_logprob_tokens": [ + -3.0486464500427246 + ], + "logprob_reward": -0.04864637553691864, + "yes_no_margin_reward": 3.0, + "token_credit_reward": -0.04864637553691864, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1029326543211937, + 0.06456128507852554, + 0.10016512125730515, + 0.13320709764957428, + 0.08147332817316055, + 0.0397452749311924, + 0.11718298494815826, + 0.09593319147825241, + 0.11624988913536072, + 0.14854909479618073 + ], + "credit_map": [ + 0.007696086075156927, + 0.009813263081014156, + 0.010710090398788452, + 0.012917987070977688, + 0.008950851857662201, + 0.009524362161755562, + 0.0221879705786705, + 0.02113203890621662, + 0.007920442149043083, + 0.006471908185631037, + 0.007003579288721085, + 0.009295432828366756, + 0.008888207376003265, + 0.008546779863536358, + 0.007914561778306961, + 0.008520377799868584, + 0.010057107545435429, + 0.008997605182230473, + 0.019030718132853508, + 0.016725704073905945, + 0.011962685734033585, + 0.010329073294997215, + 0.012240560725331306, + 0.010821659117937088, + 0.0164705328643322, + 0.013943278230726719, + 0.030246950685977936, + 0.0234804917126894, + 0.010187015868723392, + 0.011957338079810143, + 0.011819406412541866, + 0.015102090314030647, + 0.019210973754525185, + 0.0067906915210187435, + 0.020810585469007492, + 0.007085068617016077, + 0.004810716025531292, + 0.009951498359441757, + 0.005896930582821369, + 0.0069168685004115105, + 0.003658117027953267, + 0.00496817147359252, + 0.008039010688662529, + 0.0042358459904789925, + 0.0044190059415996075, + 0.00397444749251008, + 0.004549037665128708, + 0.005901636090129614, + 0.012021656148135662, + 0.015400253236293793, + 0.013035002164542675, + 0.013923686929047108, + 0.019702643156051636, + 0.012568425387144089, + 0.015528175979852676, + 0.015003151260316372, + 0.009424891322851181, + 0.007296949625015259, + 0.01048901118338108, + 0.009398175403475761, + 0.012652427889406681, + 0.011013777926564217, + 0.018084140494465828, + 0.017573818564414978, + 0.008193775080144405, + 0.010381040163338184, + 0.022310657426714897, + 0.025393890216946602, + 0.009382531978189945, + 0.01661941036581993, + 0.012229152955114841, + 0.011739428155124187, + 0.03935873508453369, + 0.008105686865746975, + 0.046837933361530304, + 0.012204001657664776, + 0.012874260544776917, + 0.010495715774595737, + 0.009199751541018486, + 0.009473007172346115 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 3, + 6 + ], + "selected_reward": -0.04864637553691864 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "striking low-angle close-up shot follows, the camera panning rapidly in an arcing trajectory to capture a futuristically designed heavy main battle tank. Its angular armored hull, constructed of a dark gray, scratched composite material, is covered in a complex web of pipes and radiator grilles. Under a gloomy sky, icy raindrops fall, drenching the concrete plaza of an abandoned industrial area. With astonishing agility belying its size, the behemoth performs a fierce, wide-angle power drift, its engine roaring deafeningly and its exhaust spewing searing streams of air. Its wide, muddy metal tracks spin frantically on the slippery concrete, grinding violently against the ground, sending forth a trail of dazzling orange-red sparks that briefly illuminate the metal details and splashing water on the tank's side. The violent turning of the tracks splashed the accumulated water on the ground violently in all directions, forming a spectacular water curtain. The whole scene was ignited by hot sparks in the cold blue-gray tones, full of primitive power, tense dynamic aesthetics and the cool atmosphere of heavy industry.", + "seed": 12853, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000027/group_5/index_4364_seed_12853_dp_11_rank_22_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 5, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.049121370818465945, + "vlm_logprob": -0.049121370818465945, + "vlm_yes_no_margin": 3.725, + "vlm_token_credit": -0.049121370818465945, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a futuristically designed heavy battle tank?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the tank's hull dark gray, scratched, and covered in pipes and radiator grilles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the tank performing a wide-angle power drift with spinning muddy tracks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are there dazzling orange-red sparks and a water curtain created by the tank's movement?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera pan rapidly in an arcing trajectory while maintaining a low-angle close-up?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.025214532390236855 + ], + "opposite_logprob_tokens": [ + -4.275214672088623 + ], + "logprob_reward": -0.025214532390236855, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.025214532390236855, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.18830528855323792, + 0.07543376833200455, + 0.0752730742096901, + 0.07546474039554596, + 0.09104390442371368, + 0.07111059129238129, + 0.08901561051607132, + 0.13344800472259521, + 0.09045706689357758, + 0.11044790595769882 + ], + "credit_map": [ + 0.03475506231188774, + 0.020903196185827255, + 0.024449899792671204, + 0.018930813297629356, + 0.016629023477435112, + 0.022890308871865273, + 0.019751885905861855, + 0.02999511919915676, + 0.00971776619553566, + 0.009157469496130943, + 0.012530853971838951, + 0.009416831657290459, + 0.007770739030092955, + 0.008893515914678574, + 0.008711427450180054, + 0.00923516321927309, + 0.008108493871986866, + 0.0077046859078109264, + 0.014938048087060452, + 0.009734828025102615, + 0.00950000062584877, + 0.008014998398721218, + 0.007734283804893494, + 0.009537734091281891, + 0.01042721513658762, + 0.0066312276758253574, + 0.010586674325168133, + 0.01023518480360508, + 0.006387120578438044, + 0.0070298402570188046, + 0.010967505164444447, + 0.013199969194829464, + 0.009416477754712105, + 0.011267676018178463, + 0.017673812806606293, + 0.01296229101717472, + 0.01173787284642458, + 0.00905606895685196, + 0.010228907689452171, + 0.008700801059603691, + 0.010572806932032108, + 0.0067574153654277325, + 0.008831970393657684, + 0.008779425173997879, + 0.004371560178697109, + 0.010905811563134193, + 0.008286830969154835, + 0.012604774907231331, + 0.010240243747830391, + 0.007472861558198929, + 0.010480697266757488, + 0.027638854458928108, + 0.009757744148373604, + 0.0050134542398154736, + 0.010008268058300018, + 0.008403480984270573, + 0.01058439165353775, + 0.018220216035842896, + 0.01882162317633629, + 0.01694602333009243, + 0.020052872598171234, + 0.010519616305828094, + 0.02317715995013714, + 0.015126103535294533, + 0.007702763192355633, + 0.004655654542148113, + 0.01579798199236393, + 0.011473198421299458, + 0.009001345373690128, + 0.004163633566349745, + 0.029204735532402992, + 0.008457758463919163, + 0.01217507105320692, + 0.010888260789215565, + 0.014187040738761425, + 0.012557663023471832, + 0.021893398836255074, + 0.013353617861866951, + 0.014007749035954475, + 0.011385097168385983 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 7, + 9 + ], + "selected_reward": -0.025214532390236855 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a futuristically designed heavy battle tank?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the tank's hull dark gray, scratched, and covered in pipes and radiator grilles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the tank performing a wide-angle power drift with spinning muddy tracks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are there dazzling orange-red sparks and a water curtain created by the tank's movement?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera pan rapidly in an arcing trajectory while maintaining a low-angle close-up?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.006760936230421066 + ], + "opposite_logprob_tokens": [ + -5.006761074066162 + ], + "logprob_reward": -0.006760936230421066, + "yes_no_margin_reward": 5.0, + "token_credit_reward": -0.006760936230421066, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1245768591761589, + 0.07015948742628098, + 0.04899973422288895, + 0.07637479901313782, + 0.11563935875892639, + 0.05217048525810242, + 0.061360668390989304, + 0.08095168322324753, + 0.25021281838417053, + 0.1195540577173233 + ], + "credit_map": [ + 0.024742428213357925, + 0.01403736975044012, + 0.017856327816843987, + 0.013148795813322067, + 0.008982783183455467, + 0.014871831983327866, + 0.011904262006282806, + 0.019033070653676987, + 0.008359101600944996, + 0.008333353325724602, + 0.00975704099982977, + 0.012369688600301743, + 0.007038509473204613, + 0.008085928857326508, + 0.007621839176863432, + 0.008594030514359474, + 0.00575691182166338, + 0.0038880757056176662, + 0.009691436775028706, + 0.00487924087792635, + 0.004094881936907768, + 0.006619708612561226, + 0.005357368849217892, + 0.00871211290359497, + 0.010480332188308239, + 0.006077381316572428, + 0.01072940044105053, + 0.00769704720005393, + 0.0066953194327652454, + 0.007757175248116255, + 0.01082499697804451, + 0.016113141551613808, + 0.011705928482115269, + 0.012916899286210537, + 0.023465389385819435, + 0.01749143749475479, + 0.016904756426811218, + 0.01163907814770937, + 0.012780915945768356, + 0.008734955452382565, + 0.007572603411972523, + 0.004613813944160938, + 0.005919198039919138, + 0.004556731320917606, + 0.0037526958622038364, + 0.008842309936881065, + 0.005458069499582052, + 0.011455061845481396, + 0.008542366325855255, + 0.008744599297642708, + 0.007111385464668274, + 0.009238357655704021, + 0.006708379369229078, + 0.005960255861282349, + 0.007630012929439545, + 0.007425314746797085, + 0.008487588725984097, + 0.009266315028071404, + 0.014749901369214058, + 0.010540582239627838, + 0.007699977140873671, + 0.0076588415540754795, + 0.013307678513228893, + 0.009240791201591492, + 0.021556712687015533, + 0.010026971809566021, + 0.04929966479539871, + 0.030538789927959442, + 0.024802636355161667, + 0.008128250017762184, + 0.08420756459236145, + 0.021652240306138992, + 0.011103144846856594, + 0.009798672050237656, + 0.013183481059968472, + 0.013947120867669582, + 0.020067084580659866, + 0.011105271987617016, + 0.019940407946705818, + 0.020408881828188896 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 0, + 9 + ], + "selected_reward": -0.006760936230421066 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a futuristically designed heavy battle tank?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the tank's hull dark gray, scratched, and covered in pipes and radiator grilles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the tank performing a wide-angle power drift with spinning muddy tracks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are there dazzling orange-red sparks and a water curtain created by the tank's movement?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera pan rapidly in an arcing trajectory while maintaining a low-angle close-up?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07897402346134186 + ], + "opposite_logprob_tokens": [ + -2.5789740085601807 + ], + "logprob_reward": -0.07897402346134186, + "yes_no_margin_reward": 2.5, + "token_credit_reward": -0.07897402346134186, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.20817963778972626, + 0.03468749672174454, + 0.04104894399642944, + 0.08189452439546585, + 0.08757717907428741, + 0.06876589357852936, + 0.060196541249752045, + 0.03866948187351227, + 0.21311748027801514, + 0.16586284339427948 + ], + "credit_map": [ + 0.029973521828651428, + 0.02025861293077469, + 0.0328231044113636, + 0.028867710381746292, + 0.013803582638502121, + 0.024613285437226295, + 0.019345492124557495, + 0.0384943000972271, + 0.004255739040672779, + 0.003949006088078022, + 0.007146561983972788, + 0.005862188991159201, + 0.0024397564120590687, + 0.003929464146494865, + 0.002735262969508767, + 0.004369518253952265, + 0.005690317135304213, + 0.004594103433191776, + 0.008738958276808262, + 0.005490574520081282, + 0.0032849612180143595, + 0.004645911511033773, + 0.004344417713582516, + 0.0042597027495503426, + 0.009576031006872654, + 0.007102389819920063, + 0.012204847298562527, + 0.010228904895484447, + 0.0059476587921381, + 0.008333050645887852, + 0.012752859853208065, + 0.015748782083392143, + 0.008996651507914066, + 0.008850788697600365, + 0.01892790012061596, + 0.013205403462052345, + 0.009873932227492332, + 0.008439320139586926, + 0.010026673786342144, + 0.009256509132683277, + 0.008303887210786343, + 0.005959538742899895, + 0.008220046758651733, + 0.0069548808969557285, + 0.004901679698377848, + 0.010967990383505821, + 0.008079796098172665, + 0.015378073789179325, + 0.0049296035431325436, + 0.00559918163344264, + 0.007295358460396528, + 0.021354304626584053, + 0.004812389146536589, + 0.0035069871228188276, + 0.007505184039473534, + 0.005193529184907675, + 0.003948523662984371, + 0.00456607760861516, + 0.007917257957160473, + 0.0044586448930203915, + 0.004181486554443836, + 0.00339772691950202, + 0.005877627991139889, + 0.0043221390806138515, + 0.018679002299904823, + 0.008791190572082996, + 0.04075659066438675, + 0.025245167315006256, + 0.021298842504620552, + 0.006253099534660578, + 0.07327631860971451, + 0.01881728507578373, + 0.014636275358498096, + 0.014311006292700768, + 0.01836112141609192, + 0.01650516502559185, + 0.03460262715816498, + 0.021030984818935394, + 0.025193383917212486, + 0.021222280338406563 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 0, + 9 + ], + "selected_reward": -0.07897402346134186 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a futuristically designed heavy battle tank?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the tank's hull dark gray, scratched, and covered in pipes and radiator grilles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the tank performing a wide-angle power drift with spinning muddy tracks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are there dazzling orange-red sparks and a water curtain created by the tank's movement?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera pan rapidly in an arcing trajectory while maintaining a low-angle close-up?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0076261102221906185 + ], + "opposite_logprob_tokens": [ + -4.882626056671143 + ], + "logprob_reward": -0.0076261102221906185, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.0076261102221906185, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.33246633410453796, + 0.0647551491856575, + 0.05251377448439598, + 0.07051872462034225, + 0.07327792048454285, + 0.07332094758749008, + 0.03226820006966591, + 0.07926168292760849, + 0.06774764508008957, + 0.15386955440044403 + ], + "credit_map": [ + 0.04714800417423248, + 0.02903609536588192, + 0.053887493908405304, + 0.031749967485666275, + 0.020979514345526695, + 0.04039103910326958, + 0.032881297171115875, + 0.07639294862747192, + 0.008687124587595463, + 0.009217984043061733, + 0.010796873830258846, + 0.008240120485424995, + 0.007289799861609936, + 0.006877437699586153, + 0.005757330916821957, + 0.007888480089604855, + 0.008355401456356049, + 0.006300881505012512, + 0.011142314411699772, + 0.0071154069155454636, + 0.006591873709112406, + 0.00413850974291563, + 0.003491631243377924, + 0.005377760156989098, + 0.007803579326719046, + 0.008860910311341286, + 0.009937116876244545, + 0.008594433777034283, + 0.006048248615115881, + 0.006774492561817169, + 0.010476413182914257, + 0.01202353648841381, + 0.00745140528306365, + 0.009271320886909962, + 0.01263460423797369, + 0.009200259111821651, + 0.01051477063447237, + 0.007955210283398628, + 0.007715263403952122, + 0.008535097353160381, + 0.011051438748836517, + 0.006448644679039717, + 0.009764039888978004, + 0.010154279880225658, + 0.004638195037841797, + 0.005629329476505518, + 0.010610793717205524, + 0.015024243853986263, + 0.0069290646351873875, + 0.003769525559619069, + 0.0036225931253284216, + 0.005200113169848919, + 0.003946807701140642, + 0.002614916069433093, + 0.00329390075057745, + 0.002891283016651869, + 0.00717254588380456, + 0.00946244690567255, + 0.01309308223426342, + 0.010252686217427254, + 0.015118880197405815, + 0.006609310396015644, + 0.008699802681803703, + 0.008852939121425152, + 0.005616798531264067, + 0.00443083792924881, + 0.011904007755219936, + 0.008159957826137543, + 0.0067743463441729546, + 0.0031042215414345264, + 0.02174202911555767, + 0.0060154590755701065, + 0.013252866454422474, + 0.014396394602954388, + 0.01681612618267536, + 0.02019508183002472, + 0.02899874374270439, + 0.024081947281956673, + 0.01919350028038025, + 0.016934901475906372 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 7 + ], + "selected_reward": -0.0076261102221906185 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a futuristically designed heavy battle tank?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the tank's hull dark gray, scratched, and covered in pipes and radiator grilles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the tank performing a wide-angle power drift with spinning muddy tracks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are there dazzling orange-red sparks and a water curtain created by the tank's movement?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera pan rapidly in an arcing trajectory while maintaining a low-angle close-up?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.12703125178813934 + ], + "opposite_logprob_tokens": [ + -2.1270313262939453 + ], + "logprob_reward": -0.12703125178813934, + "yes_no_margin_reward": 2.0, + "token_credit_reward": -0.12703125178813934, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11205030232667923, + 0.07318516820669174, + 0.0351741760969162, + 0.04519433528184891, + 0.07788405567407608, + 0.07355427742004395, + 0.056523554027080536, + 0.06738073378801346, + 0.4051610231399536, + 0.05389237776398659 + ], + "credit_map": [ + 0.016058960929512978, + 0.010944996029138565, + 0.015075815841555595, + 0.014786330983042717, + 0.012244273908436298, + 0.014427464455366135, + 0.009220308624207973, + 0.019292151555418968, + 0.013556882739067078, + 0.00813218578696251, + 0.010079575702548027, + 0.013164209201931953, + 0.006060239393264055, + 0.008298153057694435, + 0.0068825711496174335, + 0.007011350244283676, + 0.003544504987075925, + 0.004602648317813873, + 0.005056428257375956, + 0.004438432399183512, + 0.005547946318984032, + 0.004333649296313524, + 0.003442732384428382, + 0.004207836464047432, + 0.004661364946514368, + 0.00682862289249897, + 0.005039040930569172, + 0.0075908019207417965, + 0.003950077574700117, + 0.004705679602921009, + 0.006161929108202457, + 0.0062568169087171555, + 0.00930394884198904, + 0.00791726354509592, + 0.014161019586026669, + 0.012441063299775124, + 0.006574271712452173, + 0.007633273024111986, + 0.00970419216901064, + 0.01014902163296938, + 0.00989777036011219, + 0.00662551773712039, + 0.01321434136480093, + 0.011010188609361649, + 0.0037740934640169144, + 0.007681729272007942, + 0.01158929243683815, + 0.009761346504092216, + 0.005010420922189951, + 0.005250168964266777, + 0.008493379689753056, + 0.013671050779521465, + 0.004137159325182438, + 0.0035360935144126415, + 0.008444317616522312, + 0.00798096228390932, + 0.006855436135083437, + 0.010345440357923508, + 0.01281776838004589, + 0.007662653457373381, + 0.007158973719924688, + 0.00777275487780571, + 0.007268299348652363, + 0.007499408908188343, + 0.03528691455721855, + 0.020842380821704865, + 0.07462162524461746, + 0.04991060495376587, + 0.03941572830080986, + 0.015182731673121452, + 0.1335938572883606, + 0.03630715608596802, + 0.004726977087557316, + 0.004926647990942001, + 0.004719176795333624, + 0.007341618184000254, + 0.009255037643015385, + 0.0058394004590809345, + 0.00846033450216055, + 0.008623184636235237 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 0, + 4 + ], + "selected_reward": -0.12703125178813934 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000027/group_6/index_tensor([544, 544])_group_6_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000027/group_6/index_tensor([544, 544])_group_6_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..6648f3d029ed033c641a33a01cd2a6154a0c0897 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000027/group_6/index_tensor([544, 544])_group_6_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "low-angle, long, dynamic tracking shot follows a knight as he rides a majestic black steed at full speed along an ancient bluestone path. The knight is clad in a deep blue, coarse linen robe, its sleeves and hem dancing wildly in the wind. His hair is tied up in a high bun, strands of it rustling against his resolute face. He leans forward slightly, one hand firmly gripping the reins, the scabbard of the longsword at his waist swaying rhythmically with the horse's jerk. Beneath his steed, its muscles bulging, its glossy fur shimmering in the light. Its hooves fly, their collisions with the bluestone pavement sending up dust and scattered sparks. They speed past a cluster of ancient, pristine Hui-style buildings. The whitewashed walls and black tiles appear remarkably tranquil in the warm twilight. The staggered, undulating horse-head walls create a rhythmic skyline, while details of carved wooden window lattices flash by in the rapid movement. The afterglow of the setting sun dyes the entire picture with a warm golden color, lengthening the shadows of the buildings and the riders, creating a desolate and classical martial arts style atmosphere full of heroic spirit, with a strong contrast between movement and stillness.", + "seed": 3825, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000027/group_6/index_544_seed_3825_dp_12_rank_24_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 6, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.15817728713154794, + "vlm_logprob": -0.15817728713154794, + "vlm_yes_no_margin": 2.075, + "vlm_token_credit": -0.15817728713154794, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot taken from a low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a knight riding a black steed at full speed visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the knight wearing a deep blue, coarse linen robe with sleeves and hem moving in the wind?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the scene include ancient Hui-style buildings with whitewashed walls and black tiles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the lighting warm and golden, resembling the afterglow of a setting sun?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.21477197110652924 + ], + "opposite_logprob_tokens": [ + -1.8397719860076904 + ], + "logprob_reward": -0.21477197110652924, + "yes_no_margin_reward": 1.625, + "token_credit_reward": -0.21477197110652924, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11376354098320007, + 0.11167670786380768, + 0.08821001648902893, + 0.08163614571094513, + 0.09963556379079819, + 0.08597786724567413, + 0.08572372794151306, + 0.11833642423152924, + 0.15379193425178528, + 0.06124807521700859 + ], + "credit_map": [ + 0.01732955500483513, + 0.012592819519340992, + 0.015154699794948101, + 0.010957764461636543, + 0.008873329497873783, + 0.02259482629597187, + 0.014830674044787884, + 0.011429863050580025, + 0.008575326763093472, + 0.013305941596627235, + 0.016065074130892754, + 0.01422853209078312, + 0.017064368352293968, + 0.009404399432241917, + 0.02303079329431057, + 0.01000227127224207, + 0.012757338583469391, + 0.010862469673156738, + 0.009343637153506279, + 0.009649602696299553, + 0.012351270765066147, + 0.011126729659736156, + 0.014536007307469845, + 0.007582958787679672, + 0.008401011116802692, + 0.007742144167423248, + 0.015320006757974625, + 0.012098409235477448, + 0.007982192561030388, + 0.00789876189082861, + 0.015041825361549854, + 0.0071517894975841045, + 0.008494208566844463, + 0.02090306766331196, + 0.012494821101427078, + 0.011563237756490707, + 0.010846113786101341, + 0.009412525221705437, + 0.013315810821950436, + 0.012605776078999043, + 0.017298614606261253, + 0.012447548098862171, + 0.008085074834525585, + 0.009165490046143532, + 0.007725263945758343, + 0.01438091415911913, + 0.005576579365879297, + 0.011298376135528088, + 0.009535584598779678, + 0.011126628145575523, + 0.007832435891032219, + 0.009287760592997074, + 0.01424600463360548, + 0.014581459574401379, + 0.007488640956580639, + 0.011625207960605621, + 0.0079905791208148, + 0.017540618777275085, + 0.01177382655441761, + 0.01952018216252327, + 0.01593242771923542, + 0.02182704582810402, + 0.007801343686878681, + 0.015950404107570648, + 0.01491107139736414, + 0.013219698332250118, + 0.0192507766187191, + 0.018873317167162895, + 0.020465221256017685, + 0.03435644507408142, + 0.012255304493010044, + 0.020460093393921852, + 0.006753212306648493, + 0.006342646665871143, + 0.0048869214951992035, + 0.006431193090975285, + 0.008077390491962433, + 0.010616946034133434, + 0.009312194772064686, + 0.008827564306557178 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 7, + 0 + ], + "selected_reward": -0.21477197110652924 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot taken from a low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a knight riding a black steed at full speed visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the knight wearing a deep blue, coarse linen robe with sleeves and hem moving in the wind?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the scene include ancient Hui-style buildings with whitewashed walls and black tiles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the lighting warm and golden, resembling the afterglow of a setting sun?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03814457729458809 + ], + "opposite_logprob_tokens": [ + -3.288144588470459 + ], + "logprob_reward": -0.03814457729458809, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.03814457729458809, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1534416228532791, + 0.09634838253259659, + 0.030951233580708504, + 0.047108810395002365, + 0.10512740910053253, + 0.02909652143716812, + 0.04316499084234238, + 0.10953648388385773, + 0.3345724642276764, + 0.05065206065773964 + ], + "credit_map": [ + 0.01754939742386341, + 0.018803423270583153, + 0.02468189410865307, + 0.014688830822706223, + 0.014085828326642513, + 0.01760413870215416, + 0.03174690529704094, + 0.01428119745105505, + 0.009634584188461304, + 0.008939653635025024, + 0.022111963480710983, + 0.014078267849981785, + 0.008087970316410065, + 0.010288631543517113, + 0.01397429034113884, + 0.009233018383383751, + 0.004074377939105034, + 0.005025036167353392, + 0.0035719953011721373, + 0.0039602601900696754, + 0.004761686082929373, + 0.0027438171673566103, + 0.004448855295777321, + 0.002365205669775605, + 0.003899055765941739, + 0.007894010283052921, + 0.0055092754773795605, + 0.007291367277503014, + 0.006208216771483421, + 0.004748660139739513, + 0.00647321529686451, + 0.005085008218884468, + 0.014340480789542198, + 0.013487409800291061, + 0.013893401250243187, + 0.012502902187407017, + 0.012129629030823708, + 0.012146871536970139, + 0.013224453665316105, + 0.01340226549655199, + 0.004619230516254902, + 0.0032939640805125237, + 0.0035952581092715263, + 0.0038817995227873325, + 0.0036025000736117363, + 0.0035266170743852854, + 0.0027951968368142843, + 0.003781956387683749, + 0.005057075526565313, + 0.005665590520948172, + 0.0043157306499779224, + 0.005342353135347366, + 0.0051184529438614845, + 0.007829831913113594, + 0.0038395116571336985, + 0.005996442399919033, + 0.009552868083119392, + 0.011906708590686321, + 0.010752258822321892, + 0.011053947731852531, + 0.014508262276649475, + 0.02710873633623123, + 0.008900902234017849, + 0.015752805396914482, + 0.030121713876724243, + 0.0348927304148674, + 0.02835264429450035, + 0.028767086565494537, + 0.04212344065308571, + 0.0945158377289772, + 0.020535795018076897, + 0.05526323243975639, + 0.006307078525424004, + 0.006789689417928457, + 0.0043143886141479015, + 0.006014787591993809, + 0.0056191422045230865, + 0.0071375067345798016, + 0.007457219064235687, + 0.007012244779616594 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 0, + 7 + ], + "selected_reward": -0.03814457729458809 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot taken from a low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a knight riding a black steed at full speed visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the knight wearing a deep blue, coarse linen robe with sleeves and hem moving in the wind?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the scene include ancient Hui-style buildings with whitewashed walls and black tiles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the lighting warm and golden, resembling the afterglow of a setting sun?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.05498340725898743 + ], + "opposite_logprob_tokens": [ + -2.929983377456665 + ], + "logprob_reward": -0.05498340725898743, + "yes_no_margin_reward": 2.875, + "token_credit_reward": -0.05498340725898743, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.19615265727043152, + 0.0816771388053894, + 0.18064647912979126, + 0.12273062020540237, + 0.07492608577013016, + 0.062175896018743515, + 0.05072404071688652, + 0.05601361766457558, + 0.09862679243087769, + 0.07632669061422348 + ], + "credit_map": [ + 0.036026086658239365, + 0.03792424127459526, + 0.01615876704454422, + 0.023703306913375854, + 0.017428625375032425, + 0.02413739636540413, + 0.02303694747388363, + 0.017737293615937233, + 0.011819714680314064, + 0.009389913640916348, + 0.0067459107376635075, + 0.016088025644421577, + 0.010820837691426277, + 0.008278066292405128, + 0.011631961911916733, + 0.006902711465954781, + 0.032670896500349045, + 0.02292478084564209, + 0.01491559762507677, + 0.01853976584970951, + 0.02295687235891819, + 0.026490801945328712, + 0.024102913215756416, + 0.018044844269752502, + 0.012816373258829117, + 0.0157606303691864, + 0.01856287010014057, + 0.018901832401752472, + 0.014409290626645088, + 0.010049906559288502, + 0.02018500678241253, + 0.012044710107147694, + 0.011271808296442032, + 0.015928560867905617, + 0.006472080014646053, + 0.006131602916866541, + 0.007382931653410196, + 0.010533798485994339, + 0.007031576707959175, + 0.010173726826906204, + 0.013726557604968548, + 0.00657655531540513, + 0.005186144728213549, + 0.007144440896809101, + 0.006102367304265499, + 0.0103661073371768, + 0.003683190792798996, + 0.009390532970428467, + 0.005688138771802187, + 0.007782820146530867, + 0.005134060978889465, + 0.007205672562122345, + 0.0059334952384233475, + 0.007221555337309837, + 0.004609543830156326, + 0.00714875478297472, + 0.005410191603004932, + 0.006639407481998205, + 0.005301701370626688, + 0.008781136013567448, + 0.010081164538860321, + 0.008223197422921658, + 0.00448153680190444, + 0.00709528336301446, + 0.008284460753202438, + 0.02204911783337593, + 0.010380827821791172, + 0.012483200058341026, + 0.010655367746949196, + 0.01441015675663948, + 0.010696266777813435, + 0.009667393751442432, + 0.005951137747615576, + 0.013824515976011753, + 0.005586700513958931, + 0.006745252758264542, + 0.009305237792432308, + 0.013875814154744148, + 0.010677541606128216, + 0.010360494256019592 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 3 + ], + "selected_reward": -0.05498340725898743 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot taken from a low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a knight riding a black steed at full speed visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the knight wearing a deep blue, coarse linen robe with sleeves and hem moving in the wind?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the scene include ancient Hui-style buildings with whitewashed walls and black tiles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the lighting warm and golden, resembling the afterglow of a setting sun?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.20160961151123047 + ], + "opposite_logprob_tokens": [ + -1.7016096115112305 + ], + "logprob_reward": -0.20160961151123047, + "yes_no_margin_reward": 1.5, + "token_credit_reward": -0.20160961151123047, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07857071608304977, + 0.058865632861852646, + 0.07287228852510452, + 0.04171156883239746, + 0.09975086897611618, + 0.16503994166851044, + 0.03404279798269272, + 0.030033884570002556, + 0.38089045882225037, + 0.0382218211889267 + ], + "credit_map": [ + 0.011717227287590504, + 0.009597832337021828, + 0.0096132792532444, + 0.010039214044809341, + 0.006704903673380613, + 0.012368626892566681, + 0.011008929461240768, + 0.00752069428563118, + 0.005512576084583998, + 0.0067010121420025826, + 0.008735560812056065, + 0.008323037065565586, + 0.008086525835096836, + 0.005897973198443651, + 0.010739505290985107, + 0.004869435913860798, + 0.006788914557546377, + 0.007745622657239437, + 0.00925629585981369, + 0.012289843522012234, + 0.010712355375289917, + 0.007239950355142355, + 0.013195781037211418, + 0.005643523298203945, + 0.004254461266100407, + 0.0034368091728538275, + 0.006946241017431021, + 0.004815561231225729, + 0.0037882484029978514, + 0.008573322556912899, + 0.00492173433303833, + 0.004975184332579374, + 0.015018883161246777, + 0.011990542523562908, + 0.0063756126910448074, + 0.011368448846042156, + 0.01137718465179205, + 0.01788724586367607, + 0.011386396363377571, + 0.014346549287438393, + 0.03621719405055046, + 0.021166551858186722, + 0.0137483486905694, + 0.013672882691025734, + 0.014875037595629692, + 0.03478027507662773, + 0.007758416701108217, + 0.02282121777534485, + 0.002812919206917286, + 0.0037412778474390507, + 0.0030823335982859135, + 0.003554949536919594, + 0.006241876631975174, + 0.0074315485544502735, + 0.002359695266932249, + 0.004818191286176443, + 0.0026161044370383024, + 0.003654013155028224, + 0.003628713544458151, + 0.004229745827615261, + 0.003105991752818227, + 0.006352178752422333, + 0.0021872420329600573, + 0.004259894136339426, + 0.03848910331726074, + 0.02611696720123291, + 0.022330718114972115, + 0.029340464621782303, + 0.05678342282772064, + 0.12003211677074432, + 0.02242415025830269, + 0.0653735026717186, + 0.003408205695450306, + 0.0034005765337496996, + 0.0028355070389807224, + 0.003062807023525238, + 0.004339237697422504, + 0.010742219164967537, + 0.004070770926773548, + 0.006362491752952337 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 5, + 4 + ], + "selected_reward": -0.20160961151123047 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot taken from a low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a knight riding a black steed at full speed visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the knight wearing a deep blue, coarse linen robe with sleeves and hem moving in the wind?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the scene include ancient Hui-style buildings with whitewashed walls and black tiles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the lighting warm and golden, resembling the afterglow of a setting sun?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.2813768684864044 + ], + "opposite_logprob_tokens": [ + -1.406376838684082 + ], + "logprob_reward": -0.2813768684864044, + "yes_no_margin_reward": 1.125, + "token_credit_reward": -0.2813768684864044, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10548733919858932, + 0.035067468881607056, + 0.036019466817379, + 0.14153102040290833, + 0.2656400203704834, + 0.061898574233055115, + 0.03149785101413727, + 0.023166846483945847, + 0.14872832596302032, + 0.15096303820610046 + ], + "credit_map": [ + 0.02012462355196476, + 0.016803501173853874, + 0.008576595224440098, + 0.010806243866682053, + 0.011282573454082012, + 0.013930236920714378, + 0.014525522477924824, + 0.009438054636120796, + 0.0030667108949273825, + 0.0036845447029918432, + 0.005214274395257235, + 0.005954813212156296, + 0.00391598604619503, + 0.003241361817345023, + 0.006626565009355545, + 0.0033632172271609306, + 0.004440571181476116, + 0.004900305066257715, + 0.0047637708485126495, + 0.005089381244033575, + 0.004913841374218464, + 0.003926104865968227, + 0.004686754662543535, + 0.003298741765320301, + 0.02057192102074623, + 0.013273707590997219, + 0.02268887497484684, + 0.019223401322960854, + 0.01914829947054386, + 0.014080257155001163, + 0.02027245983481407, + 0.012272116728127003, + 0.041473377496004105, + 0.032831914722919464, + 0.03189821541309357, + 0.0331709086894989, + 0.02691470831632614, + 0.03244464471936226, + 0.030847638845443726, + 0.036058634519577026, + 0.012426459230482578, + 0.007974685169756413, + 0.005402238108217716, + 0.0063265180215239525, + 0.005663692019879818, + 0.011553521268069744, + 0.004388739820569754, + 0.008162724785506725, + 0.0028831358067691326, + 0.003239312442019582, + 0.002595219062641263, + 0.0034320540726184845, + 0.005562825594097376, + 0.006373437587171793, + 0.002754032611846924, + 0.004657834302634001, + 0.001924141077324748, + 0.004563387483358383, + 0.0022765432950109243, + 0.0028316809330135584, + 0.002573676872998476, + 0.0031386855989694595, + 0.0027036231476813555, + 0.0031551106367260218, + 0.014231640845537186, + 0.01423052791506052, + 0.0117747588083148, + 0.013856643810868263, + 0.014590735547244549, + 0.04425020515918732, + 0.011019599623978138, + 0.024774232879281044, + 0.011921381577849388, + 0.01670021563768387, + 0.01218534354120493, + 0.01602630689740181, + 0.022481029853224754, + 0.02496524341404438, + 0.026289211586117744, + 0.02039431780576706 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 9, + 8 + ], + "selected_reward": -0.2813768684864044 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "low-angle, long, dynamic tracking shot follows a knight as he rides a majestic black steed at full speed along an ancient bluestone path. The knight is clad in a deep blue, coarse linen robe, its sleeves and hem dancing wildly in the wind. His hair is tied up in a high bun, strands of it rustling against his resolute face. He leans forward slightly, one hand firmly gripping the reins, the scabbard of the longsword at his waist swaying rhythmically with the horse's jerk. Beneath his steed, its muscles bulging, its glossy fur shimmering in the light. Its hooves fly, their collisions with the bluestone pavement sending up dust and scattered sparks. They speed past a cluster of ancient, pristine Hui-style buildings. The whitewashed walls and black tiles appear remarkably tranquil in the warm twilight. The staggered, undulating horse-head walls create a rhythmic skyline, while details of carved wooden window lattices flash by in the rapid movement. The afterglow of the setting sun dyes the entire picture with a warm golden color, lengthening the shadows of the buildings and the riders, creating a desolate and classical martial arts style atmosphere full of heroic spirit, with a strong contrast between movement and stillness.", + "seed": 3826, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000027/group_6/index_544_seed_3826_dp_12_rank_24_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 6, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.17705197557806968, + "vlm_logprob": -0.17705197557806968, + "vlm_yes_no_margin": 2.175000011920929, + "vlm_token_credit": -0.17705197557806968, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot taken from a low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a knight riding a black steed at full speed visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the knight wearing a deep blue, coarse linen robe with sleeves and hem moving in the wind?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the scene include ancient Hui-style buildings with whitewashed walls and black tiles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the lighting warm and golden, resembling the afterglow of a setting sun?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.42059987783432007 + ], + "opposite_logprob_tokens": [ + -1.2955999374389648 + ], + "logprob_reward": -0.42059987783432007, + "yes_no_margin_reward": 0.8750000596046448, + "token_credit_reward": -0.42059987783432007, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1463140994310379, + 0.06677598506212234, + 0.023325888440012932, + 0.08705414831638336, + 0.30754074454307556, + 0.030881667509675026, + 0.061489593237638474, + 0.06828663498163223, + 0.14021553099155426, + 0.06811568886041641 + ], + "credit_map": [ + 0.010906354524195194, + 0.02119402028620243, + 0.015741845592856407, + 0.021667353808879852, + 0.024260040372610092, + 0.015057750977575779, + 0.018438538536429405, + 0.019048217684030533, + 0.0066566672176122665, + 0.0060100252740085125, + 0.009296735748648643, + 0.010614042170345783, + 0.010882112197577953, + 0.006001456640660763, + 0.006569589022547007, + 0.010745367035269737, + 0.003580676857382059, + 0.0027149326633661985, + 0.0028882583137601614, + 0.002043116604909301, + 0.0031824586912989616, + 0.0027743601240217686, + 0.0030211859848350286, + 0.003120903391391039, + 0.01647576503455639, + 0.01388098206371069, + 0.011096188798546791, + 0.007522355765104294, + 0.012257229536771774, + 0.011683061718940735, + 0.007674155291169882, + 0.006464419886469841, + 0.04719940200448036, + 0.024326540529727936, + 0.05587558075785637, + 0.026196518912911415, + 0.03391770273447037, + 0.06995164602994919, + 0.02478586696088314, + 0.02528752200305462, + 0.0036915999371558428, + 0.002581559354439378, + 0.0036637664306908846, + 0.004078646190464497, + 0.005332828965038061, + 0.0032958146184682846, + 0.0036649578250944614, + 0.0045724972151219845, + 0.007282101549208164, + 0.005903604906052351, + 0.009559590369462967, + 0.006349602714180946, + 0.007598512805998325, + 0.008499319665133953, + 0.007518347818404436, + 0.008778521791100502, + 0.006807032041251659, + 0.009990646503865719, + 0.01115614827722311, + 0.0070974333211779594, + 0.01190680917352438, + 0.009410066530108452, + 0.0061164740473032, + 0.005802036728709936, + 0.016089938580989838, + 0.00959843397140503, + 0.033110614866018295, + 0.01865776814520359, + 0.010987995192408562, + 0.025885557755827904, + 0.011049307882785797, + 0.014835921116173267, + 0.009079963900148869, + 0.009428126737475395, + 0.010943831875920296, + 0.007040721829980612, + 0.009878026321530342, + 0.0065480805933475494, + 0.006854773964732885, + 0.008342166431248188 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 0, + 8 + ], + "selected_reward": -0.42059987783432007 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot taken from a low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a knight riding a black steed at full speed visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the knight wearing a deep blue, coarse linen robe with sleeves and hem moving in the wind?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the scene include ancient Hui-style buildings with whitewashed walls and black tiles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the lighting warm and golden, resembling the afterglow of a setting sun?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.038137808442115784 + ], + "opposite_logprob_tokens": [ + -3.288137912750244 + ], + "logprob_reward": -0.038137808442115784, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.038137808442115784, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.03256456181406975, + 0.054556623101234436, + 0.018082262948155403, + 0.11920621991157532, + 0.5422329902648926, + 0.0479854978621006, + 0.060887470841407776, + 0.010915367864072323, + 0.05078216269612312, + 0.06278680264949799 + ], + "credit_map": [ + 0.003336117370054126, + 0.0032185502350330353, + 0.0037826355546712875, + 0.004589822608977556, + 0.006799512542784214, + 0.0035836060997098684, + 0.0033764236140996218, + 0.00387789448723197, + 0.006814253982156515, + 0.0028412933461368084, + 0.008472822606563568, + 0.007138443645089865, + 0.0058870683424174786, + 0.004472894594073296, + 0.005626156460493803, + 0.01330369058996439, + 0.0028083017095923424, + 0.0016141345258802176, + 0.0020661104936152697, + 0.0018778607482090592, + 0.002374783856794238, + 0.0028023577760905027, + 0.0020853106398135424, + 0.0024534023832529783, + 0.028441397473216057, + 0.012167118489742279, + 0.016178321093320847, + 0.011116878129541874, + 0.013043745420873165, + 0.01911213807761669, + 0.010892901569604874, + 0.00825370941311121, + 0.08228044956922531, + 0.043862856924533844, + 0.09377416968345642, + 0.04701043292880058, + 0.0706215351819992, + 0.11490561068058014, + 0.04565257206559181, + 0.04412542283535004, + 0.005134670063853264, + 0.00429299334064126, + 0.0060958052054047585, + 0.005319320596754551, + 0.008365150541067123, + 0.006643584929406643, + 0.005824500694870949, + 0.006309473421424627, + 0.007833674550056458, + 0.006368840113282204, + 0.009617813862860203, + 0.007728063501417637, + 0.006740901619195938, + 0.007405100855976343, + 0.007669865619391203, + 0.007523209322243929, + 0.0010340223088860512, + 0.002124217338860035, + 0.001261522644199431, + 0.000907976005692035, + 0.0022825761698186398, + 0.0013531014556065202, + 0.0009489241056144238, + 0.0010030274279415607, + 0.004545178730040789, + 0.0036476762033998966, + 0.014096897095441818, + 0.006737994030117989, + 0.0034249969758093357, + 0.009296644479036331, + 0.003600824624300003, + 0.005431948229670525, + 0.011150558479130268, + 0.008759614080190659, + 0.00652473047375679, + 0.007298016920685768, + 0.009403856471180916, + 0.005663067102432251, + 0.006708252709358931, + 0.007278704084455967 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 3, + 9 + ], + "selected_reward": -0.038137808442115784 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot taken from a low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a knight riding a black steed at full speed visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the knight wearing a deep blue, coarse linen robe with sleeves and hem moving in the wind?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the scene include ancient Hui-style buildings with whitewashed walls and black tiles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the lighting warm and golden, resembling the afterglow of a setting sun?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02330542728304863 + ], + "opposite_logprob_tokens": [ + -3.7733054161071777 + ], + "logprob_reward": -0.02330542728304863, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.02330542728304863, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.05352158099412918, + 0.06451225280761719, + 0.057435646653175354, + 0.140887051820755, + 0.15475374460220337, + 0.2831741273403168, + 0.10760002583265305, + 0.020889542996883392, + 0.04566364735364914, + 0.07156240940093994 + ], + "credit_map": [ + 0.004511720035225153, + 0.005588696338236332, + 0.006975006777793169, + 0.009207085706293583, + 0.00973498821258545, + 0.004578274209052324, + 0.00653107650578022, + 0.0063947346061468124, + 0.007313962560147047, + 0.005138047970831394, + 0.008615398779511452, + 0.008860784582793713, + 0.005489019211381674, + 0.0045850262977182865, + 0.007994333282113075, + 0.016515681520104408, + 0.006743030156940222, + 0.00732531538233161, + 0.00728875957429409, + 0.010156526230275631, + 0.007192566525191069, + 0.005571748595684767, + 0.00637474050745368, + 0.006782956887036562, + 0.030936483293771744, + 0.011708445847034454, + 0.014367271214723587, + 0.014001871459186077, + 0.022699153050780296, + 0.016803830862045288, + 0.016828598454594612, + 0.01354140229523182, + 0.02239251136779785, + 0.014772494323551655, + 0.02014004811644554, + 0.01228695921599865, + 0.023390494287014008, + 0.03365250304341316, + 0.015659235417842865, + 0.012459487654268742, + 0.03361619636416435, + 0.030438046902418137, + 0.03758388012647629, + 0.026701245456933975, + 0.060338251292705536, + 0.040799807757139206, + 0.025352301076054573, + 0.028344377875328064, + 0.015095249749720097, + 0.010715365409851074, + 0.01593788154423237, + 0.008952302858233452, + 0.017565827816724777, + 0.019441531971096992, + 0.00971344392746687, + 0.010178428143262863, + 0.0034151109866797924, + 0.002522619441151619, + 0.0033767682034522295, + 0.0024969761725515127, + 0.0020762442145496607, + 0.0027213962748646736, + 0.0019407180370762944, + 0.0023397107142955065, + 0.00712090078741312, + 0.00392635865136981, + 0.013020286336541176, + 0.005985134281218052, + 0.0023539240937680006, + 0.008408427238464355, + 0.0022780171129852533, + 0.002570599317550659, + 0.015631135553121567, + 0.0058951606042683125, + 0.013102537021040916, + 0.00842601340264082, + 0.005837408825755119, + 0.008420108817517757, + 0.0064689298160374165, + 0.007781113497912884 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 4, + 3 + ], + "selected_reward": -0.02330542728304863 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot taken from a low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a knight riding a black steed at full speed visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the knight wearing a deep blue, coarse linen robe with sleeves and hem moving in the wind?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the scene include ancient Hui-style buildings with whitewashed walls and black tiles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the lighting warm and golden, resembling the afterglow of a setting sun?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.20162314176559448 + ], + "opposite_logprob_tokens": [ + -1.7016232013702393 + ], + "logprob_reward": -0.20162314176559448, + "yes_no_margin_reward": 1.5, + "token_credit_reward": -0.20162314176559448, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.05355707183480263, + 0.047369975596666336, + 0.028073957189917564, + 0.20986278355121613, + 0.07454834133386612, + 0.1855333000421524, + 0.07359813898801804, + 0.049402061849832535, + 0.1075274869799614, + 0.170526921749115 + ], + "credit_map": [ + 0.006188653409481049, + 0.006035123486071825, + 0.007312344387173653, + 0.006783479359000921, + 0.008136264979839325, + 0.005516514647752047, + 0.0071038538590073586, + 0.00648083770647645, + 0.006338030565530062, + 0.0038234377279877663, + 0.006983887869864702, + 0.004498871508985758, + 0.009249722585082054, + 0.0045566014014184475, + 0.006395353469997644, + 0.005524073261767626, + 0.005415878724306822, + 0.004092022310942411, + 0.003430232172831893, + 0.0035019079223275185, + 0.0024798389058560133, + 0.0028647498693317175, + 0.002771995263174176, + 0.003517331788316369, + 0.05599731579422951, + 0.025897640734910965, + 0.027895117178559303, + 0.019000811502337456, + 0.013996701687574387, + 0.03400501608848572, + 0.01538920495659113, + 0.01768096536397934, + 0.009520635940134525, + 0.011013300158083439, + 0.00937184039503336, + 0.0075385174714028835, + 0.01284756138920784, + 0.009075646288692951, + 0.007665885612368584, + 0.007514950819313526, + 0.029160480946302414, + 0.017901424318552017, + 0.024042192846536636, + 0.019408538937568665, + 0.028321731835603714, + 0.026752714067697525, + 0.018898077309131622, + 0.021048150956630707, + 0.008750748820602894, + 0.006751093082129955, + 0.010292473249137402, + 0.007233303505927324, + 0.011348621919751167, + 0.012757024727761745, + 0.007883311249315739, + 0.008581563830375671, + 0.0063620018772780895, + 0.008680807426571846, + 0.006057474296540022, + 0.0040419334545731544, + 0.00844617746770382, + 0.006295868661254644, + 0.004175771959125996, + 0.005342026706784964, + 0.0076852403581142426, + 0.004499807953834534, + 0.03884449228644371, + 0.006705510895699263, + 0.004974565468728542, + 0.036945320665836334, + 0.0037165202666074038, + 0.004156026523560286, + 0.044279035180807114, + 0.02076227031648159, + 0.02135780267417431, + 0.013505300506949425, + 0.01691294088959694, + 0.025791846215724945, + 0.012188698165118694, + 0.015729036182165146 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 5, + 9 + ], + "selected_reward": -0.20162314176559448 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot taken from a low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a knight riding a black steed at full speed visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the knight wearing a deep blue, coarse linen robe with sleeves and hem moving in the wind?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the scene include ancient Hui-style buildings with whitewashed walls and black tiles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the lighting warm and golden, resembling the afterglow of a setting sun?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.20159362256526947 + ], + "opposite_logprob_tokens": [ + -1.7015936374664307 + ], + "logprob_reward": -0.20159362256526947, + "yes_no_margin_reward": 1.5, + "token_credit_reward": -0.20159362256526947, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09294869750738144, + 0.036356475204229355, + 0.021840175613760948, + 0.05060593783855438, + 0.38867563009262085, + 0.035389494150877, + 0.04432053864002228, + 0.035606347024440765, + 0.2470792829990387, + 0.04717749357223511 + ], + "credit_map": [ + 0.009120292030274868, + 0.01403891108930111, + 0.014974327757954597, + 0.011761322617530823, + 0.014718160033226013, + 0.008152053691446781, + 0.011845086701214314, + 0.008338541723787785, + 0.0036655531730502844, + 0.0029781972989439964, + 0.005496865604072809, + 0.005330536514520645, + 0.0058131287805736065, + 0.003343715565279126, + 0.00370708922855556, + 0.006021388340741396, + 0.003779028309509158, + 0.0023155524395406246, + 0.002772516803815961, + 0.0024787038564682007, + 0.003378350054845214, + 0.002582302549853921, + 0.0019953525625169277, + 0.0025383683387190104, + 0.013468644581735134, + 0.00471904594451189, + 0.005973957944661379, + 0.004055581986904144, + 0.004935925826430321, + 0.009888350032269955, + 0.0039050052873790264, + 0.0036594229750335217, + 0.05851183831691742, + 0.03296312689781189, + 0.06185634434223175, + 0.03628811612725258, + 0.04156310483813286, + 0.08072403818368912, + 0.03942257910966873, + 0.03734644129872322, + 0.004980853758752346, + 0.003436050610616803, + 0.005601829849183559, + 0.003924298100173473, + 0.0059000956825912, + 0.004089087713509798, + 0.0035833450965583324, + 0.0038739293813705444, + 0.005539225414395332, + 0.00437029218301177, + 0.006709577515721321, + 0.0034664003178477287, + 0.006785151548683643, + 0.007058560382574797, + 0.0053565832786262035, + 0.005034750793129206, + 0.005071195308119059, + 0.005741742439568043, + 0.004528541583567858, + 0.003304731799289584, + 0.00484798988327384, + 0.0057755992747843266, + 0.003412542399019003, + 0.0029240006115287542, + 0.024108272045850754, + 0.017802482470870018, + 0.07186022400856018, + 0.024971943348646164, + 0.017282593995332718, + 0.06089180335402489, + 0.014920169487595558, + 0.015241767279803753, + 0.008106386289000511, + 0.005484952591359615, + 0.006301899440586567, + 0.006625080946832895, + 0.005875132512301207, + 0.004353350028395653, + 0.003995645325630903, + 0.006435045041143894 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 8, + 0 + ], + "selected_reward": -0.20159362256526947 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "low-angle, long, dynamic tracking shot follows a knight as he rides a majestic black steed at full speed along an ancient bluestone path. The knight is clad in a deep blue, coarse linen robe, its sleeves and hem dancing wildly in the wind. His hair is tied up in a high bun, strands of it rustling against his resolute face. He leans forward slightly, one hand firmly gripping the reins, the scabbard of the longsword at his waist swaying rhythmically with the horse's jerk. Beneath his steed, its muscles bulging, its glossy fur shimmering in the light. Its hooves fly, their collisions with the bluestone pavement sending up dust and scattered sparks. They speed past a cluster of ancient, pristine Hui-style buildings. The whitewashed walls and black tiles appear remarkably tranquil in the warm twilight. The staggered, undulating horse-head walls create a rhythmic skyline, while details of carved wooden window lattices flash by in the rapid movement. The afterglow of the setting sun dyes the entire picture with a warm golden color, lengthening the shadows of the buildings and the riders, creating a desolate and classical martial arts style atmosphere full of heroic spirit, with a strong contrast between movement and stillness.", + "seed": 3827, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000027/group_6/index_544_seed_3827_dp_13_rank_26_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 6, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.09581119064241647, + "vlm_logprob": -0.09581119064241647, + "vlm_yes_no_margin": 2.775, + "vlm_token_credit": -0.09581119064241647, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot taken from a low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a knight riding a black steed at full speed visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the knight wearing a deep blue, coarse linen robe with sleeves and hem moving in the wind?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the scene include ancient Hui-style buildings with whitewashed walls and black tiles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the lighting warm and golden, resembling the afterglow of a setting sun?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.18291568756103516 + ], + "opposite_logprob_tokens": [ + -2.057915687561035 + ], + "logprob_reward": -0.18291568756103516, + "yes_no_margin_reward": 1.875, + "token_credit_reward": -0.18291568756103516, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13039547204971313, + 0.10912135243415833, + 0.18263885378837585, + 0.1427004188299179, + 0.040269944816827774, + 0.13620169460773468, + 0.07030880451202393, + 0.0388362854719162, + 0.06108648702502251, + 0.08844071626663208 + ], + "credit_map": [ + 0.018862925469875336, + 0.01593775488436222, + 0.015183540061116219, + 0.012442115694284439, + 0.014181151054799557, + 0.025855762884020805, + 0.012901457026600838, + 0.01503076497465372, + 0.014084257185459137, + 0.010264134965837002, + 0.023731069639325142, + 0.010070716962218285, + 0.007255967240780592, + 0.013705533929169178, + 0.0136607326567173, + 0.0163489431142807, + 0.01663959212601185, + 0.012459713034331799, + 0.023586492985486984, + 0.014977415092289448, + 0.025641029700636864, + 0.0388319306075573, + 0.015053491108119488, + 0.035449206829071045, + 0.010138481855392456, + 0.016567813232541084, + 0.023082543164491653, + 0.019610820338129997, + 0.02691863290965557, + 0.009610570967197418, + 0.016700809821486473, + 0.02007073350250721, + 0.004848252050578594, + 0.006477823480963707, + 0.005350138060748577, + 0.00707640266045928, + 0.003228229470551014, + 0.004958850331604481, + 0.00401372229680419, + 0.004316528793424368, + 0.01746198534965515, + 0.021880635991692543, + 0.013102784752845764, + 0.013825586996972561, + 0.014783279038965702, + 0.023085765540599823, + 0.018281765282154083, + 0.013779892586171627, + 0.0068534198217093945, + 0.009819697588682175, + 0.005078761372715235, + 0.006100070662796497, + 0.014993712306022644, + 0.012310673482716084, + 0.0066681732423603535, + 0.008484291844069958, + 0.0035251581575721502, + 0.005912271793931723, + 0.005656816065311432, + 0.005745638627558947, + 0.0035365489311516285, + 0.0050473520532250404, + 0.005515296012163162, + 0.0038972031325101852, + 0.0069100516848266125, + 0.010758565738797188, + 0.006196307484060526, + 0.0068189785815775394, + 0.010425747372210026, + 0.005688977427780628, + 0.006410181522369385, + 0.00787767767906189, + 0.009574965573847294, + 0.00718859676271677, + 0.01272704266011715, + 0.010702998377382755, + 0.007772849872708321, + 0.023558160290122032, + 0.007264908868819475, + 0.009651195257902145 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 3, + 5 + ], + "selected_reward": -0.18291568756103516 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot taken from a low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a knight riding a black steed at full speed visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the knight wearing a deep blue, coarse linen robe with sleeves and hem moving in the wind?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the scene include ancient Hui-style buildings with whitewashed walls and black tiles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the lighting warm and golden, resembling the afterglow of a setting sun?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.06208301708102226 + ], + "opposite_logprob_tokens": [ + -2.8120830059051514 + ], + "logprob_reward": -0.06208301708102226, + "yes_no_margin_reward": 2.75, + "token_credit_reward": -0.06208301708102226, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11232701689004898, + 0.09296733140945435, + 0.07591403275728226, + 0.08020798861980438, + 0.033631663769483566, + 0.17971274256706238, + 0.21578992903232574, + 0.12682010233402252, + 0.04055696353316307, + 0.04207220301032066 + ], + "credit_map": [ + 0.013098170049488544, + 0.01086517982184887, + 0.01517415139824152, + 0.008125022985041142, + 0.014397688210010529, + 0.03226896747946739, + 0.00796135701239109, + 0.010436475276947021, + 0.008728370070457458, + 0.013509141281247139, + 0.012825269252061844, + 0.009719284251332283, + 0.010336787439882755, + 0.01273877453058958, + 0.010633151978254318, + 0.01447654701769352, + 0.006035334896296263, + 0.007214393001049757, + 0.0068501755595207214, + 0.005325478967279196, + 0.00952831283211708, + 0.019431976601481438, + 0.004286990035325289, + 0.017241375520825386, + 0.0066197155974805355, + 0.00897261407226324, + 0.012968783266842365, + 0.013146674260497093, + 0.012625662609934807, + 0.007890874519944191, + 0.008958540856838226, + 0.009025120176374912, + 0.005370470229536295, + 0.003205973422154784, + 0.0027811566833406687, + 0.0030309087596833706, + 0.0031470649410039186, + 0.00844459980726242, + 0.0027937188278883696, + 0.0048577734269201756, + 0.02140379697084427, + 0.01633821614086628, + 0.02173757553100586, + 0.01844528317451477, + 0.01842079870402813, + 0.032181017100811005, + 0.027563737705349922, + 0.02362232469022274, + 0.023104732856154442, + 0.026906169950962067, + 0.013664409518241882, + 0.01816670596599579, + 0.049561213701963425, + 0.04081965610384941, + 0.018921196460723877, + 0.0246458500623703, + 0.0121684018522501, + 0.031604375690221786, + 0.009995619766414165, + 0.017143744975328445, + 0.016361741349101067, + 0.013403210788965225, + 0.013218225911259651, + 0.012924776412546635, + 0.005111087579280138, + 0.004321838729083538, + 0.0060248058289289474, + 0.0048671141266822815, + 0.0057394662871956825, + 0.004928851965814829, + 0.0049615236930549145, + 0.004602274391800165, + 0.004936919081956148, + 0.005777948070317507, + 0.0064433421939611435, + 0.004682521335780621, + 0.004307585768401623, + 0.005096597597002983, + 0.00500902347266674, + 0.005818265024572611 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 5, + 7 + ], + "selected_reward": -0.06208301708102226 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot taken from a low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a knight riding a black steed at full speed visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the knight wearing a deep blue, coarse linen robe with sleeves and hem moving in the wind?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the scene include ancient Hui-style buildings with whitewashed walls and black tiles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the lighting warm and golden, resembling the afterglow of a setting sun?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.043077558279037476 + ], + "opposite_logprob_tokens": [ + -3.1680774688720703 + ], + "logprob_reward": -0.043077558279037476, + "yes_no_margin_reward": 3.125, + "token_credit_reward": -0.043077558279037476, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12711942195892334, + 0.160856693983078, + 0.075058713555336, + 0.056555941700935364, + 0.12685999274253845, + 0.07558320462703705, + 0.09313260018825531, + 0.08460193872451782, + 0.026166215538978577, + 0.17406533658504486 + ], + "credit_map": [ + 0.01435482781380415, + 0.011278378777205944, + 0.02147827297449112, + 0.015532253310084343, + 0.013907541520893574, + 0.021633250638842583, + 0.015563074499368668, + 0.013371802866458893, + 0.017236078158020973, + 0.021486958488821983, + 0.029465392231941223, + 0.015978163108229637, + 0.014078161679208279, + 0.018089579418301582, + 0.020509537309408188, + 0.024012809619307518, + 0.006891519296914339, + 0.0064140199683606625, + 0.00570136122405529, + 0.00575436232611537, + 0.011225525289773941, + 0.016030075028538704, + 0.007381077390164137, + 0.015660768374800682, + 0.005807060282677412, + 0.0067864372394979, + 0.007458375301212072, + 0.008620359934866428, + 0.009517242200672626, + 0.005923385266214609, + 0.0056373560801148415, + 0.006805718410760164, + 0.010116459801793098, + 0.01901068352162838, + 0.019828403368592262, + 0.024117527529597282, + 0.016057154163718224, + 0.011764038354158401, + 0.013037613593041897, + 0.012928097508847713, + 0.008482910692691803, + 0.01187587808817625, + 0.007080416195094585, + 0.008831731975078583, + 0.008059542626142502, + 0.011776836588978767, + 0.01080639660358429, + 0.008669479750096798, + 0.01058549527078867, + 0.011762509122490883, + 0.006146821193397045, + 0.008531409315764904, + 0.021683817729353905, + 0.01570228300988674, + 0.007986055687069893, + 0.010734200477600098, + 0.007394953165203333, + 0.02230888232588768, + 0.0063682119362056255, + 0.013280541636049747, + 0.009308538399636745, + 0.009046155028045177, + 0.009077545255422592, + 0.007817103527486324, + 0.0030721419025212526, + 0.0038730818778276443, + 0.0023710019886493683, + 0.0030107006896287203, + 0.004370186943560839, + 0.0032286252826452255, + 0.002589602954685688, + 0.0036508713383227587, + 0.01780093088746071, + 0.014510809443891048, + 0.016184210777282715, + 0.02063903771340847, + 0.015164411626756191, + 0.043856892734766006, + 0.019265471026301384, + 0.02664356119930744 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 1, + 0 + ], + "selected_reward": -0.043077558279037476 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot taken from a low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a knight riding a black steed at full speed visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the knight wearing a deep blue, coarse linen robe with sleeves and hem moving in the wind?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the scene include ancient Hui-style buildings with whitewashed walls and black tiles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the lighting warm and golden, resembling the afterglow of a setting sun?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.17990459501743317 + ], + "opposite_logprob_tokens": [ + -1.804904580116272 + ], + "logprob_reward": -0.17990459501743317, + "yes_no_margin_reward": 1.625, + "token_credit_reward": -0.17990459501743317, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1171916276216507, + 0.13755705952644348, + 0.36101245880126953, + 0.04617036134004593, + 0.07443886250257492, + 0.08104250580072403, + 0.06772367656230927, + 0.0316273458302021, + 0.03406259045004845, + 0.049173541367053986 + ], + "credit_map": [ + 0.015879064798355103, + 0.01747211441397667, + 0.014761150814592838, + 0.012900205329060555, + 0.012387096881866455, + 0.025833919644355774, + 0.009437032975256443, + 0.008521041832864285, + 0.018115006387233734, + 0.016177337616682053, + 0.026514574885368347, + 0.013165067881345749, + 0.0123191112652421, + 0.017499366775155067, + 0.017145944759249687, + 0.01662064716219902, + 0.033424850553274155, + 0.0491458997130394, + 0.04346701130270958, + 0.03774872049689293, + 0.07596830278635025, + 0.03739378973841667, + 0.041297946125268936, + 0.04256594553589821, + 0.004698556382209063, + 0.004633090924471617, + 0.005158517975360155, + 0.00689610093832016, + 0.006400283891707659, + 0.005784743465483189, + 0.006973551120609045, + 0.005625514779239893, + 0.006083347834646702, + 0.0067735835909843445, + 0.006960602942854166, + 0.008741907775402069, + 0.011653597466647625, + 0.016760636121034622, + 0.007069054991006851, + 0.01039612665772438, + 0.01002691313624382, + 0.010933304205536842, + 0.009187353774905205, + 0.010114810429513454, + 0.008167991414666176, + 0.012077885679900646, + 0.009921059012413025, + 0.010613185353577137, + 0.0065885381773114204, + 0.009565030224621296, + 0.004370295442640781, + 0.006830129306763411, + 0.012083645910024643, + 0.014352665282785892, + 0.006424519699066877, + 0.007508853450417519, + 0.0031854461412876844, + 0.0059521435759961605, + 0.0029449057765305042, + 0.004209368489682674, + 0.003559330478310585, + 0.0044197579845786095, + 0.004379794467240572, + 0.0029765984509140253, + 0.00470650615170598, + 0.005318061914294958, + 0.0030119498260319233, + 0.003282815683633089, + 0.006969959940761328, + 0.0032073769252747297, + 0.0035277525894343853, + 0.004038169048726559, + 0.004714002832770348, + 0.005401309113949537, + 0.006674108561128378, + 0.005211375653743744, + 0.006412711460143328, + 0.008808140642940998, + 0.0055401320569217205, + 0.006411761045455933 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 1, + 0 + ], + "selected_reward": -0.17990459501743317 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot taken from a low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a knight riding a black steed at full speed visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the knight wearing a deep blue, coarse linen robe with sleeves and hem moving in the wind?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the scene include ancient Hui-style buildings with whitewashed walls and black tiles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the lighting warm and golden, resembling the afterglow of a setting sun?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.011075095273554325 + ], + "opposite_logprob_tokens": [ + -4.511075019836426 + ], + "logprob_reward": -0.011075095273554325, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.011075095273554325, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.49988651275634766, + 0.03280827775597572, + 0.05526302382349968, + 0.029721565544605255, + 0.06261374056339264, + 0.04147466644644737, + 0.05334709584712982, + 0.02889878861606121, + 0.13342371582984924, + 0.06256259232759476 + ], + "credit_map": [ + 0.06595094501972198, + 0.05273592472076416, + 0.05887109041213989, + 0.029580235481262207, + 0.060361843556165695, + 0.16284316778182983, + 0.028531905263662338, + 0.04101143404841423, + 0.004203874617815018, + 0.0028396144043654203, + 0.005561053287237883, + 0.002513129962608218, + 0.003506795270368457, + 0.0068707894533872604, + 0.0032062213867902756, + 0.004106798209249973, + 0.005513098556548357, + 0.006748538929969072, + 0.004058872815221548, + 0.00530275097116828, + 0.015832945704460144, + 0.007065925281494856, + 0.005112654063850641, + 0.0056282393634319305, + 0.003047313541173935, + 0.003173965262249112, + 0.003820930141955614, + 0.003988207317888737, + 0.004365084692835808, + 0.004011454060673714, + 0.0035642418079078197, + 0.0037503717467188835, + 0.007352744694799185, + 0.008075677789747715, + 0.00725305313244462, + 0.008313316851854324, + 0.005760032217949629, + 0.013083617202937603, + 0.005876306910067797, + 0.006898991763591766, + 0.005177206359803677, + 0.004406946711242199, + 0.003991624806076288, + 0.004619122948497534, + 0.0037436392158269882, + 0.005453708115965128, + 0.007784585934132338, + 0.006297830026596785, + 0.005567149259150028, + 0.007362420670688152, + 0.003233872586861253, + 0.004240310750901699, + 0.012284552678465843, + 0.010535168461501598, + 0.004425459075719118, + 0.005698160734027624, + 0.0028556333854794502, + 0.005819908808916807, + 0.002437449060380459, + 0.004333788529038429, + 0.004417976830154657, + 0.0029317857697606087, + 0.003629871876910329, + 0.0024723734240978956, + 0.01553105004131794, + 0.018909893929958344, + 0.01238469872623682, + 0.010624800808727741, + 0.03301436826586723, + 0.015990933403372765, + 0.009510330855846405, + 0.017457641661167145, + 0.005650418344885111, + 0.006119036581367254, + 0.007593725807964802, + 0.00660362932831049, + 0.0072288187220692635, + 0.014206325635313988, + 0.00708336615934968, + 0.008077269420027733 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 4 + ], + "selected_reward": -0.011075095273554325 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "low-angle, long, dynamic tracking shot follows a knight as he rides a majestic black steed at full speed along an ancient bluestone path. The knight is clad in a deep blue, coarse linen robe, its sleeves and hem dancing wildly in the wind. His hair is tied up in a high bun, strands of it rustling against his resolute face. He leans forward slightly, one hand firmly gripping the reins, the scabbard of the longsword at his waist swaying rhythmically with the horse's jerk. Beneath his steed, its muscles bulging, its glossy fur shimmering in the light. Its hooves fly, their collisions with the bluestone pavement sending up dust and scattered sparks. They speed past a cluster of ancient, pristine Hui-style buildings. The whitewashed walls and black tiles appear remarkably tranquil in the warm twilight. The staggered, undulating horse-head walls create a rhythmic skyline, while details of carved wooden window lattices flash by in the rapid movement. The afterglow of the setting sun dyes the entire picture with a warm golden color, lengthening the shadows of the buildings and the riders, creating a desolate and classical martial arts style atmosphere full of heroic spirit, with a strong contrast between movement and stillness.", + "seed": 3828, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000027/group_6/index_544_seed_3828_dp_13_rank_26_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 6, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.11514578014612198, + "vlm_logprob": -0.11514578014612198, + "vlm_yes_no_margin": 2.4, + "vlm_token_credit": -0.11514578014612198, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot taken from a low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a knight riding a black steed at full speed visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the knight wearing a deep blue, coarse linen robe with sleeves and hem moving in the wind?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the scene include ancient Hui-style buildings with whitewashed walls and black tiles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the lighting warm and golden, resembling the afterglow of a setting sun?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.12865683436393738 + ], + "opposite_logprob_tokens": [ + -2.3786568641662598 + ], + "logprob_reward": -0.12865683436393738, + "yes_no_margin_reward": 2.25, + "token_credit_reward": -0.12865683436393738, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.6668428778648376, + 0.07639432698488235, + 0.0331755056977272, + 0.033129774034023285, + 0.03259972855448723, + 0.012244761921465397, + 0.039186544716358185, + 0.04926058650016785, + 0.03519470617175102, + 0.021971289068460464 + ], + "credit_map": [ + 0.036026522517204285, + 0.20565341413021088, + 0.03777606040239334, + 0.025516502559185028, + 0.04509327560663223, + 0.23370058834552765, + 0.03447554260492325, + 0.04860091954469681, + 0.0078654196113348, + 0.008333143778145313, + 0.006764660123735666, + 0.009330774657428265, + 0.014968614093959332, + 0.008642636239528656, + 0.010628489777445793, + 0.00986057985574007, + 0.003482454689219594, + 0.007634936366230249, + 0.0038808651734143496, + 0.002905843313783407, + 0.006088658701628447, + 0.004041358828544617, + 0.002523085568100214, + 0.0026183000300079584, + 0.006987281609326601, + 0.004888846538960934, + 0.007514934055507183, + 0.003027567407116294, + 0.0029699779115617275, + 0.00200368813239038, + 0.002994748530909419, + 0.0027427261229604483, + 0.005238867364823818, + 0.005710084456950426, + 0.003264036728069186, + 0.0033940798602998257, + 0.004017153289169073, + 0.004341556690633297, + 0.0030464634764939547, + 0.0035874831955879927, + 0.0020111247431486845, + 0.0018183187348768115, + 0.0015832025092095137, + 0.0021415085066109896, + 0.0014201892772689462, + 0.0007714605308137834, + 0.0011983870062977076, + 0.0013005695072934031, + 0.004113859962671995, + 0.005503541324287653, + 0.004862109199166298, + 0.006703801453113556, + 0.0048688435927033424, + 0.004258377011865377, + 0.003787492634728551, + 0.005088517442345619, + 0.008803598582744598, + 0.010482870042324066, + 0.006953484378755093, + 0.007317488081753254, + 0.005335541907697916, + 0.003356555476784706, + 0.0037027420476078987, + 0.0033083062153309584, + 0.00892704352736473, + 0.00326183857396245, + 0.0030114513356238604, + 0.0022514769807457924, + 0.011038681492209435, + 0.0030045059975236654, + 0.001900381757877767, + 0.0017993279034271836, + 0.0028011223766952753, + 0.004336596932262182, + 0.0031522049102932215, + 0.0027964268811047077, + 0.0020474130287766457, + 0.002708780113607645, + 0.001961388625204563, + 0.002167356899008155 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 7 + ], + "selected_reward": -0.12865683436393738 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot taken from a low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a knight riding a black steed at full speed visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the knight wearing a deep blue, coarse linen robe with sleeves and hem moving in the wind?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the scene include ancient Hui-style buildings with whitewashed walls and black tiles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the lighting warm and golden, resembling the afterglow of a setting sun?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.048686571419239044 + ], + "opposite_logprob_tokens": [ + -3.0486865043640137 + ], + "logprob_reward": -0.048686571419239044, + "yes_no_margin_reward": 3.0, + "token_credit_reward": -0.048686571419239044, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11428596824407578, + 0.09416268765926361, + 0.0732518807053566, + 0.059103697538375854, + 0.07446017861366272, + 0.06882241368293762, + 0.16457532346248627, + 0.0988846868276596, + 0.15307094156742096, + 0.09938216209411621 + ], + "credit_map": [ + 0.010283438488841057, + 0.013173392042517662, + 0.01630261540412903, + 0.012972584925591946, + 0.01100049540400505, + 0.027082959190011024, + 0.012323647737503052, + 0.011146833188831806, + 0.011991193518042564, + 0.012592801824212074, + 0.011866610497236252, + 0.01171288825571537, + 0.01349095068871975, + 0.01235076691955328, + 0.00873146764934063, + 0.011426009237766266, + 0.00768136465921998, + 0.015806103125214577, + 0.007294909097254276, + 0.008944855071604252, + 0.011842752806842327, + 0.007458833046257496, + 0.008113322779536247, + 0.006109744776040316, + 0.00794033519923687, + 0.006829762365669012, + 0.007448114920407534, + 0.0078947264701128, + 0.007119827438145876, + 0.005176760721951723, + 0.00672123022377491, + 0.009972946718335152, + 0.011440242640674114, + 0.008799314498901367, + 0.007519624661654234, + 0.011635707691311836, + 0.01134196575731039, + 0.009600230492651463, + 0.006808813661336899, + 0.007314280606806278, + 0.00626768684014678, + 0.010082151740789413, + 0.009960517287254333, + 0.014681358821690083, + 0.010629411786794662, + 0.004753486718982458, + 0.006593453697860241, + 0.005854346789419651, + 0.031864069402217865, + 0.027592485770583153, + 0.023793956264853477, + 0.014996955171227455, + 0.014359686523675919, + 0.021464820951223373, + 0.01504945382475853, + 0.015453891828656197, + 0.013475901447236538, + 0.010154777206480503, + 0.013304146006703377, + 0.016397330909967422, + 0.011802527122199535, + 0.009908896870911121, + 0.011972554959356785, + 0.011868546716868877, + 0.0284974854439497, + 0.03023478016257286, + 0.019245905801653862, + 0.016218548640608788, + 0.033261027187108994, + 0.010414056479930878, + 0.007961304858326912, + 0.00723783066496253, + 0.013633426278829575, + 0.011006462387740612, + 0.014097902923822403, + 0.019270628690719604, + 0.012994830496609211, + 0.008899555541574955, + 0.009776787832379341, + 0.009702567011117935 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 8, + 0 + ], + "selected_reward": -0.048686571419239044 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot taken from a low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a knight riding a black steed at full speed visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the knight wearing a deep blue, coarse linen robe with sleeves and hem moving in the wind?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the scene include ancient Hui-style buildings with whitewashed walls and black tiles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the lighting warm and golden, resembling the afterglow of a setting sun?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.054970093071460724 + ], + "opposite_logprob_tokens": [ + -2.9299700260162354 + ], + "logprob_reward": -0.054970093071460724, + "yes_no_margin_reward": 2.875, + "token_credit_reward": -0.054970093071460724, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.18134117126464844, + 0.16867677867412567, + 0.09392368048429489, + 0.08614630997180939, + 0.07369177788496017, + 0.042190685868263245, + 0.04885927215218544, + 0.04307714104652405, + 0.21656931936740875, + 0.04552391171455383 + ], + "credit_map": [ + 0.013996012508869171, + 0.021457331255078316, + 0.02902643010020256, + 0.01362043246626854, + 0.012741172686219215, + 0.0600447878241539, + 0.015413491986691952, + 0.015041479840874672, + 0.0199109073728323, + 0.01447226107120514, + 0.018184075132012367, + 0.020389661192893982, + 0.03576777130365372, + 0.021575095131993294, + 0.02042040042579174, + 0.01795658841729164, + 0.009944108314812183, + 0.01536736823618412, + 0.011740329675376415, + 0.00699334591627121, + 0.02158018760383129, + 0.010797168128192425, + 0.008824474178254604, + 0.008676690980792046, + 0.02079557068645954, + 0.007217570208013058, + 0.02410479262471199, + 0.007387375459074974, + 0.0058331978507339954, + 0.0043489583767950535, + 0.007719980552792549, + 0.008738855831325054, + 0.010425941087305546, + 0.011623910628259182, + 0.007195950951427221, + 0.009124385192990303, + 0.00944511964917183, + 0.010633581317961216, + 0.0062008751556277275, + 0.009042002260684967, + 0.004004385322332382, + 0.004169096238911152, + 0.006446816958487034, + 0.011029687710106373, + 0.004403372295200825, + 0.0026459575165063143, + 0.004486649762839079, + 0.005004714708775282, + 0.006551344413310289, + 0.008619319647550583, + 0.006370723247528076, + 0.008122156374156475, + 0.005282514728605747, + 0.004667980130761862, + 0.004828151315450668, + 0.004417075775563717, + 0.006980610080063343, + 0.008540312759578228, + 0.0054375953041017056, + 0.00566779263317585, + 0.0054548648186028, + 0.004249625373631716, + 0.0033832082990556955, + 0.0033631299156695604, + 0.035523753613233566, + 0.03573410212993622, + 0.023430706933140755, + 0.02054721862077713, + 0.04919404909014702, + 0.022507142275571823, + 0.01162488479167223, + 0.018007434904575348, + 0.0063006035052239895, + 0.006646185182034969, + 0.005262788850814104, + 0.007098757661879063, + 0.00564016867429018, + 0.00484805041924119, + 0.0052690692245960236, + 0.004458284936845303 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 0, + 1 + ], + "selected_reward": -0.054970093071460724 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot taken from a low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a knight riding a black steed at full speed visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the knight wearing a deep blue, coarse linen robe with sleeves and hem moving in the wind?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the scene include ancient Hui-style buildings with whitewashed walls and black tiles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the lighting warm and golden, resembling the afterglow of a setting sun?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.28135401010513306 + ], + "opposite_logprob_tokens": [ + -1.4063539505004883 + ], + "logprob_reward": -0.28135401010513306, + "yes_no_margin_reward": 1.125, + "token_credit_reward": -0.28135401010513306, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3486199378967285, + 0.05716097354888916, + 0.15646786987781525, + 0.05353064090013504, + 0.1268072873353958, + 0.038637496531009674, + 0.026415850967168808, + 0.05311260372400284, + 0.10817307978868484, + 0.031074276193976402 + ], + "credit_map": [ + 0.018858017399907112, + 0.1032392829656601, + 0.01973550207912922, + 0.01376370806246996, + 0.023519931361079216, + 0.1274884194135666, + 0.019168538972735405, + 0.02284650132060051, + 0.008894851431250572, + 0.00929646659642458, + 0.008611143566668034, + 0.005781985353678465, + 0.008735664188861847, + 0.005421212874352932, + 0.005331144668161869, + 0.005088502541184425, + 0.018609028309583664, + 0.04596654698252678, + 0.0150063531473279, + 0.01833842508494854, + 0.01894306018948555, + 0.01747956871986389, + 0.01214656326919794, + 0.009978320449590683, + 0.012750426307320595, + 0.005389018915593624, + 0.014230477623641491, + 0.004058152437210083, + 0.004991469904780388, + 0.0029121169354766607, + 0.004612268880009651, + 0.004586704075336456, + 0.03905002027750015, + 0.022632306441664696, + 0.01166608091443777, + 0.012911277823150158, + 0.009635429829359055, + 0.013158058747649193, + 0.007448600139468908, + 0.010305513627827168, + 0.008277440443634987, + 0.005924507509917021, + 0.004948443733155727, + 0.004121688194572926, + 0.003951625432819128, + 0.0047594476491212845, + 0.002646804554387927, + 0.004007537383586168, + 0.0043907854706048965, + 0.004142033401876688, + 0.003253600560128689, + 0.0028375056572258472, + 0.003643016330897808, + 0.0025970530696213245, + 0.0025417518336325884, + 0.003010102780535817, + 0.009565710090100765, + 0.00754591915756464, + 0.007493623066693544, + 0.009362519718706608, + 0.005959316156804562, + 0.00458105094730854, + 0.004294620361179113, + 0.004309837706387043, + 0.020708050578832626, + 0.015232316218316555, + 0.010577580891549587, + 0.009207207709550858, + 0.02856149524450302, + 0.010062378831207752, + 0.005728862248361111, + 0.008095181547105312, + 0.004316046833992004, + 0.00418064696714282, + 0.004513754043728113, + 0.005039708688855171, + 0.003965297248214483, + 0.003738382598385215, + 0.002406357554718852, + 0.0029140806291252375 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 4 + ], + "selected_reward": -0.28135401010513306 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot taken from a low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a knight riding a black steed at full speed visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the knight wearing a deep blue, coarse linen robe with sleeves and hem moving in the wind?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the scene include ancient Hui-style buildings with whitewashed walls and black tiles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the lighting warm and golden, resembling the afterglow of a setting sun?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.06206139177083969 + ], + "opposite_logprob_tokens": [ + -2.812061309814453 + ], + "logprob_reward": -0.06206139177083969, + "yes_no_margin_reward": 2.75, + "token_credit_reward": -0.06206139177083969, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.810783326625824, + 0.008538271300494671, + 0.009309723973274231, + 0.007019814103841782, + 0.03914837911725044, + 0.0074163624085485935, + 0.0057994285598397255, + 0.01048176921904087, + 0.016068723052740097, + 0.08543424308300018 + ], + "credit_map": [ + 0.04378364980220795, + 0.2451828271150589, + 0.04915608465671539, + 0.03442969545722008, + 0.057284701615571976, + 0.2748652696609497, + 0.04447350651025772, + 0.061607539653778076, + 0.0015293395845219493, + 0.0013809716328978539, + 0.0009539893362671137, + 0.0009107011719606817, + 0.0009914995171129704, + 0.0012105420464649796, + 0.0006154919392429292, + 0.0009457358974032104, + 0.0011887558503076434, + 0.0022511680144816637, + 0.0009844667511060834, + 0.0011211511446163058, + 0.0012369330506771803, + 0.0009117737645283341, + 0.0007549561560153961, + 0.0008605183102190495, + 0.0009483773610554636, + 0.0012922558235004544, + 0.0007592975744046271, + 0.0007019160548225045, + 0.001324390177614987, + 0.0006393910152837634, + 0.0005949281621724367, + 0.0007592574111185968, + 0.0071077351458370686, + 0.010801342315971851, + 0.004220283590257168, + 0.004346839152276516, + 0.0033099122811108828, + 0.004753318149596453, + 0.00212528626434505, + 0.0024836622178554535, + 0.001206963905133307, + 0.001154065248556435, + 0.0011277340818196535, + 0.0007526340195909142, + 0.0011816088808700442, + 0.0008056443766690791, + 0.0005670975078828633, + 0.0006206143298186362, + 0.0009049297077581286, + 0.0010748537024483085, + 0.0008636568090878427, + 0.0006136473384685814, + 0.0007792669930495322, + 0.0007116628112271428, + 0.0004485514364205301, + 0.00040285903378389776, + 0.0020804679952561855, + 0.0012400414561852813, + 0.0015468134079128504, + 0.0020335717126727104, + 0.0010042648063972592, + 0.0007402494084089994, + 0.0009167308453470469, + 0.0009196287719532847, + 0.003740447573363781, + 0.0020298755262047052, + 0.002509781625121832, + 0.0012500520097091794, + 0.003591760527342558, + 0.001218197401612997, + 0.0008568413322791457, + 0.0008717653690837324, + 0.014262054115533829, + 0.019287891685962677, + 0.012710670940577984, + 0.010203059762716293, + 0.0063488599844276905, + 0.009883608669042587, + 0.006515031214803457, + 0.006223059259355068 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 4 + ], + "selected_reward": -0.06206139177083969 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000027/group_8/index_tensor([792, 792])_group_8_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000027/group_8/index_tensor([792, 792])_group_8_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..fae0f22b6df6e055ede0717af7d0f5a0c2a90592 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000027/group_8/index_tensor([792, 792])_group_8_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "Shot from a low angle on a precipitous cliff, a sweeping, epic wide-angle shot pans slowly to the right, revealing a majestic Chinese-style mansion perched atop the mountain like an ancient behemoth. Constructed of dark, weathered wood and a massive stone foundation, the main structure is clad in dark green glazed tiles, gleaming in the sparse morning light. Its signature eaves and ridges rise gracefully and powerfully toward the sky like dragon wings. Thick, milky white mist flows slowly around the mansion like a living river, gently entwining carved columns and weaving between the intricate eaves. It sometimes engulfs the building, leaving only the outline of the roof visible, and sometimes quietly recedes to reveal its majestic fullness. The entire scene is shrouded in a cold and serene atmosphere. The faint light of dawn struggles to penetrate the clouds, gilding the edges of the flowing mist with a faint silver hue, creating a cinematic sense of isolation, mystery, and eternal power.", + "seed": 22651, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000027/group_8/index_792_seed_22651_dp_16_rank_32_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 8, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.18659647708991542, + "vlm_logprob": -0.18659647708991542, + "vlm_yes_no_margin": 3.975, + "vlm_token_credit": -0.18659647708991542, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot taken from a low angle on a precipitous cliff?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera pan slowly to the right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a Chinese-style mansion with dark green glazed tiles visible on the mountain?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is thick, milky white mist flowing around and partially engulfing the mansion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by sparse morning light with a cold and serene atmosphere?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.6387919783592224 + ], + "opposite_logprob_tokens": [ + -0.7637919783592224 + ], + "logprob_reward": -0.6387919783592224, + "yes_no_margin_reward": 0.125, + "token_credit_reward": -0.6387919783592224, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08752992749214172, + 0.09374348074197769, + 0.10857988893985748, + 0.04530509561300278, + 0.12048249691724777, + 0.11900452524423599, + 0.13423147797584534, + 0.08612281829118729, + 0.11602757126092911, + 0.0889727771282196 + ], + "credit_map": [ + 0.005012766923755407, + 0.008597688749432564, + 0.016764093190431595, + 0.013436305336654186, + 0.006703346036374569, + 0.007073982153087854, + 0.014591875486075878, + 0.015349877066910267, + 0.0055063944309949875, + 0.009994927793741226, + 0.010659221559762955, + 0.018018370494246483, + 0.007484697736799717, + 0.0051738363690674305, + 0.01518266275525093, + 0.021723370999097824, + 0.008267119526863098, + 0.020430734381079674, + 0.011850274167954922, + 0.017511863261461258, + 0.014214390888810158, + 0.006986074615269899, + 0.01288459450006485, + 0.016434842720627785, + 0.004331182688474655, + 0.0046928985975682735, + 0.009192368946969509, + 0.008942738175392151, + 0.004062071908265352, + 0.0023256412241607904, + 0.005307643674314022, + 0.006450552027672529, + 0.00898714642971754, + 0.02348138391971588, + 0.013459708541631699, + 0.01973232813179493, + 0.014119178988039494, + 0.005653166677802801, + 0.014480913057923317, + 0.020568668842315674, + 0.009412216022610664, + 0.02342000976204872, + 0.01029478944838047, + 0.01814865507185459, + 0.01348322443664074, + 0.005177002865821123, + 0.02839076891541481, + 0.010677860118448734, + 0.011921732686460018, + 0.028702588751912117, + 0.014983870089054108, + 0.018229857087135315, + 0.01948763243854046, + 0.007459528744220734, + 0.016284141689538956, + 0.017162127420306206, + 0.006621342618018389, + 0.008255775086581707, + 0.012251135893166065, + 0.011821828782558441, + 0.011491070501506329, + 0.004572054836899042, + 0.015235170722007751, + 0.01587444357573986, + 0.011439239606261253, + 0.018745092675089836, + 0.015108485706150532, + 0.017830749973654747, + 0.012211175635457039, + 0.006310780066996813, + 0.014918454922735691, + 0.019463591277599335, + 0.006565597839653492, + 0.0069197374396026134, + 0.010069306008517742, + 0.020704424008727074, + 0.009128989651799202, + 0.005427826661616564, + 0.014352449215948582, + 0.01580444723367691 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 4, + 5 + ], + "selected_reward": -0.6387919783592224 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot taken from a low angle on a precipitous cliff?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera pan slowly to the right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a Chinese-style mansion with dark green glazed tiles visible on the mountain?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is thick, milky white mist flowing around and partially engulfing the mansion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by sparse morning light with a cold and serene atmosphere?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.2813633680343628 + ], + "opposite_logprob_tokens": [ + -1.4063633680343628 + ], + "logprob_reward": -0.2813633680343628, + "yes_no_margin_reward": 1.125, + "token_credit_reward": -0.2813633680343628, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.16984154284000397, + 0.0855516791343689, + 0.12552043795585632, + 0.04909003898501396, + 0.05720072239637375, + 0.10465772449970245, + 0.07319319248199463, + 0.10387841612100601, + 0.20237769186496735, + 0.028688481077551842 + ], + "credit_map": [ + 0.009273156523704529, + 0.015042177401483059, + 0.049176450818777084, + 0.033254362642765045, + 0.010904323309659958, + 0.008961796760559082, + 0.019910188391804695, + 0.023319104686379433, + 0.00732222618535161, + 0.008224500343203545, + 0.009614848531782627, + 0.020214656367897987, + 0.0070572406984865665, + 0.00699519133195281, + 0.018434619531035423, + 0.0076884008012712, + 0.007671944331377745, + 0.02450377307832241, + 0.01638883911073208, + 0.021881256252527237, + 0.016679391264915466, + 0.007863344624638557, + 0.01609029807150364, + 0.0144415982067585, + 0.005671359598636627, + 0.0049980259500443935, + 0.007282230071723461, + 0.008899635635316372, + 0.004697979893535376, + 0.0028476957231760025, + 0.006494827102869749, + 0.008198292925953865, + 0.006408029235899448, + 0.006184266414493322, + 0.006716735661029816, + 0.009435255080461502, + 0.005503728520125151, + 0.0030623336788266897, + 0.008149814791977406, + 0.011740565299987793, + 0.009877157397568226, + 0.014669116586446762, + 0.011777677573263645, + 0.019470250234007835, + 0.009428306482732296, + 0.003734693629667163, + 0.024776281788945198, + 0.01092425175011158, + 0.009543996304273605, + 0.0073900786228477955, + 0.012731019407510757, + 0.01245880126953125, + 0.005171207245439291, + 0.0042072501964867115, + 0.011117418296635151, + 0.010573433712124825, + 0.009169310331344604, + 0.015408881939947605, + 0.011377237737178802, + 0.013225776143372059, + 0.01666356809437275, + 0.006757988594472408, + 0.014453093521296978, + 0.0168225709348917, + 0.02118086814880371, + 0.03755439817905426, + 0.02047235704958439, + 0.0237689558416605, + 0.0317944660782814, + 0.01423362921923399, + 0.023097975179553032, + 0.03027506358921528, + 0.0020426763221621513, + 0.0021653042640537024, + 0.0029985795263201, + 0.007558694574981928, + 0.0024384567514061928, + 0.001822434482164681, + 0.0046955905854702, + 0.004966746550053358 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 0, + 2 + ], + "selected_reward": -0.2813633680343628 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot taken from a low angle on a precipitous cliff?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera pan slowly to the right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a Chinese-style mansion with dark green glazed tiles visible on the mountain?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is thick, milky white mist flowing around and partially engulfing the mansion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by sparse morning light with a cold and serene atmosphere?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0008272086852230132 + ], + "opposite_logprob_tokens": [ + -7.125827312469482 + ], + "logprob_reward": -0.0008272086852230132, + "yes_no_margin_reward": 7.125, + "token_credit_reward": -0.0008272086852230132, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2109530121088028, + 0.13301999866962433, + 0.09417486190795898, + 0.047208305448293686, + 0.1341562420129776, + 0.036355357617139816, + 0.15284256637096405, + 0.043301258236169815, + 0.10480378568172455, + 0.04318460077047348 + ], + "credit_map": [ + 0.013113829307258129, + 0.01814655028283596, + 0.06141524389386177, + 0.042164068669080734, + 0.014402533881366253, + 0.009759487584233284, + 0.025307968258857727, + 0.02664332464337349, + 0.011875465512275696, + 0.015530419535934925, + 0.017570678144693375, + 0.024134160950779915, + 0.012414402328431606, + 0.010870477184653282, + 0.025447582826018333, + 0.01517681498080492, + 0.005988385062664747, + 0.013512066565454006, + 0.010695923119783401, + 0.014481686986982822, + 0.009951144456863403, + 0.0069993543438613415, + 0.012376648373901844, + 0.02016964554786682, + 0.002528950572013855, + 0.004065841436386108, + 0.010603947564959526, + 0.008516313508152962, + 0.0034258810337632895, + 0.002098517492413521, + 0.0057986462488770485, + 0.010170206427574158, + 0.009357934817671776, + 0.026459289714694023, + 0.013082304038107395, + 0.024513782933354378, + 0.01692649908363819, + 0.00608576787635684, + 0.01578286662697792, + 0.021947801113128662, + 0.0022327324841171503, + 0.0035927779972553253, + 0.005835670977830887, + 0.009391985833644867, + 0.003314657835289836, + 0.0017131877830252051, + 0.006645323242992163, + 0.0036290199495851994, + 0.011435223743319511, + 0.027114439755678177, + 0.019119974225759506, + 0.0213799849152565, + 0.021901810541749, + 0.008255013264715672, + 0.02026893012225628, + 0.023367170244455338, + 0.0041506653651595116, + 0.006632352247834206, + 0.007511511445045471, + 0.005705770570784807, + 0.005467393435537815, + 0.0026174448430538177, + 0.006195684429258108, + 0.005020435433834791, + 0.010454462841153145, + 0.016499876976013184, + 0.01349158026278019, + 0.013532297685742378, + 0.016390666365623474, + 0.007015632931143045, + 0.014141352847218513, + 0.013277915306389332, + 0.0030546272173523903, + 0.004343581385910511, + 0.004582724068313837, + 0.009348401799798012, + 0.004792628809809685, + 0.0019050094997510314, + 0.007645339705049992, + 0.007512287236750126 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 6, + 4 + ], + "selected_reward": -0.0008272086852230132 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot taken from a low angle on a precipitous cliff?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera pan slowly to the right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a Chinese-style mansion with dark green glazed tiles visible on the mountain?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is thick, milky white mist flowing around and partially engulfing the mansion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by sparse morning light with a cold and serene atmosphere?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0009227308910340071 + ], + "opposite_logprob_tokens": [ + -7.000922679901123 + ], + "logprob_reward": -0.0009227308910340071, + "yes_no_margin_reward": 7.0, + "token_credit_reward": -0.0009227308910340071, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.21079908311367035, + 0.14377975463867188, + 0.07763306051492691, + 0.06109229102730751, + 0.14684100449085236, + 0.05419767275452614, + 0.16674944758415222, + 0.03272228315472603, + 0.06779295206069946, + 0.03839246183633804 + ], + "credit_map": [ + 0.012602303177118301, + 0.019193653017282486, + 0.05872570350766182, + 0.04184284806251526, + 0.016500988975167274, + 0.009882464073598385, + 0.024222934618592262, + 0.027828168123960495, + 0.008373514749109745, + 0.016813136637210846, + 0.02061244286596775, + 0.0376712940633297, + 0.010367353446781635, + 0.006135177798569202, + 0.02796774171292782, + 0.01583908684551716, + 0.004342240747064352, + 0.005151074379682541, + 0.011256294324994087, + 0.012994533404707909, + 0.004112649243324995, + 0.005001076031476259, + 0.011590023525059223, + 0.023185167461633682, + 0.003468463895842433, + 0.004367051646113396, + 0.009785565547645092, + 0.010152400471270084, + 0.003783191554248333, + 0.002539008855819702, + 0.009334443137049675, + 0.017662163823843002, + 0.00976755190640688, + 0.018195556476712227, + 0.017104679718613625, + 0.02587972767651081, + 0.013070682995021343, + 0.00671591330319643, + 0.020172176882624626, + 0.035934701561927795, + 0.004525181371718645, + 0.005700927227735519, + 0.007847990840673447, + 0.01118416991084814, + 0.005742775276303291, + 0.0025523588992655277, + 0.009776360355317593, + 0.0068679070100188255, + 0.0129188047721982, + 0.03104810230433941, + 0.021424707025289536, + 0.022514155134558678, + 0.022973498329520226, + 0.009105853736400604, + 0.02153126522898674, + 0.025233052670955658, + 0.0029346372466534376, + 0.004428389016538858, + 0.00496262963861227, + 0.004651601426303387, + 0.0036914206575602293, + 0.0019561974331736565, + 0.005391635466367006, + 0.004705770872533321, + 0.006712900009006262, + 0.009682875126600266, + 0.008439267985522747, + 0.009922962635755539, + 0.009069454856216908, + 0.0035909097641706467, + 0.009342153556644917, + 0.01103241741657257, + 0.0032051866874098778, + 0.002937526907771826, + 0.0034275921061635017, + 0.010159187018871307, + 0.0031804589089006186, + 0.0021242816001176834, + 0.007019078824669123, + 0.0063391453586518764 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 6, + 4 + ], + "selected_reward": -0.0009227308910340071 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot taken from a low angle on a precipitous cliff?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera pan slowly to the right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a Chinese-style mansion with dark green glazed tiles visible on the mountain?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is thick, milky white mist flowing around and partially engulfing the mansion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by sparse morning light with a cold and serene atmosphere?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.011077099479734898 + ], + "opposite_logprob_tokens": [ + -4.511076927185059 + ], + "logprob_reward": -0.011077099479734898, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.011077099479734898, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1789543479681015, + 0.15360848605632782, + 0.13034741580486298, + 0.03354950249195099, + 0.11920416355133057, + 0.058074768632650375, + 0.12932626903057098, + 0.08710142225027084, + 0.06113019958138466, + 0.04870343953371048 + ], + "credit_map": [ + 0.014244657009840012, + 0.017685048282146454, + 0.027836410328745842, + 0.024644259363412857, + 0.012544682249426842, + 0.012607340700924397, + 0.031977392733097076, + 0.0374145582318306, + 0.012711336836218834, + 0.014205445535480976, + 0.018537059426307678, + 0.033384475857019424, + 0.011814425699412823, + 0.009478585794568062, + 0.02923923172056675, + 0.02423793077468872, + 0.007905845530331135, + 0.010432750917971134, + 0.021726960316300392, + 0.02236863784492016, + 0.010842886753380299, + 0.007045355625450611, + 0.01897045224905014, + 0.031054532155394554, + 0.0028181348461657763, + 0.004422331694513559, + 0.0054047247394919395, + 0.0053540440276265144, + 0.004180257208645344, + 0.0015546097420156002, + 0.0043586911633610725, + 0.005456708837300539, + 0.009446765296161175, + 0.023492705076932907, + 0.01570749841630459, + 0.01945471577346325, + 0.015257603488862514, + 0.0056601823307573795, + 0.013263584114611149, + 0.016921106725931168, + 0.005413366481661797, + 0.005908251740038395, + 0.009597542695701122, + 0.01283499039709568, + 0.004035684745758772, + 0.0022424845956265926, + 0.012416424229741096, + 0.005626025144010782, + 0.009744668379426003, + 0.01826929859817028, + 0.019590388983488083, + 0.019904928281903267, + 0.015774039551615715, + 0.0065186177380383015, + 0.017620697617530823, + 0.021903643384575844, + 0.008832826279103756, + 0.012469914741814137, + 0.013827021233737469, + 0.014049554243683815, + 0.008960950188338757, + 0.005088095087558031, + 0.012143397703766823, + 0.011729663237929344, + 0.004510627128183842, + 0.006890247575938702, + 0.007133289705961943, + 0.014300813898444176, + 0.007506613153964281, + 0.0040525468066334724, + 0.009882662445306778, + 0.006853398866951466, + 0.0036615151911973953, + 0.0040888553485274315, + 0.005372089333832264, + 0.009766693226993084, + 0.004929682705551386, + 0.0027482726145535707, + 0.00876754429191351, + 0.00936878751963377 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 2 + ], + "selected_reward": -0.011077099479734898 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "Shot from a low angle on a precipitous cliff, a sweeping, epic wide-angle shot pans slowly to the right, revealing a majestic Chinese-style mansion perched atop the mountain like an ancient behemoth. Constructed of dark, weathered wood and a massive stone foundation, the main structure is clad in dark green glazed tiles, gleaming in the sparse morning light. Its signature eaves and ridges rise gracefully and powerfully toward the sky like dragon wings. Thick, milky white mist flows slowly around the mansion like a living river, gently entwining carved columns and weaving between the intricate eaves. It sometimes engulfs the building, leaving only the outline of the roof visible, and sometimes quietly recedes to reveal its majestic fullness. The entire scene is shrouded in a cold and serene atmosphere. The faint light of dawn struggles to penetrate the clouds, gilding the edges of the flowing mist with a faint silver hue, creating a cinematic sense of isolation, mystery, and eternal power.", + "seed": 22652, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000027/group_8/index_792_seed_22652_dp_16_rank_32_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 8, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.16963563315221108, + "vlm_logprob": -0.16963563315221108, + "vlm_yes_no_margin": 4.575, + "vlm_token_credit": -0.16963563315221108, + "vlm_yes_ratio": 0.8, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot taken from a low angle on a precipitous cliff?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera pan slowly to the right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a Chinese-style mansion with dark green glazed tiles visible on the mountain?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is thick, milky white mist flowing around and partially engulfing the mansion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by sparse morning light with a cold and serene atmosphere?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.08393089473247528 + ], + "opposite_logprob_tokens": [ + -2.7089309692382812 + ], + "logprob_reward": -0.08393089473247528, + "yes_no_margin_reward": 2.625, + "token_credit_reward": -0.08393089473247528, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.4964216947555542, + 0.14607036113739014, + 0.09492475539445877, + 0.056891415268182755, + 0.027397390455007553, + 0.01443667896091938, + 0.02537294290959835, + 0.08266369998455048, + 0.03867868334054947, + 0.01714242435991764 + ], + "credit_map": [ + 0.06752346456050873, + 0.019418874755501747, + 0.1161956638097763, + 0.03617800027132034, + 0.03841521963477135, + 0.12669922411441803, + 0.03192662447690964, + 0.06006458029150963, + 0.016777483746409416, + 0.011944534257054329, + 0.021021569147706032, + 0.016789395362138748, + 0.011242502368986607, + 0.016288336366415024, + 0.01696365885436535, + 0.03504287451505661, + 0.0138185890391469, + 0.003998999483883381, + 0.016355568543076515, + 0.007141647394746542, + 0.008619651198387146, + 0.025944514200091362, + 0.006420848425477743, + 0.012624929659068584, + 0.005081813782453537, + 0.0060754381120204926, + 0.008647092618048191, + 0.011816458776593208, + 0.005253462586551905, + 0.006526607554405928, + 0.008127260953187943, + 0.005363276228308678, + 0.002825749572366476, + 0.003753229510039091, + 0.003983315080404282, + 0.004416460637003183, + 0.0030200257897377014, + 0.0024746914859861135, + 0.004075316712260246, + 0.002848598873242736, + 0.0017890085000544786, + 0.0012460175203159451, + 0.0016510594869032502, + 0.001563343103043735, + 0.00114538858179003, + 0.003934172913432121, + 0.0014889574376866221, + 0.0016187304863706231, + 0.003249646397307515, + 0.0021062230225652456, + 0.005409563425928354, + 0.003088441677391529, + 0.001991348108276725, + 0.0030299394857138395, + 0.003416393417865038, + 0.0030813843477517366, + 0.024072909727692604, + 0.003879138035699725, + 0.02943120338022709, + 0.007142915390431881, + 0.0020538358949124813, + 0.005174492485821247, + 0.004954894073307514, + 0.005954299587756395, + 0.005312364082783461, + 0.004802368581295013, + 0.0040955329313874245, + 0.005766027141362429, + 0.00398883456364274, + 0.0031466346699744463, + 0.0065251970663666725, + 0.005041722673922777, + 0.002486739307641983, + 0.0012388755567371845, + 0.002614892553538084, + 0.002428961917757988, + 0.0009241472580470145, + 0.002879897365346551, + 0.001894674845971167, + 0.002674233866855502 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 2 + ], + "selected_reward": -0.08393089473247528 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot taken from a low angle on a precipitous cliff?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera pan slowly to the right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a Chinese-style mansion with dark green glazed tiles visible on the mountain?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is thick, milky white mist flowing around and partially engulfing the mansion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by sparse morning light with a cold and serene atmosphere?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -0.7578263282775879 + ], + "opposite_logprob_tokens": [ + -0.6328263282775879 + ], + "logprob_reward": -0.7578263282775879, + "yes_no_margin_reward": -0.125, + "token_credit_reward": -0.7578263282775879, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.5162268280982971, + 0.041830673813819885, + 0.03240096569061279, + 0.03196669742465019, + 0.03589564561843872, + 0.040165070444345474, + 0.03851637989282608, + 0.199881911277771, + 0.035058360546827316, + 0.02805740386247635 + ], + "credit_map": [ + 0.06902536749839783, + 0.018655801191926003, + 0.12767061591148376, + 0.03383210301399231, + 0.039842888712882996, + 0.1350708305835724, + 0.03319691866636276, + 0.058932360261678696, + 0.0055657499469816685, + 0.004098961595445871, + 0.00601657759398222, + 0.003850249806419015, + 0.0044855098240077496, + 0.006357875652611256, + 0.0049575334414839745, + 0.0064982217736542225, + 0.004676520358771086, + 0.0019178253132849932, + 0.004969412460923195, + 0.0031657323706895113, + 0.0029181838035583496, + 0.007740639615803957, + 0.0028944076038897038, + 0.004118248820304871, + 0.0036523041781038046, + 0.002760721370577812, + 0.005194691475480795, + 0.004995840135961771, + 0.0029489523731172085, + 0.005380553659051657, + 0.0036858173552900553, + 0.003347820835188031, + 0.003299822797998786, + 0.0037977034226059914, + 0.004021838773041964, + 0.005957574117928743, + 0.003678843379020691, + 0.0037345360033214092, + 0.006048181094229221, + 0.00535714952275157, + 0.00443612365052104, + 0.0032011026050895452, + 0.007989832200109959, + 0.005937680136412382, + 0.0027614361606538296, + 0.005021973047405481, + 0.005115759093314409, + 0.005701164714992046, + 0.004944427404552698, + 0.004451779183000326, + 0.005894958507269621, + 0.00496989581733942, + 0.00384110608138144, + 0.003542365273460746, + 0.007023847661912441, + 0.0038479992654174566, + 0.06334400922060013, + 0.008188815787434578, + 0.07917556166648865, + 0.015355161391198635, + 0.004064356908202171, + 0.013796359300613403, + 0.009615450166165829, + 0.006342221051454544, + 0.006199998315423727, + 0.0035589169710874557, + 0.004295256454497576, + 0.005373203195631504, + 0.0026970840990543365, + 0.004128328990191221, + 0.005900534801185131, + 0.002905040280893445, + 0.0053510889410972595, + 0.0022132243029773235, + 0.003883871715515852, + 0.00400153873488307, + 0.001506828237324953, + 0.0038359046448022127, + 0.0036431963089853525, + 0.0036217535380274057 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 7, + 1 + ], + "selected_reward": -0.7578263282775879 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot taken from a low angle on a precipitous cliff?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera pan slowly to the right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a Chinese-style mansion with dark green glazed tiles visible on the mountain?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is thick, milky white mist flowing around and partially engulfing the mansion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by sparse morning light with a cold and serene atmosphere?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.00035363141796551645 + ], + "opposite_logprob_tokens": [ + -8.000353813171387 + ], + "logprob_reward": -0.00035363141796551645, + "yes_no_margin_reward": 8.0, + "token_credit_reward": -0.00035363141796551645, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.19663000106811523, + 0.07740309089422226, + 0.14910276234149933, + 0.052992574870586395, + 0.07492020726203918, + 0.11416938900947571, + 0.057198479771614075, + 0.07711195200681686, + 0.09731438755989075, + 0.10315721482038498 + ], + "credit_map": [ + 0.02544027753174305, + 0.01163645088672638, + 0.038305483758449554, + 0.019728446379303932, + 0.017819436267018318, + 0.036561235785484314, + 0.019494662061333656, + 0.027644000947475433, + 0.008508729748427868, + 0.008786874823272228, + 0.011158370412886143, + 0.009435521438717842, + 0.007731194607913494, + 0.012460768222808838, + 0.006823287811130285, + 0.012498338706791401, + 0.018082652240991592, + 0.00874081440269947, + 0.024386726319789886, + 0.012872357852756977, + 0.012961392290890217, + 0.03736991062760353, + 0.01224035769701004, + 0.022448532283306122, + 0.005467734765261412, + 0.004843674134463072, + 0.007643688935786486, + 0.011601314879953861, + 0.0040979874320328236, + 0.0075041018426418304, + 0.007207464426755905, + 0.004626602865755558, + 0.0074145435355603695, + 0.009550446644425392, + 0.01163455005735159, + 0.009285714477300644, + 0.007369309663772583, + 0.010281916707754135, + 0.01059576217085123, + 0.00878796074539423, + 0.01341539341956377, + 0.014641262590885162, + 0.013406166806817055, + 0.013456000946462154, + 0.010292368941009045, + 0.014278958551585674, + 0.015813136473298073, + 0.018866101279854774, + 0.01120021939277649, + 0.004695082549005747, + 0.017979104071855545, + 0.004801803268492222, + 0.0032562571577727795, + 0.0061277528293430805, + 0.004748950712382793, + 0.004389302339404821, + 0.013811611570417881, + 0.006376377772539854, + 0.01430294755846262, + 0.011872196570038795, + 0.0035543623380362988, + 0.010443010367453098, + 0.009940513409674168, + 0.006810923106968403, + 0.013483688235282898, + 0.013873478397727013, + 0.011459721252322197, + 0.01283936109393835, + 0.010478818789124489, + 0.01036459393799305, + 0.014307849109172821, + 0.010506863705813885, + 0.01756875589489937, + 0.011698824353516102, + 0.010658792220056057, + 0.01796245388686657, + 0.0077737364917993546, + 0.011583510786294937, + 0.015988394618034363, + 0.009922737255692482 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 5 + ], + "selected_reward": -0.00035363141796551645 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot taken from a low angle on a precipitous cliff?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera pan slowly to the right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a Chinese-style mansion with dark green glazed tiles visible on the mountain?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is thick, milky white mist flowing around and partially engulfing the mansion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by sparse morning light with a cold and serene atmosphere?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0008148210472427309 + ], + "opposite_logprob_tokens": [ + -7.125814914703369 + ], + "logprob_reward": -0.0008148210472427309, + "yes_no_margin_reward": 7.125, + "token_credit_reward": -0.0008148210472427309, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.6354372501373291, + 0.06439977884292603, + 0.01911224238574505, + 0.01313193328678608, + 0.01886807009577751, + 0.020248545333743095, + 0.020020117983222008, + 0.08967757225036621, + 0.028896387666463852, + 0.09020815044641495 + ], + "credit_map": [ + 0.08465471118688583, + 0.0212774109095335, + 0.14873604476451874, + 0.04604433476924896, + 0.048990823328495026, + 0.16611671447753906, + 0.041112110018730164, + 0.07850508391857147, + 0.007289517670869827, + 0.004182700999081135, + 0.009505034424364567, + 0.007583627011626959, + 0.004277363885194063, + 0.009354456327855587, + 0.008440986275672913, + 0.013766096904873848, + 0.0026180690620094538, + 0.0012218476040288806, + 0.002882750239223242, + 0.002298555336892605, + 0.001677142921835184, + 0.004156232345849276, + 0.002090164925903082, + 0.0021674796007573605, + 0.0016180449165403843, + 0.0009467000490985811, + 0.0018739310326054692, + 0.0019423675257712603, + 0.0010283144656568766, + 0.002063828520476818, + 0.0015111700631678104, + 0.002147577004507184, + 0.0017554786754772067, + 0.002231811871752143, + 0.0022746382746845484, + 0.0026530453469604254, + 0.0018345003481954336, + 0.0033501312136650085, + 0.001913584885187447, + 0.0028548785485327244, + 0.0020565425511449575, + 0.0024279651697725058, + 0.0021281717345118523, + 0.0023122874554246664, + 0.0019108583219349384, + 0.0023200062569230795, + 0.003258130047470331, + 0.0038345849607139826, + 0.0026597941759973764, + 0.0019879962783306837, + 0.0028910881374031305, + 0.002422444522380829, + 0.0015359759563580155, + 0.0029555519577115774, + 0.002775560598820448, + 0.002791706472635269, + 0.02447826787829399, + 0.004804575350135565, + 0.0313444659113884, + 0.00797285046428442, + 0.0025547498371452093, + 0.0075973644852638245, + 0.005169130861759186, + 0.00575616117566824, + 0.004059068858623505, + 0.0032908059656620026, + 0.0031924385111778975, + 0.004404810257256031, + 0.0024502575397491455, + 0.0038763131015002728, + 0.0038358625024557114, + 0.0037868316285312176, + 0.01378617249429226, + 0.007876060903072357, + 0.009157340042293072, + 0.012761170044541359, + 0.00592789426445961, + 0.015282997861504555, + 0.012140200473368168, + 0.013276313431560993 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 7 + ], + "selected_reward": -0.0008148210472427309 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot taken from a low angle on a precipitous cliff?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera pan slowly to the right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a Chinese-style mansion with dark green glazed tiles visible on the mountain?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is thick, milky white mist flowing around and partially engulfing the mansion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by sparse morning light with a cold and serene atmosphere?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005252490285784006 + ], + "opposite_logprob_tokens": [ + -5.255252361297607 + ], + "logprob_reward": -0.005252490285784006, + "yes_no_margin_reward": 5.25, + "token_credit_reward": -0.005252490285784006, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.6069425344467163, + 0.01841496303677559, + 0.055531520396471024, + 0.055256180465221405, + 0.05374610796570778, + 0.028023378923535347, + 0.04835992306470871, + 0.06427861750125885, + 0.03411214053630829, + 0.03533469885587692 + ], + "credit_map": [ + 0.07879745215177536, + 0.02147248201072216, + 0.13859431445598602, + 0.043174825608730316, + 0.04779292643070221, + 0.1592930257320404, + 0.04102485999464989, + 0.07679259777069092, + 0.0026957376394420862, + 0.0017048888839781284, + 0.0025979760102927685, + 0.002108622109517455, + 0.0013491768622770905, + 0.0015407106839120388, + 0.002552094403654337, + 0.003865756094455719, + 0.008081097155809402, + 0.0021967485081404448, + 0.009836132638156414, + 0.004057060927152634, + 0.005039227195084095, + 0.014891283586621284, + 0.004043281078338623, + 0.007386683486402035, + 0.0062508056871593, + 0.0038852246943861246, + 0.007096229121088982, + 0.00878080539405346, + 0.004466759506613016, + 0.00929694902151823, + 0.007222477812319994, + 0.00825691968202591, + 0.005426288116723299, + 0.0047576758079230785, + 0.008457078598439693, + 0.006230718921869993, + 0.003718029474839568, + 0.012402738444507122, + 0.00454444345086813, + 0.008209135383367538, + 0.0031590485014021397, + 0.0024008313193917274, + 0.003949879202991724, + 0.0033531533554196358, + 0.002263664500787854, + 0.005950940307229757, + 0.002743365243077278, + 0.004202492069453001, + 0.006956952624022961, + 0.004486153367906809, + 0.007600222248584032, + 0.005944684613496065, + 0.003952894825488329, + 0.005869048181921244, + 0.007348387502133846, + 0.0062015727162361145, + 0.018252894282341003, + 0.003951323218643665, + 0.020771723240613937, + 0.005069644656032324, + 0.002808817196637392, + 0.004757417365908623, + 0.0048522441647946835, + 0.0038145503494888544, + 0.006422289181500673, + 0.003644485492259264, + 0.004756200592964888, + 0.004810180980712175, + 0.0022535233292728662, + 0.0045394292101264, + 0.004925648681819439, + 0.002760380506515503, + 0.0049291751347482204, + 0.003090416546911001, + 0.005367981269955635, + 0.004617092665284872, + 0.002246494172140956, + 0.004886220209300518, + 0.004880025051534176, + 0.005317292641848326 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 7, + 2 + ], + "selected_reward": -0.005252490285784006 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "No", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "Shot from a low angle on a precipitous cliff, a sweeping, epic wide-angle shot pans slowly to the right, revealing a majestic Chinese-style mansion perched atop the mountain like an ancient behemoth. Constructed of dark, weathered wood and a massive stone foundation, the main structure is clad in dark green glazed tiles, gleaming in the sparse morning light. Its signature eaves and ridges rise gracefully and powerfully toward the sky like dragon wings. Thick, milky white mist flows slowly around the mansion like a living river, gently entwining carved columns and weaving between the intricate eaves. It sometimes engulfs the building, leaving only the outline of the roof visible, and sometimes quietly recedes to reveal its majestic fullness. The entire scene is shrouded in a cold and serene atmosphere. The faint light of dawn struggles to penetrate the clouds, gilding the edges of the flowing mist with a faint silver hue, creating a cinematic sense of isolation, mystery, and eternal power.", + "seed": 22653, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000027/group_8/index_792_seed_22653_dp_17_rank_34_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 8, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.07433917835587636, + "vlm_logprob": -0.07433917835587636, + "vlm_yes_no_margin": 4.675, + "vlm_token_credit": -0.07433917835587636, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot taken from a low angle on a precipitous cliff?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera pan slowly to the right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a Chinese-style mansion with dark green glazed tiles visible on the mountain?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is thick, milky white mist flowing around and partially engulfing the mansion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by sparse morning light with a cold and serene atmosphere?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04569941759109497 + ], + "opposite_logprob_tokens": [ + -3.29569935798645 + ], + "logprob_reward": -0.04569941759109497, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.04569941759109497, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1392134130001068, + 0.05172714218497276, + 0.04407556354999542, + 0.02737439237535, + 0.05042251944541931, + 0.04602756351232529, + 0.0694156065583229, + 0.3270803689956665, + 0.1794569343328476, + 0.06520646810531616 + ], + "credit_map": [ + 0.020387209951877594, + 0.010391512885689735, + 0.01197753194719553, + 0.00784092303365469, + 0.010259808041155338, + 0.02361331321299076, + 0.013818740844726562, + 0.040924400091171265, + 0.0072570983320474625, + 0.005173013545572758, + 0.005023080389946699, + 0.004779912531375885, + 0.003964209463447332, + 0.007582572288811207, + 0.00540900556370616, + 0.012538253329694271, + 0.007116047199815512, + 0.003672902937978506, + 0.004298312123864889, + 0.0031886198557913303, + 0.0035808663815259933, + 0.0062534078024327755, + 0.007326861843466759, + 0.008638547733426094, + 0.0033160082530230284, + 0.002117996336892247, + 0.0028712786734104156, + 0.0030353295151144266, + 0.0022834737319499254, + 0.0038583544082939625, + 0.003934727516025305, + 0.00595722533762455, + 0.005740824621170759, + 0.003964539617300034, + 0.0046462309546768665, + 0.008105252869427204, + 0.0030576016288250685, + 0.006678008008748293, + 0.009264478459954262, + 0.008965583518147469, + 0.004586984869092703, + 0.00597061961889267, + 0.004081330727785826, + 0.004608944058418274, + 0.0038611218333244324, + 0.005898468662053347, + 0.008051056414842606, + 0.008969040587544441, + 0.007379551883786917, + 0.006415586452931166, + 0.007444072049111128, + 0.010127189569175243, + 0.007316881325095892, + 0.00746537372469902, + 0.012007344514131546, + 0.011259601451456547, + 0.03986719250679016, + 0.022879166528582573, + 0.02864226885139942, + 0.02949279360473156, + 0.03186029568314552, + 0.055870044976472855, + 0.03396995738148689, + 0.08449867367744446, + 0.013075667433440685, + 0.010914108715951443, + 0.017939871177077293, + 0.02409696951508522, + 0.009730672463774681, + 0.03195354714989662, + 0.01798176020383835, + 0.053764358162879944, + 0.0033239745534956455, + 0.0030869035981595516, + 0.007580544333904982, + 0.01023043692111969, + 0.0031689757015556097, + 0.0101675596088171, + 0.009528174065053463, + 0.01811990514397621 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 8, + 0 + ], + "selected_reward": -0.04569941759109497 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot taken from a low angle on a precipitous cliff?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera pan slowly to the right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a Chinese-style mansion with dark green glazed tiles visible on the mountain?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is thick, milky white mist flowing around and partially engulfing the mansion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by sparse morning light with a cold and serene atmosphere?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.3134339153766632 + ], + "opposite_logprob_tokens": [ + -1.3134338855743408 + ], + "logprob_reward": -0.3134339153766632, + "yes_no_margin_reward": 1.0, + "token_credit_reward": -0.3134339153766632, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14452213048934937, + 0.06429579854011536, + 0.08231556415557861, + 0.08222895115613937, + 0.11844545602798462, + 0.10165916383266449, + 0.09759056568145752, + 0.16032078862190247, + 0.09484867751598358, + 0.05377291142940521 + ], + "credit_map": [ + 0.022993620485067368, + 0.014136670157313347, + 0.013697599060833454, + 0.010152882896363735, + 0.018125703558325768, + 0.01941653899848461, + 0.013809357769787312, + 0.0321897454559803, + 0.010489670559763908, + 0.005924995057284832, + 0.004317163955420256, + 0.004285631235688925, + 0.008110692724585533, + 0.008749603293836117, + 0.00730894785374403, + 0.015109087340533733, + 0.01662481762468815, + 0.007012035232037306, + 0.009052515029907227, + 0.007235005963593721, + 0.006362276151776314, + 0.01126004382967949, + 0.00801141932606697, + 0.01675744168460369, + 0.015248896554112434, + 0.006436298135668039, + 0.006173836532980204, + 0.007448398973792791, + 0.005697757937014103, + 0.012440149672329426, + 0.010337570682168007, + 0.01844603940844536, + 0.010369839146733284, + 0.008740593679249287, + 0.010977772995829582, + 0.018484124913811684, + 0.01005359087139368, + 0.01664121262729168, + 0.01880386658012867, + 0.024374449625611305, + 0.009446252137422562, + 0.013122962787747383, + 0.008586801588535309, + 0.010586277581751347, + 0.009499105624854565, + 0.01323818787932396, + 0.016294041648507118, + 0.020885538309812546, + 0.01003880426287651, + 0.015564551576972008, + 0.008801823481917381, + 0.010279159992933273, + 0.007593895774334669, + 0.011209972202777863, + 0.016155103221535683, + 0.01794724352657795, + 0.019440574571490288, + 0.021402820944786072, + 0.017269356176257133, + 0.019261684268712997, + 0.01889670081436634, + 0.020392993465065956, + 0.01717609167098999, + 0.026480555534362793, + 0.005326739978045225, + 0.01222190260887146, + 0.010708386078476906, + 0.01193803921341896, + 0.010179189033806324, + 0.012420735321938992, + 0.0114482706412673, + 0.020605407655239105, + 0.005382699426263571, + 0.004060536623001099, + 0.0073780580423772335, + 0.007325822487473488, + 0.003305514110252261, + 0.007010983768850565, + 0.007384859025478363, + 0.011924436315894127 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 0, + 4 + ], + "selected_reward": -0.3134339153766632 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot taken from a low angle on a precipitous cliff?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera pan slowly to the right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a Chinese-style mansion with dark green glazed tiles visible on the mountain?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is thick, milky white mist flowing around and partially engulfing the mansion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by sparse morning light with a cold and serene atmosphere?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0004472924047149718 + ], + "opposite_logprob_tokens": [ + -7.7504472732543945 + ], + "logprob_reward": -0.0004472924047149718, + "yes_no_margin_reward": 7.75, + "token_credit_reward": -0.0004472924047149718, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2317388355731964, + 0.08361265063285828, + 0.061416272073984146, + 0.06351081281900406, + 0.09060753136873245, + 0.07766707241535187, + 0.16309697926044464, + 0.12455861270427704, + 0.0649619847536087, + 0.038829244673252106 + ], + "credit_map": [ + 0.03422636166214943, + 0.021890370175242424, + 0.023469174280762672, + 0.015799159184098244, + 0.01935700513422489, + 0.03255544975399971, + 0.027534738183021545, + 0.0569065622985363, + 0.009150282479822636, + 0.007606753148138523, + 0.0095756771042943, + 0.007287748157978058, + 0.008549957536160946, + 0.010996166616678238, + 0.012112075462937355, + 0.01833399198949337, + 0.011783365160226822, + 0.005806613247841597, + 0.005930397193878889, + 0.005674345418810844, + 0.005383316893130541, + 0.008294300176203251, + 0.007579813711345196, + 0.010964118875563145, + 0.010437383316457272, + 0.00457207765430212, + 0.005593011621385813, + 0.005477385129779577, + 0.006121000740677118, + 0.009288528934121132, + 0.008607053197920322, + 0.013414372690021992, + 0.010280812159180641, + 0.007088414393365383, + 0.009453587234020233, + 0.014777274802327156, + 0.00696070259436965, + 0.010459049604833126, + 0.017535226419568062, + 0.014052466489374638, + 0.005906155798584223, + 0.007920265197753906, + 0.00715987803414464, + 0.008618191815912724, + 0.007163247559219599, + 0.010791100561618805, + 0.015031973831355572, + 0.015076260082423687, + 0.01744701713323593, + 0.017016824334859848, + 0.017177708446979523, + 0.017834236845374107, + 0.01529101375490427, + 0.01643991284072399, + 0.037195026874542236, + 0.02469523996114731, + 0.010185823775827885, + 0.015465165488421917, + 0.014879201538860798, + 0.015119271352887154, + 0.01475383061915636, + 0.015115086920559406, + 0.017341939732432365, + 0.02169829048216343, + 0.004589039366692305, + 0.008703446947038174, + 0.008908444084227085, + 0.007661230396479368, + 0.006047403439879417, + 0.0076409392058849335, + 0.007078621070832014, + 0.014332861639559269, + 0.00371383735910058, + 0.004912588279694319, + 0.004897305276244879, + 0.004350301343947649, + 0.002773271407932043, + 0.005369776394218206, + 0.003739121835678816, + 0.009073044173419476 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 6, + 7 + ], + "selected_reward": -0.0004472924047149718 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot taken from a low angle on a precipitous cliff?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera pan slowly to the right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a Chinese-style mansion with dark green glazed tiles visible on the mountain?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is thick, milky white mist flowing around and partially engulfing the mansion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by sparse morning light with a cold and serene atmosphere?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0010417040903121233 + ], + "opposite_logprob_tokens": [ + -6.876041889190674 + ], + "logprob_reward": -0.0010417040903121233, + "yes_no_margin_reward": 6.875, + "token_credit_reward": -0.0010417040903121233, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.25464707612991333, + 0.14180387556552887, + 0.02698013372719288, + 0.027679383754730225, + 0.07439297437667847, + 0.06352268159389496, + 0.1157279759645462, + 0.10032962262630463, + 0.16270245611667633, + 0.03221382200717926 + ], + "credit_map": [ + 0.03713370859622955, + 0.01747574843466282, + 0.022589733824133873, + 0.014954504556953907, + 0.013758087530732155, + 0.0415283739566803, + 0.03148236870765686, + 0.07572456449270248, + 0.016775894910097122, + 0.011077648028731346, + 0.011936565861105919, + 0.009995670057833195, + 0.011094952002167702, + 0.02317308634519577, + 0.019417332485318184, + 0.038332730531692505, + 0.00519136805087328, + 0.002358786528930068, + 0.002018030732870102, + 0.002357935067266226, + 0.0024758647195994854, + 0.003569681430235505, + 0.003910706844180822, + 0.005097758490592241, + 0.00368253025226295, + 0.0030502616427838802, + 0.0027724714018404484, + 0.0037819454446434975, + 0.003035284113138914, + 0.0033607843797653913, + 0.0037461640313267708, + 0.004249942023307085, + 0.006369356997311115, + 0.00609869509935379, + 0.007189018186181784, + 0.011735777370631695, + 0.006432328838855028, + 0.00873024296015501, + 0.014462013728916645, + 0.013375530950725079, + 0.0064074574038386345, + 0.00812336802482605, + 0.006427654065191746, + 0.005715782288461924, + 0.005314036272466183, + 0.008666381239891052, + 0.009442389942705631, + 0.013425614684820175, + 0.016067100688815117, + 0.014240717515349388, + 0.01280047744512558, + 0.012880288064479828, + 0.008413536474108696, + 0.011951815336942673, + 0.023315120488405228, + 0.016058923676609993, + 0.013776787556707859, + 0.009620768949389458, + 0.009778181090950966, + 0.009654565714299679, + 0.011485729366540909, + 0.014655157923698425, + 0.00980774313211441, + 0.02155068702995777, + 0.010109489783644676, + 0.0066913641057908535, + 0.016468586400151253, + 0.015054515562951565, + 0.007292267866432667, + 0.032261405140161514, + 0.017352955415844917, + 0.0574718713760376, + 0.0019230664474889636, + 0.001783058512955904, + 0.003924255724996328, + 0.004928030539304018, + 0.0019631574396044016, + 0.0052409702911973, + 0.0030318102799355984, + 0.009419473819434643 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 1 + ], + "selected_reward": -0.0010417040903121233 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot taken from a low angle on a precipitous cliff?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera pan slowly to the right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a Chinese-style mansion with dark green glazed tiles visible on the mountain?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is thick, milky white mist flowing around and partially engulfing the mansion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by sparse morning light with a cold and serene atmosphere?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.011073562316596508 + ], + "opposite_logprob_tokens": [ + -4.511073589324951 + ], + "logprob_reward": -0.011073562316596508, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.011073562316596508, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.20488719642162323, + 0.08778499066829681, + 0.05024799332022667, + 0.04703548178076744, + 0.03201073035597801, + 0.06391552090644836, + 0.13398823142051697, + 0.2642556428909302, + 0.08384320139884949, + 0.03203093633055687 + ], + "credit_map": [ + 0.04108103737235069, + 0.017071174457669258, + 0.018006613478064537, + 0.010703057050704956, + 0.017333516851067543, + 0.0292705986648798, + 0.017897510901093483, + 0.053523723036050797, + 0.014860950410366058, + 0.006540501024574041, + 0.010526775382459164, + 0.007898660376667976, + 0.008240844123065472, + 0.0107687097042799, + 0.012706398963928223, + 0.01624215766787529, + 0.008592074736952782, + 0.004390411544591188, + 0.004699669778347015, + 0.0042088706977665424, + 0.004041704349219799, + 0.00723218871280551, + 0.006601972505450249, + 0.010481104254722595, + 0.0038445836398750544, + 0.005343856755644083, + 0.0040276795625686646, + 0.006059897597879171, + 0.0069402484223246574, + 0.005725465714931488, + 0.006310510914772749, + 0.008783242665231228, + 0.0042482540011405945, + 0.0032909619621932507, + 0.0032007836271077394, + 0.004833748564124107, + 0.0032592632342129946, + 0.004082770552486181, + 0.0038511333987116814, + 0.005243818275630474, + 0.007354112342000008, + 0.008117643184959888, + 0.006900610867887735, + 0.0058523546904325485, + 0.006140169221907854, + 0.008645929396152496, + 0.008823834359645844, + 0.012080874294042587, + 0.017121974378824234, + 0.01995748281478882, + 0.01612883061170578, + 0.014988814480602741, + 0.010268538258969784, + 0.01605055294930935, + 0.01862979680299759, + 0.02084224671125412, + 0.035677168518304825, + 0.02024487964808941, + 0.024827728047966957, + 0.02599758468568325, + 0.027935251593589783, + 0.042121171951293945, + 0.026058174669742584, + 0.06139371544122696, + 0.005576429422944784, + 0.010356440208852291, + 0.013222852721810341, + 0.015397817827761173, + 0.008392284624278545, + 0.0077997311018407345, + 0.009607701562345028, + 0.013489951379597187, + 0.0027197261806577444, + 0.0033213014248758554, + 0.004755988251417875, + 0.005382530391216278, + 0.002748959232121706, + 0.0027400068938732147, + 0.006041836459189653, + 0.004320588428527117 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 0, + 6 + ], + "selected_reward": -0.011073562316596508 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "Shot from a low angle on a precipitous cliff, a sweeping, epic wide-angle shot pans slowly to the right, revealing a majestic Chinese-style mansion perched atop the mountain like an ancient behemoth. Constructed of dark, weathered wood and a massive stone foundation, the main structure is clad in dark green glazed tiles, gleaming in the sparse morning light. Its signature eaves and ridges rise gracefully and powerfully toward the sky like dragon wings. Thick, milky white mist flows slowly around the mansion like a living river, gently entwining carved columns and weaving between the intricate eaves. It sometimes engulfs the building, leaving only the outline of the roof visible, and sometimes quietly recedes to reveal its majestic fullness. The entire scene is shrouded in a cold and serene atmosphere. The faint light of dawn struggles to penetrate the clouds, gilding the edges of the flowing mist with a faint silver hue, creating a cinematic sense of isolation, mystery, and eternal power.", + "seed": 22654, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000027/group_8/index_792_seed_22654_dp_17_rank_34_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 8, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.0838952729711309, + "vlm_logprob": -0.0838952729711309, + "vlm_yes_no_margin": 4.074999976158142, + "vlm_token_credit": -0.0838952729711309, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot taken from a low angle on a precipitous cliff?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera pan slowly to the right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a Chinese-style mansion with dark green glazed tiles visible on the mountain?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is thick, milky white mist flowing around and partially engulfing the mansion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by sparse morning light with a cold and serene atmosphere?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.15003713965415955 + ], + "opposite_logprob_tokens": [ + -2.0250370502471924 + ], + "logprob_reward": -0.15003713965415955, + "yes_no_margin_reward": 1.8749998807907104, + "token_credit_reward": -0.15003713965415955, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13920892775058746, + 0.09996026754379272, + 0.046966370195150375, + 0.02459070459008217, + 0.2580525279045105, + 0.0720532089471817, + 0.09557786583900452, + 0.11081718653440475, + 0.09058262407779694, + 0.06219039484858513 + ], + "credit_map": [ + 0.013409062288701534, + 0.00889655388891697, + 0.013675298541784286, + 0.01844281144440174, + 0.011196909472346306, + 0.018750954419374466, + 0.03824799507856369, + 0.016589339822530746, + 0.00809244904667139, + 0.007587825413793325, + 0.00821656920015812, + 0.009452726691961288, + 0.013091819360852242, + 0.018873631954193115, + 0.026113485917448997, + 0.008531762287020683, + 0.0033206064254045486, + 0.0060056764632463455, + 0.005621634889394045, + 0.006835410371422768, + 0.005305134225636721, + 0.0043563139624893665, + 0.007923029363155365, + 0.007598564960062504, + 0.0023180716671049595, + 0.0026536204386502504, + 0.004001385997980833, + 0.0030785230919718742, + 0.004238311201334, + 0.0021958346478641033, + 0.002948247827589512, + 0.0031567097175866365, + 0.023048704490065575, + 0.05885990336537361, + 0.023763757199048996, + 0.029588712379336357, + 0.04157634824514389, + 0.01971004344522953, + 0.03455326706171036, + 0.026951786130666733, + 0.0061553241685032845, + 0.007820216938853264, + 0.010167095810174942, + 0.009556357748806477, + 0.011400222778320312, + 0.0075082131661474705, + 0.009874874725937843, + 0.009570904076099396, + 0.006019226275384426, + 0.00792359933257103, + 0.008266320452094078, + 0.017355293035507202, + 0.006619984749704599, + 0.011015268974006176, + 0.018614113330841064, + 0.01976405829191208, + 0.005928007420152426, + 0.014617531560361385, + 0.017860598862171173, + 0.019600460305809975, + 0.011970369145274162, + 0.01030677743256092, + 0.01528932061046362, + 0.01524412352591753, + 0.006050125230103731, + 0.007959478534758091, + 0.014278990216553211, + 0.020643943920731544, + 0.00658248970285058, + 0.00985258724540472, + 0.013878137804567814, + 0.011336864903569221, + 0.0035625440068542957, + 0.004434005357325077, + 0.004563712514936924, + 0.014141267165541649, + 0.004881403874605894, + 0.007614223286509514, + 0.014696076512336731, + 0.008297164924442768 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 0, + 7 + ], + "selected_reward": -0.15003713965415955 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot taken from a low angle on a precipitous cliff?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera pan slowly to the right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a Chinese-style mansion with dark green glazed tiles visible on the mountain?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is thick, milky white mist flowing around and partially engulfing the mansion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by sparse morning light with a cold and serene atmosphere?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.2521166503429413 + ], + "opposite_logprob_tokens": [ + -1.5021166801452637 + ], + "logprob_reward": -0.2521166503429413, + "yes_no_margin_reward": 1.25, + "token_credit_reward": -0.2521166503429413, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11453328281641006, + 0.04767506942152977, + 0.05776963382959366, + 0.07436871528625488, + 0.08901200443506241, + 0.10563715547323227, + 0.09911639243364334, + 0.06730262190103531, + 0.055117566138505936, + 0.28946754336357117 + ], + "credit_map": [ + 0.0055389669723808765, + 0.003925743978470564, + 0.006799882743507624, + 0.01342732273042202, + 0.006985030137002468, + 0.026710016652941704, + 0.03879154846072197, + 0.012354785576462746, + 0.006092630326747894, + 0.005404227413237095, + 0.005439938977360725, + 0.00520608015358448, + 0.008320242166519165, + 0.004716640338301659, + 0.0067984736524522305, + 0.005696840118616819, + 0.005806494504213333, + 0.005851011257618666, + 0.010316687636077404, + 0.007529618684202433, + 0.007108157500624657, + 0.0037718862295150757, + 0.009499934501945972, + 0.007885847240686417, + 0.007542931940406561, + 0.007928614504635334, + 0.009784682653844357, + 0.011024408973753452, + 0.010195993818342686, + 0.006365887820720673, + 0.01090801041573286, + 0.010618196800351143, + 0.007429182529449463, + 0.019339874386787415, + 0.007528505753725767, + 0.013725907541811466, + 0.01270835381001234, + 0.0059371208772063255, + 0.01394098810851574, + 0.008402080275118351, + 0.00910906307399273, + 0.02374831959605217, + 0.009772470220923424, + 0.010996987111866474, + 0.01646273583173752, + 0.00880465842783451, + 0.016941195353865623, + 0.009801727719604969, + 0.006656901445239782, + 0.007487672846764326, + 0.010200931690633297, + 0.012687270529568195, + 0.007376831956207752, + 0.013675316236913204, + 0.018012918531894684, + 0.02301855757832527, + 0.004824507515877485, + 0.007089151069521904, + 0.007330548483878374, + 0.011455731466412544, + 0.005689781159162521, + 0.008159134536981583, + 0.01031608134508133, + 0.012437695637345314, + 0.004204097203910351, + 0.005547064356505871, + 0.005761610809713602, + 0.012165226973593235, + 0.006049931515008211, + 0.004993439186364412, + 0.010018940083682537, + 0.006377259735018015, + 0.014560342766344547, + 0.016857970505952835, + 0.02147567644715309, + 0.05610952153801918, + 0.01731817238032818, + 0.0406661219894886, + 0.0887705609202385, + 0.03370919078588486 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 0, + 5 + ], + "selected_reward": -0.2521166503429413 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot taken from a low angle on a precipitous cliff?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera pan slowly to the right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a Chinese-style mansion with dark green glazed tiles visible on the mountain?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is thick, milky white mist flowing around and partially engulfing the mansion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by sparse morning light with a cold and serene atmosphere?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0011872155591845512 + ], + "opposite_logprob_tokens": [ + -6.751187324523926 + ], + "logprob_reward": -0.0011872155591845512, + "yes_no_margin_reward": 6.75, + "token_credit_reward": -0.0011872155591845512, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.35047435760498047, + 0.1039191484451294, + 0.03518446907401085, + 0.025819925591349602, + 0.22795715928077698, + 0.04853038117289543, + 0.06476209312677383, + 0.07125736773014069, + 0.022383883595466614, + 0.0497112050652504 + ], + "credit_map": [ + 0.03219841048121452, + 0.017012426629662514, + 0.027891095727682114, + 0.057394690811634064, + 0.042507417500019073, + 0.03981209918856621, + 0.10845103114843369, + 0.02520717866718769, + 0.005916328635066748, + 0.007904275320470333, + 0.008042019791901112, + 0.010103795677423477, + 0.011391486041247845, + 0.018523624166846275, + 0.026065561920404434, + 0.015972066670656204, + 0.003119200235232711, + 0.004788899328559637, + 0.004209700506180525, + 0.005617306102067232, + 0.0046561965718865395, + 0.0025816711131483316, + 0.005138184875249863, + 0.0050733089447021484, + 0.002817915054038167, + 0.0025778664276003838, + 0.004137265495955944, + 0.004845930263400078, + 0.0030349872540682554, + 0.0016723095905035734, + 0.004002715460956097, + 0.00273093837313354, + 0.02078569494187832, + 0.049571357667446136, + 0.020503705367445946, + 0.027356181293725967, + 0.03322886675596237, + 0.01753859780728817, + 0.034260377287864685, + 0.024712372571229935, + 0.0029348095413297415, + 0.005109925288707018, + 0.005536200478672981, + 0.005571749992668629, + 0.005420010071247816, + 0.004835694096982479, + 0.011889379471540451, + 0.007232617121189833, + 0.003945577889680862, + 0.0052534183487296104, + 0.004792589228600264, + 0.010584070347249508, + 0.004761396907269955, + 0.007597235031425953, + 0.012924312613904476, + 0.014903492294251919, + 0.005180213134735823, + 0.0076810740865767, + 0.008613255806267262, + 0.012360284104943275, + 0.006298289634287357, + 0.007296482101082802, + 0.012640255503356457, + 0.011187506839632988, + 0.0016302132280543447, + 0.002591719152405858, + 0.002339371480047703, + 0.003971558064222336, + 0.002162184100598097, + 0.0025511954445391893, + 0.0035101231187582016, + 0.0036275209859013557, + 0.0030442881397902966, + 0.00788815412670374, + 0.003936668857932091, + 0.00949916522949934, + 0.004769858438521624, + 0.005614562425762415, + 0.00903062243014574, + 0.005927884019911289 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 1 + ], + "selected_reward": -0.0011872155591845512 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot taken from a low angle on a precipitous cliff?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera pan slowly to the right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a Chinese-style mansion with dark green glazed tiles visible on the mountain?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is thick, milky white mist flowing around and partially engulfing the mansion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by sparse morning light with a cold and serene atmosphere?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.00194193993229419 + ], + "opposite_logprob_tokens": [ + -6.251942157745361 + ], + "logprob_reward": -0.00194193993229419, + "yes_no_margin_reward": 6.25, + "token_credit_reward": -0.00194193993229419, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3198133707046509, + 0.1197374016046524, + 0.04131196439266205, + 0.02428913116455078, + 0.08324882388114929, + 0.18532317876815796, + 0.13299600780010223, + 0.03553466498851776, + 0.02650841325521469, + 0.03123706951737404 + ], + "credit_map": [ + 0.02308419905602932, + 0.010923422873020172, + 0.01886708103120327, + 0.04605880007147789, + 0.01500617153942585, + 0.05656854063272476, + 0.11424748599529266, + 0.03505765646696091, + 0.005203242413699627, + 0.00487184152007103, + 0.006077120080590248, + 0.012147895060479641, + 0.007555257063359022, + 0.029328899458050728, + 0.03966749832034111, + 0.014885635115206242, + 0.004295737016946077, + 0.004664105363190174, + 0.007910095155239105, + 0.006678195204585791, + 0.003620198694989085, + 0.002757023088634014, + 0.005930003244429827, + 0.005456605460494757, + 0.0024408299941569567, + 0.0028195895720273256, + 0.004156399518251419, + 0.0036633415147662163, + 0.003043872769922018, + 0.0016832492547109723, + 0.004096873104572296, + 0.002384975552558899, + 0.008192897774279118, + 0.017760753631591797, + 0.009031888097524643, + 0.009916705079376698, + 0.011080685071647167, + 0.006324869580566883, + 0.012581918388605118, + 0.00835910253226757, + 0.013965980149805546, + 0.03321080654859543, + 0.014881080016493797, + 0.019269440323114395, + 0.026389310136437416, + 0.017475832253694534, + 0.03755360469222069, + 0.02257712371647358, + 0.008919905871152878, + 0.008516739122569561, + 0.010029837489128113, + 0.017608726397156715, + 0.009385406970977783, + 0.018411671742796898, + 0.02526071108877659, + 0.03486301377415657, + 0.002344186184927821, + 0.004703888203948736, + 0.004779115319252014, + 0.006160040386021137, + 0.003717255312949419, + 0.0038585453294217587, + 0.00510399928316474, + 0.00486763147637248, + 0.0019222833216190338, + 0.0029565063305199146, + 0.00292628793977201, + 0.005087620113044977, + 0.0025798361748456955, + 0.002786771161481738, + 0.004408717155456543, + 0.003840390359982848, + 0.0015734932385385036, + 0.00303840939886868, + 0.0022233794443309307, + 0.006341591943055391, + 0.002756342524662614, + 0.0040717534720897675, + 0.006934348959475756, + 0.004297747742384672 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 5, + 6 + ], + "selected_reward": -0.00194193993229419 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot taken from a low angle on a precipitous cliff?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera pan slowly to the right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a Chinese-style mansion with dark green glazed tiles visible on the mountain?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is thick, milky white mist flowing around and partially engulfing the mansion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by sparse morning light with a cold and serene atmosphere?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.014193419367074966 + ], + "opposite_logprob_tokens": [ + -4.264193534851074 + ], + "logprob_reward": -0.014193419367074966, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.014193419367074966, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.17288120090961456, + 0.0680493637919426, + 0.01739063300192356, + 0.013139527291059494, + 0.5236647725105286, + 0.03405565023422241, + 0.12314551323652267, + 0.01346770953387022, + 0.017696306109428406, + 0.016509274020791054 + ], + "credit_map": [ + 0.012546369805932045, + 0.005990029778331518, + 0.008915693499147892, + 0.023698389530181885, + 0.007269520312547684, + 0.030218997970223427, + 0.061488814651966095, + 0.022753402590751648, + 0.004793294705450535, + 0.004989226348698139, + 0.00434845220297575, + 0.005810424219816923, + 0.0057803126983344555, + 0.018162770196795464, + 0.017110144719481468, + 0.007054740097373724, + 0.0016666038427501917, + 0.0019564025569707155, + 0.0018327046418562531, + 0.0022414550185203552, + 0.0020985445007681847, + 0.0016075377352535725, + 0.0032319773454219103, + 0.002755408873781562, + 0.0009736053762026131, + 0.001568731153383851, + 0.001392270321957767, + 0.001927881152369082, + 0.0015717940405011177, + 0.0009447203483432531, + 0.0029333452694118023, + 0.001827180734835565, + 0.047108929604291916, + 0.11543929576873779, + 0.04569058492779732, + 0.06198887154459953, + 0.08034901320934296, + 0.03978533297777176, + 0.07663667947053909, + 0.05666607245802879, + 0.002082039602100849, + 0.004920856561511755, + 0.002576739527285099, + 0.004281428176909685, + 0.004420253448188305, + 0.0031992776785045862, + 0.008888807147741318, + 0.003686251351609826, + 0.008100684732198715, + 0.00751976715400815, + 0.009395698085427284, + 0.01995106227695942, + 0.008786877617239952, + 0.015830636024475098, + 0.022279756143689156, + 0.03128104284405708, + 0.0008414925541728735, + 0.0019276300445199013, + 0.001623752061277628, + 0.002472642343491316, + 0.0014446317218244076, + 0.001335827400907874, + 0.0020064355339854956, + 0.0018152978736907244, + 0.0012853959342464805, + 0.0021606546360999346, + 0.001852463698014617, + 0.0032764736097306013, + 0.0016348607605323195, + 0.002204459859058261, + 0.0030282246880233288, + 0.0022537754848599434, + 0.0009041020530276, + 0.0016346924239769578, + 0.0015056453412398696, + 0.003650456201285124, + 0.0013081934303045273, + 0.0022825782652944326, + 0.0029551784973591566, + 0.0022684279829263687 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 0, + 6 + ], + "selected_reward": -0.014193419367074966 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000034/group_1/index_tensor([3118, 3118])_group_1_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000034/group_1/index_tensor([3118, 3118])_group_1_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..4b50ab2e9e48181a2faa0ff483c59142c6a01054 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000034/group_1/index_tensor([3118, 3118])_group_1_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "static medium shot, shot from a slightly upward angle, focuses squarely on the legendary Overlord of Chu. He stands amidst a desolate battlefield rendered with thick, thin, dry, and wet ink brushstrokes. His dark iron armor, shattered by the bloody battles, is covered in deep sword marks and dents, its edges curled, and several broken arrows remain lodged within. Dark blood and dirt blend, forming a smudge on the metal surface like solidified ink. His long, tied-up hair is disheveled, a few sweat-soaked strands clinging to his stern face, shrouded in dust and fatigue. In that instant, he throws his head back, the muscles in his neck tensing, and his mouth wide open, he lets out a silent yet earth-shattering roar into the oppressive, leaden sky, as if stained with ink. His eyes, wide open, burn with resentment, rage, and the endless sadness of a hero's end. The entire picture presents an extreme tragic aesthetic. The mountains in the background and the fallen weapons are blurred with light ink. Only his tense posture and desperate cry constitute the absolute center of the picture, creating a lonely and magnificent epic atmosphere.", + "seed": 7932, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000034/group_1/index_3118_seed_7932_dp_2_rank_4_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 1, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.035097304824739696, + "vlm_logprob": -0.035097304824739696, + "vlm_yes_no_margin": 3.825, + "vlm_token_credit": -0.035097304824739696, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject the legendary Overlord of Chu wearing dark iron armor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the armor visibly damaged with deep sword marks, dents, curled edges, and broken arrows lodged in it?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the character's long hair tied up, disheveled, and covered in sweat, dust, and fatigue?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the character shown throwing his head back with tensed neck muscles and an open mouth in a roaring pose?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene rendered with a mix of thick, thin, dry, and wet ink brushstrokes in a static medium shot from a slightly upward angle?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.10391230136156082 + ], + "opposite_logprob_tokens": [ + -2.353912353515625 + ], + "logprob_reward": -0.10391230136156082, + "yes_no_margin_reward": 2.25, + "token_credit_reward": -0.10391230136156082, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1834399402141571, + 0.07842403650283813, + 0.4433346688747406, + 0.07706406712532043, + 0.04279500991106033, + 0.05201953649520874, + 0.02496069110929966, + 0.034421563148498535, + 0.029323019087314606, + 0.03421745449304581 + ], + "credit_map": [ + 0.017012085765600204, + 0.02452661097049713, + 0.02710696868598461, + 0.03632211685180664, + 0.019694779068231583, + 0.023710856214165688, + 0.020645614713430405, + 0.014420918188989162, + 0.007150876801460981, + 0.0073266299441456795, + 0.01578417234122753, + 0.009245617315173149, + 0.011731884442269802, + 0.010025208815932274, + 0.009111677296459675, + 0.008047973737120628, + 0.10763945430517197, + 0.020837748423218727, + 0.16816553473472595, + 0.03402173891663551, + 0.02807929366827011, + 0.015240008011460304, + 0.028773054480552673, + 0.04057784378528595, + 0.016449036076664925, + 0.009150954894721508, + 0.006001715548336506, + 0.005873052403330803, + 0.011442171409726143, + 0.009961873292922974, + 0.007814268581569195, + 0.010370994918048382, + 0.0042974697425961494, + 0.008783792145550251, + 0.004423300735652447, + 0.003906810190528631, + 0.006336885970085859, + 0.006256498862057924, + 0.00437304750084877, + 0.004417201969772577, + 0.0062040602788329124, + 0.005448712036013603, + 0.006367792841047049, + 0.007781568914651871, + 0.007959161885082722, + 0.006002324633300304, + 0.007563954684883356, + 0.00469196354970336, + 0.003136887215077877, + 0.0024368539452552795, + 0.004074920900166035, + 0.0027304058894515038, + 0.004283970221877098, + 0.0027851618360728025, + 0.003682127920910716, + 0.0018303634133189917, + 0.004817104898393154, + 0.0036880047991871834, + 0.0036928653717041016, + 0.003913980908691883, + 0.005856805481016636, + 0.0063440511003136635, + 0.0033162073232233524, + 0.0027925430331379175, + 0.0038543143309652805, + 0.0036821013782173395, + 0.002696513431146741, + 0.0024301528465002775, + 0.004331175237894058, + 0.006705467123538256, + 0.002457321621477604, + 0.0031659728847444057, + 0.004481571260839701, + 0.003099781461060047, + 0.0049133580178022385, + 0.003671536222100258, + 0.004873997997492552, + 0.005940036848187447, + 0.0037827305495738983, + 0.0034544444642961025 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 0, + 1 + ], + "selected_reward": -0.10391230136156082 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject the legendary Overlord of Chu wearing dark iron armor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the armor visibly damaged with deep sword marks, dents, curled edges, and broken arrows lodged in it?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the character's long hair tied up, disheveled, and covered in sweat, dust, and fatigue?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the character shown throwing his head back with tensed neck muscles and an open mouth in a roaring pose?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene rendered with a mix of thick, thin, dry, and wet ink brushstrokes in a static medium shot from a slightly upward angle?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02984645590186119 + ], + "opposite_logprob_tokens": [ + -3.529846429824829 + ], + "logprob_reward": -0.02984645590186119, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.02984645590186119, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2453315109014511, + 0.09339158982038498, + 0.2424948513507843, + 0.049259692430496216, + 0.053916409611701965, + 0.11200625449419022, + 0.06588710099458694, + 0.07382910698652267, + 0.025157539173960686, + 0.038725875318050385 + ], + "credit_map": [ + 0.02593209408223629, + 0.02570805698633194, + 0.04967774078249931, + 0.054793305695056915, + 0.020934412255883217, + 0.03031138703227043, + 0.02019273117184639, + 0.01778179407119751, + 0.008857805281877518, + 0.009172185324132442, + 0.016364410519599915, + 0.009672542102634907, + 0.015164852142333984, + 0.01311701349914074, + 0.009492439217865467, + 0.011550345458090305, + 0.057290807366371155, + 0.012206293642520905, + 0.09380585700273514, + 0.017139717936515808, + 0.0167588721960783, + 0.008185996674001217, + 0.013499063439667225, + 0.0236082524061203, + 0.006911080330610275, + 0.0045313285663723946, + 0.004105737432837486, + 0.003998415544629097, + 0.007073872722685337, + 0.01221673283725977, + 0.004941814579069614, + 0.0054807113483548164, + 0.005321086384356022, + 0.010284842923283577, + 0.006034129299223423, + 0.006529628299176693, + 0.008293327875435352, + 0.007283844985067844, + 0.006176258437335491, + 0.003993290010839701, + 0.013577853329479694, + 0.010259922593832016, + 0.009452595375478268, + 0.012153787538409233, + 0.021746447309851646, + 0.022385532036423683, + 0.01430057268589735, + 0.008129545487463474, + 0.007391977123916149, + 0.007906483486294746, + 0.010541209019720554, + 0.010959354229271412, + 0.008445785380899906, + 0.00869828648865223, + 0.007886047475039959, + 0.004057961516082287, + 0.008971672505140305, + 0.007904283702373505, + 0.00815736223012209, + 0.008299894630908966, + 0.01355469599366188, + 0.01133357360959053, + 0.00885849166661501, + 0.006749132182449102, + 0.003305043326690793, + 0.004065253771841526, + 0.0024174002464860678, + 0.0025139828212559223, + 0.003868348430842161, + 0.004325016401708126, + 0.0022895121946930885, + 0.0023729833774268627, + 0.003956410568207502, + 0.003842425998300314, + 0.0074306875467300415, + 0.005529348272830248, + 0.005456176120787859, + 0.0042201015166938305, + 0.003270261688157916, + 0.005020461976528168 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 5 + ], + "selected_reward": -0.02984645590186119 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject the legendary Overlord of Chu wearing dark iron armor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the armor visibly damaged with deep sword marks, dents, curled edges, and broken arrows lodged in it?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the character's long hair tied up, disheveled, and covered in sweat, dust, and fatigue?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the character shown throwing his head back with tensed neck muscles and an open mouth in a roaring pose?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene rendered with a mix of thick, thin, dry, and wet ink brushstrokes in a static medium shot from a slightly upward angle?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.008645012974739075 + ], + "opposite_logprob_tokens": [ + -4.758645057678223 + ], + "logprob_reward": -0.008645012974739075, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.008645012974739075, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1839267909526825, + 0.031139347702264786, + 0.6078331470489502, + 0.0255520511418581, + 0.017170172184705734, + 0.03266620263457298, + 0.028786852955818176, + 0.031116079539060593, + 0.02117147482931614, + 0.020637936890125275 + ], + "credit_map": [ + 0.03165758401155472, + 0.018560532480478287, + 0.07280579209327698, + 0.02156076394021511, + 0.007256790995597839, + 0.013354826718568802, + 0.010656475089490414, + 0.008074024692177773, + 0.0033202357590198517, + 0.002812705235555768, + 0.0063596125692129135, + 0.0024894350208342075, + 0.004923731088638306, + 0.00470438227057457, + 0.0040000611916184425, + 0.00252918410114944, + 0.15366417169570923, + 0.02733573690056801, + 0.24396100640296936, + 0.04567055031657219, + 0.02677561156451702, + 0.01961221732199192, + 0.03302594646811485, + 0.05778783932328224, + 0.005319892894476652, + 0.0027649898547679186, + 0.0018837656825780869, + 0.0011203672038391232, + 0.00379678956232965, + 0.004107831045985222, + 0.002175759756937623, + 0.0043826536275446415, + 0.0022634577471762896, + 0.0024552280083298683, + 0.002514120191335678, + 0.0015620930353179574, + 0.0019697321113198996, + 0.0019040006445720792, + 0.0019591290038079023, + 0.0025424116756767035, + 0.004122384823858738, + 0.003227053675800562, + 0.00440739281475544, + 0.004602503497153521, + 0.0054872846230864525, + 0.0032190780621021986, + 0.005236528813838959, + 0.0023639786522835493, + 0.0032887232955545187, + 0.003684740513563156, + 0.006439190823584795, + 0.003322337521240115, + 0.0038516742642968893, + 0.002960183657705784, + 0.0032746929209679365, + 0.001965311588719487, + 0.0038561655674129725, + 0.0032815313898026943, + 0.0033813973423093557, + 0.003537421813234687, + 0.005835917312651873, + 0.005526143126189709, + 0.00287851644679904, + 0.002818986540660262, + 0.002753189066424966, + 0.0025720992125570774, + 0.002192082814872265, + 0.0022229375317692757, + 0.003919698763638735, + 0.00314239040017128, + 0.001886154874227941, + 0.002482922514900565, + 0.002577175386250019, + 0.002838757121935487, + 0.0018578493036329746, + 0.0019741258583962917, + 0.004480417352169752, + 0.0023121654521673918, + 0.003027273342013359, + 0.0015701717929914594 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 0, + 5 + ], + "selected_reward": -0.008645012974739075 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject the legendary Overlord of Chu wearing dark iron armor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the armor visibly damaged with deep sword marks, dents, curled edges, and broken arrows lodged in it?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the character's long hair tied up, disheveled, and covered in sweat, dust, and fatigue?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the character shown throwing his head back with tensed neck muscles and an open mouth in a roaring pose?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene rendered with a mix of thick, thin, dry, and wet ink brushstrokes in a static medium shot from a slightly upward angle?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.006737255025655031 + ], + "opposite_logprob_tokens": [ + -5.006737232208252 + ], + "logprob_reward": -0.006737255025655031, + "yes_no_margin_reward": 5.0, + "token_credit_reward": -0.006737255025655031, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07378221303224564, + 0.08308428525924683, + 0.26857227087020874, + 0.057630524039268494, + 0.04555324465036392, + 0.025948988273739815, + 0.23391444981098175, + 0.16278567910194397, + 0.03299810364842415, + 0.015730220824480057 + ], + "credit_map": [ + 0.008792867884039879, + 0.00708763487637043, + 0.009598519653081894, + 0.011785047128796577, + 0.008774252608418465, + 0.011126372031867504, + 0.0076799290254712105, + 0.008937593549489975, + 0.009127450175583363, + 0.008722429163753986, + 0.0120795127004385, + 0.00753032835200429, + 0.010944300331175327, + 0.019644789397716522, + 0.006654695607721806, + 0.008380778133869171, + 0.06610609591007233, + 0.012636593542993069, + 0.10815577208995819, + 0.018773730844259262, + 0.01362566463649273, + 0.008582641370594501, + 0.014824310317635536, + 0.025867486372590065, + 0.009578115306794643, + 0.004059928469359875, + 0.003546136897057295, + 0.003078632988035679, + 0.006362528540194035, + 0.017907358705997467, + 0.00480636628344655, + 0.008291460573673248, + 0.004455542657524347, + 0.007837111130356789, + 0.006128937005996704, + 0.004950111266225576, + 0.006468193139880896, + 0.0043865530751645565, + 0.004978495184332132, + 0.0063483016565442085, + 0.003356833476573229, + 0.001974207116290927, + 0.0036601985339075327, + 0.003717456478625536, + 0.004363794811069965, + 0.0025565982796251774, + 0.003944532945752144, + 0.0023753668647259474, + 0.025181515142321587, + 0.024015406146645546, + 0.025032924488186836, + 0.028341123834252357, + 0.04474900662899017, + 0.04660651460289955, + 0.024585634469985962, + 0.01540234126150608, + 0.024011200293898582, + 0.01805492863059044, + 0.016083434224128723, + 0.016814958304166794, + 0.02847268618643284, + 0.032905008643865585, + 0.01519039086997509, + 0.011253067292273045, + 0.004107680171728134, + 0.0040732878260314465, + 0.002862383611500263, + 0.0031830989755690098, + 0.006639714352786541, + 0.004770071245729923, + 0.003376471344381571, + 0.003985397517681122, + 0.0016548994462937117, + 0.0017859431682154536, + 0.002076159231364727, + 0.0026802693028002977, + 0.002112802816554904, + 0.00170048454310745, + 0.0016934523591771722, + 0.002026210306212306 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 6, + 7 + ], + "selected_reward": -0.006737255025655031 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject the legendary Overlord of Chu wearing dark iron armor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the armor visibly damaged with deep sword marks, dents, curled edges, and broken arrows lodged in it?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the character's long hair tied up, disheveled, and covered in sweat, dust, and fatigue?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the character shown throwing his head back with tensed neck muscles and an open mouth in a roaring pose?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene rendered with a mix of thick, thin, dry, and wet ink brushstrokes in a static medium shot from a slightly upward angle?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.026345498859882355 + ], + "opposite_logprob_tokens": [ + -3.6513454914093018 + ], + "logprob_reward": -0.026345498859882355, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.026345498859882355, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11657243221998215, + 0.09917586296796799, + 0.30713382363319397, + 0.08261489123106003, + 0.05725792050361633, + 0.03588971495628357, + 0.15898795425891876, + 0.08062933385372162, + 0.025764835998415947, + 0.03597322106361389 + ], + "credit_map": [ + 0.01226135529577732, + 0.012788771651685238, + 0.01907232403755188, + 0.01851714216172695, + 0.011744670569896698, + 0.016607986763119698, + 0.013105629943311214, + 0.012474567629396915, + 0.011208493262529373, + 0.010244065895676613, + 0.014397888444364071, + 0.012159348465502262, + 0.012361288070678711, + 0.01712769642472267, + 0.011205464601516724, + 0.010471619665622711, + 0.07552619278430939, + 0.014665725640952587, + 0.12026943266391754, + 0.024605510756373405, + 0.015121281147003174, + 0.01053786464035511, + 0.017361344769597054, + 0.02904648520052433, + 0.012993223033845425, + 0.0065852743573486805, + 0.007028678432106972, + 0.009350845590233803, + 0.010016068816184998, + 0.011370853520929813, + 0.010144449770450592, + 0.015125497244298458, + 0.005142985377460718, + 0.009845932945609093, + 0.007539399433881044, + 0.007202924229204655, + 0.00983037706464529, + 0.005863898433744907, + 0.006795245688408613, + 0.005037163849920034, + 0.004178144503384829, + 0.002707471838220954, + 0.004763109143823385, + 0.00367196137085557, + 0.006067352835088968, + 0.006927924230694771, + 0.004286923445761204, + 0.0032868296839296818, + 0.017737707123160362, + 0.018742013722658157, + 0.030493317171931267, + 0.02391858771443367, + 0.018806762993335724, + 0.01516171544790268, + 0.023563368245959282, + 0.010564486496150494, + 0.010580489411950111, + 0.008857167325913906, + 0.012435507029294968, + 0.011167214252054691, + 0.010999307967722416, + 0.009068291634321213, + 0.00893586315214634, + 0.008585496805608273, + 0.004016086459159851, + 0.0029129469767212868, + 0.002686115214601159, + 0.001961710397154093, + 0.0034536337479948997, + 0.005529352463781834, + 0.0019283476285636425, + 0.003276643343269825, + 0.003992191981524229, + 0.0032643028534948826, + 0.0047917114570736885, + 0.004915780853480101, + 0.005821173079311848, + 0.00459506269544363, + 0.0036610066890716553, + 0.0049319942481815815 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 6, + 0 + ], + "selected_reward": -0.026345498859882355 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "static medium shot, shot from a slightly upward angle, focuses squarely on the legendary Overlord of Chu. He stands amidst a desolate battlefield rendered with thick, thin, dry, and wet ink brushstrokes. His dark iron armor, shattered by the bloody battles, is covered in deep sword marks and dents, its edges curled, and several broken arrows remain lodged within. Dark blood and dirt blend, forming a smudge on the metal surface like solidified ink. His long, tied-up hair is disheveled, a few sweat-soaked strands clinging to his stern face, shrouded in dust and fatigue. In that instant, he throws his head back, the muscles in his neck tensing, and his mouth wide open, he lets out a silent yet earth-shattering roar into the oppressive, leaden sky, as if stained with ink. His eyes, wide open, burn with resentment, rage, and the endless sadness of a hero's end. The entire picture presents an extreme tragic aesthetic. The mountains in the background and the fallen weapons are blurred with light ink. Only his tense posture and desperate cry constitute the absolute center of the picture, creating a lonely and magnificent epic atmosphere.", + "seed": 7933, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000034/group_1/index_3118_seed_7933_dp_2_rank_4_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 1, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.015932766161859034, + "vlm_logprob": -0.015932766161859034, + "vlm_yes_no_margin": 4.25, + "vlm_token_credit": -0.015932766161859034, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject the legendary Overlord of Chu wearing dark iron armor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the armor visibly damaged with deep sword marks, dents, curled edges, and broken arrows lodged in it?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the character's long hair tied up, disheveled, and covered in sweat, dust, and fatigue?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the character shown throwing his head back with tensed neck muscles and an open mouth in a roaring pose?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene rendered with a mix of thick, thin, dry, and wet ink brushstrokes in a static medium shot from a slightly upward angle?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.019722314551472664 + ], + "opposite_logprob_tokens": [ + -4.1447224617004395 + ], + "logprob_reward": -0.019722314551472664, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.019722314551472664, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.088592529296875, + 0.4287344515323639, + 0.055910494178533554, + 0.08170139044523239, + 0.05613115057349205, + 0.0583244152367115, + 0.03861396014690399, + 0.03535671532154083, + 0.1097467914223671, + 0.046888090670108795 + ], + "credit_map": [ + 0.008392092771828175, + 0.007920049130916595, + 0.011453780345618725, + 0.015687089413404465, + 0.007680520415306091, + 0.003974637482315302, + 0.026883965358138084, + 0.006600399035960436, + 0.130747452378273, + 0.031059062108397484, + 0.023669013753533363, + 0.06624532490968704, + 0.011782684363424778, + 0.02819085121154785, + 0.09252064675092697, + 0.04451939836144447, + 0.008772502653300762, + 0.004839075263589621, + 0.006145359016954899, + 0.006452651694417, + 0.006280876230448484, + 0.009755541570484638, + 0.008171351626515388, + 0.005493135657161474, + 0.016798118129372597, + 0.0045790779404342175, + 0.005554303992539644, + 0.009803159162402153, + 0.0048227165825665, + 0.008456161245703697, + 0.02215760573744774, + 0.009530245326459408, + 0.006344522815197706, + 0.006438476033508778, + 0.008362332358956337, + 0.0066270469687879086, + 0.0055629657581448555, + 0.005668647587299347, + 0.007839443162083626, + 0.009287715889513493, + 0.008941182866692543, + 0.005039774347096682, + 0.007742978632450104, + 0.012468534521758556, + 0.005612736567854881, + 0.004920082166790962, + 0.007820489816367626, + 0.005778637249022722, + 0.005851455964148045, + 0.004164294805377722, + 0.0046996851451694965, + 0.005532828159630299, + 0.004419715143740177, + 0.00529513880610466, + 0.004359272308647633, + 0.004291567485779524, + 0.004479076247662306, + 0.004997326526790857, + 0.004984493367373943, + 0.0037452972028404474, + 0.0034374978858977556, + 0.0034679167438298464, + 0.0028335365932434797, + 0.0074115716852247715, + 0.012443837709724903, + 0.014005742035806179, + 0.014665317721664906, + 0.016897384077310562, + 0.008542079478502274, + 0.013802734203636646, + 0.016005748882889748, + 0.013383944518864155, + 0.004345199558883905, + 0.007363856304436922, + 0.00948660634458065, + 0.005641555413603783, + 0.004002602305263281, + 0.006529612001031637, + 0.0048313806764781475, + 0.004687277600169182 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 8, + 0 + ], + "selected_reward": -0.019722314551472664 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject the legendary Overlord of Chu wearing dark iron armor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the armor visibly damaged with deep sword marks, dents, curled edges, and broken arrows lodged in it?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the character's long hair tied up, disheveled, and covered in sweat, dust, and fatigue?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the character shown throwing his head back with tensed neck muscles and an open mouth in a roaring pose?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene rendered with a mix of thick, thin, dry, and wet ink brushstrokes in a static medium shot from a slightly upward angle?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02332732267677784 + ], + "opposite_logprob_tokens": [ + -3.773327350616455 + ], + "logprob_reward": -0.02332732267677784, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.02332732267677784, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15796300768852234, + 0.2739487886428833, + 0.13556957244873047, + 0.026504989713430405, + 0.03637252002954483, + 0.03912664204835892, + 0.06553886830806732, + 0.06673355400562286, + 0.1381056010723114, + 0.06013639643788338 + ], + "credit_map": [ + 0.013374547474086285, + 0.010596237145364285, + 0.022755848243832588, + 0.02497945912182331, + 0.009365277364850044, + 0.015808740630745888, + 0.04497659578919411, + 0.016106324270367622, + 0.06804657727479935, + 0.015578689984977245, + 0.013380697928369045, + 0.05478233844041824, + 0.008515577763319016, + 0.01516144722700119, + 0.0789853185415268, + 0.01949816755950451, + 0.01796443946659565, + 0.009671109728515148, + 0.019055206328630447, + 0.009656233713030815, + 0.009641782380640507, + 0.027530590072274208, + 0.01750209741294384, + 0.0245481189340353, + 0.0067293355241417885, + 0.0014486181316897273, + 0.0025762931909412146, + 0.003375016152858734, + 0.001652860431931913, + 0.003089813981205225, + 0.005928018596023321, + 0.001705034519545734, + 0.004662324208766222, + 0.0030343951657414436, + 0.004862915724515915, + 0.003820167388767004, + 0.0029240099247545004, + 0.0037395390681922436, + 0.005196804646402597, + 0.00813236366957426, + 0.004036176484078169, + 0.0028190382290631533, + 0.005469056777656078, + 0.0048028863966465, + 0.0029753991402685642, + 0.004294783342629671, + 0.005047360900789499, + 0.009681942872703075, + 0.007099811919033527, + 0.008334881626069546, + 0.007436895743012428, + 0.005953483283519745, + 0.008685378357768059, + 0.007301632314920425, + 0.007666371297091246, + 0.013060420751571655, + 0.005707732401788235, + 0.009325744584202766, + 0.0077460515312850475, + 0.011794273741543293, + 0.0063211447559297085, + 0.007831789553165436, + 0.007732554338872433, + 0.010274271480739117, + 0.016065916046500206, + 0.013442597351968288, + 0.017826363444328308, + 0.017797430977225304, + 0.011608423665165901, + 0.02132490463554859, + 0.019434349611401558, + 0.02060563489794731, + 0.005492118652909994, + 0.008071744814515114, + 0.010641155764460564, + 0.007919631898403168, + 0.004638963844627142, + 0.009740406647324562, + 0.006480016745626926, + 0.0071523613296449184 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 8 + ], + "selected_reward": -0.02332732267677784 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject the legendary Overlord of Chu wearing dark iron armor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the armor visibly damaged with deep sword marks, dents, curled edges, and broken arrows lodged in it?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the character's long hair tied up, disheveled, and covered in sweat, dust, and fatigue?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the character shown throwing his head back with tensed neck muscles and an open mouth in a roaring pose?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene rendered with a mix of thick, thin, dry, and wet ink brushstrokes in a static medium shot from a slightly upward angle?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.008643713779747486 + ], + "opposite_logprob_tokens": [ + -4.758643627166748 + ], + "logprob_reward": -0.008643713779747486, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.008643713779747486, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.24427007138729095, + 0.3071898818016052, + 0.058693282306194305, + 0.14951151609420776, + 0.02370450086891651, + 0.031912412494421005, + 0.035302773118019104, + 0.03360244259238243, + 0.08880562335252762, + 0.027007456868886948 + ], + "credit_map": [ + 0.030331503599882126, + 0.017136944457888603, + 0.04279528558254242, + 0.038656290620565414, + 0.01667938567698002, + 0.01591416262090206, + 0.06687892973423004, + 0.015877582132816315, + 0.054777659475803375, + 0.015150620602071285, + 0.0147333275526762, + 0.057338353246450424, + 0.012805544771254063, + 0.013245301321148872, + 0.09064223617315292, + 0.04849689081311226, + 0.008249599486589432, + 0.004571610130369663, + 0.010174406692385674, + 0.004832003265619278, + 0.005287479609251022, + 0.00744625274091959, + 0.007374899461865425, + 0.010757040232419968, + 0.03509967401623726, + 0.00655644666403532, + 0.00657364446669817, + 0.02206066995859146, + 0.0061371647752821445, + 0.010763518512248993, + 0.04902150481939316, + 0.013298913836479187, + 0.003659185953438282, + 0.0031665971037000418, + 0.002215819898992777, + 0.0028434889391064644, + 0.0024309013970196247, + 0.002935584168881178, + 0.00367448921315372, + 0.0027784365229308605, + 0.004439682234078646, + 0.0022019268944859505, + 0.005510720424354076, + 0.0039697252213954926, + 0.0031956711318343878, + 0.003138118889182806, + 0.0038058189675211906, + 0.005650750361382961, + 0.003542898688465357, + 0.004109528381377459, + 0.003434827784076333, + 0.006299270316958427, + 0.0030871625058352947, + 0.004759702831506729, + 0.00364306359551847, + 0.006426322739571333, + 0.003999481443315744, + 0.0034131095744669437, + 0.004113710951060057, + 0.004947959445416927, + 0.002383515937253833, + 0.005286744795739651, + 0.005507370922714472, + 0.003950552083551884, + 0.01266212947666645, + 0.00880841538310051, + 0.011449841782450676, + 0.007564053870737553, + 0.008793257176876068, + 0.01922222040593624, + 0.009712791070342064, + 0.01059291698038578, + 0.0025759146083146334, + 0.0038081055972725153, + 0.004210295621305704, + 0.003520748345181346, + 0.001806592452339828, + 0.0032710887026041746, + 0.00443740701302886, + 0.003377307904884219 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 3 + ], + "selected_reward": -0.008643713779747486 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject the legendary Overlord of Chu wearing dark iron armor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the armor visibly damaged with deep sword marks, dents, curled edges, and broken arrows lodged in it?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the character's long hair tied up, disheveled, and covered in sweat, dust, and fatigue?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the character shown throwing his head back with tensed neck muscles and an open mouth in a roaring pose?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene rendered with a mix of thick, thin, dry, and wet ink brushstrokes in a static medium shot from a slightly upward angle?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.018187537789344788 + ], + "opposite_logprob_tokens": [ + -4.018187522888184 + ], + "logprob_reward": -0.018187537789344788, + "yes_no_margin_reward": 4.0, + "token_credit_reward": -0.018187537789344788, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1907983124256134, + 0.34338250756263733, + 0.06074158474802971, + 0.08239350467920303, + 0.04042693227529526, + 0.033706046640872955, + 0.04026080295443535, + 0.033667195588350296, + 0.12498261034488678, + 0.04964049160480499 + ], + "credit_map": [ + 0.021392984315752983, + 0.02771235816180706, + 0.022185342386364937, + 0.02073095366358757, + 0.018918735906481743, + 0.025317292660474777, + 0.03367096185684204, + 0.02086968533694744, + 0.03749573230743408, + 0.011627813801169395, + 0.014024274423718452, + 0.07533679157495499, + 0.01379863079637289, + 0.018164264038205147, + 0.12072929739952087, + 0.052205730229616165, + 0.009998657740652561, + 0.00447783712297678, + 0.003687799908220768, + 0.005253729410469532, + 0.00533830001950264, + 0.008022218011319637, + 0.008587855845689774, + 0.015375189483165741, + 0.014148334041237831, + 0.005049880128353834, + 0.0077195921912789345, + 0.009564056992530823, + 0.0063203731551766396, + 0.013562273234128952, + 0.015942854806780815, + 0.010086139664053917, + 0.0069754053838551044, + 0.002940921811386943, + 0.004805746953934431, + 0.003653953317552805, + 0.002930621150881052, + 0.003722739638760686, + 0.00454539991915226, + 0.010852143168449402, + 0.004881234839558601, + 0.0024302052333950996, + 0.006990108173340559, + 0.004271827172487974, + 0.002968402812257409, + 0.003393087536096573, + 0.003284068778157234, + 0.005487111397087574, + 0.007768724113702774, + 0.005729681346565485, + 0.005680238362401724, + 0.004364883527159691, + 0.004294667858630419, + 0.005193144083023071, + 0.0037373939994722605, + 0.0034920682664960623, + 0.006064708810299635, + 0.004058272577822208, + 0.004770855884999037, + 0.0036953447852283716, + 0.002674211747944355, + 0.0031847672071307898, + 0.003136122366413474, + 0.006082914303988218, + 0.014927063137292862, + 0.01900939643383026, + 0.01338485348969698, + 0.014638547785580158, + 0.014352011494338512, + 0.01897372491657734, + 0.014118839986622334, + 0.015578166581690311, + 0.004651322960853577, + 0.00607477780431509, + 0.009606587700545788, + 0.006862751208245754, + 0.003599848598241806, + 0.007385294418781996, + 0.0065782442688941956, + 0.004881662782281637 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 8 + ], + "selected_reward": -0.018187537789344788 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject the legendary Overlord of Chu wearing dark iron armor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the armor visibly damaged with deep sword marks, dents, curled edges, and broken arrows lodged in it?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the character's long hair tied up, disheveled, and covered in sweat, dust, and fatigue?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the character shown throwing his head back with tensed neck muscles and an open mouth in a roaring pose?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene rendered with a mix of thick, thin, dry, and wet ink brushstrokes in a static medium shot from a slightly upward angle?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0097829420119524 + ], + "opposite_logprob_tokens": [ + -4.634782791137695 + ], + "logprob_reward": -0.0097829420119524, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.0097829420119524, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11242829263210297, + 0.31056320667266846, + 0.06439751386642456, + 0.07047412544488907, + 0.036453716456890106, + 0.046075038611888885, + 0.10724915564060211, + 0.07520568370819092, + 0.1190597414970398, + 0.05809355527162552 + ], + "credit_map": [ + 0.014830083586275578, + 0.012681600637733936, + 0.014063957147300243, + 0.018570169806480408, + 0.008490748703479767, + 0.008933596312999725, + 0.024464529007673264, + 0.010393606498837471, + 0.06766890734434128, + 0.018034188076853752, + 0.014845786616206169, + 0.06468614190816879, + 0.01049773395061493, + 0.017680184915661812, + 0.0999360904097557, + 0.017214162275195122, + 0.007742368616163731, + 0.006972743663936853, + 0.013781378977000713, + 0.0052585965022444725, + 0.008612889796495438, + 0.0073570082895457745, + 0.007093461230397224, + 0.007579062599688768, + 0.016504080966114998, + 0.0035376804880797863, + 0.0061125606298446655, + 0.008748258464038372, + 0.0030895087402313948, + 0.007937305606901646, + 0.017740435898303986, + 0.006804295349866152, + 0.00665303785353899, + 0.0035980604588985443, + 0.004787528421729803, + 0.004063987638801336, + 0.004358835984021425, + 0.0037423595786094666, + 0.003919494803994894, + 0.0053304159082472324, + 0.005886148661375046, + 0.0036768552381545305, + 0.0058455863036215305, + 0.005723976064473391, + 0.00474094832316041, + 0.005377860274165869, + 0.005376679357141256, + 0.009446985088288784, + 0.013276350684463978, + 0.01758771948516369, + 0.011833286844193935, + 0.015652872622013092, + 0.0151417450979352, + 0.01595149375498295, + 0.009099439717829227, + 0.008706250227987766, + 0.010880923829972744, + 0.009293402545154095, + 0.011956599541008472, + 0.01127544790506363, + 0.006651901639997959, + 0.007860852405428886, + 0.00863803643733263, + 0.008648516610264778, + 0.014687392860651016, + 0.016709692776203156, + 0.014238208532333374, + 0.015202092006802559, + 0.013891185633838177, + 0.012366361916065216, + 0.01673659309744835, + 0.0152282090857625, + 0.005019818898290396, + 0.007523836102336645, + 0.010723885148763657, + 0.008548642508685589, + 0.004136855714023113, + 0.006206137593835592, + 0.00956492405384779, + 0.006369448266923428 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 8, + 0 + ], + "selected_reward": -0.0097829420119524 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "static medium shot, shot from a slightly upward angle, focuses squarely on the legendary Overlord of Chu. He stands amidst a desolate battlefield rendered with thick, thin, dry, and wet ink brushstrokes. His dark iron armor, shattered by the bloody battles, is covered in deep sword marks and dents, its edges curled, and several broken arrows remain lodged within. Dark blood and dirt blend, forming a smudge on the metal surface like solidified ink. His long, tied-up hair is disheveled, a few sweat-soaked strands clinging to his stern face, shrouded in dust and fatigue. In that instant, he throws his head back, the muscles in his neck tensing, and his mouth wide open, he lets out a silent yet earth-shattering roar into the oppressive, leaden sky, as if stained with ink. His eyes, wide open, burn with resentment, rage, and the endless sadness of a hero's end. The entire picture presents an extreme tragic aesthetic. The mountains in the background and the fallen weapons are blurred with light ink. Only his tense posture and desperate cry constitute the absolute center of the picture, creating a lonely and magnificent epic atmosphere.", + "seed": 7934, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000034/group_1/index_3118_seed_7934_dp_3_rank_6_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 1, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.011506023723632098, + "vlm_logprob": -0.011506023723632098, + "vlm_yes_no_margin": 4.575, + "vlm_token_credit": -0.011506023723632098, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject the legendary Overlord of Chu wearing dark iron armor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the armor visibly damaged with deep sword marks, dents, curled edges, and broken arrows lodged in it?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the character's long hair tied up, disheveled, and covered in sweat, dust, and fatigue?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the character shown throwing his head back with tensed neck muscles and an open mouth in a roaring pose?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene rendered with a mix of thick, thin, dry, and wet ink brushstrokes in a static medium shot from a slightly upward angle?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.013035556301474571 + ], + "opposite_logprob_tokens": [ + -4.513035774230957 + ], + "logprob_reward": -0.013035556301474571, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.013035556301474571, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.20831693708896637, + 0.09651904553174973, + 0.06119951978325844, + 0.12200332432985306, + 0.0995679572224617, + 0.10850409418344498, + 0.09363352507352829, + 0.029666736721992493, + 0.11255538463592529, + 0.06803354620933533 + ], + "credit_map": [ + 0.022445639595389366, + 0.033187925815582275, + 0.01783297210931778, + 0.023676583543419838, + 0.02566063217818737, + 0.02180914580821991, + 0.04730160906910896, + 0.016402412205934525, + 0.009345310740172863, + 0.009331990964710712, + 0.00794012937694788, + 0.013403650373220444, + 0.02077467553317547, + 0.009406031109392643, + 0.015351657755672932, + 0.010965581983327866, + 0.007672098930925131, + 0.008232330903410912, + 0.005275499075651169, + 0.006494731642305851, + 0.010629997588694096, + 0.008809774182736874, + 0.007933637127280235, + 0.006151439622044563, + 0.01954476162791252, + 0.015113073401153088, + 0.010059193708002567, + 0.011694768443703651, + 0.01626555435359478, + 0.03073308989405632, + 0.010647319257259369, + 0.007945542223751545, + 0.008347693830728531, + 0.009170740842819214, + 0.01354883797466755, + 0.010933643206954002, + 0.020787067711353302, + 0.012618900276720524, + 0.010593650862574577, + 0.013567406684160233, + 0.009899587370455265, + 0.01193247176706791, + 0.008368941955268383, + 0.01534329541027546, + 0.02195560745894909, + 0.009024770930409431, + 0.02088923752307892, + 0.011090174317359924, + 0.010808889754116535, + 0.008060808293521404, + 0.013134798035025597, + 0.009520750492811203, + 0.019276093691587448, + 0.010133612900972366, + 0.012388072907924652, + 0.010310482233762741, + 0.0037253224290907383, + 0.0037453153636306524, + 0.0031018531881272793, + 0.0028922571800649166, + 0.0039034851361066103, + 0.005383247975260019, + 0.0042107682675123215, + 0.002704484621062875, + 0.01652790606021881, + 0.013037188909947872, + 0.015600151382386684, + 0.015610081143677235, + 0.009160790592432022, + 0.018580738455057144, + 0.013630238361656666, + 0.010408269241452217, + 0.011757640168070793, + 0.010892244055867195, + 0.007620085496455431, + 0.0042533716186881065, + 0.003367540892213583, + 0.020601358264684677, + 0.003818671451881528, + 0.005722623784095049 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 8 + ], + "selected_reward": -0.013035556301474571 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject the legendary Overlord of Chu wearing dark iron armor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the armor visibly damaged with deep sword marks, dents, curled edges, and broken arrows lodged in it?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the character's long hair tied up, disheveled, and covered in sweat, dust, and fatigue?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the character shown throwing his head back with tensed neck muscles and an open mouth in a roaring pose?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene rendered with a mix of thick, thin, dry, and wet ink brushstrokes in a static medium shot from a slightly upward angle?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.020601576194167137 + ], + "opposite_logprob_tokens": [ + -3.895601511001587 + ], + "logprob_reward": -0.020601576194167137, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.020601576194167137, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1797928661108017, + 0.09949661791324615, + 0.06461027264595032, + 0.1453315019607544, + 0.08253117650747299, + 0.14617682993412018, + 0.03745958209037781, + 0.015564455650746822, + 0.16295188665390015, + 0.06608472019433975 + ], + "credit_map": [ + 0.031321633607149124, + 0.036013469099998474, + 0.014024518430233002, + 0.018658017739653587, + 0.025087466463446617, + 0.020361442118883133, + 0.021232815459370613, + 0.013093500398099422, + 0.01329563558101654, + 0.015383481048047543, + 0.011851835064589977, + 0.013649621047079563, + 0.011313049122691154, + 0.00930868461728096, + 0.015086030587553978, + 0.009608280844986439, + 0.009193897247314453, + 0.012850022874772549, + 0.0063016521744430065, + 0.007054034620523453, + 0.008058050647377968, + 0.007376071531325579, + 0.007326827850192785, + 0.006449715234339237, + 0.04385034367442131, + 0.011349065229296684, + 0.010794716887176037, + 0.005742182023823261, + 0.00932246819138527, + 0.050826866179704666, + 0.006378314457833767, + 0.007067535072565079, + 0.008960837498307228, + 0.00690629193559289, + 0.007922789081931114, + 0.008026613853871822, + 0.01996818743646145, + 0.00962486956268549, + 0.01119821984320879, + 0.009923371486365795, + 0.014956105500459671, + 0.013845473527908325, + 0.010744861327111721, + 0.013827560469508171, + 0.036344461143016815, + 0.013960487209260464, + 0.02426086738705635, + 0.01823701709508896, + 0.003987358883023262, + 0.0029367089737206697, + 0.006292101461440325, + 0.0032785695511847734, + 0.004675375763326883, + 0.0043256767094135284, + 0.005311352666467428, + 0.00665243761613965, + 0.001886725309304893, + 0.0017456399509683251, + 0.001904633711092174, + 0.002194697503000498, + 0.0016650949837639928, + 0.0018524449551478028, + 0.002571030519902706, + 0.0017441880190744996, + 0.021467866376042366, + 0.01686946675181389, + 0.028373977169394493, + 0.02371722087264061, + 0.01264043990522623, + 0.027742749080061913, + 0.016456902027130127, + 0.015683269128203392, + 0.00840719137340784, + 0.009287348948419094, + 0.00688889529556036, + 0.006348638329654932, + 0.007031290791928768, + 0.014969183132052422, + 0.006562413182109594, + 0.006589755415916443 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 5 + ], + "selected_reward": -0.020601576194167137 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject the legendary Overlord of Chu wearing dark iron armor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the armor visibly damaged with deep sword marks, dents, curled edges, and broken arrows lodged in it?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the character's long hair tied up, disheveled, and covered in sweat, dust, and fatigue?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the character shown throwing his head back with tensed neck muscles and an open mouth in a roaring pose?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene rendered with a mix of thick, thin, dry, and wet ink brushstrokes in a static medium shot from a slightly upward angle?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.007630487438291311 + ], + "opposite_logprob_tokens": [ + -4.882630348205566 + ], + "logprob_reward": -0.007630487438291311, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.007630487438291311, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2929973304271698, + 0.060304440557956696, + 0.15364156663417816, + 0.06678362190723419, + 0.10638996213674545, + 0.041322581470012665, + 0.03659285232424736, + 0.029063614085316658, + 0.08407655358314514, + 0.1288275271654129 + ], + "credit_map": [ + 0.04735790193080902, + 0.04727401211857796, + 0.02679452672600746, + 0.0316210612654686, + 0.032820336520671844, + 0.02973068319261074, + 0.05953618511557579, + 0.01786261424422264, + 0.007184091955423355, + 0.009294785559177399, + 0.0074485428631305695, + 0.0066569093614816666, + 0.010533930733799934, + 0.006579999811947346, + 0.007522909436374903, + 0.005083274096250534, + 0.022691117599606514, + 0.030838603153824806, + 0.013437353074550629, + 0.0167243555188179, + 0.014150850474834442, + 0.025933442637324333, + 0.013673068955540657, + 0.01619277149438858, + 0.011694688349962234, + 0.006061253137886524, + 0.007982033304870129, + 0.006778792943805456, + 0.009037944488227367, + 0.010039063170552254, + 0.010176753625273705, + 0.005013096146285534, + 0.011936292983591557, + 0.006568473763763905, + 0.006191913969814777, + 0.008114101365208626, + 0.016610419377684593, + 0.01742253266274929, + 0.016952745616436005, + 0.02259347401559353, + 0.007963807322084904, + 0.004311378579586744, + 0.004374672193080187, + 0.004180391784757376, + 0.0068766591139137745, + 0.0032704321201890707, + 0.00675955368205905, + 0.0035856864415109158, + 0.0037300381809473038, + 0.0027926566544920206, + 0.004769514314830303, + 0.0030551105737686157, + 0.00937728676944971, + 0.004113730508834124, + 0.005105013493448496, + 0.0036495050881057978, + 0.004181512165814638, + 0.003974476829171181, + 0.002898693084716797, + 0.002785409800708294, + 0.0025360521394759417, + 0.004990336950868368, + 0.0028725583106279373, + 0.004824576433748007, + 0.012668604962527752, + 0.011023176833987236, + 0.006496960297226906, + 0.007129878271371126, + 0.009838459081947803, + 0.01985722780227661, + 0.009538381360471249, + 0.007523861713707447, + 0.013841264881193638, + 0.014938981272280216, + 0.01546674594283104, + 0.01396712101995945, + 0.014186613261699677, + 0.023134402930736542, + 0.014388286508619785, + 0.01890411786735058 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 9 + ], + "selected_reward": -0.007630487438291311 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject the legendary Overlord of Chu wearing dark iron armor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the armor visibly damaged with deep sword marks, dents, curled edges, and broken arrows lodged in it?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the character's long hair tied up, disheveled, and covered in sweat, dust, and fatigue?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the character shown throwing his head back with tensed neck muscles and an open mouth in a roaring pose?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene rendered with a mix of thick, thin, dry, and wet ink brushstrokes in a static medium shot from a slightly upward angle?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.007628476247191429 + ], + "opposite_logprob_tokens": [ + -4.882628440856934 + ], + "logprob_reward": -0.007628476247191429, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.007628476247191429, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.21097147464752197, + 0.050337787717580795, + 0.038692742586135864, + 0.45639804005622864, + 0.02914179489016533, + 0.025701036676764488, + 0.03848293423652649, + 0.04439583048224449, + 0.06509315967559814, + 0.040785111486911774 + ], + "credit_map": [ + 0.021993914619088173, + 0.029323488473892212, + 0.015475488267838955, + 0.04352807626128197, + 0.035819172859191895, + 0.022199437022209167, + 0.02757193334400654, + 0.01505997870117426, + 0.006224269047379494, + 0.010451524518430233, + 0.005464905872941017, + 0.004919188097119331, + 0.006597418338060379, + 0.005852442234754562, + 0.006009971257299185, + 0.004818068817257881, + 0.005178299732506275, + 0.0034780718851834536, + 0.0038033565506339073, + 0.004326778929680586, + 0.006894958671182394, + 0.006605682894587517, + 0.004045927431434393, + 0.0043596685864031315, + 0.1309163123369217, + 0.03597260266542435, + 0.015127286314964294, + 0.02163686975836754, + 0.0275835320353508, + 0.1663927137851715, + 0.04005054756999016, + 0.01871822588145733, + 0.002682999474927783, + 0.0021997883450239897, + 0.002581931184977293, + 0.002965412102639675, + 0.007882734760642052, + 0.004352117422968149, + 0.003152376040816307, + 0.0033244327642023563, + 0.004996133502572775, + 0.0028970662970095873, + 0.003726706374436617, + 0.0022024696227163076, + 0.003623948898166418, + 0.0022204394917935133, + 0.004114174284040928, + 0.0019200985552743077, + 0.004281696397811174, + 0.0027976834680885077, + 0.0050816964358091354, + 0.0032307226210832596, + 0.0057239169254899025, + 0.004496344365179539, + 0.0045375158078968525, + 0.008333357982337475, + 0.005701584741473198, + 0.006407035980373621, + 0.0024480968713760376, + 0.0036234729923307896, + 0.0043715243227779865, + 0.01212400570511818, + 0.006603501737117767, + 0.0031166092958301306, + 0.006658902391791344, + 0.007773886434733868, + 0.005760390777140856, + 0.010222104378044605, + 0.008464920334517956, + 0.011578453704714775, + 0.00958454143255949, + 0.005049955565482378, + 0.006682717241346836, + 0.005897925700992346, + 0.004337075166404247, + 0.0029745290521532297, + 0.00410024868324399, + 0.009031059220433235, + 0.0036683802027255297, + 0.0040931785479187965 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 0, + 8 + ], + "selected_reward": -0.007628476247191429 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject the legendary Overlord of Chu wearing dark iron armor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the armor visibly damaged with deep sword marks, dents, curled edges, and broken arrows lodged in it?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the character's long hair tied up, disheveled, and covered in sweat, dust, and fatigue?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the character shown throwing his head back with tensed neck muscles and an open mouth in a roaring pose?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene rendered with a mix of thick, thin, dry, and wet ink brushstrokes in a static medium shot from a slightly upward angle?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.008634022437036037 + ], + "opposite_logprob_tokens": [ + -4.758634090423584 + ], + "logprob_reward": -0.008634022437036037, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.008634022437036037, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2821125388145447, + 0.07623516768217087, + 0.06037135794758797, + 0.15097124874591827, + 0.07087782770395279, + 0.06344122439622879, + 0.10331309586763382, + 0.056413259357213974, + 0.06992018967866898, + 0.06634403765201569 + ], + "credit_map": [ + 0.03649735450744629, + 0.037211351096630096, + 0.017744233831763268, + 0.03350653126835823, + 0.025139011442661285, + 0.027820751070976257, + 0.07183241099119186, + 0.03236089646816254, + 0.01194928027689457, + 0.012621012516319752, + 0.007056239992380142, + 0.010623241774737835, + 0.008797883987426758, + 0.008282387629151344, + 0.011197664774954319, + 0.005707459058612585, + 0.009525955654680729, + 0.012934704311192036, + 0.0056488774716854095, + 0.00815456360578537, + 0.005999652668833733, + 0.007321788463741541, + 0.005291340872645378, + 0.005494474899023771, + 0.0314667783677578, + 0.029696136713027954, + 0.011876090429723263, + 0.012946860864758492, + 0.01909848488867283, + 0.024590400978922844, + 0.01215327251702547, + 0.009143217466771603, + 0.007841194979846478, + 0.006344383582472801, + 0.007854785770177841, + 0.004685581661760807, + 0.00940648838877678, + 0.012681999243795872, + 0.004916453268378973, + 0.017146944999694824, + 0.010490397922694683, + 0.006666594184935093, + 0.008130905218422413, + 0.004787026904523373, + 0.010566757060587406, + 0.006489829625934362, + 0.011271177791059017, + 0.005038534756749868, + 0.011480959132313728, + 0.007897624745965004, + 0.013308369554579258, + 0.010441368445754051, + 0.01478368416428566, + 0.012850171886384487, + 0.011727427132427692, + 0.02082349732518196, + 0.005119943059980869, + 0.009688138030469418, + 0.0033878448884934187, + 0.004554218612611294, + 0.006019525695592165, + 0.01717592030763626, + 0.006928575690835714, + 0.0035390884149819613, + 0.00982633139938116, + 0.007758580148220062, + 0.009804136119782925, + 0.008066585287451744, + 0.00666647357866168, + 0.01353408582508564, + 0.007621380966156721, + 0.006642618216574192, + 0.006351644638925791, + 0.009409167803823948, + 0.00883337389677763, + 0.007161236368119717, + 0.006311463192105293, + 0.014652534388005733, + 0.006834565661847591, + 0.006790050305426121 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 6 + ], + "selected_reward": -0.008634022437036037 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "static medium shot, shot from a slightly upward angle, focuses squarely on the legendary Overlord of Chu. He stands amidst a desolate battlefield rendered with thick, thin, dry, and wet ink brushstrokes. His dark iron armor, shattered by the bloody battles, is covered in deep sword marks and dents, its edges curled, and several broken arrows remain lodged within. Dark blood and dirt blend, forming a smudge on the metal surface like solidified ink. His long, tied-up hair is disheveled, a few sweat-soaked strands clinging to his stern face, shrouded in dust and fatigue. In that instant, he throws his head back, the muscles in his neck tensing, and his mouth wide open, he lets out a silent yet earth-shattering roar into the oppressive, leaden sky, as if stained with ink. His eyes, wide open, burn with resentment, rage, and the endless sadness of a hero's end. The entire picture presents an extreme tragic aesthetic. The mountains in the background and the fallen weapons are blurred with light ink. Only his tense posture and desperate cry constitute the absolute center of the picture, creating a lonely and magnificent epic atmosphere.", + "seed": 7935, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000034/group_1/index_3118_seed_7935_dp_3_rank_6_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 1, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.017334792576730252, + "vlm_logprob": -0.017334792576730252, + "vlm_yes_no_margin": 4.275, + "vlm_token_credit": -0.017334792576730252, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject the legendary Overlord of Chu wearing dark iron armor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the armor visibly damaged with deep sword marks, dents, curled edges, and broken arrows lodged in it?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the character's long hair tied up, disheveled, and covered in sweat, dust, and fatigue?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the character shown throwing his head back with tensed neck muscles and an open mouth in a roaring pose?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene rendered with a mix of thick, thin, dry, and wet ink brushstrokes in a static medium shot from a slightly upward angle?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03690384328365326 + ], + "opposite_logprob_tokens": [ + -3.4119038581848145 + ], + "logprob_reward": -0.03690384328365326, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.03690384328365326, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10419289767742157, + 0.5115019679069519, + 0.06912858784198761, + 0.03628594055771828, + 0.04220909625291824, + 0.0430247001349926, + 0.048998139798641205, + 0.04576447606086731, + 0.08023351430892944, + 0.018660681322216988 + ], + "credit_map": [ + 0.012725728563964367, + 0.00910094752907753, + 0.014410867355763912, + 0.014676679857075214, + 0.01573517546057701, + 0.007554046809673309, + 0.01301439106464386, + 0.01697506010532379, + 0.08540403097867966, + 0.027360323816537857, + 0.137137308716774, + 0.0472993478178978, + 0.07363822311162949, + 0.02905500866472721, + 0.06151055544614792, + 0.050097137689590454, + 0.008798536844551563, + 0.006869018543511629, + 0.011484588496387005, + 0.009041311219334602, + 0.008948327042162418, + 0.0053796907886862755, + 0.010815070942044258, + 0.0077920518815517426, + 0.004638798534870148, + 0.0023803862277418375, + 0.0032152505591511726, + 0.004950658418238163, + 0.005798779893666506, + 0.0029188732150942087, + 0.005323578603565693, + 0.007059616036713123, + 0.0047143311239778996, + 0.0034303413704037666, + 0.005930259358137846, + 0.003714553778991103, + 0.006746264640241861, + 0.0038782611954957247, + 0.009328782558441162, + 0.004466300830245018, + 0.003368495497852564, + 0.004819528199732304, + 0.004728804808109999, + 0.006290328688919544, + 0.009668080136179924, + 0.003504994558170438, + 0.005225060973316431, + 0.005419405177235603, + 0.00469986442476511, + 0.006897777784615755, + 0.006146269850432873, + 0.005055330693721771, + 0.009932185523211956, + 0.004041896667331457, + 0.006813596468418837, + 0.005411219317466021, + 0.00437316345050931, + 0.005950927268713713, + 0.0053950343281030655, + 0.005136795341968536, + 0.00611281581223011, + 0.0037597068585455418, + 0.010187284089624882, + 0.004848747979849577, + 0.0065422966144979, + 0.008355485275387764, + 0.007845329120755196, + 0.01559455506503582, + 0.013794442638754845, + 0.006335468962788582, + 0.011615389958024025, + 0.010150549933314323, + 0.0017037686193361878, + 0.00175697251688689, + 0.003333062632009387, + 0.0023953807540237904, + 0.0030298717319965363, + 0.0016791498055681586, + 0.0025710836052894592, + 0.002191392006352544 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 8 + ], + "selected_reward": -0.03690384328365326 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject the legendary Overlord of Chu wearing dark iron armor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the armor visibly damaged with deep sword marks, dents, curled edges, and broken arrows lodged in it?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the character's long hair tied up, disheveled, and covered in sweat, dust, and fatigue?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the character shown throwing his head back with tensed neck muscles and an open mouth in a roaring pose?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene rendered with a mix of thick, thin, dry, and wet ink brushstrokes in a static medium shot from a slightly upward angle?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.023322198539972305 + ], + "opposite_logprob_tokens": [ + -3.773322105407715 + ], + "logprob_reward": -0.023322198539972305, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.023322198539972305, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.20204460620880127, + 0.17753346264362335, + 0.11918558925390244, + 0.1183106079697609, + 0.0396956242620945, + 0.07994204759597778, + 0.050684601068496704, + 0.07175689190626144, + 0.07230282574892044, + 0.06854373216629028 + ], + "credit_map": [ + 0.019767429679632187, + 0.018852248787879944, + 0.03221915289759636, + 0.02715330757200718, + 0.035128191113471985, + 0.01705504208803177, + 0.028328420594334602, + 0.023540809750556946, + 0.01653575338423252, + 0.015539990738034248, + 0.04316705837845802, + 0.01422139536589384, + 0.030220922082662582, + 0.01993652619421482, + 0.020236685872077942, + 0.017675142735242844, + 0.016050074249505997, + 0.0084620900452137, + 0.023347022011876106, + 0.014105626381933689, + 0.015811525285243988, + 0.010271222330629826, + 0.01853141002357006, + 0.012606615200638771, + 0.01801048405468464, + 0.006923672743141651, + 0.01089716050773859, + 0.014179623685777187, + 0.015735525637865067, + 0.008558310568332672, + 0.016490869224071503, + 0.027514971792697906, + 0.005531568080186844, + 0.0032868720591068268, + 0.005883835721760988, + 0.004127770196646452, + 0.006039249245077372, + 0.0033103220630437136, + 0.0072557865642011166, + 0.004260222893208265, + 0.007580641657114029, + 0.009889802895486355, + 0.012590617872774601, + 0.011909931898117065, + 0.012911256402730942, + 0.006686350796371698, + 0.010199028067290783, + 0.008174419403076172, + 0.0064733512699604034, + 0.007509871385991573, + 0.009162148460745811, + 0.005679359193891287, + 0.006362009793519974, + 0.0049902405589818954, + 0.006313710007816553, + 0.004193909000605345, + 0.0067274016328155994, + 0.008981349878013134, + 0.009846752509474754, + 0.008761788718402386, + 0.008601359091699123, + 0.00573068717494607, + 0.016091005876660347, + 0.0070165470242500305, + 0.007455909159034491, + 0.012875176034867764, + 0.008133799768984318, + 0.009943358600139618, + 0.010026674717664719, + 0.007061205338686705, + 0.008652125485241413, + 0.008154580369591713, + 0.005492907948791981, + 0.006465788930654526, + 0.006012580823153257, + 0.009600714780390263, + 0.01696202903985977, + 0.0063079968094825745, + 0.009967970661818981, + 0.007733739446848631 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 2 + ], + "selected_reward": -0.023322198539972305 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject the legendary Overlord of Chu wearing dark iron armor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the armor visibly damaged with deep sword marks, dents, curled edges, and broken arrows lodged in it?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the character's long hair tied up, disheveled, and covered in sweat, dust, and fatigue?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the character shown throwing his head back with tensed neck muscles and an open mouth in a roaring pose?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene rendered with a mix of thick, thin, dry, and wet ink brushstrokes in a static medium shot from a slightly upward angle?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.006739504635334015 + ], + "opposite_logprob_tokens": [ + -5.006739616394043 + ], + "logprob_reward": -0.006739504635334015, + "yes_no_margin_reward": 5.0, + "token_credit_reward": -0.006739504635334015, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.24027778208255768, + 0.15093348920345306, + 0.08695805072784424, + 0.039888523519039154, + 0.03877422958612442, + 0.11367806047201157, + 0.07739479839801788, + 0.08598614484071732, + 0.09978079795837402, + 0.06632811576128006 + ], + "credit_map": [ + 0.021951381117105484, + 0.0278872549533844, + 0.03067830204963684, + 0.03470708802342415, + 0.030932549387216568, + 0.01787669211626053, + 0.03447146713733673, + 0.04177306592464447, + 0.020372973755002022, + 0.01857783831655979, + 0.04360225051641464, + 0.015702780336141586, + 0.013560286723077297, + 0.013313616625964642, + 0.01279543712735176, + 0.013008305802941322, + 0.009821669198572636, + 0.006923245266079903, + 0.010548433288931847, + 0.007953369989991188, + 0.012848141603171825, + 0.00959049817174673, + 0.014411330223083496, + 0.01486137229949236, + 0.004359038546681404, + 0.003555362345650792, + 0.005375540349632502, + 0.0053459652699530125, + 0.00670167850330472, + 0.0026100215036422014, + 0.005926920101046562, + 0.006013996433466673, + 0.005924801807850599, + 0.0034007856156677008, + 0.0067747654393315315, + 0.006328001152724028, + 0.0051289708353579044, + 0.0035876056645065546, + 0.0044786068610847, + 0.0031506905797868967, + 0.010569867677986622, + 0.010078842751681805, + 0.017514506354928017, + 0.015353452414274216, + 0.02199670299887657, + 0.009184938855469227, + 0.01498367264866829, + 0.013996080495417118, + 0.009920050390064716, + 0.009839799255132675, + 0.014247531071305275, + 0.007360722403973341, + 0.011957494542002678, + 0.007357825990766287, + 0.010965620167553425, + 0.00574575737118721, + 0.007700435351580381, + 0.006268516182899475, + 0.015035652555525303, + 0.011556954123079777, + 0.011001452803611755, + 0.007635957561433315, + 0.016321435570716858, + 0.010465738363564014, + 0.010448331944644451, + 0.007719252724200487, + 0.024324603378772736, + 0.014372112229466438, + 0.011034829542040825, + 0.010326908901333809, + 0.013682030141353607, + 0.007872719317674637, + 0.006361653562635183, + 0.009149506688117981, + 0.006620833184570074, + 0.009045561775565147, + 0.012979717925190926, + 0.005085644777864218, + 0.010694092139601707, + 0.006391102913767099 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 5 + ], + "selected_reward": -0.006739504635334015 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject the legendary Overlord of Chu wearing dark iron armor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the armor visibly damaged with deep sword marks, dents, curled edges, and broken arrows lodged in it?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the character's long hair tied up, disheveled, and covered in sweat, dust, and fatigue?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the character shown throwing his head back with tensed neck muscles and an open mouth in a roaring pose?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene rendered with a mix of thick, thin, dry, and wet ink brushstrokes in a static medium shot from a slightly upward angle?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.008636504411697388 + ], + "opposite_logprob_tokens": [ + -4.758636474609375 + ], + "logprob_reward": -0.008636504411697388, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.008636504411697388, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12187380343675613, + 0.4640245735645294, + 0.04010632634162903, + 0.02477545663714409, + 0.03112741746008396, + 0.031034015119075775, + 0.07050549238920212, + 0.08346766233444214, + 0.08044090867042542, + 0.05264437198638916 + ], + "credit_map": [ + 0.0125951012596488, + 0.009354513138532639, + 0.014518278650939465, + 0.022078735753893852, + 0.018235407769680023, + 0.008368321694433689, + 0.015741286799311638, + 0.020982148125767708, + 0.09478864073753357, + 0.02716204896569252, + 0.16508223116397858, + 0.03120679035782814, + 0.04125721752643585, + 0.0220718402415514, + 0.04569828137755394, + 0.036757487803697586, + 0.003916075918823481, + 0.004000208340585232, + 0.00599795812740922, + 0.004791458137333393, + 0.004825298674404621, + 0.00292395381256938, + 0.007991350255906582, + 0.005660017486661673, + 0.0039008602034300566, + 0.0018530762754380703, + 0.00288287247531116, + 0.0030539585277438164, + 0.003187373047694564, + 0.0016038373578339815, + 0.002887045731768012, + 0.005406430922448635, + 0.004124658182263374, + 0.004568436183035374, + 0.004110282752662897, + 0.0038519836962223053, + 0.003775272285565734, + 0.0023814444430172443, + 0.005009111948311329, + 0.003306223079562187, + 0.00408719340339303, + 0.0031545283272862434, + 0.003530146088451147, + 0.0035881302319467068, + 0.005293348338454962, + 0.0027249602135270834, + 0.0049033453688025475, + 0.0037523626815527678, + 0.007396929431706667, + 0.00624388363212347, + 0.015431837178766727, + 0.008100045844912529, + 0.00992873776704073, + 0.007744202855974436, + 0.00670563755556941, + 0.008954220451414585, + 0.010928700678050518, + 0.011018775403499603, + 0.017056869342923164, + 0.008041654713451862, + 0.008559495210647583, + 0.009725246578454971, + 0.009126623161137104, + 0.00901029258966446, + 0.006634433288127184, + 0.008587432093918324, + 0.010495488531887531, + 0.015090071596205235, + 0.011287037283182144, + 0.006646386813372374, + 0.010994985699653625, + 0.010705065913498402, + 0.00389575376175344, + 0.0061653172597289085, + 0.006499222479760647, + 0.009886628948152065, + 0.009150314144790173, + 0.004319978877902031, + 0.007623898796737194, + 0.005103251431137323 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 7 + ], + "selected_reward": -0.008636504411697388 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject the legendary Overlord of Chu wearing dark iron armor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the armor visibly damaged with deep sword marks, dents, curled edges, and broken arrows lodged in it?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the character's long hair tied up, disheveled, and covered in sweat, dust, and fatigue?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the character shown throwing his head back with tensed neck muscles and an open mouth in a roaring pose?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene rendered with a mix of thick, thin, dry, and wet ink brushstrokes in a static medium shot from a slightly upward angle?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01107191201299429 + ], + "opposite_logprob_tokens": [ + -4.511071681976318 + ], + "logprob_reward": -0.01107191201299429, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.01107191201299429, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.22129730880260468, + 0.32721391320228577, + 0.05081573501229286, + 0.06107588857412338, + 0.05562535673379898, + 0.06099964678287506, + 0.05113492161035538, + 0.10757696628570557, + 0.03393076732754707, + 0.03032943606376648 + ], + "credit_map": [ + 0.025730786845088005, + 0.0188226830214262, + 0.03120594099164009, + 0.03683748468756676, + 0.039545826613903046, + 0.017219258472323418, + 0.027164457365870476, + 0.024770919233560562, + 0.04318149760365486, + 0.027516430243849754, + 0.04564759507775307, + 0.02874494157731533, + 0.0688764750957489, + 0.029052602127194405, + 0.04934312775731087, + 0.034851282835006714, + 0.007206966169178486, + 0.005008358974009752, + 0.010151531547307968, + 0.004992994945496321, + 0.006489304360002279, + 0.005271104164421558, + 0.005806994158774614, + 0.005888488609343767, + 0.007498781196773052, + 0.0057230121456086636, + 0.008234390057623386, + 0.008600631728768349, + 0.011894271709024906, + 0.004004344809800386, + 0.007288903929293156, + 0.00783156231045723, + 0.006891172379255295, + 0.006240825168788433, + 0.008664236404001713, + 0.00674579618498683, + 0.007066009566187859, + 0.0039736139588057995, + 0.009275420568883419, + 0.006768290419131517, + 0.0076717897318303585, + 0.006494678556919098, + 0.011529156938195229, + 0.006340361665934324, + 0.009344169870018959, + 0.005859055556356907, + 0.007391499821096659, + 0.006368946749716997, + 0.00632709264755249, + 0.00781857781112194, + 0.009432130493223667, + 0.006587144453078508, + 0.006330844480544329, + 0.00525891687721014, + 0.005441672168672085, + 0.003938548266887665, + 0.012926864437758923, + 0.01690554991364479, + 0.016297336667776108, + 0.014255708083510399, + 0.011499190703034401, + 0.010394117794930935, + 0.013394367881119251, + 0.011903843842446804, + 0.0038272489327937365, + 0.004970356356352568, + 0.006235218606889248, + 0.0047345575876533985, + 0.004123263526707888, + 0.0033722720108926296, + 0.0034354436211287975, + 0.0032324136700481176, + 0.0031314550433307886, + 0.004290811251848936, + 0.003087563207373023, + 0.00339006376452744, + 0.0059478264302015305, + 0.0029086547438055277, + 0.004170004744082689, + 0.003403064562007785 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 7 + ], + "selected_reward": -0.01107191201299429 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000034/group_10/index_tensor([7444, 7444])_group_10_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000034/group_10/index_tensor([7444, 7444])_group_10_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..11dfe4226e0ae620989d12b56f380bd49a0c19fc --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000034/group_10/index_tensor([7444, 7444])_group_10_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "From a low angle, the camera slowly and deliberately zooms in on a calm South Asian woman standing on the deck of a yacht. Her thick, dark curls, framed by the bright afternoon sun, create a luminous halo around her silhouette. She wears delicate, round, wire-rimmed glasses, the lenses occasionally catching a glint of sunlight, framing her serene face and soft, contented smile. Her attire is a comfortable, oversized, thick cream knit sweater, its rich texture palpable in the clear light. Steady in her hands, she holds a simple white ceramic cup, her body swaying in gentle, rhythmic movements, perfectly synchronized with the vessel's gentle swaying on the water. As the camera continues its steady inward advance, the background reveals the gleaming, polished teak deck and the vast, sparkling azure sea meeting the clear sky on the horizon. The entire scene is drenched in the warm, golden afternoon light, casting soft, flattering shadows and creating an overwhelming atmosphere of tranquility and blissful solitude.", + "seed": 3960, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000034/group_10/index_7444_seed_3960_dp_20_rank_40_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 10, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.1271683882456273, + "vlm_logprob": -0.1271683882456273, + "vlm_yes_no_margin": 3.65, + "vlm_token_credit": -0.1271683882456273, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a South Asian woman with thick, dark curls?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the woman wearing a thick cream knit sweater and round wire-rimmed glasses?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the woman holding a simple white ceramic cup while standing on a yacht deck?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera start from a low angle and slowly zoom in on the subject?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background visible as a sparkling azure sea meeting a clear sky on the horizon?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.024617338553071022 + ], + "opposite_logprob_tokens": [ + -3.7746174335479736 + ], + "logprob_reward": -0.024617338553071022, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.024617338553071022, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1934247463941574, + 0.07942593097686768, + 0.04384595528244972, + 0.17349232733249664, + 0.07372667640447617, + 0.06947766989469528, + 0.04399411380290985, + 0.07639843970537186, + 0.09400736540555954, + 0.1522066742181778 + ], + "credit_map": [ + 0.02623983658850193, + 0.03145948797464371, + 0.017940621823072433, + 0.025105377659201622, + 0.02066020667552948, + 0.01787838339805603, + 0.027192894369363785, + 0.02694794349372387, + 0.011165129020810127, + 0.010393135249614716, + 0.009024755097925663, + 0.007102368865162134, + 0.009564989246428013, + 0.011843325570225716, + 0.010691001079976559, + 0.00964122824370861, + 0.005495843943208456, + 0.004707193002104759, + 0.006900458130985498, + 0.005445865914225578, + 0.005941887851804495, + 0.006179207935929298, + 0.00453899335116148, + 0.004636501893401146, + 0.036761511117219925, + 0.03479693830013275, + 0.02191000059247017, + 0.016645729541778564, + 0.012709235772490501, + 0.013316212221980095, + 0.02341732569038868, + 0.013935388997197151, + 0.007374727167189121, + 0.007320369593799114, + 0.010112813673913479, + 0.01057655643671751, + 0.011148693971335888, + 0.007933550514280796, + 0.00868152640759945, + 0.010578437708318233, + 0.007688163314014673, + 0.0075102257542312145, + 0.009936511516571045, + 0.015212460421025753, + 0.00835273414850235, + 0.005418117623776197, + 0.0072875851765275, + 0.00807187333703041, + 0.005154760554432869, + 0.005378749221563339, + 0.007455614395439625, + 0.006594166159629822, + 0.0049556042067706585, + 0.00420021079480648, + 0.004883273970335722, + 0.0053717344999313354, + 0.008777271956205368, + 0.009277011267840862, + 0.01078555267304182, + 0.009365024976432323, + 0.010437631979584694, + 0.006545874290168285, + 0.010594476945698261, + 0.010615598410367966, + 0.014217332005500793, + 0.01421414315700531, + 0.015158375725150108, + 0.015533290803432465, + 0.008270521648228168, + 0.008158096112310886, + 0.00983695313334465, + 0.008618655614554882, + 0.02842758223414421, + 0.03251521661877632, + 0.013460875488817692, + 0.014025303535163403, + 0.01826789230108261, + 0.01457783579826355, + 0.01797219179570675, + 0.012959775514900684 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 9 + ], + "selected_reward": -0.024617338553071022 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a South Asian woman with thick, dark curls?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the woman wearing a thick cream knit sweater and round wire-rimmed glasses?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the woman holding a simple white ceramic cup while standing on a yacht deck?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera start from a low angle and slowly zoom in on the subject?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background visible as a sparkling azure sea meeting a clear sky on the horizon?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.008651395328342915 + ], + "opposite_logprob_tokens": [ + -4.758651256561279 + ], + "logprob_reward": -0.008651395328342915, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.008651395328342915, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.28047555685043335, + 0.05490092560648918, + 0.05031243711709976, + 0.13938409090042114, + 0.07586415112018585, + 0.13879217207431793, + 0.12319409102201462, + 0.05262347683310509, + 0.05005275830626488, + 0.03440031409263611 + ], + "credit_map": [ + 0.03517565876245499, + 0.04279523342847824, + 0.02596081607043743, + 0.04986913129687309, + 0.026128917932510376, + 0.025477098301053047, + 0.03903067111968994, + 0.03603802248835564, + 0.005631915759295225, + 0.005878076422959566, + 0.008961301296949387, + 0.007061294745653868, + 0.007249577436596155, + 0.006256968714296818, + 0.008023438975214958, + 0.005838355049490929, + 0.006879155524075031, + 0.0062218704260885715, + 0.005928438156843185, + 0.004180910997092724, + 0.005761243402957916, + 0.009762893430888653, + 0.005005537066608667, + 0.0065723867155611515, + 0.02858765423297882, + 0.027655651792883873, + 0.021682633087038994, + 0.012409196235239506, + 0.010230876505374908, + 0.012964512221515179, + 0.014756700955331326, + 0.011096867732703686, + 0.006804724223911762, + 0.006256754510104656, + 0.011520992033183575, + 0.012693800032138824, + 0.011586043983697891, + 0.009025635197758675, + 0.007639073766767979, + 0.01033712737262249, + 0.01783384568989277, + 0.025027528405189514, + 0.012354487553238869, + 0.012321172282099724, + 0.01914048381149769, + 0.014393270015716553, + 0.014627580530941486, + 0.023093806579709053, + 0.01594230905175209, + 0.01760503649711609, + 0.01354878768324852, + 0.013546721078455448, + 0.014101105742156506, + 0.01020834594964981, + 0.013640503399074078, + 0.0246012844145298, + 0.00669707590714097, + 0.006887953262776136, + 0.005360335577279329, + 0.004327073227614164, + 0.005629983264952898, + 0.0059633636847138405, + 0.007495404686778784, + 0.010262285359203815, + 0.006140290293842554, + 0.0064832414500415325, + 0.009991605766117573, + 0.012721528299152851, + 0.0035866594407707453, + 0.003231093753129244, + 0.0052038319408893585, + 0.0026945110876113176, + 0.0037069395184516907, + 0.004501186311244965, + 0.0038043491076678038, + 0.005428127478808165, + 0.005678886082023382, + 0.0037808462511748075, + 0.00447052763774991, + 0.0030294505413621664 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 5 + ], + "selected_reward": -0.008651395328342915 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a South Asian woman with thick, dark curls?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the woman wearing a thick cream knit sweater and round wire-rimmed glasses?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the woman holding a simple white ceramic cup while standing on a yacht deck?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera start from a low angle and slowly zoom in on the subject?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background visible as a sparkling azure sea meeting a clear sky on the horizon?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.023289820179343224 + ], + "opposite_logprob_tokens": [ + -3.773289918899536 + ], + "logprob_reward": -0.023289820179343224, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.023289820179343224, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3868033289909363, + 0.0788588598370552, + 0.06053714454174042, + 0.12766820192337036, + 0.09705349802970886, + 0.04469454288482666, + 0.07710966467857361, + 0.045426517724990845, + 0.04597393050789833, + 0.03587440773844719 + ], + "credit_map": [ + 0.05256262421607971, + 0.061443913727998734, + 0.03990233317017555, + 0.06589239090681076, + 0.032464269548654556, + 0.030856486409902573, + 0.054485928267240524, + 0.04919537529349327, + 0.007277222815901041, + 0.008230434730648994, + 0.012276783585548401, + 0.011333324015140533, + 0.01119090523570776, + 0.012179655954241753, + 0.00940270908176899, + 0.0069678244180977345, + 0.0072383275255560875, + 0.00651812506839633, + 0.008955041877925396, + 0.007475865539163351, + 0.006947276648133993, + 0.011359592899680138, + 0.005304466933012009, + 0.006738446187227964, + 0.021771907806396484, + 0.022672124207019806, + 0.02463170327246189, + 0.013473952189087868, + 0.00916767492890358, + 0.011647844687104225, + 0.01376380119472742, + 0.01053919643163681, + 0.009747489355504513, + 0.009192433208227158, + 0.018557356670498848, + 0.01974942535161972, + 0.012910652905702591, + 0.007721278350800276, + 0.010892597027122974, + 0.00828226376324892, + 0.005043536424636841, + 0.006549779791384935, + 0.006061137188225985, + 0.00594562292098999, + 0.0036444650031626225, + 0.006718968972563744, + 0.004849573131650686, + 0.005881459452211857, + 0.009561596438288689, + 0.010072989389300346, + 0.017152195796370506, + 0.012283945456147194, + 0.007904660888016224, + 0.006523321382701397, + 0.005542255938053131, + 0.008068694733083248, + 0.006658848375082016, + 0.007133744657039642, + 0.006802320014685392, + 0.008571716956794262, + 0.003838950302451849, + 0.0038260493893176317, + 0.004467679653316736, + 0.004127205815166235, + 0.006483328994363546, + 0.006962264887988567, + 0.004505900200456381, + 0.005751468241214752, + 0.006070373114198446, + 0.0038594179786741734, + 0.005917841102927923, + 0.006423340179026127, + 0.0039055312518030405, + 0.004496661014854908, + 0.0034896486904472113, + 0.0049342927522957325, + 0.0051141162402927876, + 0.004098834469914436, + 0.006042559165507555, + 0.003792764386162162 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 4 + ], + "selected_reward": -0.023289820179343224 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a South Asian woman with thick, dark curls?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the woman wearing a thick cream knit sweater and round wire-rimmed glasses?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the woman holding a simple white ceramic cup while standing on a yacht deck?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera start from a low angle and slowly zoom in on the subject?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background visible as a sparkling azure sea meeting a clear sky on the horizon?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.5760927200317383 + ], + "opposite_logprob_tokens": [ + -0.8260927200317383 + ], + "logprob_reward": -0.5760927200317383, + "yes_no_margin_reward": 0.25, + "token_credit_reward": -0.5760927200317383, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09311775118112564, + 0.07939686626195908, + 0.057461898773908615, + 0.061219651252031326, + 0.040607914328575134, + 0.04301783815026283, + 0.08326097577810287, + 0.08522158116102219, + 0.4512961208820343, + 0.0053993589244782925 + ], + "credit_map": [ + 0.00971435010433197, + 0.010851934552192688, + 0.0076582785695791245, + 0.011597725562751293, + 0.012166335247457027, + 0.007576495874673128, + 0.018367983400821686, + 0.015184647403657436, + 0.007245603483170271, + 0.007575925439596176, + 0.013493428938090801, + 0.012768342159688473, + 0.01226977352052927, + 0.006320060696452856, + 0.011622383259236813, + 0.00810134969651699, + 0.006779336836189032, + 0.006189963314682245, + 0.009384291246533394, + 0.005539126694202423, + 0.004041639622300863, + 0.011117488145828247, + 0.004940434359014034, + 0.009469621814787388, + 0.006385767366737127, + 0.006018049083650112, + 0.011139562353491783, + 0.007533456198871136, + 0.005092316307127476, + 0.008529744111001492, + 0.008084313943982124, + 0.008436438627541065, + 0.004893606994301081, + 0.004298676736652851, + 0.007881412282586098, + 0.0077362521551549435, + 0.004927024710923433, + 0.0038155270740389824, + 0.003676571184769273, + 0.003378842957317829, + 0.004537144210189581, + 0.005570203997194767, + 0.005624132696539164, + 0.0061674583703279495, + 0.004656709264963865, + 0.003978541120886803, + 0.004535586107522249, + 0.007948060519993305, + 0.012455896474421024, + 0.01165270246565342, + 0.02019684575498104, + 0.013253902085125446, + 0.007920888252556324, + 0.004477180540561676, + 0.0054769436828792095, + 0.007826615124940872, + 0.008352149277925491, + 0.009367323480546474, + 0.02009250968694687, + 0.0213198009878397, + 0.006154926959425211, + 0.0063946922309696674, + 0.0069288755767047405, + 0.006611302029341459, + 0.047352150082588196, + 0.052996113896369934, + 0.10086055099964142, + 0.13485164940357208, + 0.02647959627211094, + 0.02626093290746212, + 0.03587030619382858, + 0.026624813675880432, + 0.0005351366708055139, + 0.0006069691735319793, + 0.000855097605381161, + 0.0010914038866758347, + 0.0006104337517172098, + 0.0004650240298360586, + 0.0007867359672673047, + 0.0004485576646402478 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 0, + 7 + ], + "selected_reward": -0.5760927200317383 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a South Asian woman with thick, dark curls?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the woman wearing a thick cream knit sweater and round wire-rimmed glasses?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the woman holding a simple white ceramic cup while standing on a yacht deck?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera start from a low angle and slowly zoom in on the subject?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background visible as a sparkling azure sea meeting a clear sky on the horizon?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.003190667135640979 + ], + "opposite_logprob_tokens": [ + -5.753190517425537 + ], + "logprob_reward": -0.003190667135640979, + "yes_no_margin_reward": 5.75, + "token_credit_reward": -0.003190667135640979, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12238229066133499, + 0.07665720582008362, + 0.04386449605226517, + 0.11782092601060867, + 0.04856564477086067, + 0.030675552785396576, + 0.07221416383981705, + 0.07938867807388306, + 0.32857316732406616, + 0.07985786348581314 + ], + "credit_map": [ + 0.016757899895310402, + 0.018062641844153404, + 0.011867198161780834, + 0.016660647466778755, + 0.01355202216655016, + 0.012090396136045456, + 0.01827310398221016, + 0.015118371695280075, + 0.008127600885927677, + 0.008715654723346233, + 0.01553382072597742, + 0.013770288787782192, + 0.007709437981247902, + 0.007389983627945185, + 0.008325646631419659, + 0.007084776181727648, + 0.005488269962370396, + 0.0051201628521084785, + 0.010139870457351208, + 0.006394661962985992, + 0.002874902915209532, + 0.005724907387048006, + 0.0033775828778743744, + 0.0047441390343010426, + 0.02777508832514286, + 0.028468362987041473, + 0.018143316730856895, + 0.011154570616781712, + 0.010066946968436241, + 0.0082251550629735, + 0.0071585229597985744, + 0.006828955840319395, + 0.005509713664650917, + 0.004972329828888178, + 0.008955053053796291, + 0.008710998110473156, + 0.006196062080562115, + 0.004042185842990875, + 0.004940643906593323, + 0.005238653160631657, + 0.003508978523313999, + 0.004771987441927195, + 0.006181288976222277, + 0.0038420576602220535, + 0.0036607610527426004, + 0.0022554562892764807, + 0.0035034511238336563, + 0.002951571252197027, + 0.008070574142038822, + 0.0070227887481451035, + 0.01289000827819109, + 0.009994562715291977, + 0.00835085567086935, + 0.0063779656775295734, + 0.006857785861939192, + 0.012649625539779663, + 0.008553321473300457, + 0.00886882096529007, + 0.019064849242568016, + 0.020164310932159424, + 0.006427483167499304, + 0.0049585807137191296, + 0.006457156967371702, + 0.00489415368065238, + 0.035811666399240494, + 0.0395725816488266, + 0.07399538904428482, + 0.09776586294174194, + 0.018848558887839317, + 0.018575267866253853, + 0.025331314653158188, + 0.01867252215743065, + 0.015170926228165627, + 0.017483796924352646, + 0.01612505130469799, + 0.012900561094284058, + 0.0041992091573774815, + 0.00384642300195992, + 0.005237618926912546, + 0.004894275683909655 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 0, + 3 + ], + "selected_reward": -0.003190667135640979 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "From a low angle, the camera slowly and deliberately zooms in on a calm South Asian woman standing on the deck of a yacht. Her thick, dark curls, framed by the bright afternoon sun, create a luminous halo around her silhouette. She wears delicate, round, wire-rimmed glasses, the lenses occasionally catching a glint of sunlight, framing her serene face and soft, contented smile. Her attire is a comfortable, oversized, thick cream knit sweater, its rich texture palpable in the clear light. Steady in her hands, she holds a simple white ceramic cup, her body swaying in gentle, rhythmic movements, perfectly synchronized with the vessel's gentle swaying on the water. As the camera continues its steady inward advance, the background reveals the gleaming, polished teak deck and the vast, sparkling azure sea meeting the clear sky on the horizon. The entire scene is drenched in the warm, golden afternoon light, casting soft, flattering shadows and creating an overwhelming atmosphere of tranquility and blissful solitude.", + "seed": 3961, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000034/group_10/index_7444_seed_3961_dp_20_rank_40_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 10, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.015234108688309789, + "vlm_logprob": -0.015234108688309789, + "vlm_yes_no_margin": 4.975, + "vlm_token_credit": -0.015234108688309789, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a South Asian woman with thick, dark curls?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the woman wearing a thick cream knit sweater and round wire-rimmed glasses?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the woman holding a simple white ceramic cup while standing on a yacht deck?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera start from a low angle and slowly zoom in on the subject?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background visible as a sparkling azure sea meeting a clear sky on the horizon?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.008069413714110851 + ], + "opposite_logprob_tokens": [ + -5.0080695152282715 + ], + "logprob_reward": -0.008069413714110851, + "yes_no_margin_reward": 5.0, + "token_credit_reward": -0.008069413714110851, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.18463493883609772, + 0.06183183938264847, + 0.1013275682926178, + 0.08891461789608002, + 0.13941103219985962, + 0.1732185035943985, + 0.10387985408306122, + 0.04796123504638672, + 0.02457069419324398, + 0.0742497369647026 + ], + "credit_map": [ + 0.02423013001680374, + 0.02139982394874096, + 0.031437620520591736, + 0.014871664345264435, + 0.018956143409013748, + 0.02272304892539978, + 0.018837329000234604, + 0.032179173082113266, + 0.009173565544188023, + 0.006475960835814476, + 0.015796039253473282, + 0.005819626618176699, + 0.006915254984050989, + 0.006815420463681221, + 0.005608840379863977, + 0.005227130372077227, + 0.00998940784484148, + 0.013653201051056385, + 0.013265346176922321, + 0.012799457646906376, + 0.016372133046388626, + 0.013062285259366035, + 0.014751020818948746, + 0.00743471784517169, + 0.006718288641422987, + 0.012475170195102692, + 0.007958756759762764, + 0.008288099430501461, + 0.008345071226358414, + 0.01179218664765358, + 0.008480049669742584, + 0.02485699951648712, + 0.01211696770042181, + 0.011939979158341885, + 0.0364462174475193, + 0.01738651469349861, + 0.02109229937195778, + 0.018761880695819855, + 0.010504713281989098, + 0.011162462644279003, + 0.010569524019956589, + 0.00925883837044239, + 0.017764417454600334, + 0.0095209376886487, + 0.051551204174757004, + 0.05042523518204689, + 0.01359646674245596, + 0.010531880892813206, + 0.01263660192489624, + 0.015865391120314598, + 0.01626967266201973, + 0.01467300858348608, + 0.013006908819079399, + 0.013056143186986446, + 0.00840790569782257, + 0.009964227676391602, + 0.004570626653730869, + 0.006808184552937746, + 0.006708996836096048, + 0.005874120630323887, + 0.004154295194894075, + 0.005246308166533709, + 0.00519534619525075, + 0.009403352625668049, + 0.0027162726037204266, + 0.0026768106035888195, + 0.0027230579871684313, + 0.0032891975715756416, + 0.004430878907442093, + 0.003558125114068389, + 0.002849995158612728, + 0.002326355315744877, + 0.009130680933594704, + 0.011204740032553673, + 0.00830128975212574, + 0.010416911914944649, + 0.010561773553490639, + 0.008271820843219757, + 0.008211538195610046, + 0.008150977082550526 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 5, + 4 + ], + "selected_reward": -0.008069413714110851 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a South Asian woman with thick, dark curls?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the woman wearing a thick cream knit sweater and round wire-rimmed glasses?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the woman holding a simple white ceramic cup while standing on a yacht deck?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera start from a low angle and slowly zoom in on the subject?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background visible as a sparkling azure sea meeting a clear sky on the horizon?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0017210922669619322 + ], + "opposite_logprob_tokens": [ + -6.376720905303955 + ], + "logprob_reward": -0.0017210922669619322, + "yes_no_margin_reward": 6.375, + "token_credit_reward": -0.0017210922669619322, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12255863845348358, + 0.11075586080551147, + 0.12402742356061935, + 0.09110783785581589, + 0.10547465085983276, + 0.29353898763656616, + 0.04544826224446297, + 0.05458974838256836, + 0.028245344758033752, + 0.02425321936607361 + ], + "credit_map": [ + 0.015183944255113602, + 0.0113373389467597, + 0.01737833209335804, + 0.013696949928998947, + 0.012745183892548084, + 0.015297832898795605, + 0.011823783628642559, + 0.025095274671912193, + 0.01184605248272419, + 0.012873781844973564, + 0.01904369331896305, + 0.011584664694964886, + 0.018434111028909683, + 0.012736298143863678, + 0.01233881339430809, + 0.011898454278707504, + 0.010588638484477997, + 0.014823481440544128, + 0.018412306904792786, + 0.011007406748831272, + 0.024297986179590225, + 0.01834784634411335, + 0.016505377367138863, + 0.010044386610388756, + 0.007491646334528923, + 0.011991686187684536, + 0.008398360572755337, + 0.008749907836318016, + 0.009351170621812344, + 0.011813237331807613, + 0.008686876855790615, + 0.024624956771731377, + 0.007901085540652275, + 0.009895685128867626, + 0.022808784618973732, + 0.011726802214980125, + 0.01896536350250244, + 0.015580243431031704, + 0.008208585903048515, + 0.010388117283582687, + 0.017027754336595535, + 0.013688264414668083, + 0.038568031042814255, + 0.014659930020570755, + 0.0895892009139061, + 0.0885629951953888, + 0.018249614164233208, + 0.013193230144679546, + 0.004212645348161459, + 0.006091013550758362, + 0.006535886786878109, + 0.004872691351920366, + 0.007649334147572517, + 0.007836301811039448, + 0.0036110070068389177, + 0.0046393838711082935, + 0.004968215245753527, + 0.007493243552744389, + 0.006471868604421616, + 0.006696391850709915, + 0.005429739598184824, + 0.006298299413174391, + 0.006598516367375851, + 0.010633479803800583, + 0.0024730220902711153, + 0.0026805195957422256, + 0.0033522630110383034, + 0.004912390373647213, + 0.005185183137655258, + 0.003611958585679531, + 0.0035212337970733643, + 0.002508776495233178, + 0.0029673839453607798, + 0.0027885164599865675, + 0.003076536813750863, + 0.0024119880981743336, + 0.001874294481240213, + 0.0030515710823237896, + 0.0020711699035018682, + 0.006011759862303734 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 2, + 0 + ], + "selected_reward": -0.0017210922669619322 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a South Asian woman with thick, dark curls?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the woman wearing a thick cream knit sweater and round wire-rimmed glasses?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the woman holding a simple white ceramic cup while standing on a yacht deck?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera start from a low angle and slowly zoom in on the subject?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background visible as a sparkling azure sea meeting a clear sky on the horizon?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.008636504411697388 + ], + "opposite_logprob_tokens": [ + -4.758636474609375 + ], + "logprob_reward": -0.008636504411697388, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.008636504411697388, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.4634394347667694, + 0.062145307660102844, + 0.06789760291576385, + 0.06395580619573593, + 0.08103138953447342, + 0.13440096378326416, + 0.04398152232170105, + 0.03909652680158615, + 0.014701908454298973, + 0.029349513351917267 + ], + "credit_map": [ + 0.0530739352107048, + 0.04806886240839958, + 0.07142119854688644, + 0.031293679028749466, + 0.0427614189684391, + 0.061791449785232544, + 0.04587122052907944, + 0.10915770381689072, + 0.007869323715567589, + 0.006022485438734293, + 0.014708895236253738, + 0.006575948558747768, + 0.007883013226091862, + 0.007856163196265697, + 0.006188369356095791, + 0.005041108466684818, + 0.0061017596162855625, + 0.010784602724015713, + 0.007350603584200144, + 0.009925886988639832, + 0.0090399831533432, + 0.0089501291513443, + 0.00878771673887968, + 0.006956921890377998, + 0.006153791677206755, + 0.007535070646554232, + 0.005421728361397982, + 0.0066396985203027725, + 0.0074255806393921375, + 0.008148274384438992, + 0.00662626838311553, + 0.016005398705601692, + 0.006492074579000473, + 0.008417073637247086, + 0.02351040579378605, + 0.008802233263850212, + 0.011013743467628956, + 0.010142084211111069, + 0.005381233524531126, + 0.007272545248270035, + 0.008351102471351624, + 0.006309180520474911, + 0.01651773229241371, + 0.009748534299433231, + 0.037663381546735764, + 0.037207696586847305, + 0.010961358435451984, + 0.007641971111297607, + 0.0028864967171102762, + 0.0051574320532381535, + 0.0038229282945394516, + 0.005525992717593908, + 0.007861760444939137, + 0.006793048698455095, + 0.005682868417352438, + 0.006250993348658085, + 0.004228889010846615, + 0.006595134269446135, + 0.005679870489984751, + 0.005841727368533611, + 0.0036923657171428204, + 0.003927958197891712, + 0.003930526785552502, + 0.0052000549621880054, + 0.001330415834672749, + 0.0019449712708592415, + 0.002515331842005253, + 0.002091754460707307, + 0.002082117833197117, + 0.0016848546219989657, + 0.0017047937726601958, + 0.0013476688181981444, + 0.003285239217802882, + 0.0034280368126928806, + 0.004382540937513113, + 0.0020430528093129396, + 0.0019056583987548947, + 0.0037354135420173407, + 0.0025819356087595224, + 0.007987636141479015 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 5, + 4 + ], + "selected_reward": -0.008636504411697388 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a South Asian woman with thick, dark curls?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the woman wearing a thick cream knit sweater and round wire-rimmed glasses?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the woman holding a simple white ceramic cup while standing on a yacht deck?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera start from a low angle and slowly zoom in on the subject?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background visible as a sparkling azure sea meeting a clear sky on the horizon?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.054928794503211975 + ], + "opposite_logprob_tokens": [ + -2.929928779602051 + ], + "logprob_reward": -0.054928794503211975, + "yes_no_margin_reward": 2.875, + "token_credit_reward": -0.054928794503211975, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.21890392899513245, + 0.06544572860002518, + 0.10326380282640457, + 0.11655056476593018, + 0.14185591042041779, + 0.24857798218727112, + 0.03488944470882416, + 0.03678752854466438, + 0.029059479013085365, + 0.004665710497647524 + ], + "credit_map": [ + 0.025224514305591583, + 0.021385308355093002, + 0.036447834223508835, + 0.014060823246836662, + 0.022287743166089058, + 0.02995668537914753, + 0.021338745951652527, + 0.04820225015282631, + 0.00992378406226635, + 0.006776141468435526, + 0.011482746340334415, + 0.008669627830386162, + 0.009031292051076889, + 0.006653925403952599, + 0.006826433818787336, + 0.006081768311560154, + 0.010270125232636929, + 0.017255134880542755, + 0.010443508625030518, + 0.012331821955740452, + 0.01735609769821167, + 0.013463595882058144, + 0.01405737642198801, + 0.008086122572422028, + 0.010221200063824654, + 0.010672197677195072, + 0.0155362943187356, + 0.013650881126523018, + 0.011739525012671947, + 0.014428351074457169, + 0.011958698742091656, + 0.028343401849269867, + 0.012232848443090916, + 0.011800139211118221, + 0.02905486896634102, + 0.016783257946372032, + 0.024844517931342125, + 0.021742159500718117, + 0.011639416217803955, + 0.013758673332631588, + 0.015055809170007706, + 0.014371984638273716, + 0.024993931874632835, + 0.0143525879830122, + 0.07351700216531754, + 0.07233934849500656, + 0.020555268973112106, + 0.013391992077231407, + 0.0031302787829190493, + 0.00442767608910799, + 0.0038286447525024414, + 0.0038967779837548733, + 0.00683976523578167, + 0.005970577709376812, + 0.0031331174541264772, + 0.003662600414827466, + 0.0032588120084255934, + 0.006821915972977877, + 0.0037453428376466036, + 0.005161971319466829, + 0.004070141818374395, + 0.004301421344280243, + 0.00517978286370635, + 0.004248136188834906, + 0.0028834743425250053, + 0.0034926976077258587, + 0.0029663504101336002, + 0.00395089341327548, + 0.0042544580064713955, + 0.0038766777142882347, + 0.0030025618616491556, + 0.004632363095879555, + 0.0005580196739174426, + 0.0005139841232448816, + 0.0009525187197141349, + 0.00039752942393533885, + 0.0003890056104864925, + 0.0005150288925506175, + 0.0004563253023661673, + 0.0008832979947328568 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 0, + 4 + ], + "selected_reward": -0.054928794503211975 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a South Asian woman with thick, dark curls?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the woman wearing a thick cream knit sweater and round wire-rimmed glasses?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the woman holding a simple white ceramic cup while standing on a yacht deck?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera start from a low angle and slowly zoom in on the subject?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background visible as a sparkling azure sea meeting a clear sky on the horizon?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0028147385455667973 + ], + "opposite_logprob_tokens": [ + -5.877814769744873 + ], + "logprob_reward": -0.0028147385455667973, + "yes_no_margin_reward": 5.875, + "token_credit_reward": -0.0028147385455667973, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14740216732025146, + 0.15228283405303955, + 0.18121324479579926, + 0.07549482583999634, + 0.09420987963676453, + 0.1662946194410324, + 0.05980295315384865, + 0.057470064610242844, + 0.03869171813130379, + 0.02713763900101185 + ], + "credit_map": [ + 0.019708821550011635, + 0.02073640190064907, + 0.02759009227156639, + 0.015484952367842197, + 0.015250851400196552, + 0.017347566783428192, + 0.01306928787380457, + 0.01821422204375267, + 0.022575683891773224, + 0.010243400000035763, + 0.05202953889966011, + 0.015432837419211864, + 0.010994226671755314, + 0.01912643574178219, + 0.009410637430846691, + 0.012470087967813015, + 0.021773839369416237, + 0.013854117132723331, + 0.053097181022167206, + 0.018599683418869972, + 0.013827516697347164, + 0.022926611825823784, + 0.015930362045764923, + 0.021203933283686638, + 0.005646212957799435, + 0.006931164767593145, + 0.010531756095588207, + 0.01130886934697628, + 0.007710758596658707, + 0.009549086913466454, + 0.006477136630564928, + 0.01733984611928463, + 0.005977050866931677, + 0.0064966678619384766, + 0.01766352728009224, + 0.009123262949287891, + 0.016128940507769585, + 0.016360504552721977, + 0.008896119892597198, + 0.013563811779022217, + 0.009054160676896572, + 0.007177806459367275, + 0.01587665267288685, + 0.006714178249239922, + 0.05310476943850517, + 0.052701838314533234, + 0.011724133044481277, + 0.009941084310412407, + 0.00474436953663826, + 0.0058360775001347065, + 0.0049591162241995335, + 0.006363471504300833, + 0.011277302168309689, + 0.009049137122929096, + 0.0066888537257909775, + 0.010884631425142288, + 0.0059778946451842785, + 0.006554558407515287, + 0.008063445799052715, + 0.008420931175351143, + 0.005909039173275232, + 0.006210072431713343, + 0.009005306288599968, + 0.007328816223889589, + 0.00475477147847414, + 0.004437864758074284, + 0.0046155499294400215, + 0.003934404347091913, + 0.004018849693238735, + 0.005048986990004778, + 0.004630631301552057, + 0.00725066103041172, + 0.0032956551294773817, + 0.002191549865528941, + 0.0053048161789774895, + 0.0024907970800995827, + 0.0019013907294720411, + 0.00323776644654572, + 0.00247737905010581, + 0.006238284986466169 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 5, + 1 + ], + "selected_reward": -0.0028147385455667973 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "From a low angle, the camera slowly and deliberately zooms in on a calm South Asian woman standing on the deck of a yacht. Her thick, dark curls, framed by the bright afternoon sun, create a luminous halo around her silhouette. She wears delicate, round, wire-rimmed glasses, the lenses occasionally catching a glint of sunlight, framing her serene face and soft, contented smile. Her attire is a comfortable, oversized, thick cream knit sweater, its rich texture palpable in the clear light. Steady in her hands, she holds a simple white ceramic cup, her body swaying in gentle, rhythmic movements, perfectly synchronized with the vessel's gentle swaying on the water. As the camera continues its steady inward advance, the background reveals the gleaming, polished teak deck and the vast, sparkling azure sea meeting the clear sky on the horizon. The entire scene is drenched in the warm, golden afternoon light, casting soft, flattering shadows and creating an overwhelming atmosphere of tranquility and blissful solitude.", + "seed": 3962, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000034/group_10/index_7444_seed_3962_dp_21_rank_42_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 10, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.10472541060298682, + "vlm_logprob": -0.10472541060298682, + "vlm_yes_no_margin": 3.7, + "vlm_token_credit": -0.10472541060298682, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a South Asian woman with thick, dark curls?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the woman wearing a thick cream knit sweater and round wire-rimmed glasses?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the woman holding a simple white ceramic cup while standing on a yacht deck?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera start from a low angle and slowly zoom in on the subject?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background visible as a sparkling azure sea meeting a clear sky on the horizon?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01757793501019478 + ], + "opposite_logprob_tokens": [ + -4.142578125 + ], + "logprob_reward": -0.01757793501019478, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.01757793501019478, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14976079761981964, + 0.0575096495449543, + 0.10623347759246826, + 0.07879190146923065, + 0.26969829201698303, + 0.16607829928398132, + 0.03567756339907646, + 0.01550903171300888, + 0.0362374521791935, + 0.08450345695018768 + ], + "credit_map": [ + 0.017214972525835037, + 0.025367051362991333, + 0.014344406314194202, + 0.01703672669827938, + 0.014714371412992477, + 0.02281421422958374, + 0.01654834672808647, + 0.021720703691244125, + 0.00396976713091135, + 0.010111038573086262, + 0.009754914790391922, + 0.008619170635938644, + 0.0050163897685706615, + 0.008070546202361584, + 0.007021963596343994, + 0.004945866297930479, + 0.011360522359609604, + 0.014415218494832516, + 0.02333536557853222, + 0.016230983659625053, + 0.009538790211081505, + 0.014588404446840286, + 0.00941908173263073, + 0.007345127407461405, + 0.00852455385029316, + 0.00912965927273035, + 0.012798383831977844, + 0.013893427327275276, + 0.010834192857146263, + 0.012575882486999035, + 0.007073197979480028, + 0.00396260479465127, + 0.041467055678367615, + 0.03988435119390488, + 0.034192219376564026, + 0.0261188093572855, + 0.03831041604280472, + 0.03804881125688553, + 0.032163310796022415, + 0.019513342529535294, + 0.031691212207078934, + 0.020076071843504906, + 0.016880469396710396, + 0.021343840286135674, + 0.016672471538186073, + 0.027901707217097282, + 0.018479950726032257, + 0.013032585382461548, + 0.0028795944526791573, + 0.004600147716701031, + 0.004196034278720617, + 0.006742901634424925, + 0.004250575788319111, + 0.0051509058102965355, + 0.005040689371526241, + 0.0028167173732072115, + 0.002268407726660371, + 0.0015378588577732444, + 0.002104269340634346, + 0.0015026513719931245, + 0.0014406094560399652, + 0.002590272342786193, + 0.0018265219405293465, + 0.002238442189991474, + 0.0037565636448562145, + 0.0038097864016890526, + 0.004633266478776932, + 0.004137963987886906, + 0.007050759624689817, + 0.005996951833367348, + 0.003572573186829686, + 0.003279588185250759, + 0.00660918140783906, + 0.008149427361786366, + 0.010064736008644104, + 0.007133199367672205, + 0.02202208898961544, + 0.02076304703950882, + 0.005279418546706438, + 0.004482367541640997 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 5, + 0 + ], + "selected_reward": -0.01757793501019478 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a South Asian woman with thick, dark curls?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the woman wearing a thick cream knit sweater and round wire-rimmed glasses?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the woman holding a simple white ceramic cup while standing on a yacht deck?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera start from a low angle and slowly zoom in on the subject?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background visible as a sparkling azure sea meeting a clear sky on the horizon?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.009796871803700924 + ], + "opposite_logprob_tokens": [ + -4.634797096252441 + ], + "logprob_reward": -0.009796871803700924, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.009796871803700924, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.18635894358158112, + 0.09590136259794235, + 0.155761256814003, + 0.07862505316734314, + 0.04924878478050232, + 0.08338350057601929, + 0.02442893572151661, + 0.03495602682232857, + 0.017364501953125, + 0.2739715576171875 + ], + "credit_map": [ + 0.019389452412724495, + 0.022859247401356697, + 0.02271266095340252, + 0.022771762683987617, + 0.02267872355878353, + 0.0281594879925251, + 0.024689411744475365, + 0.023098191246390343, + 0.007399238646030426, + 0.012447278946638107, + 0.0164654478430748, + 0.013293915428221226, + 0.009537415578961372, + 0.01388563122600317, + 0.014007806777954102, + 0.008864632807672024, + 0.013645977713167667, + 0.019992182031273842, + 0.028177933767437935, + 0.03410743549466133, + 0.013078533113002777, + 0.020107751712203026, + 0.015979869291186333, + 0.010671565309166908, + 0.007493041455745697, + 0.011376045644283295, + 0.013053991831839085, + 0.012392503209412098, + 0.008952002041041851, + 0.010686833411455154, + 0.008917906321585178, + 0.005752720404416323, + 0.007162925321608782, + 0.00853315182030201, + 0.00549172842875123, + 0.007084573619067669, + 0.0053876363672316074, + 0.006472258362919092, + 0.005300282500684261, + 0.0038162274286150932, + 0.012312556616961956, + 0.01028086431324482, + 0.00870150700211525, + 0.012144419364631176, + 0.007328248117119074, + 0.013036909513175488, + 0.010690420866012573, + 0.008888580836355686, + 0.0025931550189852715, + 0.0021625820081681013, + 0.004731588996946812, + 0.0025982502847909927, + 0.00237386510707438, + 0.00366782839410007, + 0.0026596083771437407, + 0.003642057301476598, + 0.004201511852443218, + 0.0037207796704024076, + 0.003737524850293994, + 0.004023618996143341, + 0.0058228373527526855, + 0.0051320320926606655, + 0.004297923296689987, + 0.0040197973139584064, + 0.0012621574569493532, + 0.0012575614964589477, + 0.002415363909676671, + 0.0026209200732409954, + 0.0022004498168826103, + 0.0023103905841708183, + 0.003208576003089547, + 0.0020890836603939533, + 0.0037970514968037605, + 0.003774958197027445, + 0.006187416613101959, + 0.00464866915717721, + 0.12650254368782043, + 0.12290377169847488, + 0.003389436984434724, + 0.002767715370282531 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 0, + 2 + ], + "selected_reward": -0.009796871803700924 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a South Asian woman with thick, dark curls?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the woman wearing a thick cream knit sweater and round wire-rimmed glasses?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the woman holding a simple white ceramic cup while standing on a yacht deck?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera start from a low angle and slowly zoom in on the subject?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background visible as a sparkling azure sea meeting a clear sky on the horizon?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.016072474420070648 + ], + "opposite_logprob_tokens": [ + -4.1410722732543945 + ], + "logprob_reward": -0.016072474420070648, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.016072474420070648, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0904281958937645, + 0.039250101894140244, + 0.04719999432563782, + 0.06020902842283249, + 0.11330781877040863, + 0.03000197559595108, + 0.04267346113920212, + 0.017372656613588333, + 0.02348179928958416, + 0.5360749363899231 + ], + "credit_map": [ + 0.006989300716668367, + 0.008968932554125786, + 0.01224262174218893, + 0.010326536372303963, + 0.008144146762788296, + 0.013424024917185307, + 0.013795378617942333, + 0.0165372546762228, + 0.0030696506146341562, + 0.0041406527161598206, + 0.005045550409704447, + 0.004367177374660969, + 0.0029641524888575077, + 0.005473356228321791, + 0.0092041976749897, + 0.004985360894352198, + 0.004476431757211685, + 0.00482236547395587, + 0.007985996082425117, + 0.010182882659137249, + 0.004134852439165115, + 0.006088899914175272, + 0.005266434978693724, + 0.0042421286925673485, + 0.005445417482405901, + 0.003980930428951979, + 0.006853312253952026, + 0.010580957867205143, + 0.00739419087767601, + 0.009526379406452179, + 0.009364352561533451, + 0.007063484285026789, + 0.008699841797351837, + 0.014462989754974842, + 0.014060135930776596, + 0.01743091270327568, + 0.013388105668127537, + 0.014473206363618374, + 0.014381618238985538, + 0.016411010175943375, + 0.006282446905970573, + 0.0034676359500736, + 0.0030160569585859776, + 0.0033664361108094454, + 0.0032340025063604116, + 0.004740953911095858, + 0.0032931820023804903, + 0.00260126031935215, + 0.005076519213616848, + 0.004197430331259966, + 0.009596049785614014, + 0.003855941817164421, + 0.003044306766241789, + 0.006303703878074884, + 0.00391285540536046, + 0.006686656270176172, + 0.0027819168753921986, + 0.0013087689876556396, + 0.0022971071302890778, + 0.0016770666697993875, + 0.001643729628995061, + 0.002847658935934305, + 0.0016552259912714362, + 0.003161181230098009, + 0.0018497691489756107, + 0.00143155618570745, + 0.0022964882664382458, + 0.003029748797416687, + 0.0028716102242469788, + 0.0030727030243724585, + 0.005210100673139095, + 0.003719824831932783, + 0.005499117076396942, + 0.004558633081614971, + 0.005075403489172459, + 0.0042054797522723675, + 0.25613337755203247, + 0.250031977891922, + 0.006039452273398638, + 0.004531531129032373 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 4, + 0 + ], + "selected_reward": -0.016072474420070648 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a South Asian woman with thick, dark curls?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the woman wearing a thick cream knit sweater and round wire-rimmed glasses?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the woman holding a simple white ceramic cup while standing on a yacht deck?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera start from a low angle and slowly zoom in on the subject?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background visible as a sparkling azure sea meeting a clear sky on the horizon?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.4742329716682434 + ], + "opposite_logprob_tokens": [ + -0.9742329716682434 + ], + "logprob_reward": -0.4742329716682434, + "yes_no_margin_reward": 0.5, + "token_credit_reward": -0.4742329716682434, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.22149088978767395, + 0.06571998447179794, + 0.06619935482740402, + 0.07542155683040619, + 0.07864632457494736, + 0.1088949516415596, + 0.03525947034358978, + 0.011072001419961452, + 0.03316213935613632, + 0.30413326621055603 + ], + "credit_map": [ + 0.023175902664661407, + 0.03391226381063461, + 0.022628795355558395, + 0.025999058037996292, + 0.02374013513326645, + 0.03342299163341522, + 0.0238370168954134, + 0.03477473929524422, + 0.004401358775794506, + 0.015949247404932976, + 0.008846077136695385, + 0.009522735141217709, + 0.005305238999426365, + 0.00942061934620142, + 0.00658523291349411, + 0.005689467303454876, + 0.004767504520714283, + 0.007330296561121941, + 0.011894444935023785, + 0.0159156396985054, + 0.0065007335506379604, + 0.008439908735454082, + 0.0069601573050022125, + 0.004390673246234655, + 0.00907242950052023, + 0.011866310611367226, + 0.009946870617568493, + 0.009524158202111721, + 0.009346471168100834, + 0.00966495368629694, + 0.010236999951303005, + 0.005763362627476454, + 0.00843116082251072, + 0.011988951824605465, + 0.010329867713153362, + 0.012480213306844234, + 0.010404546745121479, + 0.011874862015247345, + 0.007817688398063183, + 0.005319036543369293, + 0.01383691094815731, + 0.008782326243817806, + 0.014619099907577038, + 0.01196834072470665, + 0.010455882176756859, + 0.01703222282230854, + 0.015059197321534157, + 0.017140967771410942, + 0.0031620922964066267, + 0.004507952835410833, + 0.0061958469450473785, + 0.005035801790654659, + 0.0043055033311247826, + 0.005628453101962805, + 0.004142242018133402, + 0.00228157639503479, + 0.001605680212378502, + 0.001107179676182568, + 0.001051395433023572, + 0.0016009114915505052, + 0.0011141854338347912, + 0.0019316076068207622, + 0.0014438338112086058, + 0.001217207289300859, + 0.0023688499350100756, + 0.003723792266100645, + 0.004869104363024235, + 0.004745045211166143, + 0.003981810994446278, + 0.004287004936486483, + 0.005880631972104311, + 0.0033059027045965195, + 0.004933133255690336, + 0.0044647883623838425, + 0.00801826361566782, + 0.006298049818724394, + 0.13488195836544037, + 0.13201338052749634, + 0.0077551910653710365, + 0.005768492817878723 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 0, + 5 + ], + "selected_reward": -0.4742329716682434 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a South Asian woman with thick, dark curls?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the woman wearing a thick cream knit sweater and round wire-rimmed glasses?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the woman holding a simple white ceramic cup while standing on a yacht deck?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera start from a low angle and slowly zoom in on the subject?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background visible as a sparkling azure sea meeting a clear sky on the horizon?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005946800112724304 + ], + "opposite_logprob_tokens": [ + -5.130946636199951 + ], + "logprob_reward": -0.005946800112724304, + "yes_no_margin_reward": 5.125, + "token_credit_reward": -0.005946800112724304, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.03400536626577377, + 0.014028702862560749, + 0.014490569941699505, + 0.009408287703990936, + 0.008556104265153408, + 0.04037713259458542, + 0.007403462193906307, + 0.012159342877566814, + 0.009320960380136967, + 0.8502500057220459 + ], + "credit_map": [ + 0.004257339984178543, + 0.004739552736282349, + 0.005786463152617216, + 0.005564047023653984, + 0.00317289587110281, + 0.005002239253371954, + 0.0029755912255495787, + 0.002507241442799568, + 0.0015147149097174406, + 0.0025465821381658316, + 0.001816612551920116, + 0.0017281339969485998, + 0.0014179194113239646, + 0.0019006581278517842, + 0.0015396414091810584, + 0.0015644405502825975, + 0.0014203746104612947, + 0.001702473615296185, + 0.0028672905173152685, + 0.0017983188154175878, + 0.001613569911569357, + 0.002078401390463114, + 0.0015306754503399134, + 0.0014794658636674285, + 0.0009949980303645134, + 0.0016720409039407969, + 0.0013153301551938057, + 0.0012900553410872817, + 0.0011458644876256585, + 0.0011679161107167602, + 0.0011286841472610831, + 0.0006933993427082896, + 0.0008815046167001128, + 0.0008773385779932141, + 0.0016384365735575557, + 0.0011764182709157467, + 0.0009519431041553617, + 0.001235474250279367, + 0.0008303977083414793, + 0.0009645915706641972, + 0.0035680211149156094, + 0.005027356091886759, + 0.004942112136632204, + 0.004529594909399748, + 0.003105840180069208, + 0.006305341143161058, + 0.005018102936446667, + 0.00788076687604189, + 0.0005802710074931383, + 0.0012205278035253286, + 0.000767555960919708, + 0.0010464375372976065, + 0.001160118612460792, + 0.0011538974940776825, + 0.0008802387164905667, + 0.0005944158183410764, + 0.0013756435364484787, + 0.0013193971244618297, + 0.0014716697623953223, + 0.0016735054086893797, + 0.0018623971845954657, + 0.0018317074282094836, + 0.0016690550837665796, + 0.0009559687459841371, + 0.0006596966413781047, + 0.0008566698525100946, + 0.00092117435997352, + 0.001317685004323721, + 0.0016596176428720355, + 0.001445266418159008, + 0.0014783473452553153, + 0.0009825039887800813, + 0.008500930853188038, + 0.006643946748226881, + 0.005464055109769106, + 0.006361524574458599, + 0.40673649311065674, + 0.39787155389785767, + 0.010749157518148422, + 0.007922410033643246 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 5, + 0 + ], + "selected_reward": -0.005946800112724304 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "From a low angle, the camera slowly and deliberately zooms in on a calm South Asian woman standing on the deck of a yacht. Her thick, dark curls, framed by the bright afternoon sun, create a luminous halo around her silhouette. She wears delicate, round, wire-rimmed glasses, the lenses occasionally catching a glint of sunlight, framing her serene face and soft, contented smile. Her attire is a comfortable, oversized, thick cream knit sweater, its rich texture palpable in the clear light. Steady in her hands, she holds a simple white ceramic cup, her body swaying in gentle, rhythmic movements, perfectly synchronized with the vessel's gentle swaying on the water. As the camera continues its steady inward advance, the background reveals the gleaming, polished teak deck and the vast, sparkling azure sea meeting the clear sky on the horizon. The entire scene is drenched in the warm, golden afternoon light, casting soft, flattering shadows and creating an overwhelming atmosphere of tranquility and blissful solitude.", + "seed": 3963, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000034/group_10/index_7444_seed_3963_dp_21_rank_42_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 10, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.09631019541993738, + "vlm_logprob": -0.09631019541993738, + "vlm_yes_no_margin": 3.75, + "vlm_token_credit": -0.09631019541993738, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a South Asian woman with thick, dark curls?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the woman wearing a thick cream knit sweater and round wire-rimmed glasses?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the woman holding a simple white ceramic cup while standing on a yacht deck?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera start from a low angle and slowly zoom in on the subject?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background visible as a sparkling azure sea meeting a clear sky on the horizon?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.015524257905781269 + ], + "opposite_logprob_tokens": [ + -4.265524387359619 + ], + "logprob_reward": -0.015524257905781269, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.015524257905781269, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12704287469387054, + 0.12015800178050995, + 0.07634148746728897, + 0.037643030285835266, + 0.3579423427581787, + 0.13222146034240723, + 0.04779619351029396, + 0.03445214033126831, + 0.041110891848802567, + 0.025291558355093002 + ], + "credit_map": [ + 0.017956018447875977, + 0.01727209985256195, + 0.01507331058382988, + 0.009635856375098228, + 0.02156139351427555, + 0.014381307177245617, + 0.021533681079745293, + 0.009629211388528347, + 0.0135171664878726, + 0.013292745687067509, + 0.017323745414614677, + 0.012881312519311905, + 0.014405687339603901, + 0.015677692368626595, + 0.014260142110288143, + 0.018799500539898872, + 0.0066595617681741714, + 0.010468282736837864, + 0.009552822448313236, + 0.016624385491013527, + 0.006172168534249067, + 0.011251868680119514, + 0.010595457628369331, + 0.005016939248889685, + 0.0033019862603396177, + 0.004302095156162977, + 0.004477575421333313, + 0.0077563016675412655, + 0.004133936017751694, + 0.0060911523178219795, + 0.003938355948776007, + 0.003641625866293907, + 0.029461026191711426, + 0.05252734571695328, + 0.06533832103013992, + 0.034873154014348984, + 0.041297297924757004, + 0.049480099231004715, + 0.03160370513796806, + 0.05336138233542442, + 0.01472799014300108, + 0.017540793865919113, + 0.024395687505602837, + 0.012405875138938427, + 0.01671932078897953, + 0.01749519631266594, + 0.0128574064001441, + 0.016079191118478775, + 0.004784394055604935, + 0.007614525966346264, + 0.005309517495334148, + 0.005017549265176058, + 0.003798458958044648, + 0.006446633022278547, + 0.006559010129421949, + 0.008266100659966469, + 0.0032873698510229588, + 0.007179007399827242, + 0.0042442684061825275, + 0.0032633361406624317, + 0.004645396023988724, + 0.004560495726764202, + 0.0037885340861976147, + 0.003483732230961323, + 0.003708102274686098, + 0.008651637472212315, + 0.004602568689733744, + 0.006891947239637375, + 0.0035889113787561655, + 0.005708960350602865, + 0.004530922044068575, + 0.003427841467782855, + 0.0021177444141358137, + 0.003602026728913188, + 0.002384827472269535, + 0.004831043072044849, + 0.0018525230698287487, + 0.003753630444407463, + 0.0034868773072957993, + 0.003262884682044387 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 5, + 0 + ], + "selected_reward": -0.015524257905781269 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a South Asian woman with thick, dark curls?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the woman wearing a thick cream knit sweater and round wire-rimmed glasses?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the woman holding a simple white ceramic cup while standing on a yacht deck?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera start from a low angle and slowly zoom in on the subject?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background visible as a sparkling azure sea meeting a clear sky on the horizon?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.00979274045675993 + ], + "opposite_logprob_tokens": [ + -4.634792804718018 + ], + "logprob_reward": -0.00979274045675993, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.00979274045675993, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.24929817020893097, + 0.12409478425979614, + 0.14311827719211578, + 0.05889374017715454, + 0.06360365450382233, + 0.0635698139667511, + 0.08252976089715958, + 0.08181274682283401, + 0.04619944840669632, + 0.08687961101531982 + ], + "credit_map": [ + 0.02757909521460533, + 0.03453688323497772, + 0.029151374474167824, + 0.030593832954764366, + 0.029145624488592148, + 0.026339774951338768, + 0.04817534238100052, + 0.023776236921548843, + 0.015462595038115978, + 0.010806129314005375, + 0.02007106877863407, + 0.014584948308765888, + 0.010527478531002998, + 0.01570994220674038, + 0.015989180654287338, + 0.020943446084856987, + 0.011724812909960747, + 0.021249093115329742, + 0.019321369007229805, + 0.029913708567619324, + 0.008814715780317783, + 0.01978147402405739, + 0.022612130269408226, + 0.009700966067612171, + 0.00663344282656908, + 0.006281848065555096, + 0.006382593885064125, + 0.008074639365077019, + 0.007061829324811697, + 0.008222059346735477, + 0.007413371466100216, + 0.008823956362903118, + 0.005641666706651449, + 0.007755167782306671, + 0.007407423108816147, + 0.011538938619196415, + 0.006995865143835545, + 0.008417806588113308, + 0.008928420022130013, + 0.00691836467012763, + 0.007169164251536131, + 0.004779761657118797, + 0.013172438368201256, + 0.005741755478084087, + 0.007907510735094547, + 0.007931227795779705, + 0.006443386431783438, + 0.010424572974443436, + 0.007342634256929159, + 0.013079789467155933, + 0.008532331325113773, + 0.0074988435953855515, + 0.00777553953230381, + 0.010973777621984482, + 0.014205067418515682, + 0.013121782802045345, + 0.00578734278678894, + 0.013413319364190102, + 0.007716277148574591, + 0.017310855910182, + 0.00852765329182148, + 0.011511486954987049, + 0.013053502887487411, + 0.004492313135415316, + 0.004195561632514, + 0.00869406946003437, + 0.006937223486602306, + 0.00680144177749753, + 0.0037056959699839354, + 0.006222591735422611, + 0.006141164805740118, + 0.003501699073240161, + 0.006659518927335739, + 0.014004754833877087, + 0.007472336292266846, + 0.012354256585240364, + 0.00919321272522211, + 0.012029671110212803, + 0.009524833410978317, + 0.015641028061509132 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 1 + ], + "selected_reward": -0.00979274045675993 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a South Asian woman with thick, dark curls?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the woman wearing a thick cream knit sweater and round wire-rimmed glasses?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the woman holding a simple white ceramic cup while standing on a yacht deck?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera start from a low angle and slowly zoom in on the subject?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background visible as a sparkling azure sea meeting a clear sky on the horizon?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.023287491872906685 + ], + "opposite_logprob_tokens": [ + -3.773287534713745 + ], + "logprob_reward": -0.023287491872906685, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.023287491872906685, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.23951834440231323, + 0.16185976564884186, + 0.14039145410060883, + 0.12008058279752731, + 0.06309229880571365, + 0.0579521469771862, + 0.0871564969420433, + 0.061692994087934494, + 0.02919338457286358, + 0.03906247019767761 + ], + "credit_map": [ + 0.024758711457252502, + 0.03371892124414444, + 0.024368269369006157, + 0.037365347146987915, + 0.026287585496902466, + 0.025883525609970093, + 0.03785507008433342, + 0.029280919581651688, + 0.02351495437324047, + 0.014729855582118034, + 0.02390260621905327, + 0.02217603474855423, + 0.022771058604121208, + 0.021027101203799248, + 0.017757166177034378, + 0.01598101668059826, + 0.012358525767922401, + 0.01747037284076214, + 0.021365148946642876, + 0.029466424137353897, + 0.012647637166082859, + 0.021205466240644455, + 0.014804902486503124, + 0.011072995141148567, + 0.01273315493017435, + 0.012315932661294937, + 0.01283774059265852, + 0.021445801481604576, + 0.013610419817268848, + 0.018733898177742958, + 0.016069162636995316, + 0.012334478087723255, + 0.005289679393172264, + 0.00669207563623786, + 0.008182087913155556, + 0.006978979334235191, + 0.0075364080257713795, + 0.009345637634396553, + 0.006404037121683359, + 0.012663397938013077, + 0.007170731667429209, + 0.006588878110051155, + 0.008977029472589493, + 0.006019540596753359, + 0.008027401752769947, + 0.008486664853990078, + 0.006646054796874523, + 0.006035849917680025, + 0.008084296248853207, + 0.013033675029873848, + 0.011845923960208893, + 0.008283093571662903, + 0.007870513014495373, + 0.010796417482197285, + 0.015267633832991123, + 0.011974947527050972, + 0.0050767334178090096, + 0.009203415364027023, + 0.008623075671494007, + 0.011578068137168884, + 0.0057156821712851524, + 0.008811848238110542, + 0.00840700138360262, + 0.004277173895388842, + 0.0032660120632499456, + 0.004060083534568548, + 0.0038584561552852392, + 0.004127939231693745, + 0.002184649696573615, + 0.004057148937135935, + 0.003623044351115823, + 0.004016050137579441, + 0.003882960882037878, + 0.004911220632493496, + 0.003750259755179286, + 0.006038293242454529, + 0.0030888954643160105, + 0.0055181547068059444, + 0.005175366997718811, + 0.006697321776300669 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 2 + ], + "selected_reward": -0.023287491872906685 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a South Asian woman with thick, dark curls?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the woman wearing a thick cream knit sweater and round wire-rimmed glasses?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the woman holding a simple white ceramic cup while standing on a yacht deck?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera start from a low angle and slowly zoom in on the subject?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background visible as a sparkling azure sea meeting a clear sky on the horizon?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.42885276675224304 + ], + "opposite_logprob_tokens": [ + -1.0538527965545654 + ], + "logprob_reward": -0.42885276675224304, + "yes_no_margin_reward": 0.625, + "token_credit_reward": -0.42885276675224304, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09050443768501282, + 0.1062694787979126, + 0.32715868949890137, + 0.04589073359966278, + 0.09302381426095963, + 0.05216443911194801, + 0.0707622617483139, + 0.027302386239171028, + 0.049100760370492935, + 0.1378229558467865 + ], + "credit_map": [ + 0.011177832260727882, + 0.009289783425629139, + 0.010761996731162071, + 0.008301645517349243, + 0.011092244647443295, + 0.011679247952997684, + 0.012158364057540894, + 0.016043324023485184, + 0.012213654816150665, + 0.009582798928022385, + 0.01284776907414198, + 0.013333064503967762, + 0.010062671266496181, + 0.014623773284256458, + 0.014727153815329075, + 0.018878603354096413, + 0.043015941977500916, + 0.033027201890945435, + 0.06303667277097702, + 0.03676295280456543, + 0.04467054083943367, + 0.04339945688843727, + 0.024940764531493187, + 0.03830515965819359, + 0.0036839789245277643, + 0.0055156792514026165, + 0.004739728756248951, + 0.008983141742646694, + 0.0034039567690342665, + 0.00624077208340168, + 0.007044882047921419, + 0.006278598681092262, + 0.007434704806655645, + 0.010874911211431026, + 0.019581308588385582, + 0.010050125420093536, + 0.010567848570644855, + 0.012940212152898312, + 0.007287279237061739, + 0.014287433587014675, + 0.00414395984262228, + 0.006388877984136343, + 0.00721733458340168, + 0.0063656289130449295, + 0.009303637780249119, + 0.006456287112087011, + 0.006023789290338755, + 0.00626492640003562, + 0.006618851330131292, + 0.010094976983964443, + 0.008717212826013565, + 0.005932914558798075, + 0.00505555747076869, + 0.009709362871944904, + 0.011996319517493248, + 0.012637076899409294, + 0.002024845452979207, + 0.004816560074687004, + 0.0040335808880627155, + 0.0030780769884586334, + 0.0034966806415468454, + 0.003755238838493824, + 0.003012110246345401, + 0.0030852933414280415, + 0.00425959937274456, + 0.0087659927085042, + 0.006161799654364586, + 0.006327420473098755, + 0.0032179304398596287, + 0.006753087975084782, + 0.004738895688205957, + 0.008876039646565914, + 0.01544087566435337, + 0.022049762308597565, + 0.011134876869618893, + 0.024640316143631935, + 0.013380389660596848, + 0.022202465683221817, + 0.01532755047082901, + 0.013646737672388554 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 9, + 1 + ], + "selected_reward": -0.42885276675224304 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a South Asian woman with thick, dark curls?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the woman wearing a thick cream knit sweater and round wire-rimmed glasses?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the woman holding a simple white ceramic cup while standing on a yacht deck?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera start from a low angle and slowly zoom in on the subject?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background visible as a sparkling azure sea meeting a clear sky on the horizon?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0040937201119959354 + ], + "opposite_logprob_tokens": [ + -5.504093647003174 + ], + "logprob_reward": -0.0040937201119959354, + "yes_no_margin_reward": 5.5, + "token_credit_reward": -0.0040937201119959354, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.22654859721660614, + 0.118745818734169, + 0.11339526623487473, + 0.03706430271267891, + 0.09148728102445602, + 0.13230472803115845, + 0.14489762485027313, + 0.02694021537899971, + 0.06623376905918121, + 0.04238240793347359 + ], + "credit_map": [ + 0.032831691205501556, + 0.016126468777656555, + 0.04305792227387428, + 0.024964001029729843, + 0.024029361084103584, + 0.031623680144548416, + 0.037316665053367615, + 0.016598818823695183, + 0.013135310262441635, + 0.008245927281677723, + 0.019541798159480095, + 0.011761476285755634, + 0.011549003422260284, + 0.014813990332186222, + 0.011560410261154175, + 0.028137894347310066, + 0.012046225368976593, + 0.00981771107763052, + 0.01068039983510971, + 0.015381278470158577, + 0.013527161441743374, + 0.01505892165005207, + 0.011613900773227215, + 0.025269674137234688, + 0.0037867745850235224, + 0.003751408075913787, + 0.0057050250470638275, + 0.005582252983003855, + 0.0036609391681849957, + 0.005536404903978109, + 0.004723068326711655, + 0.00431843101978302, + 0.008625209331512451, + 0.006962419021874666, + 0.011175710707902908, + 0.010710429400205612, + 0.010364574380218983, + 0.013823348097503185, + 0.008991068229079247, + 0.02083451673388481, + 0.013303876854479313, + 0.0149032948538661, + 0.028422649949789047, + 0.01405696477741003, + 0.01060779020190239, + 0.01648346520960331, + 0.012413796968758106, + 0.02211288921535015, + 0.020538590848445892, + 0.029413893818855286, + 0.01948152668774128, + 0.0160110704600811, + 0.013586044311523438, + 0.017846226692199707, + 0.017206856980919838, + 0.01081340853124857, + 0.0022776497062295675, + 0.0030214113648980856, + 0.004542431328445673, + 0.0029572690837085247, + 0.0035524112172424793, + 0.003906879108399153, + 0.002484070137143135, + 0.0041980925016105175, + 0.006118196062743664, + 0.0073563470505177975, + 0.011822455562651157, + 0.006487783044576645, + 0.004564276896417141, + 0.009784108959138393, + 0.007203646469861269, + 0.012896958738565445, + 0.00413864478468895, + 0.0066829342395067215, + 0.003450988093391061, + 0.007199050392955542, + 0.0025993480812758207, + 0.006049127783626318, + 0.005984301213175058, + 0.006278014276176691 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 6, + 5 + ], + "selected_reward": -0.0040937201119959354 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000034/group_11/index_tensor([3203, 3203])_group_11_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000034/group_11/index_tensor([3203, 3203])_group_11_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..c25e21366bebed478ef241d9c266ffe8da1d6f1c --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000034/group_11/index_tensor([3203, 3203])_group_11_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "static, extreme close-up captures a realistic rendering of a weathered, red-painted metal surface gleaming in intense, concentrated heat. The light is harsh and direct, creating a bright hotspot on the semi-gloss finish. From a slightly higher angle, the camera observes a small portion of paint begin to expand, slowly and deliberately rising from a flat surface into a taut, hemispherical blister, its surface stretching and thinning, the color subtly fading at its apex. A sudden, almost imperceptible vibration rips through the metal, and in an instant, a network of hairline cracks splits apart within the fragile, heat-baked blister. In a mesmerizing slow-motion sequence, the fractured paint fragments begin to curl upward and away from the breaking point, their sharp edges lifting like burnt parchment to reveal the dark, matte-polished gunmetal gray of the original metal beneath, creating an atmosphere of intense, clinical observation focused on the physics of material decay.", + "seed": 31846, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000034/group_11/index_3203_seed_31846_dp_22_rank_44_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 11, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.020546634495258332, + "vlm_logprob": -0.020546634495258332, + "vlm_yes_no_margin": 4.575, + "vlm_token_credit": -0.020546634495258332, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of a weathered, red-painted metal surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is there a bright hotspot on the semi-gloss finish caused by harsh, direct light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a paint blister rise from the surface into a taut hemispherical shape before cracking?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do fractured paint fragments curl upward to reveal dark, matte-polished gunmetal gray metal beneath?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera positioned at a slightly higher angle observing the scene?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.012001684866845608 + ], + "opposite_logprob_tokens": [ + -6.1370015144348145 + ], + "logprob_reward": -0.012001684866845608, + "yes_no_margin_reward": 6.125, + "token_credit_reward": -0.012001684866845608, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0653361901640892, + 0.09204411506652832, + 0.0856347382068634, + 0.08484416455030441, + 0.13154131174087524, + 0.11799172312021255, + 0.21449014544487, + 0.15106847882270813, + 0.027762988582253456, + 0.029286138713359833 + ], + "credit_map": [ + 0.007283974904567003, + 0.00991967972368002, + 0.006692909635603428, + 0.00642347102984786, + 0.006305590737611055, + 0.01095551811158657, + 0.007901918143033981, + 0.009853130206465721, + 0.01142098382115364, + 0.01007769163697958, + 0.01657595857977867, + 0.011827917769551277, + 0.008526372723281384, + 0.010491509921848774, + 0.007527986541390419, + 0.015595698729157448, + 0.01426931656897068, + 0.007728219497948885, + 0.011793294921517372, + 0.009173264726996422, + 0.0057746488600969315, + 0.010463963262736797, + 0.005892506334930658, + 0.020539522171020508, + 0.013819481246173382, + 0.004263445734977722, + 0.01250678114593029, + 0.008193626999855042, + 0.004816514439880848, + 0.00995486881583929, + 0.004626013338565826, + 0.026663433760404587, + 0.022069917991757393, + 0.010707474313676357, + 0.017892440780997276, + 0.010815244168043137, + 0.008446821942925453, + 0.016669413074851036, + 0.00989731028676033, + 0.03504268825054169, + 0.017493274062871933, + 0.01444199588149786, + 0.02017231658101082, + 0.0131479911506176, + 0.01450600940734148, + 0.010286988690495491, + 0.013866085559129715, + 0.014077053405344486, + 0.03562304750084877, + 0.024988964200019836, + 0.022991232573986053, + 0.020198572427034378, + 0.021916629746556282, + 0.03468186408281326, + 0.020205074921250343, + 0.03388476371765137, + 0.02308758720755577, + 0.013844778761267662, + 0.022680038586258888, + 0.014873466454446316, + 0.01687842793762684, + 0.019815875217318535, + 0.015331846661865711, + 0.02455645054578781, + 0.004090319853276014, + 0.0023780642077326775, + 0.003567367559298873, + 0.00257395813241601, + 0.00299203023314476, + 0.0034954347647726536, + 0.0026402496732771397, + 0.006025565322488546, + 0.0038419044576585293, + 0.0031899409368634224, + 0.003152060555294156, + 0.0039887987077236176, + 0.004369994625449181, + 0.0033953003585338593, + 0.004611450247466564, + 0.002736689057201147 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 7, + 4 + ], + "selected_reward": -0.012001684866845608 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of a weathered, red-painted metal surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is there a bright hotspot on the semi-gloss finish caused by harsh, direct light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a paint blister rise from the surface into a taut hemispherical shape before cracking?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do fractured paint fragments curl upward to reveal dark, matte-polished gunmetal gray metal beneath?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera positioned at a slightly higher angle observing the scene?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03813666105270386 + ], + "opposite_logprob_tokens": [ + -3.2881367206573486 + ], + "logprob_reward": -0.03813666105270386, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.03813666105270386, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10834676027297974, + 0.17512957751750946, + 0.15336383879184723, + 0.12296684086322784, + 0.05481646955013275, + 0.0565972663462162, + 0.06508567184209824, + 0.13228580355644226, + 0.10821026563644409, + 0.023197518661618233 + ], + "credit_map": [ + 0.01126259658485651, + 0.013075223192572594, + 0.01317930594086647, + 0.01244583074003458, + 0.01129782572388649, + 0.01570468582212925, + 0.01387610100209713, + 0.017505180090665817, + 0.02762642689049244, + 0.018469927832484245, + 0.025321461260318756, + 0.021246323361992836, + 0.016983281821012497, + 0.020216573029756546, + 0.01696162112057209, + 0.02830394171178341, + 0.025243768468499184, + 0.01629987731575966, + 0.024322422221302986, + 0.013009500689804554, + 0.01069780346006155, + 0.026621846482157707, + 0.009591170586645603, + 0.02757743000984192, + 0.02375887893140316, + 0.007641289383172989, + 0.01393388956785202, + 0.009357915259897709, + 0.008580055087804794, + 0.010855206288397312, + 0.007670758757740259, + 0.04116882383823395, + 0.009586832486093044, + 0.00489810248836875, + 0.007592090871185064, + 0.004604141227900982, + 0.005107794422656298, + 0.00646050414070487, + 0.005145243369042873, + 0.011421754956245422, + 0.008394899778068066, + 0.006948353722691536, + 0.007921820506453514, + 0.0051854331977665424, + 0.004491868428885937, + 0.008256809785962105, + 0.005641480442136526, + 0.009756594896316528, + 0.009614940732717514, + 0.0072831930592656136, + 0.005649962928146124, + 0.007336849812418222, + 0.008193985559046268, + 0.009076539427042007, + 0.007570736575871706, + 0.010359461419284344, + 0.017484547570347786, + 0.012293257750570774, + 0.015356596559286118, + 0.021526169031858444, + 0.0183598380535841, + 0.012871919199824333, + 0.02237696945667267, + 0.012016496621072292, + 0.01576884463429451, + 0.009073171764612198, + 0.012403273954987526, + 0.010199714452028275, + 0.008882734924554825, + 0.013143460266292095, + 0.011524940840899944, + 0.027214109897613525, + 0.0029423448722809553, + 0.0025106684770435095, + 0.002575765596702695, + 0.002312281634658575, + 0.0038151326589286327, + 0.0023634880781173706, + 0.004628187511116266, + 0.0020496470388025045 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 2, + 7 + ], + "selected_reward": -0.03813666105270386 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of a weathered, red-painted metal surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is there a bright hotspot on the semi-gloss finish caused by harsh, direct light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a paint blister rise from the surface into a taut hemispherical shape before cracking?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do fractured paint fragments curl upward to reveal dark, matte-polished gunmetal gray metal beneath?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera positioned at a slightly higher angle observing the scene?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.020590832456946373 + ], + "opposite_logprob_tokens": [ + -3.8955907821655273 + ], + "logprob_reward": -0.020590832456946373, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.020590832456946373, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08876068890094757, + 0.05552729219198227, + 0.0846911370754242, + 0.19409433007240295, + 0.036437101662158966, + 0.06335198879241943, + 0.09156937897205353, + 0.24819786846637726, + 0.09624581784009933, + 0.0411243811249733 + ], + "credit_map": [ + 0.010973597876727581, + 0.011301648803055286, + 0.012118417769670486, + 0.010114355012774467, + 0.007039716001600027, + 0.01161497924476862, + 0.012732676230370998, + 0.012865296564996243, + 0.005293665919452906, + 0.005718155764043331, + 0.0059937830083072186, + 0.006145421415567398, + 0.0063610682263970375, + 0.008043097332119942, + 0.010415141470730305, + 0.007556963711977005, + 0.01612498238682747, + 0.005214756354689598, + 0.012043388560414314, + 0.006052684970200062, + 0.005297279451042414, + 0.015592323616147041, + 0.00739385187625885, + 0.01697186566889286, + 0.03643809258937836, + 0.011143786832690239, + 0.024740729480981827, + 0.01296890527009964, + 0.01054870430380106, + 0.01695641130208969, + 0.008867020718753338, + 0.07243068516254425, + 0.00480971485376358, + 0.004051913972944021, + 0.005041557364165783, + 0.005517189856618643, + 0.004333318676799536, + 0.0030812828335911036, + 0.005054153501987457, + 0.004547970835119486, + 0.00951702706515789, + 0.005282798316329718, + 0.009414972737431526, + 0.008179867640137672, + 0.006731508765369654, + 0.006605967413634062, + 0.009641832672059536, + 0.007978016510605812, + 0.012972325086593628, + 0.010716727003455162, + 0.009184630587697029, + 0.011346600018441677, + 0.01211948599666357, + 0.012766439467668533, + 0.010902376845479012, + 0.011560792103409767, + 0.03146660327911377, + 0.02253754809498787, + 0.02483617328107357, + 0.04135055094957352, + 0.03667103499174118, + 0.02073829621076584, + 0.04678366705775261, + 0.023813996464014053, + 0.014214150607585907, + 0.008017138577997684, + 0.012294902466237545, + 0.007528782356530428, + 0.00813496857881546, + 0.011530495248734951, + 0.010913236066699028, + 0.02361214905977249, + 0.004968808963894844, + 0.005364343989640474, + 0.004940801300108433, + 0.005256236996501684, + 0.005452560260891914, + 0.006024335511028767, + 0.005918434821069241, + 0.003198861377313733 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 3, + 8 + ], + "selected_reward": -0.020590832456946373 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of a weathered, red-painted metal surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is there a bright hotspot on the semi-gloss finish caused by harsh, direct light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a paint blister rise from the surface into a taut hemispherical shape before cracking?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do fractured paint fragments curl upward to reveal dark, matte-polished gunmetal gray metal beneath?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera positioned at a slightly higher angle observing the scene?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0022011594846844673 + ], + "opposite_logprob_tokens": [ + -6.127201080322266 + ], + "logprob_reward": -0.0022011594846844673, + "yes_no_margin_reward": 6.125, + "token_credit_reward": -0.0022011594846844673, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13207021355628967, + 0.09903033077716827, + 0.1241253912448883, + 0.14539438486099243, + 0.04012411832809448, + 0.038926515728235245, + 0.08133809268474579, + 0.15304064750671387, + 0.12241973727941513, + 0.0635305866599083 + ], + "credit_map": [ + 0.014068861491978168, + 0.01567349210381508, + 0.016201213002204895, + 0.014305784367024899, + 0.010263768956065178, + 0.03132041171193123, + 0.01493842527270317, + 0.015298254787921906, + 0.011792558245360851, + 0.012145407497882843, + 0.015675725415349007, + 0.013466877862811089, + 0.00921592116355896, + 0.00980088859796524, + 0.010530254803597927, + 0.016402697190642357, + 0.021000519394874573, + 0.0119360676035285, + 0.017141243442893028, + 0.010796703398227692, + 0.009586474858224392, + 0.020056772977113724, + 0.010879145003855228, + 0.022728461772203445, + 0.02481953427195549, + 0.008439583703875542, + 0.01984395831823349, + 0.011260990984737873, + 0.008030595257878304, + 0.012774159200489521, + 0.008731621317565441, + 0.05149393528699875, + 0.005645324010401964, + 0.00432988116517663, + 0.005940063390880823, + 0.004742641467601061, + 0.004357811063528061, + 0.005079805850982666, + 0.004439759999513626, + 0.0055888332426548, + 0.0055825007148087025, + 0.004843279253691435, + 0.004357791505753994, + 0.0056573920883238316, + 0.0031463291961699724, + 0.004448215011507273, + 0.005054436158388853, + 0.005836572963744402, + 0.012527192942798138, + 0.01129311602562666, + 0.007300826720893383, + 0.008266056887805462, + 0.010003777220845222, + 0.012882466427981853, + 0.009389383718371391, + 0.009675269946455956, + 0.020741140469908714, + 0.012714081443846226, + 0.020930001512169838, + 0.021814385429024696, + 0.021628694608807564, + 0.01518834475427866, + 0.024987315759062767, + 0.015036676079034805, + 0.017165951430797577, + 0.009826557710766792, + 0.014554807916283607, + 0.011632942594587803, + 0.011063818819820881, + 0.01510333176702261, + 0.016706708818674088, + 0.026365624740719795, + 0.0089264502748847, + 0.008729341439902782, + 0.0076998150907456875, + 0.007249224465340376, + 0.00869917031377554, + 0.00798554252833128, + 0.00927390344440937, + 0.004967137705534697 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 3, + 0 + ], + "selected_reward": -0.0022011594846844673 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of a weathered, red-painted metal surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is there a bright hotspot on the semi-gloss finish caused by harsh, direct light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a paint blister rise from the surface into a taut hemispherical shape before cracking?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do fractured paint fragments curl upward to reveal dark, matte-polished gunmetal gray metal beneath?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera positioned at a slightly higher angle observing the scene?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02980283461511135 + ], + "opposite_logprob_tokens": [ + -3.5298027992248535 + ], + "logprob_reward": -0.02980283461511135, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.02980283461511135, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.28870153427124023, + 0.06863676756620407, + 0.10283883661031723, + 0.04441201314330101, + 0.04750903323292732, + 0.06498522311449051, + 0.13881438970565796, + 0.10668300092220306, + 0.039639879018068314, + 0.09777931123971939 + ], + "credit_map": [ + 0.029651910066604614, + 0.03770468011498451, + 0.036836620420217514, + 0.03632928431034088, + 0.020809825509786606, + 0.05916542559862137, + 0.028043365105986595, + 0.04016043618321419, + 0.007657912094146013, + 0.006972293835133314, + 0.013221398927271366, + 0.007760001812130213, + 0.007362554781138897, + 0.00714094378054142, + 0.009366463869810104, + 0.009155204519629478, + 0.016592107713222504, + 0.008497023023664951, + 0.01962924934923649, + 0.011693001724779606, + 0.006770562846213579, + 0.016082147136330605, + 0.007985045202076435, + 0.01558971218764782, + 0.006796394940465689, + 0.003348381258547306, + 0.006631604861468077, + 0.003269439795985818, + 0.0036651156842708588, + 0.004337642341852188, + 0.00335879297927022, + 0.01300464104861021, + 0.006663789972662926, + 0.005042589269578457, + 0.006247370503842831, + 0.004527167417109013, + 0.006070149131119251, + 0.0051847193390131, + 0.005729653872549534, + 0.008043600246310234, + 0.008631711825728416, + 0.009383047930896282, + 0.0074674217030406, + 0.008928133174777031, + 0.008164222352206707, + 0.008044100366532803, + 0.0067766038700938225, + 0.007589986082166433, + 0.021025532856583595, + 0.016791049391031265, + 0.012811163440346718, + 0.013813608326017857, + 0.0183442160487175, + 0.01998235657811165, + 0.01711801253259182, + 0.018928460776805878, + 0.015886759385466576, + 0.011529778130352497, + 0.01611357554793358, + 0.013415123336017132, + 0.013800338841974735, + 0.011576566845178604, + 0.010587290860712528, + 0.013773581944406033, + 0.005273984745144844, + 0.00400544423609972, + 0.004437485244125128, + 0.004992486443370581, + 0.004679822828620672, + 0.003598713083192706, + 0.005033356603235006, + 0.007618589326739311, + 0.012511427514255047, + 0.012774372473359108, + 0.009268858470022678, + 0.01191043108701706, + 0.015834778547286987, + 0.013377388939261436, + 0.014860392548143864, + 0.007241674233227968 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 6, + 7 + ], + "selected_reward": -0.02980283461511135 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "static, extreme close-up captures a realistic rendering of a weathered, red-painted metal surface gleaming in intense, concentrated heat. The light is harsh and direct, creating a bright hotspot on the semi-gloss finish. From a slightly higher angle, the camera observes a small portion of paint begin to expand, slowly and deliberately rising from a flat surface into a taut, hemispherical blister, its surface stretching and thinning, the color subtly fading at its apex. A sudden, almost imperceptible vibration rips through the metal, and in an instant, a network of hairline cracks splits apart within the fragile, heat-baked blister. In a mesmerizing slow-motion sequence, the fractured paint fragments begin to curl upward and away from the breaking point, their sharp edges lifting like burnt parchment to reveal the dark, matte-polished gunmetal gray of the original metal beneath, creating an atmosphere of intense, clinical observation focused on the physics of material decay.", + "seed": 31847, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000034/group_11/index_3203_seed_31847_dp_22_rank_44_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 11, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.018951001390814782, + "vlm_logprob": -0.018951001390814782, + "vlm_yes_no_margin": 4.7, + "vlm_token_credit": -0.018951001390814782, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of a weathered, red-painted metal surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is there a bright hotspot on the semi-gloss finish caused by harsh, direct light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a paint blister rise from the surface into a taut hemispherical shape before cracking?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do fractured paint fragments curl upward to reveal dark, matte-polished gunmetal gray metal beneath?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera positioned at a slightly higher angle observing the scene?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.012288101948797703 + ], + "opposite_logprob_tokens": [ + -6.0122880935668945 + ], + "logprob_reward": -0.012288101948797703, + "yes_no_margin_reward": 6.0, + "token_credit_reward": -0.012288101948797703, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15128667652606964, + 0.024637483060359955, + 0.023870578035712242, + 0.027250465005636215, + 0.028989026322960854, + 0.10741548240184784, + 0.09767088294029236, + 0.07945095002651215, + 0.3610629439353943, + 0.09836551547050476 + ], + "credit_map": [ + 0.009017781354486942, + 0.003641857299953699, + 0.0038845259696245193, + 0.003357533598318696, + 0.004759658128023148, + 0.06384281814098358, + 0.004028120543807745, + 0.058754388242959976, + 0.0022696047089993954, + 0.0021036211401224136, + 0.0030168548692017794, + 0.0015745839336887002, + 0.00367603893391788, + 0.005317019764333963, + 0.0026228695642203093, + 0.004056890495121479, + 0.0024891989305615425, + 0.0036680917255580425, + 0.0026089376769959927, + 0.002245859708636999, + 0.0034894042182713747, + 0.0039060492999851704, + 0.002361949300393462, + 0.0031010888051241636, + 0.0022491735871881247, + 0.0031705123838037252, + 0.0033570691011846066, + 0.0035767261870205402, + 0.004393588285893202, + 0.0047059571370482445, + 0.002859972883015871, + 0.0029374645091593266, + 0.003227266250178218, + 0.0027247632388025522, + 0.004406476393342018, + 0.003358315909281373, + 0.003314351662993431, + 0.004216132219880819, + 0.0037745945155620575, + 0.0039671240374445915, + 0.00930488109588623, + 0.005099304020404816, + 0.008790340274572372, + 0.016811834648251534, + 0.006561285816133022, + 0.018352575600147247, + 0.008330627344548702, + 0.034164633601903915, + 0.006918325088918209, + 0.006359598133713007, + 0.010135903023183346, + 0.01961234211921692, + 0.00791897252202034, + 0.013064220547676086, + 0.008333786390721798, + 0.025327742099761963, + 0.007617824710905552, + 0.004412654787302017, + 0.01700117066502571, + 0.02256484143435955, + 0.004614039324223995, + 0.009726359508931637, + 0.004628302529454231, + 0.008885757997632027, + 0.038885753601789474, + 0.015955017879605293, + 0.05705295503139496, + 0.023800915107131004, + 0.02734282612800598, + 0.06756027787923813, + 0.03640126809477806, + 0.09406393766403198, + 0.006940705701708794, + 0.006740563083440065, + 0.009591138921678066, + 0.008658583275973797, + 0.007979948073625565, + 0.023894336074590683, + 0.008783834055066109, + 0.025776410475373268 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 0, + 5 + ], + "selected_reward": -0.012288101948797703 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of a weathered, red-painted metal surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is there a bright hotspot on the semi-gloss finish caused by harsh, direct light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a paint blister rise from the surface into a taut hemispherical shape before cracking?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do fractured paint fragments curl upward to reveal dark, matte-polished gunmetal gray metal beneath?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera positioned at a slightly higher angle observing the scene?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.029837431386113167 + ], + "opposite_logprob_tokens": [ + -3.5298373699188232 + ], + "logprob_reward": -0.029837431386113167, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.029837431386113167, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.04083247110247612, + 0.06081937253475189, + 0.05065501108765602, + 0.04827893525362015, + 0.047338489443063736, + 0.35358482599258423, + 0.07848900556564331, + 0.19421495497226715, + 0.0791485384106636, + 0.04663845896720886 + ], + "credit_map": [ + 0.003614085726439953, + 0.005296419374644756, + 0.0031973544973880053, + 0.0052747274748981, + 0.006640539038926363, + 0.005876258481293917, + 0.005119753535836935, + 0.005813330877572298, + 0.007190174888819456, + 0.006106835789978504, + 0.006155494134873152, + 0.003449568524956703, + 0.00960427988320589, + 0.014259162358939648, + 0.004627317655831575, + 0.009426531381905079, + 0.00613653939217329, + 0.005135711748152971, + 0.008031648583710194, + 0.005579712800681591, + 0.006211483851075172, + 0.007114194333553314, + 0.006189979612827301, + 0.006255740765482187, + 0.004474628251045942, + 0.004265511874109507, + 0.00607663718983531, + 0.005525555927306414, + 0.00846038106828928, + 0.007694091647863388, + 0.004627234768122435, + 0.0071548884734511375, + 0.004662377759814262, + 0.006351768970489502, + 0.006769830826669931, + 0.005524931475520134, + 0.00523237232118845, + 0.00419429037719965, + 0.006328882649540901, + 0.008274028077721596, + 0.03177153691649437, + 0.01979498378932476, + 0.019978288561105728, + 0.09810422360897064, + 0.015298175625503063, + 0.04010260850191116, + 0.03122114948928356, + 0.09731383621692657, + 0.00368016236461699, + 0.004053391050547361, + 0.0063805659301579, + 0.018796568736433983, + 0.0031514246948063374, + 0.010388360358774662, + 0.006367955822497606, + 0.02567056193947792, + 0.051734112203121185, + 0.01187270786613226, + 0.035032328218221664, + 0.018419628962874413, + 0.009424513205885887, + 0.02714778482913971, + 0.01464792899787426, + 0.025935940444469452, + 0.007744454313069582, + 0.0035507292486727238, + 0.011800079606473446, + 0.004770950879901648, + 0.006807246245443821, + 0.011771801859140396, + 0.011136287823319435, + 0.021566977724432945, + 0.002991969231516123, + 0.002861183136701584, + 0.0046256789937615395, + 0.004465836565941572, + 0.0031714742071926594, + 0.011233558878302574, + 0.004741629585623741, + 0.012547126971185207 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 7, + 8 + ], + "selected_reward": -0.029837431386113167 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of a weathered, red-painted metal surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is there a bright hotspot on the semi-gloss finish caused by harsh, direct light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a paint blister rise from the surface into a taut hemispherical shape before cracking?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do fractured paint fragments curl upward to reveal dark, matte-polished gunmetal gray metal beneath?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera positioned at a slightly higher angle observing the scene?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.006746372207999229 + ], + "opposite_logprob_tokens": [ + -5.006746292114258 + ], + "logprob_reward": -0.006746372207999229, + "yes_no_margin_reward": 5.0, + "token_credit_reward": -0.006746372207999229, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.03458641469478607, + 0.03561094403266907, + 0.03157760947942734, + 0.08426345139741898, + 0.03994952142238617, + 0.1580137312412262, + 0.14764584600925446, + 0.3235233426094055, + 0.06238784268498421, + 0.08244132250547409 + ], + "credit_map": [ + 0.003534073242917657, + 0.0024810004979372025, + 0.003988734446465969, + 0.0023743975907564163, + 0.004222853109240532, + 0.006863197777420282, + 0.003976654261350632, + 0.007145501207560301, + 0.003470690455287695, + 0.0032327487133443356, + 0.0038202879950404167, + 0.0027638610918074846, + 0.007321289740502834, + 0.005930175539106131, + 0.003734079422429204, + 0.005337809678167105, + 0.0037366957403719425, + 0.0025875915307551622, + 0.005855718161910772, + 0.0031696257647126913, + 0.003730504307895899, + 0.004644503351300955, + 0.0035035053733736277, + 0.004349466413259506, + 0.00542827881872654, + 0.007658443879336119, + 0.008790599182248116, + 0.015675675123929977, + 0.012952880933880806, + 0.012205050326883793, + 0.010477588512003422, + 0.011074930429458618, + 0.0036473353393375874, + 0.0035870159044861794, + 0.005623296368867159, + 0.007184217218309641, + 0.004890476353466511, + 0.004139558877795935, + 0.005968204699456692, + 0.004909415263682604, + 0.01425864640623331, + 0.009694268926978111, + 0.009340922348201275, + 0.03887661173939705, + 0.012974618002772331, + 0.025745531544089317, + 0.014173909090459347, + 0.03294922038912773, + 0.006717666517943144, + 0.007788261864334345, + 0.011114360764622688, + 0.03719682991504669, + 0.006891596131026745, + 0.019290132448077202, + 0.012779484502971172, + 0.04586750268936157, + 0.08269371092319489, + 0.020065471529960632, + 0.07167499512434006, + 0.02549159713089466, + 0.016128327697515488, + 0.04713108390569687, + 0.02497502788901329, + 0.03536312282085419, + 0.0072273975238204, + 0.004504895303398371, + 0.006631059106439352, + 0.004556226544082165, + 0.009317625313997269, + 0.01303289458155632, + 0.007515064440667629, + 0.009602680802345276, + 0.005918106064200401, + 0.005377229303121567, + 0.009620079770684242, + 0.00810754019767046, + 0.007601517718285322, + 0.014738989993929863, + 0.011830588802695274, + 0.019247276708483696 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 5, + 6 + ], + "selected_reward": -0.006746372207999229 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of a weathered, red-painted metal surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is there a bright hotspot on the semi-gloss finish caused by harsh, direct light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a paint blister rise from the surface into a taut hemispherical shape before cracking?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do fractured paint fragments curl upward to reveal dark, matte-polished gunmetal gray metal beneath?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera positioned at a slightly higher angle observing the scene?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.00282187107950449 + ], + "opposite_logprob_tokens": [ + -5.877821922302246 + ], + "logprob_reward": -0.00282187107950449, + "yes_no_margin_reward": 5.875, + "token_credit_reward": -0.00282187107950449, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1397063434123993, + 0.018126560375094414, + 0.023354142904281616, + 0.02674013562500477, + 0.022135786712169647, + 0.2175658792257309, + 0.07225780189037323, + 0.09333128482103348, + 0.2916654050350189, + 0.09511672705411911 + ], + "credit_map": [ + 0.007331717759370804, + 0.00472137238830328, + 0.006568931508809328, + 0.0032703070901334286, + 0.005423631519079208, + 0.05646754428744316, + 0.00416203960776329, + 0.05176078528165817, + 0.0018290103180333972, + 0.0014093335485085845, + 0.0018141486216336489, + 0.0009347865707240999, + 0.001960350666195154, + 0.005445774178951979, + 0.0012109556701034307, + 0.0035221995785832405, + 0.0024496789555996656, + 0.003080721478909254, + 0.0029915852937847376, + 0.0030013422947376966, + 0.0037427092902362347, + 0.002982456237077713, + 0.0026742357295006514, + 0.0024314129259437323, + 0.002297025639563799, + 0.002602442167699337, + 0.003188331378623843, + 0.0036361070815473795, + 0.0036019152030348778, + 0.005346306599676609, + 0.0030272023286670446, + 0.003040805459022522, + 0.002082751365378499, + 0.0017050383612513542, + 0.004116546828299761, + 0.0023941078688949347, + 0.0029626290779560804, + 0.003931416664272547, + 0.002809245837852359, + 0.002134049544110894, + 0.018265653401613235, + 0.016999557614326477, + 0.014837456867098808, + 0.05481327697634697, + 0.02107912488281727, + 0.0365772619843483, + 0.020328354090452194, + 0.03466518968343735, + 0.0077776312828063965, + 0.0068085952661931515, + 0.00855246838182211, + 0.01180025190114975, + 0.008059813641011715, + 0.011463119648396969, + 0.008373002521693707, + 0.009422920644283295, + 0.01669146493077278, + 0.006016541738063097, + 0.00679729925468564, + 0.007995991967618465, + 0.005961718503385782, + 0.022736074402928352, + 0.004460765514522791, + 0.022671427577733994, + 0.03295553848147392, + 0.014394496567547321, + 0.0461568683385849, + 0.02120942994952202, + 0.025903893634676933, + 0.05774787440896034, + 0.025797804817557335, + 0.06749949604272842, + 0.006828248035162687, + 0.005964925512671471, + 0.010468127205967903, + 0.010472345165908337, + 0.007522807456552982, + 0.017242509871721268, + 0.014806506223976612, + 0.02181125618517399 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 5, + 0 + ], + "selected_reward": -0.00282187107950449 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of a weathered, red-painted metal surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is there a bright hotspot on the semi-gloss finish caused by harsh, direct light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a paint blister rise from the surface into a taut hemispherical shape before cracking?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do fractured paint fragments curl upward to reveal dark, matte-polished gunmetal gray metal beneath?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera positioned at a slightly higher angle observing the scene?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04306123033165932 + ], + "opposite_logprob_tokens": [ + -3.1680612564086914 + ], + "logprob_reward": -0.04306123033165932, + "yes_no_margin_reward": 3.125, + "token_credit_reward": -0.04306123033165932, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3136378824710846, + 0.010630491189658642, + 0.0267502311617136, + 0.024666940793395042, + 0.019101914018392563, + 0.11652299761772156, + 0.09119214862585068, + 0.21855758130550385, + 0.16414865851402283, + 0.014791151508688927 + ], + "credit_map": [ + 0.015244168229401112, + 0.008454321883618832, + 0.0121487220749259, + 0.00678445678204298, + 0.008639905601739883, + 0.13288861513137817, + 0.008220871910452843, + 0.12125685065984726, + 0.001084326533600688, + 0.0010869128163903952, + 0.0012630015844479203, + 0.0009617808391340077, + 0.0012171459384262562, + 0.0023158167023211718, + 0.0011394935427233577, + 0.001562013290822506, + 0.0029525989666581154, + 0.0029003415256738663, + 0.00530467601493001, + 0.0024999461602419615, + 0.0032117394730448723, + 0.0031734423246234655, + 0.002528803190216422, + 0.004178685136139393, + 0.0021401506382972, + 0.0024638448376208544, + 0.002324160421267152, + 0.003201920771971345, + 0.005917072761803865, + 0.002599121304228902, + 0.0028629456646740437, + 0.0031577248591929674, + 0.001789996400475502, + 0.0018680654466152191, + 0.0025102836079895496, + 0.00301087973639369, + 0.0028818531427532434, + 0.002437583403661847, + 0.002332439413294196, + 0.0022708128672093153, + 0.007865432649850845, + 0.005553885363042355, + 0.008724953047931194, + 0.01707303337752819, + 0.01080907229334116, + 0.025437097996473312, + 0.008733577094972134, + 0.03232593834400177, + 0.006432756315916777, + 0.006479011848568916, + 0.008241537027060986, + 0.02063070982694626, + 0.006230574566870928, + 0.011496777646243572, + 0.007899437099695206, + 0.02378135174512863, + 0.054771166294813156, + 0.013148556463420391, + 0.04816146939992905, + 0.020331816747784615, + 0.011389750987291336, + 0.03226726874709129, + 0.017123380675911903, + 0.02136417105793953, + 0.01675773225724697, + 0.006736463401466608, + 0.02699660323560238, + 0.01046506967395544, + 0.011683041229844093, + 0.028149118646979332, + 0.018384024500846863, + 0.044976603239774704, + 0.0018516762647777796, + 0.001537380856461823, + 0.0016109227435663342, + 0.001332193030975759, + 0.0017097062664106488, + 0.0025646276772022247, + 0.002205467550083995, + 0.0019791757222265005 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 7, + 8 + ], + "selected_reward": -0.04306123033165932 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "static, extreme close-up captures a realistic rendering of a weathered, red-painted metal surface gleaming in intense, concentrated heat. The light is harsh and direct, creating a bright hotspot on the semi-gloss finish. From a slightly higher angle, the camera observes a small portion of paint begin to expand, slowly and deliberately rising from a flat surface into a taut, hemispherical blister, its surface stretching and thinning, the color subtly fading at its apex. A sudden, almost imperceptible vibration rips through the metal, and in an instant, a network of hairline cracks splits apart within the fragile, heat-baked blister. In a mesmerizing slow-motion sequence, the fractured paint fragments begin to curl upward and away from the breaking point, their sharp edges lifting like burnt parchment to reveal the dark, matte-polished gunmetal gray of the original metal beneath, creating an atmosphere of intense, clinical observation focused on the physics of material decay.", + "seed": 31848, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000034/group_11/index_3203_seed_31848_dp_23_rank_46_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 11, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.016936011775396765, + "vlm_logprob": -0.016936011775396765, + "vlm_yes_no_margin": 4.65, + "vlm_token_credit": -0.016936011775396765, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of a weathered, red-painted metal surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is there a bright hotspot on the semi-gloss finish caused by harsh, direct light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a paint blister rise from the surface into a taut hemispherical shape before cracking?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do fractured paint fragments curl upward to reveal dark, matte-polished gunmetal gray metal beneath?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera positioned at a slightly higher angle observing the scene?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.010064095258712769 + ], + "opposite_logprob_tokens": [ + -5.510064125061035 + ], + "logprob_reward": -0.010064095258712769, + "yes_no_margin_reward": 5.5, + "token_credit_reward": -0.010064095258712769, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1115407794713974, + 0.1183815598487854, + 0.2769853174686432, + 0.05294821038842201, + 0.09212879091501236, + 0.05176996812224388, + 0.04033859446644783, + 0.03249118849635124, + 0.05148863047361374, + 0.17192696034908295 + ], + "credit_map": [ + 0.016099892556667328, + 0.010952494107186794, + 0.01950737275183201, + 0.010605575516819954, + 0.01476055383682251, + 0.014601926319301128, + 0.012081504799425602, + 0.0129314623773098, + 0.017962925136089325, + 0.010367467068135738, + 0.013728639110922813, + 0.01465743687003851, + 0.011070598848164082, + 0.019396591931581497, + 0.016716090962290764, + 0.01448181550949812, + 0.07211487740278244, + 0.015283198095858097, + 0.013686751946806908, + 0.012025231495499611, + 0.01840277947485447, + 0.10304277390241623, + 0.014912668615579605, + 0.027517037466168404, + 0.007026264443993568, + 0.002425874350592494, + 0.005553599912673235, + 0.0032215642277151346, + 0.002355285920202732, + 0.01780722290277481, + 0.0034352599177509546, + 0.011123144067823887, + 0.009959382005035877, + 0.005943072959780693, + 0.010964181274175644, + 0.008017678745090961, + 0.007393815089017153, + 0.01437821239233017, + 0.011712923645973206, + 0.02375953458249569, + 0.005855469033122063, + 0.003419652348384261, + 0.007427559234201908, + 0.0065688034519553185, + 0.006729671265929937, + 0.008471256121993065, + 0.006409265100955963, + 0.006888294592499733, + 0.009486480616033077, + 0.003023760858923197, + 0.004473943263292313, + 0.0028237695805728436, + 0.005357928574085236, + 0.007705231662839651, + 0.003012490225955844, + 0.004454994574189186, + 0.004292359575629234, + 0.00223892112262547, + 0.004731812980026007, + 0.003153761150315404, + 0.003495587734505534, + 0.007292924914509058, + 0.003378107212483883, + 0.00390771497040987, + 0.011120613664388657, + 0.00140588718932122, + 0.005088966805487871, + 0.004632194060832262, + 0.003330028150230646, + 0.017746053636074066, + 0.003444956848397851, + 0.004719932563602924, + 0.0399821512401104, + 0.01085236668586731, + 0.011470071971416473, + 0.0184298288077116, + 0.012892119586467743, + 0.05255960673093796, + 0.010696735233068466, + 0.015044093132019043 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 9, + 1 + ], + "selected_reward": -0.010064095258712769 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of a weathered, red-painted metal surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is there a bright hotspot on the semi-gloss finish caused by harsh, direct light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a paint blister rise from the surface into a taut hemispherical shape before cracking?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do fractured paint fragments curl upward to reveal dark, matte-polished gunmetal gray metal beneath?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera positioned at a slightly higher angle observing the scene?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03373648226261139 + ], + "opposite_logprob_tokens": [ + -3.40873646736145 + ], + "logprob_reward": -0.03373648226261139, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.03373648226261139, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3389635384082794, + 0.08632976561784744, + 0.22396741807460785, + 0.10970453917980194, + 0.07584881037473679, + 0.05121755972504616, + 0.04276898875832558, + 0.019013404846191406, + 0.017947321757674217, + 0.034238673746585846 + ], + "credit_map": [ + 0.06497836858034134, + 0.037291135638952255, + 0.026927998289465904, + 0.029382579028606415, + 0.02631489560008049, + 0.09681146591901779, + 0.02831890620291233, + 0.028938185423612595, + 0.013094194233417511, + 0.0063591767102479935, + 0.012986063957214355, + 0.008235301822423935, + 0.00816029030829668, + 0.01451283972710371, + 0.011535766534507275, + 0.01144612580537796, + 0.052415017038583755, + 0.014676005579531193, + 0.015362595207989216, + 0.014754538424313068, + 0.016135970130562782, + 0.07621331512928009, + 0.014138013124465942, + 0.020271949470043182, + 0.03355233743786812, + 0.005825664848089218, + 0.00989026390016079, + 0.006848352029919624, + 0.005596083588898182, + 0.02268640138208866, + 0.00802316889166832, + 0.017282262444496155, + 0.009357800707221031, + 0.003287574043497443, + 0.006554324645549059, + 0.005032740067690611, + 0.0039005689322948456, + 0.01365392655134201, + 0.006156171206384897, + 0.027905704453587532, + 0.015264506451785564, + 0.002378418343141675, + 0.003638958092778921, + 0.003080216934904456, + 0.0029910255689173937, + 0.01677520014345646, + 0.0026574213989079, + 0.004431812558323145, + 0.015301062725484371, + 0.0012926026247441769, + 0.0024730924051254988, + 0.0013731736689805984, + 0.0017601370345801115, + 0.016515856608748436, + 0.0014874537009745836, + 0.002565610222518444, + 0.0032068367581814528, + 0.0014212805544957519, + 0.001425000256858766, + 0.002471779240295291, + 0.001822682679630816, + 0.0031510326080024242, + 0.0028416579589247704, + 0.0026731339748948812, + 0.0030127475038170815, + 0.0010385983623564243, + 0.002297345781698823, + 0.0018191344570368528, + 0.0022607771679759026, + 0.0033524632453918457, + 0.0019862062763422728, + 0.0021800477989017963, + 0.006355165969580412, + 0.002595917321741581, + 0.0038870512507855892, + 0.003642018185928464, + 0.002448095241561532, + 0.008768719621002674, + 0.003369132289662957, + 0.0031725759617984295 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 3 + ], + "selected_reward": -0.03373648226261139 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of a weathered, red-painted metal surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is there a bright hotspot on the semi-gloss finish caused by harsh, direct light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a paint blister rise from the surface into a taut hemispherical shape before cracking?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do fractured paint fragments curl upward to reveal dark, matte-polished gunmetal gray metal beneath?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera positioned at a slightly higher angle observing the scene?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.016074467450380325 + ], + "opposite_logprob_tokens": [ + -4.1410746574401855 + ], + "logprob_reward": -0.016074467450380325, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.016074467450380325, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11488892883062363, + 0.12138577550649643, + 0.45541664958000183, + 0.09157390147447586, + 0.033499617129564285, + 0.03070581890642643, + 0.03901592269539833, + 0.044923342764377594, + 0.038914140313863754, + 0.02967592142522335 + ], + "credit_map": [ + 0.013311277143657207, + 0.008148282766342163, + 0.006709509063512087, + 0.006261125206947327, + 0.02146398276090622, + 0.04295557737350464, + 0.008492331951856613, + 0.007546846754848957, + 0.013482015579938889, + 0.010774575173854828, + 0.019179165363311768, + 0.0132933733984828, + 0.009972947649657726, + 0.017386987805366516, + 0.01623084582388401, + 0.021065864711999893, + 0.12900768220424652, + 0.02243708074092865, + 0.01994132250547409, + 0.014666661620140076, + 0.028616638854146004, + 0.18513958156108856, + 0.0189345795661211, + 0.03667309135198593, + 0.029643310233950615, + 0.004324143752455711, + 0.006237806286662817, + 0.003782459534704685, + 0.0042503345757722855, + 0.02722352184355259, + 0.006452222820371389, + 0.009660103358328342, + 0.0037564255762845278, + 0.0022377129644155502, + 0.006203529890626669, + 0.003956870641559362, + 0.0028257386293262243, + 0.005663941614329815, + 0.0030769312288612127, + 0.0057784696109592915, + 0.005969079211354256, + 0.002112957648932934, + 0.0037956663873046637, + 0.0030994205735623837, + 0.0030775247141718864, + 0.006679479032754898, + 0.002936495002359152, + 0.0030351977329701185, + 0.011367284692823887, + 0.0017901388928294182, + 0.003484798828139901, + 0.0018738973885774612, + 0.0027266698889434338, + 0.011992826126515865, + 0.002014383440837264, + 0.0037659259978681803, + 0.0049380892887711525, + 0.0038978748489171267, + 0.006926991976797581, + 0.006997184827923775, + 0.005026669707149267, + 0.006332207936793566, + 0.006331616081297398, + 0.004472707863897085, + 0.007843498140573502, + 0.002001999644562602, + 0.0022034267894923687, + 0.0034204255789518356, + 0.005042353644967079, + 0.011246908456087112, + 0.0025959352497011423, + 0.004559592809528112, + 0.005628525745123625, + 0.00262147793546319, + 0.0031750770285725594, + 0.003262314712628722, + 0.002835573861375451, + 0.007507431320846081, + 0.002380021382123232, + 0.0022654985077679157 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 1, + 0 + ], + "selected_reward": -0.016074467450380325 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of a weathered, red-painted metal surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is there a bright hotspot on the semi-gloss finish caused by harsh, direct light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a paint blister rise from the surface into a taut hemispherical shape before cracking?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do fractured paint fragments curl upward to reveal dark, matte-polished gunmetal gray metal beneath?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera positioned at a slightly higher angle observing the scene?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0015151937259361148 + ], + "opposite_logprob_tokens": [ + -6.5015153884887695 + ], + "logprob_reward": -0.0015151937259361148, + "yes_no_margin_reward": 6.5, + "token_credit_reward": -0.0015151937259361148, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12267903238534927, + 0.10165566205978394, + 0.27972114086151123, + 0.10796774178743362, + 0.10529427975416183, + 0.06002040207386017, + 0.06848389655351639, + 0.04942231997847557, + 0.03541528433561325, + 0.0693402960896492 + ], + "credit_map": [ + 0.011170684359967709, + 0.01174652948975563, + 0.008884599432349205, + 0.012663133442401886, + 0.020433466881513596, + 0.033055923879146576, + 0.01203174702823162, + 0.012692928314208984, + 0.015589134767651558, + 0.007761992048472166, + 0.013951806351542473, + 0.008016540668904781, + 0.009297950193285942, + 0.023354416713118553, + 0.011262024752795696, + 0.012421786785125732, + 0.07630760222673416, + 0.015317612327635288, + 0.013359606266021729, + 0.012783481739461422, + 0.018802620470523834, + 0.1080121099948883, + 0.013264814391732216, + 0.021873246878385544, + 0.03186618536710739, + 0.00470711337402463, + 0.012108049355447292, + 0.00526607409119606, + 0.004626852925866842, + 0.022149046882987022, + 0.00632735900580883, + 0.020917043089866638, + 0.015492449514567852, + 0.00664433091878891, + 0.016263781115412712, + 0.008251041173934937, + 0.006458755116909742, + 0.014543814584612846, + 0.012532743625342846, + 0.02510734833776951, + 0.0122897420078516, + 0.004889066331088543, + 0.009640752337872982, + 0.0063619595021009445, + 0.005142954643815756, + 0.009341754019260406, + 0.005188492592424154, + 0.007165670394897461, + 0.02146119251847267, + 0.0029496049974113703, + 0.004798219073563814, + 0.003446177812293172, + 0.005510816350579262, + 0.021119169890880585, + 0.0042872000485658646, + 0.00491150701418519, + 0.0038182553835213184, + 0.0046857078559696674, + 0.005361204035580158, + 0.008667447604238987, + 0.006058053113520145, + 0.007124374620616436, + 0.00714279618114233, + 0.006564475130289793, + 0.0030749477446079254, + 0.0032778496388345957, + 0.004624063149094582, + 0.004406462423503399, + 0.005563313141465187, + 0.00423003314062953, + 0.004703871440142393, + 0.005534736439585686, + 0.0128810228779912, + 0.0062511214055120945, + 0.006357814185321331, + 0.00806332752108574, + 0.006891260389238596, + 0.017527764663100243, + 0.005494802724570036, + 0.005873176269233227 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 0, + 3 + ], + "selected_reward": -0.0015151937259361148 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of a weathered, red-painted metal surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is there a bright hotspot on the semi-gloss finish caused by harsh, direct light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a paint blister rise from the surface into a taut hemispherical shape before cracking?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do fractured paint fragments curl upward to reveal dark, matte-polished gunmetal gray metal beneath?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera positioned at a slightly higher angle observing the scene?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.023289820179343224 + ], + "opposite_logprob_tokens": [ + -3.773289918899536 + ], + "logprob_reward": -0.023289820179343224, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.023289820179343224, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1405077427625656, + 0.10464567691087723, + 0.36765244603157043, + 0.09670664370059967, + 0.04264303669333458, + 0.02463088370859623, + 0.038694679737091064, + 0.028190413489937782, + 0.09385522454977036, + 0.06247315555810928 + ], + "credit_map": [ + 0.024532757699489594, + 0.01367750484496355, + 0.02070085145533085, + 0.012369812466204166, + 0.012701808474957943, + 0.02498014084994793, + 0.014767829328775406, + 0.016777055338025093, + 0.019753333181142807, + 0.007296261377632618, + 0.007721449248492718, + 0.007146956864744425, + 0.010028624907135963, + 0.0327942930161953, + 0.009030153974890709, + 0.010874614119529724, + 0.10311849415302277, + 0.018207380548119545, + 0.01730979233980179, + 0.011936054565012455, + 0.0230437982827425, + 0.14792592823505402, + 0.015690024942159653, + 0.030421024188399315, + 0.02648025192320347, + 0.004183266311883926, + 0.008934929035604, + 0.00362083176150918, + 0.0043006716296076775, + 0.03477948158979416, + 0.0044568562880158424, + 0.009950362145900726, + 0.0038981824181973934, + 0.0019070846028625965, + 0.004137448035180569, + 0.0030049020424485207, + 0.0025584131944924593, + 0.008828610181808472, + 0.004022127948701382, + 0.01428627036511898, + 0.0037059085443615913, + 0.001948826014995575, + 0.003396488493308425, + 0.002477872185409069, + 0.002795949811115861, + 0.003574779024347663, + 0.0027866377495229244, + 0.003944424446672201, + 0.012855250388383865, + 0.001485251123085618, + 0.0027473156806081533, + 0.001658857217989862, + 0.002153148176148534, + 0.014046885073184967, + 0.0016751944785937667, + 0.0020727827213704586, + 0.007029340136796236, + 0.0014949612086638808, + 0.005562011618167162, + 0.0031591642182320356, + 0.002547268522903323, + 0.003535430645570159, + 0.002368845511227846, + 0.00249339547008276, + 0.021488642320036888, + 0.0034182905219495296, + 0.0052777561359107494, + 0.008166812360286713, + 0.0075034876354038715, + 0.03294093906879425, + 0.0062128580175340176, + 0.008846446871757507, + 0.013905821368098259, + 0.0035807036329060793, + 0.0045682769268751144, + 0.006475889589637518, + 0.004403115715831518, + 0.01929827779531479, + 0.004412519745528698, + 0.005828553345054388 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 0, + 1 + ], + "selected_reward": -0.023289820179343224 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "static, extreme close-up captures a realistic rendering of a weathered, red-painted metal surface gleaming in intense, concentrated heat. The light is harsh and direct, creating a bright hotspot on the semi-gloss finish. From a slightly higher angle, the camera observes a small portion of paint begin to expand, slowly and deliberately rising from a flat surface into a taut, hemispherical blister, its surface stretching and thinning, the color subtly fading at its apex. A sudden, almost imperceptible vibration rips through the metal, and in an instant, a network of hairline cracks splits apart within the fragile, heat-baked blister. In a mesmerizing slow-motion sequence, the fractured paint fragments begin to curl upward and away from the breaking point, their sharp edges lifting like burnt parchment to reveal the dark, matte-polished gunmetal gray of the original metal beneath, creating an atmosphere of intense, clinical observation focused on the physics of material decay.", + "seed": 31849, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000034/group_11/index_3203_seed_31849_dp_23_rank_46_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 11, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.024770016549155116, + "vlm_logprob": -0.024770016549155116, + "vlm_yes_no_margin": 4.325, + "vlm_token_credit": -0.024770016549155116, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of a weathered, red-painted metal surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is there a bright hotspot on the semi-gloss finish caused by harsh, direct light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a paint blister rise from the surface into a taut hemispherical shape before cracking?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do fractured paint fragments curl upward to reveal dark, matte-polished gunmetal gray metal beneath?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera positioned at a slightly higher angle observing the scene?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.010131007060408592 + ], + "opposite_logprob_tokens": [ + -6.010130882263184 + ], + "logprob_reward": -0.010131007060408592, + "yes_no_margin_reward": 6.0, + "token_credit_reward": -0.010131007060408592, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1420091837644577, + 0.06252201646566391, + 0.06723707914352417, + 0.11128532886505127, + 0.034133702516555786, + 0.029969502240419388, + 0.09518521279096603, + 0.04344971850514412, + 0.357507586479187, + 0.05670066922903061 + ], + "credit_map": [ + 0.013363898731768131, + 0.014192342758178711, + 0.013630339875817299, + 0.011906098574399948, + 0.02883540838956833, + 0.02997913956642151, + 0.015082305297255516, + 0.01501965057104826, + 0.007294199895113707, + 0.005556746385991573, + 0.005835382733494043, + 0.005670086480677128, + 0.005631749052554369, + 0.010259916074573994, + 0.012159503996372223, + 0.010114433243870735, + 0.007493170909583569, + 0.006913068704307079, + 0.006135039962828159, + 0.007670722436159849, + 0.006061856169253588, + 0.00986743625253439, + 0.011194148100912571, + 0.011901645921170712, + 0.014694014564156532, + 0.010012760758399963, + 0.013109738007187843, + 0.011225833557546139, + 0.0064891912043094635, + 0.017448941245675087, + 0.012723684310913086, + 0.025581175461411476, + 0.002993186702951789, + 0.004107949323952198, + 0.002674822462722659, + 0.0035786679945886135, + 0.0034853590186685324, + 0.0050435452722013, + 0.004635489545762539, + 0.007614686153829098, + 0.004509247839450836, + 0.003414009464904666, + 0.0035553527995944023, + 0.003075605258345604, + 0.0028909575194120407, + 0.004712272435426712, + 0.002899966901168227, + 0.004912094213068485, + 0.009214099496603012, + 0.007299225311726332, + 0.00698806531727314, + 0.02215804159641266, + 0.005382575560361147, + 0.026825645938515663, + 0.010345243848860264, + 0.0069723306223750114, + 0.004643231630325317, + 0.005686137359589338, + 0.004327616188675165, + 0.006381274200975895, + 0.0051077790558338165, + 0.003161120228469372, + 0.010257900692522526, + 0.003884663339704275, + 0.022201165556907654, + 0.013591788709163666, + 0.13305434584617615, + 0.10564500093460083, + 0.006300793029367924, + 0.03102070838212967, + 0.02261863835155964, + 0.023075150325894356, + 0.0042633116245269775, + 0.007211449556052685, + 0.009130621328949928, + 0.010433808900415897, + 0.0036872285418212414, + 0.006397012155503035, + 0.0070237270556390285, + 0.008553512394428253 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 0, + 3 + ], + "selected_reward": -0.010131007060408592 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of a weathered, red-painted metal surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is there a bright hotspot on the semi-gloss finish caused by harsh, direct light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a paint blister rise from the surface into a taut hemispherical shape before cracking?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do fractured paint fragments curl upward to reveal dark, matte-polished gunmetal gray metal beneath?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera positioned at a slightly higher angle observing the scene?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04869383946061134 + ], + "opposite_logprob_tokens": [ + -3.048693895339966 + ], + "logprob_reward": -0.04869383946061134, + "yes_no_margin_reward": 3.0, + "token_credit_reward": -0.04869383946061134, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.20320984721183777, + 0.12040462344884872, + 0.0821511447429657, + 0.0823841318488121, + 0.03859630972146988, + 0.03993912786245346, + 0.04620540887117386, + 0.03445727378129959, + 0.304795503616333, + 0.04785662144422531 + ], + "credit_map": [ + 0.017071476206183434, + 0.013134916312992573, + 0.015415086410939693, + 0.019030876457691193, + 0.029767325147986412, + 0.049684107303619385, + 0.02305615320801735, + 0.03604990616440773, + 0.011246641166508198, + 0.013307349756360054, + 0.014984565787017345, + 0.008642572909593582, + 0.014236031100153923, + 0.018007079139351845, + 0.02467605099081993, + 0.015304332599043846, + 0.009735739789903164, + 0.011088402010500431, + 0.009383844211697578, + 0.008851341903209686, + 0.006746298633515835, + 0.011509544216096401, + 0.008905167691409588, + 0.01593080163002014, + 0.010828832164406776, + 0.010287820361554623, + 0.00896107591688633, + 0.008486769162118435, + 0.0049375989474356174, + 0.012088298797607422, + 0.009653829969465733, + 0.01713990978896618, + 0.0037224190309643745, + 0.004175214562565088, + 0.004409116692841053, + 0.0045759896747767925, + 0.005223001819103956, + 0.005155361723154783, + 0.005628352984786034, + 0.0057068513706326485, + 0.005053797736763954, + 0.005104099400341511, + 0.006032790523022413, + 0.004720068536698818, + 0.002735208487138152, + 0.0057392483577132225, + 0.003942559473216534, + 0.006611356977373362, + 0.005584615282714367, + 0.005035891197621822, + 0.005634871311485767, + 0.010909031145274639, + 0.003190102521330118, + 0.006072785705327988, + 0.005497663281857967, + 0.004280447028577328, + 0.003211385803297162, + 0.0032368572428822517, + 0.0045981016010046005, + 0.008310467004776001, + 0.003088222350925207, + 0.002353067509829998, + 0.005662280600517988, + 0.003996893297880888, + 0.00855239201337099, + 0.011452210135757923, + 0.026989569887518883, + 0.10524559020996094, + 0.011742422357201576, + 0.009074386209249496, + 0.11088840663433075, + 0.02085050381720066, + 0.005006982013583183, + 0.004261464346200228, + 0.008338648825883865, + 0.006475349422544241, + 0.0031197962816804647, + 0.006957920268177986, + 0.006093253847211599, + 0.007603204343467951 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 0, + 1 + ], + "selected_reward": -0.04869383946061134 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of a weathered, red-painted metal surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is there a bright hotspot on the semi-gloss finish caused by harsh, direct light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a paint blister rise from the surface into a taut hemispherical shape before cracking?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do fractured paint fragments curl upward to reveal dark, matte-polished gunmetal gray metal beneath?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera positioned at a slightly higher angle observing the scene?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02330123446881771 + ], + "opposite_logprob_tokens": [ + -3.773301124572754 + ], + "logprob_reward": -0.02330123446881771, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.02330123446881771, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10164126753807068, + 0.15198804438114166, + 0.05402513965964317, + 0.031119372695684433, + 0.034662678837776184, + 0.0408000573515892, + 0.034344617277383804, + 0.10551019012928009, + 0.4104683995246887, + 0.03544033318758011 + ], + "credit_map": [ + 0.007855444215238094, + 0.009828845039010048, + 0.009628045372664928, + 0.010759472846984863, + 0.0183876845985651, + 0.015818018466234207, + 0.013332943432033062, + 0.016030803322792053, + 0.010943261906504631, + 0.009412580169737339, + 0.012107986025512218, + 0.021577272564172745, + 0.01256596203893423, + 0.02025066502392292, + 0.04506569355726242, + 0.020064599812030792, + 0.004888567142188549, + 0.007193677127361298, + 0.005782824940979481, + 0.005839048884809017, + 0.004569725599139929, + 0.007696229498833418, + 0.009462052956223488, + 0.008593003265559673, + 0.00533805089071393, + 0.0028138740453869104, + 0.0029837931506335735, + 0.0034187494311481714, + 0.00229755905456841, + 0.004113460890948772, + 0.004434776958078146, + 0.005719106178730726, + 0.004482969641685486, + 0.002200397429987788, + 0.002625236287713051, + 0.002930013695731759, + 0.0029309322126209736, + 0.006273036822676659, + 0.003476819721981883, + 0.00974326953291893, + 0.004869794473052025, + 0.005560666788369417, + 0.004459751304239035, + 0.00512318080291152, + 0.006095826625823975, + 0.004559610970318317, + 0.006268548779189587, + 0.0038626755122095346, + 0.003328528720885515, + 0.004880051128566265, + 0.003562087658792734, + 0.004082404542714357, + 0.00425054132938385, + 0.005337141919881105, + 0.005110191646963358, + 0.0037936666049063206, + 0.010900075547397137, + 0.01147378794848919, + 0.01111884880810976, + 0.02482670359313488, + 0.00918380357325077, + 0.008649577386677265, + 0.0167522095143795, + 0.012605172581970692, + 0.02204594947397709, + 0.010361515916883945, + 0.0820520669221878, + 0.0786895900964737, + 0.017023075371980667, + 0.009898814372718334, + 0.1604209989309311, + 0.029976308345794678, + 0.002296523889526725, + 0.0036999450530856848, + 0.0043449546210467815, + 0.006950880866497755, + 0.0027595090214163065, + 0.003648822195827961, + 0.006442977115511894, + 0.005296716932207346 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 1, + 7 + ], + "selected_reward": -0.02330123446881771 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of a weathered, red-painted metal surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is there a bright hotspot on the semi-gloss finish caused by harsh, direct light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a paint blister rise from the surface into a taut hemispherical shape before cracking?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do fractured paint fragments curl upward to reveal dark, matte-polished gunmetal gray metal beneath?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera positioned at a slightly higher angle observing the scene?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0036206187214702368 + ], + "opposite_logprob_tokens": [ + -5.628620624542236 + ], + "logprob_reward": -0.0036206187214702368, + "yes_no_margin_reward": 5.625, + "token_credit_reward": -0.0036206187214702368, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11231144517660141, + 0.14953690767288208, + 0.06769097596406937, + 0.05455468222498894, + 0.041896868497133255, + 0.06691531091928482, + 0.04391223564743996, + 0.08008340001106262, + 0.3399025499820709, + 0.043195564299821854 + ], + "credit_map": [ + 0.008512553758919239, + 0.012457573786377907, + 0.007906033657491207, + 0.012166708707809448, + 0.017826946452260017, + 0.019932355731725693, + 0.01650026999413967, + 0.017009010538458824, + 0.011020089499652386, + 0.013753125444054604, + 0.015176817774772644, + 0.02290480025112629, + 0.010436673648655415, + 0.018727833405137062, + 0.03595399484038353, + 0.02156359702348709, + 0.0054132672958076, + 0.008153557777404785, + 0.004663598258048296, + 0.01274308655411005, + 0.007290965877473354, + 0.009260439313948154, + 0.013446816243231297, + 0.006719247438013554, + 0.007260902784764767, + 0.00411897711455822, + 0.004505458287894726, + 0.004852452781051397, + 0.003588606370612979, + 0.009175320155918598, + 0.008282765746116638, + 0.012770201079547405, + 0.003893250832334161, + 0.004693777300417423, + 0.0044749281369149685, + 0.005388602614402771, + 0.005132465623319149, + 0.005899230018258095, + 0.005535643082112074, + 0.006878974381834269, + 0.006195737048983574, + 0.009010681882500648, + 0.006838291883468628, + 0.007747418712824583, + 0.012076901271939278, + 0.007561969570815563, + 0.010450783185660839, + 0.0070335338823497295, + 0.004355959128588438, + 0.005547935143113136, + 0.004430514294654131, + 0.006244785618036985, + 0.005421164445579052, + 0.008875638246536255, + 0.005335637368261814, + 0.003700606757774949, + 0.007045161444693804, + 0.007609778083860874, + 0.007948840036988258, + 0.021621983498334885, + 0.00538838654756546, + 0.007170779164880514, + 0.012081465683877468, + 0.01121701393276453, + 0.019323129206895828, + 0.010448633693158627, + 0.07059301435947418, + 0.07192783802747726, + 0.014213042333722115, + 0.01094325352460146, + 0.12191353738307953, + 0.020540114492177963, + 0.003242711303755641, + 0.004159944597631693, + 0.005222600884735584, + 0.007241158280521631, + 0.002792831277474761, + 0.003977350424975157, + 0.008739490993320942, + 0.00781947746872902 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 1, + 0 + ], + "selected_reward": -0.0036206187214702368 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of a weathered, red-painted metal surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is there a bright hotspot on the semi-gloss finish caused by harsh, direct light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a paint blister rise from the surface into a taut hemispherical shape before cracking?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do fractured paint fragments curl upward to reveal dark, matte-polished gunmetal gray metal beneath?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera positioned at a slightly higher angle observing the scene?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0381033830344677 + ], + "opposite_logprob_tokens": [ + -3.2881033420562744 + ], + "logprob_reward": -0.0381033830344677, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.0381033830344677, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13518595695495605, + 0.07886987179517746, + 0.09684953838586807, + 0.11212270706892014, + 0.03959996625781059, + 0.04112769663333893, + 0.0366741307079792, + 0.032816775143146515, + 0.3638601303100586, + 0.06289327144622803 + ], + "credit_map": [ + 0.013672687113285065, + 0.014160881750285625, + 0.013552432879805565, + 0.015399373136460781, + 0.014420727267861366, + 0.01916370540857315, + 0.01968756504356861, + 0.02512858435511589, + 0.007793030701577663, + 0.010455738753080368, + 0.008207682520151138, + 0.0059190778993070126, + 0.007183815352618694, + 0.010402888059616089, + 0.019513478502631187, + 0.009394168853759766, + 0.010399380698800087, + 0.017207114025950432, + 0.011585528030991554, + 0.010154115036129951, + 0.007776584941893816, + 0.01277348306030035, + 0.012117731384932995, + 0.0148356007412076, + 0.015158623456954956, + 0.008155766874551773, + 0.011315546929836273, + 0.01201113685965538, + 0.006817672401666641, + 0.017943192273378372, + 0.014069722034037113, + 0.02665105275809765, + 0.0036834245547652245, + 0.00564672751352191, + 0.004496983252465725, + 0.00424951920285821, + 0.004485445097088814, + 0.0055845919996500015, + 0.00552774453535676, + 0.005925529170781374, + 0.005993783939629793, + 0.004887454677373171, + 0.005455940030515194, + 0.0056833806447684765, + 0.004903801251202822, + 0.005353313405066729, + 0.004400955978780985, + 0.004449067637324333, + 0.006366950925439596, + 0.003887262661010027, + 0.005086712539196014, + 0.00538661191239953, + 0.003688535187393427, + 0.004504977725446224, + 0.004686937667429447, + 0.0030661411583423615, + 0.002324103843420744, + 0.0034564484376460314, + 0.0034156031906604767, + 0.006819943431764841, + 0.0030894444789737463, + 0.0020656660199165344, + 0.008177587762475014, + 0.0034679798409342766, + 0.013542440719902515, + 0.013307582587003708, + 0.08327112346887589, + 0.12274829298257828, + 0.00973690114915371, + 0.015639549121260643, + 0.086325503885746, + 0.019288722425699234, + 0.005739351734519005, + 0.007190782576799393, + 0.011280517093837261, + 0.012621025554835796, + 0.003995389211922884, + 0.004673531278967857, + 0.009272627532482147, + 0.008120047859847546 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 0, + 3 + ], + "selected_reward": -0.0381033830344677 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000034/group_12/index_tensor([3907, 3907])_group_12_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000034/group_12/index_tensor([3907, 3907])_group_12_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..2979aec6f8a2c9e4cbfdcf92005b8861f140a5e8 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000034/group_12/index_tensor([3907, 3907])_group_12_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "cinematic, low-angle tracking shot follows a cyclist as he navigates a futuristic city at dusk. The young rider, clad in a seamless, skin-tight black cycling suit made of high-tech synthetic material, accentuates his muscular form, honed by years of training. He wears an aerodynamic helmet and gazes intently ahead, his jaw set in a resolute concentration. The connected, smart shared bike he rides boasts a sleek, one-piece frame crafted from matte black carbon fiber, free of any unnecessary wires. From front to back, thin, data-like streams of cyan-blue light pulse rhythmically, flowing along the frame's contours and forming mesmerizing halos around the rapidly spinning wheels. In the background, towering skyscrapers, adorned with giant holographic billboards, intertwine their vibrant neon lights with the deep twilight sky, casting distorted and iridescent reflections on the rain-soaked asphalt. The rider leans forward, pedaling with a steady, powerful rhythm. The light from the bike's data stream illuminates the movements of his legs with exceptional clarity, leaving a fleeting blue trail on the road. The 8K ultra-high-definition camera perfectly captures every detail—from the tiny water droplets splashed by the wheels to the light reflecting off the rider's muscular tones—creating a dynamic and hopeful atmosphere that blends cutting-edge technology with human perseverance.", + "seed": 3066, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000034/group_12/index_3907_seed_3066_dp_24_rank_48_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 12, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.038915779627859595, + "vlm_logprob": -0.038915779627859595, + "vlm_yes_no_margin": 3.925, + "vlm_token_credit": -0.038915779627859595, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot taken from a low angle while tracking the cyclist?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the cyclist wearing a seamless, skin-tight black cycling suit made of high-tech synthetic material?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the bike feature a sleek, one-piece matte black carbon fiber frame with thin cyan-blue data-like light streams pulsing along it?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a futuristic city at dusk with towering skyscrapers, holographic billboards, and rain-soaked asphalt reflecting neon lights?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the rider leaning forward and pedaling with a steady, powerful rhythm?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.059362731873989105 + ], + "opposite_logprob_tokens": [ + -3.5593626499176025 + ], + "logprob_reward": -0.059362731873989105, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.059362731873989105, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.16025029122829437, + 0.09009800851345062, + 0.1080731749534607, + 0.10527559369802475, + 0.12170074880123138, + 0.05805385112762451, + 0.09846064448356628, + 0.04857325553894043, + 0.09219707548618317, + 0.11731734871864319 + ], + "credit_map": [ + 0.04366878420114517, + 0.013929404318332672, + 0.032263025641441345, + 0.008844038471579552, + 0.00888042338192463, + 0.012924186885356903, + 0.0162234865128994, + 0.023516938090324402, + 0.015950331464409828, + 0.010515968315303326, + 0.013596490025520325, + 0.008352091535925865, + 0.007644881960004568, + 0.0137812914326787, + 0.008395139127969742, + 0.011861813254654408, + 0.019776087254285812, + 0.015365813858807087, + 0.01690566912293434, + 0.007951664738357067, + 0.01240579690784216, + 0.015643661841750145, + 0.009122337214648724, + 0.010902135632932186, + 0.013630179688334465, + 0.011969095095992088, + 0.017061050981283188, + 0.009773983620107174, + 0.005719061475247145, + 0.01827920228242874, + 0.010214789770543575, + 0.018628230318427086, + 0.017454029992222786, + 0.018963288515806198, + 0.015504744835197926, + 0.012873376719653606, + 0.011431973427534103, + 0.014513490721583366, + 0.012479186058044434, + 0.018480660393834114, + 0.0054565477184951305, + 0.00560715002939105, + 0.009096442721784115, + 0.005297896917909384, + 0.00718328170478344, + 0.009858283214271069, + 0.008385304361581802, + 0.007168943528085947, + 0.01168234832584858, + 0.010028569027781487, + 0.01351514644920826, + 0.005601791664958, + 0.012849711813032627, + 0.02328014001250267, + 0.0089911799877882, + 0.012511755339801311, + 0.010486973449587822, + 0.005824113264679909, + 0.008320796303451061, + 0.005098708905279636, + 0.00586885679513216, + 0.005851960275322199, + 0.00331937987357378, + 0.003802464809268713, + 0.007319263648241758, + 0.008880941197276115, + 0.021588465198874474, + 0.005561568774282932, + 0.0058015682734549046, + 0.033334992825984955, + 0.0032966178841888905, + 0.006413656752556562, + 0.018282772973179817, + 0.01427711546421051, + 0.027685515582561493, + 0.015573610551655293, + 0.011710792779922485, + 0.009891792200505733, + 0.009927856735885143, + 0.009967891499400139 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 9 + ], + "selected_reward": -0.059362731873989105 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot taken from a low angle while tracking the cyclist?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the cyclist wearing a seamless, skin-tight black cycling suit made of high-tech synthetic material?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the bike feature a sleek, one-piece matte black carbon fiber frame with thin cyan-blue data-like light streams pulsing along it?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a futuristic city at dusk with towering skyscrapers, holographic billboards, and rain-soaked asphalt reflecting neon lights?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the rider leaning forward and pedaling with a steady, powerful rhythm?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.10039108246564865 + ], + "opposite_logprob_tokens": [ + -2.350391149520874 + ], + "logprob_reward": -0.10039108246564865, + "yes_no_margin_reward": 2.25, + "token_credit_reward": -0.10039108246564865, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10316844284534454, + 0.11512947082519531, + 0.05272122472524643, + 0.023472988978028297, + 0.049790944904088974, + 0.2175392359495163, + 0.03151601180434227, + 0.03518912196159363, + 0.33869585394859314, + 0.032776638865470886 + ], + "credit_map": [ + 0.025006791576743126, + 0.012105342000722885, + 0.018525684252381325, + 0.009872972033917904, + 0.009230319410562515, + 0.009150763973593712, + 0.01076759397983551, + 0.008508981205523014, + 0.017182480543851852, + 0.01713668182492256, + 0.018692394718527794, + 0.009177436120808125, + 0.011371864005923271, + 0.01700211688876152, + 0.011989063583314419, + 0.012577440589666367, + 0.007042899262160063, + 0.008450862020254135, + 0.006748346146196127, + 0.00479929381981492, + 0.004912548698484898, + 0.008033626712858677, + 0.005425236187875271, + 0.00730841513723135, + 0.0031071940902620554, + 0.0027676383033394814, + 0.004673035815358162, + 0.002478850306943059, + 0.001797437435016036, + 0.0029711846727877855, + 0.0022936477325856686, + 0.0033840027172118425, + 0.007828508503735065, + 0.006601938512176275, + 0.008626663126051426, + 0.004251522943377495, + 0.00472733611240983, + 0.010267033241689205, + 0.004485284443944693, + 0.0030026580207049847, + 0.017515284940600395, + 0.015182164497673512, + 0.010784718208014965, + 0.011035428382456303, + 0.06917333602905273, + 0.06724432110786438, + 0.01641315408051014, + 0.010190845467150211, + 0.0038213219959288836, + 0.0033282337244600058, + 0.004333378281444311, + 0.002560423919931054, + 0.003956950735300779, + 0.005719693843275309, + 0.003846745938062668, + 0.003949266858398914, + 0.00526325823739171, + 0.008736097253859043, + 0.004508967977017164, + 0.002901853993535042, + 0.004817447625100613, + 0.004236039239913225, + 0.0020667649805545807, + 0.002658698009327054, + 0.01913033239543438, + 0.0230565145611763, + 0.08925852924585342, + 0.017420977354049683, + 0.022396204993128777, + 0.1344258338212967, + 0.010986178182065487, + 0.022021297365427017, + 0.005070587620139122, + 0.003514604875817895, + 0.008633189834654331, + 0.0039274259470403194, + 0.002954778727144003, + 0.0038475203327834606, + 0.002470817882567644, + 0.002357716904953122 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 5, + 1 + ], + "selected_reward": -0.10039108246564865 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot taken from a low angle while tracking the cyclist?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the cyclist wearing a seamless, skin-tight black cycling suit made of high-tech synthetic material?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the bike feature a sleek, one-piece matte black carbon fiber frame with thin cyan-blue data-like light streams pulsing along it?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a futuristic city at dusk with towering skyscrapers, holographic billboards, and rain-soaked asphalt reflecting neon lights?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the rider leaning forward and pedaling with a steady, powerful rhythm?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01254394743591547 + ], + "opposite_logprob_tokens": [ + -4.38754415512085 + ], + "logprob_reward": -0.01254394743591547, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.01254394743591547, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13502860069274902, + 0.06105133891105652, + 0.051402751356363297, + 0.04697674140334129, + 0.04205595329403877, + 0.14899396896362305, + 0.02764374203979969, + 0.022721627727150917, + 0.33730146288871765, + 0.12682387232780457 + ], + "credit_map": [ + 0.027160916477441788, + 0.017318762838840485, + 0.026310566812753677, + 0.009653721004724503, + 0.015100146643817425, + 0.010511490516364574, + 0.01611616462469101, + 0.012856832705438137, + 0.010141517035663128, + 0.008105109445750713, + 0.00963926687836647, + 0.00456964410841465, + 0.004682373721152544, + 0.01146574318408966, + 0.004979065153747797, + 0.007468618452548981, + 0.007898678071796894, + 0.004458459559828043, + 0.00961263943463564, + 0.004326609894633293, + 0.0036692582070827484, + 0.010529791004955769, + 0.0050038970075547695, + 0.005903415381908417, + 0.0075350129045546055, + 0.008604559116065502, + 0.0062573072500526905, + 0.003916571382433176, + 0.004948224872350693, + 0.005403495393693447, + 0.00494414335116744, + 0.0053674266673624516, + 0.00617006653919816, + 0.00448952428996563, + 0.00872050877660513, + 0.0024132796097546816, + 0.004535369575023651, + 0.008505288511514664, + 0.003825741121545434, + 0.003396175568923354, + 0.010898515582084656, + 0.010994414798915386, + 0.009408089332282543, + 0.006615245249122381, + 0.049959830939769745, + 0.044884152710437775, + 0.01037487480789423, + 0.0058588506653904915, + 0.004075888078659773, + 0.00316948932595551, + 0.004142357036471367, + 0.0021537160500884056, + 0.0028582834638655186, + 0.005027011036872864, + 0.002281145891174674, + 0.003935852088034153, + 0.0031746625900268555, + 0.004833270329982042, + 0.0029063725378364325, + 0.0016327560879290104, + 0.0032861707732081413, + 0.002627665176987648, + 0.0023934131022542715, + 0.0018673165468499064, + 0.020240332931280136, + 0.021740851923823357, + 0.09049530327320099, + 0.017494911327958107, + 0.020107978954911232, + 0.13423429429531097, + 0.011109791696071625, + 0.021877985447645187, + 0.02122749388217926, + 0.01589115895330906, + 0.030530648306012154, + 0.015397361479699612, + 0.013433614745736122, + 0.011992642655968666, + 0.009236453101038933, + 0.009114492684602737 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 5, + 0 + ], + "selected_reward": -0.01254394743591547 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot taken from a low angle while tracking the cyclist?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the cyclist wearing a seamless, skin-tight black cycling suit made of high-tech synthetic material?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the bike feature a sleek, one-piece matte black carbon fiber frame with thin cyan-blue data-like light streams pulsing along it?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a futuristic city at dusk with towering skyscrapers, holographic billboards, and rain-soaked asphalt reflecting neon lights?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the rider leaning forward and pedaling with a steady, powerful rhythm?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01818777248263359 + ], + "opposite_logprob_tokens": [ + -4.018187999725342 + ], + "logprob_reward": -0.01818777248263359, + "yes_no_margin_reward": 4.0, + "token_credit_reward": -0.01818777248263359, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.02502094954252243, + 0.024557383731007576, + 0.05142258107662201, + 0.05612226575613022, + 0.019047250971198082, + 0.6714519262313843, + 0.013323317281901836, + 0.020161721855401993, + 0.02895374223589897, + 0.08993884176015854 + ], + "credit_map": [ + 0.0064660548232495785, + 0.0018395911902189255, + 0.005797809921205044, + 0.0020315302535891533, + 0.0019701849669218063, + 0.0027974885888397694, + 0.0016127575654536486, + 0.002505535027012229, + 0.0044951848685741425, + 0.0021589789539575577, + 0.003289451589807868, + 0.0016177059151232243, + 0.001971195451915264, + 0.007030258886516094, + 0.0017071247566491365, + 0.0022874854039400816, + 0.009944602847099304, + 0.003019527532160282, + 0.009923987090587616, + 0.002535342937335372, + 0.003339046146720648, + 0.01608845591545105, + 0.002522544004023075, + 0.004049085546284914, + 0.008050467818975449, + 0.004629487171769142, + 0.008860127069056034, + 0.003474608063697815, + 0.00371163128875196, + 0.01823834329843521, + 0.0032258278224617243, + 0.005931779742240906, + 0.0027460064738988876, + 0.0016530968714505434, + 0.004164541140198708, + 0.001275810063816607, + 0.0016826351638883352, + 0.003498159581795335, + 0.001786489738151431, + 0.0022405143827199936, + 0.05075005814433098, + 0.04289348050951958, + 0.04254690930247307, + 0.021507082507014275, + 0.22240377962589264, + 0.22488568723201752, + 0.044138580560684204, + 0.022326387465000153, + 0.0029760259203612804, + 0.0013263302389532328, + 0.0028710891492664814, + 0.0007654191576875746, + 0.001277895993553102, + 0.0018667519325390458, + 0.0009980543982237577, + 0.0012417511316016316, + 0.004475154913961887, + 0.002772298641502857, + 0.0036505667958408594, + 0.0016996086342260242, + 0.0020413047168403864, + 0.0026317129377275705, + 0.0013709794729948044, + 0.0015200990019366145, + 0.0020393410231918097, + 0.0020455922931432724, + 0.007378352805972099, + 0.0015438876580446959, + 0.0017478837398812175, + 0.011308997869491577, + 0.0009719079243950546, + 0.0019177822396159172, + 0.014048732817173004, + 0.013601505197584629, + 0.022311141714453697, + 0.010924320667982101, + 0.007607864681631327, + 0.010931946337223053, + 0.0049482169561088085, + 0.005565122701227665 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 9, + 3 + ], + "selected_reward": -0.01818777248263359 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot taken from a low angle while tracking the cyclist?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the cyclist wearing a seamless, skin-tight black cycling suit made of high-tech synthetic material?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the bike feature a sleek, one-piece matte black carbon fiber frame with thin cyan-blue data-like light streams pulsing along it?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a futuristic city at dusk with towering skyscrapers, holographic billboards, and rain-soaked asphalt reflecting neon lights?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the rider leaning forward and pedaling with a steady, powerful rhythm?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004093363881111145 + ], + "opposite_logprob_tokens": [ + -5.504093170166016 + ], + "logprob_reward": -0.004093363881111145, + "yes_no_margin_reward": 5.5, + "token_credit_reward": -0.004093363881111145, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1375952810049057, + 0.08033312112092972, + 0.08909785002470016, + 0.06117977201938629, + 0.04016208276152611, + 0.23370981216430664, + 0.060802169144153595, + 0.058275043964385986, + 0.12814196944236755, + 0.11070291697978973 + ], + "credit_map": [ + 0.018690994009375572, + 0.016450027003884315, + 0.01640314795076847, + 0.01218869537115097, + 0.016095856204628944, + 0.010558880865573883, + 0.02193743735551834, + 0.0252702534198761, + 0.011513377539813519, + 0.010223936289548874, + 0.01155635342001915, + 0.006465518847107887, + 0.011589987203478813, + 0.014147783629596233, + 0.008355625905096531, + 0.0064805359579622746, + 0.01473982073366642, + 0.012736978009343147, + 0.014705531299114227, + 0.007650676183402538, + 0.008158024400472641, + 0.014287554658949375, + 0.006510477513074875, + 0.010308787226676941, + 0.00942543987184763, + 0.007385180797427893, + 0.010140417143702507, + 0.0037381616421043873, + 0.004642139188945293, + 0.012623758055269718, + 0.005201259627938271, + 0.008023414760828018, + 0.005565594416111708, + 0.006347587797790766, + 0.006874576210975647, + 0.0035237178672105074, + 0.004164667800068855, + 0.006076030433177948, + 0.003974787425249815, + 0.0036351217422634363, + 0.016364559531211853, + 0.015904467552900314, + 0.01461173314601183, + 0.008176589384675026, + 0.07902411371469498, + 0.07503392547369003, + 0.015910202637314796, + 0.00868422631174326, + 0.005827238783240318, + 0.006914448458701372, + 0.00786584336310625, + 0.003929034806787968, + 0.007548486813902855, + 0.0134239187464118, + 0.006592456251382828, + 0.008700740523636341, + 0.008733565919101238, + 0.013955645263195038, + 0.0059567829594016075, + 0.005060267634689808, + 0.009247448295354843, + 0.00684884749352932, + 0.004080432001501322, + 0.004392055328935385, + 0.007732902187854052, + 0.008540487848222256, + 0.0356399305164814, + 0.006533426698297262, + 0.00808313861489296, + 0.048993151634931564, + 0.004250833299010992, + 0.008368092589080334, + 0.01626160554587841, + 0.014330136589705944, + 0.023619109764695168, + 0.01284262165427208, + 0.014677775092422962, + 0.011392527259886265, + 0.007927494123578072, + 0.009651648811995983 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 0, + 8 + ], + "selected_reward": -0.004093363881111145 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "cinematic, low-angle tracking shot follows a cyclist as he navigates a futuristic city at dusk. The young rider, clad in a seamless, skin-tight black cycling suit made of high-tech synthetic material, accentuates his muscular form, honed by years of training. He wears an aerodynamic helmet and gazes intently ahead, his jaw set in a resolute concentration. The connected, smart shared bike he rides boasts a sleek, one-piece frame crafted from matte black carbon fiber, free of any unnecessary wires. From front to back, thin, data-like streams of cyan-blue light pulse rhythmically, flowing along the frame's contours and forming mesmerizing halos around the rapidly spinning wheels. In the background, towering skyscrapers, adorned with giant holographic billboards, intertwine their vibrant neon lights with the deep twilight sky, casting distorted and iridescent reflections on the rain-soaked asphalt. The rider leans forward, pedaling with a steady, powerful rhythm. The light from the bike's data stream illuminates the movements of his legs with exceptional clarity, leaving a fleeting blue trail on the road. The 8K ultra-high-definition camera perfectly captures every detail—from the tiny water droplets splashed by the wheels to the light reflecting off the rider's muscular tones—creating a dynamic and hopeful atmosphere that blends cutting-edge technology with human perseverance.", + "seed": 3067, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000034/group_12/index_3907_seed_3067_dp_24_rank_48_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 12, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.04949755165725946, + "vlm_logprob": -0.04949755165725946, + "vlm_yes_no_margin": 3.8, + "vlm_token_credit": -0.04949755165725946, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot taken from a low angle while tracking the cyclist?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the cyclist wearing a seamless, skin-tight black cycling suit made of high-tech synthetic material?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the bike feature a sleek, one-piece matte black carbon fiber frame with thin cyan-blue data-like light streams pulsing along it?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a futuristic city at dusk with towering skyscrapers, holographic billboards, and rain-soaked asphalt reflecting neon lights?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the rider leaning forward and pedaling with a steady, powerful rhythm?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04612274840474129 + ], + "opposite_logprob_tokens": [ + -4.4211225509643555 + ], + "logprob_reward": -0.04612274840474129, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.04612274840474129, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10437428206205368, + 0.05886762961745262, + 0.13178667426109314, + 0.12153539061546326, + 0.16650812327861786, + 0.08896706253290176, + 0.0634775459766388, + 0.12404660135507584, + 0.08211784064769745, + 0.05831883102655411 + ], + "credit_map": [ + 0.011387225240468979, + 0.009839503094553947, + 0.0253446102142334, + 0.013521174900233746, + 0.00868216436356306, + 0.008879875764250755, + 0.015310404822230339, + 0.01140932273119688, + 0.007405493408441544, + 0.007157506421208382, + 0.010504825972020626, + 0.0061820209957659245, + 0.005984393879771233, + 0.00643423106521368, + 0.008870899677276611, + 0.006328258663415909, + 0.015656812116503716, + 0.010508297942578793, + 0.03187507018446922, + 0.00830590259283781, + 0.027812696993350983, + 0.016162795946002007, + 0.010146917775273323, + 0.011318190023303032, + 0.016782132908701897, + 0.026846008375287056, + 0.01572621613740921, + 0.010850830934941769, + 0.007490368559956551, + 0.023372020572423935, + 0.008191019296646118, + 0.0122767873108387, + 0.014001613482832909, + 0.009571854025125504, + 0.052515264600515366, + 0.0071845389902591705, + 0.00804686639457941, + 0.01348927617073059, + 0.016032902523875237, + 0.04566580429673195, + 0.020475855097174644, + 0.01043249387294054, + 0.017941899597644806, + 0.007734011858701706, + 0.0061036571860313416, + 0.006929484196007252, + 0.0062021720223128796, + 0.01314748264849186, + 0.007189041003584862, + 0.006262800190597773, + 0.009006117470562458, + 0.013046102598309517, + 0.006920089013874531, + 0.005780140869319439, + 0.007127240765839815, + 0.008146014995872974, + 0.015543618239462376, + 0.013579368591308594, + 0.019241049885749817, + 0.017712727189064026, + 0.014087394811213017, + 0.013159755617380142, + 0.011125377379357815, + 0.019597310572862625, + 0.006010513287037611, + 0.008172521367669106, + 0.006635255180299282, + 0.01520285289734602, + 0.008225028403103352, + 0.007057658862322569, + 0.014783497899770737, + 0.016030509024858475, + 0.007029393687844276, + 0.004081490449607372, + 0.01158080156892538, + 0.005574528593569994, + 0.0060786474496126175, + 0.007124761585146189, + 0.0056905243545770645, + 0.011158683337271214 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 2, + 7 + ], + "selected_reward": -0.04612274840474129 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot taken from a low angle while tracking the cyclist?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the cyclist wearing a seamless, skin-tight black cycling suit made of high-tech synthetic material?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the bike feature a sleek, one-piece matte black carbon fiber frame with thin cyan-blue data-like light streams pulsing along it?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a futuristic city at dusk with towering skyscrapers, holographic billboards, and rain-soaked asphalt reflecting neon lights?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the rider leaning forward and pedaling with a steady, powerful rhythm?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.16043917834758759 + ], + "opposite_logprob_tokens": [ + -1.910439133644104 + ], + "logprob_reward": -0.16043917834758759, + "yes_no_margin_reward": 1.75, + "token_credit_reward": -0.16043917834758759, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1931576132774353, + 0.05282382667064667, + 0.130604088306427, + 0.040647123008966446, + 0.10067316144704819, + 0.0923919603228569, + 0.1309594064950943, + 0.02676979824900627, + 0.19122262299060822, + 0.0407504104077816 + ], + "credit_map": [ + 0.02077900990843773, + 0.020038312301039696, + 0.034586258232593536, + 0.02543330192565918, + 0.029636431485414505, + 0.016594452783465385, + 0.028385214507579803, + 0.01770462468266487, + 0.006859389133751392, + 0.0038963358383625746, + 0.014447050169110298, + 0.004583972506225109, + 0.0035839471966028214, + 0.006082505453377962, + 0.005957086570560932, + 0.007413540966808796, + 0.013691866770386696, + 0.007161966059356928, + 0.04301861673593521, + 0.009344062767922878, + 0.024509770795702934, + 0.011687724851071835, + 0.008620528504252434, + 0.012569545768201351, + 0.003950045444071293, + 0.006189972627907991, + 0.004281886853277683, + 0.0053560687229037285, + 0.004434403497725725, + 0.006737649440765381, + 0.004662274848669767, + 0.005034821107983589, + 0.008547184988856316, + 0.007449677214026451, + 0.032478924840688705, + 0.0062887659296393394, + 0.00871808361262083, + 0.008741917088627815, + 0.007209139410406351, + 0.02123945951461792, + 0.014572056010365486, + 0.009727408178150654, + 0.010961901396512985, + 0.013870315626263618, + 0.008490490727126598, + 0.009988587349653244, + 0.011878518387675285, + 0.012902681715786457, + 0.013952204957604408, + 0.011319412849843502, + 0.02803477644920349, + 0.01258051861077547, + 0.014186359010636806, + 0.01670335978269577, + 0.011704741045832634, + 0.022478027269244194, + 0.002710193395614624, + 0.003581392578780651, + 0.0030584605410695076, + 0.002885609632357955, + 0.004267542622983456, + 0.0035156658850610256, + 0.0035962206311523914, + 0.003154712961986661, + 0.013766174204647541, + 0.01106867752969265, + 0.05345245823264122, + 0.014210287481546402, + 0.01915409415960312, + 0.013878955505788326, + 0.014359124936163425, + 0.05133284628391266, + 0.004179679788649082, + 0.004056756850332022, + 0.005779492203146219, + 0.005248019937425852, + 0.005369231104850769, + 0.004255204927176237, + 0.004282074514776468, + 0.00757994968444109 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 6 + ], + "selected_reward": -0.16043917834758759 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot taken from a low angle while tracking the cyclist?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the cyclist wearing a seamless, skin-tight black cycling suit made of high-tech synthetic material?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the bike feature a sleek, one-piece matte black carbon fiber frame with thin cyan-blue data-like light streams pulsing along it?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a futuristic city at dusk with towering skyscrapers, holographic billboards, and rain-soaked asphalt reflecting neon lights?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the rider leaning forward and pedaling with a steady, powerful rhythm?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.014199766330420971 + ], + "opposite_logprob_tokens": [ + -4.264199733734131 + ], + "logprob_reward": -0.014199766330420971, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.014199766330420971, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1434800922870636, + 0.04391021281480789, + 0.14032359421253204, + 0.06663749366998672, + 0.06605677306652069, + 0.09512883424758911, + 0.07879535853862762, + 0.04675489664077759, + 0.2191140502691269, + 0.09979868680238724 + ], + "credit_map": [ + 0.020181182771921158, + 0.014348473399877548, + 0.017899075523018837, + 0.018866432830691338, + 0.024732155725359917, + 0.012014335952699184, + 0.021230394020676613, + 0.01420802902430296, + 0.006098505109548569, + 0.004691633861511946, + 0.01008196733891964, + 0.0034755030646920204, + 0.004614366218447685, + 0.004170700907707214, + 0.005658614449203014, + 0.005118921399116516, + 0.013936862349510193, + 0.009943715296685696, + 0.04379337280988693, + 0.010077451355755329, + 0.029250016435980797, + 0.011259956285357475, + 0.01026584580540657, + 0.011796370148658752, + 0.00802437774837017, + 0.011882703751325607, + 0.006210563238710165, + 0.007104267366230488, + 0.004872649908065796, + 0.015976358205080032, + 0.004760455805808306, + 0.007806118577718735, + 0.005285460501909256, + 0.0035670544020831585, + 0.019257964566349983, + 0.0038330110255628824, + 0.002947371918708086, + 0.004480562638491392, + 0.005635702982544899, + 0.02104964293539524, + 0.01022374164313078, + 0.01777266524732113, + 0.009209074079990387, + 0.011351053602993488, + 0.007324787322431803, + 0.009825124405324459, + 0.008908946998417377, + 0.020513441413640976, + 0.007344962563365698, + 0.011024627834558487, + 0.011858184821903706, + 0.008574243634939194, + 0.016322480514645576, + 0.009768975898623466, + 0.006468174513429403, + 0.007433711551129818, + 0.004923456348478794, + 0.008186115883290768, + 0.007713085971772671, + 0.00519429100677371, + 0.00540159409865737, + 0.0056364405900239944, + 0.0038250565994530916, + 0.005874857772141695, + 0.016259754076600075, + 0.012011517770588398, + 0.06400883942842484, + 0.019853297621011734, + 0.021729232743382454, + 0.015708904713392258, + 0.014435839839279652, + 0.05510665848851204, + 0.008383461274206638, + 0.0050703310407698154, + 0.028217343613505363, + 0.008619718253612518, + 0.0048594349063932896, + 0.01054797600954771, + 0.007107965182512999, + 0.02699245698750019 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 0, + 2 + ], + "selected_reward": -0.014199766330420971 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot taken from a low angle while tracking the cyclist?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the cyclist wearing a seamless, skin-tight black cycling suit made of high-tech synthetic material?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the bike feature a sleek, one-piece matte black carbon fiber frame with thin cyan-blue data-like light streams pulsing along it?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a futuristic city at dusk with towering skyscrapers, holographic billboards, and rain-soaked asphalt reflecting neon lights?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the rider leaning forward and pedaling with a steady, powerful rhythm?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.014193302020430565 + ], + "opposite_logprob_tokens": [ + -4.264193534851074 + ], + "logprob_reward": -0.014193302020430565, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.014193302020430565, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2944088280200958, + 0.17886808514595032, + 0.09630108624696732, + 0.08245295286178589, + 0.11866119503974915, + 0.026414059102535248, + 0.023864420130848885, + 0.10196390002965927, + 0.03720298781991005, + 0.03986241668462753 + ], + "credit_map": [ + 0.05108695104718208, + 0.022486267611384392, + 0.08058398962020874, + 0.02400849759578705, + 0.020080946385860443, + 0.025524932891130447, + 0.03574404865503311, + 0.034893207252025604, + 0.02901109680533409, + 0.009525828063488007, + 0.05587796866893768, + 0.0106142433360219, + 0.014417216181755066, + 0.02205558307468891, + 0.017770063132047653, + 0.019596101716160774, + 0.011310477741062641, + 0.00802477728575468, + 0.028143294155597687, + 0.006151880137622356, + 0.018886664882302284, + 0.007671916391700506, + 0.0071623255498707294, + 0.008949765004217625, + 0.01098637841641903, + 0.01898527331650257, + 0.007815188728272915, + 0.008065481670200825, + 0.005371278151869774, + 0.017604772001504898, + 0.006033633835613728, + 0.007590950466692448, + 0.010954129509627819, + 0.007695464417338371, + 0.039456624537706375, + 0.006187891121953726, + 0.004824776202440262, + 0.010401328094303608, + 0.005533471703529358, + 0.03360752761363983, + 0.003980044275522232, + 0.003231324488297105, + 0.003693928010761738, + 0.0022674158681184053, + 0.0022863189224153757, + 0.003149133874103427, + 0.003733169985935092, + 0.0040727281011641026, + 0.002323517808690667, + 0.003034026362001896, + 0.004159612115472555, + 0.002732801716774702, + 0.004574780352413654, + 0.002602302934974432, + 0.002002932596951723, + 0.0024344483390450478, + 0.013343295082449913, + 0.03225855529308319, + 0.012965197674930096, + 0.006959549151360989, + 0.007382889743894339, + 0.011057424359023571, + 0.005918304435908794, + 0.012078694067895412, + 0.002115385141223669, + 0.002816863590851426, + 0.003674411913380027, + 0.006211448460817337, + 0.0034230949822813272, + 0.002233905717730522, + 0.009946655482053757, + 0.006781223230063915, + 0.0051263622008264065, + 0.004543856717646122, + 0.004140231758356094, + 0.004043040797114372, + 0.006217877846211195, + 0.005065549165010452, + 0.00389089435338974, + 0.006834608037024736 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 4 + ], + "selected_reward": -0.014193302020430565 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot taken from a low angle while tracking the cyclist?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the cyclist wearing a seamless, skin-tight black cycling suit made of high-tech synthetic material?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the bike feature a sleek, one-piece matte black carbon fiber frame with thin cyan-blue data-like light streams pulsing along it?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a futuristic city at dusk with towering skyscrapers, holographic billboards, and rain-soaked asphalt reflecting neon lights?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the rider leaning forward and pedaling with a steady, powerful rhythm?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.012532763183116913 + ], + "opposite_logprob_tokens": [ + -4.387532711029053 + ], + "logprob_reward": -0.012532763183116913, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.012532763183116913, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.18725067377090454, + 0.04228351637721062, + 0.2103319615125656, + 0.03760164603590965, + 0.10581045597791672, + 0.08645325899124146, + 0.09483379125595093, + 0.04350389167666435, + 0.10460327565670013, + 0.08732753247022629 + ], + "credit_map": [ + 0.018267350271344185, + 0.028873058035969734, + 0.02376243658363819, + 0.031076399609446526, + 0.02473839372396469, + 0.012281366623938084, + 0.029608331620693207, + 0.018643315881490707, + 0.0046382020227611065, + 0.004509355407208204, + 0.011307045817375183, + 0.003755495883524418, + 0.0037183472886681557, + 0.0035513676702976227, + 0.00618585292249918, + 0.004617846570909023, + 0.01857384853065014, + 0.022502213716506958, + 0.054453544318675995, + 0.015792371705174446, + 0.060404274612665176, + 0.012660712003707886, + 0.01163731049746275, + 0.01430766936391592, + 0.006030690856277943, + 0.006365390028804541, + 0.004090307746082544, + 0.005160634405910969, + 0.0035922599490731955, + 0.00491322623565793, + 0.0031711747869849205, + 0.004277956672012806, + 0.008685121312737465, + 0.0066992067731916904, + 0.03220648691058159, + 0.00539500592276454, + 0.005688935983926058, + 0.009390868246555328, + 0.008281736634671688, + 0.029463082551956177, + 0.0076079838909208775, + 0.022245749831199646, + 0.007033709902316332, + 0.008594711311161518, + 0.005811355076730251, + 0.009308678098022938, + 0.007128698751330376, + 0.018722359091043472, + 0.00862060021609068, + 0.008728524670004845, + 0.018030116334557533, + 0.016704011708498, + 0.012546174228191376, + 0.009451362304389477, + 0.009567423723638058, + 0.011185572482645512, + 0.004688563756644726, + 0.0029791519045829773, + 0.006822717376053333, + 0.006250620353966951, + 0.0067390985786914825, + 0.006138117518275976, + 0.004485649056732655, + 0.005399968009442091, + 0.008875112980604172, + 0.005196257960051298, + 0.02868463657796383, + 0.009011239744722843, + 0.009370919317007065, + 0.006561758928000927, + 0.008186335675418377, + 0.028717001900076866, + 0.008519860915839672, + 0.007280607242137194, + 0.01722045987844467, + 0.009897871874272823, + 0.008810101076960564, + 0.011547649279236794, + 0.007847451604902744, + 0.01620352454483509 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 0, + 4 + ], + "selected_reward": -0.012532763183116913 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "cinematic, low-angle tracking shot follows a cyclist as he navigates a futuristic city at dusk. The young rider, clad in a seamless, skin-tight black cycling suit made of high-tech synthetic material, accentuates his muscular form, honed by years of training. He wears an aerodynamic helmet and gazes intently ahead, his jaw set in a resolute concentration. The connected, smart shared bike he rides boasts a sleek, one-piece frame crafted from matte black carbon fiber, free of any unnecessary wires. From front to back, thin, data-like streams of cyan-blue light pulse rhythmically, flowing along the frame's contours and forming mesmerizing halos around the rapidly spinning wheels. In the background, towering skyscrapers, adorned with giant holographic billboards, intertwine their vibrant neon lights with the deep twilight sky, casting distorted and iridescent reflections on the rain-soaked asphalt. The rider leans forward, pedaling with a steady, powerful rhythm. The light from the bike's data stream illuminates the movements of his legs with exceptional clarity, leaving a fleeting blue trail on the road. The 8K ultra-high-definition camera perfectly captures every detail—from the tiny water droplets splashed by the wheels to the light reflecting off the rider's muscular tones—creating a dynamic and hopeful atmosphere that blends cutting-edge technology with human perseverance.", + "seed": 3068, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000034/group_12/index_3907_seed_3068_dp_25_rank_50_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 12, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.03193258885294199, + "vlm_logprob": -0.03193258885294199, + "vlm_yes_no_margin": 4.375, + "vlm_token_credit": -0.03193258885294199, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot taken from a low angle while tracking the cyclist?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the cyclist wearing a seamless, skin-tight black cycling suit made of high-tech synthetic material?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the bike feature a sleek, one-piece matte black carbon fiber frame with thin cyan-blue data-like light streams pulsing along it?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a futuristic city at dusk with towering skyscrapers, holographic billboards, and rain-soaked asphalt reflecting neon lights?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the rider leaning forward and pedaling with a steady, powerful rhythm?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04043861851096153 + ], + "opposite_logprob_tokens": [ + -4.290438652038574 + ], + "logprob_reward": -0.04043861851096153, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.04043861851096153, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.19314028322696686, + 0.13282832503318787, + 0.1126689687371254, + 0.07227196544408798, + 0.03582364693284035, + 0.159773051738739, + 0.06488429754972458, + 0.09598192572593689, + 0.09242112189531326, + 0.04020644351840019 + ], + "credit_map": [ + 0.05233500897884369, + 0.019245080649852753, + 0.018140504136681557, + 0.03257302939891815, + 0.014566871337592602, + 0.018963729962706566, + 0.01731974072754383, + 0.019996320828795433, + 0.031781964004039764, + 0.007123274262994528, + 0.05221938341856003, + 0.01003275252878666, + 0.003913302905857563, + 0.01275564730167389, + 0.009639527648687363, + 0.005362476222217083, + 0.012243871577084064, + 0.027778666466474533, + 0.010296197608113289, + 0.005902836099267006, + 0.012045136652886868, + 0.015007338486611843, + 0.01937607303261757, + 0.010018856264650822, + 0.008359900675714016, + 0.0074797384440898895, + 0.008873650804162025, + 0.006310393568128347, + 0.013511549681425095, + 0.011541650630533695, + 0.008094735443592072, + 0.008100348524749279, + 0.007120050489902496, + 0.003956349100917578, + 0.0077773164957761765, + 0.0033947450574487448, + 0.0024937998969107866, + 0.006286504678428173, + 0.0022489409893751144, + 0.002545941388234496, + 0.013418142683804035, + 0.01968028023838997, + 0.029066918417811394, + 0.02344152331352234, + 0.01914537511765957, + 0.023817135021090508, + 0.018432756885886192, + 0.012770910747349262, + 0.012094220146536827, + 0.007028237450867891, + 0.012362407520413399, + 0.005895831156522036, + 0.006368520203977823, + 0.011060787364840508, + 0.004482502117753029, + 0.005591784603893757, + 0.006335014011710882, + 0.004783385433256626, + 0.03396828845143318, + 0.00455306563526392, + 0.004212874453514814, + 0.030829232186079025, + 0.003661317750811577, + 0.007638752926141024, + 0.01704665832221508, + 0.018033958971500397, + 0.016507361084222794, + 0.007325643673539162, + 0.006749937776476145, + 0.012486905790865421, + 0.007622655481100082, + 0.00664799427613616, + 0.007781964726746082, + 0.004819292575120926, + 0.010461715050041676, + 0.0028286082670092583, + 0.0028842389583587646, + 0.005230333190411329, + 0.004076922312378883, + 0.002123370999470353 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 5, + 1 + ], + "selected_reward": -0.04043861851096153 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot taken from a low angle while tracking the cyclist?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the cyclist wearing a seamless, skin-tight black cycling suit made of high-tech synthetic material?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the bike feature a sleek, one-piece matte black carbon fiber frame with thin cyan-blue data-like light streams pulsing along it?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a futuristic city at dusk with towering skyscrapers, holographic billboards, and rain-soaked asphalt reflecting neon lights?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the rider leaning forward and pedaling with a steady, powerful rhythm?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.10035744309425354 + ], + "opposite_logprob_tokens": [ + -2.3503575325012207 + ], + "logprob_reward": -0.10035744309425354, + "yes_no_margin_reward": 2.25, + "token_credit_reward": -0.10035744309425354, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.18738846480846405, + 0.12770269811153412, + 0.07883492857217789, + 0.03577878698706627, + 0.02844279818236828, + 0.1393522024154663, + 0.057360466569662094, + 0.2354951947927475, + 0.06033378466963768, + 0.04931061342358589 + ], + "credit_map": [ + 0.03986341133713722, + 0.01134337019175291, + 0.027407031506299973, + 0.036896903067827225, + 0.009564993903040886, + 0.03281273692846298, + 0.016477206721901894, + 0.013022816739976406, + 0.01928620971739292, + 0.011145683005452156, + 0.0071135638281702995, + 0.01695651188492775, + 0.008430740796029568, + 0.008361851796507835, + 0.04213676229119301, + 0.01427137665450573, + 0.0077941264025866985, + 0.018670428544282913, + 0.007831042632460594, + 0.004949323832988739, + 0.00799091812223196, + 0.01127091608941555, + 0.013251928612589836, + 0.00707624526694417, + 0.00638013519346714, + 0.006964671425521374, + 0.004401025362312794, + 0.0036878795363008976, + 0.004234660882502794, + 0.0036054488737136126, + 0.0030229901894927025, + 0.0034819745924323797, + 0.003628154518082738, + 0.0029085897840559483, + 0.004586850758641958, + 0.0031956664752215147, + 0.003714645281434059, + 0.004523029085248709, + 0.00335821695625782, + 0.0025276460219174623, + 0.013054613955318928, + 0.015773393213748932, + 0.03673606365919113, + 0.011821730062365532, + 0.010515724308788776, + 0.032638709992170334, + 0.011160283349454403, + 0.007651674095541239, + 0.007719370070844889, + 0.00777538912370801, + 0.005456318613141775, + 0.0049483319744467735, + 0.009306247346103191, + 0.009843913838267326, + 0.006874818820506334, + 0.005436073988676071, + 0.014628004282712936, + 0.015410580672323704, + 0.07331772893667221, + 0.01272449642419815, + 0.01892842724919319, + 0.06540293246507645, + 0.015141687355935574, + 0.019941341131925583, + 0.008117555640637875, + 0.008127326145768166, + 0.010241159237921238, + 0.004740801639854908, + 0.007618428207933903, + 0.008206913247704506, + 0.0062112584710121155, + 0.007070343475788832, + 0.005627814214676619, + 0.01165700238198042, + 0.007724622264504433, + 0.003427846124395728, + 0.006353519391268492, + 0.00618457468226552, + 0.005126870237290859, + 0.0032083638943731785 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 0, + 5 + ], + "selected_reward": -0.10035744309425354 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot taken from a low angle while tracking the cyclist?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the cyclist wearing a seamless, skin-tight black cycling suit made of high-tech synthetic material?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the bike feature a sleek, one-piece matte black carbon fiber frame with thin cyan-blue data-like light streams pulsing along it?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a futuristic city at dusk with towering skyscrapers, holographic billboards, and rain-soaked asphalt reflecting neon lights?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the rider leaning forward and pedaling with a steady, powerful rhythm?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.007629304192960262 + ], + "opposite_logprob_tokens": [ + -4.88262939453125 + ], + "logprob_reward": -0.007629304192960262, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.007629304192960262, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.32597360014915466, + 0.217269167304039, + 0.07710471749305725, + 0.040690865367650986, + 0.01840454339981079, + 0.057246580719947815, + 0.07012996077537537, + 0.08548681437969208, + 0.0751166120171547, + 0.03257707506418228 + ], + "credit_map": [ + 0.08040265738964081, + 0.023631615564227104, + 0.025709355250000954, + 0.048670630902051926, + 0.033763207495212555, + 0.03535760939121246, + 0.03828728571534157, + 0.04015124961733818, + 0.061361655592918396, + 0.014980903826653957, + 0.08290232717990875, + 0.01829567365348339, + 0.005154863931238651, + 0.017627596855163574, + 0.010265807621181011, + 0.006680336315184832, + 0.008220204152166843, + 0.0068922583013772964, + 0.00898051355034113, + 0.009347544983029366, + 0.011364854872226715, + 0.010178743861615658, + 0.010557913221418858, + 0.011562686413526535, + 0.007298633456230164, + 0.005467840936034918, + 0.008795350790023804, + 0.003537378739565611, + 0.003088330151513219, + 0.005785190965980291, + 0.0033242888748645782, + 0.003393852384760976, + 0.0020871537271887064, + 0.0017359720077365637, + 0.0027602154295891523, + 0.002695067087188363, + 0.0016473468858748674, + 0.0032088449224829674, + 0.002244207076728344, + 0.002025736728683114, + 0.005388814024627209, + 0.0062417746521532536, + 0.009694586507976055, + 0.0066519128158688545, + 0.006818542256951332, + 0.013358612544834614, + 0.005980338901281357, + 0.0031119973864406347, + 0.008796644397079945, + 0.007094407454133034, + 0.008738124743103981, + 0.00785630289465189, + 0.014482217840850353, + 0.0074051120318472385, + 0.0071782879531383514, + 0.008578864857554436, + 0.006706171203404665, + 0.005973509047180414, + 0.02463441714644432, + 0.00564628466963768, + 0.011819855310022831, + 0.018054984509944916, + 0.005687074735760689, + 0.006964517291635275, + 0.00969394575804472, + 0.01139030884951353, + 0.011626954190433025, + 0.00686487415805459, + 0.009474188089370728, + 0.009923282079398632, + 0.007633121684193611, + 0.00850993487983942, + 0.004461449570953846, + 0.008067145012319088, + 0.006273805163800716, + 0.0020552165806293488, + 0.0021967997308820486, + 0.00391449686139822, + 0.003884878009557724, + 0.0017232809914276004 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 7 + ], + "selected_reward": -0.007629304192960262 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot taken from a low angle while tracking the cyclist?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the cyclist wearing a seamless, skin-tight black cycling suit made of high-tech synthetic material?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the bike feature a sleek, one-piece matte black carbon fiber frame with thin cyan-blue data-like light streams pulsing along it?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a futuristic city at dusk with towering skyscrapers, holographic billboards, and rain-soaked asphalt reflecting neon lights?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the rider leaning forward and pedaling with a steady, powerful rhythm?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.007627174723893404 + ], + "opposite_logprob_tokens": [ + -4.882627010345459 + ], + "logprob_reward": -0.007627174723893404, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.007627174723893404, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.16392655670642853, + 0.1534799486398697, + 0.18182653188705444, + 0.1152113676071167, + 0.04731651023030281, + 0.09184705466032028, + 0.10043614357709885, + 0.10509242117404938, + 0.02400260977447033, + 0.016860822215676308 + ], + "credit_map": [ + 0.041288405656814575, + 0.013669786974787712, + 0.01934479922056198, + 0.04463927075266838, + 0.008115014992654324, + 0.016085678711533546, + 0.013707115314900875, + 0.007076474837958813, + 0.04113726690411568, + 0.010869167745113373, + 0.060602810233831406, + 0.011248747818171978, + 0.0035752388648688793, + 0.014707912690937519, + 0.006257979664951563, + 0.005080830305814743, + 0.021511241793632507, + 0.02664448507130146, + 0.0157614778727293, + 0.009556845761835575, + 0.01978546380996704, + 0.01491193100810051, + 0.05278877913951874, + 0.02086629904806614, + 0.019329141825437546, + 0.015470133163034916, + 0.021834533661603928, + 0.00847411435097456, + 0.009429733268916607, + 0.026068056002259254, + 0.006632191129028797, + 0.007973468862473965, + 0.013245019130408764, + 0.0052818721160292625, + 0.013351338915526867, + 0.004667941946536303, + 0.002880533691495657, + 0.003139723325148225, + 0.002327636117115617, + 0.0024224428925663233, + 0.007987498305737972, + 0.011623822152614594, + 0.015642352402210236, + 0.014011550694704056, + 0.011244536377489567, + 0.014733003452420235, + 0.01114840991795063, + 0.005455885082483292, + 0.013339467346668243, + 0.009283577091991901, + 0.01123680081218481, + 0.01022230926901102, + 0.02228345163166523, + 0.010769912041723728, + 0.010919919237494469, + 0.012380705215036869, + 0.0070397090166807175, + 0.005064333789050579, + 0.035602718591690063, + 0.0048549179919064045, + 0.004688717890530825, + 0.03539344295859337, + 0.004293515812605619, + 0.008155073039233685, + 0.004472437314689159, + 0.0027777880895882845, + 0.0030362035613507032, + 0.0047949934378266335, + 0.0014914211351424456, + 0.003915672656148672, + 0.0014827661216259003, + 0.0020313262939453125, + 0.002463778480887413, + 0.005442568100988865, + 0.002254754537716508, + 0.0010288303019478917, + 0.001352605759166181, + 0.0017761965282261372, + 0.001702327048406005, + 0.0008397627389058471 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 0, + 1 + ], + "selected_reward": -0.007627174723893404 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot taken from a low angle while tracking the cyclist?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the cyclist wearing a seamless, skin-tight black cycling suit made of high-tech synthetic material?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the bike feature a sleek, one-piece matte black carbon fiber frame with thin cyan-blue data-like light streams pulsing along it?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a futuristic city at dusk with towering skyscrapers, holographic billboards, and rain-soaked asphalt reflecting neon lights?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the rider leaning forward and pedaling with a steady, powerful rhythm?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0036104037426412106 + ], + "opposite_logprob_tokens": [ + -5.628610610961914 + ], + "logprob_reward": -0.0036104037426412106, + "yes_no_margin_reward": 5.625, + "token_credit_reward": -0.0036104037426412106, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.37460869550704956, + 0.10272254049777985, + 0.15872454643249512, + 0.05296986922621727, + 0.026435639709234238, + 0.06967081129550934, + 0.048733144998550415, + 0.053350817412137985, + 0.06616382300853729, + 0.04662017151713371 + ], + "credit_map": [ + 0.11425905674695969, + 0.02892364375293255, + 0.021783486008644104, + 0.10582245141267776, + 0.023996546864509583, + 0.022228728979825974, + 0.036739904433488846, + 0.020854827016592026, + 0.020186929032206535, + 0.009400391951203346, + 0.020734449848532677, + 0.011322865262627602, + 0.00998005736619234, + 0.009528690949082375, + 0.011922083795070648, + 0.009647072292864323, + 0.018342873081564903, + 0.032683081924915314, + 0.015231617726385593, + 0.01976514421403408, + 0.01538645476102829, + 0.015256346203386784, + 0.01862049475312233, + 0.023438531905412674, + 0.00951213575899601, + 0.008652543649077415, + 0.008695325814187527, + 0.005028192885220051, + 0.005601544864475727, + 0.005795263219624758, + 0.0045800288207829, + 0.005104833282530308, + 0.004857951309531927, + 0.002921072533354163, + 0.004513928666710854, + 0.002989985980093479, + 0.00272803520783782, + 0.004127188585698605, + 0.002225337317213416, + 0.0020721398759633303, + 0.006575595587491989, + 0.007560776546597481, + 0.014082616195082664, + 0.011180290021002293, + 0.007918384857475758, + 0.007870837114751339, + 0.00900996569544077, + 0.0054723420180380344, + 0.006204421166330576, + 0.003932982217520475, + 0.006661843042820692, + 0.004913304932415485, + 0.00931656826287508, + 0.006323900539427996, + 0.0059741479344666, + 0.005405977368354797, + 0.0034592212177813053, + 0.0034066608641296625, + 0.016367213800549507, + 0.002608047565445304, + 0.004706274718046188, + 0.014936459250748158, + 0.003603816032409668, + 0.00426312442868948, + 0.006093885283917189, + 0.0065972283482551575, + 0.00946156308054924, + 0.0064749871380627155, + 0.011988392099738121, + 0.007932505570352077, + 0.009467906318604946, + 0.008147353306412697, + 0.004466515500098467, + 0.008460956625640392, + 0.005517607554793358, + 0.003248641500249505, + 0.009726645424962044, + 0.004507026635110378, + 0.006922339554876089, + 0.0037704382557421923 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 1 + ], + "selected_reward": -0.0036104037426412106 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "cinematic, low-angle tracking shot follows a cyclist as he navigates a futuristic city at dusk. The young rider, clad in a seamless, skin-tight black cycling suit made of high-tech synthetic material, accentuates his muscular form, honed by years of training. He wears an aerodynamic helmet and gazes intently ahead, his jaw set in a resolute concentration. The connected, smart shared bike he rides boasts a sleek, one-piece frame crafted from matte black carbon fiber, free of any unnecessary wires. From front to back, thin, data-like streams of cyan-blue light pulse rhythmically, flowing along the frame's contours and forming mesmerizing halos around the rapidly spinning wheels. In the background, towering skyscrapers, adorned with giant holographic billboards, intertwine their vibrant neon lights with the deep twilight sky, casting distorted and iridescent reflections on the rain-soaked asphalt. The rider leans forward, pedaling with a steady, powerful rhythm. The light from the bike's data stream illuminates the movements of his legs with exceptional clarity, leaving a fleeting blue trail on the road. The 8K ultra-high-definition camera perfectly captures every detail—from the tiny water droplets splashed by the wheels to the light reflecting off the rider's muscular tones—creating a dynamic and hopeful atmosphere that blends cutting-edge technology with human perseverance.", + "seed": 3069, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000034/group_12/index_3907_seed_3069_dp_25_rank_50_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 12, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.03827803712338209, + "vlm_logprob": -0.03827803712338209, + "vlm_yes_no_margin": 4.1, + "vlm_token_credit": -0.03827803712338209, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot taken from a low angle while tracking the cyclist?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the cyclist wearing a seamless, skin-tight black cycling suit made of high-tech synthetic material?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the bike feature a sleek, one-piece matte black carbon fiber frame with thin cyan-blue data-like light streams pulsing along it?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a futuristic city at dusk with towering skyscrapers, holographic billboards, and rain-soaked asphalt reflecting neon lights?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the rider leaning forward and pedaling with a steady, powerful rhythm?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03738638013601303 + ], + "opposite_logprob_tokens": [ + -4.287386417388916 + ], + "logprob_reward": -0.03738638013601303, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.03738638013601303, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.16287195682525635, + 0.0777939110994339, + 0.0536356084048748, + 0.034538835287094116, + 0.05973638966679573, + 0.04915633052587509, + 0.08704590052366257, + 0.09259232878684998, + 0.2397204041481018, + 0.14290837943553925 + ], + "credit_map": [ + 0.014560137875378132, + 0.013980216346681118, + 0.013297044672071934, + 0.01712208054959774, + 0.02141830138862133, + 0.02563881315290928, + 0.016099808737635612, + 0.04075554385781288, + 0.007318421732634306, + 0.006923601496964693, + 0.008363216184079647, + 0.010724006220698357, + 0.021973369643092155, + 0.009409821592271328, + 0.008462432771921158, + 0.0046190437860786915, + 0.004694764968007803, + 0.009177534841001034, + 0.006261616479605436, + 0.008411280810832977, + 0.0064071048982441425, + 0.006649013143032789, + 0.007621396332979202, + 0.004412895534187555, + 0.006768975406885147, + 0.006406599190086126, + 0.0035548233427107334, + 0.003792105708271265, + 0.0037311904598027468, + 0.0032493649050593376, + 0.003333546919748187, + 0.0037022288888692856, + 0.0075533208437263966, + 0.004596387501806021, + 0.014087908901274204, + 0.007642464246600866, + 0.004674315918236971, + 0.008856420405209064, + 0.005024039186537266, + 0.0073015326634049416, + 0.005201566964387894, + 0.004186798352748156, + 0.006557437125593424, + 0.004831107798963785, + 0.010346708819270134, + 0.007307932712137699, + 0.004213600419461727, + 0.0065111806616187096, + 0.01700449548661709, + 0.010612049140036106, + 0.008822262287139893, + 0.016088053584098816, + 0.007770481985062361, + 0.007587138097733259, + 0.009346498176455498, + 0.009814922697842121, + 0.008639601990580559, + 0.01329437643289566, + 0.007751045282930136, + 0.009645689278841019, + 0.011338449083268642, + 0.007872570306062698, + 0.015240845270454884, + 0.018809756264090538, + 0.012393527664244175, + 0.019787607714533806, + 0.01659802533686161, + 0.014134477823972702, + 0.023902475833892822, + 0.06966366618871689, + 0.020042642951011658, + 0.06319795548915863, + 0.010292875580489635, + 0.00935103464871645, + 0.0326312817633152, + 0.021449096500873566, + 0.016788411885499954, + 0.016128838062286377, + 0.01876179315149784, + 0.01750505156815052 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 0, + 9 + ], + "selected_reward": -0.03738638013601303 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot taken from a low angle while tracking the cyclist?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the cyclist wearing a seamless, skin-tight black cycling suit made of high-tech synthetic material?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the bike feature a sleek, one-piece matte black carbon fiber frame with thin cyan-blue data-like light streams pulsing along it?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a futuristic city at dusk with towering skyscrapers, holographic billboards, and rain-soaked asphalt reflecting neon lights?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the rider leaning forward and pedaling with a steady, powerful rhythm?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.12710663676261902 + ], + "opposite_logprob_tokens": [ + -2.1271066665649414 + ], + "logprob_reward": -0.12710663676261902, + "yes_no_margin_reward": 2.0, + "token_credit_reward": -0.12710663676261902, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.20581656694412231, + 0.06277811527252197, + 0.06336715817451477, + 0.10296498239040375, + 0.10080254077911377, + 0.025145018473267555, + 0.09119202941656113, + 0.14751547574996948, + 0.08560634404420853, + 0.11481183767318726 + ], + "credit_map": [ + 0.019296113401651382, + 0.026777859777212143, + 0.024638041853904724, + 0.02840103581547737, + 0.01978696696460247, + 0.0331013947725296, + 0.019577112048864365, + 0.03423803672194481, + 0.004920515697449446, + 0.005362246185541153, + 0.006198020651936531, + 0.007152204867452383, + 0.01650397665798664, + 0.0064831506460905075, + 0.010809693485498428, + 0.0053483047522604465, + 0.005548339337110519, + 0.012134441174566746, + 0.006455674301832914, + 0.008679579012095928, + 0.009096265770494938, + 0.01016524713486433, + 0.005765934009104967, + 0.005521677900105715, + 0.03125400468707085, + 0.013702135533094406, + 0.013874112628400326, + 0.00977905374020338, + 0.00596818421036005, + 0.007616298273205757, + 0.00855077151209116, + 0.012220417149364948, + 0.009461561217904091, + 0.004074245225638151, + 0.004176001064479351, + 0.006808441132307053, + 0.004762218799442053, + 0.03404377028346062, + 0.008425080217421055, + 0.029051216319203377, + 0.003306108294054866, + 0.0031864275224506855, + 0.002418343210592866, + 0.003939397167414427, + 0.0036035922821611166, + 0.0032074758782982826, + 0.0023147177416831255, + 0.0031689582392573357, + 0.03426303341984749, + 0.005459390580654144, + 0.007119965739548206, + 0.006146378815174103, + 0.004429627675563097, + 0.0182200875133276, + 0.007276018615812063, + 0.008277523331344128, + 0.01162030827254057, + 0.019442062824964523, + 0.01585954800248146, + 0.0292724147439003, + 0.014713914133608341, + 0.011403718031942844, + 0.02043069526553154, + 0.024772820994257927, + 0.006069309078156948, + 0.010804926976561546, + 0.009287061169743538, + 0.009118632413446903, + 0.007997372187674046, + 0.019105195999145508, + 0.009588384069502354, + 0.013635464012622833, + 0.0077057830058038235, + 0.008939642459154129, + 0.025451913475990295, + 0.017269762232899666, + 0.013700293377041817, + 0.01133123692125082, + 0.020640388131141663, + 0.009772824123501778 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 7, + 9 + ], + "selected_reward": -0.12710663676261902 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot taken from a low angle while tracking the cyclist?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the cyclist wearing a seamless, skin-tight black cycling suit made of high-tech synthetic material?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the bike feature a sleek, one-piece matte black carbon fiber frame with thin cyan-blue data-like light streams pulsing along it?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a futuristic city at dusk with towering skyscrapers, holographic billboards, and rain-soaked asphalt reflecting neon lights?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the rider leaning forward and pedaling with a steady, powerful rhythm?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0067365444265306 + ], + "opposite_logprob_tokens": [ + -5.006736755371094 + ], + "logprob_reward": -0.0067365444265306, + "yes_no_margin_reward": 5.0, + "token_credit_reward": -0.0067365444265306, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09465253353118896, + 0.0733816847205162, + 0.04814513027667999, + 0.07374724000692368, + 0.18558084964752197, + 0.033913202583789825, + 0.1129927858710289, + 0.06745584309101105, + 0.09893966466188431, + 0.2111910581588745 + ], + "credit_map": [ + 0.009465925395488739, + 0.011810997501015663, + 0.009892886504530907, + 0.012653440237045288, + 0.01345934346318245, + 0.013518283143639565, + 0.011480957269668579, + 0.012370705604553223, + 0.0069304779171943665, + 0.007101407274603844, + 0.005965778138488531, + 0.008114009164273739, + 0.020391061902046204, + 0.008335278369486332, + 0.01114069577306509, + 0.005402973387390375, + 0.004169878549873829, + 0.009727856144309044, + 0.004941270221024752, + 0.0086439847946167, + 0.004456609021872282, + 0.007191918790340424, + 0.005293399561196566, + 0.0037202113308012486, + 0.019619280472397804, + 0.011643143370747566, + 0.011040076613426208, + 0.007093373686075211, + 0.003952475730329752, + 0.0059234448708593845, + 0.005202583037316799, + 0.0092728640884161, + 0.016419170424342155, + 0.0043219830840826035, + 0.008424356579780579, + 0.010090136900544167, + 0.012021259404718876, + 0.07787968963384628, + 0.005023836158216, + 0.05140041187405586, + 0.0043431599624454975, + 0.0034647127613425255, + 0.003205309621989727, + 0.00395049387589097, + 0.006147722713649273, + 0.004055391065776348, + 0.0035731331445276737, + 0.0051732794381678104, + 0.04434777423739433, + 0.005314575508236885, + 0.009170285426080227, + 0.004875658545643091, + 0.005732022691518068, + 0.026581211015582085, + 0.00883761327713728, + 0.008133652620017529, + 0.008168870583176613, + 0.013133523054420948, + 0.006376712117344141, + 0.008749447762966156, + 0.009020335972309113, + 0.004331386182457209, + 0.009768558666110039, + 0.007907015271484852, + 0.005193658638745546, + 0.0052444953471422195, + 0.008633327670395374, + 0.007340265903621912, + 0.008389758877456188, + 0.03001490980386734, + 0.00703020766377449, + 0.027093037962913513, + 0.015902942046523094, + 0.0150464391335845, + 0.043813787400722504, + 0.029609952121973038, + 0.023782599717378616, + 0.024385223165154457, + 0.03444463387131691, + 0.024205481633543968 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 4, + 6 + ], + "selected_reward": -0.0067365444265306 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot taken from a low angle while tracking the cyclist?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the cyclist wearing a seamless, skin-tight black cycling suit made of high-tech synthetic material?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the bike feature a sleek, one-piece matte black carbon fiber frame with thin cyan-blue data-like light streams pulsing along it?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a futuristic city at dusk with towering skyscrapers, holographic billboards, and rain-soaked asphalt reflecting neon lights?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the rider leaning forward and pedaling with a steady, powerful rhythm?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.012537590228021145 + ], + "opposite_logprob_tokens": [ + -4.387537479400635 + ], + "logprob_reward": -0.012537590228021145, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.012537590228021145, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.28136709332466125, + 0.0917954221367836, + 0.1234336867928505, + 0.05630747228860855, + 0.03316420316696167, + 0.05506226792931557, + 0.05279511585831642, + 0.05943151190876961, + 0.1152501180768013, + 0.13139311969280243 + ], + "credit_map": [ + 0.02187231183052063, + 0.022072846069931984, + 0.04086722433567047, + 0.025609593838453293, + 0.04583296179771423, + 0.04344199597835541, + 0.04907401651144028, + 0.03259614109992981, + 0.005513511132448912, + 0.007968060672283173, + 0.011075480841100216, + 0.010806814767420292, + 0.015375258401036263, + 0.01005609706044197, + 0.020593246445059776, + 0.01040695235133171, + 0.009922021999955177, + 0.02391470968723297, + 0.009957793168723583, + 0.017228180542588234, + 0.019546937197446823, + 0.020795494318008423, + 0.011106564663350582, + 0.010961990803480148, + 0.009852391667664051, + 0.007557136006653309, + 0.0064805978909134865, + 0.006541009992361069, + 0.00540148327127099, + 0.005313900299370289, + 0.008805222809314728, + 0.006355732213705778, + 0.0031531681306660175, + 0.0028580036014318466, + 0.004688899964094162, + 0.004014601465314627, + 0.002992303343489766, + 0.00794562790542841, + 0.0029875817708671093, + 0.0045240167528390884, + 0.007732958532869816, + 0.005008439067751169, + 0.007179996930062771, + 0.005951880477368832, + 0.00863299798220396, + 0.006841214373707771, + 0.006395931355655193, + 0.0073188478127121925, + 0.014764958992600441, + 0.0035047305282205343, + 0.004291234537959099, + 0.0036398929078131914, + 0.00448855385184288, + 0.009507295675575733, + 0.005824078805744648, + 0.006774373818188906, + 0.004608914256095886, + 0.009027763269841671, + 0.006094851531088352, + 0.010432664304971695, + 0.0068967449478805065, + 0.005045067984610796, + 0.008650140836834908, + 0.008675366640090942, + 0.008348639123141766, + 0.011165464296936989, + 0.014005779288709164, + 0.012611126527190208, + 0.012717019766569138, + 0.02677333541214466, + 0.013120434246957302, + 0.016508329659700394, + 0.013943148776888847, + 0.0165877528488636, + 0.021440748125314713, + 0.011082686483860016, + 0.015393462963402271, + 0.020488673821091652, + 0.0105214174836874, + 0.0219352375715971 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 2 + ], + "selected_reward": -0.012537590228021145 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot taken from a low angle while tracking the cyclist?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the cyclist wearing a seamless, skin-tight black cycling suit made of high-tech synthetic material?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the bike feature a sleek, one-piece matte black carbon fiber frame with thin cyan-blue data-like light streams pulsing along it?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a futuristic city at dusk with towering skyscrapers, holographic billboards, and rain-soaked asphalt reflecting neon lights?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the rider leaning forward and pedaling with a steady, powerful rhythm?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.007623034063726664 + ], + "opposite_logprob_tokens": [ + -4.882623195648193 + ], + "logprob_reward": -0.007623034063726664, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.007623034063726664, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0830552875995636, + 0.026093170046806335, + 0.0496273972094059, + 0.07967790961265564, + 0.16539059579372406, + 0.13111037015914917, + 0.14457246661186218, + 0.16787858307361603, + 0.10284288972616196, + 0.04975135996937752 + ], + "credit_map": [ + 0.007939756847918034, + 0.010793708264827728, + 0.00875008199363947, + 0.013244500383734703, + 0.007481981068849564, + 0.01148171816021204, + 0.011065584607422352, + 0.012297949753701687, + 0.002458172617480159, + 0.0036621387116611004, + 0.0025349664501845837, + 0.003457661485299468, + 0.003769433591514826, + 0.0022085378877818584, + 0.0057699354365468025, + 0.0022323208395391703, + 0.003943263087421656, + 0.009973317384719849, + 0.0055945212952792645, + 0.006901825312525034, + 0.007966602221131325, + 0.005808405578136444, + 0.004615771118551493, + 0.00482369028031826, + 0.0106424605473876, + 0.013271473348140717, + 0.009539076127111912, + 0.009467515163123608, + 0.00962089840322733, + 0.011836244724690914, + 0.006818893365561962, + 0.008481341414153576, + 0.0144355408847332, + 0.007688107434660196, + 0.015028196386992931, + 0.01708884909749031, + 0.013847185298800468, + 0.052580155432224274, + 0.012556069530546665, + 0.032166488468647, + 0.014634273946285248, + 0.011879518628120422, + 0.013692439533770084, + 0.01659785956144333, + 0.018152933567762375, + 0.029713625088334084, + 0.011042379774153233, + 0.01539731863886118, + 0.03233426809310913, + 0.013198255561292171, + 0.016099805012345314, + 0.0216367207467556, + 0.011246448382735252, + 0.013452747836709023, + 0.013964958488941193, + 0.022639255970716476, + 0.02109026536345482, + 0.03298035264015198, + 0.011755812913179398, + 0.025779426097869873, + 0.024978535249829292, + 0.009604344144463539, + 0.02390850894153118, + 0.01778133027255535, + 0.005781750660389662, + 0.004922274500131607, + 0.007722456008195877, + 0.008402456529438496, + 0.011959466151893139, + 0.02922292798757553, + 0.008934426121413708, + 0.025897128507494926, + 0.00571632944047451, + 0.004385247360914946, + 0.005913307890295982, + 0.005056461319327354, + 0.008889866061508656, + 0.0067128147929906845, + 0.005362239666283131, + 0.007715090177953243 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 4, + 6 + ], + "selected_reward": -0.007623034063726664 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000034/group_15/index_tensor([9829, 9829])_group_15_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000034/group_15/index_tensor([9829, 9829])_group_15_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..1fb03c14be1733bb32188d0ecf49edf29c1b0398 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000034/group_15/index_tensor([9829, 9829])_group_15_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "dramatic, extreme top-down close-up slowly zooms in on a sheet of weathered parchment, lying flat on a black, polished wooden table. The entire scene is bathed in warm, low-key lighting cast by a single, flickering candle just out of frame, its wavering light causing shadows to dance and highlighting rich textures. The parchment itself is a map of time, its surface a pale ochre stained with dark brown water stains and frayed along its delicately curled edges. Across this ancient canvas, a complex column of Chinese characters, written in a flowing calligraphic script, shimmers with wet ink. Initially static, the elegant forms begin to stir; subtle, unsettling pulsations ripple through the broadest brushstrokes. Then, with a silent, desperate energy, the figures begin to twist and struggle. Their elegant lines twist unnaturally, stretching as if exhausted, then violently contracting, the liquid ink bleeding and reforming as the roots of each symbol pull apart and collide in a chaotic, living dance. The effect is both beautiful and terrifying, creating a tense, mysterious atmosphere of both ancient magic awakening and an entity's struggle to escape its two-dimensional prison.", + "seed": 23996, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000034/group_15/index_9829_seed_23996_dp_30_rank_60_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 15, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.1554971156641841, + "vlm_logprob": -0.1554971156641841, + "vlm_yes_no_margin": 2.299999988079071, + "vlm_token_credit": -0.1554971156641841, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned in an extreme top-down close-up view?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera slowly zoom in on the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene lit by a single flickering candle source casting warm, low-key light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the Chinese calligraphy characters on the parchment begin to twist, stretch, and struggle as if alive?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the parchment surface pale ochre with dark brown water stains and frayed, curled edges?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.12611208856105804 + ], + "opposite_logprob_tokens": [ + -2.3761119842529297 + ], + "logprob_reward": -0.12611208856105804, + "yes_no_margin_reward": 2.25, + "token_credit_reward": -0.12611208856105804, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.18365104496479034, + 0.038552358746528625, + 0.08349534124135971, + 0.0470670610666275, + 0.04612645134329796, + 0.024397043511271477, + 0.032052136957645416, + 0.07720644772052765, + 0.29915767908096313, + 0.16829438507556915 + ], + "credit_map": [ + 0.0233071930706501, + 0.031093748286366463, + 0.018758602440357208, + 0.038892656564712524, + 0.014279989525675774, + 0.017410926520824432, + 0.01958654075860977, + 0.02032138593494892, + 0.004244854208081961, + 0.005856329575181007, + 0.003235854208469391, + 0.00432439474388957, + 0.0038348333910107613, + 0.005319175776094198, + 0.007001832127571106, + 0.004735080990940332, + 0.0059680514968931675, + 0.0065412758849561214, + 0.010885457508265972, + 0.02952245995402336, + 0.0046548992395401, + 0.005094509571790695, + 0.011006791144609451, + 0.009821896441280842, + 0.006575297564268112, + 0.004791542422026396, + 0.0057619912549853325, + 0.0049569448456168175, + 0.004660872742533684, + 0.0079198582097888, + 0.006961352191865444, + 0.0054392023012042046, + 0.003218804020434618, + 0.00699245510622859, + 0.00396260479465127, + 0.007707054726779461, + 0.005699953064322472, + 0.006687731947749853, + 0.006014577578753233, + 0.005843266379088163, + 0.0021388872992247343, + 0.0036895680241286755, + 0.002387799322605133, + 0.005219664424657822, + 0.0022103458177298307, + 0.0028270354960113764, + 0.003119294997304678, + 0.002804448362439871, + 0.002818781416863203, + 0.0029225389007478952, + 0.007691096048802137, + 0.004512147977948189, + 0.0022902791388332844, + 0.0026062913239002228, + 0.005030564498156309, + 0.004180439282208681, + 0.0038595120422542095, + 0.007036634720861912, + 0.01487718429416418, + 0.02632458135485649, + 0.004261693451553583, + 0.004211498890072107, + 0.005852957256138325, + 0.010782389901578426, + 0.03332996368408203, + 0.02964659035205841, + 0.047841619700193405, + 0.04887748882174492, + 0.029897773638367653, + 0.024691501632332802, + 0.04410123452544212, + 0.040771521627902985, + 0.018901808187365532, + 0.022776495665311813, + 0.02100743167102337, + 0.031698159873485565, + 0.014682207256555557, + 0.013820426538586617, + 0.015786508098244667, + 0.029621345922350883 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 0, + 9 + ], + "selected_reward": -0.12611208856105804 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned in an extreme top-down close-up view?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera slowly zoom in on the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene lit by a single flickering candle source casting warm, low-key light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the Chinese calligraphy characters on the parchment begin to twist, stretch, and struggle as if alive?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the parchment surface pale ochre with dark brown water stains and frayed, curled edges?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.3486819863319397 + ], + "opposite_logprob_tokens": [ + -1.223681926727295 + ], + "logprob_reward": -0.3486819863319397, + "yes_no_margin_reward": 0.8749999403953552, + "token_credit_reward": -0.3486819863319397, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.27597978711128235, + 0.052209582179784775, + 0.09734392166137695, + 0.16379958391189575, + 0.11348534375429153, + 0.0637330636382103, + 0.03721790388226509, + 0.04464852064847946, + 0.0662161111831665, + 0.08536618202924728 + ], + "credit_map": [ + 0.03610563650727272, + 0.04409179463982582, + 0.02465205267071724, + 0.036325979977846146, + 0.02359735406935215, + 0.027515891939401627, + 0.041756052523851395, + 0.04193501174449921, + 0.005293203517794609, + 0.005970847327262163, + 0.0039732800796628, + 0.006446299143135548, + 0.004257442895323038, + 0.009901735000312328, + 0.007153806742280722, + 0.009212963283061981, + 0.006151609122753143, + 0.0080801323056221, + 0.012440234422683716, + 0.03230493888258934, + 0.005037322640419006, + 0.007454784587025642, + 0.011176986619830132, + 0.014697917737066746, + 0.013801123946905136, + 0.02715258114039898, + 0.022228283807635307, + 0.029302839189767838, + 0.01599246822297573, + 0.012712185271084309, + 0.022703060880303383, + 0.019907057285308838, + 0.009231165051460266, + 0.009551257826387882, + 0.008198185823857784, + 0.011389936320483685, + 0.023844802752137184, + 0.01897634193301201, + 0.009888533502817154, + 0.022405123338103294, + 0.005630483850836754, + 0.010271322913467884, + 0.006114448420703411, + 0.012942287139594555, + 0.00647903373464942, + 0.006053122226148844, + 0.009785907343029976, + 0.00645646546036005, + 0.003343024989590049, + 0.005510414019227028, + 0.004629132803529501, + 0.008471553213894367, + 0.003266997169703245, + 0.0031778314150869846, + 0.004174640402197838, + 0.004644312895834446, + 0.0039811101742088795, + 0.00426986301317811, + 0.009346945211291313, + 0.009899369440972805, + 0.003499377518892288, + 0.004097463563084602, + 0.0055679623037576675, + 0.003986430820077658, + 0.007773025892674923, + 0.009212726727128029, + 0.00772432517260313, + 0.009689330123364925, + 0.009407094679772854, + 0.007800383027642965, + 0.007863558828830719, + 0.006745663471519947, + 0.010781803168356419, + 0.010613570921123028, + 0.0127305518835783, + 0.012455160729587078, + 0.009863189421594143, + 0.010382386855781078, + 0.008505198173224926, + 0.01003432460129261 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 4 + ], + "selected_reward": -0.3486819863319397 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned in an extreme top-down close-up view?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera slowly zoom in on the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene lit by a single flickering candle source casting warm, low-key light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the Chinese calligraphy characters on the parchment begin to twist, stretch, and struggle as if alive?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the parchment surface pale ochre with dark brown water stains and frayed, curled edges?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.11297227442264557 + ], + "opposite_logprob_tokens": [ + -2.2379722595214844 + ], + "logprob_reward": -0.11297227442264557, + "yes_no_margin_reward": 2.125, + "token_credit_reward": -0.11297227442264557, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12767565250396729, + 0.1672046035528183, + 0.09912919998168945, + 0.11951051652431488, + 0.08508153259754181, + 0.01505043264478445, + 0.031847838312387466, + 0.04981537163257599, + 0.23774348199367523, + 0.06694125384092331 + ], + "credit_map": [ + 0.012812341563403606, + 0.021840697154402733, + 0.012175660580396652, + 0.023848379030823708, + 0.013100518845021725, + 0.013592003844678402, + 0.015634309500455856, + 0.014671762473881245, + 0.019240299239754677, + 0.021626250818371773, + 0.014888377860188484, + 0.01934853009879589, + 0.019724763929843903, + 0.021204646676778793, + 0.03139466047286987, + 0.019777102395892143, + 0.008799724280834198, + 0.013823725283145905, + 0.009936762973666191, + 0.020744098350405693, + 0.008697357028722763, + 0.011665343306958675, + 0.015594406053423882, + 0.009867792017757893, + 0.017920607700943947, + 0.01065297331660986, + 0.018332740291953087, + 0.011299630627036095, + 0.009631357155740261, + 0.014627380296587944, + 0.01947801560163498, + 0.017567826434969902, + 0.008906184695661068, + 0.011254504323005676, + 0.007227463647723198, + 0.011219358071684837, + 0.008579772897064686, + 0.013448752462863922, + 0.014931808225810528, + 0.009513694792985916, + 0.0011073386995121837, + 0.0022681145928800106, + 0.001999549102038145, + 0.0025105506647378206, + 0.0014536429662257433, + 0.0018046465702354908, + 0.0016727186739444733, + 0.0022338710259646177, + 0.0023668899666517973, + 0.002123134909197688, + 0.00971357710659504, + 0.0027817387599498034, + 0.0024077484849840403, + 0.002336423844099045, + 0.0024292338639497757, + 0.007689096499234438, + 0.00255330977961421, + 0.004954385571181774, + 0.00939822569489479, + 0.016484763473272324, + 0.003099982161074877, + 0.0032790289260447025, + 0.004323325119912624, + 0.005722354631870985, + 0.02797868475317955, + 0.027033675462007523, + 0.038531672209501266, + 0.03924716264009476, + 0.02211974933743477, + 0.019126465544104576, + 0.032506655901670456, + 0.03119942732155323, + 0.007287726737558842, + 0.007192643824964762, + 0.008030983619391918, + 0.013462422415614128, + 0.0057893902994692326, + 0.00559378694742918, + 0.0063840229995548725, + 0.013200284913182259 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 1, + 0 + ], + "selected_reward": -0.11297227442264557 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned in an extreme top-down close-up view?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera slowly zoom in on the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene lit by a single flickering candle source casting warm, low-key light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the Chinese calligraphy characters on the parchment begin to twist, stretch, and struggle as if alive?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the parchment surface pale ochre with dark brown water stains and frayed, curled edges?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.17994053661823273 + ], + "opposite_logprob_tokens": [ + -1.8049405813217163 + ], + "logprob_reward": -0.17994053661823273, + "yes_no_margin_reward": 1.625, + "token_credit_reward": -0.17994053661823273, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13137631118297577, + 0.131172776222229, + 0.04688848927617073, + 0.15888220071792603, + 0.08850682526826859, + 0.030187278985977173, + 0.04824380204081535, + 0.06196722388267517, + 0.20937591791152954, + 0.09339910000562668 + ], + "credit_map": [ + 0.01759318634867668, + 0.017080536112189293, + 0.015351113863289356, + 0.014322425238788128, + 0.01291265245527029, + 0.01939084753394127, + 0.01539770606905222, + 0.019327841699123383, + 0.013437209650874138, + 0.016524212434887886, + 0.011842449195683002, + 0.015324406325817108, + 0.01503725629299879, + 0.01857358030974865, + 0.02279704250395298, + 0.017636613920331, + 0.003964098170399666, + 0.006407914217561483, + 0.004960345104336739, + 0.008938057348132133, + 0.0032939729280769825, + 0.007382430601865053, + 0.007145875133574009, + 0.004795797634869814, + 0.014422150328755379, + 0.022692985832691193, + 0.027793316170573235, + 0.02819875441491604, + 0.008845135569572449, + 0.011998731642961502, + 0.024682823568582535, + 0.020248299464583397, + 0.009490709751844406, + 0.010551059618592262, + 0.008245747536420822, + 0.013802336528897285, + 0.009478756226599216, + 0.011144895106554031, + 0.0159896407276392, + 0.009803679771721363, + 0.00250453338958323, + 0.0054585011675953865, + 0.0038666613399982452, + 0.006269247271120548, + 0.002472051652148366, + 0.0020115107763558626, + 0.003910096827894449, + 0.0036946781910955906, + 0.003987868782132864, + 0.007382738403975964, + 0.00826819334179163, + 0.007411873433738947, + 0.0034672252368181944, + 0.004645669832825661, + 0.0072643933817744255, + 0.00581583846360445, + 0.00712987594306469, + 0.00607984559610486, + 0.008063709363341331, + 0.007310079876333475, + 0.006460749078541994, + 0.008071931079030037, + 0.006493206601589918, + 0.012357829138636589, + 0.025046411901712418, + 0.02033226378262043, + 0.03294779732823372, + 0.034371230751276016, + 0.01904863677918911, + 0.01751856505870819, + 0.029752032831311226, + 0.030358975753188133, + 0.010151877999305725, + 0.011901704594492912, + 0.012306876480579376, + 0.01855633221566677, + 0.007209733594208956, + 0.00772215286269784, + 0.008569272235035896, + 0.016981150954961777 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 3, + 0 + ], + "selected_reward": -0.17994053661823273 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned in an extreme top-down close-up view?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera slowly zoom in on the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene lit by a single flickering candle source casting warm, low-key light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the Chinese calligraphy characters on the parchment begin to twist, stretch, and struggle as if alive?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the parchment surface pale ochre with dark brown water stains and frayed, curled edges?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.00977869238704443 + ], + "opposite_logprob_tokens": [ + -4.6347784996032715 + ], + "logprob_reward": -0.00977869238704443, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.00977869238704443, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.18419837951660156, + 0.07632328569889069, + 0.09903495758771896, + 0.04771965369582176, + 0.27237239480018616, + 0.07277031242847443, + 0.04322222247719765, + 0.058795612305402756, + 0.07481728494167328, + 0.07074593752622604 + ], + "credit_map": [ + 0.020815212279558182, + 0.032484784722328186, + 0.01959679275751114, + 0.03861599788069725, + 0.015994327142834663, + 0.015421410091221333, + 0.020424386486411095, + 0.02084546722471714, + 0.009824869222939014, + 0.009658070281147957, + 0.008702791295945644, + 0.008842239156365395, + 0.0055215382017195225, + 0.012077491730451584, + 0.01192531455308199, + 0.009770971722900867, + 0.006184850353747606, + 0.009965987876057625, + 0.008177653886377811, + 0.031571514904499054, + 0.005745214410126209, + 0.0077677699737250805, + 0.011217603459954262, + 0.018404364585876465, + 0.004521451890468597, + 0.007249330170452595, + 0.0062649864703416824, + 0.007898836396634579, + 0.005141005851328373, + 0.0038257604464888573, + 0.00700472854077816, + 0.005813551601022482, + 0.022533908486366272, + 0.03484363108873367, + 0.02356911264359951, + 0.041372764855623245, + 0.032227661460638046, + 0.037442419677972794, + 0.031758129596710205, + 0.04862476512789726, + 0.006525212898850441, + 0.008637739345431328, + 0.010060902684926987, + 0.007995682768523693, + 0.013152046129107475, + 0.008062953129410744, + 0.010496022179722786, + 0.007839756086468697, + 0.0031456644646823406, + 0.005617428570985794, + 0.011732865124940872, + 0.004590378608554602, + 0.0025035818107426167, + 0.002899806946516037, + 0.0051483577117323875, + 0.007584136910736561, + 0.00764637254178524, + 0.006038988009095192, + 0.009194727055728436, + 0.007142557762563229, + 0.004685739986598492, + 0.004779906943440437, + 0.003842161037027836, + 0.01546515803784132, + 0.006166304461658001, + 0.009257207624614239, + 0.011493420228362083, + 0.014347691088914871, + 0.00881408341228962, + 0.008263048715889454, + 0.009951733984053135, + 0.006523803807795048, + 0.006482415832579136, + 0.012022426351904869, + 0.010363264940679073, + 0.011863944120705128, + 0.00738462945446372, + 0.006346186622977257, + 0.009581494145095348, + 0.006701574195176363 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 0, + 2 + ], + "selected_reward": -0.00977869238704443 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "dramatic, extreme top-down close-up slowly zooms in on a sheet of weathered parchment, lying flat on a black, polished wooden table. The entire scene is bathed in warm, low-key lighting cast by a single, flickering candle just out of frame, its wavering light causing shadows to dance and highlighting rich textures. The parchment itself is a map of time, its surface a pale ochre stained with dark brown water stains and frayed along its delicately curled edges. Across this ancient canvas, a complex column of Chinese characters, written in a flowing calligraphic script, shimmers with wet ink. Initially static, the elegant forms begin to stir; subtle, unsettling pulsations ripple through the broadest brushstrokes. Then, with a silent, desperate energy, the figures begin to twist and struggle. Their elegant lines twist unnaturally, stretching as if exhausted, then violently contracting, the liquid ink bleeding and reforming as the roots of each symbol pull apart and collide in a chaotic, living dance. The effect is both beautiful and terrifying, creating a tense, mysterious atmosphere of both ancient magic awakening and an entity's struggle to escape its two-dimensional prison.", + "seed": 23997, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000034/group_15/index_9829_seed_23997_dp_30_rank_60_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 15, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.29702120684087274, + "vlm_logprob": -0.29702120684087274, + "vlm_yes_no_margin": 1.5749999761581421, + "vlm_token_credit": -0.29702120684087274, + "vlm_yes_ratio": 0.8, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned in an extreme top-down close-up view?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera slowly zoom in on the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene lit by a single flickering candle source casting warm, low-key light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the Chinese calligraphy characters on the parchment begin to twist, stretch, and struggle as if alive?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the parchment surface pale ochre with dark brown water stains and frayed, curled edges?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.26069775223731995 + ], + "opposite_logprob_tokens": [ + -1.6356977224349976 + ], + "logprob_reward": -0.26069775223731995, + "yes_no_margin_reward": 1.375, + "token_credit_reward": -0.26069775223731995, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.05156775191426277, + 0.7068719267845154, + 0.08229832351207733, + 0.04229941964149475, + 0.01694687455892563, + 0.033623792231082916, + 0.0074938624165952206, + 0.0217407438904047, + 0.014598927460610867, + 0.022558297961950302 + ], + "credit_map": [ + 0.004520497750490904, + 0.0041255224496126175, + 0.01070940587669611, + 0.006180762313306332, + 0.005452327895909548, + 0.008611349388957024, + 0.003946891985833645, + 0.008020997047424316, + 0.04743592441082001, + 0.04533972218632698, + 0.02875567227602005, + 0.051657628268003464, + 0.03758309409022331, + 0.21465791761875153, + 0.22290129959583282, + 0.058540720492601395, + 0.007513528689742088, + 0.005971115548163652, + 0.003519168822094798, + 0.005837913136929274, + 0.004656548611819744, + 0.025236327201128006, + 0.024219196289777756, + 0.005344536155462265, + 0.0035486093256622553, + 0.002067499328404665, + 0.010915208607912064, + 0.00293574295938015, + 0.00600910559296608, + 0.009418784640729427, + 0.0020396013278514147, + 0.00536486878991127, + 0.0020292506087571383, + 0.0016034864820539951, + 0.0021456805989146233, + 0.003216197481378913, + 0.002150688087567687, + 0.0025676873046904802, + 0.002042773412540555, + 0.0011911102337762713, + 0.002456801710650325, + 0.004823809489607811, + 0.0036981566809117794, + 0.0032086563296616077, + 0.007161479908972979, + 0.003320506541058421, + 0.0033806124702095985, + 0.005573771428316832, + 0.0016620998503640294, + 0.0011190505465492606, + 0.0008013771148398519, + 0.0008283111383207142, + 0.0009876909898594022, + 0.0005647471989504993, + 0.0007144779083319008, + 0.0008161084842868149, + 0.004525610711425543, + 0.0021482885349541903, + 0.0018734015757218003, + 0.004050839226692915, + 0.0025927803944796324, + 0.0015134350396692753, + 0.002554361242800951, + 0.0024820284452289343, + 0.002791751641780138, + 0.0014428379945456982, + 0.0020059256348758936, + 0.002499715657904744, + 0.0012733273906633258, + 0.0011077665258198977, + 0.0015840705018490553, + 0.0018935339758172631, + 0.0013236415106803179, + 0.004624920431524515, + 0.004591119941323996, + 0.00831245630979538, + 0.0008513235370628536, + 0.000606900779530406, + 0.0011506890878081322, + 0.001097247120924294 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 2, + 0 + ], + "selected_reward": -0.26069775223731995 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned in an extreme top-down close-up view?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera slowly zoom in on the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene lit by a single flickering candle source casting warm, low-key light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the Chinese calligraphy characters on the parchment begin to twist, stretch, and struggle as if alive?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the parchment surface pale ochre with dark brown water stains and frayed, curled edges?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.22563250362873077 + ], + "opposite_logprob_tokens": [ + -1.6006325483322144 + ], + "logprob_reward": -0.22563250362873077, + "yes_no_margin_reward": 1.375, + "token_credit_reward": -0.22563250362873077, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.05338224023580551, + 0.38278695940971375, + 0.0948178619146347, + 0.09841848909854889, + 0.17702479660511017, + 0.08698207885026932, + 0.006369102280586958, + 0.00809427723288536, + 0.004307548049837351, + 0.08781667053699493 + ], + "credit_map": [ + 0.0032995392102748156, + 0.0032243768218904734, + 0.01709185168147087, + 0.006202151533216238, + 0.005493463482707739, + 0.005163671914488077, + 0.005284913815557957, + 0.007622269447892904, + 0.024998804554343224, + 0.03038410097360611, + 0.012961427681148052, + 0.021107366308569908, + 0.03142426908016205, + 0.1536901593208313, + 0.09403310716152191, + 0.014187725260853767, + 0.006598749663680792, + 0.006884795613586903, + 0.003003104357048869, + 0.006491214968264103, + 0.006046878639608622, + 0.03183959797024727, + 0.026885688304901123, + 0.00706782191991806, + 0.006498280912637711, + 0.005857698153704405, + 0.02957063727080822, + 0.006533714476972818, + 0.007512862794101238, + 0.03171776980161667, + 0.003126439405605197, + 0.007601081393659115, + 0.013539036735892296, + 0.020753663033246994, + 0.016367336735129356, + 0.022810587659478188, + 0.033744603395462036, + 0.04098895564675331, + 0.014513645321130753, + 0.014306970871984959, + 0.005274943076074123, + 0.015968821942806244, + 0.006118378601968288, + 0.009421777911484241, + 0.023674046620726585, + 0.009189285337924957, + 0.007120317779481411, + 0.010214508511126041, + 0.0009031909285113215, + 0.0006832227227278054, + 0.0006821402930654585, + 0.0011076185619458556, + 0.0011182682355865836, + 0.0006347160669974983, + 0.0006998667377047241, + 0.0005400787922553718, + 0.0016743657179176807, + 0.000791520404163748, + 0.0007480507483705878, + 0.0014328474644571543, + 0.0010155210038647056, + 0.0006095789140090346, + 0.0008982066065073013, + 0.0009241857915185392, + 0.001119512366130948, + 0.0005002595135010779, + 0.0006563325878232718, + 0.0005059656105004251, + 0.000412392255384475, + 0.00028771706274710596, + 0.000323822139762342, + 0.0005015463684685528, + 0.005718561355024576, + 0.011650465428829193, + 0.02314630337059498, + 0.0349210761487484, + 0.002478764858096838, + 0.0015161011833697557, + 0.004290605895221233, + 0.0040947915986180305 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 4, + 3 + ], + "selected_reward": -0.22563250362873077 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned in an extreme top-down close-up view?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera slowly zoom in on the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene lit by a single flickering candle source casting warm, low-key light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the Chinese calligraphy characters on the parchment begin to twist, stretch, and struggle as if alive?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the parchment surface pale ochre with dark brown water stains and frayed, curled edges?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.14283332228660583 + ], + "opposite_logprob_tokens": [ + -2.0178332328796387 + ], + "logprob_reward": -0.14283332228660583, + "yes_no_margin_reward": 1.8749998807907104, + "token_credit_reward": -0.14283332228660583, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12361504882574081, + 0.05727406591176987, + 0.071977898478508, + 0.2361508309841156, + 0.12850619852542877, + 0.1586945652961731, + 0.03593776002526283, + 0.08342433720827103, + 0.01849851757287979, + 0.08592075854539871 + ], + "credit_map": [ + 0.012233266606926918, + 0.00950503908097744, + 0.030333859845995903, + 0.009426593780517578, + 0.009690669365227222, + 0.027943575754761696, + 0.011003329418599606, + 0.013478721491992474, + 0.004628206603229046, + 0.00460261944681406, + 0.003964605275541544, + 0.003406977280974388, + 0.0035369761753827333, + 0.019260717555880547, + 0.015732131898403168, + 0.0021418307442218065, + 0.013781026005744934, + 0.005532659590244293, + 0.006542473565787077, + 0.007200072985142469, + 0.0059343371540308, + 0.006443054415285587, + 0.01835356280207634, + 0.00819071102887392, + 0.011382625438272953, + 0.014786501415073872, + 0.07939980924129486, + 0.016061700880527496, + 0.011705203913152218, + 0.07457778602838516, + 0.008768838830292225, + 0.019468357786536217, + 0.008639681152999401, + 0.014317662455141544, + 0.009096737951040268, + 0.0185528751462698, + 0.023338191211223602, + 0.03208046406507492, + 0.012042549438774586, + 0.010438039898872375, + 0.010905232280492783, + 0.03184354677796364, + 0.012569310143589973, + 0.020296579226851463, + 0.04759091138839722, + 0.01883867383003235, + 0.012387020513415337, + 0.0042632995173335075, + 0.01038677804172039, + 0.003518253332003951, + 0.0032851274590939283, + 0.00593618955463171, + 0.0035166332963854074, + 0.0026849517598748207, + 0.002878776052966714, + 0.0037310512270778418, + 0.016332978382706642, + 0.011917227879166603, + 0.0076101417653262615, + 0.018983999267220497, + 0.007490760181099176, + 0.005313184577971697, + 0.008717880584299564, + 0.00705816550180316, + 0.0028110111597925425, + 0.0016497456235811114, + 0.0034389144275337458, + 0.002947497181594372, + 0.002539748093113303, + 0.0013271942734718323, + 0.0020036085043102503, + 0.0017807993572205305, + 0.007741008419543505, + 0.018075212836265564, + 0.025202639400959015, + 0.01919374242424965, + 0.005440316628664732, + 0.0033522879239171743, + 0.0024880303535610437, + 0.0044275252148509026 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 5, + 4 + ], + "selected_reward": -0.14283332228660583 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned in an extreme top-down close-up view?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera slowly zoom in on the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene lit by a single flickering candle source casting warm, low-key light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the Chinese calligraphy characters on the parchment begin to twist, stretch, and struggle as if alive?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the parchment surface pale ochre with dark brown water stains and frayed, curled edges?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -0.8261452913284302 + ], + "opposite_logprob_tokens": [ + -0.5761452913284302 + ], + "logprob_reward": -0.8261452913284302, + "yes_no_margin_reward": -0.25, + "token_credit_reward": -0.8261452913284302, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.04927647486329079, + 0.4692935347557068, + 0.07193878293037415, + 0.14778584241867065, + 0.04126555845141411, + 0.10994862765073776, + 0.014646291732788086, + 0.01490081287920475, + 0.00514701334759593, + 0.07579706609249115 + ], + "credit_map": [ + 0.005249773617833853, + 0.003800897393375635, + 0.011613796465098858, + 0.003292480017989874, + 0.004352387972176075, + 0.010378251783549786, + 0.004253932740539312, + 0.006334953010082245, + 0.03164590150117874, + 0.0347270593047142, + 0.01709693856537342, + 0.029404090717434883, + 0.03006569668650627, + 0.1688121110200882, + 0.13208264112472534, + 0.02545909397304058, + 0.006927249021828175, + 0.00506117008626461, + 0.0027836766093969345, + 0.004988390486687422, + 0.00433387840166688, + 0.021245690062642097, + 0.021319687366485596, + 0.005279036238789558, + 0.008162662386894226, + 0.009556208737194538, + 0.043230943381786346, + 0.00886576622724533, + 0.012134905904531479, + 0.05203996226191521, + 0.004677900113165379, + 0.009117497131228447, + 0.0030169186647981405, + 0.0037510604597628117, + 0.009243052452802658, + 0.005476390942931175, + 0.004793860483914614, + 0.0023400061763823032, + 0.005634805653244257, + 0.007009468041360378, + 0.005519606173038483, + 0.01638895645737648, + 0.02296493761241436, + 0.021192220970988274, + 0.023547222837805748, + 0.009797514416277409, + 0.005538302939385176, + 0.004999866243451834, + 0.0016627830918878317, + 0.0020767971873283386, + 0.0016198507510125637, + 0.0017129445914179087, + 0.003453347599133849, + 0.001379164052195847, + 0.0017306990921497345, + 0.0010107059497386217, + 0.003345274366438389, + 0.0016940707573667169, + 0.0015403615543618798, + 0.0033140024170279503, + 0.0009766760049387813, + 0.0009224263485521078, + 0.001649691490456462, + 0.0014583109878003597, + 0.0005929818144068122, + 0.0006743327830918133, + 0.0008711858536116779, + 0.0009793206118047237, + 0.000658828706946224, + 0.0003596627793740481, + 0.0005766496178694069, + 0.00043405103497207165, + 0.006056983023881912, + 0.006535734515637159, + 0.02522665075957775, + 0.024883897975087166, + 0.0035315800923854113, + 0.001858494826592505, + 0.002852199599146843, + 0.004851518664509058 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 3, + 5 + ], + "selected_reward": -0.8261452913284302 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned in an extreme top-down close-up view?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera slowly zoom in on the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene lit by a single flickering candle source casting warm, low-key light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the Chinese calligraphy characters on the parchment begin to twist, stretch, and struggle as if alive?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the parchment surface pale ochre with dark brown water stains and frayed, curled edges?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.029797164723277092 + ], + "opposite_logprob_tokens": [ + -3.529797077178955 + ], + "logprob_reward": -0.029797164723277092, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.029797164723277092, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07488991320133209, + 0.4350438117980957, + 0.19203273952007294, + 0.07488394528627396, + 0.03824622929096222, + 0.07292551547288895, + 0.011557593010365963, + 0.014739759266376495, + 0.006749863736331463, + 0.07893054932355881 + ], + "credit_map": [ + 0.007460249587893486, + 0.006563662085682154, + 0.019014447927474976, + 0.006664039101451635, + 0.006640504579991102, + 0.006235088221728802, + 0.0061407387256622314, + 0.01617119275033474, + 0.02838546223938465, + 0.03397081047296524, + 0.016079094260931015, + 0.023936444893479347, + 0.03429103642702103, + 0.17120996117591858, + 0.11072024703025818, + 0.016450831666588783, + 0.012458436191082, + 0.01729614846408367, + 0.006477505434304476, + 0.009592217393219471, + 0.016626132652163506, + 0.07981667667627335, + 0.04173990711569786, + 0.00802575796842575, + 0.005657814908772707, + 0.0034958787728101015, + 0.023005034774541855, + 0.005876736715435982, + 0.007575880736112595, + 0.013055686838924885, + 0.004166450817137957, + 0.012050469405949116, + 0.0038695638068020344, + 0.004740212112665176, + 0.0043283989652991295, + 0.006584757007658482, + 0.0066476562060415745, + 0.006021065171808004, + 0.0036628234665840864, + 0.0023917548824101686, + 0.004338318947702646, + 0.012918607331812382, + 0.01114579476416111, + 0.00938164722174406, + 0.018283545970916748, + 0.0077201128005981445, + 0.005048631690442562, + 0.004088865127414465, + 0.001286617829464376, + 0.0012316834181547165, + 0.0013554670149460435, + 0.0022793440148234367, + 0.0021349159069359303, + 0.0009135431027971208, + 0.00137491209898144, + 0.000981111079454422, + 0.0035080809611827135, + 0.002229374600574374, + 0.0013882744824513793, + 0.0028440363239496946, + 0.0010348346550017595, + 0.0008833649917505682, + 0.0014461672399193048, + 0.001405628165230155, + 0.0008704809006303549, + 0.0008704240317456424, + 0.0009150424739345908, + 0.0008877281215973198, + 0.001341352122835815, + 0.0007052717264741659, + 0.0006561946356669068, + 0.0005033703637309372, + 0.004342623054981232, + 0.021433383226394653, + 0.011378522962331772, + 0.027381451800465584, + 0.004416787996888161, + 0.0032755499705672264, + 0.00381146976724267, + 0.0028907705564051867 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 2, + 9 + ], + "selected_reward": -0.029797164723277092 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "No", + "Yes" + ] + }, + { + "prompt": "dramatic, extreme top-down close-up slowly zooms in on a sheet of weathered parchment, lying flat on a black, polished wooden table. The entire scene is bathed in warm, low-key lighting cast by a single, flickering candle just out of frame, its wavering light causing shadows to dance and highlighting rich textures. The parchment itself is a map of time, its surface a pale ochre stained with dark brown water stains and frayed along its delicately curled edges. Across this ancient canvas, a complex column of Chinese characters, written in a flowing calligraphic script, shimmers with wet ink. Initially static, the elegant forms begin to stir; subtle, unsettling pulsations ripple through the broadest brushstrokes. Then, with a silent, desperate energy, the figures begin to twist and struggle. Their elegant lines twist unnaturally, stretching as if exhausted, then violently contracting, the liquid ink bleeding and reforming as the roots of each symbol pull apart and collide in a chaotic, living dance. The effect is both beautiful and terrifying, creating a tense, mysterious atmosphere of both ancient magic awakening and an entity's struggle to escape its two-dimensional prison.", + "seed": 23998, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000034/group_15/index_9829_seed_23998_dp_31_rank_62_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 15, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.10959816984832287, + "vlm_logprob": -0.10959816984832287, + "vlm_yes_no_margin": 2.6, + "vlm_token_credit": -0.10959816984832287, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned in an extreme top-down close-up view?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera slowly zoom in on the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene lit by a single flickering candle source casting warm, low-key light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the Chinese calligraphy characters on the parchment begin to twist, stretch, and struggle as if alive?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the parchment surface pale ochre with dark brown water stains and frayed, curled edges?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.10675016045570374 + ], + "opposite_logprob_tokens": [ + -2.481750249862671 + ], + "logprob_reward": -0.10675016045570374, + "yes_no_margin_reward": 2.375, + "token_credit_reward": -0.10675016045570374, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1987331211566925, + 0.07735785096883774, + 0.14117662608623505, + 0.15720032155513763, + 0.11030039191246033, + 0.12433336675167084, + 0.0590798445045948, + 0.015745827928185463, + 0.03168519213795662, + 0.08438742160797119 + ], + "credit_map": [ + 0.020889554172754288, + 0.026671327650547028, + 0.03141327574849129, + 0.023739207535982132, + 0.023313293233513832, + 0.023356806486845016, + 0.02102961577475071, + 0.02832004241645336, + 0.011298428289592266, + 0.011441949754953384, + 0.0088680824264884, + 0.009908394888043404, + 0.009244553744792938, + 0.00893041305243969, + 0.008085393346846104, + 0.009580641984939575, + 0.019718503579497337, + 0.01697429269552231, + 0.016703765839338303, + 0.02061120979487896, + 0.012870647013187408, + 0.02073819376528263, + 0.017937788739800453, + 0.015622223727405071, + 0.009329330176115036, + 0.010735556483268738, + 0.010324040427803993, + 0.06057732179760933, + 0.008214994333684444, + 0.009030911140143871, + 0.022128576412796974, + 0.026859594509005547, + 0.012140472419559956, + 0.024526681751012802, + 0.009902658872306347, + 0.011388844810426235, + 0.021274356171488762, + 0.00958250556141138, + 0.012577883899211884, + 0.008906987495720387, + 0.016276979818940163, + 0.01480892114341259, + 0.014505142346024513, + 0.016729019582271576, + 0.024155041202902794, + 0.01322627067565918, + 0.015905063599348068, + 0.008726922795176506, + 0.006949576083570719, + 0.008628671057522297, + 0.004873925354331732, + 0.009043588303029537, + 0.014340616762638092, + 0.005496544297784567, + 0.005611551459878683, + 0.004135369323194027, + 0.002278036903589964, + 0.001793196308426559, + 0.00201960070990026, + 0.001974182203412056, + 0.0023094983771443367, + 0.0020942518021911383, + 0.0018700837390497327, + 0.0014069783501327038, + 0.0030307485722005367, + 0.003819496603682637, + 0.0032059422228485346, + 0.003786335000768304, + 0.004164370242506266, + 0.003987730015069246, + 0.007134464103728533, + 0.0025561051443219185, + 0.009218419902026653, + 0.018055453896522522, + 0.00992636103183031, + 0.009790671057999134, + 0.010718252509832382, + 0.00915384292602539, + 0.009055118076503277, + 0.00846930406987667 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 2 + ], + "selected_reward": -0.10675016045570374 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned in an extreme top-down close-up view?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera slowly zoom in on the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene lit by a single flickering candle source casting warm, low-key light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the Chinese calligraphy characters on the parchment begin to twist, stretch, and struggle as if alive?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the parchment surface pale ochre with dark brown water stains and frayed, curled edges?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.25212955474853516 + ], + "opposite_logprob_tokens": [ + -1.5021295547485352 + ], + "logprob_reward": -0.25212955474853516, + "yes_no_margin_reward": 1.25, + "token_credit_reward": -0.25212955474853516, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11497122049331665, + 0.04695730656385422, + 0.13929782807826996, + 0.18864575028419495, + 0.049751754850149155, + 0.04610561951994896, + 0.19596219062805176, + 0.032360322773456573, + 0.06801773607730865, + 0.11793025583028793 + ], + "credit_map": [ + 0.015065294690430164, + 0.013546514324843884, + 0.01498731691390276, + 0.013257568702101707, + 0.018615247681736946, + 0.009982699528336525, + 0.015295411460101604, + 0.014221174642443657, + 0.007861600257456303, + 0.005019607488065958, + 0.00557587668299675, + 0.005411558318883181, + 0.004740847274661064, + 0.005856927949935198, + 0.006708808243274689, + 0.005782080814242363, + 0.017459947615861893, + 0.014187543652951717, + 0.021599357947707176, + 0.019154991954565048, + 0.012251113541424274, + 0.018028633669018745, + 0.019136881455779076, + 0.017479371279478073, + 0.01575450785458088, + 0.016905395314097404, + 0.0152836162596941, + 0.06868577748537064, + 0.01089255791157484, + 0.011813749559223652, + 0.02478627860546112, + 0.02452387474477291, + 0.0048369052819907665, + 0.00812492985278368, + 0.005527841858565807, + 0.005937060806900263, + 0.011285325512290001, + 0.004017652478069067, + 0.006553528364747763, + 0.0034685127902776003, + 0.0056292531080543995, + 0.005854768678545952, + 0.004121157340705395, + 0.006257758010178804, + 0.010733242146670818, + 0.003926322795450687, + 0.007124523166567087, + 0.0024585952050983906, + 0.021295320242643356, + 0.029739946126937866, + 0.01579182967543602, + 0.0294113140553236, + 0.04960251972079277, + 0.01919451355934143, + 0.020108245313167572, + 0.010818497277796268, + 0.00410980312153697, + 0.003500312799587846, + 0.00447789765894413, + 0.00477562053129077, + 0.00463502062484622, + 0.003523613093420863, + 0.004005317110568285, + 0.00333273783326149, + 0.005880668759346008, + 0.006543298251926899, + 0.008593861013650894, + 0.007905460894107819, + 0.00837088655680418, + 0.005365604534745216, + 0.02044478803873062, + 0.004913171287626028, + 0.010348387993872166, + 0.02005971595644951, + 0.011258282698690891, + 0.01642805151641369, + 0.014482679776847363, + 0.010548115707933903, + 0.023798413574695587, + 0.01100661139935255 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 3, + 2 + ], + "selected_reward": -0.25212955474853516 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned in an extreme top-down close-up view?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera slowly zoom in on the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene lit by a single flickering candle source casting warm, low-key light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the Chinese calligraphy characters on the parchment begin to twist, stretch, and struggle as if alive?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the parchment surface pale ochre with dark brown water stains and frayed, curled edges?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.10033102333545685 + ], + "opposite_logprob_tokens": [ + -2.3503310680389404 + ], + "logprob_reward": -0.10033102333545685, + "yes_no_margin_reward": 2.25, + "token_credit_reward": -0.10033102333545685, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.17917270958423615, + 0.042551588267087936, + 0.043617118149995804, + 0.38450735807418823, + 0.0647694393992424, + 0.08749343454837799, + 0.046146683394908905, + 0.050072528421878815, + 0.033517986536026, + 0.068151094019413 + ], + "credit_map": [ + 0.021131837740540504, + 0.01837521605193615, + 0.03111015073955059, + 0.01970924437046051, + 0.019621197134256363, + 0.022365672513842583, + 0.017372433096170425, + 0.02948695793747902, + 0.006294903811067343, + 0.0045849718153476715, + 0.007830020971596241, + 0.004418451804667711, + 0.00325009529478848, + 0.0054318867623806, + 0.005026334896683693, + 0.005714923609048128, + 0.005249270237982273, + 0.0036457872483879328, + 0.007748572155833244, + 0.005633159540593624, + 0.0034073616843670607, + 0.0054433103650808334, + 0.006988828536123037, + 0.00550082977861166, + 0.023823361843824387, + 0.028365585952997208, + 0.019325602799654007, + 0.1641141027212143, + 0.015585871413350105, + 0.014061754569411278, + 0.05319580063223839, + 0.06603531539440155, + 0.006110494490712881, + 0.0077662174589931965, + 0.00469548674300313, + 0.012084227055311203, + 0.0073480564169585705, + 0.004902874119579792, + 0.01654135063290596, + 0.0053207301534712315, + 0.0104506965726614, + 0.015485208481550217, + 0.009102427400648594, + 0.012695186771452427, + 0.016898903995752335, + 0.007664720062166452, + 0.010090885683894157, + 0.005105405114591122, + 0.006158412899821997, + 0.006089695729315281, + 0.004349434748291969, + 0.0060607073828577995, + 0.011165320873260498, + 0.004387460183352232, + 0.0057442388497292995, + 0.0021914150565862656, + 0.005610523279756308, + 0.005489119794219732, + 0.007282652892172337, + 0.007016095332801342, + 0.007575681898742914, + 0.006814535707235336, + 0.0054525351151824, + 0.00483138719573617, + 0.0031234908383339643, + 0.0038160732947289944, + 0.004746967926621437, + 0.0037049928214401007, + 0.0039850822649896145, + 0.002846389776095748, + 0.007909806445240974, + 0.0033851847983896732, + 0.008247625082731247, + 0.012757711112499237, + 0.006888322066515684, + 0.00822495948523283, + 0.00921833235770464, + 0.005893446039408445, + 0.008033191785216331, + 0.00888750609010458 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 0, + 5 + ], + "selected_reward": -0.10033102333545685 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned in an extreme top-down close-up view?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera slowly zoom in on the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene lit by a single flickering candle source casting warm, low-key light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the Chinese calligraphy characters on the parchment begin to twist, stretch, and struggle as if alive?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the parchment surface pale ochre with dark brown water stains and frayed, curled edges?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07900354266166687 + ], + "opposite_logprob_tokens": [ + -2.5790035724639893 + ], + "logprob_reward": -0.07900354266166687, + "yes_no_margin_reward": 2.5, + "token_credit_reward": -0.07900354266166687, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11025342345237732, + 0.03895285353064537, + 0.09622541069984436, + 0.2658931612968445, + 0.062356170266866684, + 0.05726027116179466, + 0.0985516756772995, + 0.042600180953741074, + 0.1384614259004593, + 0.08944537490606308 + ], + "credit_map": [ + 0.011830642819404602, + 0.010986262932419777, + 0.01861041598021984, + 0.01165750715881586, + 0.014807515777647495, + 0.012979519553482533, + 0.012197407893836498, + 0.017184162512421608, + 0.005535654723644257, + 0.0037721297703683376, + 0.005835837684571743, + 0.0045557389967143536, + 0.0035578333772718906, + 0.005670924670994282, + 0.0044412692077457905, + 0.005583468358963728, + 0.013154307380318642, + 0.008514895103871822, + 0.012632091529667377, + 0.011160346679389477, + 0.01022359449416399, + 0.011219064705073833, + 0.015349493362009525, + 0.013971622101962566, + 0.01756911166012287, + 0.020604291930794716, + 0.02035415545105934, + 0.10606025904417038, + 0.01197289489209652, + 0.01347182784229517, + 0.03578348830342293, + 0.04007716104388237, + 0.005888679996132851, + 0.009557107463479042, + 0.005046780686825514, + 0.007216399535536766, + 0.012950718402862549, + 0.004968181252479553, + 0.01082996279001236, + 0.0058983443304896355, + 0.005885726772248745, + 0.009045504033565521, + 0.007201558444648981, + 0.009370729327201843, + 0.008202062919735909, + 0.0061821467243134975, + 0.007513510994613171, + 0.0038590359035879374, + 0.01128404214978218, + 0.015785489231348038, + 0.007995812222361565, + 0.015679627656936646, + 0.02299603819847107, + 0.00918530486524105, + 0.009642664343118668, + 0.0059827095828950405, + 0.006194532383233309, + 0.004806490149348974, + 0.005196586716920137, + 0.006598914507776499, + 0.006070658098906279, + 0.004943672567605972, + 0.005180399864912033, + 0.003608930390328169, + 0.01429294515401125, + 0.01414328534156084, + 0.01180508453398943, + 0.01661636307835579, + 0.022209221497178078, + 0.014575131237506866, + 0.03344319015741348, + 0.011376209557056427, + 0.010447422973811626, + 0.017925919964909554, + 0.006821839604526758, + 0.01234277244657278, + 0.013222011737525463, + 0.007582406047731638, + 0.01118886936455965, + 0.009914142079651356 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 8, + 0 + ], + "selected_reward": -0.07900354266166687 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned in an extreme top-down close-up view?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera slowly zoom in on the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene lit by a single flickering candle source casting warm, low-key light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the Chinese calligraphy characters on the parchment begin to twist, stretch, and struggle as if alive?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the parchment surface pale ochre with dark brown water stains and frayed, curled edges?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.009776568040251732 + ], + "opposite_logprob_tokens": [ + -4.634776592254639 + ], + "logprob_reward": -0.009776568040251732, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.009776568040251732, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15489763021469116, + 0.1607133001089096, + 0.06403719633817673, + 0.08174155652523041, + 0.08078641444444656, + 0.06389535218477249, + 0.06011926010251045, + 0.05845583230257034, + 0.05147626996040344, + 0.2238771915435791 + ], + "credit_map": [ + 0.016202379018068314, + 0.01956399716436863, + 0.026226546615362167, + 0.013953532092273235, + 0.01425153948366642, + 0.021227477118372917, + 0.020910492166876793, + 0.022561663761734962, + 0.019080396741628647, + 0.025354066863656044, + 0.020334163680672646, + 0.018574178218841553, + 0.02154938131570816, + 0.01789225824177265, + 0.017136776819825172, + 0.02079208754003048, + 0.008861626498401165, + 0.006305737420916557, + 0.00982669834047556, + 0.0075822798535227776, + 0.005339100491255522, + 0.008177753537893295, + 0.009159868583083153, + 0.008784126490354538, + 0.005263790488243103, + 0.007261835504323244, + 0.006052232347428799, + 0.030130432918667793, + 0.004385894630104303, + 0.004145563114434481, + 0.01082965824753046, + 0.013672147877514362, + 0.007971379905939102, + 0.016967648640275, + 0.005973182152956724, + 0.01051364652812481, + 0.015836823731660843, + 0.007972034625709057, + 0.010225963778793812, + 0.005325739737600088, + 0.008654515258967876, + 0.008223188109695911, + 0.008453858084976673, + 0.009716861881315708, + 0.008766761980950832, + 0.006482748780399561, + 0.007927456870675087, + 0.005669955629855394, + 0.007065345998853445, + 0.006861464586108923, + 0.007202209439128637, + 0.00870245136320591, + 0.012494615279138088, + 0.005095238797366619, + 0.007385430857539177, + 0.005312500987201929, + 0.006369623821228743, + 0.009876021184027195, + 0.00664006220176816, + 0.008115577511489391, + 0.009841243736445904, + 0.00638786144554615, + 0.006539518013596535, + 0.004685927182435989, + 0.004400866571813822, + 0.004680689889937639, + 0.007350465748459101, + 0.0067366063594818115, + 0.0059425365179777145, + 0.003969318233430386, + 0.013633745722472668, + 0.004762039519846439, + 0.025621531531214714, + 0.04786049574613571, + 0.016514645889401436, + 0.03376828506588936, + 0.03186367452144623, + 0.014858195558190346, + 0.02371368184685707, + 0.02967667020857334 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 1, + 0 + ], + "selected_reward": -0.009776568040251732 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "dramatic, extreme top-down close-up slowly zooms in on a sheet of weathered parchment, lying flat on a black, polished wooden table. The entire scene is bathed in warm, low-key lighting cast by a single, flickering candle just out of frame, its wavering light causing shadows to dance and highlighting rich textures. The parchment itself is a map of time, its surface a pale ochre stained with dark brown water stains and frayed along its delicately curled edges. Across this ancient canvas, a complex column of Chinese characters, written in a flowing calligraphic script, shimmers with wet ink. Initially static, the elegant forms begin to stir; subtle, unsettling pulsations ripple through the broadest brushstrokes. Then, with a silent, desperate energy, the figures begin to twist and struggle. Their elegant lines twist unnaturally, stretching as if exhausted, then violently contracting, the liquid ink bleeding and reforming as the roots of each symbol pull apart and collide in a chaotic, living dance. The effect is both beautiful and terrifying, creating a tense, mysterious atmosphere of both ancient magic awakening and an entity's struggle to escape its two-dimensional prison.", + "seed": 23999, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000034/group_15/index_9829_seed_23999_dp_31_rank_62_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 15, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.22247273549437524, + "vlm_logprob": -0.22247273549437524, + "vlm_yes_no_margin": 2.05, + "vlm_token_credit": -0.22247273549437524, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned in an extreme top-down close-up view?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera slowly zoom in on the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene lit by a single flickering candle source casting warm, low-key light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the Chinese calligraphy characters on the parchment begin to twist, stretch, and struggle as if alive?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the parchment surface pale ochre with dark brown water stains and frayed, curled edges?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.2704135775566101 + ], + "opposite_logprob_tokens": [ + -1.6454136371612549 + ], + "logprob_reward": -0.2704135775566101, + "yes_no_margin_reward": 1.375, + "token_credit_reward": -0.2704135775566101, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.29600852727890015, + 0.1354829967021942, + 0.06334494054317474, + 0.05349152535200119, + 0.22180834412574768, + 0.035919953137636185, + 0.06233067438006401, + 0.0632784515619278, + 0.033775512129068375, + 0.03455902263522148 + ], + "credit_map": [ + 0.053694650530815125, + 0.03320164605975151, + 0.032204996794462204, + 0.02069387584924698, + 0.050158336758613586, + 0.04742952063679695, + 0.02277371473610401, + 0.035851817578077316, + 0.013363544829189777, + 0.013330969959497452, + 0.022342558950185776, + 0.01682383194565773, + 0.014739331789314747, + 0.021683096885681152, + 0.021155789494514465, + 0.012043887749314308, + 0.005581662058830261, + 0.008307218551635742, + 0.007354033179581165, + 0.007763794623315334, + 0.006497515831142664, + 0.008013831451535225, + 0.011447533033788204, + 0.008379359729588032, + 0.007538387086242437, + 0.005176525097340345, + 0.004817259963601828, + 0.008268882520496845, + 0.010290224105119705, + 0.007860478945076466, + 0.005198410712182522, + 0.0043413578532636166, + 0.02137881889939308, + 0.02663007564842701, + 0.022494647651910782, + 0.03632885962724686, + 0.033209674060344696, + 0.022350436076521873, + 0.02777903899550438, + 0.03163681551814079, + 0.003453989513218403, + 0.0034290559124201536, + 0.0035456621553748846, + 0.005877183750271797, + 0.00563902547582984, + 0.004179147072136402, + 0.004863525740802288, + 0.00493236631155014, + 0.007398388348519802, + 0.01020553894340992, + 0.004157970659434795, + 0.009033969603478909, + 0.0068924687802791595, + 0.005661478266119957, + 0.01171094086021185, + 0.007269926369190216, + 0.0072251721285283566, + 0.009712829254567623, + 0.007433818653225899, + 0.01154246460646391, + 0.008962583728134632, + 0.004340053535997868, + 0.008033613674342632, + 0.006027923431247473, + 0.0036995410919189453, + 0.006578948348760605, + 0.0034983735531568527, + 0.005064499098807573, + 0.003653785213828087, + 0.0021151674445718527, + 0.0057794940657913685, + 0.003385707503184676, + 0.004806722979992628, + 0.005696249660104513, + 0.003060184419155121, + 0.005374146159738302, + 0.004688170272856951, + 0.0024756730999797583, + 0.005085688550025225, + 0.0033721912186592817 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 1 + ], + "selected_reward": -0.2704135775566101 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned in an extreme top-down close-up view?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera slowly zoom in on the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene lit by a single flickering candle source casting warm, low-key light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the Chinese calligraphy characters on the parchment begin to twist, stretch, and struggle as if alive?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the parchment surface pale ochre with dark brown water stains and frayed, curled edges?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.3134925663471222 + ], + "opposite_logprob_tokens": [ + -1.3134925365447998 + ], + "logprob_reward": -0.3134925663471222, + "yes_no_margin_reward": 1.0, + "token_credit_reward": -0.3134925663471222, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.21002116799354553, + 0.08986006677150726, + 0.0909360870718956, + 0.18723788857460022, + 0.17569132149219513, + 0.05994485691189766, + 0.04686955735087395, + 0.03722413256764412, + 0.04070165380835533, + 0.06151330843567848 + ], + "credit_map": [ + 0.03172062337398529, + 0.0259561724960804, + 0.026422185823321342, + 0.017177319154143333, + 0.032879866659641266, + 0.03658989816904068, + 0.013865988701581955, + 0.025409115478396416, + 0.013728600926697254, + 0.008915583603084087, + 0.007873525843024254, + 0.008046762086451054, + 0.015826459974050522, + 0.014003862626850605, + 0.010647877119481564, + 0.010817390866577625, + 0.008875560015439987, + 0.011159971356391907, + 0.009025431238114834, + 0.011011708527803421, + 0.010876663029193878, + 0.012902434915304184, + 0.013048780150711536, + 0.014035526663064957, + 0.02107933536171913, + 0.01914946176111698, + 0.013377026654779911, + 0.024939555674791336, + 0.05258302390575409, + 0.025095172226428986, + 0.012797868810594082, + 0.018216434866189957, + 0.016987349838018417, + 0.022641802206635475, + 0.01878250762820244, + 0.028265155851840973, + 0.023929359391331673, + 0.019182469695806503, + 0.021380696445703506, + 0.02452196180820465, + 0.006422378588467836, + 0.00761215528473258, + 0.008287467993795872, + 0.010534390807151794, + 0.005502653308212757, + 0.0053264060989022255, + 0.006728922948241234, + 0.009530478157103062, + 0.004214730579406023, + 0.006208904553204775, + 0.004294993355870247, + 0.0099493945017457, + 0.008523792959749699, + 0.0031912284903228283, + 0.007094480562955141, + 0.003392027923837304, + 0.0037543857470154762, + 0.004978375509381294, + 0.0037471663672477007, + 0.007671660743653774, + 0.005455804988741875, + 0.0029651601798832417, + 0.005313432309776545, + 0.003338142763823271, + 0.004348120652139187, + 0.005785553716123104, + 0.004066838882863522, + 0.005690207704901695, + 0.0077831316739320755, + 0.0033836220391094685, + 0.005953125189989805, + 0.0036910506896674633, + 0.009115779772400856, + 0.011026599444448948, + 0.0058546182699501514, + 0.009454101324081421, + 0.007247246336191893, + 0.003625744953751564, + 0.009566139429807663, + 0.005623075645416975 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 4 + ], + "selected_reward": -0.3134925663471222 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned in an extreme top-down close-up view?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera slowly zoom in on the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene lit by a single flickering candle source casting warm, low-key light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the Chinese calligraphy characters on the parchment begin to twist, stretch, and struggle as if alive?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the parchment surface pale ochre with dark brown water stains and frayed, curled edges?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0381188727915287 + ], + "opposite_logprob_tokens": [ + -3.288118839263916 + ], + "logprob_reward": -0.0381188727915287, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.0381188727915287, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1492469757795334, + 0.042375512421131134, + 0.23600266873836517, + 0.2963716983795166, + 0.09671113640069962, + 0.04861803725361824, + 0.03197840228676796, + 0.014119892381131649, + 0.04890057072043419, + 0.035675037652254105 + ], + "credit_map": [ + 0.021008117124438286, + 0.017194848507642746, + 0.016048887744545937, + 0.013673623092472553, + 0.021582983434200287, + 0.025918370112776756, + 0.015061236917972565, + 0.01875891350209713, + 0.0044496748596429825, + 0.004307654220610857, + 0.006767566781491041, + 0.006738228723406792, + 0.004602216184139252, + 0.004986145533621311, + 0.0065810359083116055, + 0.003942997194826603, + 0.025690987706184387, + 0.022347353398799896, + 0.027231473475694656, + 0.028944291174411774, + 0.050846461206674576, + 0.038838569074869156, + 0.018370017409324646, + 0.02373354509472847, + 0.03206667676568031, + 0.02972627803683281, + 0.018885238096117973, + 0.0335521399974823, + 0.09886260330677032, + 0.0399152971804142, + 0.019229965284466743, + 0.024133561179041862, + 0.008966432884335518, + 0.011117909103631973, + 0.008244423195719719, + 0.011576935648918152, + 0.023706374689936638, + 0.011631348170340061, + 0.010568193159997463, + 0.010899532586336136, + 0.005444483831524849, + 0.00483404565602541, + 0.007267007604241371, + 0.007609101478010416, + 0.006080898456275463, + 0.0049553681164979935, + 0.00497741624712944, + 0.00744971726089716, + 0.0031137890182435513, + 0.0034248267766088247, + 0.002953089540824294, + 0.005761040840297937, + 0.00409125117585063, + 0.003984928131103516, + 0.0047356910072267056, + 0.003913788124918938, + 0.0015063327737152576, + 0.0013704127632081509, + 0.0012786497827619314, + 0.0018746014684438705, + 0.002699699718505144, + 0.0013802085304632783, + 0.0026529913302510977, + 0.0013569977600127459, + 0.005580686964094639, + 0.006851432844996452, + 0.004130564630031586, + 0.0075887818820774555, + 0.012023965828120708, + 0.003348770085722208, + 0.005953615996986628, + 0.0034227590076625347, + 0.004727684427052736, + 0.005724865477532148, + 0.003252701135352254, + 0.006192876026034355, + 0.004057412967085838, + 0.0026907112915068865, + 0.005191803444176912, + 0.003836986841633916 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 2, + 0 + ], + "selected_reward": -0.0381188727915287 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned in an extreme top-down close-up view?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera slowly zoom in on the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene lit by a single flickering candle source casting warm, low-key light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the Chinese calligraphy characters on the parchment begin to twist, stretch, and struggle as if alive?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the parchment surface pale ochre with dark brown water stains and frayed, curled edges?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.47427451610565186 + ], + "opposite_logprob_tokens": [ + -0.9742745161056519 + ], + "logprob_reward": -0.47427451610565186, + "yes_no_margin_reward": 0.5, + "token_credit_reward": -0.47427451610565186, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.16525642573833466, + 0.0438002310693264, + 0.2162957489490509, + 0.24683599174022675, + 0.16006870567798615, + 0.0551481619477272, + 0.04185708612203598, + 0.01723264716565609, + 0.02655390091240406, + 0.026951132342219353 + ], + "credit_map": [ + 0.0253964401781559, + 0.020365580916404724, + 0.01780427247285843, + 0.016393594443798065, + 0.02372671663761139, + 0.02615647204220295, + 0.01581413298845291, + 0.01959921233355999, + 0.004026923794299364, + 0.00417557405307889, + 0.0064905076287686825, + 0.004655926954001188, + 0.0056898826733231544, + 0.005831154994666576, + 0.006620706059038639, + 0.0063095553778111935, + 0.023624686524271965, + 0.02089543454349041, + 0.021417541429400444, + 0.026626842096447945, + 0.04161114618182182, + 0.036057062447071075, + 0.02329174242913723, + 0.022771304473280907, + 0.02820241078734398, + 0.025875620543956757, + 0.015845317393541336, + 0.02766132913529873, + 0.08146399259567261, + 0.031836409121751785, + 0.013120237737894058, + 0.02283068187534809, + 0.013699078932404518, + 0.019539859145879745, + 0.013236507773399353, + 0.0258122980594635, + 0.02421514317393303, + 0.01603306084871292, + 0.02827387861907482, + 0.019258877262473106, + 0.004376883152872324, + 0.007846834138035774, + 0.004668571520596743, + 0.010557535104453564, + 0.00486355647444725, + 0.004290413577109575, + 0.012570385821163654, + 0.00597397843375802, + 0.00408810144290328, + 0.004250260069966316, + 0.004853148013353348, + 0.007247020956128836, + 0.004783587995916605, + 0.004831037484109402, + 0.00642756512388587, + 0.005376364104449749, + 0.0020602652803063393, + 0.0025637478101998568, + 0.0015557578299194574, + 0.0023545431904494762, + 0.0032693487592041492, + 0.0011940159602090716, + 0.0024192892014980316, + 0.0018156793667003512, + 0.0029594909865409136, + 0.0045774755999445915, + 0.0022353234235197306, + 0.003544487990438938, + 0.004434212576597929, + 0.002118184231221676, + 0.003930929582566023, + 0.002753797685727477, + 0.002996777417138219, + 0.003797029610723257, + 0.0027549276128411293, + 0.004035371821373701, + 0.005010770633816719, + 0.001992863602936268, + 0.004041776992380619, + 0.0023216125555336475 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 2, + 0 + ], + "selected_reward": -0.47427451610565186 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned in an extreme top-down close-up view?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera slowly zoom in on the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene lit by a single flickering candle source casting warm, low-key light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the Chinese calligraphy characters on the parchment begin to twist, stretch, and struggle as if alive?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the parchment surface pale ochre with dark brown water stains and frayed, curled edges?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.016064144670963287 + ], + "opposite_logprob_tokens": [ + -4.141064167022705 + ], + "logprob_reward": -0.016064144670963287, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.016064144670963287, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1771591603755951, + 0.16492044925689697, + 0.10511043667793274, + 0.09804710745811462, + 0.14268095791339874, + 0.1564026176929474, + 0.06664467602968216, + 0.033804893493652344, + 0.027406027540564537, + 0.027823656797409058 + ], + "credit_map": [ + 0.023740489035844803, + 0.018830319866538048, + 0.027385998517274857, + 0.018855096772313118, + 0.02162940241396427, + 0.026405498385429382, + 0.014206081628799438, + 0.026106271892786026, + 0.019992299377918243, + 0.021339302882552147, + 0.018835676833987236, + 0.016976412385702133, + 0.019963737577199936, + 0.024771885946393013, + 0.02149815857410431, + 0.02154296450316906, + 0.010682601481676102, + 0.012281322851777077, + 0.011103327386081219, + 0.012627940624952316, + 0.008494098670780659, + 0.014848767779767513, + 0.01850135065615177, + 0.01657102070748806, + 0.011582421138882637, + 0.009264403954148293, + 0.00837485771626234, + 0.010087717324495316, + 0.025700032711029053, + 0.01306813396513462, + 0.00890044029802084, + 0.011069103144109249, + 0.01704258844256401, + 0.017244115471839905, + 0.016512827947735786, + 0.018727706745266914, + 0.01588430628180504, + 0.01887388899922371, + 0.024736689403653145, + 0.013658839277923107, + 0.01492296438664198, + 0.01979820244014263, + 0.020821023732423782, + 0.027400990948081017, + 0.012229404412209988, + 0.014424219727516174, + 0.02832152508199215, + 0.018484294414520264, + 0.008552628569304943, + 0.007682060357183218, + 0.006786026526242495, + 0.010833495296537876, + 0.008028198964893818, + 0.007024433463811874, + 0.009233823046088219, + 0.00850401259958744, + 0.003839887445792556, + 0.004770370200276375, + 0.003087228164076805, + 0.00473312521353364, + 0.004424314945936203, + 0.0026333611458539963, + 0.0061487965285778046, + 0.004167809151113033, + 0.0035652907099574804, + 0.003945954609662294, + 0.002794117433950305, + 0.0034674324560910463, + 0.0033998836297541857, + 0.0018318479415029287, + 0.0046602594666182995, + 0.0037412410601973534, + 0.0034717009402811527, + 0.0026287222281098366, + 0.003232227638363838, + 0.004689592402428389, + 0.003974721767008305, + 0.0021018965635448694, + 0.005211965646594763, + 0.0025128310080617666 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 5 + ], + "selected_reward": -0.016064144670963287 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000034/group_2/index_tensor([6618, 6618])_group_2_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000034/group_2/index_tensor([6618, 6618])_group_2_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..708949ca8b3cc4e9839c0a2a20dd26134ee15cdb --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000034/group_2/index_tensor([6618, 6618])_group_2_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "shot shot from an extremely low angle, the camera slowly tilting upward to emphasize a sense of majestic oppression. The image depicts an ancient, dense primeval forest, drenched in the cold, clear blue light of a full moon at midnight. Within this stillness, a massive, humanoid creature, composed of the rough, deeply grooved bark of an ancient tree, occupies the center of the frame, its massive form dotted with moist moss. Countless glowing vines, like the creature's own veins, entwine and weave through the cracks of its bark, rhythmically pulsing with a mysterious, inner blue-violet light. With a dull, grinding sound, as if from deep within the earth, the creature, with an extremely slow and solemn gesture, carefully unfolds its arms, resembling those of enormous, dead branches; the movement is so deliberate and powerful that a subtle disturbance in the air is palpable. High-contrast moonlight streams down from above, interplaying with the glow of the vines within the creature, casting ever-changing, flowing, dramatic shadows across its textured body. The contraction and relaxation of every muscle – if you can call it that – are clearly outlined by the light and shadow, creating an intense and awe-inspiring cinematic lighting effect and an atmosphere that is ancient, mysterious, and full of vitality.", + "seed": 2867, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000034/group_2/index_6618_seed_2867_dp_4_rank_8_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 2, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.09641902288421988, + "vlm_logprob": -0.09641902288421988, + "vlm_yes_no_margin": 3.45, + "vlm_token_credit": -0.09641902288421988, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot from an extremely low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera slowly tilt upward during the shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the central subject a massive humanoid creature made of rough, deeply grooved tree bark?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are countless glowing vines entwined in the creature's bark, pulsing with blue-violet light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the creature slowly and deliberately unfold its arms resembling enormous dead branches?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.2819681763648987 + ], + "opposite_logprob_tokens": [ + -1.656968116760254 + ], + "logprob_reward": -0.2819681763648987, + "yes_no_margin_reward": 1.375, + "token_credit_reward": -0.2819681763648987, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.4962379038333893, + 0.056263089179992676, + 0.11256036907434464, + 0.03699140623211861, + 0.045659251511096954, + 0.05267686769366264, + 0.04186199977993965, + 0.07333578914403915, + 0.06142309308052063, + 0.022990243509411812 + ], + "credit_map": [ + 0.025033876299858093, + 0.014884055592119694, + 0.022779423743486404, + 0.012829664163291454, + 0.018844984471797943, + 0.19786104559898376, + 0.02910076081752777, + 0.17490407824516296, + 0.005258891265839338, + 0.005418010987341404, + 0.007840822450816631, + 0.010462330654263496, + 0.004638066980987787, + 0.009183663874864578, + 0.0054991962388157845, + 0.007962106727063656, + 0.01148850005120039, + 0.016994213685393333, + 0.015662454068660736, + 0.014991058968007565, + 0.014442367479205132, + 0.012342294678092003, + 0.009634199552237988, + 0.017005277797579765, + 0.004747893661260605, + 0.006783653981983662, + 0.004580117762088776, + 0.0047201961278915405, + 0.004180127754807472, + 0.0023604165762662888, + 0.004472224973142147, + 0.005146774463355541, + 0.003580729942768812, + 0.005152480211108923, + 0.006835302338004112, + 0.0035997380036860704, + 0.005899728275835514, + 0.00736679881811142, + 0.004809246398508549, + 0.008415224961936474, + 0.010250498540699482, + 0.005549474153667688, + 0.004887549206614494, + 0.003621115814894438, + 0.0048078130930662155, + 0.0167133379727602, + 0.003218867350369692, + 0.0036282124929130077, + 0.002819221932440996, + 0.003143589710816741, + 0.003759710816666484, + 0.004317957907915115, + 0.0032522762194275856, + 0.010421914048492908, + 0.006303253583610058, + 0.00784407276660204, + 0.0038262121379375458, + 0.0036324290558695793, + 0.0035658583510667086, + 0.0037417984567582607, + 0.0029675632249563932, + 0.02540157549083233, + 0.004260636400431395, + 0.025939716026186943, + 0.005289258901029825, + 0.006411385256797075, + 0.0070354752242565155, + 0.0073940204456448555, + 0.006807013414800167, + 0.007819673046469688, + 0.01061196904629469, + 0.01005430519580841, + 0.0032058688811957836, + 0.002004389651119709, + 0.002054211450740695, + 0.003609220962971449, + 0.0020009512081742287, + 0.0023581022396683693, + 0.004299824126064777, + 0.0034576761536300182 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 7 + ], + "selected_reward": -0.2819681763648987 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot from an extremely low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera slowly tilt upward during the shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the central subject a massive humanoid creature made of rough, deeply grooved tree bark?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are countless glowing vines entwined in the creature's bark, pulsing with blue-violet light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the creature slowly and deliberately unfold its arms resembling enormous dead branches?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.1604122817516327 + ], + "opposite_logprob_tokens": [ + -1.910412311553955 + ], + "logprob_reward": -0.1604122817516327, + "yes_no_margin_reward": 1.75, + "token_credit_reward": -0.1604122817516327, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07761411368846893, + 0.06513020396232605, + 0.07152996957302094, + 0.0809098482131958, + 0.08942002803087234, + 0.2053324580192566, + 0.028796391561627388, + 0.1512923389673233, + 0.20150309801101685, + 0.028471587225794792 + ], + "credit_map": [ + 0.010541778057813644, + 0.007948487997055054, + 0.010410642251372337, + 0.007183782756328583, + 0.008063727989792824, + 0.009218110702931881, + 0.013149712234735489, + 0.011097874492406845, + 0.0100203026086092, + 0.00712598068639636, + 0.009464384987950325, + 0.007588159292936325, + 0.00585660757496953, + 0.010097116231918335, + 0.005856877658516169, + 0.009120776318013668, + 0.008772010914981365, + 0.011109782382845879, + 0.008594091050326824, + 0.006076363380998373, + 0.009134937077760696, + 0.007398431189358234, + 0.007507790345698595, + 0.012936565093696117, + 0.008501355536282063, + 0.016162093728780746, + 0.010520056821405888, + 0.009840836748480797, + 0.010446506552398205, + 0.005046316888183355, + 0.007691453211009502, + 0.012701236642897129, + 0.00850955955684185, + 0.01129163708537817, + 0.01020862814038992, + 0.008588039316236973, + 0.01126924715936184, + 0.014628737233579159, + 0.007913334295153618, + 0.017010845243930817, + 0.035004645586013794, + 0.012467259541153908, + 0.029166989028453827, + 0.014174210838973522, + 0.009317249990999699, + 0.08119993656873703, + 0.008702444843947887, + 0.015299705788493156, + 0.002866319613531232, + 0.0027849352918565273, + 0.0029002337250858545, + 0.0025798273272812366, + 0.002424464328214526, + 0.0044443802908062935, + 0.0043401881121098995, + 0.006456043105572462, + 0.007846962660551071, + 0.006660712882876396, + 0.0059625557623803616, + 0.006314748432487249, + 0.0044559696689248085, + 0.054151102900505066, + 0.007190533448010683, + 0.0587097592651844, + 0.018918167799711227, + 0.02366088517010212, + 0.026148974895477295, + 0.020141806453466415, + 0.021978842094540596, + 0.027751941233873367, + 0.03472209721803665, + 0.02818039059638977, + 0.003990220837295055, + 0.00264002475887537, + 0.0036406898871064186, + 0.0037765728775411844, + 0.002945457585155964, + 0.003414492355659604, + 0.0050738160498440266, + 0.00299031101167202 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 8, + 7 + ], + "selected_reward": -0.1604122817516327 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot from an extremely low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera slowly tilt upward during the shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the central subject a massive humanoid creature made of rough, deeply grooved tree bark?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are countless glowing vines entwined in the creature's bark, pulsing with blue-violet light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the creature slowly and deliberately unfold its arms resembling enormous dead branches?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004653101786971092 + ], + "opposite_logprob_tokens": [ + -5.379652976989746 + ], + "logprob_reward": -0.004653101786971092, + "yes_no_margin_reward": 5.375, + "token_credit_reward": -0.004653101786971092, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3576774597167969, + 0.10029792040586472, + 0.19516879320144653, + 0.040431201457977295, + 0.03242993354797363, + 0.03382725268602371, + 0.017878009006381035, + 0.10679859668016434, + 0.09556256234645844, + 0.019928211346268654 + ], + "credit_map": [ + 0.02039589360356331, + 0.015245834365487099, + 0.017779547721147537, + 0.015043343417346478, + 0.016680311411619186, + 0.12256120890378952, + 0.028905285522341728, + 0.12106601893901825, + 0.011920860968530178, + 0.010800475254654884, + 0.011618963442742825, + 0.018953094258904457, + 0.011036708019673824, + 0.014510036446154118, + 0.008794889785349369, + 0.012662879191339016, + 0.02056111767888069, + 0.03993803262710571, + 0.02421066351234913, + 0.023431072011590004, + 0.02878817357122898, + 0.019928142428398132, + 0.014295277185738087, + 0.024016328155994415, + 0.0033780750818550587, + 0.008477184921503067, + 0.004536632914096117, + 0.007125784642994404, + 0.004056018311530352, + 0.0034488379023969173, + 0.0033032416831701994, + 0.006105426698923111, + 0.003340162103995681, + 0.003981277346611023, + 0.004486674442887306, + 0.0030433719512075186, + 0.0031411468517035246, + 0.005732709541916847, + 0.002774058375507593, + 0.00593053363263607, + 0.005899303127080202, + 0.0031638091895729303, + 0.0037952361162751913, + 0.0022793705575168133, + 0.002468094928190112, + 0.01039297878742218, + 0.0021627035457640886, + 0.0036657548043876886, + 0.0016678697429597378, + 0.0015163224888965487, + 0.0019347671186551452, + 0.001391658210195601, + 0.0012158354511484504, + 0.004318169318139553, + 0.0025864203926175833, + 0.003246965818107128, + 0.005586366169154644, + 0.004586154595017433, + 0.003490562317892909, + 0.004492847714573145, + 0.0032954213675111532, + 0.03919537737965584, + 0.005285077728331089, + 0.04086678475141525, + 0.009971702471375465, + 0.00903000496327877, + 0.013741566799581051, + 0.011191267520189285, + 0.009903148747980595, + 0.0131600983440876, + 0.015866298228502274, + 0.012698465958237648, + 0.00402757478877902, + 0.002404939616099, + 0.0023305939976125956, + 0.002348121954128146, + 0.0017551424680277705, + 0.0023246535565704107, + 0.002535097999498248, + 0.0022020870819687843 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 7 + ], + "selected_reward": -0.004653101786971092 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot from an extremely low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera slowly tilt upward during the shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the central subject a massive humanoid creature made of rough, deeply grooved tree bark?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are countless glowing vines entwined in the creature's bark, pulsing with blue-violet light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the creature slowly and deliberately unfold its arms resembling enormous dead branches?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0052565219812095165 + ], + "opposite_logprob_tokens": [ + -5.255256652832031 + ], + "logprob_reward": -0.0052565219812095165, + "yes_no_margin_reward": 5.25, + "token_credit_reward": -0.0052565219812095165, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.803621768951416, + 0.0539952777326107, + 0.02873547561466694, + 0.025689933449029922, + 0.009361924603581429, + 0.015433240681886673, + 0.010655802674591541, + 0.011948098428547382, + 0.030889436602592468, + 0.009668981656432152 + ], + "credit_map": [ + 0.039695948362350464, + 0.02229209430515766, + 0.031178630888462067, + 0.022539323195815086, + 0.031872279942035675, + 0.2880050241947174, + 0.07809730619192123, + 0.28994128108024597, + 0.0070245107635855675, + 0.007049387786537409, + 0.006288641132414341, + 0.00603296235203743, + 0.004668944049626589, + 0.010946046561002731, + 0.00473166536539793, + 0.0072531248442828655, + 0.003290384542196989, + 0.005454171448945999, + 0.003575698472559452, + 0.0035140991676598787, + 0.003389839781448245, + 0.002574820537120104, + 0.002970683854073286, + 0.003965780138969421, + 0.0028057866729795933, + 0.003982612397521734, + 0.0040619042702019215, + 0.003339997725561261, + 0.003155251033604145, + 0.0017770994454622269, + 0.0025004628114402294, + 0.004066820256412029, + 0.0010153298499062657, + 0.0010837726294994354, + 0.0015478020068258047, + 0.0009839796693995595, + 0.0008799344068393111, + 0.0009873241651803255, + 0.0008491058833897114, + 0.002014676807448268, + 0.0019689376931637526, + 0.0011193108512088656, + 0.001964171417057514, + 0.0010150758316740394, + 0.001291102496907115, + 0.005430466029793024, + 0.001137531828135252, + 0.0015066465130075812, + 0.001323927310295403, + 0.0009376390953548253, + 0.001132997334934771, + 0.0011662589386105537, + 0.0008689602836966515, + 0.0015580697217956185, + 0.001926921191625297, + 0.0017410292057320476, + 0.0010735249379649758, + 0.0010242430726066232, + 0.0010980474762618542, + 0.0018909773789346218, + 0.0009681850206106901, + 0.002761938376352191, + 0.0016767128836363554, + 0.0014544703299179673, + 0.0031243590638041496, + 0.002836965722963214, + 0.004282402340322733, + 0.004069807939231396, + 0.0032535919453948736, + 0.004278882406651974, + 0.004961885046213865, + 0.004081543534994125, + 0.001296832924708724, + 0.0017107785679399967, + 0.0010794874979183078, + 0.0014319857582449913, + 0.0008972190553322434, + 0.0008830349543131888, + 0.0015842259163036942, + 0.0007854176801629364 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 8 + ], + "selected_reward": -0.0052565219812095165 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot from an extremely low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera slowly tilt upward during the shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the central subject a massive humanoid creature made of rough, deeply grooved tree bark?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are countless glowing vines entwined in the creature's bark, pulsing with blue-violet light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the creature slowly and deliberately unfold its arms resembling enormous dead branches?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.029805032536387444 + ], + "opposite_logprob_tokens": [ + -3.5298049449920654 + ], + "logprob_reward": -0.029805032536387444, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.029805032536387444, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2723902761936188, + 0.10834657400846481, + 0.09920656681060791, + 0.08538040518760681, + 0.08679160475730896, + 0.11764023452997208, + 0.05561451613903046, + 0.08325944095849991, + 0.07825960963964462, + 0.013110780157148838 + ], + "credit_map": [ + 0.012632518075406551, + 0.006136714946478605, + 0.007014922797679901, + 0.0068395850248634815, + 0.013501957058906555, + 0.11186273396015167, + 0.02675015479326248, + 0.08765167742967606, + 0.009702219627797604, + 0.012766108848154545, + 0.011462979018688202, + 0.022355062887072563, + 0.013250916264951229, + 0.011706497520208359, + 0.011891137808561325, + 0.015211650170385838, + 0.011400720104575157, + 0.019515283405780792, + 0.011115560308098793, + 0.008539595641195774, + 0.01614038087427616, + 0.011212274432182312, + 0.008560756221413612, + 0.01272199209779501, + 0.008267069235444069, + 0.020964423194527626, + 0.008003910072147846, + 0.010830939747393131, + 0.010368253104388714, + 0.006546997930854559, + 0.006560834124684334, + 0.013837977312505245, + 0.011704422533512115, + 0.011845695786178112, + 0.014988432638347149, + 0.008037338964641094, + 0.010696018114686012, + 0.008855825290083885, + 0.009984604083001614, + 0.010679269209504128, + 0.01530691422522068, + 0.00671443110331893, + 0.019867904484272003, + 0.008784567937254906, + 0.004827726166695356, + 0.04755987226963043, + 0.005025716498494148, + 0.009553097188472748, + 0.004248148761689663, + 0.0032268192153424025, + 0.005864517297595739, + 0.005974749103188515, + 0.0032218473497778177, + 0.013588634319603443, + 0.009587089531123638, + 0.009902707301080227, + 0.004401758778840303, + 0.003254974028095603, + 0.003371903207153082, + 0.005452940706163645, + 0.003169327275827527, + 0.029301630333065987, + 0.006037269718945026, + 0.028269637376070023, + 0.007171325385570526, + 0.0071073477156460285, + 0.013043636456131935, + 0.009331882931292057, + 0.009666431695222855, + 0.010453995317220688, + 0.011672921478748322, + 0.0098120691254735, + 0.0019607802387326956, + 0.0014470480382442474, + 0.0011994920205324888, + 0.0017188148340210319, + 0.0012420834973454475, + 0.0020901570096611977, + 0.0016210920875892043, + 0.0018313118489459157 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 5, + 1 + ], + "selected_reward": -0.029805032536387444 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "shot shot from an extremely low angle, the camera slowly tilting upward to emphasize a sense of majestic oppression. The image depicts an ancient, dense primeval forest, drenched in the cold, clear blue light of a full moon at midnight. Within this stillness, a massive, humanoid creature, composed of the rough, deeply grooved bark of an ancient tree, occupies the center of the frame, its massive form dotted with moist moss. Countless glowing vines, like the creature's own veins, entwine and weave through the cracks of its bark, rhythmically pulsing with a mysterious, inner blue-violet light. With a dull, grinding sound, as if from deep within the earth, the creature, with an extremely slow and solemn gesture, carefully unfolds its arms, resembling those of enormous, dead branches; the movement is so deliberate and powerful that a subtle disturbance in the air is palpable. High-contrast moonlight streams down from above, interplaying with the glow of the vines within the creature, casting ever-changing, flowing, dramatic shadows across its textured body. The contraction and relaxation of every muscle – if you can call it that – are clearly outlined by the light and shadow, creating an intense and awe-inspiring cinematic lighting effect and an atmosphere that is ancient, mysterious, and full of vitality.", + "seed": 2868, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000034/group_2/index_6618_seed_2868_dp_4_rank_8_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 2, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.04084634720347822, + "vlm_logprob": -0.04084634720347822, + "vlm_yes_no_margin": 4.0, + "vlm_token_credit": -0.04084634720347822, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot from an extremely low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera slowly tilt upward during the shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the central subject a massive humanoid creature made of rough, deeply grooved tree bark?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are countless glowing vines entwined in the creature's bark, pulsing with blue-violet light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the creature slowly and deliberately unfold its arms resembling enormous dead branches?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.1080884039402008 + ], + "opposite_logprob_tokens": [ + -2.608088493347168 + ], + "logprob_reward": -0.1080884039402008, + "yes_no_margin_reward": 2.5, + "token_credit_reward": -0.1080884039402008, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07598526775836945, + 0.028183991089463234, + 0.016799891367554665, + 0.036402489989995956, + 0.12895941734313965, + 0.030317196622490883, + 0.6304976940155029, + 0.01291071716696024, + 0.008065287955105305, + 0.031878162175416946 + ], + "credit_map": [ + 0.01235886663198471, + 0.0051694088615477085, + 0.0185442753136158, + 0.004361157771199942, + 0.004532465245574713, + 0.005450400989502668, + 0.009444798342883587, + 0.016123883426189423, + 0.002540348330512643, + 0.003968758042901754, + 0.0036577063146978617, + 0.002286554081365466, + 0.003889539046213031, + 0.003675834508612752, + 0.0035564033314585686, + 0.0046088434755802155, + 0.001262611709535122, + 0.0018780792597681284, + 0.00499764597043395, + 0.0017281343461945653, + 0.0013547156704589725, + 0.0016064784722402692, + 0.0011103719007223845, + 0.002861850429326296, + 0.0029846399556845427, + 0.0030790953896939754, + 0.0071703107096254826, + 0.010370067320764065, + 0.002302375854924321, + 0.003763464977964759, + 0.002895868383347988, + 0.003836663207039237, + 0.02288389578461647, + 0.011417883448302746, + 0.011244435794651508, + 0.006428527645766735, + 0.01110557559877634, + 0.04737391695380211, + 0.007107928860932589, + 0.0113972257822752, + 0.006755424663424492, + 0.003065967233851552, + 0.002394401002675295, + 0.0025230853352695704, + 0.004563498310744762, + 0.006862634792923927, + 0.001968515571206808, + 0.0021836645901203156, + 0.18924973905086517, + 0.06389555335044861, + 0.021837826818227768, + 0.027312854304909706, + 0.028539132326841354, + 0.24866455793380737, + 0.02376713789999485, + 0.027230802923440933, + 0.002469402737915516, + 0.0015822439454495907, + 0.0008381298976019025, + 0.0013125052209943533, + 0.0019617064390331507, + 0.00231354171410203, + 0.0013297434197738767, + 0.0011034425115212798, + 0.0009499068255536258, + 0.0011002701940014958, + 0.0008200298761948943, + 0.0008436903008259833, + 0.0013901174534112215, + 0.0009859675774350762, + 0.0013383520999923348, + 0.0006369527545757592, + 0.004967677406966686, + 0.004224464762955904, + 0.0032090216409415007, + 0.0043281083926558495, + 0.005603792145848274, + 0.003930306062102318, + 0.003260193392634392, + 0.002354593947529793 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 4, + 0 + ], + "selected_reward": -0.1080884039402008 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot from an extremely low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera slowly tilt upward during the shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the central subject a massive humanoid creature made of rough, deeply grooved tree bark?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are countless glowing vines entwined in the creature's bark, pulsing with blue-violet light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the creature slowly and deliberately unfold its arms resembling enormous dead branches?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.06207640469074249 + ], + "opposite_logprob_tokens": [ + -2.8120763301849365 + ], + "logprob_reward": -0.06207640469074249, + "yes_no_margin_reward": 2.75, + "token_credit_reward": -0.06207640469074249, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10050748288631439, + 0.02334301359951496, + 0.0503971204161644, + 0.03937748447060585, + 0.49108511209487915, + 0.04410524666309357, + 0.22533346712589264, + 0.013001113198697567, + 0.006602099165320396, + 0.006247872952371836 + ], + "credit_map": [ + 0.01712808385491371, + 0.006542641203850508, + 0.02548050880432129, + 0.005186518654227257, + 0.006127777975052595, + 0.008162986487150192, + 0.011260340921580791, + 0.02061861753463745, + 0.001954818842932582, + 0.0031369582284241915, + 0.003584655001759529, + 0.002104869345203042, + 0.0031967663671821356, + 0.0029025287367403507, + 0.002640599152073264, + 0.003821815364062786, + 0.004331542644649744, + 0.006761881988495588, + 0.012129852548241615, + 0.008259242400527, + 0.0037807009648531675, + 0.003568722400814295, + 0.004013939760625362, + 0.007551233284175396, + 0.003638720838353038, + 0.006910172291100025, + 0.0061660036444664, + 0.007559323217719793, + 0.0040365830063819885, + 0.003436002414673567, + 0.003619743511080742, + 0.004010932985693216, + 0.09105744957923889, + 0.038764163851737976, + 0.04860283061861992, + 0.021838391199707985, + 0.03450740501284599, + 0.18866711854934692, + 0.024987494572997093, + 0.04266020283102989, + 0.010927896946668625, + 0.003963265102356672, + 0.00268375501036644, + 0.003734944388270378, + 0.004487001337110996, + 0.01112516038119793, + 0.0033721744548529387, + 0.0038110450841486454, + 0.06433837115764618, + 0.02206549048423767, + 0.00919909868389368, + 0.011386760510504246, + 0.01306022610515356, + 0.08412934839725494, + 0.011139865964651108, + 0.010014289990067482, + 0.002401473466306925, + 0.001686485600657761, + 0.0012008462799713016, + 0.001185513800010085, + 0.0017769410042092204, + 0.0015350371832028031, + 0.0020922780968248844, + 0.0011225377675145864, + 0.0007275817915797234, + 0.0007980432128533721, + 0.0007676459499634802, + 0.0007135506602935493, + 0.0013555745827034116, + 0.0007039032061584294, + 0.0008559600682929158, + 0.0006798387039452791, + 0.0005212635733187199, + 0.0007906037499196827, + 0.0006967568770051003, + 0.0005325782112777233, + 0.001673696213401854, + 0.0008286508964374661, + 0.0007179431850090623, + 0.0004863793437834829 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 6, + 0 + ], + "selected_reward": -0.06207640469074249 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot from an extremely low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera slowly tilt upward during the shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the central subject a massive humanoid creature made of rough, deeply grooved tree bark?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are countless glowing vines entwined in the creature's bark, pulsing with blue-violet light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the creature slowly and deliberately unfold its arms resembling enormous dead branches?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.003622637828812003 + ], + "opposite_logprob_tokens": [ + -5.628622531890869 + ], + "logprob_reward": -0.003622637828812003, + "yes_no_margin_reward": 5.625, + "token_credit_reward": -0.003622637828812003, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08811643719673157, + 0.10613161325454712, + 0.03327890858054161, + 0.025186687707901, + 0.26160484552383423, + 0.02420153096318245, + 0.32933348417282104, + 0.026462271809577942, + 0.015836596488952637, + 0.08984754979610443 + ], + "credit_map": [ + 0.01467716135084629, + 0.007299154996871948, + 0.02129012905061245, + 0.006573502440005541, + 0.008548630401492119, + 0.00786155927926302, + 0.007909750565886497, + 0.013956550508737564, + 0.012749182060360909, + 0.02084907703101635, + 0.016457565128803253, + 0.008675415068864822, + 0.012064268812537193, + 0.009974873624742031, + 0.012098491191864014, + 0.013262742199003696, + 0.00344324903562665, + 0.003265335224568844, + 0.008266108110547066, + 0.0037714545615017414, + 0.0022456536535173655, + 0.00340589159168303, + 0.002412928268313408, + 0.006468289531767368, + 0.0020186773035675287, + 0.002915174700319767, + 0.005765245761722326, + 0.006947306916117668, + 0.001890007988549769, + 0.0020064713899046183, + 0.0023076781071722507, + 0.00133612472563982, + 0.04689083248376846, + 0.025112425908446312, + 0.022454407066106796, + 0.011984756216406822, + 0.02680640108883381, + 0.08757499605417252, + 0.016529152169823647, + 0.024251876398921013, + 0.0028267628513276577, + 0.0033111674711108208, + 0.001931642065756023, + 0.002456486225128174, + 0.004268435295671225, + 0.003966442309319973, + 0.0029702403116971254, + 0.0024703529197722673, + 0.09553789347410202, + 0.032467156648635864, + 0.012699327431619167, + 0.015436078421771526, + 0.018291356042027473, + 0.12478616833686829, + 0.015170144848525524, + 0.014945371076464653, + 0.004994208458811045, + 0.002759048016741872, + 0.0022764913737773895, + 0.0025804233737289906, + 0.0033026598393917084, + 0.0032608970068395138, + 0.005236194934695959, + 0.002052348107099533, + 0.0017866425914689898, + 0.002543803071603179, + 0.0014724666252732277, + 0.0015140357427299023, + 0.003272254718467593, + 0.002026173286139965, + 0.0018630602862685919, + 0.0013581610983237624, + 0.012467893771827221, + 0.012355425395071507, + 0.009357765316963196, + 0.010692903771996498, + 0.020034246146678925, + 0.010336989536881447, + 0.008676988072693348, + 0.005925335921347141 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 4, + 1 + ], + "selected_reward": -0.003622637828812003 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot from an extremely low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera slowly tilt upward during the shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the central subject a massive humanoid creature made of rough, deeply grooved tree bark?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are countless glowing vines entwined in the creature's bark, pulsing with blue-violet light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the creature slowly and deliberately unfold its arms resembling enormous dead branches?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004097281955182552 + ], + "opposite_logprob_tokens": [ + -5.5040974617004395 + ], + "logprob_reward": -0.004097281955182552, + "yes_no_margin_reward": 5.5, + "token_credit_reward": -0.004097281955182552, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12288741767406464, + 0.053511422127485275, + 0.17944952845573425, + 0.06644962728023529, + 0.3868766129016876, + 0.05095186084508896, + 0.06241428479552269, + 0.014853744767606258, + 0.012693244963884354, + 0.049912307411432266 + ], + "credit_map": [ + 0.011439396999776363, + 0.018385129049420357, + 0.01957898586988449, + 0.010561881586909294, + 0.017405780032277107, + 0.011177358217537403, + 0.015390126034617424, + 0.018948763608932495, + 0.006160927005112171, + 0.009146053344011307, + 0.007553961127996445, + 0.003766958136111498, + 0.007365293335169554, + 0.005411171354353428, + 0.006287648342549801, + 0.007819408550858498, + 0.01727578230202198, + 0.018174922093749046, + 0.05866825208067894, + 0.02185956947505474, + 0.009528134018182755, + 0.012474779039621353, + 0.01110928412526846, + 0.03035880997776985, + 0.0056369914673268795, + 0.007485620677471161, + 0.014444843865931034, + 0.021048517897725105, + 0.004891740623861551, + 0.005011525936424732, + 0.004537965636700392, + 0.003392426995560527, + 0.07344228029251099, + 0.03191281482577324, + 0.0360121987760067, + 0.017378292977809906, + 0.028150631114840508, + 0.14762483537197113, + 0.02017582394182682, + 0.03217974305152893, + 0.01174970343708992, + 0.005026858765631914, + 0.002532523823902011, + 0.004241395276039839, + 0.006989141460508108, + 0.012120818719267845, + 0.004326218273490667, + 0.0039652008563280106, + 0.01319640688598156, + 0.006504787597805262, + 0.004143525380641222, + 0.005608478561043739, + 0.005925127305090427, + 0.016725484281778336, + 0.005688617005944252, + 0.004621855914592743, + 0.003376161213964224, + 0.0018621416529640555, + 0.0013387186918407679, + 0.0013980251969769597, + 0.0023279672022908926, + 0.0016305821482092142, + 0.0019902943167835474, + 0.0009298549266532063, + 0.002006793860346079, + 0.0018050438957288861, + 0.0012172670103609562, + 0.001239015138708055, + 0.001904864446260035, + 0.002232055878266692, + 0.0012028864584863186, + 0.0010853172279894352, + 0.00886223092675209, + 0.005585741717368364, + 0.006594800390303135, + 0.004170202650129795, + 0.00829975213855505, + 0.006353472359478474, + 0.005209785420447588, + 0.004836322274059057 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 2, + 0 + ], + "selected_reward": -0.004097281955182552 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot from an extremely low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera slowly tilt upward during the shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the central subject a massive humanoid creature made of rough, deeply grooved tree bark?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are countless glowing vines entwined in the creature's bark, pulsing with blue-violet light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the creature slowly and deliberately unfold its arms resembling enormous dead branches?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.026347007602453232 + ], + "opposite_logprob_tokens": [ + -3.6513469219207764 + ], + "logprob_reward": -0.026347007602453232, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.026347007602453232, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.05406752973794937, + 0.024359775707125664, + 0.03020963817834854, + 0.03773387148976326, + 0.4479970335960388, + 0.05846581608057022, + 0.14292538166046143, + 0.03142509236931801, + 0.04229210317134857, + 0.13052375614643097 + ], + "credit_map": [ + 0.008235500194132328, + 0.004560078959912062, + 0.012808416970074177, + 0.003030557883903384, + 0.004130194429308176, + 0.004038888495415449, + 0.007215943653136492, + 0.010047951713204384, + 0.0027539250440895557, + 0.004576414357870817, + 0.003648924408480525, + 0.0019029889954254031, + 0.003135194769129157, + 0.0025302914436906576, + 0.003116464940831065, + 0.002695573028177023, + 0.0025082232896238565, + 0.0038988261949270964, + 0.007017004303634167, + 0.004688999615609646, + 0.002487102523446083, + 0.002564474940299988, + 0.002588021568953991, + 0.0044569880701601505, + 0.0028136433102190495, + 0.005187141243368387, + 0.008227921091020107, + 0.010311528109014034, + 0.0029425411485135555, + 0.003153934609144926, + 0.0029887964483350515, + 0.0021083711180835962, + 0.0839303582906723, + 0.03762586787343025, + 0.04199415445327759, + 0.020747918635606766, + 0.03228047117590904, + 0.1700582653284073, + 0.023358864709734917, + 0.038001175969839096, + 0.006972532253712416, + 0.006659801118075848, + 0.005517762620002031, + 0.006710526999086142, + 0.009010820649564266, + 0.00733482651412487, + 0.008134509436786175, + 0.008125041611492634, + 0.04210059717297554, + 0.015301249921321869, + 0.005688893608748913, + 0.005265653133392334, + 0.008253021165728569, + 0.05498082563281059, + 0.0054499381221830845, + 0.005885221995413303, + 0.004413780756294727, + 0.003861816134303808, + 0.0020619195420295, + 0.002864265814423561, + 0.004863486625254154, + 0.006482084281742573, + 0.004066211637109518, + 0.002811530837789178, + 0.004168978426605463, + 0.005530652590095997, + 0.004495270550251007, + 0.004931517411023378, + 0.008404935710132122, + 0.006019992288202047, + 0.00522778881713748, + 0.003512970870360732, + 0.014369388110935688, + 0.016592638567090034, + 0.017491957172751427, + 0.01920677348971367, + 0.024587130174040794, + 0.013286326080560684, + 0.014927650801837444, + 0.010061908513307571 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 6, + 9 + ], + "selected_reward": -0.026347007602453232 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "shot shot from an extremely low angle, the camera slowly tilting upward to emphasize a sense of majestic oppression. The image depicts an ancient, dense primeval forest, drenched in the cold, clear blue light of a full moon at midnight. Within this stillness, a massive, humanoid creature, composed of the rough, deeply grooved bark of an ancient tree, occupies the center of the frame, its massive form dotted with moist moss. Countless glowing vines, like the creature's own veins, entwine and weave through the cracks of its bark, rhythmically pulsing with a mysterious, inner blue-violet light. With a dull, grinding sound, as if from deep within the earth, the creature, with an extremely slow and solemn gesture, carefully unfolds its arms, resembling those of enormous, dead branches; the movement is so deliberate and powerful that a subtle disturbance in the air is palpable. High-contrast moonlight streams down from above, interplaying with the glow of the vines within the creature, casting ever-changing, flowing, dramatic shadows across its textured body. The contraction and relaxation of every muscle – if you can call it that – are clearly outlined by the light and shadow, creating an intense and awe-inspiring cinematic lighting effect and an atmosphere that is ancient, mysterious, and full of vitality.", + "seed": 2869, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000034/group_2/index_6618_seed_2869_dp_5_rank_10_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 2, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.05302314292639494, + "vlm_logprob": -0.05302314292639494, + "vlm_yes_no_margin": 3.6, + "vlm_token_credit": -0.05302314292639494, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot from an extremely low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera slowly tilt upward during the shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the central subject a massive humanoid creature made of rough, deeply grooved tree bark?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are countless glowing vines entwined in the creature's bark, pulsing with blue-violet light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the creature slowly and deliberately unfold its arms resembling enormous dead branches?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.08797551691532135 + ], + "opposite_logprob_tokens": [ + -2.96297550201416 + ], + "logprob_reward": -0.08797551691532135, + "yes_no_margin_reward": 2.875, + "token_credit_reward": -0.08797551691532135, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1287907361984253, + 0.033739153295755386, + 0.11327259242534637, + 0.05157849192619324, + 0.3183402419090271, + 0.04104633629322052, + 0.16417592763900757, + 0.0662037581205368, + 0.0611010417342186, + 0.021751709282398224 + ], + "credit_map": [ + 0.014239056967198849, + 0.016195494681596756, + 0.018966728821396828, + 0.015558600425720215, + 0.016127798706293106, + 0.012095431797206402, + 0.01813160628080368, + 0.017476029694080353, + 0.0031667761504650116, + 0.003445846028625965, + 0.004427443258464336, + 0.005489324685186148, + 0.004166437778621912, + 0.0026426143012940884, + 0.004613154102116823, + 0.0057875593192875385, + 0.010573071427643299, + 0.014247823506593704, + 0.01350485347211361, + 0.012856190092861652, + 0.017955612391233444, + 0.014347097836434841, + 0.009436355903744698, + 0.0203515887260437, + 0.005429989192634821, + 0.005067429970949888, + 0.012223535217344761, + 0.004526504315435886, + 0.005024061072617769, + 0.006107889115810394, + 0.0038978904485702515, + 0.009301193989813328, + 0.024345535784959793, + 0.08950693160295486, + 0.01767166517674923, + 0.01468056533485651, + 0.04922240599989891, + 0.09517637640237808, + 0.00988542940467596, + 0.017851315438747406, + 0.004412689711898565, + 0.004330500960350037, + 0.005250368267297745, + 0.005508062429726124, + 0.0054610478691756725, + 0.005103162955492735, + 0.0066680326126515865, + 0.004312470555305481, + 0.020262669771909714, + 0.03052780032157898, + 0.018430761992931366, + 0.010585498996078968, + 0.010239722207188606, + 0.01385642122477293, + 0.009790273383259773, + 0.05048278346657753, + 0.008693946525454521, + 0.008601232431828976, + 0.008157494477927685, + 0.009350919164717197, + 0.0058525181375443935, + 0.005896925926208496, + 0.01282298844307661, + 0.006827730219811201, + 0.006574586499482393, + 0.011141139082610607, + 0.006863749586045742, + 0.008058276027441025, + 0.005596281494945288, + 0.007478633895516396, + 0.007539199665188789, + 0.007849176414310932, + 0.0020060986280441284, + 0.0029938186053186655, + 0.0022122925147414207, + 0.003504561260342598, + 0.0025780443102121353, + 0.0029011843726038933, + 0.003027537604793906, + 0.002528171753510833 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 6, + 0 + ], + "selected_reward": -0.08797551691532135 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot from an extremely low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera slowly tilt upward during the shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the central subject a massive humanoid creature made of rough, deeply grooved tree bark?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are countless glowing vines entwined in the creature's bark, pulsing with blue-violet light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the creature slowly and deliberately unfold its arms resembling enormous dead branches?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.033728644251823425 + ], + "opposite_logprob_tokens": [ + -3.40872859954834 + ], + "logprob_reward": -0.033728644251823425, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.033728644251823425, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09047094732522964, + 0.026988478377461433, + 0.07019965350627899, + 0.09123367816209793, + 0.43209293484687805, + 0.04266763851046562, + 0.1487828642129898, + 0.04898038133978844, + 0.026080437004566193, + 0.022502945736050606 + ], + "credit_map": [ + 0.010534132830798626, + 0.01009231898933649, + 0.01186649315059185, + 0.010500337928533554, + 0.012746037915349007, + 0.006784915458410978, + 0.015148426406085491, + 0.012798286974430084, + 0.003995339386165142, + 0.002624690532684326, + 0.0030010854825377464, + 0.0029800483025610447, + 0.002962609985843301, + 0.002354142488911748, + 0.0024901777505874634, + 0.0065803867764770985, + 0.007246377877891064, + 0.008484391495585442, + 0.009769034571945667, + 0.009498009458184242, + 0.009970182552933693, + 0.007906400598585606, + 0.006234611384570599, + 0.011090646497905254, + 0.00822531245648861, + 0.008515631780028343, + 0.02095002308487892, + 0.008080178871750832, + 0.009591510519385338, + 0.01060455571860075, + 0.007027163170278072, + 0.01823931187391281, + 0.031086662784218788, + 0.12289584428071976, + 0.022457515820860863, + 0.018981225788593292, + 0.06692244112491608, + 0.1321662962436676, + 0.012632626108825207, + 0.024950377643108368, + 0.004838029853999615, + 0.004668933805078268, + 0.0052748373709619045, + 0.005865620449185371, + 0.007468582596629858, + 0.0051947832107543945, + 0.005308175925165415, + 0.004048679023981094, + 0.02555842511355877, + 0.027411464601755142, + 0.01545739360153675, + 0.012309271842241287, + 0.011302548460662365, + 0.009952405467629433, + 0.01158533152192831, + 0.035206034779548645, + 0.005687833763659, + 0.004607882350683212, + 0.004128551576286554, + 0.0047250366769731045, + 0.0073309652507305145, + 0.0064316499046981335, + 0.010041861794888973, + 0.006026603747159243, + 0.00262855994515121, + 0.0023344329092651606, + 0.003381734946742654, + 0.0031043956987559795, + 0.004177874885499477, + 0.0041104271076619625, + 0.0035097599029541016, + 0.0028332530055195093, + 0.0022047003731131554, + 0.002637631958350539, + 0.002576271304860711, + 0.004752448759973049, + 0.0025036639999598265, + 0.0023640505969524384, + 0.0031811296939849854, + 0.002283051609992981 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 6, + 3 + ], + "selected_reward": -0.033728644251823425 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot from an extremely low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera slowly tilt upward during the shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the central subject a massive humanoid creature made of rough, deeply grooved tree bark?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are countless glowing vines entwined in the creature's bark, pulsing with blue-violet light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the creature slowly and deliberately unfold its arms resembling enormous dead branches?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005266008898615837 + ], + "opposite_logprob_tokens": [ + -5.255266189575195 + ], + "logprob_reward": -0.005266008898615837, + "yes_no_margin_reward": 5.25, + "token_credit_reward": -0.005266008898615837, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.24973155558109283, + 0.05333133786916733, + 0.058354105800390244, + 0.059669844806194305, + 0.12828482687473297, + 0.051299210637807846, + 0.14217205345630646, + 0.06871048361063004, + 0.06820488721132278, + 0.1202416941523552 + ], + "credit_map": [ + 0.030687088146805763, + 0.04080173373222351, + 0.02089284174144268, + 0.02552993968129158, + 0.04407985135912895, + 0.020559854805469513, + 0.03976478427648544, + 0.027415461838245392, + 0.006975021678954363, + 0.007127487566322088, + 0.007546296808868647, + 0.007267543114721775, + 0.005311283748596907, + 0.0035519693046808243, + 0.008058641105890274, + 0.0074930922128260136, + 0.010395091027021408, + 0.008497328497469425, + 0.006232138257473707, + 0.006246451288461685, + 0.00706658186390996, + 0.006602882873266935, + 0.00438187038525939, + 0.008931761607527733, + 0.003909112419933081, + 0.00694781681522727, + 0.007324673235416412, + 0.004328638780862093, + 0.011477827094495296, + 0.0073099504224956036, + 0.004268724005669355, + 0.014103099703788757, + 0.013979187235236168, + 0.030838370323181152, + 0.00966282095760107, + 0.008384780026972294, + 0.017471810802817345, + 0.03167376294732094, + 0.007656553760170937, + 0.008617543615400791, + 0.010014667175710201, + 0.004765412304550409, + 0.004644449334591627, + 0.008950907737016678, + 0.004401359241455793, + 0.004820826929062605, + 0.006453265901654959, + 0.007248320151120424, + 0.019698182120919228, + 0.029279440641403198, + 0.01491832546889782, + 0.009196292608976364, + 0.010067665949463844, + 0.010786335915327072, + 0.008570975624024868, + 0.039654843509197235, + 0.007964540272951126, + 0.009102849289774895, + 0.007398461457341909, + 0.009350546635687351, + 0.007200988940894604, + 0.00863693468272686, + 0.010467625223100185, + 0.008588544093072414, + 0.008748769760131836, + 0.008558882400393486, + 0.006125392857939005, + 0.009550105780363083, + 0.009745810180902481, + 0.00983611773699522, + 0.006803519558161497, + 0.008836288936436176, + 0.010100314393639565, + 0.015213259495794773, + 0.013318266719579697, + 0.030042126774787903, + 0.010238129645586014, + 0.013257917016744614, + 0.01479244139045477, + 0.01327924057841301 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 6, + 4 + ], + "selected_reward": -0.005266008898615837 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot from an extremely low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera slowly tilt upward during the shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the central subject a massive humanoid creature made of rough, deeply grooved tree bark?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are countless glowing vines entwined in the creature's bark, pulsing with blue-violet light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the creature slowly and deliberately unfold its arms resembling enormous dead branches?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.011081107892096043 + ], + "opposite_logprob_tokens": [ + -4.511081218719482 + ], + "logprob_reward": -0.011081107892096043, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.011081107892096043, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14383484423160553, + 0.12236268818378448, + 0.057897597551345825, + 0.06236067786812782, + 0.19620145857334137, + 0.027282923460006714, + 0.07359664887189865, + 0.058273542672395706, + 0.05671416595578194, + 0.20147539675235748 + ], + "credit_map": [ + 0.015668721869587898, + 0.01936156675219536, + 0.014687563292682171, + 0.01794893853366375, + 0.020810019224882126, + 0.019293617457151413, + 0.017711743712425232, + 0.018352679908275604, + 0.01797332428395748, + 0.014201025478541851, + 0.013692473992705345, + 0.01733110286295414, + 0.012970511801540852, + 0.007831895723938942, + 0.015047590248286724, + 0.023314762860536575, + 0.006394798867404461, + 0.008484999649226665, + 0.00895675178617239, + 0.008611170575022697, + 0.00952162966132164, + 0.0052296980284154415, + 0.004661669954657555, + 0.006036880426108837, + 0.004446296952664852, + 0.004579807631671429, + 0.013327470049262047, + 0.005197819788008928, + 0.006272335536777973, + 0.0076432907953858376, + 0.005403707269579172, + 0.015489951707422733, + 0.01645970158278942, + 0.052597418427467346, + 0.010379286482930183, + 0.009219205006957054, + 0.028455354273319244, + 0.05961223319172859, + 0.007533536292612553, + 0.011944721452891827, + 0.004973851144313812, + 0.002583532128483057, + 0.0035395368468016386, + 0.00388004956766963, + 0.0028527122922241688, + 0.0028634259942919016, + 0.003284069010987878, + 0.0033057478722184896, + 0.011326543055474758, + 0.013976182788610458, + 0.007611522916704416, + 0.005637672729790211, + 0.005212696734815836, + 0.005120720714330673, + 0.005048476159572601, + 0.019662832841277122, + 0.006988706067204475, + 0.008159013465046883, + 0.007846755906939507, + 0.007688099052757025, + 0.00429803179576993, + 0.008200488984584808, + 0.007741882000118494, + 0.007350564934313297, + 0.005234984681010246, + 0.009450958110392094, + 0.006304111797362566, + 0.007852723821997643, + 0.004262151662260294, + 0.008720221929252148, + 0.008245047181844711, + 0.006643969099968672, + 0.015373959206044674, + 0.02055301144719124, + 0.0200913418084383, + 0.05277835205197334, + 0.021995753049850464, + 0.01945899799466133, + 0.03137405589222908, + 0.019849924370646477 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 4, + 0 + ], + "selected_reward": -0.011081107892096043 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot from an extremely low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera slowly tilt upward during the shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the central subject a massive humanoid creature made of rough, deeply grooved tree bark?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are countless glowing vines entwined in the creature's bark, pulsing with blue-violet light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the creature slowly and deliberately unfold its arms resembling enormous dead branches?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.12706443667411804 + ], + "opposite_logprob_tokens": [ + -2.1270644664764404 + ], + "logprob_reward": -0.12706443667411804, + "yes_no_margin_reward": 2.0, + "token_credit_reward": -0.12706443667411804, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.047398753464221954, + 0.020520033314824104, + 0.02680150605738163, + 0.048000507056713104, + 0.7455116510391235, + 0.021826837211847305, + 0.032299064099788666, + 0.016984015703201294, + 0.0215617585927248, + 0.019095905125141144 + ], + "credit_map": [ + 0.004959952086210251, + 0.005908855702728033, + 0.004548199940472841, + 0.006675420328974724, + 0.005027805455029011, + 0.009070304222404957, + 0.006071910262107849, + 0.0051363068632781506, + 0.0019559618085622787, + 0.002803226700052619, + 0.0024721198715269566, + 0.002647469053044915, + 0.003169938223436475, + 0.0019062942592427135, + 0.002744821133092046, + 0.0028202023822814226, + 0.0035791315604001284, + 0.005039108917117119, + 0.002576654776930809, + 0.0026554383803159, + 0.005897591356188059, + 0.002526074880734086, + 0.0023350450210273266, + 0.002192461397498846, + 0.00509859062731266, + 0.004529756959527731, + 0.009757805615663528, + 0.0054748146794736385, + 0.004729846026748419, + 0.005429641809314489, + 0.005169438198208809, + 0.007810610346496105, + 0.05125853419303894, + 0.2158014178276062, + 0.035313207656145096, + 0.031133387237787247, + 0.11590855568647385, + 0.23354358971118927, + 0.01858292520046234, + 0.0439700223505497, + 0.0023151603527367115, + 0.002543969079852104, + 0.0018336116336286068, + 0.005578355863690376, + 0.0019411511020734906, + 0.0027167382650077343, + 0.0025898111052811146, + 0.0023080394603312016, + 0.003920943476259708, + 0.007069957442581654, + 0.0031307311728596687, + 0.0030400496907532215, + 0.002778769703581929, + 0.002529578283429146, + 0.003223580541089177, + 0.006605456117540598, + 0.0022298977710306644, + 0.0018389606848359108, + 0.002049781149253249, + 0.001967740710824728, + 0.0019392495742067695, + 0.0024702565278857946, + 0.0024337845388799906, + 0.0020543448626995087, + 0.0022937043104320765, + 0.0030063847079873085, + 0.002108100801706314, + 0.0022893461864441633, + 0.002479874063283205, + 0.0035697307903319597, + 0.003144842339679599, + 0.002669774228706956, + 0.0018670830177143216, + 0.002713355701416731, + 0.0023939809761941433, + 0.003483796026557684, + 0.0018617340829223394, + 0.002383333630859852, + 0.0021512508392333984, + 0.0022413714323192835 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 3, + 0 + ], + "selected_reward": -0.12706443667411804 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "shot shot from an extremely low angle, the camera slowly tilting upward to emphasize a sense of majestic oppression. The image depicts an ancient, dense primeval forest, drenched in the cold, clear blue light of a full moon at midnight. Within this stillness, a massive, humanoid creature, composed of the rough, deeply grooved bark of an ancient tree, occupies the center of the frame, its massive form dotted with moist moss. Countless glowing vines, like the creature's own veins, entwine and weave through the cracks of its bark, rhythmically pulsing with a mysterious, inner blue-violet light. With a dull, grinding sound, as if from deep within the earth, the creature, with an extremely slow and solemn gesture, carefully unfolds its arms, resembling those of enormous, dead branches; the movement is so deliberate and powerful that a subtle disturbance in the air is palpable. High-contrast moonlight streams down from above, interplaying with the glow of the vines within the creature, casting ever-changing, flowing, dramatic shadows across its textured body. The contraction and relaxation of every muscle – if you can call it that – are clearly outlined by the light and shadow, creating an intense and awe-inspiring cinematic lighting effect and an atmosphere that is ancient, mysterious, and full of vitality.", + "seed": 2870, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000034/group_2/index_6618_seed_2870_dp_5_rank_10_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 2, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.13858654592186212, + "vlm_logprob": -0.13858654592186212, + "vlm_yes_no_margin": 2.999999988079071, + "vlm_token_credit": -0.13858654592186212, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot from an extremely low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera slowly tilt upward during the shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the central subject a massive humanoid creature made of rough, deeply grooved tree bark?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are countless glowing vines entwined in the creature's bark, pulsing with blue-violet light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the creature slowly and deliberately unfold its arms resembling enormous dead branches?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.37380582094192505 + ], + "opposite_logprob_tokens": [ + -1.3738057613372803 + ], + "logprob_reward": -0.37380582094192505, + "yes_no_margin_reward": 0.9999999403953552, + "token_credit_reward": -0.37380582094192505, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.16965346038341522, + 0.05589206516742706, + 0.02959708496928215, + 0.14235667884349823, + 0.06302032619714737, + 0.11516837775707245, + 0.2504074275493622, + 0.062637560069561, + 0.06571387499570847, + 0.045553144067525864 + ], + "credit_map": [ + 0.017236480489373207, + 0.022307394072413445, + 0.014370477758347988, + 0.0191928930580616, + 0.018706263974308968, + 0.02403957024216652, + 0.03400575369596481, + 0.019794631749391556, + 0.00611934345215559, + 0.00853660050779581, + 0.0052452958188951015, + 0.006007135380059481, + 0.007027399726212025, + 0.00487165804952383, + 0.01271255686879158, + 0.005372075829654932, + 0.0028012467082589865, + 0.006269816774874926, + 0.004197415895760059, + 0.0034001879394054413, + 0.00281165586784482, + 0.0028928667306900024, + 0.004292846657335758, + 0.0029310493264347315, + 0.012363302521407604, + 0.017429277300834656, + 0.018238548189401627, + 0.02094360627233982, + 0.017608867958188057, + 0.016913574188947678, + 0.01889651268720627, + 0.019962981343269348, + 0.006442537065595388, + 0.00912997592240572, + 0.011100858449935913, + 0.012042250484228134, + 0.003228214103728533, + 0.0049711912870407104, + 0.006296878214925528, + 0.009808418340981007, + 0.010719750076532364, + 0.030767042189836502, + 0.01793668046593666, + 0.014589807949960232, + 0.007893871515989304, + 0.009992108680307865, + 0.00858942698687315, + 0.014679692685604095, + 0.020277870818972588, + 0.026438061147928238, + 0.008830979466438293, + 0.011243973858654499, + 0.09540793299674988, + 0.053397782146930695, + 0.02160562388598919, + 0.013205213472247124, + 0.007001310586929321, + 0.011090587824583054, + 0.008270378224551678, + 0.008686142042279243, + 0.008018264546990395, + 0.0038624743465334177, + 0.009510133415460587, + 0.0061982739716768265, + 0.008447858504951, + 0.011530403047800064, + 0.00615414185449481, + 0.00605414342135191, + 0.006354107055813074, + 0.010466247797012329, + 0.008815537206828594, + 0.007891436107456684, + 0.004023559391498566, + 0.0038159165997058153, + 0.0025414996780455112, + 0.003996729850769043, + 0.014279558323323727, + 0.003579754149541259, + 0.0036342341918498278, + 0.009681889787316322 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 0, + 3 + ], + "selected_reward": -0.37380582094192505 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot from an extremely low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera slowly tilt upward during the shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the central subject a massive humanoid creature made of rough, deeply grooved tree bark?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are countless glowing vines entwined in the creature's bark, pulsing with blue-violet light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the creature slowly and deliberately unfold its arms resembling enormous dead branches?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.2521543800830841 + ], + "opposite_logprob_tokens": [ + -1.5021543502807617 + ], + "logprob_reward": -0.2521543800830841, + "yes_no_margin_reward": 1.25, + "token_credit_reward": -0.2521543800830841, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1793142408132553, + 0.07782259583473206, + 0.06539051234722137, + 0.12034689635038376, + 0.035630688071250916, + 0.03869199752807617, + 0.22961080074310303, + 0.08556914329528809, + 0.05435873195528984, + 0.11326443403959274 + ], + "credit_map": [ + 0.02078455686569214, + 0.022438954561948776, + 0.014516075141727924, + 0.01948816515505314, + 0.01781579665839672, + 0.022088775411248207, + 0.0313819982111454, + 0.03079991042613983, + 0.006425593513995409, + 0.011628394946455956, + 0.006122100166976452, + 0.008845260366797447, + 0.007555563934147358, + 0.007157787214964628, + 0.018885763362050056, + 0.011202124878764153, + 0.005997566971927881, + 0.011983164586126804, + 0.006634331773966551, + 0.00844621192663908, + 0.005697314161807299, + 0.0077295671217143536, + 0.01126230601221323, + 0.007640053052455187, + 0.009137569926679134, + 0.014009114354848862, + 0.014803574420511723, + 0.019321803003549576, + 0.012186044827103615, + 0.016179407015442848, + 0.014752153307199478, + 0.019957231357693672, + 0.0036015247460454702, + 0.006342738401144743, + 0.0036438247188925743, + 0.004335733130574226, + 0.003055667271837592, + 0.0030107691418379545, + 0.0044863419607281685, + 0.007154092192649841, + 0.0038255793042480946, + 0.005850088316947222, + 0.005146019160747528, + 0.006691087037324905, + 0.0033743383828550577, + 0.004241060931235552, + 0.0044757104478776455, + 0.00508811604231596, + 0.0188759658485651, + 0.020277926698327065, + 0.016832346096634865, + 0.018263880163431168, + 0.06265527009963989, + 0.04552086815237999, + 0.020097242668271065, + 0.027087274938821793, + 0.009563811123371124, + 0.01255886536091566, + 0.01400427334010601, + 0.008562554605305195, + 0.009742188267409801, + 0.006701459176838398, + 0.011653758585453033, + 0.012782230973243713, + 0.005858832038938999, + 0.008103452622890472, + 0.004461268894374371, + 0.005567977204918861, + 0.006932723335921764, + 0.01058411318808794, + 0.006649488117545843, + 0.006200879346579313, + 0.011718765832483768, + 0.0103503642603755, + 0.006412850692868233, + 0.009992938488721848, + 0.034291744232177734, + 0.008701669983565807, + 0.008782525546848774, + 0.023013576865196228 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 0, + 3 + ], + "selected_reward": -0.2521543800830841 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot from an extremely low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera slowly tilt upward during the shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the central subject a massive humanoid creature made of rough, deeply grooved tree bark?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are countless glowing vines entwined in the creature's bark, pulsing with blue-violet light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the creature slowly and deliberately unfold its arms resembling enormous dead branches?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004107610322535038 + ], + "opposite_logprob_tokens": [ + -5.504107475280762 + ], + "logprob_reward": -0.004107610322535038, + "yes_no_margin_reward": 5.5, + "token_credit_reward": -0.004107610322535038, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09647909551858902, + 0.05207502841949463, + 0.069486103951931, + 0.10411137342453003, + 0.059718962758779526, + 0.17826630175113678, + 0.26575663685798645, + 0.07745394855737686, + 0.022429049015045166, + 0.07422347366809845 + ], + "credit_map": [ + 0.01700258068740368, + 0.012097426690161228, + 0.006850056815892458, + 0.009833930991590023, + 0.00911751575767994, + 0.014769993722438812, + 0.014205217361450195, + 0.012602378614246845, + 0.007660991046577692, + 0.0068582333624362946, + 0.005154972895979881, + 0.005298708099871874, + 0.004419358447194099, + 0.006539497058838606, + 0.009853919968008995, + 0.006289348006248474, + 0.008244993165135384, + 0.011629752814769745, + 0.008199146948754787, + 0.022124189883470535, + 0.002883998677134514, + 0.004004975315183401, + 0.005119324196130037, + 0.007279723882675171, + 0.007891750894486904, + 0.011891630478203297, + 0.011347792111337185, + 0.01511884294450283, + 0.012434616684913635, + 0.012656179256737232, + 0.01523024681955576, + 0.01754031889140606, + 0.0058084637857973576, + 0.004524897783994675, + 0.01028425619006157, + 0.013318531215190887, + 0.006353533361107111, + 0.004203996155411005, + 0.005474769044667482, + 0.009750518016517162, + 0.01657012663781643, + 0.04850020259618759, + 0.026403097435832024, + 0.024457741528749466, + 0.012712397612631321, + 0.013037839904427528, + 0.01647324487566948, + 0.02011164091527462, + 0.02138204500079155, + 0.02580474503338337, + 0.010607938282191753, + 0.012061934918165207, + 0.09532568603754044, + 0.06221189722418785, + 0.020852748304605484, + 0.017509620636701584, + 0.007381952367722988, + 0.012638211250305176, + 0.012996260076761246, + 0.009971541352570057, + 0.007946725003421307, + 0.005047906190156937, + 0.011927115730941296, + 0.009544235654175282, + 0.0025088766124099493, + 0.002771774074062705, + 0.002850213088095188, + 0.0024416313972324133, + 0.0024694688618183136, + 0.0038633549120277166, + 0.0027445987798273563, + 0.0027791315224021673, + 0.006625455338507891, + 0.007083880715072155, + 0.0030875226948410273, + 0.0071684918366372585, + 0.0204316359013319, + 0.0056759086437523365, + 0.006823582109063864, + 0.017326992005109787 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 5, + 3 + ], + "selected_reward": -0.004107610322535038 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot from an extremely low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera slowly tilt upward during the shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the central subject a massive humanoid creature made of rough, deeply grooved tree bark?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are countless glowing vines entwined in the creature's bark, pulsing with blue-violet light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the creature slowly and deliberately unfold its arms resembling enormous dead branches?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.014204349368810654 + ], + "opposite_logprob_tokens": [ + -4.264204502105713 + ], + "logprob_reward": -0.014204349368810654, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.014204349368810654, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.059272851794958115, + 0.048534754663705826, + 0.08720044791698456, + 0.12769116461277008, + 0.22310782968997955, + 0.07738702744245529, + 0.26179248094558716, + 0.06717821210622787, + 0.024669254198670387, + 0.023165958002209663 + ], + "credit_map": [ + 0.007478600367903709, + 0.007677980232983828, + 0.004558292683213949, + 0.006704133003950119, + 0.0050915880128741264, + 0.010421480983495712, + 0.005194114521145821, + 0.01214666198939085, + 0.005781461019068956, + 0.006951423827558756, + 0.004279311280697584, + 0.006095381919294596, + 0.00433270214125514, + 0.005808187648653984, + 0.010941606014966965, + 0.00434468500316143, + 0.009512764401733875, + 0.016968371346592903, + 0.010455857031047344, + 0.019403574988245964, + 0.005852988921105862, + 0.007522507570683956, + 0.005893087014555931, + 0.011591300368309021, + 0.010683195665478706, + 0.017300108447670937, + 0.01583077572286129, + 0.018218213692307472, + 0.013670282438397408, + 0.01619855873286724, + 0.015242483466863632, + 0.020547544583678246, + 0.023965464904904366, + 0.037244465202093124, + 0.0324416309595108, + 0.03323570266366005, + 0.015270614996552467, + 0.014483162201941013, + 0.02907598577439785, + 0.03739078715443611, + 0.006658248137682676, + 0.016015570610761642, + 0.008441709913313389, + 0.015225674025714397, + 0.007480780128389597, + 0.009889901615679264, + 0.007327754516154528, + 0.006347383838146925, + 0.02100585214793682, + 0.02512269653379917, + 0.00941556878387928, + 0.010978484526276588, + 0.09636096656322479, + 0.06229707971215248, + 0.020216217264533043, + 0.01639562100172043, + 0.00740385428071022, + 0.008147290907800198, + 0.006368701346218586, + 0.006443243008106947, + 0.010729262605309486, + 0.013137257657945156, + 0.005711581092327833, + 0.009237018413841724, + 0.0027468204498291016, + 0.003038863418623805, + 0.0030162930488586426, + 0.0029360861517488956, + 0.00473044766113162, + 0.0033263759687542915, + 0.002512046368792653, + 0.002362322760745883, + 0.0022903603967279196, + 0.00242691021412611, + 0.0025209353771060705, + 0.0027179161552339792, + 0.0036908695474267006, + 0.0026959478855133057, + 0.003263638587668538, + 0.003559378208592534 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 4, + 3 + ], + "selected_reward": -0.014204349368810654 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot from an extremely low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera slowly tilt upward during the shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the central subject a massive humanoid creature made of rough, deeply grooved tree bark?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are countless glowing vines entwined in the creature's bark, pulsing with blue-violet light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the creature slowly and deliberately unfold its arms resembling enormous dead branches?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04866056889295578 + ], + "opposite_logprob_tokens": [ + -3.0486605167388916 + ], + "logprob_reward": -0.04866056889295578, + "yes_no_margin_reward": 3.0, + "token_credit_reward": -0.04866056889295578, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07481260597705841, + 0.10599873214960098, + 0.03858298435807228, + 0.11956333369016647, + 0.09035919606685638, + 0.09409385174512863, + 0.1493305116891861, + 0.09308595210313797, + 0.07605526596307755, + 0.15811757743358612 + ], + "credit_map": [ + 0.006849356926977634, + 0.009142998605966568, + 0.006757005117833614, + 0.007937774062156677, + 0.006611345335841179, + 0.012856286019086838, + 0.012590473517775536, + 0.012067366391420364, + 0.009692457504570484, + 0.015617257915437222, + 0.011904201470315456, + 0.012335026636719704, + 0.00981636717915535, + 0.01013663038611412, + 0.027328558266162872, + 0.009168234653770924, + 0.003638420021161437, + 0.006196983158588409, + 0.004946193192154169, + 0.011168472468852997, + 0.002927863970398903, + 0.0033665827941149473, + 0.0031514496076852083, + 0.003187020542100072, + 0.010593999177217484, + 0.015565402805805206, + 0.0148024195805192, + 0.01957179792225361, + 0.011801449581980705, + 0.015211198478937149, + 0.014600066468119621, + 0.017416998744010925, + 0.011187925934791565, + 0.011830003000795841, + 0.013139652088284492, + 0.014072991907596588, + 0.007924651727080345, + 0.006977187003940344, + 0.010292286984622478, + 0.014934496022760868, + 0.0085407430306077, + 0.02227104641497135, + 0.013631956651806831, + 0.01123768463730812, + 0.007621548138558865, + 0.006810474209487438, + 0.008442656137049198, + 0.01553773507475853, + 0.011775430291891098, + 0.01506638340651989, + 0.0077639794908463955, + 0.008407817222177982, + 0.05013895779848099, + 0.031794432550668716, + 0.012024848721921444, + 0.01235865242779255, + 0.008984431624412537, + 0.013300622813403606, + 0.010249759070575237, + 0.013250674121081829, + 0.013620262034237385, + 0.008005665615200996, + 0.014185675419867039, + 0.011488854885101318, + 0.008230358362197876, + 0.010532715357840061, + 0.008875326253473759, + 0.009284812957048416, + 0.008793939836323261, + 0.01203613355755806, + 0.00955934077501297, + 0.008742636069655418, + 0.014553959481418133, + 0.01766122132539749, + 0.009061307646334171, + 0.01873103342950344, + 0.03500380367040634, + 0.012811163440346718, + 0.017089851200580597, + 0.03320523723959923 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 6, + 3 + ], + "selected_reward": -0.04866056889295578 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000034/group_3/index_tensor([3278, 3278])_group_3_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000034/group_3/index_tensor([3278, 3278])_group_3_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..81966a8a355ef1fa331c17fc49c584e023a7d54e --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000034/group_3/index_tensor([3278, 3278])_group_3_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "dynamic, upward-looking tracking shot follows a mythical phoenix soaring through the deep, inky night sky. Enveloped in a luminous, warm golden flame, the phoenix gracefully flickers and dances with its movements. Each feather seems forged from pure light, radiating a fiery, divine glow. Its vast, magnificent wings are fully extended, their edges gleaming like rubies and sapphires in the flames. Its elegant head is held high, its eyes gleaming with wisdom and eternity. Each flap of its wings is graceful yet powerful, propelling it in an elegant arc across a night dotted with countless shimmering stars. Behind it, a long, iridescent ribbon of light, composed of countless shimmering particles of light, trails out, a brilliant rainbow-like trail. This ribbon blends blazing orange, deep purple, and brilliant cyan, slowly dissipating into the darkness. The phoenix itself is the sole source of light in the scene, its golden radiance tinging the surrounding night sky with a warm hue, creating an epic atmosphere of sacredness, grandeur, and eternal vitality.", + "seed": 22244, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000034/group_3/index_3278_seed_22244_dp_6_rank_12_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 3, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.021674728579819204, + "vlm_logprob": -0.021674728579819204, + "vlm_yes_no_margin": 4.9, + "vlm_token_credit": -0.021674728579819204, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a mythical phoenix with wings fully extended?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the phoenix enveloped in luminous, warm golden flame with feathers appearing to be made of pure light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the phoenix move in a graceful, dancing motion propelling it in an elegant arc across the sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is there a long, iridescent ribbon of light trailing behind the phoenix composed of orange, purple, and cyan particles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera executing a dynamic, upward-looking tracking shot following the phoenix?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0034073886927217245 + ], + "opposite_logprob_tokens": [ + -6.0034074783325195 + ], + "logprob_reward": -0.0034073886927217245, + "yes_no_margin_reward": 6.0, + "token_credit_reward": -0.0034073886927217245, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.16693024337291718, + 0.08966981619596481, + 0.10786605626344681, + 0.13027019798755646, + 0.09957176446914673, + 0.07510283589363098, + 0.09496629238128662, + 0.10640811920166016, + 0.06580320000648499, + 0.06341149657964706 + ], + "credit_map": [ + 0.030018551275134087, + 0.015986859798431396, + 0.0206863172352314, + 0.019488045945763588, + 0.016342036426067352, + 0.01731148734688759, + 0.020561158657073975, + 0.02653578296303749, + 0.014811142347753048, + 0.008601073175668716, + 0.00911426730453968, + 0.010031397454440594, + 0.010426593944430351, + 0.009040000848472118, + 0.013824079185724258, + 0.013821261934936047, + 0.01392457541078329, + 0.009319490753114223, + 0.008574441075325012, + 0.007793774362653494, + 0.007546655368059874, + 0.03205803781747818, + 0.008102886378765106, + 0.02054620161652565, + 0.012141338549554348, + 0.013943611644208431, + 0.01941167376935482, + 0.014428100548684597, + 0.01945212297141552, + 0.017817659303545952, + 0.019551310688257217, + 0.013524371199309826, + 0.010009245947003365, + 0.006745305843651295, + 0.027679504826664925, + 0.01168023981153965, + 0.008103939704596996, + 0.00997478049248457, + 0.009466921910643578, + 0.015911826863884926, + 0.0067923166789114475, + 0.004364520777016878, + 0.006083239335566759, + 0.008466675877571106, + 0.004554185550659895, + 0.02506505325436592, + 0.011380997486412525, + 0.008395846001803875, + 0.008178259246051311, + 0.006138222757726908, + 0.019011732190847397, + 0.011799282394349575, + 0.007670701015740633, + 0.00971523392945528, + 0.013031430542469025, + 0.01942143402993679, + 0.009921345859766006, + 0.008925560861825943, + 0.016527492552995682, + 0.012693722732365131, + 0.011823363602161407, + 0.01535825990140438, + 0.014149725437164307, + 0.01700863614678383, + 0.00830464530736208, + 0.005042690318077803, + 0.00947090145200491, + 0.008125354535877705, + 0.007192092482000589, + 0.008089949376881123, + 0.010144306346774101, + 0.009433262050151825, + 0.009905092418193817, + 0.0064120120368897915, + 0.009726455435156822, + 0.006315167993307114, + 0.009258822537958622, + 0.005238011945039034, + 0.008020923472940922, + 0.008535011671483517 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 2 + ], + "selected_reward": -0.0034073886927217245 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a mythical phoenix with wings fully extended?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the phoenix enveloped in luminous, warm golden flame with feathers appearing to be made of pure light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the phoenix move in a graceful, dancing motion propelling it in an elegant arc across the sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is there a long, iridescent ribbon of light trailing behind the phoenix composed of orange, purple, and cyan particles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera executing a dynamic, upward-looking tracking shot following the phoenix?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0013440155889838934 + ], + "opposite_logprob_tokens": [ + -6.626344203948975 + ], + "logprob_reward": -0.0013440155889838934, + "yes_no_margin_reward": 6.625, + "token_credit_reward": -0.0013440155889838934, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.17126061022281647, + 0.03573017567396164, + 0.21717245876789093, + 0.11270667612552643, + 0.06615976244211197, + 0.07363618910312653, + 0.06455905735492706, + 0.0446760319173336, + 0.09864897280931473, + 0.11545007675886154 + ], + "credit_map": [ + 0.024732448160648346, + 0.017804039642214775, + 0.017805488780140877, + 0.01907411962747574, + 0.017752764746546745, + 0.021433012560009956, + 0.02910085767507553, + 0.0235578790307045, + 0.0071321758441627026, + 0.003943272400647402, + 0.003805345855653286, + 0.004020965192466974, + 0.004133053123950958, + 0.0033932419028133154, + 0.0029824881348758936, + 0.006319635082036257, + 0.028086131438612938, + 0.013433409854769707, + 0.012256107293069363, + 0.00681976368650794, + 0.009702750481665134, + 0.08598017692565918, + 0.015968866646289825, + 0.044925257563591, + 0.01167700532823801, + 0.007969639264047146, + 0.017220158129930496, + 0.012519351206719875, + 0.012255597859621048, + 0.020899202674627304, + 0.014644681476056576, + 0.015521037392318249, + 0.006343384739011526, + 0.007018615957349539, + 0.015619776211678982, + 0.007278462406247854, + 0.0068216812796890736, + 0.007371813524514437, + 0.007155661936849356, + 0.008550365455448627, + 0.008883019909262657, + 0.0091355349868536, + 0.011318089440464973, + 0.010081423446536064, + 0.009373444132506847, + 0.009988579899072647, + 0.008661007508635521, + 0.006195098161697388, + 0.005919999442994595, + 0.004380174446851015, + 0.011285086162388325, + 0.010391001589596272, + 0.003765285015106201, + 0.006736680865287781, + 0.006538815330713987, + 0.015542017295956612, + 0.0041297199204564095, + 0.002675882074981928, + 0.007618745788931847, + 0.009240851737558842, + 0.00285809813067317, + 0.0037240791134536266, + 0.005422758404165506, + 0.009005897678434849, + 0.012845760211348534, + 0.017512449994683266, + 0.013406087644398212, + 0.008590747602283955, + 0.015932831913232803, + 0.010010968893766403, + 0.008894640952348709, + 0.01145548839122057, + 0.01023031584918499, + 0.008892194367945194, + 0.016630638390779495, + 0.023072589188814163, + 0.014832920394837856, + 0.021814746782183647, + 0.009159550070762634, + 0.010817124508321285 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 0, + 9 + ], + "selected_reward": -0.0013440155889838934 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a mythical phoenix with wings fully extended?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the phoenix enveloped in luminous, warm golden flame with feathers appearing to be made of pure light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the phoenix move in a graceful, dancing motion propelling it in an elegant arc across the sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is there a long, iridescent ribbon of light trailing behind the phoenix composed of orange, purple, and cyan particles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera executing a dynamic, upward-looking tracking shot following the phoenix?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02057623490691185 + ], + "opposite_logprob_tokens": [ + -3.895576238632202 + ], + "logprob_reward": -0.02057623490691185, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.02057623490691185, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15644831955432892, + 0.09370237588882446, + 0.062329135835170746, + 0.22701530158519745, + 0.060012996196746826, + 0.14928312599658966, + 0.0922008529305458, + 0.03732622414827347, + 0.06432578712701797, + 0.057355862110853195 + ], + "credit_map": [ + 0.026503784582018852, + 0.010660787113010883, + 0.018065400421619415, + 0.02123943716287613, + 0.016035014763474464, + 0.021140918135643005, + 0.017243066802620888, + 0.02555990405380726, + 0.01583239436149597, + 0.00928282830864191, + 0.007192846387624741, + 0.008892636746168137, + 0.015053628012537956, + 0.009862387552857399, + 0.015410569496452808, + 0.01217508316040039, + 0.009548082947731018, + 0.008651914075016975, + 0.006955106742680073, + 0.0059857722371816635, + 0.010552196763455868, + 0.006693705450743437, + 0.008375227451324463, + 0.005567128770053387, + 0.020323369652032852, + 0.017169402912259102, + 0.032175760716199875, + 0.026866480708122253, + 0.028196491301059723, + 0.04193314537405968, + 0.028130633756518364, + 0.0322200246155262, + 0.0076586417853832245, + 0.005151285324245691, + 0.013460592366755009, + 0.0066469525918364525, + 0.005671207793056965, + 0.00721485773101449, + 0.004695266019552946, + 0.009514190256595612, + 0.013019677251577377, + 0.016581017524003983, + 0.01588214933872223, + 0.018222298473119736, + 0.0163145512342453, + 0.03388757258653641, + 0.022944694384932518, + 0.01243116706609726, + 0.009374769404530525, + 0.006123298779129982, + 0.016121981665492058, + 0.014885577373206615, + 0.0054045552387833595, + 0.010630594566464424, + 0.010524986311793327, + 0.019135085865855217, + 0.0044737025164067745, + 0.002585549373179674, + 0.005866013467311859, + 0.006217667367309332, + 0.002506841206923127, + 0.005604657810181379, + 0.00489779282361269, + 0.005173996090888977, + 0.006066166330128908, + 0.00940025132149458, + 0.013425327837467194, + 0.0067688291892409325, + 0.007815024815499783, + 0.006166925188153982, + 0.005533972755074501, + 0.00914929062128067, + 0.006075004581362009, + 0.005488263908773661, + 0.009961619973182678, + 0.007611387874931097, + 0.004980094730854034, + 0.011898549273610115, + 0.005105063784867525, + 0.006235883105546236 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 0, + 5 + ], + "selected_reward": -0.02057623490691185 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a mythical phoenix with wings fully extended?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the phoenix enveloped in luminous, warm golden flame with feathers appearing to be made of pure light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the phoenix move in a graceful, dancing motion propelling it in an elegant arc across the sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is there a long, iridescent ribbon of light trailing behind the phoenix composed of orange, purple, and cyan particles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera executing a dynamic, upward-looking tracking shot following the phoenix?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004091701935976744 + ], + "opposite_logprob_tokens": [ + -5.504091739654541 + ], + "logprob_reward": -0.004091701935976744, + "yes_no_margin_reward": 5.5, + "token_credit_reward": -0.004091701935976744, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.48148205876350403, + 0.06065456569194794, + 0.1708102822303772, + 0.13772541284561157, + 0.03387805074453354, + 0.03033149428665638, + 0.016841653734445572, + 0.017062345519661903, + 0.027692724019289017, + 0.023521455004811287 + ], + "credit_map": [ + 0.06136080622673035, + 0.04463092237710953, + 0.06258755922317505, + 0.08257832378149033, + 0.04594375565648079, + 0.060037169605493546, + 0.07029078155755997, + 0.054052747786045074, + 0.013067571446299553, + 0.0041089230217039585, + 0.005119778215885162, + 0.006026387680321932, + 0.0045713214203715324, + 0.0074791801162064075, + 0.006965953856706619, + 0.013315451331436634, + 0.022371483966708183, + 0.010228250175714493, + 0.00841430015861988, + 0.004653689917176962, + 0.007507152855396271, + 0.07114619761705399, + 0.010353351011872292, + 0.03613585606217384, + 0.01552383229136467, + 0.018079420551657677, + 0.014010850340127945, + 0.014641823247075081, + 0.015729881823062897, + 0.02868732251226902, + 0.014716736972332, + 0.01633555255830288, + 0.004492823965847492, + 0.003812091890722513, + 0.005893438123166561, + 0.0035569057799875736, + 0.0031512880232185125, + 0.004038148559629917, + 0.005044692661613226, + 0.0038886612746864557, + 0.0024416542146354914, + 0.0025660612154752016, + 0.0035509911831468344, + 0.002787476871162653, + 0.0022755246609449387, + 0.008852018974721432, + 0.0033926530741155148, + 0.0044651152566075325, + 0.0016759892459958792, + 0.000986520666629076, + 0.0035678858403116465, + 0.0021467984188348055, + 0.0009702335228212178, + 0.0017448249273002148, + 0.0019946489483118057, + 0.003754752455279231, + 0.0016000621253624558, + 0.001195206306874752, + 0.002771461382508278, + 0.0014532693894580007, + 0.0017052325420081615, + 0.0025900336913764477, + 0.002912559313699603, + 0.0028345203027129173, + 0.0031887427903711796, + 0.004546465817838907, + 0.003585794707760215, + 0.0030742897652089596, + 0.004077963065356016, + 0.002503215102478862, + 0.0031094136647880077, + 0.0036068400368094444, + 0.0033026833552867174, + 0.002235915046185255, + 0.004491810686886311, + 0.002708453219383955, + 0.002452602842822671, + 0.0028247712180018425, + 0.002167204860597849, + 0.003338012844324112 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 3 + ], + "selected_reward": -0.004091701935976744 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a mythical phoenix with wings fully extended?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the phoenix enveloped in luminous, warm golden flame with feathers appearing to be made of pure light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the phoenix move in a graceful, dancing motion propelling it in an elegant arc across the sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is there a long, iridescent ribbon of light trailing behind the phoenix composed of orange, purple, and cyan particles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera executing a dynamic, upward-looking tracking shot following the phoenix?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0789543017745018 + ], + "opposite_logprob_tokens": [ + -2.5789542198181152 + ], + "logprob_reward": -0.0789543017745018, + "yes_no_margin_reward": 2.5, + "token_credit_reward": -0.0789543017745018, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10602717101573944, + 0.07517632842063904, + 0.2848460078239441, + 0.18421652913093567, + 0.02287324331700802, + 0.15743088722229004, + 0.05894249677658081, + 0.020974373444914818, + 0.04450003430247307, + 0.0450129508972168 + ], + "credit_map": [ + 0.017986299470067024, + 0.010011776350438595, + 0.011195089668035507, + 0.009492509067058563, + 0.012189386412501335, + 0.01555741298943758, + 0.01453858520835638, + 0.015056116506457329, + 0.010159069672226906, + 0.0092142503708601, + 0.006941843777894974, + 0.0066315121948719025, + 0.013019832782447338, + 0.008858824148774147, + 0.011490454897284508, + 0.008860547095537186, + 0.033269137144088745, + 0.017578676342964172, + 0.017868468537926674, + 0.007913880050182343, + 0.010400696657598019, + 0.11426304280757904, + 0.01753859966993332, + 0.06601349264383316, + 0.01467140018939972, + 0.020554162561893463, + 0.02749980241060257, + 0.020234961062669754, + 0.02469773404300213, + 0.027984708547592163, + 0.023773284628987312, + 0.02480047382414341, + 0.0028751520439982414, + 0.0018344686832278967, + 0.005485163070261478, + 0.0023864188697189093, + 0.0021262478549033403, + 0.00239023519679904, + 0.00278816488571465, + 0.0029873959720134735, + 0.013305744156241417, + 0.006847340147942305, + 0.00924841407686472, + 0.006892737466841936, + 0.008886095136404037, + 0.06708703190088272, + 0.02628081664443016, + 0.018882717937231064, + 0.006987758446484804, + 0.004297284875065088, + 0.009138248860836029, + 0.006108958274126053, + 0.0039360662922263145, + 0.008369848132133484, + 0.007598700933158398, + 0.012505637481808662, + 0.002437023678794503, + 0.0013683115830644965, + 0.004686844069510698, + 0.0026545273140072823, + 0.0015651326393708587, + 0.0023419775534421206, + 0.0024178451858460903, + 0.0035027109552174807, + 0.004094221629202366, + 0.003335464745759964, + 0.008207115344703197, + 0.0072865113615989685, + 0.003018778981640935, + 0.0054536075331270695, + 0.0047741723246872425, + 0.008330166339874268, + 0.006353145930916071, + 0.003389024641364813, + 0.007012851536273956, + 0.006124549079686403, + 0.005206409841775894, + 0.006348378490656614, + 0.004934494383633137, + 0.005644103046506643 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 3, + 5 + ], + "selected_reward": -0.0789543017745018 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "dynamic, upward-looking tracking shot follows a mythical phoenix soaring through the deep, inky night sky. Enveloped in a luminous, warm golden flame, the phoenix gracefully flickers and dances with its movements. Each feather seems forged from pure light, radiating a fiery, divine glow. Its vast, magnificent wings are fully extended, their edges gleaming like rubies and sapphires in the flames. Its elegant head is held high, its eyes gleaming with wisdom and eternity. Each flap of its wings is graceful yet powerful, propelling it in an elegant arc across a night dotted with countless shimmering stars. Behind it, a long, iridescent ribbon of light, composed of countless shimmering particles of light, trails out, a brilliant rainbow-like trail. This ribbon blends blazing orange, deep purple, and brilliant cyan, slowly dissipating into the darkness. The phoenix itself is the sole source of light in the scene, its golden radiance tinging the surrounding night sky with a warm hue, creating an epic atmosphere of sacredness, grandeur, and eternal vitality.", + "seed": 22245, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000034/group_3/index_3278_seed_22245_dp_6_rank_12_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 3, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.029530081315897405, + "vlm_logprob": -0.029530081315897405, + "vlm_yes_no_margin": 4.325, + "vlm_token_credit": -0.029530081315897405, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a mythical phoenix with wings fully extended?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the phoenix enveloped in luminous, warm golden flame with feathers appearing to be made of pure light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the phoenix move in a graceful, dancing motion propelling it in an elegant arc across the sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is there a long, iridescent ribbon of light trailing behind the phoenix composed of orange, purple, and cyan particles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera executing a dynamic, upward-looking tracking shot following the phoenix?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0047096991911530495 + ], + "opposite_logprob_tokens": [ + -5.754709720611572 + ], + "logprob_reward": -0.0047096991911530495, + "yes_no_margin_reward": 5.75, + "token_credit_reward": -0.0047096991911530495, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12452516704797745, + 0.17653287947177887, + 0.10128763318061829, + 0.08760374039411545, + 0.03456384688615799, + 0.02951895445585251, + 0.07525050640106201, + 0.10624194890260696, + 0.062301330268383026, + 0.20217400789260864 + ], + "credit_map": [ + 0.013422640971839428, + 0.017165767028927803, + 0.014841667376458645, + 0.016760367900133133, + 0.016785165295004845, + 0.015781357884407043, + 0.01859164796769619, + 0.011176560074090958, + 0.04059736430644989, + 0.00836094468832016, + 0.06413136422634125, + 0.0251455195248127, + 0.006578371860086918, + 0.00939993280917406, + 0.01109279878437519, + 0.011226593516767025, + 0.011746012605726719, + 0.010313338600099087, + 0.017974115908145905, + 0.016941867768764496, + 0.010299531742930412, + 0.010396999306976795, + 0.012150556780397892, + 0.011465203016996384, + 0.011638468131422997, + 0.005820372607558966, + 0.025455046445131302, + 0.009823854081332684, + 0.007573320996016264, + 0.007057663518935442, + 0.010313613340258598, + 0.009921395219862461, + 0.0036622711922973394, + 0.0026235354598611593, + 0.005237475037574768, + 0.004523183219134808, + 0.004079898353666067, + 0.0037072449922561646, + 0.0054818796925246716, + 0.005248356610536575, + 0.00296897254884243, + 0.0026640237774699926, + 0.0058130305260419846, + 0.0036395208444446325, + 0.002446755999699235, + 0.004808011930435896, + 0.0036497802939265966, + 0.0035288583021610975, + 0.010663427412509918, + 0.010683191940188408, + 0.013114248402416706, + 0.006155830342322588, + 0.007891532965004444, + 0.011330408975481987, + 0.006553086917847395, + 0.008858785033226013, + 0.01084620226174593, + 0.009323653765022755, + 0.017016543075442314, + 0.013457821682095528, + 0.01914321444928646, + 0.012482210993766785, + 0.012610426172614098, + 0.011361873708665371, + 0.007243314292281866, + 0.004948019050061703, + 0.00978722795844078, + 0.008248304016888142, + 0.009641753509640694, + 0.00932442955672741, + 0.00655349949374795, + 0.006554781459271908, + 0.03456231579184532, + 0.01634684018790722, + 0.029451407492160797, + 0.035598427057266235, + 0.01788097247481346, + 0.01828257367014885, + 0.025457167997956276, + 0.024594301357865334 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 1, + 0 + ], + "selected_reward": -0.0047096991911530495 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a mythical phoenix with wings fully extended?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the phoenix enveloped in luminous, warm golden flame with feathers appearing to be made of pure light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the phoenix move in a graceful, dancing motion propelling it in an elegant arc across the sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is there a long, iridescent ribbon of light trailing behind the phoenix composed of orange, purple, and cyan particles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera executing a dynamic, upward-looking tracking shot following the phoenix?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0019468179671093822 + ], + "opposite_logprob_tokens": [ + -6.251946926116943 + ], + "logprob_reward": -0.0019468179671093822, + "yes_no_margin_reward": 6.25, + "token_credit_reward": -0.0019468179671093822, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12909266352653503, + 0.21593092381954193, + 0.11427304893732071, + 0.06088371202349663, + 0.032152581959962845, + 0.05902222916483879, + 0.026288986206054688, + 0.19373281300067902, + 0.12047164142131805, + 0.048151422291994095 + ], + "credit_map": [ + 0.017521139234304428, + 0.014500339515507221, + 0.014999918639659882, + 0.015100166201591492, + 0.017578866332769394, + 0.01609523966908455, + 0.018186835572123528, + 0.015110145322978497, + 0.049158331006765366, + 0.010774567723274231, + 0.08221350610256195, + 0.01906474307179451, + 0.008512104861438274, + 0.027675263583660126, + 0.010028568096458912, + 0.008503821678459644, + 0.02197994291782379, + 0.012820604257285595, + 0.008398204110562801, + 0.01501526404172182, + 0.008004509843885899, + 0.014020729809999466, + 0.02655375748872757, + 0.007480029948055744, + 0.007097709458321333, + 0.004534789826720953, + 0.01667650043964386, + 0.006946950685232878, + 0.005207994487136602, + 0.006521131843328476, + 0.006089914124459028, + 0.007808719761669636, + 0.00301531795412302, + 0.00261003989726305, + 0.004911596886813641, + 0.004284361843019724, + 0.004288946744054556, + 0.004019260872155428, + 0.004311045166105032, + 0.004712010733783245, + 0.006707518361508846, + 0.004584657493978739, + 0.00983575638383627, + 0.005593432113528252, + 0.003911908250302076, + 0.010885138064622879, + 0.009732403792440891, + 0.00777141097933054, + 0.002644498599693179, + 0.0028292639181017876, + 0.00488020945340395, + 0.003373105311766267, + 0.004496335051953793, + 0.004003413952887058, + 0.0024941344745457172, + 0.0015680205542594194, + 0.020831283181905746, + 0.015133347362279892, + 0.04499617591500282, + 0.02476152405142784, + 0.028507418930530548, + 0.024220289662480354, + 0.018389765173196793, + 0.01689300499856472, + 0.014730441384017467, + 0.009965375065803528, + 0.021208185702562332, + 0.018449656665325165, + 0.016162103042006493, + 0.016628190875053406, + 0.012591425329446793, + 0.010736260563135147, + 0.004859922453761101, + 0.006365557201206684, + 0.006476180627942085, + 0.00981434341520071, + 0.005816841032356024, + 0.0042852843180298805, + 0.0057585700415074825, + 0.004774720408022404 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 7, + 0 + ], + "selected_reward": -0.0019468179671093822 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a mythical phoenix with wings fully extended?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the phoenix enveloped in luminous, warm golden flame with feathers appearing to be made of pure light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the phoenix move in a graceful, dancing motion propelling it in an elegant arc across the sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is there a long, iridescent ribbon of light trailing behind the phoenix composed of orange, purple, and cyan particles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera executing a dynamic, upward-looking tracking shot following the phoenix?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.05494414269924164 + ], + "opposite_logprob_tokens": [ + -2.9299440383911133 + ], + "logprob_reward": -0.05494414269924164, + "yes_no_margin_reward": 2.875, + "token_credit_reward": -0.05494414269924164, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13230355083942413, + 0.06305018067359924, + 0.256043016910553, + 0.0633145123720169, + 0.028750767931342125, + 0.03367907181382179, + 0.06460685282945633, + 0.1693711280822754, + 0.045153070241212845, + 0.14372782409191132 + ], + "credit_map": [ + 0.017359666526317596, + 0.03058285266160965, + 0.011273466050624847, + 0.012369676493108273, + 0.023363953456282616, + 0.01351888943463564, + 0.012486123479902744, + 0.011348926462233067, + 0.011449698358774185, + 0.006916208658367395, + 0.009625370614230633, + 0.007938502356410027, + 0.005214647389948368, + 0.006415045820176601, + 0.008266040123999119, + 0.007224665954709053, + 0.03483576700091362, + 0.02247600257396698, + 0.02778112329542637, + 0.033265672624111176, + 0.016271542757749557, + 0.03728608414530754, + 0.06897982954978943, + 0.015147000551223755, + 0.008339163847267628, + 0.0040319147519767284, + 0.01883593574166298, + 0.0067023904994130135, + 0.00478457473218441, + 0.005907231941819191, + 0.007991624996066093, + 0.006721672601997852, + 0.0029707271605730057, + 0.002518978202715516, + 0.0046661305241286755, + 0.003810261841863394, + 0.003955881576985121, + 0.0030122387688606977, + 0.00414029136300087, + 0.0036762573290616274, + 0.0033263342920690775, + 0.003125510411337018, + 0.007651180494576693, + 0.003538606921210885, + 0.003500326070934534, + 0.004415505100041628, + 0.0051108067855238914, + 0.0030108036007732153, + 0.007041650824248791, + 0.008565842173993587, + 0.01256926916539669, + 0.0059995027258992195, + 0.008306988514959812, + 0.007716713473200798, + 0.009732553735375404, + 0.0046743368729949, + 0.0201653353869915, + 0.012331189587712288, + 0.038666997104883194, + 0.020529571920633316, + 0.027260171249508858, + 0.013862403109669685, + 0.01825500652194023, + 0.018300458788871765, + 0.005663890391588211, + 0.003140527056530118, + 0.0070166331715881824, + 0.007113551255315542, + 0.007722855545580387, + 0.0035992746707051992, + 0.005012422800064087, + 0.0058839162811636925, + 0.023372171446681023, + 0.016657855361700058, + 0.018204038962721825, + 0.029234562069177628, + 0.015087269246578217, + 0.011268813163042068, + 0.016482556238770485, + 0.013420557603240013 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 7, + 9 + ], + "selected_reward": -0.05494414269924164 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a mythical phoenix with wings fully extended?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the phoenix enveloped in luminous, warm golden flame with feathers appearing to be made of pure light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the phoenix move in a graceful, dancing motion propelling it in an elegant arc across the sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is there a long, iridescent ribbon of light trailing behind the phoenix composed of orange, purple, and cyan particles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera executing a dynamic, upward-looking tracking shot following the phoenix?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.016060274094343185 + ], + "opposite_logprob_tokens": [ + -4.1410603523254395 + ], + "logprob_reward": -0.016060274094343185, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.016060274094343185, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.17833057045936584, + 0.1422734558582306, + 0.14761655032634735, + 0.0789816603064537, + 0.09742845594882965, + 0.0732971802353859, + 0.04342909902334213, + 0.062264833599328995, + 0.1372985988855362, + 0.039079587906599045 + ], + "credit_map": [ + 0.012877123430371284, + 0.06105324625968933, + 0.006973049137741327, + 0.008856134489178658, + 0.05849206820130348, + 0.01057227049022913, + 0.01325095072388649, + 0.0062557305209338665, + 0.012936938554048538, + 0.007495770696550608, + 0.03476333990693092, + 0.014106622897088528, + 0.010214242152869701, + 0.04085317999124527, + 0.01359438244253397, + 0.00830897968262434, + 0.027580469846725464, + 0.017239956185221672, + 0.01645912416279316, + 0.021989116445183754, + 0.010056805796921253, + 0.014952508732676506, + 0.026890283450484276, + 0.012448293156921864, + 0.008088257163763046, + 0.0066434540785849094, + 0.01083053182810545, + 0.012867959216237068, + 0.00633618701249361, + 0.011708761565387249, + 0.010467611253261566, + 0.012038894928991795, + 0.009088987484574318, + 0.009117739275097847, + 0.016259096562862396, + 0.011707364581525326, + 0.013125602155923843, + 0.013899339362978935, + 0.014057779684662819, + 0.010172557085752487, + 0.008252179250121117, + 0.006543018389493227, + 0.014979739673435688, + 0.0072226375341415405, + 0.0061217667534947395, + 0.011921504512429237, + 0.008641751483082771, + 0.009614585898816586, + 0.004832952283322811, + 0.004228377249091864, + 0.006139155011624098, + 0.004855936858803034, + 0.007110808975994587, + 0.005447503179311752, + 0.005047123413532972, + 0.005767237860709429, + 0.004764828365296125, + 0.004705408122390509, + 0.015676015987992287, + 0.010248279199004173, + 0.007225825916975737, + 0.0071681784465909, + 0.006644360721111298, + 0.005831936374306679, + 0.023224512115120888, + 0.011243524961173534, + 0.019057197496294975, + 0.023726051673293114, + 0.01699570007622242, + 0.011765295639634132, + 0.015737712383270264, + 0.015548602677881718, + 0.004233864601701498, + 0.004619404207915068, + 0.0046599009074270725, + 0.008197289891541004, + 0.0033744985703378916, + 0.003880397416651249, + 0.00490246806293726, + 0.0052117640152573586 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 1 + ], + "selected_reward": -0.016060274094343185 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a mythical phoenix with wings fully extended?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the phoenix enveloped in luminous, warm golden flame with feathers appearing to be made of pure light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the phoenix move in a graceful, dancing motion propelling it in an elegant arc across the sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is there a long, iridescent ribbon of light trailing behind the phoenix composed of orange, purple, and cyan particles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera executing a dynamic, upward-looking tracking shot following the phoenix?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.06998947262763977 + ], + "opposite_logprob_tokens": [ + -2.6949894428253174 + ], + "logprob_reward": -0.06998947262763977, + "yes_no_margin_reward": 2.625, + "token_credit_reward": -0.06998947262763977, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.19353440403938293, + 0.04841609299182892, + 0.15438687801361084, + 0.07633497565984726, + 0.05649247765541077, + 0.0331864058971405, + 0.0766591727733612, + 0.2520606815814972, + 0.06282997131347656, + 0.04609885439276695 + ], + "credit_map": [ + 0.02899831160902977, + 0.03968288004398346, + 0.01590137369930744, + 0.01618603989481926, + 0.042705655097961426, + 0.016583452001214027, + 0.01665307581424713, + 0.016823647543787956, + 0.008407472632825375, + 0.005293759983032942, + 0.005613184068351984, + 0.004785505123436451, + 0.004192717839032412, + 0.009914645925164223, + 0.0056374697014689445, + 0.004571345169097185, + 0.02732306718826294, + 0.016476081684231758, + 0.012322014197707176, + 0.018957961350679398, + 0.00986761786043644, + 0.020183101296424866, + 0.04006273299455643, + 0.00919431447982788, + 0.010226051323115826, + 0.005021862220019102, + 0.022073477506637573, + 0.009057563729584217, + 0.006146928295493126, + 0.007350249215960503, + 0.008955901488661766, + 0.007502949330955744, + 0.005579234100878239, + 0.004994208924472332, + 0.009839258156716824, + 0.007522585801780224, + 0.007899523712694645, + 0.00523348106071353, + 0.0073607442900538445, + 0.008063449524343014, + 0.00307206972502172, + 0.003092232160270214, + 0.005529341753572226, + 0.0038140600081533194, + 0.002561593195423484, + 0.0036821889225393534, + 0.007048687431961298, + 0.00438624108210206, + 0.00811384990811348, + 0.011122344061732292, + 0.01673092506825924, + 0.006620774045586586, + 0.0084862494841218, + 0.008547693490982056, + 0.012435125187039375, + 0.0046022264286875725, + 0.03082393854856491, + 0.0187920480966568, + 0.0648428425192833, + 0.029017556458711624, + 0.03243699669837952, + 0.021961333230137825, + 0.025886325165629387, + 0.028299683704972267, + 0.007726851850748062, + 0.004977508448064327, + 0.012827071361243725, + 0.00879386905580759, + 0.006211285945028067, + 0.00935723539441824, + 0.006581934634596109, + 0.006354218814522028, + 0.0069909896701574326, + 0.0038976764772087336, + 0.007658300921320915, + 0.008338911458849907, + 0.004420445300638676, + 0.0035121487453579903, + 0.005685379728674889, + 0.0055950055830180645 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 0, + 2 + ], + "selected_reward": -0.06998947262763977 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "dynamic, upward-looking tracking shot follows a mythical phoenix soaring through the deep, inky night sky. Enveloped in a luminous, warm golden flame, the phoenix gracefully flickers and dances with its movements. Each feather seems forged from pure light, radiating a fiery, divine glow. Its vast, magnificent wings are fully extended, their edges gleaming like rubies and sapphires in the flames. Its elegant head is held high, its eyes gleaming with wisdom and eternity. Each flap of its wings is graceful yet powerful, propelling it in an elegant arc across a night dotted with countless shimmering stars. Behind it, a long, iridescent ribbon of light, composed of countless shimmering particles of light, trails out, a brilliant rainbow-like trail. This ribbon blends blazing orange, deep purple, and brilliant cyan, slowly dissipating into the darkness. The phoenix itself is the sole source of light in the scene, its golden radiance tinging the surrounding night sky with a warm hue, creating an epic atmosphere of sacredness, grandeur, and eternal vitality.", + "seed": 22246, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000034/group_3/index_3278_seed_22246_dp_7_rank_14_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 3, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.02040095446864143, + "vlm_logprob": -0.02040095446864143, + "vlm_yes_no_margin": 4.774999952316284, + "vlm_token_credit": -0.02040095446864143, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a mythical phoenix with wings fully extended?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the phoenix enveloped in luminous, warm golden flame with feathers appearing to be made of pure light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the phoenix move in a graceful, dancing motion propelling it in an elegant arc across the sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is there a long, iridescent ribbon of light trailing behind the phoenix composed of orange, purple, and cyan particles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera executing a dynamic, upward-looking tracking shot following the phoenix?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005274902563542128 + ], + "opposite_logprob_tokens": [ + -5.380274772644043 + ], + "logprob_reward": -0.005274902563542128, + "yes_no_margin_reward": 5.375, + "token_credit_reward": -0.005274902563542128, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06496191024780273, + 0.04994502663612366, + 0.13986781239509583, + 0.0607340969145298, + 0.1647953987121582, + 0.11936358362436295, + 0.055157896131277084, + 0.09494562447071075, + 0.1164349690079689, + 0.13379371166229248 + ], + "credit_map": [ + 0.01551659032702446, + 0.007553086616098881, + 0.004210253246128559, + 0.004309060052037239, + 0.005354476626962423, + 0.010944098234176636, + 0.00695816520601511, + 0.010116173885762691, + 0.005600729025900364, + 0.0025215616915374994, + 0.006023493129760027, + 0.004691209644079208, + 0.0029930712189525366, + 0.01411901693791151, + 0.002871220000088215, + 0.011124726384878159, + 0.021792342886328697, + 0.01783759519457817, + 0.017994020134210587, + 0.02650967240333557, + 0.013515636324882507, + 0.022238774225115776, + 0.008184735663235188, + 0.011795026250183582, + 0.01292963046580553, + 0.0075543224811553955, + 0.01632457785308361, + 0.007157936692237854, + 0.004007726442068815, + 0.005075688939541578, + 0.0032126386649906635, + 0.00447157071903348, + 0.024241093546152115, + 0.01476225070655346, + 0.033373553305864334, + 0.015518330037593842, + 0.012335670180618763, + 0.016787584871053696, + 0.020836038514971733, + 0.0269408468157053, + 0.018719475716352463, + 0.013820568099617958, + 0.016512829810380936, + 0.012315187603235245, + 0.012073561549186707, + 0.01832534559071064, + 0.009154537692666054, + 0.018442075699567795, + 0.008961991406977177, + 0.006063658744096756, + 0.007693423889577389, + 0.005234791897237301, + 0.0049754902720451355, + 0.009791260585188866, + 0.005074683576822281, + 0.00736259063705802, + 0.010053548961877823, + 0.011530715972185135, + 0.006423224229365587, + 0.0058293514885008335, + 0.0072892773896455765, + 0.016158433631062508, + 0.005897676572203636, + 0.03176339343190193, + 0.012761527672410011, + 0.011197310872375965, + 0.014700074680149555, + 0.007633907254785299, + 0.011919384822249413, + 0.018169941380620003, + 0.0077938842587172985, + 0.032258931547403336, + 0.020869459956884384, + 0.013220801949501038, + 0.018071068450808525, + 0.014140922576189041, + 0.016845600679516792, + 0.01470186561346054, + 0.01243523508310318, + 0.023508742451667786 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 2, + 9 + ], + "selected_reward": -0.005274902563542128 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a mythical phoenix with wings fully extended?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the phoenix enveloped in luminous, warm golden flame with feathers appearing to be made of pure light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the phoenix move in a graceful, dancing motion propelling it in an elegant arc across the sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is there a long, iridescent ribbon of light trailing behind the phoenix composed of orange, purple, and cyan particles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera executing a dynamic, upward-looking tracking shot following the phoenix?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0009261847590096295 + ], + "opposite_logprob_tokens": [ + -7.0009260177612305 + ], + "logprob_reward": -0.0009261847590096295, + "yes_no_margin_reward": 7.0, + "token_credit_reward": -0.0009261847590096295, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09017131477594376, + 0.07766786962747574, + 0.15265333652496338, + 0.04084120690822601, + 0.07083688676357269, + 0.07336696237325668, + 0.1605672687292099, + 0.07896692305803299, + 0.0988178700208664, + 0.15611037611961365 + ], + "credit_map": [ + 0.015980413183569908, + 0.010025068186223507, + 0.007917895913124084, + 0.008560236543416977, + 0.010077123530209064, + 0.015726789832115173, + 0.009261551313102245, + 0.012622232548892498, + 0.009686063975095749, + 0.005694656167179346, + 0.00941410567611456, + 0.006764136254787445, + 0.0057055288925766945, + 0.017039015889167786, + 0.005773180164396763, + 0.01759118214249611, + 0.02599271945655346, + 0.026799572631716728, + 0.011457969434559345, + 0.013034876435995102, + 0.022831471636891365, + 0.026285747066140175, + 0.011659565381705761, + 0.014591416344046593, + 0.006007124669849873, + 0.004350220784544945, + 0.00647325161844492, + 0.004963784944266081, + 0.0032367820385843515, + 0.006092722062021494, + 0.004548888653516769, + 0.005168433766812086, + 0.008261303417384624, + 0.005721675232052803, + 0.011548842303454876, + 0.008703534491360188, + 0.005046416074037552, + 0.008485786616802216, + 0.007722889538854361, + 0.015346442349255085, + 0.010970793664455414, + 0.008284577168524265, + 0.006065261550247669, + 0.009126938879489899, + 0.007303519174456596, + 0.01142185740172863, + 0.006076277699321508, + 0.01411773543804884, + 0.05356781929731369, + 0.007301175035536289, + 0.004608543124049902, + 0.006435638293623924, + 0.0077882250770926476, + 0.06232421472668648, + 0.004888234660029411, + 0.013653422705829144, + 0.00891351792961359, + 0.009600942954421043, + 0.006967587862163782, + 0.004990729503333569, + 0.0073258583433926105, + 0.012786036357283592, + 0.007026957347989082, + 0.021355293691158295, + 0.017015490680933, + 0.007746986113488674, + 0.012978700920939445, + 0.006504791788756847, + 0.006766348145902157, + 0.015122083015739918, + 0.005275346338748932, + 0.027408117428421974, + 0.029440846294164658, + 0.018466891720891, + 0.018735114485025406, + 0.014253142289817333, + 0.01565958373248577, + 0.028248503804206848, + 0.011956739239394665, + 0.019349562004208565 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 9, + 2 + ], + "selected_reward": -0.0009261847590096295 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a mythical phoenix with wings fully extended?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the phoenix enveloped in luminous, warm golden flame with feathers appearing to be made of pure light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the phoenix move in a graceful, dancing motion propelling it in an elegant arc across the sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is there a long, iridescent ribbon of light trailing behind the phoenix composed of orange, purple, and cyan particles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera executing a dynamic, upward-looking tracking shot following the phoenix?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.018182972446084023 + ], + "opposite_logprob_tokens": [ + -4.018182754516602 + ], + "logprob_reward": -0.018182972446084023, + "yes_no_margin_reward": 3.999999761581421, + "token_credit_reward": -0.018182972446084023, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07084818929433823, + 0.08270620554685593, + 0.15678666532039642, + 0.0500161312520504, + 0.1740260124206543, + 0.12831278145313263, + 0.03526844084262848, + 0.11371888965368271, + 0.12389872968196869, + 0.06441794335842133 + ], + "credit_map": [ + 0.011832831427454948, + 0.007846666499972343, + 0.00577924307435751, + 0.004738828632980585, + 0.008452355861663818, + 0.014906947501003742, + 0.007053555455058813, + 0.010237760841846466, + 0.01873631961643696, + 0.004943591542541981, + 0.009837300516664982, + 0.005843665450811386, + 0.004832656122744083, + 0.009679507464170456, + 0.004720616620033979, + 0.024112548679113388, + 0.025105662643909454, + 0.025478282943367958, + 0.013067424297332764, + 0.013550061732530594, + 0.020963657647371292, + 0.03302877023816109, + 0.012881715781986713, + 0.012711076997220516, + 0.005852792877703905, + 0.006806161720305681, + 0.008428719826042652, + 0.006562936119735241, + 0.004446189384907484, + 0.007347092963755131, + 0.004423838108778, + 0.006148400250822306, + 0.02889891341328621, + 0.018761537969112396, + 0.02491547353565693, + 0.020959513261914253, + 0.014965115115046501, + 0.02117878384888172, + 0.016339369118213654, + 0.028007300570607185, + 0.019190235063433647, + 0.013288208283483982, + 0.010936504229903221, + 0.024713128805160522, + 0.01214230339974165, + 0.019884895533323288, + 0.01620319113135338, + 0.011954319663345814, + 0.007465977221727371, + 0.003436296945437789, + 0.002963004633784294, + 0.003978588152676821, + 0.002717715920880437, + 0.007564205676317215, + 0.0028332138899713755, + 0.004309436772018671, + 0.014113456942141056, + 0.010354082100093365, + 0.008958005346357822, + 0.010373569093644619, + 0.009914372116327286, + 0.018950585275888443, + 0.009140878915786743, + 0.03191394358873367, + 0.02098659984767437, + 0.011764083988964558, + 0.016568372026085854, + 0.010553489439189434, + 0.012423444539308548, + 0.016720281913876534, + 0.01198563352227211, + 0.02289682812988758, + 0.011035099625587463, + 0.004508034326136112, + 0.009868871420621872, + 0.00365339289419353, + 0.0044938139617443085, + 0.006287154741585255, + 0.003945628181099892, + 0.020625941455364227 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 2, + 5 + ], + "selected_reward": -0.018182972446084023 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a mythical phoenix with wings fully extended?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the phoenix enveloped in luminous, warm golden flame with feathers appearing to be made of pure light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the phoenix move in a graceful, dancing motion propelling it in an elegant arc across the sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is there a long, iridescent ribbon of light trailing behind the phoenix composed of orange, purple, and cyan particles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera executing a dynamic, upward-looking tracking shot following the phoenix?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.007622679229825735 + ], + "opposite_logprob_tokens": [ + -4.882622718811035 + ], + "logprob_reward": -0.007622679229825735, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.007622679229825735, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06983921676874161, + 0.07876480370759964, + 0.07669050991535187, + 0.15606386959552765, + 0.0901646688580513, + 0.06454088538885117, + 0.02942286804318428, + 0.17760993540287018, + 0.20129899680614471, + 0.055604204535484314 + ], + "credit_map": [ + 0.01074225828051567, + 0.006252769846469164, + 0.006441697012633085, + 0.0054778167977929115, + 0.006888920906931162, + 0.015514337457716465, + 0.008555743843317032, + 0.009965679608285427, + 0.008245281875133514, + 0.004580455366522074, + 0.00648146728053689, + 0.005242714658379555, + 0.004914165008813143, + 0.02640283852815628, + 0.006705448962748051, + 0.016192438080906868, + 0.010336535051465034, + 0.009996001608669758, + 0.007964135147631168, + 0.006611701101064682, + 0.008298313245177269, + 0.019086139276623726, + 0.007073601242154837, + 0.007324094884097576, + 0.015637533739209175, + 0.01639370061457157, + 0.01953888311982155, + 0.018481997773051262, + 0.016560973599553108, + 0.031288065016269684, + 0.017889490351080894, + 0.020273251459002495, + 0.009957723319530487, + 0.006380927748978138, + 0.011448536068201065, + 0.009032726287841797, + 0.007932646200060844, + 0.01299954205751419, + 0.010088442824780941, + 0.02232413925230503, + 0.009810966439545155, + 0.007539426442235708, + 0.005207729060202837, + 0.009367807768285275, + 0.0067992969416081905, + 0.00882013887166977, + 0.007758771535009146, + 0.009236753918230534, + 0.00642678327858448, + 0.0025580115616321564, + 0.0020465829875320196, + 0.00202766852453351, + 0.0025383613537997007, + 0.006863386835902929, + 0.0019218095112591982, + 0.0050402674823999405, + 0.016502313315868378, + 0.012330607511103153, + 0.010553649626672268, + 0.009318253956735134, + 0.010028236545622349, + 0.03475584834814072, + 0.013146761804819107, + 0.07097429037094116, + 0.027810581028461456, + 0.019086172804236412, + 0.025669537484645844, + 0.016252592206001282, + 0.019200749695301056, + 0.029010849073529243, + 0.01861741952598095, + 0.04565112292766571, + 0.010750854387879372, + 0.0040142531506717205, + 0.006806815974414349, + 0.0034752734936773777, + 0.00548554677516222, + 0.012988122180104256, + 0.004254526924341917, + 0.00782882608473301 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 7, + 3 + ], + "selected_reward": -0.007622679229825735 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a mythical phoenix with wings fully extended?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the phoenix enveloped in luminous, warm golden flame with feathers appearing to be made of pure light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the phoenix move in a graceful, dancing motion propelling it in an elegant arc across the sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is there a long, iridescent ribbon of light trailing behind the phoenix composed of orange, purple, and cyan particles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera executing a dynamic, upward-looking tracking shot following the phoenix?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.06999803334474564 + ], + "opposite_logprob_tokens": [ + -2.694998025894165 + ], + "logprob_reward": -0.06999803334474564, + "yes_no_margin_reward": 2.625, + "token_credit_reward": -0.06999803334474564, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0949554443359375, + 0.12324468791484833, + 0.131487175822258, + 0.06603721529245377, + 0.13904400169849396, + 0.0857749953866005, + 0.065297432243824, + 0.16443981230258942, + 0.04891294240951538, + 0.08080623298883438 + ], + "credit_map": [ + 0.024317186325788498, + 0.007117066998034716, + 0.006444860249757767, + 0.004899855237454176, + 0.009689239785075188, + 0.0200314000248909, + 0.006874378304928541, + 0.015581455081701279, + 0.027710039168596268, + 0.007028529420495033, + 0.015190285630524158, + 0.00986634474247694, + 0.007647303398698568, + 0.01369162742048502, + 0.007349491119384766, + 0.03476107120513916, + 0.027532678097486496, + 0.021233828738331795, + 0.010739962570369244, + 0.010094721801578999, + 0.01698964647948742, + 0.026350758969783783, + 0.008548766374588013, + 0.009996808134019375, + 0.008979965932667255, + 0.009317178279161453, + 0.009707696735858917, + 0.007453863974660635, + 0.005966613534837961, + 0.011220662854611874, + 0.006124656181782484, + 0.00726658059284091, + 0.016808249056339264, + 0.010240372270345688, + 0.020053472369909286, + 0.017680389806628227, + 0.008898557163774967, + 0.018605567514896393, + 0.015527711249887943, + 0.031229693442583084, + 0.013239927589893341, + 0.0075033376924693584, + 0.00970715656876564, + 0.00853085145354271, + 0.007342944387346506, + 0.010559671558439732, + 0.008709879592061043, + 0.020181231200695038, + 0.020634276792407036, + 0.003134682774543762, + 0.002837009960785508, + 0.003406044328585267, + 0.003235127544030547, + 0.024350307881832123, + 0.002238075016066432, + 0.005461912602186203, + 0.015741216018795967, + 0.011601981706917286, + 0.00954370852559805, + 0.008683037012815475, + 0.009876525029540062, + 0.03260986506938934, + 0.011051458306610584, + 0.06533200293779373, + 0.005316735710948706, + 0.005419264547526836, + 0.006697612348943949, + 0.0039859553799033165, + 0.005884633399546146, + 0.006910905707627535, + 0.004238339141011238, + 0.010459496639668941, + 0.011958606541156769, + 0.007358850911259651, + 0.013048237189650536, + 0.004685512278228998, + 0.006544759962707758, + 0.011039551347494125, + 0.004667137283831835, + 0.021503573283553123 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 4, + 2 + ], + "selected_reward": -0.06999803334474564 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "dynamic, upward-looking tracking shot follows a mythical phoenix soaring through the deep, inky night sky. Enveloped in a luminous, warm golden flame, the phoenix gracefully flickers and dances with its movements. Each feather seems forged from pure light, radiating a fiery, divine glow. Its vast, magnificent wings are fully extended, their edges gleaming like rubies and sapphires in the flames. Its elegant head is held high, its eyes gleaming with wisdom and eternity. Each flap of its wings is graceful yet powerful, propelling it in an elegant arc across a night dotted with countless shimmering stars. Behind it, a long, iridescent ribbon of light, composed of countless shimmering particles of light, trails out, a brilliant rainbow-like trail. This ribbon blends blazing orange, deep purple, and brilliant cyan, slowly dissipating into the darkness. The phoenix itself is the sole source of light in the scene, its golden radiance tinging the surrounding night sky with a warm hue, creating an epic atmosphere of sacredness, grandeur, and eternal vitality.", + "seed": 22247, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000034/group_3/index_3278_seed_22247_dp_7_rank_14_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 3, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.02926926324144006, + "vlm_logprob": -0.02926926324144006, + "vlm_yes_no_margin": 4.7, + "vlm_token_credit": -0.02926926324144006, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a mythical phoenix with wings fully extended?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the phoenix enveloped in luminous, warm golden flame with feathers appearing to be made of pure light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the phoenix move in a graceful, dancing motion propelling it in an elegant arc across the sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is there a long, iridescent ribbon of light trailing behind the phoenix composed of orange, purple, and cyan particles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera executing a dynamic, upward-looking tracking shot following the phoenix?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0038265592884272337 + ], + "opposite_logprob_tokens": [ + -6.00382661819458 + ], + "logprob_reward": -0.0038265592884272337, + "yes_no_margin_reward": 6.0, + "token_credit_reward": -0.0038265592884272337, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11080313473939896, + 0.13000375032424927, + 0.255374550819397, + 0.059670135378837585, + 0.1612488180398941, + 0.04847758263349533, + 0.054713305085897446, + 0.08825734257698059, + 0.04276575520634651, + 0.04868566617369652 + ], + "credit_map": [ + 0.013670993968844414, + 0.011832831427454948, + 0.016124049201607704, + 0.012254170142114162, + 0.015604797750711441, + 0.012318438850343227, + 0.009312380105257034, + 0.019685452803969383, + 0.015248902142047882, + 0.015599476173520088, + 0.02504449151456356, + 0.016139226034283638, + 0.015655837953090668, + 0.020851360633969307, + 0.011353947222232819, + 0.010110506787896156, + 0.012031927704811096, + 0.008168404921889305, + 0.0818856880068779, + 0.014573216438293457, + 0.00923536904156208, + 0.1067003533244133, + 0.0068227737210690975, + 0.0159568153321743, + 0.0065068951807916164, + 0.013496535830199718, + 0.00415535643696785, + 0.004664892330765724, + 0.009355306625366211, + 0.007475681137293577, + 0.0067520588636398315, + 0.0072634038515388966, + 0.048387229442596436, + 0.008124439045786858, + 0.06019929051399231, + 0.012612364254891872, + 0.00854887068271637, + 0.006998568773269653, + 0.007790559437125921, + 0.008587485179305077, + 0.007490812335163355, + 0.009336885996162891, + 0.004312261939048767, + 0.004973781760782003, + 0.004998786374926567, + 0.0052746995352208614, + 0.005909330677241087, + 0.006181016564369202, + 0.007434864062815905, + 0.009472541511058807, + 0.006320375483483076, + 0.005521723534911871, + 0.007380202412605286, + 0.005767410155385733, + 0.007723379880189896, + 0.0050928001292049885, + 0.03408341482281685, + 0.006541067268699408, + 0.007691582199186087, + 0.005108288489282131, + 0.00521603599190712, + 0.0057685887441039085, + 0.0063853817991912365, + 0.017462966963648796, + 0.005303603131324053, + 0.0058780936524271965, + 0.006046343594789505, + 0.0038897257763892412, + 0.004823960363864899, + 0.006811497267335653, + 0.0048620933666825294, + 0.005150432698428631, + 0.009139562025666237, + 0.006242099218070507, + 0.005926403217017651, + 0.006540960632264614, + 0.004420202691107988, + 0.005266799591481686, + 0.0060544926673173904, + 0.005095144268125296 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 4, + 1 + ], + "selected_reward": -0.0038265592884272337 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a mythical phoenix with wings fully extended?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the phoenix enveloped in luminous, warm golden flame with feathers appearing to be made of pure light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the phoenix move in a graceful, dancing motion propelling it in an elegant arc across the sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is there a long, iridescent ribbon of light trailing behind the phoenix composed of orange, purple, and cyan particles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera executing a dynamic, upward-looking tracking shot following the phoenix?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.000929281348362565 + ], + "opposite_logprob_tokens": [ + -7.000929355621338 + ], + "logprob_reward": -0.000929281348362565, + "yes_no_margin_reward": 7.0, + "token_credit_reward": -0.000929281348362565, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09333980083465576, + 0.03124600276350975, + 0.15133647620677948, + 0.05740547552704811, + 0.37627673149108887, + 0.06697072088718414, + 0.05468784272670746, + 0.04408242553472519, + 0.06368879228830338, + 0.06096566841006279 + ], + "credit_map": [ + 0.010464323684573174, + 0.012040083296597004, + 0.012492499314248562, + 0.01159662939608097, + 0.012050619348883629, + 0.008735056035220623, + 0.012732402421534061, + 0.013228186406195164, + 0.005918022245168686, + 0.003321903757750988, + 0.005436664447188377, + 0.003209750633686781, + 0.002973308088257909, + 0.0038740329910069704, + 0.002705749124288559, + 0.003806570079177618, + 0.007920674979686737, + 0.006946663837879896, + 0.04411754384636879, + 0.008745170198380947, + 0.008388183079659939, + 0.058423422276973724, + 0.006190088577568531, + 0.010604729875922203, + 0.005188449751585722, + 0.010100282728672028, + 0.004459866788238287, + 0.005169489420950413, + 0.00755061162635684, + 0.008192562498152256, + 0.010736183263361454, + 0.006008028518408537, + 0.11940938234329224, + 0.015963712707161903, + 0.15019364655017853, + 0.028471680358052254, + 0.01996966078877449, + 0.012135494500398636, + 0.012934583239257336, + 0.017198586836457253, + 0.010088920593261719, + 0.008742637000977993, + 0.00715071614831686, + 0.009543882682919502, + 0.007523368112742901, + 0.006104565225541592, + 0.00960482656955719, + 0.008211805485188961, + 0.008073942735791206, + 0.007303610444068909, + 0.004802089184522629, + 0.005148651078343391, + 0.0072546424344182014, + 0.006022799294441938, + 0.00658273883163929, + 0.009499368257820606, + 0.011505604721605778, + 0.004063752014189959, + 0.005669618025422096, + 0.0031823385506868362, + 0.00417530769482255, + 0.004467997699975967, + 0.004263462498784065, + 0.006754344794899225, + 0.0089065907523036, + 0.0066702160984277725, + 0.007412922568619251, + 0.005993175785988569, + 0.005332284606993198, + 0.01222944911569357, + 0.007320076692849398, + 0.00982408132404089, + 0.013808641582727432, + 0.007468254305422306, + 0.005854061339050531, + 0.005191412754356861, + 0.004733988083899021, + 0.007969465106725693, + 0.005173496901988983, + 0.010766341350972652 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 2, + 0 + ], + "selected_reward": -0.000929281348362565 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a mythical phoenix with wings fully extended?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the phoenix enveloped in luminous, warm golden flame with feathers appearing to be made of pure light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the phoenix move in a graceful, dancing motion propelling it in an elegant arc across the sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is there a long, iridescent ribbon of light trailing behind the phoenix composed of orange, purple, and cyan particles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera executing a dynamic, upward-looking tracking shot following the phoenix?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03369210660457611 + ], + "opposite_logprob_tokens": [ + -3.4086921215057373 + ], + "logprob_reward": -0.03369210660457611, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.03369210660457611, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.038070738315582275, + 0.03787780925631523, + 0.09543197602033615, + 0.07383553683757782, + 0.6013022661209106, + 0.04595088213682175, + 0.01748061552643776, + 0.03160594403743744, + 0.0369940884411335, + 0.02145010605454445 + ], + "credit_map": [ + 0.004557352978736162, + 0.008454729802906513, + 0.004718375857919455, + 0.004239929839968681, + 0.004494938068091869, + 0.004065630026161671, + 0.0036929724738001823, + 0.00384681043215096, + 0.005361615680158138, + 0.00214185519143939, + 0.010477644391357899, + 0.0029907990247011185, + 0.0017308199312537909, + 0.010173714719712734, + 0.0014389986172318459, + 0.0035623630974441767, + 0.005153954960405827, + 0.0049695284105837345, + 0.026880236342549324, + 0.005420447327196598, + 0.005685117561370134, + 0.03571905195713043, + 0.003977659624069929, + 0.0076259770430624485, + 0.008852296508848667, + 0.018013525754213333, + 0.005164078436791897, + 0.005704060662537813, + 0.01233295165002346, + 0.008436193689703941, + 0.007009494584053755, + 0.008322937414050102, + 0.18981271982192993, + 0.025676866993308067, + 0.2442166656255722, + 0.04651958495378494, + 0.029459862038493156, + 0.015271594747900963, + 0.021931136026978493, + 0.028413835912942886, + 0.006610654294490814, + 0.01123175397515297, + 0.004206934943795204, + 0.005661387462168932, + 0.00456482358276844, + 0.004200028255581856, + 0.005803738255053759, + 0.003671560436487198, + 0.0026224758476018906, + 0.002914152108132839, + 0.001584660611115396, + 0.0015210045967251062, + 0.001819016644731164, + 0.0018450445495545864, + 0.0023081661202013493, + 0.0028660946991294622, + 0.011013058014214039, + 0.0022853075060993433, + 0.0027288473211228848, + 0.00223726243712008, + 0.0017497928347438574, + 0.0025412654504179955, + 0.0024472372606396675, + 0.006603172980248928, + 0.004949615336954594, + 0.003962887451052666, + 0.005156695377081633, + 0.003288149368017912, + 0.0029245621990412474, + 0.006590156815946102, + 0.00421288562938571, + 0.005909136030822992, + 0.002970925997942686, + 0.003428111784160137, + 0.0023668124340474606, + 0.003104210365563631, + 0.0021748621948063374, + 0.0024549230001866817, + 0.002667357213795185, + 0.00228290306404233 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 2, + 3 + ], + "selected_reward": -0.03369210660457611 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a mythical phoenix with wings fully extended?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the phoenix enveloped in luminous, warm golden flame with feathers appearing to be made of pure light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the phoenix move in a graceful, dancing motion propelling it in an elegant arc across the sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is there a long, iridescent ribbon of light trailing behind the phoenix composed of orange, purple, and cyan particles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera executing a dynamic, upward-looking tracking shot following the phoenix?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.007622442673891783 + ], + "opposite_logprob_tokens": [ + -4.882622241973877 + ], + "logprob_reward": -0.007622442673891783, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.007622442673891783, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.17436939477920532, + 0.04473651945590973, + 0.1411242038011551, + 0.08155415207147598, + 0.34486454725265503, + 0.09624304622411728, + 0.02857358753681183, + 0.03943219035863876, + 0.027295179665088654, + 0.021807165816426277 + ], + "credit_map": [ + 0.017726052552461624, + 0.02126171439886093, + 0.021216003224253654, + 0.01984564960002899, + 0.022871164605021477, + 0.016012756153941154, + 0.03007415682077408, + 0.025361888110637665, + 0.004249901045113802, + 0.0025708472821861506, + 0.014421455562114716, + 0.002750395331531763, + 0.0032075021881610155, + 0.01346104871481657, + 0.0015343207633122802, + 0.0025410479865968227, + 0.006294646766036749, + 0.0037436699494719505, + 0.04717608541250229, + 0.007973987609148026, + 0.004415668081492186, + 0.059722650796175, + 0.003297831630334258, + 0.008499665185809135, + 0.011097212322056293, + 0.018348822370171547, + 0.006732183042913675, + 0.006699081975966692, + 0.011792538687586784, + 0.009978665970265865, + 0.00839146040380001, + 0.008514187298715115, + 0.1093907356262207, + 0.014439423568546772, + 0.13856521248817444, + 0.026313992217183113, + 0.016957946121692657, + 0.009434636682271957, + 0.01380067691206932, + 0.015961917117238045, + 0.012404036708176136, + 0.009060204960405827, + 0.011049225926399231, + 0.015038909390568733, + 0.008943426422774792, + 0.009267953224480152, + 0.012721291743218899, + 0.017757998779416084, + 0.004816144704818726, + 0.004022915381938219, + 0.0021925929468125105, + 0.0025137807242572308, + 0.003050412517040968, + 0.0040336656384170055, + 0.0038666920736432076, + 0.004077384248375893, + 0.013490729965269566, + 0.0031540978234261274, + 0.003305003046989441, + 0.0020995154045522213, + 0.0021580674219876528, + 0.0029519263189285994, + 0.002834897954016924, + 0.009437951259315014, + 0.0036241873167455196, + 0.0028146214317530394, + 0.004319703672081232, + 0.002249173354357481, + 0.0019246443407610059, + 0.00529470993205905, + 0.002905769506469369, + 0.004162370227277279, + 0.003148391144350171, + 0.0027075724210590124, + 0.0016593091422691941, + 0.002773515647277236, + 0.0018289197469130158, + 0.0030072396621108055, + 0.0023713894188404083, + 0.004310828633606434 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 0, + 2 + ], + "selected_reward": -0.007622442673891783 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a mythical phoenix with wings fully extended?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the phoenix enveloped in luminous, warm golden flame with feathers appearing to be made of pure light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the phoenix move in a graceful, dancing motion propelling it in an elegant arc across the sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is there a long, iridescent ribbon of light trailing behind the phoenix composed of orange, purple, and cyan particles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera executing a dynamic, upward-looking tracking shot following the phoenix?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.1002759262919426 + ], + "opposite_logprob_tokens": [ + -2.350275993347168 + ], + "logprob_reward": -0.1002759262919426, + "yes_no_margin_reward": 2.25, + "token_credit_reward": -0.1002759262919426, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0898214802145958, + 0.09449731558561325, + 0.17338255047798157, + 0.11084277927875519, + 0.15419034659862518, + 0.10157351195812225, + 0.036445412784814835, + 0.07213037461042404, + 0.14075274765491486, + 0.026363400742411613 + ], + "credit_map": [ + 0.010623224079608917, + 0.012582090683281422, + 0.011429300531744957, + 0.010987886227667332, + 0.010268721729516983, + 0.009282312355935574, + 0.009591243229806423, + 0.015056712552905083, + 0.012533481232821941, + 0.010920489206910133, + 0.01640772633254528, + 0.010047037154436111, + 0.00946453120559454, + 0.01782245934009552, + 0.00956773106008768, + 0.007733876816928387, + 0.009866133332252502, + 0.012613550759851933, + 0.03965720906853676, + 0.010711515322327614, + 0.013098636642098427, + 0.049505919218063354, + 0.010591340251266956, + 0.027338264510035515, + 0.012322594411671162, + 0.020292581990361214, + 0.00997137650847435, + 0.0089322654530406, + 0.017056455835700035, + 0.012867044657468796, + 0.013805008493363857, + 0.015595465898513794, + 0.04367896541953087, + 0.010682595893740654, + 0.052509669214487076, + 0.012133050709962845, + 0.008896883577108383, + 0.008768637664616108, + 0.008462932892143726, + 0.00905761681497097, + 0.013209247961640358, + 0.022812264040112495, + 0.009741128422319889, + 0.012907241471111774, + 0.013105835765600204, + 0.008593483828008175, + 0.01153749879449606, + 0.009666825644671917, + 0.006942181382328272, + 0.005458396393805742, + 0.003914692439138889, + 0.0026008221320807934, + 0.0031305348966270685, + 0.005578265525400639, + 0.0029948751907795668, + 0.0058256457559764385, + 0.02640203759074211, + 0.005056702997535467, + 0.006545230280607939, + 0.004572021774947643, + 0.004101640544831753, + 0.0052961623296141624, + 0.00536947650834918, + 0.014787109568715096, + 0.019588634371757507, + 0.014153171330690384, + 0.01751318760216236, + 0.013873427174985409, + 0.011313799768686295, + 0.025171872228384018, + 0.016161294654011726, + 0.022977370768785477, + 0.004467076621949673, + 0.0045030065812170506, + 0.001989833079278469, + 0.0027528759092092514, + 0.00271457526832819, + 0.0026616323739290237, + 0.0033474494703114033, + 0.003926953300833702 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 4, + 8 + ], + "selected_reward": -0.1002759262919426 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000034/group_4/index_tensor([8349, 8349])_group_4_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000034/group_4/index_tensor([8349, 8349])_group_4_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..c749e607671cd0b0d9f868c509b6a9966ae53274 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000034/group_4/index_tensor([8349, 8349])_group_4_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "dynamic, low-angle tracking shot follows a massive main battle tank as it hurtles forward at breakneck speed across a vast, rain-soaked, muddy field. Under a gloomy gray sky, the war machine, painted in a deep, matte olive green camouflage, churns its broad, sturdy metal tracks furiously, each rut digging deep into the ground and violently throwing huge chunks of sticky, dark brown mud and murky water high and low to the rear and sides in a spectacular, chaotic curtain of mud. The heavy vehicle heaves and pitches violently across the uneven terrain, its massive turret pointing steadily forward, its long barrel swaying slightly as the vehicle sways. Soft, diffused lighting accentuates the cool texture of the tank's metal hull and the wet sheen of the splattered mud. The entire scene, shrouded in a chilling hue, is imbued with a sense of raw power, tension, and unstoppable momentum.", + "seed": 15330, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000034/group_4/index_8349_seed_15330_dp_8_rank_16_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 4, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.05953494133427739, + "vlm_logprob": -0.05953494133427739, + "vlm_yes_no_margin": 3.875, + "vlm_token_credit": -0.05953494133427739, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a massive battle tank painted in deep matte olive green camouflage?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the tank moving at high speed across a vast, rain-soaked, muddy field?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are huge chunks of dark brown mud and murky water being thrown high and low by the tank's tracks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the sky gloomy and gray with soft, diffused lighting?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a dynamic low-angle tracking shot following the tank from behind or the side?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03615424782037735 + ], + "opposite_logprob_tokens": [ + -3.661154270172119 + ], + "logprob_reward": -0.03615424782037735, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.03615424782037735, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10961194336414337, + 0.09792964905500412, + 0.05072690173983574, + 0.034865498542785645, + 0.03491382300853729, + 0.07192076742649078, + 0.04227714240550995, + 0.20143567025661469, + 0.0465337410569191, + 0.3097849488258362 + ], + "credit_map": [ + 0.01686881296336651, + 0.018066255375742912, + 0.014693180099129677, + 0.01012018509209156, + 0.02010400779545307, + 0.011386329308152199, + 0.0077290297485888, + 0.010644142515957355, + 0.01627608761191368, + 0.0201155636459589, + 0.008093991316854954, + 0.012357635423541069, + 0.011835908517241478, + 0.013810820877552032, + 0.006571665406227112, + 0.008867975324392319, + 0.006674397736787796, + 0.008239970542490482, + 0.006291156634688377, + 0.009245029650628567, + 0.005002626217901707, + 0.005048392806202173, + 0.005842382553964853, + 0.004382943734526634, + 0.004597769118845463, + 0.0038006568793207407, + 0.004794109147042036, + 0.003657520515844226, + 0.00483194412663579, + 0.0032232736703008413, + 0.004973709117621183, + 0.0049865152686834335, + 0.004374702461063862, + 0.004309594631195068, + 0.004603558219969273, + 0.005595523398369551, + 0.004259715788066387, + 0.004534043371677399, + 0.0035287898499518633, + 0.003707894589751959, + 0.008233415894210339, + 0.005420187953859568, + 0.016303040087223053, + 0.00712995370849967, + 0.013787337578833103, + 0.0073808846063911915, + 0.0059571657329797745, + 0.007708780467510223, + 0.0037472371477633715, + 0.003511617425829172, + 0.006672922987490892, + 0.004223763011395931, + 0.004950408823788166, + 0.004665665794163942, + 0.0058798883110284805, + 0.008625635877251625, + 0.04232972860336304, + 0.030178876593708992, + 0.02004246972501278, + 0.015569932758808136, + 0.027454884722828865, + 0.022384807467460632, + 0.021308207884430885, + 0.022166751325130463, + 0.0053731235675513744, + 0.005842674057930708, + 0.007361955009400845, + 0.0073753041215240955, + 0.004898513667285442, + 0.0038583120331168175, + 0.007669821381568909, + 0.004154040012508631, + 0.04230383411049843, + 0.030536796897649765, + 0.0335189551115036, + 0.04141221195459366, + 0.058594878762960434, + 0.05039351060986519, + 0.03324366733431816, + 0.019781077280640602 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 7, + 0 + ], + "selected_reward": -0.03615424782037735 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a massive battle tank painted in deep matte olive green camouflage?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the tank moving at high speed across a vast, rain-soaked, muddy field?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are huge chunks of dark brown mud and murky water being thrown high and low by the tank's tracks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the sky gloomy and gray with soft, diffused lighting?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a dynamic low-angle tracking shot following the tank from behind or the side?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.22560158371925354 + ], + "opposite_logprob_tokens": [ + -1.6006015539169312 + ], + "logprob_reward": -0.22560158371925354, + "yes_no_margin_reward": 1.375, + "token_credit_reward": -0.22560158371925354, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3693387806415558, + 0.123113714158535, + 0.07145199179649353, + 0.05906078964471817, + 0.10826294869184494, + 0.04901259019970894, + 0.04167698696255684, + 0.06198912486433983, + 0.050779204815626144, + 0.0653138980269432 + ], + "credit_map": [ + 0.05240380018949509, + 0.03929422050714493, + 0.0734972432255745, + 0.04423587769269943, + 0.04376186802983284, + 0.047322217375040054, + 0.03455071896314621, + 0.034272801131010056, + 0.01783006638288498, + 0.014460943639278412, + 0.013299672864377499, + 0.01757546328008175, + 0.015641380101442337, + 0.01900414004921913, + 0.01225681509822607, + 0.013045229949057102, + 0.008906296454370022, + 0.010823510587215424, + 0.009526267647743225, + 0.013447756879031658, + 0.006807471625506878, + 0.008749575354158878, + 0.00724664656445384, + 0.00594447273761034, + 0.0071099912747740746, + 0.006025560665875673, + 0.01098711509257555, + 0.008415387943387032, + 0.005306430626660585, + 0.0057053943164646626, + 0.0076181055046617985, + 0.00789280328899622, + 0.012325439602136612, + 0.015110323205590248, + 0.0205211341381073, + 0.02109513245522976, + 0.007698941510170698, + 0.0073646605014801025, + 0.011897217482328415, + 0.012250088155269623, + 0.00525125116109848, + 0.0035741408355534077, + 0.014453270472586155, + 0.00533613096922636, + 0.005277976859360933, + 0.006579666864126921, + 0.0037403854075819254, + 0.004799768328666687, + 0.00416335416957736, + 0.005841802805662155, + 0.00585858803242445, + 0.006163681857287884, + 0.003694770857691765, + 0.005116211250424385, + 0.0039760153740644455, + 0.006862559821456671, + 0.007964919321238995, + 0.0077770729549229145, + 0.008203096687793732, + 0.007485042791813612, + 0.009056376293301582, + 0.006214211229234934, + 0.006821849849075079, + 0.00846655759960413, + 0.007620984688401222, + 0.006620363797992468, + 0.007847845554351807, + 0.008598348125815392, + 0.003653008257970214, + 0.003705905517563224, + 0.00778208440169692, + 0.004950664937496185, + 0.00457640690729022, + 0.006665600463747978, + 0.009350824169814587, + 0.013331948779523373, + 0.011329133994877338, + 0.006524448283016682, + 0.008812240324914455, + 0.004723289981484413 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 4 + ], + "selected_reward": -0.22560158371925354 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a massive battle tank painted in deep matte olive green camouflage?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the tank moving at high speed across a vast, rain-soaked, muddy field?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are huge chunks of dark brown mud and murky water being thrown high and low by the tank's tracks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the sky gloomy and gray with soft, diffused lighting?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a dynamic low-angle tracking shot following the tank from behind or the side?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0036266762763261795 + ], + "opposite_logprob_tokens": [ + -5.628626823425293 + ], + "logprob_reward": -0.0036266762763261795, + "yes_no_margin_reward": 5.625, + "token_credit_reward": -0.0036266762763261795, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2873802185058594, + 0.0683324858546257, + 0.05489374324679375, + 0.10987638682126999, + 0.11942058056592941, + 0.038846537470817566, + 0.047325339168310165, + 0.1571742445230484, + 0.044263068586587906, + 0.07248741388320923 + ], + "credit_map": [ + 0.04299331083893776, + 0.03680983930826187, + 0.05438584089279175, + 0.03347926214337349, + 0.033033497631549835, + 0.03315419703722, + 0.02465173415839672, + 0.0288725383579731, + 0.010328056290745735, + 0.013609311543405056, + 0.00882803089916706, + 0.008184813894331455, + 0.007948559708893299, + 0.007438108790665865, + 0.005565179046243429, + 0.006430424749851227, + 0.006427365820854902, + 0.00533765135332942, + 0.008195580914616585, + 0.00760970264673233, + 0.005723194684833288, + 0.006854783743619919, + 0.009103523567318916, + 0.005641941912472248, + 0.013694092631340027, + 0.010031242854893208, + 0.02819805033504963, + 0.01181115210056305, + 0.010827888734638691, + 0.010653802193701267, + 0.011308841407299042, + 0.013351316563785076, + 0.01568831130862236, + 0.0158219076693058, + 0.020966336131095886, + 0.02400106191635132, + 0.011884715408086777, + 0.01069485954940319, + 0.010409815236926079, + 0.009953573346138, + 0.00505205150693655, + 0.0035240259021520615, + 0.00986046064645052, + 0.004138079471886158, + 0.006124467588961124, + 0.004126585554331541, + 0.002975243842229247, + 0.0030456206295639277, + 0.004893203731626272, + 0.00680223572999239, + 0.009140449576079845, + 0.005192673299461603, + 0.00340333697386086, + 0.005708798766136169, + 0.004746171645820141, + 0.00743846595287323, + 0.010092693381011486, + 0.016099048778414726, + 0.018059343099594116, + 0.015633001923561096, + 0.0290286336094141, + 0.01891290955245495, + 0.021013837307691574, + 0.028334779664874077, + 0.004611223470419645, + 0.004853654187172651, + 0.008149456232786179, + 0.007314651273190975, + 0.0023995398078113794, + 0.003149963216856122, + 0.0068558878265321255, + 0.006928692571818829, + 0.009836004115641117, + 0.004895887337625027, + 0.008348546922206879, + 0.012641177512705326, + 0.010190262459218502, + 0.012569138780236244, + 0.009393349289894104, + 0.004613043740391731 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 7, + 4 + ], + "selected_reward": -0.0036266762763261795 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a massive battle tank painted in deep matte olive green camouflage?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the tank moving at high speed across a vast, rain-soaked, muddy field?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are huge chunks of dark brown mud and murky water being thrown high and low by the tank's tracks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the sky gloomy and gray with soft, diffused lighting?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a dynamic low-angle tracking shot following the tank from behind or the side?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005953317973762751 + ], + "opposite_logprob_tokens": [ + -5.130953311920166 + ], + "logprob_reward": -0.005953317973762751, + "yes_no_margin_reward": 5.125, + "token_credit_reward": -0.005953317973762751, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.17211319506168365, + 0.08466826379299164, + 0.04161473736166954, + 0.02709376811981201, + 0.42512819170951843, + 0.08131591230630875, + 0.030241426080465317, + 0.04684535786509514, + 0.0151728680357337, + 0.07580626755952835 + ], + "credit_map": [ + 0.034139130264520645, + 0.047411926090717316, + 0.01931915245950222, + 0.01231112889945507, + 0.018537847325205803, + 0.010837910696864128, + 0.011386832222342491, + 0.01816927269101143, + 0.01756991259753704, + 0.02404223009943962, + 0.007585757412016392, + 0.008175436407327652, + 0.00898436363786459, + 0.007854729890823364, + 0.004499581176787615, + 0.005956254433840513, + 0.0045853531919419765, + 0.0027963188476860523, + 0.0064051588997244835, + 0.006596902851015329, + 0.0039278301410377026, + 0.004248185083270073, + 0.008435066789388657, + 0.004619921091943979, + 0.004025216214358807, + 0.0030254304874688387, + 0.006778606679290533, + 0.0028453213162720203, + 0.002680985489860177, + 0.0022704743314534426, + 0.002795217093080282, + 0.0026725181378424168, + 0.06276969611644745, + 0.09410035610198975, + 0.06105339154601097, + 0.07921501994132996, + 0.042511213570833206, + 0.029293734580278397, + 0.03161279857158661, + 0.024571968242526054, + 0.012488552369177341, + 0.006854103412479162, + 0.019971363246440887, + 0.010935178026556969, + 0.006502275355160236, + 0.009302658960223198, + 0.007850142195820808, + 0.00741164106875658, + 0.002403463702648878, + 0.002720411866903305, + 0.005137132480740547, + 0.004405415151268244, + 0.0025758002884685993, + 0.0032158587127923965, + 0.003594842739403248, + 0.0061885034665465355, + 0.009165270254015923, + 0.006041158922016621, + 0.005807247944176197, + 0.005543950013816357, + 0.0046456726267933846, + 0.00479816272854805, + 0.005400564521551132, + 0.0054433299228549, + 0.001808261382393539, + 0.001935756765305996, + 0.002632958348840475, + 0.002362669911235571, + 0.0010622360277920961, + 0.0014464457053691149, + 0.002064625732600689, + 0.00185991486068815, + 0.01041910145431757, + 0.004576351493597031, + 0.008829265832901001, + 0.017273547127842903, + 0.00627949507907033, + 0.01537338737398386, + 0.008892152458429337, + 0.004162956960499287 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 0, + 1 + ], + "selected_reward": -0.005953317973762751 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a massive battle tank painted in deep matte olive green camouflage?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the tank moving at high speed across a vast, rain-soaked, muddy field?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are huge chunks of dark brown mud and murky water being thrown high and low by the tank's tracks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the sky gloomy and gray with soft, diffused lighting?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a dynamic low-angle tracking shot following the tank from behind or the side?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.026338880881667137 + ], + "opposite_logprob_tokens": [ + -3.651338815689087 + ], + "logprob_reward": -0.026338880881667137, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.026338880881667137, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.17455869913101196, + 0.06479363888502121, + 0.07975210249423981, + 0.14736947417259216, + 0.0979938730597496, + 0.10661391913890839, + 0.06489474326372147, + 0.10994277894496918, + 0.07216750085353851, + 0.08191322535276413 + ], + "credit_map": [ + 0.02510734647512436, + 0.023681743070483208, + 0.034200090914964676, + 0.019474811851978302, + 0.018477987498044968, + 0.015579935163259506, + 0.01740988902747631, + 0.02062690258026123, + 0.010727090761065483, + 0.014868536964058876, + 0.006091089453548193, + 0.007355951238423586, + 0.007127637043595314, + 0.005910802166908979, + 0.005265409592539072, + 0.0074471235275268555, + 0.011249140836298466, + 0.015205800533294678, + 0.0140922199934721, + 0.01168957632035017, + 0.00546610914170742, + 0.007005542982369661, + 0.007490808609873056, + 0.007552904076874256, + 0.017177000641822815, + 0.015663672238588333, + 0.02228018268942833, + 0.017649097368121147, + 0.01639178767800331, + 0.015397613868117332, + 0.024065952748060226, + 0.01874416135251522, + 0.012235001660883427, + 0.015820249915122986, + 0.016986338421702385, + 0.020224925130605698, + 0.007104414515197277, + 0.006493743509054184, + 0.00936884619295597, + 0.0097603565081954, + 0.012244505807757378, + 0.008202828466892242, + 0.024549243971705437, + 0.011529948562383652, + 0.02280549332499504, + 0.011003763414919376, + 0.0072864932008087635, + 0.008991645649075508, + 0.008332506753504276, + 0.009913288988173008, + 0.009611447341740131, + 0.007628201972693205, + 0.006739379838109016, + 0.0084757748991251, + 0.005545431282371283, + 0.008648714981973171, + 0.01124540064483881, + 0.01356204692274332, + 0.016029251739382744, + 0.010669254697859287, + 0.017078611999750137, + 0.013309583067893982, + 0.014852677471935749, + 0.013195950537919998, + 0.009768994525074959, + 0.011717965826392174, + 0.00837249867618084, + 0.008677562698721886, + 0.006057471502572298, + 0.007094854488968849, + 0.011319378390908241, + 0.009158783592283726, + 0.011427044868469238, + 0.009978423826396465, + 0.01017253752797842, + 0.01185374241322279, + 0.008969499729573727, + 0.010452451184391975, + 0.013267417438328266, + 0.005792111624032259 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 7 + ], + "selected_reward": -0.026338880881667137 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "dynamic, low-angle tracking shot follows a massive main battle tank as it hurtles forward at breakneck speed across a vast, rain-soaked, muddy field. Under a gloomy gray sky, the war machine, painted in a deep, matte olive green camouflage, churns its broad, sturdy metal tracks furiously, each rut digging deep into the ground and violently throwing huge chunks of sticky, dark brown mud and murky water high and low to the rear and sides in a spectacular, chaotic curtain of mud. The heavy vehicle heaves and pitches violently across the uneven terrain, its massive turret pointing steadily forward, its long barrel swaying slightly as the vehicle sways. Soft, diffused lighting accentuates the cool texture of the tank's metal hull and the wet sheen of the splattered mud. The entire scene, shrouded in a chilling hue, is imbued with a sense of raw power, tension, and unstoppable momentum.", + "seed": 15331, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000034/group_4/index_8349_seed_15331_dp_8_rank_16_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 4, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.025452891271561383, + "vlm_logprob": -0.025452891271561383, + "vlm_yes_no_margin": 4.2, + "vlm_token_credit": -0.025452891271561383, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a massive battle tank painted in deep matte olive green camouflage?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the tank moving at high speed across a vast, rain-soaked, muddy field?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are huge chunks of dark brown mud and murky water being thrown high and low by the tank's tracks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the sky gloomy and gray with soft, diffused lighting?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a dynamic low-angle tracking shot following the tank from behind or the side?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.05163879320025444 + ], + "opposite_logprob_tokens": [ + -3.1766388416290283 + ], + "logprob_reward": -0.05163879320025444, + "yes_no_margin_reward": 3.125, + "token_credit_reward": -0.05163879320025444, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14689837396144867, + 0.05055832862854004, + 0.060903821140527725, + 0.07032973319292068, + 0.03839936479926109, + 0.13242417573928833, + 0.10436563938856125, + 0.3028144836425781, + 0.048229753971099854, + 0.04507634416222572 + ], + "credit_map": [ + 0.021602759137749672, + 0.028715141117572784, + 0.02322801761329174, + 0.01482368167489767, + 0.011090893298387527, + 0.01961563713848591, + 0.010867206379771233, + 0.016955042257905006, + 0.007656251545995474, + 0.011656178161501884, + 0.006952397990971804, + 0.005966580472886562, + 0.005553325638175011, + 0.00618711719289422, + 0.0031291234772652388, + 0.003457358106970787, + 0.008626721799373627, + 0.013257208280265331, + 0.008700255304574966, + 0.0048560211434960365, + 0.006349705625325441, + 0.006755884271115065, + 0.006257927510887384, + 0.006100109778344631, + 0.008753317408263683, + 0.006059003062546253, + 0.009964528493583202, + 0.006337952334433794, + 0.01725427806377411, + 0.010608455166220665, + 0.005381874740123749, + 0.005970333702862263, + 0.0050520701333880424, + 0.0037344235461205244, + 0.004122862592339516, + 0.0039911530911922455, + 0.0056427037343382835, + 0.004284172784537077, + 0.0061242892406880856, + 0.005447692703455687, + 0.020663654431700706, + 0.008407377637922764, + 0.013237364590168, + 0.017422381788492203, + 0.011510743759572506, + 0.022435028105974197, + 0.01572098210453987, + 0.023026656359434128, + 0.010778329335153103, + 0.00927754770964384, + 0.011997373774647713, + 0.015932375565171242, + 0.014161803759634495, + 0.012773728929460049, + 0.01359359547495842, + 0.015850894153118134, + 0.031525734812021255, + 0.01492911484092474, + 0.029379380866885185, + 0.042935874313116074, + 0.043670184910297394, + 0.053661245852708817, + 0.04497000575065613, + 0.04174298420548439, + 0.004084070213139057, + 0.002460062736645341, + 0.00789859052747488, + 0.007252547424286604, + 0.004164181184023619, + 0.00514264078810811, + 0.007847986184060574, + 0.009379678405821323, + 0.007252268027514219, + 0.004519592970609665, + 0.007251421455293894, + 0.005680333822965622, + 0.004094783682376146, + 0.005344995763152838, + 0.004663954954594374, + 0.006268998607993126 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 0, + 5 + ], + "selected_reward": -0.05163879320025444 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a massive battle tank painted in deep matte olive green camouflage?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the tank moving at high speed across a vast, rain-soaked, muddy field?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are huge chunks of dark brown mud and murky water being thrown high and low by the tank's tracks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the sky gloomy and gray with soft, diffused lighting?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a dynamic low-angle tracking shot following the tank from behind or the side?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04310096427798271 + ], + "opposite_logprob_tokens": [ + -3.1681010723114014 + ], + "logprob_reward": -0.04310096427798271, + "yes_no_margin_reward": 3.125, + "token_credit_reward": -0.04310096427798271, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13811562955379486, + 0.17374467849731445, + 0.05528455227613449, + 0.20533473789691925, + 0.08448277413845062, + 0.05654922500252724, + 0.0746946632862091, + 0.057256147265434265, + 0.08904428035020828, + 0.06549334526062012 + ], + "credit_map": [ + 0.022800855338573456, + 0.029958873987197876, + 0.026118764653801918, + 0.01096013281494379, + 0.009428207762539387, + 0.019651340320706367, + 0.009192575700581074, + 0.010004876181483269, + 0.04323693737387657, + 0.04876631870865822, + 0.01672416739165783, + 0.011517636477947235, + 0.012109508737921715, + 0.02335597202181816, + 0.010734781622886658, + 0.007299367338418961, + 0.008923950605094433, + 0.010709221474826336, + 0.008446704596281052, + 0.006547941360622644, + 0.0063539654947817326, + 0.0049871583469212055, + 0.0041883173398673534, + 0.005127293057739735, + 0.03399878740310669, + 0.021757278591394424, + 0.043847937136888504, + 0.027317561209201813, + 0.016759680584073067, + 0.029726430773735046, + 0.014742632396519184, + 0.01718442142009735, + 0.011632944457232952, + 0.007502617314457893, + 0.007013251539319754, + 0.008213578723371029, + 0.01461890060454607, + 0.009584705345332623, + 0.012531168758869171, + 0.013385603204369545, + 0.00910367164760828, + 0.004509876482188702, + 0.006610923912376165, + 0.005641846917569637, + 0.005696824286133051, + 0.010273381136357784, + 0.006510928273200989, + 0.008201775141060352, + 0.007395891472697258, + 0.005365606397390366, + 0.007409119978547096, + 0.01229903195053339, + 0.00449707917869091, + 0.004919788334518671, + 0.019602766260504723, + 0.01320537831634283, + 0.006172090768814087, + 0.002770720049738884, + 0.005162173882126808, + 0.008673074655234814, + 0.0062537482008337975, + 0.009600045159459114, + 0.009251509793102741, + 0.009372781030833721, + 0.010533486492931843, + 0.003220969345420599, + 0.010971673764288425, + 0.011339818127453327, + 0.00674548139795661, + 0.01088789850473404, + 0.013293853960931301, + 0.022051101550459862, + 0.010226010344922543, + 0.006883764639496803, + 0.010722468607127666, + 0.007634647190570831, + 0.006730640307068825, + 0.007677383720874786, + 0.007091108243912458, + 0.008527319878339767 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 1, + 0 + ], + "selected_reward": -0.04310096427798271 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a massive battle tank painted in deep matte olive green camouflage?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the tank moving at high speed across a vast, rain-soaked, muddy field?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are huge chunks of dark brown mud and murky water being thrown high and low by the tank's tracks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the sky gloomy and gray with soft, diffused lighting?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a dynamic low-angle tracking shot following the tank from behind or the side?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.00249824533239007 + ], + "opposite_logprob_tokens": [ + -6.002498149871826 + ], + "logprob_reward": -0.00249824533239007, + "yes_no_margin_reward": 6.0, + "token_credit_reward": -0.00249824533239007, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.18998529016971588, + 0.17986632883548737, + 0.07048054039478302, + 0.08978486806154251, + 0.050866711884737015, + 0.0607142299413681, + 0.050547610968351364, + 0.04191074147820473, + 0.07834240049123764, + 0.18750126659870148 + ], + "credit_map": [ + 0.02829359844326973, + 0.0402815043926239, + 0.03335757181048393, + 0.01910652220249176, + 0.01358824223279953, + 0.02408924698829651, + 0.010988790541887283, + 0.02027982659637928, + 0.026526814326643944, + 0.041687265038490295, + 0.03241358697414398, + 0.018982283771038055, + 0.01396753080189228, + 0.02081076242029667, + 0.011746860109269619, + 0.013731240294873714, + 0.010076542384922504, + 0.01061499584466219, + 0.010381205007433891, + 0.007415322586894035, + 0.009337570518255234, + 0.006654203403741121, + 0.008067109622061253, + 0.007933592423796654, + 0.01301652379333973, + 0.00977263879030943, + 0.01964222453534603, + 0.009975937195122242, + 0.007925677113234997, + 0.015550542622804642, + 0.0065150074660778046, + 0.007386323064565659, + 0.0059689246118068695, + 0.004179930780082941, + 0.004514335189014673, + 0.005504482891410589, + 0.007905108854174614, + 0.005681488662958145, + 0.008576460182666779, + 0.008535985834896564, + 0.008544626645743847, + 0.004379397723823786, + 0.007531123235821724, + 0.007213626988232136, + 0.006372253410518169, + 0.008816970512270927, + 0.007872280664741993, + 0.009983953088521957, + 0.005960030946880579, + 0.004761591088026762, + 0.009107355028390884, + 0.00686700688675046, + 0.005362382624298334, + 0.004015975631773472, + 0.007140734698623419, + 0.007332534063607454, + 0.004288993775844574, + 0.0017450953600928187, + 0.005737737286835909, + 0.005314841400831938, + 0.004049462731927633, + 0.005501169711351395, + 0.00643920386210084, + 0.008834239095449448, + 0.010329061187803745, + 0.003202912164852023, + 0.010471910238265991, + 0.008714134804904461, + 0.004608253017067909, + 0.008006290532648563, + 0.013329457491636276, + 0.019680386409163475, + 0.014655484817922115, + 0.0168027151376009, + 0.03528289496898651, + 0.03172357752919197, + 0.012261847034096718, + 0.020655706524848938, + 0.02886103093624115, + 0.027258021757006645 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 1 + ], + "selected_reward": -0.00249824533239007 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a massive battle tank painted in deep matte olive green camouflage?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the tank moving at high speed across a vast, rain-soaked, muddy field?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are huge chunks of dark brown mud and murky water being thrown high and low by the tank's tracks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the sky gloomy and gray with soft, diffused lighting?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a dynamic low-angle tracking shot following the tank from behind or the side?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.006741872988641262 + ], + "opposite_logprob_tokens": [ + -5.006742000579834 + ], + "logprob_reward": -0.006741872988641262, + "yes_no_margin_reward": 5.0, + "token_credit_reward": -0.006741872988641262, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10774191468954086, + 0.08977091312408447, + 0.1773986965417862, + 0.0361085906624794, + 0.06463846564292908, + 0.044853851199150085, + 0.17904327809810638, + 0.03745828941464424, + 0.039453692734241486, + 0.22353234887123108 + ], + "credit_map": [ + 0.015431923791766167, + 0.022528911009430885, + 0.014660135842859745, + 0.012425299733877182, + 0.008662733249366283, + 0.013974977657198906, + 0.007212144788354635, + 0.01284578163176775, + 0.014159733429551125, + 0.0213099904358387, + 0.009831543080508709, + 0.01158286165446043, + 0.0072108181193470955, + 0.014730353839695454, + 0.004959820304065943, + 0.005985799245536327, + 0.03349659964442253, + 0.04559965059161186, + 0.023674102500081062, + 0.01425858587026596, + 0.01365539152175188, + 0.022946663200855255, + 0.011618713848292828, + 0.012148989364504814, + 0.005726666189730167, + 0.0042363922111690044, + 0.004341860767453909, + 0.004912399221211672, + 0.003257880685850978, + 0.006945861503481865, + 0.002949039451777935, + 0.00373848807066679, + 0.008602448739111423, + 0.00543784536421299, + 0.005042203236371279, + 0.006377127021551132, + 0.010879226960241795, + 0.007644651457667351, + 0.009926509112119675, + 0.010728457011282444, + 0.004920894745737314, + 0.004467044956982136, + 0.005882650148123503, + 0.007257856894284487, + 0.005405955016613007, + 0.004188886843621731, + 0.00636921776458621, + 0.006361345760524273, + 0.01954858936369419, + 0.015034841373562813, + 0.025733081623911858, + 0.030101953074336052, + 0.01644006557762623, + 0.015568042173981667, + 0.02816818095743656, + 0.028448525816202164, + 0.0036085264291614294, + 0.001874017296358943, + 0.0028540482744574547, + 0.005879728123545647, + 0.004972186870872974, + 0.006731775123625994, + 0.0065024239011108875, + 0.005035583395510912, + 0.004309902898967266, + 0.002126601291820407, + 0.00661763409152627, + 0.005769961979240179, + 0.0026890833396464586, + 0.004999327007681131, + 0.006230201106518507, + 0.006710979621857405, + 0.032497115433216095, + 0.022263143211603165, + 0.02957955189049244, + 0.027857938781380653, + 0.025606201961636543, + 0.034524139016866684, + 0.0224276315420866, + 0.0287766233086586 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 6, + 2 + ], + "selected_reward": -0.006741872988641262 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a massive battle tank painted in deep matte olive green camouflage?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the tank moving at high speed across a vast, rain-soaked, muddy field?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are huge chunks of dark brown mud and murky water being thrown high and low by the tank's tracks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the sky gloomy and gray with soft, diffused lighting?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a dynamic low-angle tracking shot following the tank from behind or the side?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.023284580558538437 + ], + "opposite_logprob_tokens": [ + -3.773284673690796 + ], + "logprob_reward": -0.023284580558538437, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.023284580558538437, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14865900576114655, + 0.0597180537879467, + 0.07297269999980927, + 0.10199198871850967, + 0.09805240482091904, + 0.1600734144449234, + 0.07029059529304504, + 0.06925194710493088, + 0.1576390564441681, + 0.06135082617402077 + ], + "credit_map": [ + 0.023429330438375473, + 0.034754980355501175, + 0.021816346794366837, + 0.014006765559315681, + 0.010847797617316246, + 0.021091455593705177, + 0.009757909923791885, + 0.012954403646290302, + 0.008570853620767593, + 0.01320688147097826, + 0.00877552293241024, + 0.005798254627734423, + 0.005746123380959034, + 0.0075296154245734215, + 0.0053149801678955555, + 0.0047758230939507484, + 0.012489760294556618, + 0.011726356111466885, + 0.011333031579852104, + 0.006796311587095261, + 0.006327791139483452, + 0.005344416480511427, + 0.010633046738803387, + 0.008321991190314293, + 0.01610517129302025, + 0.00980287604033947, + 0.019921639934182167, + 0.014449241571128368, + 0.007487539667636156, + 0.014229647815227509, + 0.009930239990353584, + 0.010065628215670586, + 0.019952289760112762, + 0.007317682728171349, + 0.011403738521039486, + 0.009502138942480087, + 0.010236673057079315, + 0.014448115602135658, + 0.01167384348809719, + 0.013517924584448338, + 0.022079506888985634, + 0.015092020854353905, + 0.02059321478009224, + 0.018476709723472595, + 0.023460114374756813, + 0.016532015055418015, + 0.018313894048333168, + 0.025525931268930435, + 0.009074339643120766, + 0.0052994186989963055, + 0.01223274227231741, + 0.010657420381903648, + 0.005554416216909885, + 0.006987551227211952, + 0.008526545017957687, + 0.011958167888224125, + 0.009059705771505833, + 0.0024703925009816885, + 0.008858754299581051, + 0.009975248947739601, + 0.004147625993937254, + 0.007785302121192217, + 0.01112054567784071, + 0.015834374353289604, + 0.014272681437432766, + 0.007444988936185837, + 0.02273343876004219, + 0.01851136051118374, + 0.015727998688817024, + 0.01723199151456356, + 0.029552437365055084, + 0.03216416388750076, + 0.007815643213689327, + 0.005667202640324831, + 0.010425818152725697, + 0.008340009488165379, + 0.005600066389888525, + 0.007000776939094067, + 0.008689877577126026, + 0.007811430841684341 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 8, + 0 + ], + "selected_reward": -0.023284580558538437 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "dynamic, low-angle tracking shot follows a massive main battle tank as it hurtles forward at breakneck speed across a vast, rain-soaked, muddy field. Under a gloomy gray sky, the war machine, painted in a deep, matte olive green camouflage, churns its broad, sturdy metal tracks furiously, each rut digging deep into the ground and violently throwing huge chunks of sticky, dark brown mud and murky water high and low to the rear and sides in a spectacular, chaotic curtain of mud. The heavy vehicle heaves and pitches violently across the uneven terrain, its massive turret pointing steadily forward, its long barrel swaying slightly as the vehicle sways. Soft, diffused lighting accentuates the cool texture of the tank's metal hull and the wet sheen of the splattered mud. The entire scene, shrouded in a chilling hue, is imbued with a sense of raw power, tension, and unstoppable momentum.", + "seed": 15332, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000034/group_4/index_8349_seed_15332_dp_9_rank_18_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 4, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.02602203101851046, + "vlm_logprob": -0.02602203101851046, + "vlm_yes_no_margin": 4.425, + "vlm_token_credit": -0.02602203101851046, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a massive battle tank painted in deep matte olive green camouflage?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the tank moving at high speed across a vast, rain-soaked, muddy field?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are huge chunks of dark brown mud and murky water being thrown high and low by the tank's tracks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the sky gloomy and gray with soft, diffused lighting?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a dynamic low-angle tracking shot following the tank from behind or the side?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.026563413441181183 + ], + "opposite_logprob_tokens": [ + -3.9015634059906006 + ], + "logprob_reward": -0.026563413441181183, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.026563413441181183, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2185211032629013, + 0.13846063613891602, + 0.09178010374307632, + 0.08931178599596024, + 0.0854991152882576, + 0.05655673146247864, + 0.03703012689948082, + 0.05281384289264679, + 0.09714672714471817, + 0.1328798085451126 + ], + "credit_map": [ + 0.03476439788937569, + 0.026086168363690376, + 0.03075290471315384, + 0.024615975096821785, + 0.02249077335000038, + 0.02915140613913536, + 0.027047334238886833, + 0.02361215278506279, + 0.0223523136228323, + 0.019411498680710793, + 0.01654709503054619, + 0.015280816704034805, + 0.014788372442126274, + 0.01900099404156208, + 0.01608549803495407, + 0.014994058758020401, + 0.019940251484513283, + 0.008870133198797703, + 0.01220110896974802, + 0.013203991577029228, + 0.0056481314823031425, + 0.012581105343997478, + 0.008907179348170757, + 0.010428206995129585, + 0.010964794084429741, + 0.009075820446014404, + 0.030273761600255966, + 0.00972351711243391, + 0.0055510513484478, + 0.010421275161206722, + 0.00677774241194129, + 0.0065238336101174355, + 0.008336284197866917, + 0.007738840766251087, + 0.01725246198475361, + 0.01607500948011875, + 0.007990292273461819, + 0.0065181381069123745, + 0.010822979733347893, + 0.01076511200517416, + 0.005696580279618502, + 0.006990980356931686, + 0.006928946357220411, + 0.010696303099393845, + 0.007209513336420059, + 0.004593325313180685, + 0.007728539872914553, + 0.00671254051849246, + 0.004324621520936489, + 0.005794726777821779, + 0.006018063984811306, + 0.0048108454793691635, + 0.0038458609487861395, + 0.003724188543856144, + 0.004493693821132183, + 0.004018126986920834, + 0.0058744000270962715, + 0.005191810894757509, + 0.006245685275644064, + 0.012718376703560352, + 0.007010831031948328, + 0.00420495355501771, + 0.005980262532830238, + 0.005587526597082615, + 0.008950506336987019, + 0.013991680927574635, + 0.013427633792161942, + 0.013077342882752419, + 0.012279155664145947, + 0.015831178054213524, + 0.013396451249718666, + 0.006192777771502733, + 0.009908435866236687, + 0.011831051670014858, + 0.018529126420617104, + 0.026483919471502304, + 0.014081540517508984, + 0.012446208856999874, + 0.024172259494662285, + 0.015427282080054283 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 9 + ], + "selected_reward": -0.026563413441181183 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a massive battle tank painted in deep matte olive green camouflage?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the tank moving at high speed across a vast, rain-soaked, muddy field?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are huge chunks of dark brown mud and murky water being thrown high and low by the tank's tracks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the sky gloomy and gray with soft, diffused lighting?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a dynamic low-angle tracking shot following the tank from behind or the side?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07900365442037582 + ], + "opposite_logprob_tokens": [ + -2.5790035724639893 + ], + "logprob_reward": -0.07900365442037582, + "yes_no_margin_reward": 2.5, + "token_credit_reward": -0.07900365442037582, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06150362640619278, + 0.3444788455963135, + 0.04446273297071457, + 0.10300467908382416, + 0.07937947660684586, + 0.2680056393146515, + 0.0404791422188282, + 0.011375458911061287, + 0.01648598723113537, + 0.030824430286884308 + ], + "credit_map": [ + 0.008069274015724659, + 0.009846406057476997, + 0.00832489226013422, + 0.006501373834908009, + 0.00637467298656702, + 0.00827101431787014, + 0.007000151555985212, + 0.007115835323929787, + 0.07862518727779388, + 0.07152154296636581, + 0.02833610214293003, + 0.023572955280542374, + 0.02467401698231697, + 0.06233683601021767, + 0.025426501408219337, + 0.02998567558825016, + 0.011735834181308746, + 0.0037642577663064003, + 0.004997184034436941, + 0.00396525114774704, + 0.003216114593669772, + 0.007112415507435799, + 0.005586637184023857, + 0.004085035994648933, + 0.011922468431293964, + 0.009396879002451897, + 0.026198258623480797, + 0.014774913899600506, + 0.007916610687971115, + 0.013091606087982655, + 0.011021259240806103, + 0.008682671003043652, + 0.007860459387302399, + 0.007846632041037083, + 0.015818122774362564, + 0.010224944911897182, + 0.007907683029770851, + 0.011031965725123882, + 0.007962433621287346, + 0.010727223008871078, + 0.02599894069135189, + 0.024095920845866203, + 0.023913655430078506, + 0.04727308824658394, + 0.03543910011649132, + 0.018824050202965736, + 0.0487615242600441, + 0.04369935393333435, + 0.005734754726290703, + 0.008006732910871506, + 0.0048368326388299465, + 0.005313692148774862, + 0.004768014419823885, + 0.004399762488901615, + 0.0033820883836597204, + 0.004037259612232447, + 0.0008935996447689831, + 0.0014941425761207938, + 0.0012117642909288406, + 0.0018954864935949445, + 0.0019103089580312371, + 0.0009772805497050285, + 0.001034085638821125, + 0.0019587883725762367, + 0.0018836555536836386, + 0.002850739751011133, + 0.00262621627189219, + 0.002174975583329797, + 0.0017206151969730854, + 0.0017310085240751505, + 0.0021399441175162792, + 0.0013588311849161983, + 0.0023851264268159866, + 0.0021475236862897873, + 0.008722895756363869, + 0.0037474290002137423, + 0.002328207716345787, + 0.0020325847435742617, + 0.0076459418050944805, + 0.0018147181253880262 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 5, + 3 + ], + "selected_reward": -0.07900365442037582 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a massive battle tank painted in deep matte olive green camouflage?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the tank moving at high speed across a vast, rain-soaked, muddy field?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are huge chunks of dark brown mud and murky water being thrown high and low by the tank's tracks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the sky gloomy and gray with soft, diffused lighting?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a dynamic low-angle tracking shot following the tank from behind or the side?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0017216873820871115 + ], + "opposite_logprob_tokens": [ + -6.3767218589782715 + ], + "logprob_reward": -0.0017216873820871115, + "yes_no_margin_reward": 6.375, + "token_credit_reward": -0.0017216873820871115, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07163269072771072, + 0.41321635246276855, + 0.03296999633312225, + 0.14629925787448883, + 0.07103974372148514, + 0.06625562161207199, + 0.0457175187766552, + 0.05064015835523605, + 0.06267978996038437, + 0.03954878821969032 + ], + "credit_map": [ + 0.009352611377835274, + 0.009507604874670506, + 0.008152840659022331, + 0.006561764981597662, + 0.008704382926225662, + 0.01104892697185278, + 0.008831802755594254, + 0.00947276409715414, + 0.08640066534280777, + 0.07837119698524475, + 0.04041410610079765, + 0.03234247863292694, + 0.033835332840681076, + 0.06919436156749725, + 0.03505746275186539, + 0.0376008078455925, + 0.00629241531714797, + 0.00340359122492373, + 0.004981172736734152, + 0.004696707706898451, + 0.002110525965690613, + 0.004176706541329622, + 0.0034211736638098955, + 0.003887707134708762, + 0.018332451581954956, + 0.013646959327161312, + 0.05204301327466965, + 0.01668340526521206, + 0.00721239298582077, + 0.015129979699850082, + 0.012225240468978882, + 0.011025827378034592, + 0.006262980867177248, + 0.0075338599272072315, + 0.013396683149039745, + 0.011566737666726112, + 0.009513071738183498, + 0.00645314808934927, + 0.009059924632310867, + 0.007253347896039486, + 0.006584451533854008, + 0.006945159751921892, + 0.0060867308638989925, + 0.010597164742648602, + 0.010130641981959343, + 0.0050988467410206795, + 0.010354320518672466, + 0.010458311066031456, + 0.006963088642805815, + 0.009105904959142208, + 0.004717200994491577, + 0.0037095381412655115, + 0.006638697814196348, + 0.006209126207977533, + 0.00397625332698226, + 0.004397714044898748, + 0.005275902803987265, + 0.006584027782082558, + 0.006497033406049013, + 0.008406761102378368, + 0.00531324278563261, + 0.006345639005303383, + 0.006208791863173246, + 0.0060087647289037704, + 0.007320800330489874, + 0.0107225701212883, + 0.010406294837594032, + 0.008297850377857685, + 0.00705392612144351, + 0.006186938378959894, + 0.0075866603292524815, + 0.0051047527231276035, + 0.0028911239933222532, + 0.003795275231823325, + 0.005154172424226999, + 0.006301277782768011, + 0.005128843244165182, + 0.005232411436736584, + 0.0057980772107839584, + 0.005247611552476883 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 3, + 0 + ], + "selected_reward": -0.0017216873820871115 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a massive battle tank painted in deep matte olive green camouflage?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the tank moving at high speed across a vast, rain-soaked, muddy field?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are huge chunks of dark brown mud and murky water being thrown high and low by the tank's tracks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the sky gloomy and gray with soft, diffused lighting?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a dynamic low-angle tracking shot following the tank from behind or the side?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004641236271709204 + ], + "opposite_logprob_tokens": [ + -5.379641056060791 + ], + "logprob_reward": -0.004641236271709204, + "yes_no_margin_reward": 5.375, + "token_credit_reward": -0.004641236271709204, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.28537625074386597, + 0.12639747560024261, + 0.12511569261550903, + 0.14477534592151642, + 0.0788324847817421, + 0.03994709253311157, + 0.0242240559309721, + 0.04369925707578659, + 0.04914535954594612, + 0.08248701691627502 + ], + "credit_map": [ + 0.05048366263508797, + 0.056156691163778305, + 0.028265953063964844, + 0.022938277572393417, + 0.02493305318057537, + 0.04964917525649071, + 0.023920586332678795, + 0.029028810560703278, + 0.02034049853682518, + 0.018700946122407913, + 0.020175538957118988, + 0.01550424937158823, + 0.010037342086434364, + 0.017389237880706787, + 0.013782097958028316, + 0.010467561893165112, + 0.0303768552839756, + 0.01515199150890112, + 0.020167574286460876, + 0.011424712836742401, + 0.008002614602446556, + 0.018581770360469818, + 0.01036202721297741, + 0.01104813814163208, + 0.021557137370109558, + 0.017305849120020866, + 0.04785227030515671, + 0.011461062356829643, + 0.006794493645429611, + 0.023042570799589157, + 0.009943302720785141, + 0.006818641908466816, + 0.008903445675969124, + 0.00831525307148695, + 0.01634267158806324, + 0.009580567479133606, + 0.01068407017737627, + 0.010164787992835045, + 0.008189099840819836, + 0.006652586627751589, + 0.004193953704088926, + 0.006071481388062239, + 0.0038765077479183674, + 0.005209073424339294, + 0.004388376139104366, + 0.003015947761014104, + 0.006492943502962589, + 0.0066988044418394566, + 0.0033701371867209673, + 0.005351015366613865, + 0.003356130560860038, + 0.0029889647848904133, + 0.00284872786141932, + 0.002171587198972702, + 0.0021227956749498844, + 0.0020146938040852547, + 0.005484249908477068, + 0.0049497452564537525, + 0.005698274355381727, + 0.008801296353340149, + 0.005035371985286474, + 0.004184447694569826, + 0.0052041420713067055, + 0.004341724328696728, + 0.004475494846701622, + 0.004922827705740929, + 0.011377423070371151, + 0.008174235932528973, + 0.004474475048482418, + 0.0041943383403122425, + 0.006567732896655798, + 0.004958828445523977, + 0.006570145022124052, + 0.00808019656687975, + 0.012009422294795513, + 0.013406587764620781, + 0.00908470619469881, + 0.011508562602102757, + 0.013006255961954594, + 0.008821128867566586 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 1 + ], + "selected_reward": -0.004641236271709204 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a massive battle tank painted in deep matte olive green camouflage?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the tank moving at high speed across a vast, rain-soaked, muddy field?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are huge chunks of dark brown mud and murky water being thrown high and low by the tank's tracks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the sky gloomy and gray with soft, diffused lighting?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a dynamic low-angle tracking shot following the tank from behind or the side?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.018180163577198982 + ], + "opposite_logprob_tokens": [ + -4.0181803703308105 + ], + "logprob_reward": -0.018180163577198982, + "yes_no_margin_reward": 4.0, + "token_credit_reward": -0.018180163577198982, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08775155991315842, + 0.12683095037937164, + 0.03407057374715805, + 0.07932495325803757, + 0.10041410475969315, + 0.3176674544811249, + 0.041735172271728516, + 0.05618773773312569, + 0.11358137428760529, + 0.042436085641384125 + ], + "credit_map": [ + 0.012669835239648819, + 0.011573683470487595, + 0.009660842828452587, + 0.00938629824668169, + 0.011377271264791489, + 0.013262392953038216, + 0.010053742676973343, + 0.009767490439116955, + 0.02810218371450901, + 0.021402766928076744, + 0.014713702723383904, + 0.010787581093609333, + 0.008373608812689781, + 0.020571187138557434, + 0.01160496287047863, + 0.011274967342615128, + 0.007664878387004137, + 0.0034572607837617397, + 0.004373973235487938, + 0.0034752702340483665, + 0.002580747241154313, + 0.004800710827112198, + 0.004877821542322636, + 0.0028399124275892973, + 0.010018628090620041, + 0.007810959126800299, + 0.02571048028767109, + 0.007700102869421244, + 0.00527734262868762, + 0.00753283966332674, + 0.007497169077396393, + 0.007777433376759291, + 0.010237799026072025, + 0.013918141834437847, + 0.018013274297118187, + 0.013262813910841942, + 0.01266448013484478, + 0.011886914260685444, + 0.011492501944303513, + 0.00893817562609911, + 0.02990391105413437, + 0.027414752170443535, + 0.029614867642521858, + 0.055678386241197586, + 0.04036695882678032, + 0.023069562390446663, + 0.05844472348690033, + 0.05317429080605507, + 0.005075728986412287, + 0.006919052917510271, + 0.005839183926582336, + 0.005494037643074989, + 0.004568901844322681, + 0.006500056944787502, + 0.003772490657866001, + 0.0035657209809869528, + 0.00531562278047204, + 0.006416521035134792, + 0.006443733815103769, + 0.011979501694440842, + 0.0066582513973116875, + 0.005321676842868328, + 0.00784680899232626, + 0.006205620709806681, + 0.01479119248688221, + 0.01678094081580639, + 0.0250029806047678, + 0.01722048595547676, + 0.00966662634164095, + 0.00945961382240057, + 0.012584267184138298, + 0.008075255900621414, + 0.0036033468786627054, + 0.004366003908216953, + 0.005960571113973856, + 0.0052238814532756805, + 0.0059470064006745815, + 0.0069752903655171394, + 0.006143432110548019, + 0.004216555971652269 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 1, + 8 + ], + "selected_reward": -0.018180163577198982 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "dynamic, low-angle tracking shot follows a massive main battle tank as it hurtles forward at breakneck speed across a vast, rain-soaked, muddy field. Under a gloomy gray sky, the war machine, painted in a deep, matte olive green camouflage, churns its broad, sturdy metal tracks furiously, each rut digging deep into the ground and violently throwing huge chunks of sticky, dark brown mud and murky water high and low to the rear and sides in a spectacular, chaotic curtain of mud. The heavy vehicle heaves and pitches violently across the uneven terrain, its massive turret pointing steadily forward, its long barrel swaying slightly as the vehicle sways. Soft, diffused lighting accentuates the cool texture of the tank's metal hull and the wet sheen of the splattered mud. The entire scene, shrouded in a chilling hue, is imbued with a sense of raw power, tension, and unstoppable momentum.", + "seed": 15333, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000034/group_4/index_8349_seed_15333_dp_9_rank_18_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 4, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.12332379631698132, + "vlm_logprob": -0.12332379631698132, + "vlm_yes_no_margin": 3.0, + "vlm_token_credit": -0.12332379631698132, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a massive battle tank painted in deep matte olive green camouflage?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the tank moving at high speed across a vast, rain-soaked, muddy field?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are huge chunks of dark brown mud and murky water being thrown high and low by the tank's tracks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the sky gloomy and gray with soft, diffused lighting?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a dynamic low-angle tracking shot following the tank from behind or the side?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.05960063636302948 + ], + "opposite_logprob_tokens": [ + -2.934600591659546 + ], + "logprob_reward": -0.05960063636302948, + "yes_no_margin_reward": 2.875, + "token_credit_reward": -0.05960063636302948, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2714768946170807, + 0.1839095503091812, + 0.06600846350193024, + 0.12316247820854187, + 0.071774423122406, + 0.045301057398319244, + 0.10685469210147858, + 0.05505487322807312, + 0.049008194357156754, + 0.02744933031499386 + ], + "credit_map": [ + 0.03924998641014099, + 0.026753416284918785, + 0.03309560939669609, + 0.0296709593385458, + 0.02021525613963604, + 0.030450895428657532, + 0.035054586827754974, + 0.056986190378665924, + 0.019481919705867767, + 0.021392669528722763, + 0.019834620878100395, + 0.017588112503290176, + 0.016732005402445793, + 0.022965196520090103, + 0.01573086343705654, + 0.05018417164683342, + 0.006607096176594496, + 0.005231770221143961, + 0.008495081216096878, + 0.01139634195715189, + 0.008217079564929008, + 0.005396517924964428, + 0.009534304030239582, + 0.011130272410809994, + 0.012780682183802128, + 0.008676011115312576, + 0.02297998033463955, + 0.01914110593497753, + 0.0098662618547678, + 0.010607351548969746, + 0.022505735978484154, + 0.016605345532298088, + 0.00820192601531744, + 0.008726585656404495, + 0.009156518615782261, + 0.009945559315383434, + 0.008602221496403217, + 0.009658231399953365, + 0.011000623926520348, + 0.006482755299657583, + 0.008327252231538296, + 0.00497756851837039, + 0.00615584384649992, + 0.0041544451378285885, + 0.005566517356783152, + 0.0060281637124717236, + 0.00486106937751174, + 0.005230193957686424, + 0.023375799879431725, + 0.007796294987201691, + 0.007474248297512531, + 0.004493402782827616, + 0.01769718900322914, + 0.03130347654223442, + 0.005021826364099979, + 0.009692455641925335, + 0.006478962488472462, + 0.00755984615534544, + 0.006660587154328823, + 0.007332241628319025, + 0.0067184241488575935, + 0.00682353088632226, + 0.00621806038543582, + 0.007263215258717537, + 0.005100628826767206, + 0.004668696783483028, + 0.0050195613875985146, + 0.005997541360557079, + 0.004810442216694355, + 0.0057213203981518745, + 0.008365482091903687, + 0.009324521757662296, + 0.0038193243090063334, + 0.003231215523555875, + 0.002706213854253292, + 0.003577044466510415, + 0.004451248794794083, + 0.003983131609857082, + 0.003078745910897851, + 0.0026024049147963524 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 3 + ], + "selected_reward": -0.05960063636302948 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a massive battle tank painted in deep matte olive green camouflage?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the tank moving at high speed across a vast, rain-soaked, muddy field?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are huge chunks of dark brown mud and murky water being thrown high and low by the tank's tracks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the sky gloomy and gray with soft, diffused lighting?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a dynamic low-angle tracking shot following the tank from behind or the side?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.4742853343486786 + ], + "opposite_logprob_tokens": [ + -0.974285364151001 + ], + "logprob_reward": -0.4742853343486786, + "yes_no_margin_reward": 0.5, + "token_credit_reward": -0.4742853343486786, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2603321671485901, + 0.3051653206348419, + 0.09517567604780197, + 0.10457204282283783, + 0.07848293334245682, + 0.038562241941690445, + 0.027055438607931137, + 0.03004567138850689, + 0.03727874159812927, + 0.023329749703407288 + ], + "credit_map": [ + 0.03733813762664795, + 0.018933789804577827, + 0.032670099288225174, + 0.028393838554620743, + 0.022357318550348282, + 0.03688967972993851, + 0.02679227851331234, + 0.05695705860853195, + 0.03572902828454971, + 0.02641146630048752, + 0.0398213155567646, + 0.029545437544584274, + 0.027250630781054497, + 0.039011429995298386, + 0.025016000494360924, + 0.0823800265789032, + 0.010256840847432613, + 0.010735687799751759, + 0.013239944353699684, + 0.011316982097923756, + 0.011786970309913158, + 0.009965131990611553, + 0.012693841010332108, + 0.015180294401943684, + 0.011068934574723244, + 0.008125477470457554, + 0.01816549152135849, + 0.01710817962884903, + 0.008751550689339638, + 0.008840451017022133, + 0.019533295184373856, + 0.01297866739332676, + 0.007211547344923019, + 0.008158702403306961, + 0.009958782233297825, + 0.016286831349134445, + 0.010061034932732582, + 0.007230509538203478, + 0.011163334362208843, + 0.00841219536960125, + 0.005092294421046972, + 0.006601438391953707, + 0.00620561558753252, + 0.0035827409010380507, + 0.0049762362614274025, + 0.0052070398814976215, + 0.004132937639951706, + 0.0027639400213956833, + 0.005435887724161148, + 0.004448562860488892, + 0.002675506519153714, + 0.0019416355062276125, + 0.0033004041761159897, + 0.004781071096658707, + 0.001752649899572134, + 0.0027197240851819515, + 0.0033525319304317236, + 0.003980347421020269, + 0.0040465835481882095, + 0.0036476431414484978, + 0.0034081884659826756, + 0.004671346861869097, + 0.0034483482595533133, + 0.003490684786811471, + 0.004160867538303137, + 0.0030671486165374517, + 0.004938045982271433, + 0.004238994792103767, + 0.005531954113394022, + 0.0038395607843995094, + 0.006200042553246021, + 0.00530213164165616, + 0.0025415816344320774, + 0.0029912088066339493, + 0.0024881770368665457, + 0.0039026986341923475, + 0.003396206069737673, + 0.0027682301588356495, + 0.002993628615513444, + 0.0022480208426713943 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 3 + ], + "selected_reward": -0.4742853343486786 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a massive battle tank painted in deep matte olive green camouflage?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the tank moving at high speed across a vast, rain-soaked, muddy field?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are huge chunks of dark brown mud and murky water being thrown high and low by the tank's tracks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the sky gloomy and gray with soft, diffused lighting?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a dynamic low-angle tracking shot following the tank from behind or the side?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005959006026387215 + ], + "opposite_logprob_tokens": [ + -5.1309590339660645 + ], + "logprob_reward": -0.005959006026387215, + "yes_no_margin_reward": 5.125, + "token_credit_reward": -0.005959006026387215, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.19422107934951782, + 0.28274470567703247, + 0.08883638679981232, + 0.04570329189300537, + 0.04364367201924324, + 0.032774753868579865, + 0.09311102330684662, + 0.06907689571380615, + 0.12915728986263275, + 0.02073090150952339 + ], + "credit_map": [ + 0.030008016154170036, + 0.01620725356042385, + 0.02384174056351185, + 0.020690040662884712, + 0.014762993901968002, + 0.02323364093899727, + 0.021315133199095726, + 0.044162265956401825, + 0.03788512572646141, + 0.025241881608963013, + 0.03359272703528404, + 0.02624797448515892, + 0.02505607157945633, + 0.037198472768068314, + 0.02387944981455803, + 0.07364297658205032, + 0.009459503926336765, + 0.01043569017201662, + 0.01175532117486, + 0.00836654007434845, + 0.012124025262892246, + 0.008716456592082977, + 0.011766594834625721, + 0.016212256625294685, + 0.0048464201390743256, + 0.0036273545119911432, + 0.007505092304199934, + 0.0076711163856089115, + 0.0034446746576577425, + 0.0039907582104206085, + 0.007315447088330984, + 0.007302428130060434, + 0.004922260995954275, + 0.00637973565608263, + 0.0036275358870625496, + 0.006949963513761759, + 0.007140063680708408, + 0.006197662558406591, + 0.003486163215711713, + 0.004940290469676256, + 0.006536218337714672, + 0.004068841226398945, + 0.0051347543485462666, + 0.0030523468740284443, + 0.004803475923836231, + 0.0037697877269238234, + 0.0029658633284270763, + 0.0024434637743979692, + 0.01609240099787712, + 0.007398467976599932, + 0.00880495086312294, + 0.006639581173658371, + 0.014178061857819557, + 0.021427568048238754, + 0.007430997211486101, + 0.01113899890333414, + 0.008411434479057789, + 0.010489405132830143, + 0.007335175294429064, + 0.009090790525078773, + 0.010190466418862343, + 0.008248623460531235, + 0.007827268913388252, + 0.007483736611902714, + 0.01290432084351778, + 0.009066685102880001, + 0.01481618918478489, + 0.014645393006503582, + 0.014150986447930336, + 0.014950739219784737, + 0.02217099443078041, + 0.02645198069512844, + 0.002582331420853734, + 0.0022476406302303076, + 0.0026958161033689976, + 0.0031359682325273752, + 0.0027575232088565826, + 0.0016906860982999206, + 0.0033669667318463326, + 0.002253969432786107 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 8 + ], + "selected_reward": -0.005959006026387215 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a massive battle tank painted in deep matte olive green camouflage?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the tank moving at high speed across a vast, rain-soaked, muddy field?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are huge chunks of dark brown mud and murky water being thrown high and low by the tank's tracks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the sky gloomy and gray with soft, diffused lighting?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a dynamic low-angle tracking shot following the tank from behind or the side?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.043082013726234436 + ], + "opposite_logprob_tokens": [ + -3.1680819988250732 + ], + "logprob_reward": -0.043082013726234436, + "yes_no_margin_reward": 3.125, + "token_credit_reward": -0.043082013726234436, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14699263870716095, + 0.2037847340106964, + 0.07252822071313858, + 0.05335943028330803, + 0.04285896569490433, + 0.14697250723838806, + 0.290219247341156, + 0.01619654707610607, + 0.01780742220580578, + 0.009280283004045486 + ], + "credit_map": [ + 0.023338619619607925, + 0.012853682972490788, + 0.01784539595246315, + 0.015805665403604507, + 0.01060507446527481, + 0.017577432096004486, + 0.01608113758265972, + 0.03288563713431358, + 0.02364373207092285, + 0.01807928830385208, + 0.022733261808753014, + 0.02056225575506687, + 0.017419131472706795, + 0.02521854266524315, + 0.02034885808825493, + 0.05577966570854187, + 0.008439657278358936, + 0.0083127086982131, + 0.009903974831104279, + 0.009574443101882935, + 0.008101418614387512, + 0.00754493148997426, + 0.010617995634675026, + 0.01003308780491352, + 0.005697940476238728, + 0.004668282810598612, + 0.0097538186237216, + 0.008631010539829731, + 0.004193073604255915, + 0.004346264060586691, + 0.009880556724965572, + 0.00618848018348217, + 0.004843799397349358, + 0.00697914557531476, + 0.003931668121367693, + 0.0047198026441037655, + 0.008405284024775028, + 0.005782956723123789, + 0.0033671529963612556, + 0.004829158075153828, + 0.02352660335600376, + 0.016192451119422913, + 0.028221487998962402, + 0.017011312767863274, + 0.01846925914287567, + 0.014116303995251656, + 0.016256975010037422, + 0.013178103603422642, + 0.06278453767299652, + 0.01858067326247692, + 0.020258571952581406, + 0.013618274591863155, + 0.046756450086832047, + 0.08949632942676544, + 0.01402890495955944, + 0.0246955044567585, + 0.002469516359269619, + 0.0016901087947189808, + 0.001452527241781354, + 0.0017575022066012025, + 0.002129477448761463, + 0.003592950291931629, + 0.0015012280782684684, + 0.0016032361891120672, + 0.00155935762450099, + 0.0016861301846802235, + 0.0019150800071656704, + 0.0023935609497129917, + 0.0026443270035088062, + 0.002222285373136401, + 0.002383557613939047, + 0.0030031243804842234, + 0.0011886995052918792, + 0.0009148212266154587, + 0.0008757813484407961, + 0.0012895704712718725, + 0.0014595891116186976, + 0.0011148680932819843, + 0.0016313655069097877, + 0.0008055879734456539 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 1, + 0 + ], + "selected_reward": -0.043082013726234436 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a massive battle tank painted in deep matte olive green camouflage?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the tank moving at high speed across a vast, rain-soaked, muddy field?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are huge chunks of dark brown mud and murky water being thrown high and low by the tank's tracks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the sky gloomy and gray with soft, diffused lighting?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a dynamic low-angle tracking shot following the tank from behind or the side?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03369199112057686 + ], + "opposite_logprob_tokens": [ + -3.408691883087158 + ], + "logprob_reward": -0.03369199112057686, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.03369199112057686, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.18368637561798096, + 0.08027935773134232, + 0.08107859641313553, + 0.20681677758693695, + 0.08338800072669983, + 0.12055955082178116, + 0.02669185772538185, + 0.062040459364652634, + 0.1240091398358345, + 0.03144984692335129 + ], + "credit_map": [ + 0.02653367444872856, + 0.01825042814016342, + 0.02097129262983799, + 0.02127126045525074, + 0.012814327143132687, + 0.021060168743133545, + 0.020705292001366615, + 0.04207994043827057, + 0.011734808795154095, + 0.006976010277867317, + 0.01212927233427763, + 0.007022433914244175, + 0.00884680263698101, + 0.009953725151717663, + 0.006923324428498745, + 0.016692982986569405, + 0.01028242614120245, + 0.008255615830421448, + 0.01093650609254837, + 0.008373428136110306, + 0.012260004878044128, + 0.008470301516354084, + 0.010355374775826931, + 0.012144941836595535, + 0.022952938452363014, + 0.021158933639526367, + 0.032499827444553375, + 0.02897586300969124, + 0.020973024889826775, + 0.02057424932718277, + 0.03609553351998329, + 0.023586411029100418, + 0.01040642149746418, + 0.010285862721502781, + 0.010148566216230392, + 0.010745911858975887, + 0.008231996558606625, + 0.011642899364233017, + 0.013345855288207531, + 0.008580494672060013, + 0.014067139476537704, + 0.016989443451166153, + 0.022501055151224136, + 0.011887924745678902, + 0.013149443082511425, + 0.01457797922194004, + 0.015454680658876896, + 0.011931885965168476, + 0.003338060574606061, + 0.0031268030870705843, + 0.003950168378651142, + 0.0028891186229884624, + 0.003051044652238488, + 0.003910294268280268, + 0.0029903051909059286, + 0.003436062717810273, + 0.006852608174085617, + 0.008660388179123402, + 0.00767027772963047, + 0.0077200159430503845, + 0.008845495991408825, + 0.008323553949594498, + 0.006900385022163391, + 0.007067731581628323, + 0.013498136773705482, + 0.011125755496323109, + 0.010127895511686802, + 0.016628529876470566, + 0.01193393673747778, + 0.013858705759048462, + 0.020543552935123444, + 0.026292623952031136, + 0.0038812244310975075, + 0.0033668046817183495, + 0.0040718973614275455, + 0.003641619347035885, + 0.0058394488878548145, + 0.0028520291671156883, + 0.004751679953187704, + 0.0030451409984380007 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 0, + 8 + ], + "selected_reward": -0.03369199112057686 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000034/group_5/index_tensor([2799, 2799])_group_5_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000034/group_5/index_tensor([2799, 2799])_group_5_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..a575ee4340ec73f044385acfc4d981e5407f3ecf --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000034/group_5/index_tensor([2799, 2799])_group_5_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "sharp side-angle macro lens, slowly tilted downwards to follow the movement of a water droplet, captures with exquisite detail a perfectly spherical, crystalline droplet resting on a nearly imperceptible, high-gloss, hydrophobic, invisible paint protection film that adheres perfectly to the car body. Beneath the film, a deep midnight blue paint shimmers with countless tiny metallic particles, creating a fluid, muscular, and elegant curve. Illuminated by bright, even studio overhead lighting, the top of the droplet forms a dazzling highlight, while its clear interior, like a miniature fisheye lens, refracts and distorts the blurred contours of its surroundings. Driven by the subtle effects of surface tension, the droplet begins to slide gracefully and silkily down the curve of the car body, its base slightly elongated by gravity, leaving behind a nearly invisible, microscopic wet trail. The entire process is steady and deliberate. The light precisely outlines the transparent contours of the water droplets and the smooth texture of the paint surface, creating a tranquil, pure and high-tech visual atmosphere, as if a quiet celebration of the beauty of industrial design and natural physical phenomena.", + "seed": 22045, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000034/group_5/index_2799_seed_22045_dp_10_rank_20_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 5, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.05997808873653412, + "vlm_logprob": -0.05997808873653412, + "vlm_yes_no_margin": 3.0, + "vlm_token_credit": -0.05997808873653412, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a sharp side-angle macro view of a perfectly spherical, crystalline water droplet?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the droplet resting on a high-gloss, hydrophobic, invisible paint protection film over a deep midnight blue car body with metallic particles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the droplet slide gracefully and silkily down the curve, leaving a nearly invisible microscopic wet trail?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene illuminated by bright, even studio overhead lighting that creates a dazzling highlight on the droplet's top?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera slowly tilt downwards to follow the movement of the droplet?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.06268993765115738 + ], + "opposite_logprob_tokens": [ + -2.937690019607544 + ], + "logprob_reward": -0.06268993765115738, + "yes_no_margin_reward": 2.875, + "token_credit_reward": -0.06268993765115738, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08536345511674881, + 0.1077040508389473, + 0.13325296342372894, + 0.07892463356256485, + 0.3844095468521118, + 0.04343166947364807, + 0.06666623055934906, + 0.031164588406682014, + 0.05228021740913391, + 0.016802655532956123 + ], + "credit_map": [ + 0.00794046651571989, + 0.00949784368276596, + 0.013063873164355755, + 0.008296862244606018, + 0.013904929161071777, + 0.01563212461769581, + 0.009443338960409164, + 0.007584020961076021, + 0.009284330531954765, + 0.008973473682999611, + 0.02996462769806385, + 0.008043867535889149, + 0.008943135850131512, + 0.021490560844540596, + 0.00773041695356369, + 0.013273641467094421, + 0.007080430630594492, + 0.008585663512349129, + 0.04959886148571968, + 0.0068266624584794044, + 0.005579620599746704, + 0.046694524586200714, + 0.0038773135747760534, + 0.005009890533983707, + 0.011434703134000301, + 0.013045725412666798, + 0.012810914777219296, + 0.006266293115913868, + 0.008826556615531445, + 0.015090885572135448, + 0.006305367685854435, + 0.005144187714904547, + 0.03654656559228897, + 0.009418779984116554, + 0.05456385761499405, + 0.010457716882228851, + 0.042929988354444504, + 0.13369056582450867, + 0.02309942990541458, + 0.07370267808437347, + 0.007139984052628279, + 0.004980198107659817, + 0.005077284295111895, + 0.0037614398170262575, + 0.00383805762976408, + 0.005071865860372782, + 0.0052816420793533325, + 0.00828119833022356, + 0.009758827276527882, + 0.009886502288281918, + 0.008438005112111568, + 0.010587030090391636, + 0.004382214043289423, + 0.0059463344514369965, + 0.008459396660327911, + 0.009207922033965588, + 0.003635162953287363, + 0.0030096624977886677, + 0.0034180928487330675, + 0.003101817797869444, + 0.0018002860015258193, + 0.002869617659598589, + 0.00666500860825181, + 0.0066649410873651505, + 0.0105931106954813, + 0.003594964975491166, + 0.005439335014671087, + 0.004402578342705965, + 0.0043698875233531, + 0.005050833337008953, + 0.009854240342974663, + 0.008975266478955746, + 0.0022656614892184734, + 0.0012606644304469228, + 0.002083135535940528, + 0.002108251443132758, + 0.0012742625549435616, + 0.001280579948797822, + 0.0031844854820519686, + 0.00334561406634748 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 2, + 1 + ], + "selected_reward": -0.06268993765115738 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a sharp side-angle macro view of a perfectly spherical, crystalline water droplet?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the droplet resting on a high-gloss, hydrophobic, invisible paint protection film over a deep midnight blue car body with metallic particles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the droplet slide gracefully and silkily down the curve, leaving a nearly invisible microscopic wet trail?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene illuminated by bright, even studio overhead lighting that creates a dazzling highlight on the droplet's top?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera slowly tilt downwards to follow the movement of the droplet?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03816511854529381 + ], + "opposite_logprob_tokens": [ + -3.2881650924682617 + ], + "logprob_reward": -0.03816511854529381, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.03816511854529381, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.037418048828840256, + 0.04429623484611511, + 0.03396204859018326, + 0.053818270564079285, + 0.6989490985870361, + 0.026029039174318314, + 0.019579585641622543, + 0.03364056348800659, + 0.04091646894812584, + 0.011390742845833302 + ], + "credit_map": [ + 0.0031899891328066587, + 0.004134560469537973, + 0.0052503724582493305, + 0.0041357078589499, + 0.005248383618891239, + 0.00732399383559823, + 0.004234724678099155, + 0.003900314914062619, + 0.004546113777905703, + 0.004317983519285917, + 0.01082809828221798, + 0.003595895366743207, + 0.0038745307829231024, + 0.005341765005141497, + 0.004033587872982025, + 0.007758253253996372, + 0.0019267327152192593, + 0.0017749005928635597, + 0.01271698996424675, + 0.0015408048639073968, + 0.0013270055642351508, + 0.012309263460338116, + 0.0008694696007296443, + 0.0014968767063692212, + 0.008186021819710732, + 0.009324881248176098, + 0.007951068691909313, + 0.004163945559412241, + 0.006634653080254793, + 0.009367500431835651, + 0.004406643565744162, + 0.0037835543043911457, + 0.07120372354984283, + 0.02205703966319561, + 0.0936514288187027, + 0.018925506621599197, + 0.07586798816919327, + 0.23337998986244202, + 0.04325231537222862, + 0.14061099290847778, + 0.004013516008853912, + 0.002608068287372589, + 0.004218642134219408, + 0.0021262369118630886, + 0.0020239781588315964, + 0.0032407219987362623, + 0.0033943180460482836, + 0.004403556697070599, + 0.0023256198037415743, + 0.002367459936067462, + 0.0021620485931634903, + 0.002187464851886034, + 0.0018029329366981983, + 0.0020529404282569885, + 0.0028315300587564707, + 0.003849588567391038, + 0.0037521435879170895, + 0.0036430489271879196, + 0.003551337867975235, + 0.0038444751407951117, + 0.0018729593139141798, + 0.0032757287845015526, + 0.007438342086970806, + 0.006262524519115686, + 0.006012761499732733, + 0.0023422411177307367, + 0.0033534362446516752, + 0.003786179469898343, + 0.0037100662011653185, + 0.0035565816797316074, + 0.009830726310610771, + 0.008324472233653069, + 0.001351029030047357, + 0.0006474907859228551, + 0.0011321186320856214, + 0.0013464585645124316, + 0.0008728682878427207, + 0.001467174501158297, + 0.002204061718657613, + 0.0023695402778685093 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 3, + 1 + ], + "selected_reward": -0.03816511854529381 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a sharp side-angle macro view of a perfectly spherical, crystalline water droplet?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the droplet resting on a high-gloss, hydrophobic, invisible paint protection film over a deep midnight blue car body with metallic particles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the droplet slide gracefully and silkily down the curve, leaving a nearly invisible microscopic wet trail?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene illuminated by bright, even studio overhead lighting that creates a dazzling highlight on the droplet's top?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera slowly tilt downwards to follow the movement of the droplet?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01607857272028923 + ], + "opposite_logprob_tokens": [ + -4.141078472137451 + ], + "logprob_reward": -0.01607857272028923, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.01607857272028923, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.05336228013038635, + 0.05165010690689087, + 0.02327929250895977, + 0.020215725526213646, + 0.7136152982711792, + 0.038055844604969025, + 0.008860909380018711, + 0.02249486930668354, + 0.054886482656002045, + 0.01357919443398714 + ], + "credit_map": [ + 0.004207063000649214, + 0.005762906279414892, + 0.006152405869215727, + 0.007076410576701164, + 0.007390208076685667, + 0.007133308332413435, + 0.00858899112790823, + 0.007050985004752874, + 0.0067086732015013695, + 0.006686396896839142, + 0.007187527138739824, + 0.005487574264407158, + 0.009143395349383354, + 0.005260987672954798, + 0.005569486878812313, + 0.005606065038591623, + 0.001986450282856822, + 0.0025484643410891294, + 0.006325708236545324, + 0.0015735524939373136, + 0.0021249635610729456, + 0.006157932337373495, + 0.0011200623121112585, + 0.0014421581290662289, + 0.0026801028288900852, + 0.003756166435778141, + 0.0031739643309265375, + 0.002060768660157919, + 0.002813867758959532, + 0.0020289793610572815, + 0.0014825885882601142, + 0.0022192876785993576, + 0.07452332228422165, + 0.022698594257235527, + 0.09395061433315277, + 0.019088264554739, + 0.07717156410217285, + 0.23520028591156006, + 0.04480225592851639, + 0.14618037641048431, + 0.005600521340966225, + 0.005721523892134428, + 0.004579629749059677, + 0.004458341747522354, + 0.002878401428461075, + 0.004228571429848671, + 0.0038506556302309036, + 0.006738197058439255, + 0.0010542290983721614, + 0.0011293151183053851, + 0.0015178255271166563, + 0.0011319321347400546, + 0.0006433451198972762, + 0.0010116409976035357, + 0.0011885450221598148, + 0.0011840761872008443, + 0.0029198203701525927, + 0.0019630701281130314, + 0.0025480103213340044, + 0.0025159460492432117, + 0.0014879112131893635, + 0.0023050750605762005, + 0.004354894161224365, + 0.004400142468512058, + 0.00941989105194807, + 0.0038953896146267653, + 0.005050370469689369, + 0.005356587003916502, + 0.004285563714802265, + 0.004914978519082069, + 0.012230686843395233, + 0.009733013808727264, + 0.0013861025217920542, + 0.001177905942313373, + 0.001419194508343935, + 0.0017936269287019968, + 0.001037726178765297, + 0.0009353276691399515, + 0.0026970168109983206, + 0.0031322939321398735 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 8, + 0 + ], + "selected_reward": -0.01607857272028923 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a sharp side-angle macro view of a perfectly spherical, crystalline water droplet?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the droplet resting on a high-gloss, hydrophobic, invisible paint protection film over a deep midnight blue car body with metallic particles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the droplet slide gracefully and silkily down the curve, leaving a nearly invisible microscopic wet trail?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene illuminated by bright, even studio overhead lighting that creates a dazzling highlight on the droplet's top?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera slowly tilt downwards to follow the movement of the droplet?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07003682106733322 + ], + "opposite_logprob_tokens": [ + -2.6950368881225586 + ], + "logprob_reward": -0.07003682106733322, + "yes_no_margin_reward": 2.625, + "token_credit_reward": -0.07003682106733322, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.04639297351241112, + 0.020941009745001793, + 0.059258244931697845, + 0.07156392186880112, + 0.2607628107070923, + 0.06822101026773453, + 0.03437505662441254, + 0.040099237114191055, + 0.3792026937007904, + 0.01918305829167366 + ], + "credit_map": [ + 0.004816973116248846, + 0.006768288090825081, + 0.004438800271600485, + 0.005963216535747051, + 0.007408647798001766, + 0.005484268069267273, + 0.006868188269436359, + 0.00464458717033267, + 0.0020040131639689207, + 0.002053226577118039, + 0.0025715611409395933, + 0.003421217203140259, + 0.0031073796562850475, + 0.00437622657045722, + 0.002035146811977029, + 0.001372235594317317, + 0.00632791593670845, + 0.004265040624886751, + 0.01868518255650997, + 0.0029671108350157738, + 0.003480838378891349, + 0.01623336412012577, + 0.002440409269183874, + 0.004858378786593676, + 0.013062713667750359, + 0.008348372764885426, + 0.010867029428482056, + 0.004597674589604139, + 0.010027701035141945, + 0.012112710624933243, + 0.006080114748328924, + 0.006467602215707302, + 0.022700194269418716, + 0.0052512092515826225, + 0.04210692271590233, + 0.006756751798093319, + 0.029384136199951172, + 0.0948401391506195, + 0.014669335447251797, + 0.045054104179143906, + 0.011078190989792347, + 0.006744334474205971, + 0.010884522460401058, + 0.005517062731087208, + 0.0045772395096719265, + 0.007863027043640614, + 0.008550149388611317, + 0.013006479479372501, + 0.004616291727870703, + 0.005271215923130512, + 0.00548181589692831, + 0.004918281454592943, + 0.0026320472825318575, + 0.0033524599857628345, + 0.003547761356458068, + 0.004555180203169584, + 0.0051062037236988544, + 0.004094333853572607, + 0.00446890527382493, + 0.0045040627010166645, + 0.002332051284611225, + 0.0037798702251166105, + 0.008102997206151485, + 0.007710810750722885, + 0.059372883290052414, + 0.021272143349051476, + 0.032692864537239075, + 0.03436100855469704, + 0.03374725207686424, + 0.03484645113348961, + 0.08870247006416321, + 0.074207603931427, + 0.0026143668219447136, + 0.0014414332108572125, + 0.0027229920960962772, + 0.0027436562813818455, + 0.0013658823445439339, + 0.0016331329243257642, + 0.003343504387885332, + 0.0033180892933160067 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 4, + 3 + ], + "selected_reward": -0.07003682106733322 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a sharp side-angle macro view of a perfectly spherical, crystalline water droplet?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the droplet resting on a high-gloss, hydrophobic, invisible paint protection film over a deep midnight blue car body with metallic particles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the droplet slide gracefully and silkily down the curve, leaving a nearly invisible microscopic wet trail?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene illuminated by bright, even studio overhead lighting that creates a dazzling highlight on the droplet's top?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera slowly tilt downwards to follow the movement of the droplet?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.11291999369859695 + ], + "opposite_logprob_tokens": [ + -2.237920045852661 + ], + "logprob_reward": -0.11291999369859695, + "yes_no_margin_reward": 2.125, + "token_credit_reward": -0.11291999369859695, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.007857543416321278, + 0.010661030188202858, + 0.11046180874109268, + 0.04679110646247864, + 0.7723889946937561, + 0.020033881068229675, + 0.009889241307973862, + 0.003861233126372099, + 0.0144658163189888, + 0.003589373081922531 + ], + "credit_map": [ + 0.0009067349601536989, + 0.0008865544805303216, + 0.001093797036446631, + 0.0009093211847357452, + 0.0012998038437217474, + 0.0011203998001292348, + 0.001003753743134439, + 0.0006371769704855978, + 0.0008414062322117388, + 0.0010055063758045435, + 0.003157333005219698, + 0.0007019602344371378, + 0.0012691299198195338, + 0.0022794625256210566, + 0.0007715559331700206, + 0.0006346751470118761, + 0.006127922795712948, + 0.006913044024258852, + 0.03930390998721123, + 0.005805767606943846, + 0.00462394068017602, + 0.03857320174574852, + 0.0036837910301983356, + 0.005430214572697878, + 0.007375899702310562, + 0.008720742538571358, + 0.006886247545480728, + 0.003965470474213362, + 0.0053732977248728275, + 0.007405991666018963, + 0.003666880540549755, + 0.00339656719006598, + 0.07893267273902893, + 0.02268042042851448, + 0.10479971021413803, + 0.020229659974575043, + 0.0839625895023346, + 0.25877419114112854, + 0.047707099467515945, + 0.15530259907245636, + 0.0029275836423039436, + 0.0036946292966604233, + 0.0031571134459227324, + 0.002112781163305044, + 0.0019107750849798322, + 0.0021625501103699207, + 0.0016173110343515873, + 0.0024511348456144333, + 0.0012178304605185986, + 0.0017840472282841802, + 0.0011469656601548195, + 0.0016368209617212415, + 0.0008540262933820486, + 0.0009099689195863903, + 0.001079646055586636, + 0.0012599341571331024, + 0.00045054638758301735, + 0.0003392671642359346, + 0.0003831506473943591, + 0.0004705658939201385, + 0.00019806121417786926, + 0.0003126715309917927, + 0.0009367724996991456, + 0.0007701973081566393, + 0.0019214871572330594, + 0.001040950184687972, + 0.0011216523125767708, + 0.0009784277062863111, + 0.0014226756757125258, + 0.0008495625224895775, + 0.003313618479296565, + 0.00381744047626853, + 0.0002724697405938059, + 0.00024408233002759516, + 0.0003325671423226595, + 0.0003686993441078812, + 0.00031114419107325375, + 0.00021461323194671422, + 0.0008256540750153363, + 0.0010201424593105912 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 2, + 3 + ], + "selected_reward": -0.11291999369859695 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "sharp side-angle macro lens, slowly tilted downwards to follow the movement of a water droplet, captures with exquisite detail a perfectly spherical, crystalline droplet resting on a nearly imperceptible, high-gloss, hydrophobic, invisible paint protection film that adheres perfectly to the car body. Beneath the film, a deep midnight blue paint shimmers with countless tiny metallic particles, creating a fluid, muscular, and elegant curve. Illuminated by bright, even studio overhead lighting, the top of the droplet forms a dazzling highlight, while its clear interior, like a miniature fisheye lens, refracts and distorts the blurred contours of its surroundings. Driven by the subtle effects of surface tension, the droplet begins to slide gracefully and silkily down the curve of the car body, its base slightly elongated by gravity, leaving behind a nearly invisible, microscopic wet trail. The entire process is steady and deliberate. The light precisely outlines the transparent contours of the water droplets and the smooth texture of the paint surface, creating a tranquil, pure and high-tech visual atmosphere, as if a quiet celebration of the beauty of industrial design and natural physical phenomena.", + "seed": 22046, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000034/group_5/index_2799_seed_22046_dp_10_rank_20_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 5, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.16390617080032827, + "vlm_logprob": -0.16390617080032827, + "vlm_yes_no_margin": 2.325, + "vlm_token_credit": -0.16390617080032827, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a sharp side-angle macro view of a perfectly spherical, crystalline water droplet?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the droplet resting on a high-gloss, hydrophobic, invisible paint protection film over a deep midnight blue car body with metallic particles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the droplet slide gracefully and silkily down the curve, leaving a nearly invisible microscopic wet trail?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene illuminated by bright, even studio overhead lighting that creates a dazzling highlight on the droplet's top?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera slowly tilt downwards to follow the movement of the droplet?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.2939019799232483 + ], + "opposite_logprob_tokens": [ + -1.4189019203186035 + ], + "logprob_reward": -0.2939019799232483, + "yes_no_margin_reward": 1.125, + "token_credit_reward": -0.2939019799232483, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13446857035160065, + 0.024162398651242256, + 0.3012882471084595, + 0.06908033788204193, + 0.09210395812988281, + 0.03600351884961128, + 0.10956718772649765, + 0.13404913246631622, + 0.04419691115617752, + 0.05507967248558998 + ], + "credit_map": [ + 0.024610860273241997, + 0.013433880172669888, + 0.014522570185363293, + 0.015848226845264435, + 0.016515877097845078, + 0.01939086802303791, + 0.01709209568798542, + 0.013054200448095798, + 0.002962770638987422, + 0.001921895076520741, + 0.002703889971598983, + 0.00256810849532485, + 0.0032562781125307083, + 0.002599917585030198, + 0.0052156527526676655, + 0.002933888230472803, + 0.032972995191812515, + 0.020335331559181213, + 0.03691421449184418, + 0.047465112060308456, + 0.03198523819446564, + 0.048568252474069595, + 0.05620049312710762, + 0.026846613734960556, + 0.006142918020486832, + 0.006817585788667202, + 0.006966863293200731, + 0.011316145770251751, + 0.008190355263650417, + 0.007752479985356331, + 0.015659082680940628, + 0.006234914064407349, + 0.0100458525121212, + 0.012535317800939083, + 0.01147571112960577, + 0.012041080743074417, + 0.009047421626746655, + 0.018535219132900238, + 0.010035542771220207, + 0.008387826383113861, + 0.004337789490818977, + 0.0028949794359505177, + 0.00417627627030015, + 0.005890609696507454, + 0.004111090209335089, + 0.00547401700168848, + 0.006406113039702177, + 0.002712645335122943, + 0.018212657421827316, + 0.010671532712876797, + 0.01681150496006012, + 0.013915329240262508, + 0.007965524680912495, + 0.01782500371336937, + 0.015722956508398056, + 0.00844268873333931, + 0.011819717474281788, + 0.006347452290356159, + 0.049966875463724136, + 0.009767022915184498, + 0.006377521902322769, + 0.0105739776045084, + 0.006726541556417942, + 0.03247002512216568, + 0.007192943710833788, + 0.0038642401341348886, + 0.004501781892031431, + 0.006021076813340187, + 0.0036508990451693535, + 0.007436917629092932, + 0.005760333966463804, + 0.005768724717199802, + 0.008529732935130596, + 0.004584432579576969, + 0.004718247801065445, + 0.0070798066444695, + 0.007653480861335993, + 0.009100382216274738, + 0.005965584889054298, + 0.007448004558682442 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 0, + 7 + ], + "selected_reward": -0.2939019799232483 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a sharp side-angle macro view of a perfectly spherical, crystalline water droplet?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the droplet resting on a high-gloss, hydrophobic, invisible paint protection film over a deep midnight blue car body with metallic particles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the droplet slide gracefully and silkily down the curve, leaving a nearly invisible microscopic wet trail?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene illuminated by bright, even studio overhead lighting that creates a dazzling highlight on the droplet's top?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera slowly tilt downwards to follow the movement of the droplet?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0381888709962368 + ], + "opposite_logprob_tokens": [ + -3.288188934326172 + ], + "logprob_reward": -0.0381888709962368, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.0381888709962368, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.24373731017112732, + 0.043250203132629395, + 0.15371167659759521, + 0.07280196994543076, + 0.12264887243509293, + 0.07169295102357864, + 0.06906160712242126, + 0.09074601531028748, + 0.0558650940656662, + 0.0764843001961708 + ], + "credit_map": [ + 0.04193146154284477, + 0.024357646703720093, + 0.023328233510255814, + 0.03779292106628418, + 0.02032952941954136, + 0.0475456602871418, + 0.02694748155772686, + 0.02150438167154789, + 0.005801792722195387, + 0.0034630007576197386, + 0.004236745648086071, + 0.005305724684149027, + 0.00486562866717577, + 0.004902822431176901, + 0.00974057987332344, + 0.004933909513056278, + 0.015874208882451057, + 0.0113043412566185, + 0.01897764578461647, + 0.02480410598218441, + 0.017115099355578423, + 0.022316552698612213, + 0.03060215897858143, + 0.012717553414404392, + 0.009030204266309738, + 0.007501046638935804, + 0.007997825741767883, + 0.010517152957618237, + 0.008029903285205364, + 0.00994435977190733, + 0.012546361424028873, + 0.007235121913254261, + 0.02028195932507515, + 0.010277366265654564, + 0.022796958684921265, + 0.01363673061132431, + 0.013374028727412224, + 0.020562347024679184, + 0.012706002220511436, + 0.009013470262289047, + 0.008157760836184025, + 0.007057862356305122, + 0.007173717021942139, + 0.008645127527415752, + 0.008204406127333641, + 0.01671436056494713, + 0.008962254971265793, + 0.006777460221201181, + 0.010530024766921997, + 0.006071725394576788, + 0.014370386488735676, + 0.007450223434716463, + 0.005908966530114412, + 0.010547702200710773, + 0.007169840391725302, + 0.007012731395661831, + 0.007583548780530691, + 0.005149156786501408, + 0.02845386601984501, + 0.0068496293388307095, + 0.005690113641321659, + 0.00783618539571762, + 0.006252855993807316, + 0.022930659353733063, + 0.010513232089579105, + 0.0045042484998703, + 0.005083010997623205, + 0.008235516026616096, + 0.005363058764487505, + 0.007625019643455744, + 0.008306081406772137, + 0.006234926637262106, + 0.012034312821924686, + 0.007132622413337231, + 0.008491473272442818, + 0.008220795542001724, + 0.008091319352388382, + 0.015224387869238853, + 0.008690033107995987, + 0.008599359542131424 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 4 + ], + "selected_reward": -0.0381888709962368 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a sharp side-angle macro view of a perfectly spherical, crystalline water droplet?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the droplet resting on a high-gloss, hydrophobic, invisible paint protection film over a deep midnight blue car body with metallic particles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the droplet slide gracefully and silkily down the curve, leaving a nearly invisible microscopic wet trail?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene illuminated by bright, even studio overhead lighting that creates a dazzling highlight on the droplet's top?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera slowly tilt downwards to follow the movement of the droplet?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.009790970012545586 + ], + "opposite_logprob_tokens": [ + -4.634790897369385 + ], + "logprob_reward": -0.009790970012545586, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.009790970012545586, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06787393242120743, + 0.012610649690032005, + 0.14344164729118347, + 0.022902805358171463, + 0.04801972210407257, + 0.21104083955287933, + 0.032787710428237915, + 0.30549150705337524, + 0.08798927068710327, + 0.06784194707870483 + ], + "credit_map": [ + 0.008867491967976093, + 0.005297395400702953, + 0.007219910155981779, + 0.015499875880777836, + 0.007900995202362537, + 0.005983320996165276, + 0.011238649487495422, + 0.005866291467100382, + 0.002235007006675005, + 0.0008309803670272231, + 0.0011958193499594927, + 0.00131055002566427, + 0.001051290542818606, + 0.001808472559787333, + 0.0019933576695621014, + 0.0021851714700460434, + 0.015888376161456108, + 0.009242533706128597, + 0.017631016671657562, + 0.022509658709168434, + 0.015466771088540554, + 0.023983722552657127, + 0.02593717910349369, + 0.012782387435436249, + 0.0025664407294243574, + 0.0021394998766481876, + 0.0023276282008737326, + 0.0029395585879683495, + 0.0024926746264100075, + 0.0038624173030257225, + 0.004416401963680983, + 0.0021581831388175488, + 0.005935761611908674, + 0.004932637792080641, + 0.00672509241849184, + 0.004680725280195475, + 0.006191571708768606, + 0.007103170268237591, + 0.006924023386090994, + 0.0055267405696213245, + 0.025860557332634926, + 0.01765844039618969, + 0.02423028089106083, + 0.025503087788820267, + 0.02392829768359661, + 0.045414768159389496, + 0.03157951682806015, + 0.016865892335772514, + 0.0037060854956507683, + 0.00285509810782969, + 0.003970073536038399, + 0.004264129791408777, + 0.003899384057149291, + 0.005036570597440004, + 0.006136731710284948, + 0.0029196348041296005, + 0.025809872895479202, + 0.012960865162312984, + 0.12169738113880157, + 0.021735195070505142, + 0.012554384768009186, + 0.02187585085630417, + 0.014065920375287533, + 0.0747920423746109, + 0.011776532046496868, + 0.007802893407642841, + 0.00830258708447218, + 0.01601613499224186, + 0.011094825342297554, + 0.008505560457706451, + 0.014859133400022984, + 0.009631603956222534, + 0.009127789177000523, + 0.005520004779100418, + 0.005914291366934776, + 0.010466564446687698, + 0.00960538350045681, + 0.008632265031337738, + 0.009544210508465767, + 0.009031441062688828 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 5, + 2 + ], + "selected_reward": -0.009790970012545586 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a sharp side-angle macro view of a perfectly spherical, crystalline water droplet?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the droplet resting on a high-gloss, hydrophobic, invisible paint protection film over a deep midnight blue car body with metallic particles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the droplet slide gracefully and silkily down the curve, leaving a nearly invisible microscopic wet trail?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene illuminated by bright, even studio overhead lighting that creates a dazzling highlight on the droplet's top?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera slowly tilt downwards to follow the movement of the droplet?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.22556951642036438 + ], + "opposite_logprob_tokens": [ + -1.600569486618042 + ], + "logprob_reward": -0.22556951642036438, + "yes_no_margin_reward": 1.375, + "token_credit_reward": -0.22556951642036438, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.05329604819417, + 0.02483678050339222, + 0.17017754912376404, + 0.01641177386045456, + 0.07985685765743256, + 0.032364372164011, + 0.05724278464913368, + 0.46829673647880554, + 0.027145657688379288, + 0.07037139683961868 + ], + "credit_map": [ + 0.004803565796464682, + 0.003940881229937077, + 0.006351296324282885, + 0.006062561646103859, + 0.00651506194844842, + 0.009140540845692158, + 0.010818684473633766, + 0.0056634582579135895, + 0.00276941922493279, + 0.0025345501489937305, + 0.002810238627716899, + 0.002559548942372203, + 0.0032784505747258663, + 0.003645140677690506, + 0.004752904176712036, + 0.00248652882874012, + 0.017065413296222687, + 0.011599640361964703, + 0.02092345803976059, + 0.02754749357700348, + 0.018736889585852623, + 0.02580035850405693, + 0.03383061662316322, + 0.014673680998384953, + 0.002937261713668704, + 0.0017425664700567722, + 0.001721285399980843, + 0.002137494971975684, + 0.0014989639166742563, + 0.0020704662892967463, + 0.0027298510540276766, + 0.0015738840447738767, + 0.010282285511493683, + 0.009345580823719501, + 0.009756289422512054, + 0.007137320004403591, + 0.00690468680113554, + 0.017340753227472305, + 0.007739280816167593, + 0.011350660584867, + 0.003875443711876869, + 0.002645013388246298, + 0.003553864313289523, + 0.004787938203662634, + 0.0038140411488711834, + 0.0037858011201024055, + 0.007394648622721434, + 0.002507622353732586, + 0.007083837408572435, + 0.0053475177846848965, + 0.006814168766140938, + 0.009638426825404167, + 0.005900124553591013, + 0.007206988055258989, + 0.010992413386702538, + 0.00425930880010128, + 0.03885265439748764, + 0.026302382349967957, + 0.18551045656204224, + 0.031332120299339294, + 0.017719849944114685, + 0.03896966949105263, + 0.015218386426568031, + 0.11439122259616852, + 0.004547993652522564, + 0.0021493721287697554, + 0.003945438656955957, + 0.003962640650570393, + 0.002411924535408616, + 0.003619290189817548, + 0.00350006646476686, + 0.0030089330393821, + 0.012375792488455772, + 0.007267317734658718, + 0.007760701235383749, + 0.008106373250484467, + 0.00715666962787509, + 0.011560210026800632, + 0.004946643020957708, + 0.011197686195373535 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 2, + 4 + ], + "selected_reward": -0.22556951642036438 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a sharp side-angle macro view of a perfectly spherical, crystalline water droplet?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the droplet resting on a high-gloss, hydrophobic, invisible paint protection film over a deep midnight blue car body with metallic particles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the droplet slide gracefully and silkily down the curve, leaving a nearly invisible microscopic wet trail?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene illuminated by bright, even studio overhead lighting that creates a dazzling highlight on the droplet's top?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera slowly tilt downwards to follow the movement of the droplet?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.2520795166492462 + ], + "opposite_logprob_tokens": [ + -1.5020794868469238 + ], + "logprob_reward": -0.2520795166492462, + "yes_no_margin_reward": 1.25, + "token_credit_reward": -0.2520795166492462, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10029826313257217, + 0.014498037286102772, + 0.05958455055952072, + 0.024302180856466293, + 0.028241366147994995, + 0.19466625154018402, + 0.07306740432977676, + 0.2880142033100128, + 0.0932735726237297, + 0.12405416369438171 + ], + "credit_map": [ + 0.012372308410704136, + 0.007487189956009388, + 0.009464333765208721, + 0.01432404387742281, + 0.011290100403130054, + 0.012953000143170357, + 0.0210890956223011, + 0.011318200267851353, + 0.0016412653494626284, + 0.0012064996408298612, + 0.0015358497621491551, + 0.0013520772336050868, + 0.0014073720667511225, + 0.002622018801048398, + 0.002581553300842643, + 0.002151402411982417, + 0.00563203077763319, + 0.004880946129560471, + 0.007009696681052446, + 0.009735779836773872, + 0.006456516217440367, + 0.00952381081879139, + 0.011246901005506516, + 0.005098871886730194, + 0.003089726436883211, + 0.001770565053448081, + 0.0030635392758995295, + 0.004335635807365179, + 0.0026054102927446365, + 0.0023955516517162323, + 0.005015479866415262, + 0.002026275033131242, + 0.003913708031177521, + 0.0025946481619030237, + 0.003896836657077074, + 0.003118199994787574, + 0.0025847407523542643, + 0.004811484832316637, + 0.0038242805749177933, + 0.003497468773275614, + 0.023067334666848183, + 0.015426906757056713, + 0.022719072178006172, + 0.02653380110859871, + 0.02321895956993103, + 0.03322656452655792, + 0.0366029255092144, + 0.013870694674551487, + 0.009600935503840446, + 0.00555649446323514, + 0.014001062139868736, + 0.009937991388142109, + 0.007090545259416103, + 0.008444256149232388, + 0.01151329092681408, + 0.006922835484147072, + 0.025364676490426064, + 0.018603306263685226, + 0.10427363961935043, + 0.022155433893203735, + 0.014516590163111687, + 0.02667328529059887, + 0.018303293734788895, + 0.05812400206923485, + 0.011425396427512169, + 0.010348714888095856, + 0.012844912707805634, + 0.013345836661756039, + 0.010950276628136635, + 0.014795156195759773, + 0.011134017258882523, + 0.008429265581071377, + 0.01589471660554409, + 0.010456045158207417, + 0.012974312528967857, + 0.019655128940939903, + 0.018510805442929268, + 0.01676744967699051, + 0.018117045983672142, + 0.01167866587638855 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 5, + 9 + ], + "selected_reward": -0.2520795166492462 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "sharp side-angle macro lens, slowly tilted downwards to follow the movement of a water droplet, captures with exquisite detail a perfectly spherical, crystalline droplet resting on a nearly imperceptible, high-gloss, hydrophobic, invisible paint protection film that adheres perfectly to the car body. Beneath the film, a deep midnight blue paint shimmers with countless tiny metallic particles, creating a fluid, muscular, and elegant curve. Illuminated by bright, even studio overhead lighting, the top of the droplet forms a dazzling highlight, while its clear interior, like a miniature fisheye lens, refracts and distorts the blurred contours of its surroundings. Driven by the subtle effects of surface tension, the droplet begins to slide gracefully and silkily down the curve of the car body, its base slightly elongated by gravity, leaving behind a nearly invisible, microscopic wet trail. The entire process is steady and deliberate. The light precisely outlines the transparent contours of the water droplets and the smooth texture of the paint surface, creating a tranquil, pure and high-tech visual atmosphere, as if a quiet celebration of the beauty of industrial design and natural physical phenomena.", + "seed": 22047, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000034/group_5/index_2799_seed_22047_dp_11_rank_22_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 5, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.05998803041875363, + "vlm_logprob": -0.05998803041875363, + "vlm_yes_no_margin": 3.2, + "vlm_token_credit": -0.05998803041875363, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a sharp side-angle macro view of a perfectly spherical, crystalline water droplet?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the droplet resting on a high-gloss, hydrophobic, invisible paint protection film over a deep midnight blue car body with metallic particles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the droplet slide gracefully and silkily down the curve, leaving a nearly invisible microscopic wet trail?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene illuminated by bright, even studio overhead lighting that creates a dazzling highlight on the droplet's top?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera slowly tilt downwards to follow the movement of the droplet?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07307226955890656 + ], + "opposite_logprob_tokens": [ + -2.8230721950531006 + ], + "logprob_reward": -0.07307226955890656, + "yes_no_margin_reward": 2.75, + "token_credit_reward": -0.07307226955890656, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.058094825595617294, + 0.063740074634552, + 0.0536058247089386, + 0.07270314544439316, + 0.04775198921561241, + 0.035092055797576904, + 0.13259916007518768, + 0.204576775431633, + 0.06566768139600754, + 0.26616841554641724 + ], + "credit_map": [ + 0.003545676590874791, + 0.00437408359721303, + 0.005261870566755533, + 0.006237285677343607, + 0.005994931794703007, + 0.01018254179507494, + 0.008288588374853134, + 0.01420985534787178, + 0.004396605305373669, + 0.00433442322537303, + 0.003518209792673588, + 0.004582292400300503, + 0.008832939900457859, + 0.015079133212566376, + 0.005864886566996574, + 0.017131591215729713, + 0.003962439950555563, + 0.005217310041189194, + 0.004059390164911747, + 0.007270416244864464, + 0.008502330631017685, + 0.0071147773414850235, + 0.011560356244444847, + 0.005918814800679684, + 0.004970064852386713, + 0.006783058866858482, + 0.00943051464855671, + 0.006124250590801239, + 0.010693136602640152, + 0.010192453861236572, + 0.013282272964715958, + 0.011227397248148918, + 0.003570119384676218, + 0.007352043408900499, + 0.005998172331601381, + 0.004787928890436888, + 0.007025368511676788, + 0.005712130106985569, + 0.007593597751110792, + 0.005712633952498436, + 0.0027286848053336143, + 0.007642737124115229, + 0.004287649877369404, + 0.0021564397029578686, + 0.005701171699911356, + 0.004115742165595293, + 0.004726964980363846, + 0.003732670797035098, + 0.016854366287589073, + 0.025182824581861496, + 0.011190983466804028, + 0.009233749471604824, + 0.01568961888551712, + 0.01598150096833706, + 0.02214949205517769, + 0.016316646710038185, + 0.029172973707318306, + 0.04078398272395134, + 0.01967351697385311, + 0.01373413484543562, + 0.019601428881287575, + 0.028156373649835587, + 0.030700165778398514, + 0.02275422029197216, + 0.009414568543434143, + 0.006683417595922947, + 0.007711874321103096, + 0.009728582575917244, + 0.0053697689436376095, + 0.0059891194105148315, + 0.008458365686237812, + 0.01231198851019144, + 0.03240758925676346, + 0.04351934418082237, + 0.03151564300060272, + 0.028174569830298424, + 0.02787909470498562, + 0.037307098507881165, + 0.029226787388324738, + 0.036138325929641724 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 7, + 6 + ], + "selected_reward": -0.07307226955890656 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a sharp side-angle macro view of a perfectly spherical, crystalline water droplet?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the droplet resting on a high-gloss, hydrophobic, invisible paint protection film over a deep midnight blue car body with metallic particles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the droplet slide gracefully and silkily down the curve, leaving a nearly invisible microscopic wet trail?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene illuminated by bright, even studio overhead lighting that creates a dazzling highlight on the droplet's top?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera slowly tilt downwards to follow the movement of the droplet?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02063567377626896 + ], + "opposite_logprob_tokens": [ + -3.8956356048583984 + ], + "logprob_reward": -0.02063567377626896, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.02063567377626896, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15559697151184082, + 0.3107120394706726, + 0.0555649995803833, + 0.04424673318862915, + 0.05595161020755768, + 0.049732647836208344, + 0.04968196898698807, + 0.09865481406450272, + 0.04012491554021835, + 0.13973331451416016 + ], + "credit_map": [ + 0.009290034882724285, + 0.00884341262280941, + 0.010477675125002861, + 0.011161470785737038, + 0.013790401630103588, + 0.025175215676426888, + 0.026605714112520218, + 0.05025302991271019, + 0.00907097477465868, + 0.008379640989005566, + 0.00597128551453352, + 0.011066297069191933, + 0.015874288976192474, + 0.11422906816005707, + 0.01857057400047779, + 0.12754985690116882, + 0.006854251027107239, + 0.007181372959166765, + 0.00527350464835763, + 0.005900315009057522, + 0.007400543428957462, + 0.007063170894980431, + 0.007793748285621405, + 0.008098090998828411, + 0.0039026655722409487, + 0.0043222676031291485, + 0.005892093759030104, + 0.003967217169702053, + 0.004783483687788248, + 0.0055968728847801685, + 0.00779860932379961, + 0.007983524352312088, + 0.004342213738709688, + 0.008749723434448242, + 0.0069596171379089355, + 0.00524402828887105, + 0.007968826219439507, + 0.0064117866568267345, + 0.0076165106147527695, + 0.00865890458226204, + 0.005169389769434929, + 0.008358213119208813, + 0.005922200158238411, + 0.003445379203185439, + 0.008476887829601765, + 0.0053526731207966805, + 0.0072597237303853035, + 0.005748180206865072, + 0.008172058500349522, + 0.0056511713191866875, + 0.0049348738975822926, + 0.0036387776490300894, + 0.005536649376153946, + 0.004855944775044918, + 0.0077446033246815205, + 0.009147893637418747, + 0.012839569710195065, + 0.019922221079468727, + 0.014974601566791534, + 0.00924823060631752, + 0.007484618574380875, + 0.013996287249028683, + 0.009509428404271603, + 0.010679858736693859, + 0.005480439402163029, + 0.004862794186919928, + 0.004798592999577522, + 0.00405589584261179, + 0.004387443419545889, + 0.00518425926566124, + 0.005113603081554174, + 0.006241887342184782, + 0.015596847049891949, + 0.0191547479480505, + 0.017408285290002823, + 0.013956506736576557, + 0.015507703647017479, + 0.017288699746131897, + 0.015802903100848198, + 0.025017613545060158 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 9 + ], + "selected_reward": -0.02063567377626896 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a sharp side-angle macro view of a perfectly spherical, crystalline water droplet?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the droplet resting on a high-gloss, hydrophobic, invisible paint protection film over a deep midnight blue car body with metallic particles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the droplet slide gracefully and silkily down the curve, leaving a nearly invisible microscopic wet trail?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene illuminated by bright, even studio overhead lighting that creates a dazzling highlight on the droplet's top?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera slowly tilt downwards to follow the movement of the droplet?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01607857272028923 + ], + "opposite_logprob_tokens": [ + -4.141078472137451 + ], + "logprob_reward": -0.01607857272028923, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.01607857272028923, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1044122725725174, + 0.3179559111595154, + 0.04768608510494232, + 0.029093610122799873, + 0.029070619493722916, + 0.033318571746349335, + 0.12288932502269745, + 0.09450840950012207, + 0.13092699646949768, + 0.09013813734054565 + ], + "credit_map": [ + 0.0092175817117095, + 0.0127116022631526, + 0.010255799628794193, + 0.014789639040827751, + 0.016220780089497566, + 0.013201798312366009, + 0.012503600679337978, + 0.015511475503444672, + 0.008064420893788338, + 0.004924213048070669, + 0.004639084916561842, + 0.007734216749668121, + 0.01133499015122652, + 0.12403570860624313, + 0.017812024801969528, + 0.13941127061843872, + 0.004746306221932173, + 0.006861267611384392, + 0.0035209967754781246, + 0.005240467377007008, + 0.007194293662905693, + 0.0062929363921284676, + 0.0068265353329479694, + 0.007003279868513346, + 0.002614716300740838, + 0.003135305130854249, + 0.0030988515354692936, + 0.0024034834932535887, + 0.004142532125115395, + 0.004276024177670479, + 0.005687372758984566, + 0.0037353269290179014, + 0.0021498273126780987, + 0.0033330637961626053, + 0.0027594484854489565, + 0.002963451901450753, + 0.006784995552152395, + 0.003824487328529358, + 0.003982708789408207, + 0.0032726360950618982, + 0.0040323929861187935, + 0.006371594965457916, + 0.0029256620910018682, + 0.0029299352318048477, + 0.0039039927069097757, + 0.004402189049869776, + 0.004938697442412376, + 0.0038141077384352684, + 0.02006002701818943, + 0.027698419988155365, + 0.0099873598664999, + 0.008753020316362381, + 0.012203183956444263, + 0.016116373240947723, + 0.012675885111093521, + 0.01539506297558546, + 0.018273796886205673, + 0.018981656059622765, + 0.010395330376923084, + 0.0067138285376131535, + 0.006920642219483852, + 0.014238889329135418, + 0.007145653944462538, + 0.011838605627417564, + 0.018577463924884796, + 0.017796335741877556, + 0.017774173989892006, + 0.016382265836000443, + 0.015150525607168674, + 0.014654921367764473, + 0.015529272146522999, + 0.015062046237289906, + 0.009589020162820816, + 0.008062712848186493, + 0.010793724097311497, + 0.008867756463587284, + 0.01176049467176199, + 0.009919529780745506, + 0.012568918988108635, + 0.01857597753405571 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 8, + 6 + ], + "selected_reward": -0.01607857272028923 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a sharp side-angle macro view of a perfectly spherical, crystalline water droplet?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the droplet resting on a high-gloss, hydrophobic, invisible paint protection film over a deep midnight blue car body with metallic particles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the droplet slide gracefully and silkily down the curve, leaving a nearly invisible microscopic wet trail?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene illuminated by bright, even studio overhead lighting that creates a dazzling highlight on the droplet's top?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera slowly tilt downwards to follow the movement of the droplet?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02981162816286087 + ], + "opposite_logprob_tokens": [ + -3.5298116207122803 + ], + "logprob_reward": -0.02981162816286087, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.02981162816286087, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.05711643770337105, + 0.24484883248806, + 0.04128041863441467, + 0.049416374415159225, + 0.072767473757267, + 0.03331317380070686, + 0.04363250732421875, + 0.16800948977470398, + 0.14743036031723022, + 0.14218494296073914 + ], + "credit_map": [ + 0.005015128757804632, + 0.005288096144795418, + 0.004210397135466337, + 0.004551986698061228, + 0.005547900218516588, + 0.010188591666519642, + 0.00686598988249898, + 0.015448346734046936, + 0.01041007973253727, + 0.010330207645893097, + 0.007202738430351019, + 0.00866569485515356, + 0.018810315057635307, + 0.08203906565904617, + 0.014462505467236042, + 0.0929282158613205, + 0.0036776429042220116, + 0.004648387432098389, + 0.003506443230435252, + 0.0042782071977853775, + 0.0059997146017849445, + 0.005228226073086262, + 0.0070375846698880196, + 0.006904208566993475, + 0.004287781659513712, + 0.005366814788430929, + 0.006745473947376013, + 0.004370369482785463, + 0.006000072695314884, + 0.006616164930164814, + 0.007759172935038805, + 0.00827051978558302, + 0.006227165460586548, + 0.012261171825230122, + 0.008549608290195465, + 0.006020601373165846, + 0.01133284717798233, + 0.009226209484040737, + 0.009756282903254032, + 0.009393579326570034, + 0.003389693796634674, + 0.008602546527981758, + 0.0027580109890550375, + 0.0027956899721175432, + 0.004040536005049944, + 0.004665223415941, + 0.003347973106428981, + 0.0037134976591914892, + 0.005693443119525909, + 0.006327951792627573, + 0.0034656806383281946, + 0.0030423554126173258, + 0.005095881409943104, + 0.005120906978845596, + 0.007867244072258472, + 0.007019044831395149, + 0.026661742478609085, + 0.0369495153427124, + 0.01597660966217518, + 0.010419214144349098, + 0.014689462259411812, + 0.025075694546103477, + 0.021838989108800888, + 0.016398269683122635, + 0.016680289059877396, + 0.020153407007455826, + 0.023345915600657463, + 0.020208604633808136, + 0.013371933251619339, + 0.01577092334628105, + 0.020790986716747284, + 0.017108306288719177, + 0.01690351963043213, + 0.02314022183418274, + 0.015659820288419724, + 0.015614829026162624, + 0.01448899321258068, + 0.01940568909049034, + 0.015468941070139408, + 0.02150292880833149 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 7, + 8 + ], + "selected_reward": -0.02981162816286087 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a sharp side-angle macro view of a perfectly spherical, crystalline water droplet?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the droplet resting on a high-gloss, hydrophobic, invisible paint protection film over a deep midnight blue car body with metallic particles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the droplet slide gracefully and silkily down the curve, leaving a nearly invisible microscopic wet trail?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene illuminated by bright, even studio overhead lighting that creates a dazzling highlight on the droplet's top?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera slowly tilt downwards to follow the movement of the droplet?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.1603420078754425 + ], + "opposite_logprob_tokens": [ + -1.9103419780731201 + ], + "logprob_reward": -0.1603420078754425, + "yes_no_margin_reward": 1.75, + "token_credit_reward": -0.1603420078754425, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.055297382175922394, + 0.12986601889133453, + 0.02522066794335842, + 0.03986692056059837, + 0.040757421404123306, + 0.02992279827594757, + 0.17626404762268066, + 0.35043007135391235, + 0.11934579908847809, + 0.03302888199687004 + ], + "credit_map": [ + 0.004966620821505785, + 0.006751240696758032, + 0.0053432104177773, + 0.007971244864165783, + 0.005210974719375372, + 0.006184553727507591, + 0.009551740251481533, + 0.009317794814705849, + 0.006228242069482803, + 0.004141652025282383, + 0.005357434973120689, + 0.005505939945578575, + 0.010601739399135113, + 0.042136166244745255, + 0.007746301591396332, + 0.04814853519201279, + 0.0019961686339229345, + 0.0027170355897396803, + 0.001873335218988359, + 0.0019613325130194426, + 0.004603786394000053, + 0.003768005408346653, + 0.002403860678896308, + 0.005897144787013531, + 0.0040519991889595985, + 0.00486290268599987, + 0.005282348021864891, + 0.0028222277760505676, + 0.0052442243322730064, + 0.005104142241179943, + 0.005178437102586031, + 0.007320639211684465, + 0.003801585640758276, + 0.006146538536995649, + 0.0059034936130046844, + 0.00417398102581501, + 0.005448588170111179, + 0.005082688294351101, + 0.005218163598328829, + 0.004982379265129566, + 0.0032773257698863745, + 0.006171866320073605, + 0.0026502623222768307, + 0.0021445616148412228, + 0.004866794683039188, + 0.004261927213519812, + 0.0029800941701978445, + 0.0035699671134352684, + 0.027473509311676025, + 0.04011780768632889, + 0.014408529736101627, + 0.01125038880854845, + 0.019704632461071014, + 0.02336186356842518, + 0.018184520304203033, + 0.021762795746326447, + 0.062159355729818344, + 0.07895952463150024, + 0.03379426524043083, + 0.021850164979696274, + 0.025240566581487656, + 0.054136972874403, + 0.03620252385735512, + 0.038086697459220886, + 0.009523636661469936, + 0.01632966846227646, + 0.019320417195558548, + 0.016314145177602768, + 0.011838177219033241, + 0.012669015675783157, + 0.019054751843214035, + 0.014295991510152817, + 0.003637713845819235, + 0.0054174927063286304, + 0.003026631660759449, + 0.0031649796292185783, + 0.004188790451735258, + 0.004513968713581562, + 0.003395674517378211, + 0.0056836302392184734 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 6, + 1 + ], + "selected_reward": -0.1603420078754425 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "sharp side-angle macro lens, slowly tilted downwards to follow the movement of a water droplet, captures with exquisite detail a perfectly spherical, crystalline droplet resting on a nearly imperceptible, high-gloss, hydrophobic, invisible paint protection film that adheres perfectly to the car body. Beneath the film, a deep midnight blue paint shimmers with countless tiny metallic particles, creating a fluid, muscular, and elegant curve. Illuminated by bright, even studio overhead lighting, the top of the droplet forms a dazzling highlight, while its clear interior, like a miniature fisheye lens, refracts and distorts the blurred contours of its surroundings. Driven by the subtle effects of surface tension, the droplet begins to slide gracefully and silkily down the curve of the car body, its base slightly elongated by gravity, leaving behind a nearly invisible, microscopic wet trail. The entire process is steady and deliberate. The light precisely outlines the transparent contours of the water droplets and the smooth texture of the paint surface, creating a tranquil, pure and high-tech visual atmosphere, as if a quiet celebration of the beauty of industrial design and natural physical phenomena.", + "seed": 22048, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000034/group_5/index_2799_seed_22048_dp_11_rank_22_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 5, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.07153315478935837, + "vlm_logprob": -0.07153315478935837, + "vlm_yes_no_margin": 3.4, + "vlm_token_credit": -0.07153315478935837, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a sharp side-angle macro view of a perfectly spherical, crystalline water droplet?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the droplet resting on a high-gloss, hydrophobic, invisible paint protection film over a deep midnight blue car body with metallic particles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the droplet slide gracefully and silkily down the curve, leaving a nearly invisible microscopic wet trail?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene illuminated by bright, even studio overhead lighting that creates a dazzling highlight on the droplet's top?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera slowly tilt downwards to follow the movement of the droplet?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02932148426771164 + ], + "opposite_logprob_tokens": [ + -4.029321670532227 + ], + "logprob_reward": -0.02932148426771164, + "yes_no_margin_reward": 4.0, + "token_credit_reward": -0.02932148426771164, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10038454085588455, + 0.039411067962646484, + 0.012001318857073784, + 0.01709052361547947, + 0.02281627431511879, + 0.027033375576138496, + 0.02171892300248146, + 0.5383679270744324, + 0.025557005777955055, + 0.19561909139156342 + ], + "credit_map": [ + 0.013416778296232224, + 0.011245271191000938, + 0.014507351443171501, + 0.011670693755149841, + 0.013579574413597584, + 0.01688382774591446, + 0.01343787182122469, + 0.005643162876367569, + 0.0036182370968163013, + 0.003808915615081787, + 0.0056043826043605804, + 0.007708222605288029, + 0.006084837019443512, + 0.005232307128608227, + 0.005057149566709995, + 0.0022970126010477543, + 0.0012120743049308658, + 0.0013055909657850862, + 0.0013461909256875515, + 0.0018325530691072345, + 0.0016180529491975904, + 0.0018646985990926623, + 0.0017132596112787724, + 0.0011088971514254808, + 0.001953044906258583, + 0.001945160562172532, + 0.0023009185679256916, + 0.002184538869187236, + 0.0017625960754230618, + 0.0027972941752523184, + 0.0021344274282455444, + 0.002012542448937893, + 0.002092721639201045, + 0.0021981881000101566, + 0.0028415594715625048, + 0.004116065800189972, + 0.002983157755807042, + 0.0027622950728982687, + 0.004016450140625238, + 0.0018058334244415164, + 0.002663730178028345, + 0.004736714530736208, + 0.0037024749908596277, + 0.004489011596888304, + 0.003333265660330653, + 0.0031223748810589314, + 0.003150062868371606, + 0.0018357365624979138, + 0.0024724439717829227, + 0.004866633098572493, + 0.0025856317952275276, + 0.0029464964754879475, + 0.002713550115004182, + 0.002206662902608514, + 0.002285431604832411, + 0.001642070710659027, + 0.04607834666967392, + 0.08632893860340118, + 0.06510495394468307, + 0.053629882633686066, + 0.06504525989294052, + 0.09098703414201736, + 0.0507531464099884, + 0.08044029027223587, + 0.0027961237356066704, + 0.0039145322516560555, + 0.0031399440485984087, + 0.0027115540578961372, + 0.00396656384691596, + 0.0030251420103013515, + 0.003272874280810356, + 0.0027302689850330353, + 0.0253573227673769, + 0.03846758231520653, + 0.022523658350110054, + 0.018272100016474724, + 0.02859337627887726, + 0.016969826072454453, + 0.023476960137486458, + 0.021958239376544952 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 9, + 0 + ], + "selected_reward": -0.02932148426771164 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a sharp side-angle macro view of a perfectly spherical, crystalline water droplet?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the droplet resting on a high-gloss, hydrophobic, invisible paint protection film over a deep midnight blue car body with metallic particles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the droplet slide gracefully and silkily down the curve, leaving a nearly invisible microscopic wet trail?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene illuminated by bright, even studio overhead lighting that creates a dazzling highlight on the droplet's top?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera slowly tilt downwards to follow the movement of the droplet?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0431685596704483 + ], + "opposite_logprob_tokens": [ + -3.168168544769287 + ], + "logprob_reward": -0.0431685596704483, + "yes_no_margin_reward": 3.125, + "token_credit_reward": -0.0431685596704483, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10693056136369705, + 0.11053843051195145, + 0.05271857976913452, + 0.04018789529800415, + 0.02649402990937233, + 0.11145979166030884, + 0.02638351358473301, + 0.4173588156700134, + 0.04261366277933121, + 0.0653146505355835 + ], + "credit_map": [ + 0.009500900283455849, + 0.010224147699773312, + 0.019143328070640564, + 0.010786888189613819, + 0.02127988263964653, + 0.015124257653951645, + 0.01423618569970131, + 0.006634972989559174, + 0.010058843530714512, + 0.01080669742077589, + 0.01317192055284977, + 0.02315336838364601, + 0.013981135562062263, + 0.01697710156440735, + 0.011535792611539364, + 0.010853568091988564, + 0.009765081107616425, + 0.0058388723991811275, + 0.006306915078312159, + 0.005194122903048992, + 0.005090037826448679, + 0.010127516463398933, + 0.005405195988714695, + 0.004990840796381235, + 0.007519977632910013, + 0.005133047234266996, + 0.005695969332009554, + 0.003631891217082739, + 0.0034356452524662018, + 0.007153092883527279, + 0.003917992580682039, + 0.0037002810277044773, + 0.002406880958005786, + 0.004209279082715511, + 0.003390472149476409, + 0.0047488100826740265, + 0.002522808499634266, + 0.003239865880459547, + 0.00385166029445827, + 0.0021242517977952957, + 0.011448509059846401, + 0.017340218648314476, + 0.014900140464305878, + 0.023355500772595406, + 0.013840531930327415, + 0.011829271912574768, + 0.010100258514285088, + 0.008645364083349705, + 0.0028519812040030956, + 0.004889652598649263, + 0.003209182294085622, + 0.004055826459079981, + 0.0029958884697407484, + 0.002851329743862152, + 0.0029881466180086136, + 0.002541505265980959, + 0.03660088777542114, + 0.06290654838085175, + 0.05028265342116356, + 0.04051131010055542, + 0.04799502342939377, + 0.07375062257051468, + 0.03708929568529129, + 0.06822247803211212, + 0.004394411575049162, + 0.009125908836722374, + 0.005139200948178768, + 0.0038438220508396626, + 0.0071055400185287, + 0.003361669834703207, + 0.004473193548619747, + 0.005169915501028299, + 0.010006126016378403, + 0.009677057154476643, + 0.008410787209868431, + 0.00794775690883398, + 0.009649869985878468, + 0.006951684132218361, + 0.005998794920742512, + 0.0066725765354931355 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 5, + 1 + ], + "selected_reward": -0.0431685596704483 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a sharp side-angle macro view of a perfectly spherical, crystalline water droplet?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the droplet resting on a high-gloss, hydrophobic, invisible paint protection film over a deep midnight blue car body with metallic particles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the droplet slide gracefully and silkily down the curve, leaving a nearly invisible microscopic wet trail?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene illuminated by bright, even studio overhead lighting that creates a dazzling highlight on the droplet's top?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera slowly tilt downwards to follow the movement of the droplet?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0067427014000713825 + ], + "opposite_logprob_tokens": [ + -5.006742477416992 + ], + "logprob_reward": -0.0067427014000713825, + "yes_no_margin_reward": 5.0, + "token_credit_reward": -0.0067427014000713825, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.030504604801535606, + 0.028771305456757545, + 0.016962839290499687, + 0.011653576977550983, + 0.011354480870068073, + 0.031003229320049286, + 0.038095276802778244, + 0.682937741279602, + 0.1273055225610733, + 0.02141142450273037 + ], + "credit_map": [ + 0.0038783461786806583, + 0.0032794089056551456, + 0.004249420017004013, + 0.0039525399915874004, + 0.0038086071144789457, + 0.004030449781566858, + 0.004544436931610107, + 0.002761396812275052, + 0.003443001303821802, + 0.002836079802364111, + 0.003595770103856921, + 0.005140000954270363, + 0.004108448512852192, + 0.004277477506548166, + 0.00335253169760108, + 0.002017994411289692, + 0.002051855903118849, + 0.0019225451396778226, + 0.0017856581835076213, + 0.0025555440224707127, + 0.0018079159781336784, + 0.002561039524152875, + 0.0025269445031881332, + 0.001751335454173386, + 0.0015210662968456745, + 0.0014507288578897715, + 0.001556338625960052, + 0.0014559044502675533, + 0.0012094565900042653, + 0.001981099834665656, + 0.0014296731678768992, + 0.0010493090376257896, + 0.0010854407446458936, + 0.0018775931093841791, + 0.0014929514145478606, + 0.0016335054533556104, + 0.0014164545573294163, + 0.001473889802582562, + 0.001699745887890458, + 0.0006748999003320932, + 0.0030670729465782642, + 0.005559167359024286, + 0.0037126783281564713, + 0.006673785857856274, + 0.0033882921561598778, + 0.00259441533125937, + 0.004056011792272329, + 0.0019518061308190227, + 0.004330059979110956, + 0.008336219936609268, + 0.004709069151431322, + 0.005016071256250143, + 0.0040637431666255, + 0.00456266338005662, + 0.004030694719403982, + 0.003046755911782384, + 0.0591825433075428, + 0.10441336035728455, + 0.08321712911128998, + 0.06708478182554245, + 0.08196363598108292, + 0.11739810556173325, + 0.062292877584695816, + 0.1073853150010109, + 0.01479195524007082, + 0.02578313834965229, + 0.014398027211427689, + 0.010874251835048199, + 0.020213250070810318, + 0.012277255766093731, + 0.012491094879806042, + 0.016476547345519066, + 0.003167178016155958, + 0.004264977760612965, + 0.0025418605655431747, + 0.0021561344619840384, + 0.0035785974469035864, + 0.002064397791400552, + 0.001903597847558558, + 0.001734680263325572 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 8, + 6 + ], + "selected_reward": -0.0067427014000713825 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a sharp side-angle macro view of a perfectly spherical, crystalline water droplet?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the droplet resting on a high-gloss, hydrophobic, invisible paint protection film over a deep midnight blue car body with metallic particles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the droplet slide gracefully and silkily down the curve, leaving a nearly invisible microscopic wet trail?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene illuminated by bright, even studio overhead lighting that creates a dazzling highlight on the droplet's top?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera slowly tilt downwards to follow the movement of the droplet?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.026351187378168106 + ], + "opposite_logprob_tokens": [ + -3.6513512134552 + ], + "logprob_reward": -0.026351187378168106, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.026351187378168106, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14972327649593353, + 0.06804317981004715, + 0.12661918997764587, + 0.07545168697834015, + 0.04150906205177307, + 0.03282898664474487, + 0.07974883168935776, + 0.2679094076156616, + 0.06537481397390366, + 0.09279157966375351 + ], + "credit_map": [ + 0.02338363043963909, + 0.020396223291754723, + 0.017898758873343468, + 0.016415227204561234, + 0.018558228388428688, + 0.02078677900135517, + 0.023112423717975616, + 0.009172002784907818, + 0.006730356719344854, + 0.006893544923514128, + 0.011272898875176907, + 0.010133162140846252, + 0.009190673008561134, + 0.010991987772285938, + 0.008861996233463287, + 0.0039685554802417755, + 0.01716306060552597, + 0.014435243792831898, + 0.013270048424601555, + 0.014648565091192722, + 0.012904861010611057, + 0.02018437348306179, + 0.015238934196531773, + 0.018774092197418213, + 0.011454363353550434, + 0.009192290715873241, + 0.010111377574503422, + 0.007858955301344395, + 0.006318827625364065, + 0.013106364756822586, + 0.0088125541806221, + 0.008596948347985744, + 0.0030434506479650736, + 0.00555019173771143, + 0.004901689477264881, + 0.006646689958870411, + 0.006895395927131176, + 0.004137793090194464, + 0.008168923668563366, + 0.0021649247501045465, + 0.0026079185772687197, + 0.00533130532130599, + 0.004413701593875885, + 0.00612536771222949, + 0.004590413998812437, + 0.002957671880722046, + 0.004831627011299133, + 0.0019709791522473097, + 0.010176691226661205, + 0.018621182069182396, + 0.009571380913257599, + 0.009764249436557293, + 0.009373231790959835, + 0.00986760389059782, + 0.0071601360104978085, + 0.005214361008256674, + 0.022559097036719322, + 0.04388022422790527, + 0.03165130689740181, + 0.02664007991552353, + 0.033643413335084915, + 0.045185476541519165, + 0.025554314255714417, + 0.03879551962018013, + 0.007321602664887905, + 0.013084431178867817, + 0.008393613621592522, + 0.005564530845731497, + 0.00968244019895792, + 0.006190293934196234, + 0.007270829286426306, + 0.007867074571549892, + 0.013175707310438156, + 0.019144896417856216, + 0.011100482195615768, + 0.00903678871691227, + 0.012476852163672447, + 0.009326902218163013, + 0.00838438980281353, + 0.010145560838282108 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 0, + 2 + ], + "selected_reward": -0.026351187378168106 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a sharp side-angle macro view of a perfectly spherical, crystalline water droplet?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the droplet resting on a high-gloss, hydrophobic, invisible paint protection film over a deep midnight blue car body with metallic particles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the droplet slide gracefully and silkily down the curve, leaving a nearly invisible microscopic wet trail?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene illuminated by bright, even studio overhead lighting that creates a dazzling highlight on the droplet's top?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera slowly tilt downwards to follow the movement of the droplet?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.25208184123039246 + ], + "opposite_logprob_tokens": [ + -1.5020818710327148 + ], + "logprob_reward": -0.25208184123039246, + "yes_no_margin_reward": 1.25, + "token_credit_reward": -0.25208184123039246, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09766658395528793, + 0.03939857333898544, + 0.025168586522340775, + 0.01729428581893444, + 0.0339486263692379, + 0.0771726742386818, + 0.057442329823970795, + 0.31243330240249634, + 0.16273890435695648, + 0.17673620581626892 + ], + "credit_map": [ + 0.012702742591500282, + 0.008460001088678837, + 0.015424289740622044, + 0.01710217073559761, + 0.012630429118871689, + 0.01099120918661356, + 0.014154129661619663, + 0.006201602052897215, + 0.003915914334356785, + 0.0034748222678899765, + 0.005228631664067507, + 0.006958378944545984, + 0.0064880503341555595, + 0.005643051583319902, + 0.005075907334685326, + 0.002613816177472472, + 0.003503663931041956, + 0.0031229674350470304, + 0.0032110903412103653, + 0.0032403634395450354, + 0.0037281708791851997, + 0.00442082853987813, + 0.0022352528758347034, + 0.001706247334368527, + 0.0014898208901286125, + 0.002747540595009923, + 0.0025052179116755724, + 0.002784339478239417, + 0.0018030829960480332, + 0.00211503729224205, + 0.002799473935738206, + 0.0010497723706066608, + 0.003218125319108367, + 0.005972389131784439, + 0.003790494753047824, + 0.005328697618097067, + 0.004805472679436207, + 0.0031338531989604235, + 0.005615123547613621, + 0.002084472682327032, + 0.005631349980831146, + 0.015499861910939217, + 0.009548941627144814, + 0.014542154036462307, + 0.011138465255498886, + 0.007218014914542437, + 0.009363089688122272, + 0.004230794031172991, + 0.006876755505800247, + 0.008746643550693989, + 0.008024351671338081, + 0.010755296796560287, + 0.006699004210531712, + 0.004691986832767725, + 0.007854228839278221, + 0.003794055664911866, + 0.032896023243665695, + 0.040808942168951035, + 0.04416041076183319, + 0.03323003277182579, + 0.03891851752996445, + 0.04509183391928673, + 0.028124207630753517, + 0.049203336238861084, + 0.019475292414426804, + 0.02994072064757347, + 0.020359400659799576, + 0.013580852188169956, + 0.02673339657485485, + 0.015077527612447739, + 0.014560342766344547, + 0.023011375218629837, + 0.020662520080804825, + 0.03502795100212097, + 0.020712530240416527, + 0.01849626563489437, + 0.02550443820655346, + 0.022799918428063393, + 0.020325154066085815, + 0.013207421638071537 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 9, + 8 + ], + "selected_reward": -0.25208184123039246 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000034/group_6/index_tensor([281, 281])_group_6_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000034/group_6/index_tensor([281, 281])_group_6_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..279038150697b0854131008a0c0b13573b8f5ecb --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000034/group_6/index_tensor([281, 281])_group_6_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "In an extreme close-up, shot handheld for a subtle and natural shakiness, the camera slowly tilts downward, following the graceful, slow rotation and fall of a minimalist wireless earbud case. The case is crafted from a matte, space gray polymer with a nearly imperceptible fine-grained texture, set against a pure, out-of-focus, deep black background. A precise and dramatic, cool sidelight shines from the upper left, creating a sharp, bright highlight along the rounded edge. Each slow rotation of the light allows the light to flow along its smooth curves, highlighting the flawless industrial design. A soft, reflected light subtly illuminates the other side, revealing rich shadows and material details without detracting from the overall depth. The image, shrouded in a cool, luxurious monochrome, exudes a focused, refined, and technologically modern aesthetic, as if time has been slowed down to showcase the object's exquisite craftsmanship and pure form.", + "seed": 2514, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000034/group_6/index_281_seed_2514_dp_12_rank_24_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 6, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.2489060267806053, + "vlm_logprob": -0.2489060267806053, + "vlm_yes_no_margin": 2.199999952316284, + "vlm_token_credit": -0.2489060267806053, + "vlm_yes_ratio": 0.8, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a minimalist wireless earbud case?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the case made of matte space gray polymer with a fine-grained texture?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the background a pure, out-of-focus, deep black?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a cool sidelight shine from the upper left creating a sharp highlight on the rounded edge?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera slowly tilt downward while the case rotates and falls?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.035016026347875595 + ], + "opposite_logprob_tokens": [ + -3.6600160598754883 + ], + "logprob_reward": -0.035016026347875595, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.035016026347875595, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1311882883310318, + 0.08397182077169418, + 0.137460395693779, + 0.04409676417708397, + 0.12040919065475464, + 0.10454461723566055, + 0.11699759215116501, + 0.050294533371925354, + 0.16948935389518738, + 0.041547421365976334 + ], + "credit_map": [ + 0.015349142253398895, + 0.01842045970261097, + 0.018683508038520813, + 0.017636839300394058, + 0.011356084607541561, + 0.01684185490012169, + 0.015558620914816856, + 0.01734178327023983, + 0.008129081688821316, + 0.008621998131275177, + 0.011871728114783764, + 0.015334554947912693, + 0.006596220191568136, + 0.013077905401587486, + 0.011765382252633572, + 0.008574954234063625, + 0.012759382836520672, + 0.010508055798709393, + 0.03381524980068207, + 0.01933482103049755, + 0.009958208538591862, + 0.026266148313879967, + 0.012580035254359245, + 0.01223849132657051, + 0.0042709303088486195, + 0.005443986505270004, + 0.0070996591821312904, + 0.008004829287528992, + 0.003542115679010749, + 0.0054996623657643795, + 0.006153255235403776, + 0.004082330968230963, + 0.007825387641787529, + 0.011775768361985683, + 0.032536301761865616, + 0.013597624376416206, + 0.006848837714642286, + 0.028157461434602737, + 0.011655616573989391, + 0.0080121960490942, + 0.020183848217129707, + 0.00783198606222868, + 0.01608155481517315, + 0.015040019527077675, + 0.006077897734940052, + 0.01393886934965849, + 0.01657131314277649, + 0.008819127455353737, + 0.013752947561442852, + 0.009557221084833145, + 0.017015699297189713, + 0.013299990445375443, + 0.006955370772629976, + 0.014405961148440838, + 0.018649183213710785, + 0.02336122654378414, + 0.007563652936369181, + 0.003736383281648159, + 0.00677068904042244, + 0.007916408590972424, + 0.0037913103587925434, + 0.008401399478316307, + 0.0070093912072479725, + 0.00510529475286603, + 0.014312569983303547, + 0.02859382890164852, + 0.023026471957564354, + 0.019828511402010918, + 0.011950417421758175, + 0.021873490884900093, + 0.028029659762978554, + 0.021874399855732918, + 0.004127999301999807, + 0.003886153921484947, + 0.006554299965500832, + 0.005600007716566324, + 0.003851077286526561, + 0.0047583854757249355, + 0.006551485974341631, + 0.006218011025339365 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 2, + 0 + ], + "selected_reward": -0.035016026347875595 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a minimalist wireless earbud case?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the case made of matte space gray polymer with a fine-grained texture?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the background a pure, out-of-focus, deep black?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a cool sidelight shine from the upper left creating a sharp highlight on the rounded edge?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera slowly tilt downward while the case rotates and falls?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.1130433976650238 + ], + "opposite_logprob_tokens": [ + -2.2380433082580566 + ], + "logprob_reward": -0.1130433976650238, + "yes_no_margin_reward": 2.125, + "token_credit_reward": -0.1130433976650238, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1712937355041504, + 0.23242995142936707, + 0.049159739166498184, + 0.0426483228802681, + 0.1283130645751953, + 0.04574822634458542, + 0.09588078409433365, + 0.06268342584371567, + 0.13200931251049042, + 0.039833419024944305 + ], + "credit_map": [ + 0.026842422783374786, + 0.019235171377658844, + 0.03827277943491936, + 0.013455426320433617, + 0.012134125456213951, + 0.029596274718642235, + 0.014328298158943653, + 0.017429260537028313, + 0.016397958621382713, + 0.024123823270201683, + 0.04377945512533188, + 0.04273456335067749, + 0.016311634331941605, + 0.042241163551807404, + 0.02812548726797104, + 0.018715865910053253, + 0.0043525914661586285, + 0.003763513872399926, + 0.009392878971993923, + 0.0074848891235888, + 0.003854768816381693, + 0.010100343264639378, + 0.004724458325654268, + 0.0054862950928509235, + 0.005177000537514687, + 0.0036866965238004923, + 0.006567039992660284, + 0.006170213222503662, + 0.0028275465592741966, + 0.005528443958610296, + 0.005759690888226032, + 0.006931692361831665, + 0.012177283875644207, + 0.013302222825586796, + 0.0315624363720417, + 0.01697002537548542, + 0.006536003202199936, + 0.021709656342864037, + 0.014016838744282722, + 0.012038596905767918, + 0.005870495922863483, + 0.00232522701844573, + 0.011243750341236591, + 0.005388197489082813, + 0.0021923885215073824, + 0.005854762624949217, + 0.005415002349764109, + 0.007458401843905449, + 0.00793521385639906, + 0.009297986514866352, + 0.0203428752720356, + 0.009920574724674225, + 0.006437893025577068, + 0.013586210086941719, + 0.015275754034519196, + 0.013084277510643005, + 0.00928041897714138, + 0.0037885888013988733, + 0.012163634411990643, + 0.012515993788838387, + 0.0035342364571988583, + 0.007717175409197807, + 0.007976130582392216, + 0.005707251373678446, + 0.014623777009546757, + 0.02218073047697544, + 0.020468033850193024, + 0.013156993314623833, + 0.007936804555356503, + 0.017481839284300804, + 0.020707381889224052, + 0.015453750267624855, + 0.0035284296609461308, + 0.005615399219095707, + 0.0061961328610777855, + 0.004246379714459181, + 0.00358951510861516, + 0.004840380512177944, + 0.005515806842595339, + 0.006301375571638346 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 8 + ], + "selected_reward": -0.1130433976650238 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a minimalist wireless earbud case?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the case made of matte space gray polymer with a fine-grained texture?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the background a pure, out-of-focus, deep black?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a cool sidelight shine from the upper left creating a sharp highlight on the rounded edge?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera slowly tilt downward while the case rotates and falls?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -0.8983523845672607 + ], + "opposite_logprob_tokens": [ + -0.5233523845672607 + ], + "logprob_reward": -0.8983523845672607, + "yes_no_margin_reward": -0.375, + "token_credit_reward": -0.8983523845672607, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.3630828857421875, + 0.09967997670173645, + 0.145620658993721, + 0.05239742994308472, + 0.09497009962797165, + 0.10501837730407715, + 0.05286780372262001, + 0.026595665141940117, + 0.04210247844457626, + 0.017664631828665733 + ], + "credit_map": [ + 0.04272453114390373, + 0.0416061133146286, + 0.08531314879655838, + 0.03225572407245636, + 0.032264139503240585, + 0.05819358304142952, + 0.03377659618854523, + 0.03694906085729599, + 0.008295109495520592, + 0.010397464036941528, + 0.027261657640337944, + 0.00869616586714983, + 0.004300946835428476, + 0.020050449296832085, + 0.0076722269877791405, + 0.013005952350795269, + 0.026967868208885193, + 0.00786213856190443, + 0.03517867252230644, + 0.01063031516969204, + 0.006472054403275251, + 0.03928203880786896, + 0.007706522010266781, + 0.011521033942699432, + 0.009941108524799347, + 0.003587981453165412, + 0.01429411955177784, + 0.006305858958512545, + 0.002324759727343917, + 0.0076108709909021854, + 0.003440506523475051, + 0.004892223048955202, + 0.015323062427341938, + 0.0052971928380429745, + 0.028004052117466927, + 0.007700690068304539, + 0.0032728221267461777, + 0.02191833406686783, + 0.006748868152499199, + 0.006705073174089193, + 0.023187756538391113, + 0.0062474352307617664, + 0.018062977120280266, + 0.014355501160025597, + 0.004482434596866369, + 0.014746093191206455, + 0.012580608949065208, + 0.011355554684996605, + 0.004725316539406776, + 0.002957871649414301, + 0.01665809005498886, + 0.005446045659482479, + 0.0033001541160047054, + 0.009530632756650448, + 0.005007770378142595, + 0.005241917911916971, + 0.004897174425423145, + 0.0015798135427758098, + 0.00459825387224555, + 0.004114444367587566, + 0.0014071052428334951, + 0.0037806753534823656, + 0.002691007684916258, + 0.003527189837768674, + 0.004315120168030262, + 0.006337969563901424, + 0.005361374467611313, + 0.005050152540206909, + 0.002744539873674512, + 0.006069835741072893, + 0.007150665391236544, + 0.005072816740721464, + 0.0018518593860790133, + 0.0024804852437227964, + 0.0027279844507575035, + 0.002509830053895712, + 0.0016354551771655679, + 0.0021490035578608513, + 0.0027919765561819077, + 0.001518036238849163 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 5 + ], + "selected_reward": -0.8983523845672607 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a minimalist wireless earbud case?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the case made of matte space gray polymer with a fine-grained texture?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the background a pure, out-of-focus, deep black?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a cool sidelight shine from the upper left creating a sharp highlight on the rounded edge?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera slowly tilt downward while the case rotates and falls?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.018204864114522934 + ], + "opposite_logprob_tokens": [ + -4.018204689025879 + ], + "logprob_reward": -0.018204864114522934, + "yes_no_margin_reward": 3.999999761581421, + "token_credit_reward": -0.018204864114522934, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3463791310787201, + 0.08090940117835999, + 0.12162104994058609, + 0.041395217180252075, + 0.07756488025188446, + 0.0988745242357254, + 0.04655330255627632, + 0.0501536950469017, + 0.08896271139383316, + 0.04758601263165474 + ], + "credit_map": [ + 0.04393282160162926, + 0.042197808623313904, + 0.10442866384983063, + 0.02057381346821785, + 0.017849572002887726, + 0.07137040048837662, + 0.015447665005922318, + 0.03057839535176754, + 0.008592595346271992, + 0.011899106204509735, + 0.015455882996320724, + 0.011569475755095482, + 0.005390768405050039, + 0.010450479574501514, + 0.009795667603611946, + 0.007755417842417955, + 0.008931217715144157, + 0.008223259821534157, + 0.030836498364806175, + 0.015303028747439384, + 0.010030405595898628, + 0.029819747433066368, + 0.010310412384569645, + 0.00816648080945015, + 0.0054404824040830135, + 0.003320920979604125, + 0.007908221334218979, + 0.008261080831289291, + 0.0023903455585241318, + 0.004854880273342133, + 0.005236244760453701, + 0.003983043599873781, + 0.008431816473603249, + 0.0059623257257044315, + 0.016061611473560333, + 0.01141529344022274, + 0.004807284101843834, + 0.00884799100458622, + 0.009378510527312756, + 0.012660049833357334, + 0.020879119634628296, + 0.006357063073664904, + 0.01774384081363678, + 0.012441417202353477, + 0.004644939675927162, + 0.013678502291440964, + 0.012799154035747051, + 0.010330481454730034, + 0.006904188543558121, + 0.0034375567920506, + 0.009506726637482643, + 0.005197422578930855, + 0.0032439688220620155, + 0.0070472219958901405, + 0.003962963353842497, + 0.007253252435475588, + 0.006127848289906979, + 0.0028155697509646416, + 0.011700736358761787, + 0.008194307796657085, + 0.002778839087113738, + 0.007693672552704811, + 0.0060872891917824745, + 0.004755432717502117, + 0.00773891806602478, + 0.014778208918869495, + 0.012144225649535656, + 0.010116391815245152, + 0.005956883076578379, + 0.01221462432295084, + 0.014066282659769058, + 0.011947172693908215, + 0.004148803185671568, + 0.0027704022359102964, + 0.00884217955172062, + 0.006593355908989906, + 0.0025020826142281294, + 0.01110704056918621, + 0.006002513226121664, + 0.005619638599455357 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 5 + ], + "selected_reward": -0.018204864114522934 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a minimalist wireless earbud case?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the case made of matte space gray polymer with a fine-grained texture?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the background a pure, out-of-focus, deep black?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a cool sidelight shine from the upper left creating a sharp highlight on the rounded edge?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera slowly tilt downward while the case rotates and falls?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.1799134612083435 + ], + "opposite_logprob_tokens": [ + -1.8049135208129883 + ], + "logprob_reward": -0.1799134612083435, + "yes_no_margin_reward": 1.625, + "token_credit_reward": -0.1799134612083435, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07889498770236969, + 0.09448453783988953, + 0.15649349987506866, + 0.06925451755523682, + 0.06163705885410309, + 0.13381589949131012, + 0.05008811503648758, + 0.11638560146093369, + 0.13113996386528015, + 0.10780581086874008 + ], + "credit_map": [ + 0.008964707143604755, + 0.008526290766894817, + 0.01148771122097969, + 0.00797541905194521, + 0.009358386509120464, + 0.008197265677154064, + 0.012310907244682312, + 0.012074297294020653, + 0.009560097008943558, + 0.008283601142466068, + 0.013058857060968876, + 0.017951449379324913, + 0.006906531751155853, + 0.012148541398346424, + 0.01340717263519764, + 0.013168287463486195, + 0.010968075133860111, + 0.009154210798442364, + 0.04550888016819954, + 0.015981605276465416, + 0.009787056595087051, + 0.04144463688135147, + 0.01282945740967989, + 0.010819579474627972, + 0.009396793320775032, + 0.006964270956814289, + 0.010540979914367199, + 0.012952672317624092, + 0.004761200863867998, + 0.008387962356209755, + 0.009984765201807022, + 0.006265876814723015, + 0.00794956460595131, + 0.003036602633073926, + 0.014814251102507114, + 0.005208190530538559, + 0.0018208208493888378, + 0.013685119338333607, + 0.004346140660345554, + 0.010776365175843239, + 0.02636481635272503, + 0.009244678542017937, + 0.020810311660170555, + 0.01780974306166172, + 0.007109616417437792, + 0.017845092341303825, + 0.020702864974737167, + 0.013928784057497978, + 0.007719256915152073, + 0.0025748903863132, + 0.01055912021547556, + 0.005064062308520079, + 0.004344423301517963, + 0.006598205771297216, + 0.006792378146201372, + 0.006435777992010117, + 0.017679531127214432, + 0.007015023846179247, + 0.01797778159379959, + 0.020496947690844536, + 0.007478002924472094, + 0.01589256152510643, + 0.016571251675486565, + 0.01327450480312109, + 0.013880755752325058, + 0.020173996686935425, + 0.019287237897515297, + 0.01591150276362896, + 0.007883087731897831, + 0.01487671211361885, + 0.020759064704179764, + 0.018367595970630646, + 0.008138692937791348, + 0.009002148173749447, + 0.020250877365469933, + 0.018891550600528717, + 0.009785084053874016, + 0.016063565388321877, + 0.01705751195549965, + 0.00861637108027935 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 5, + 8 + ], + "selected_reward": -0.1799134612083435 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "No", + "Yes", + "Yes" + ] + }, + { + "prompt": "In an extreme close-up, shot handheld for a subtle and natural shakiness, the camera slowly tilts downward, following the graceful, slow rotation and fall of a minimalist wireless earbud case. The case is crafted from a matte, space gray polymer with a nearly imperceptible fine-grained texture, set against a pure, out-of-focus, deep black background. A precise and dramatic, cool sidelight shines from the upper left, creating a sharp, bright highlight along the rounded edge. Each slow rotation of the light allows the light to flow along its smooth curves, highlighting the flawless industrial design. A soft, reflected light subtly illuminates the other side, revealing rich shadows and material details without detracting from the overall depth. The image, shrouded in a cool, luxurious monochrome, exudes a focused, refined, and technologically modern aesthetic, as if time has been slowed down to showcase the object's exquisite craftsmanship and pure form.", + "seed": 2515, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000034/group_6/index_281_seed_2515_dp_12_rank_24_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 6, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.23555343970656395, + "vlm_logprob": -0.23555343970656395, + "vlm_yes_no_margin": 2.425, + "vlm_token_credit": -0.23555343970656395, + "vlm_yes_ratio": 0.8, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a minimalist wireless earbud case?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the case made of matte space gray polymer with a fine-grained texture?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the background a pure, out-of-focus, deep black?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a cool sidelight shine from the upper left creating a sharp highlight on the rounded edge?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera slowly tilt downward while the case rotates and falls?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.020866278558969498 + ], + "opposite_logprob_tokens": [ + -4.645866394042969 + ], + "logprob_reward": -0.020866278558969498, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.020866278558969498, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.35350221395492554, + 0.07100030779838562, + 0.04461641609668732, + 0.03508754447102547, + 0.058836136013269424, + 0.04407208785414696, + 0.1542966514825821, + 0.06597532331943512, + 0.05003298819065094, + 0.12258035689592361 + ], + "credit_map": [ + 0.016877898946404457, + 0.027718765661120415, + 0.06568572670221329, + 0.02953723631799221, + 0.03926461189985275, + 0.10089842975139618, + 0.03629395738244057, + 0.03722558170557022, + 0.004476848989725113, + 0.010676056146621704, + 0.010329185985028744, + 0.005758413579314947, + 0.007246427237987518, + 0.006555791944265366, + 0.01453381311148405, + 0.011423777788877487, + 0.0023095845244824886, + 0.005869465880095959, + 0.007919986732304096, + 0.005581994540989399, + 0.007572645787149668, + 0.0031817785929888487, + 0.007288849912583828, + 0.004892112221568823, + 0.001986280083656311, + 0.005194382276386023, + 0.0059426696971058846, + 0.0068204994313418865, + 0.0039772288873791695, + 0.0024049377534538507, + 0.0038509562145918608, + 0.004910594318062067, + 0.0035482835955917835, + 0.00798568595200777, + 0.015005653724074364, + 0.009032216854393482, + 0.004945504479110241, + 0.0040359580889344215, + 0.007641511503607035, + 0.006641319487243891, + 0.00201496877707541, + 0.003906200872734189, + 0.013656383380293846, + 0.011745041236281395, + 0.0025244143325835466, + 0.0016601739916950464, + 0.0036088498309254646, + 0.004956054035574198, + 0.005846202373504639, + 0.027748169377446175, + 0.05213845148682594, + 0.025307554751634598, + 0.008087818510830402, + 0.007707344368100166, + 0.014941110275685787, + 0.01251998357474804, + 0.003948661033064127, + 0.00688185216858983, + 0.010298223234713078, + 0.011915026232600212, + 0.00694086030125618, + 0.0039367759600281715, + 0.014793804846704006, + 0.007260117679834366, + 0.0024736199993640184, + 0.0033284216187894344, + 0.011382275260984898, + 0.00827177707105875, + 0.008982713334262371, + 0.002429493935778737, + 0.00852335523813963, + 0.004641334060579538, + 0.006788307800889015, + 0.009487135335803032, + 0.018046662211418152, + 0.016216790303587914, + 0.02951423078775406, + 0.009182610549032688, + 0.020760681480169296, + 0.01258393656462431 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 6, + 9 + ], + "selected_reward": -0.020866278558969498 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a minimalist wireless earbud case?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the case made of matte space gray polymer with a fine-grained texture?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the background a pure, out-of-focus, deep black?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a cool sidelight shine from the upper left creating a sharp highlight on the rounded edge?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera slowly tilt downward while the case rotates and falls?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.20172107219696045 + ], + "opposite_logprob_tokens": [ + -1.7017210721969604 + ], + "logprob_reward": -0.20172107219696045, + "yes_no_margin_reward": 1.5, + "token_credit_reward": -0.20172107219696045, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.324564129114151, + 0.059375058859586716, + 0.034911494702100754, + 0.04260122776031494, + 0.05005314573645592, + 0.12460018694400787, + 0.164605051279068, + 0.04825975373387337, + 0.03406636416912079, + 0.11696360260248184 + ], + "credit_map": [ + 0.01912888139486313, + 0.019107259809970856, + 0.07040182501077652, + 0.025898868218064308, + 0.01826087012887001, + 0.1049247682094574, + 0.028091365471482277, + 0.03875027596950531, + 0.004663760308176279, + 0.007667808793485165, + 0.010523985140025616, + 0.006536971777677536, + 0.0060761976055800915, + 0.005526342429220676, + 0.011140795424580574, + 0.007239197846502066, + 0.002974065952003002, + 0.0062004756182432175, + 0.005364539101719856, + 0.0047948164865374565, + 0.004086113069206476, + 0.002503106603398919, + 0.004660268314182758, + 0.004328109323978424, + 0.0037975322920829058, + 0.004685057792812586, + 0.006585065275430679, + 0.005434355698525906, + 0.005705388728529215, + 0.0048685441724956036, + 0.005764724221080542, + 0.005760557949542999, + 0.0028353941161185503, + 0.01060602255165577, + 0.013403328135609627, + 0.006220532115548849, + 0.0027964136097580194, + 0.002994570415467024, + 0.005212380550801754, + 0.005984502378851175, + 0.0037048528902232647, + 0.02733626216650009, + 0.04667242243885994, + 0.021551212295889854, + 0.004539141431450844, + 0.003907093778252602, + 0.010110298171639442, + 0.006778908893465996, + 0.0052633024752140045, + 0.021939994767308235, + 0.04422985762357712, + 0.027836430817842484, + 0.010950044728815556, + 0.010952350683510303, + 0.02532372996211052, + 0.018109336495399475, + 0.00255516916513443, + 0.0037739898543804884, + 0.007963262498378754, + 0.005323741119354963, + 0.008463758043944836, + 0.003250859910622239, + 0.010602214373648167, + 0.006326758768409491, + 0.0019541047513484955, + 0.002464552642777562, + 0.006216197274625301, + 0.005936820525676012, + 0.0055062309838831425, + 0.002263641683384776, + 0.00574107700958848, + 0.003983738366514444, + 0.00774799007922411, + 0.0102424630895257, + 0.016184967011213303, + 0.016173584386706352, + 0.018932484090328217, + 0.014393620193004608, + 0.01883309707045555, + 0.014455391094088554 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 6, + 5 + ], + "selected_reward": -0.20172107219696045 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a minimalist wireless earbud case?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the case made of matte space gray polymer with a fine-grained texture?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the background a pure, out-of-focus, deep black?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a cool sidelight shine from the upper left creating a sharp highlight on the rounded edge?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera slowly tilt downward while the case rotates and falls?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -0.8261362314224243 + ], + "opposite_logprob_tokens": [ + -0.5761362314224243 + ], + "logprob_reward": -0.8261362314224243, + "yes_no_margin_reward": -0.25, + "token_credit_reward": -0.8261362314224243, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.3856903910636902, + 0.12821879982948303, + 0.05256727337837219, + 0.0642600730061531, + 0.09991991519927979, + 0.08573216199874878, + 0.07390822470188141, + 0.03965259715914726, + 0.02016078680753708, + 0.04988979175686836 + ], + "credit_map": [ + 0.014196420088410378, + 0.017160072922706604, + 0.09821358323097229, + 0.020488010719418526, + 0.016305923461914062, + 0.14483827352523804, + 0.02688044123351574, + 0.04760769382119179, + 0.007793146651238203, + 0.011964200995862484, + 0.02684730291366577, + 0.01500683929771185, + 0.009389459155499935, + 0.011674830690026283, + 0.025722680613398552, + 0.019820330664515495, + 0.003831299487501383, + 0.005423352587968111, + 0.00853707641363144, + 0.005730298347771168, + 0.011446216143667698, + 0.003589118830859661, + 0.005369197111576796, + 0.008640714921057224, + 0.004965437110513449, + 0.009846556931734085, + 0.014717353507876396, + 0.009277595207095146, + 0.006445605307817459, + 0.004011302255094051, + 0.0059776934795081615, + 0.009018528275191784, + 0.005177484359592199, + 0.019855642691254616, + 0.03305116295814514, + 0.018801627680659294, + 0.005128406919538975, + 0.003888620063662529, + 0.00711824418976903, + 0.006898729596287012, + 0.008492283523082733, + 0.012524583376944065, + 0.023032836616039276, + 0.017203671857714653, + 0.004541538655757904, + 0.004191395360976458, + 0.008273892104625702, + 0.007471967954188585, + 0.004175025038421154, + 0.008269060403108597, + 0.018270356580615044, + 0.01422051154077053, + 0.004179377108812332, + 0.004344993736594915, + 0.012300388887524605, + 0.008148507215082645, + 0.002306663664057851, + 0.003351921681314707, + 0.007118596229702234, + 0.006885794457048178, + 0.004410418216139078, + 0.0026886346749961376, + 0.007120419293642044, + 0.005770149175077677, + 0.001107283984310925, + 0.0017570680938661098, + 0.003628074424341321, + 0.0034444027114659548, + 0.0026060615200549364, + 0.0017052947077900171, + 0.0032115853391587734, + 0.002701016142964363, + 0.005391391925513744, + 0.0028410619124770164, + 0.007432600483298302, + 0.006164709106087685, + 0.008570191450417042, + 0.0034230828750878572, + 0.007535138167440891, + 0.008531617000699043 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 4 + ], + "selected_reward": -0.8261362314224243 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a minimalist wireless earbud case?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the case made of matte space gray polymer with a fine-grained texture?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the background a pure, out-of-focus, deep black?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a cool sidelight shine from the upper left creating a sharp highlight on the rounded edge?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera slowly tilt downward while the case rotates and falls?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.016082093119621277 + ], + "opposite_logprob_tokens": [ + -4.141082286834717 + ], + "logprob_reward": -0.016082093119621277, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.016082093119621277, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.28496354818344116, + 0.12878479063510895, + 0.046165745705366135, + 0.05451023206114769, + 0.10689995437860489, + 0.0834544450044632, + 0.09425771236419678, + 0.08552788197994232, + 0.039434581995010376, + 0.07600107789039612 + ], + "credit_map": [ + 0.018545903265476227, + 0.026414314284920692, + 0.05474790185689926, + 0.02494611032307148, + 0.02043517306447029, + 0.07422934472560883, + 0.030550574883818626, + 0.035094261169433594, + 0.010916347615420818, + 0.015020838938653469, + 0.01909375935792923, + 0.012499634176492691, + 0.014571760781109333, + 0.011210714466869831, + 0.024591343477368355, + 0.020880401134490967, + 0.0033532509114593267, + 0.00637742318212986, + 0.00723431259393692, + 0.004850989207625389, + 0.007583730388432741, + 0.002822075504809618, + 0.0059720356948673725, + 0.0079719303175807, + 0.003662419505417347, + 0.0056950850412249565, + 0.010176420211791992, + 0.01144613791257143, + 0.0064381202682852745, + 0.0025565207470208406, + 0.005101982969790697, + 0.009433547034859657, + 0.007334189955145121, + 0.022982921451330185, + 0.03630426526069641, + 0.01833343505859375, + 0.003994518890976906, + 0.0035261220764368773, + 0.006494969129562378, + 0.007929540239274502, + 0.003737384220585227, + 0.021316148340702057, + 0.028218990191817284, + 0.010030784644186497, + 0.004843361675739288, + 0.00324462098069489, + 0.006819532252848148, + 0.005243632011115551, + 0.0034380266442894936, + 0.018522031605243683, + 0.028421131893992424, + 0.01335730217397213, + 0.0063016945496201515, + 0.0056716096587479115, + 0.01072742324322462, + 0.007818497717380524, + 0.006341052707284689, + 0.009306357242166996, + 0.014340752735733986, + 0.013430951163172722, + 0.010613299906253815, + 0.00528203509747982, + 0.014039572328329086, + 0.012173867784440517, + 0.003094170242547989, + 0.003274473361670971, + 0.007222525775432587, + 0.005457363557070494, + 0.004875880666077137, + 0.003642380004748702, + 0.006549005396664143, + 0.005318786017596722, + 0.005269980523735285, + 0.004966334439814091, + 0.012885729782283306, + 0.013068119995296001, + 0.009771851822733879, + 0.008719299919903278, + 0.011594423092901707, + 0.009725343436002731 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 4 + ], + "selected_reward": -0.016082093119621277 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a minimalist wireless earbud case?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the case made of matte space gray polymer with a fine-grained texture?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the background a pure, out-of-focus, deep black?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a cool sidelight shine from the upper left creating a sharp highlight on the rounded edge?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera slowly tilt downward while the case rotates and falls?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.11296152323484421 + ], + "opposite_logprob_tokens": [ + -2.237961530685425 + ], + "logprob_reward": -0.11296152323484421, + "yes_no_margin_reward": 2.125, + "token_credit_reward": -0.11296152323484421, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09579284489154816, + 0.10324886441230774, + 0.068858802318573, + 0.05254815146327019, + 0.15001361072063446, + 0.1897132694721222, + 0.13191115856170654, + 0.12581492960453033, + 0.029360927641391754, + 0.05273749306797981 + ], + "credit_map": [ + 0.005543033592402935, + 0.010343782603740692, + 0.013862730003893375, + 0.011738277971744537, + 0.009509780444204807, + 0.02037750743329525, + 0.013188175857067108, + 0.011229556053876877, + 0.005317485425621271, + 0.020472366362810135, + 0.015782825648784637, + 0.011679942719638348, + 0.010976689867675304, + 0.010285714641213417, + 0.017851833254098892, + 0.010882005095481873, + 0.00431219395250082, + 0.007299012504518032, + 0.011380625888705254, + 0.006393409799784422, + 0.018604271113872528, + 0.0042297556065022945, + 0.008122362196445465, + 0.008517177775502205, + 0.004216890782117844, + 0.0070220232009887695, + 0.009714847430586815, + 0.008875854313373566, + 0.004921031184494495, + 0.003815515898168087, + 0.0057800887152552605, + 0.008201899006962776, + 0.00844122376292944, + 0.0218519177287817, + 0.03585134074091911, + 0.020835772156715393, + 0.011531619355082512, + 0.010311463847756386, + 0.02037608064711094, + 0.020814193412661552, + 0.006254065316170454, + 0.05222676694393158, + 0.07232797890901566, + 0.02335899882018566, + 0.008994613774120808, + 0.004484097473323345, + 0.013352264650166035, + 0.008714481256902218, + 0.005230973940342665, + 0.02575746178627014, + 0.05083058401942253, + 0.02495145983994007, + 0.004650670103728771, + 0.004004478454589844, + 0.009764411486685276, + 0.006721121724694967, + 0.00762189133092761, + 0.01273905485868454, + 0.01929224655032158, + 0.01756756380200386, + 0.012363015674054623, + 0.007968167774379253, + 0.02754085883498192, + 0.020722122862935066, + 0.0022393614053726196, + 0.0029630998615175486, + 0.004857019521296024, + 0.005342685151845217, + 0.0033313583116978407, + 0.00214035133831203, + 0.00510774040594697, + 0.0033793114125728607, + 0.003483918495476246, + 0.004355241544544697, + 0.008665644563734531, + 0.00949897337704897, + 0.007901602424681187, + 0.0043286108411848545, + 0.007539455778896809, + 0.006964047439396381 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 4, + 6 + ], + "selected_reward": -0.11296152323484421 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "No", + "Yes", + "Yes" + ] + }, + { + "prompt": "In an extreme close-up, shot handheld for a subtle and natural shakiness, the camera slowly tilts downward, following the graceful, slow rotation and fall of a minimalist wireless earbud case. The case is crafted from a matte, space gray polymer with a nearly imperceptible fine-grained texture, set against a pure, out-of-focus, deep black background. A precise and dramatic, cool sidelight shines from the upper left, creating a sharp, bright highlight along the rounded edge. Each slow rotation of the light allows the light to flow along its smooth curves, highlighting the flawless industrial design. A soft, reflected light subtly illuminates the other side, revealing rich shadows and material details without detracting from the overall depth. The image, shrouded in a cool, luxurious monochrome, exudes a focused, refined, and technologically modern aesthetic, as if time has been slowed down to showcase the object's exquisite craftsmanship and pure form.", + "seed": 2516, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000034/group_6/index_281_seed_2516_dp_13_rank_26_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 6, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.06797616956755519, + "vlm_logprob": -0.06797616956755519, + "vlm_yes_no_margin": 3.275, + "vlm_token_credit": -0.06797616956755519, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a minimalist wireless earbud case?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the case made of matte space gray polymer with a fine-grained texture?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the background a pure, out-of-focus, deep black?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a cool sidelight shine from the upper left creating a sharp highlight on the rounded edge?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera slowly tilt downward while the case rotates and falls?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04053112119436264 + ], + "opposite_logprob_tokens": [ + -3.6655311584472656 + ], + "logprob_reward": -0.04053112119436264, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.04053112119436264, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1884392499923706, + 0.12487416714429855, + 0.06954149156808853, + 0.022735023871064186, + 0.06872475147247314, + 0.08707218617200851, + 0.02816367894411087, + 0.0474226176738739, + 0.0925430878996849, + 0.27048376202583313 + ], + "credit_map": [ + 0.0060590957291424274, + 0.07622078061103821, + 0.005961745511740446, + 0.02274421975016594, + 0.036528438329696655, + 0.010672739706933498, + 0.008811897598206997, + 0.021440330892801285, + 0.008568418212234974, + 0.04067332670092583, + 0.006783547345548868, + 0.014153484255075455, + 0.02095598727464676, + 0.010051094926893711, + 0.011332236230373383, + 0.012356068938970566, + 0.005789005197584629, + 0.021683180704712868, + 0.004012641962617636, + 0.00930050853639841, + 0.01225017849355936, + 0.005264512728899717, + 0.004269363358616829, + 0.006972101517021656, + 0.003223058069124818, + 0.003534630872309208, + 0.002753437263891101, + 0.002720313146710396, + 0.003204710315912962, + 0.002614089986309409, + 0.0019197838846594095, + 0.0027650000993162394, + 0.013671248219907284, + 0.014716524630784988, + 0.006184285506606102, + 0.004871591459959745, + 0.00835453812032938, + 0.007649334147572517, + 0.004019332583993673, + 0.009257898665964603, + 0.010817574337124825, + 0.019069496542215347, + 0.006163904909044504, + 0.005389775615185499, + 0.014179964549839497, + 0.009034202434122562, + 0.010007567703723907, + 0.012409703806042671, + 0.0029674379620701075, + 0.0026185361202806234, + 0.00219272100366652, + 0.002397173782810569, + 0.003989508841186762, + 0.004259075038135052, + 0.006002264097332954, + 0.003736962331458926, + 0.004203329794108868, + 0.0032392549328505993, + 0.0033955040853470564, + 0.003192282747477293, + 0.0058324020355939865, + 0.009090457111597061, + 0.008452700451016426, + 0.010016684420406818, + 0.006267454009503126, + 0.009113180451095104, + 0.00748782092705369, + 0.01335144229233265, + 0.010229635052382946, + 0.015699198469519615, + 0.01314424630254507, + 0.01725011318922043, + 0.019203094765543938, + 0.027536122128367424, + 0.023419329896569252, + 0.03435777872800827, + 0.04210523143410683, + 0.045781828463077545, + 0.027259834110736847, + 0.05082055181264877 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 0, + 1 + ], + "selected_reward": -0.04053112119436264 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a minimalist wireless earbud case?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the case made of matte space gray polymer with a fine-grained texture?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the background a pure, out-of-focus, deep black?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a cool sidelight shine from the upper left creating a sharp highlight on the rounded edge?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera slowly tilt downward while the case rotates and falls?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.16053573787212372 + ], + "opposite_logprob_tokens": [ + -1.9105356931686401 + ], + "logprob_reward": -0.16053573787212372, + "yes_no_margin_reward": 1.75, + "token_credit_reward": -0.16053573787212372, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1178557425737381, + 0.2899706959724426, + 0.17914432287216187, + 0.059773560613393784, + 0.03178982436656952, + 0.03862485662102699, + 0.01293969340622425, + 0.04357985779643059, + 0.09685663878917694, + 0.12946484982967377 + ], + "credit_map": [ + 0.004856097046285868, + 0.038445986807346344, + 0.007600296754390001, + 0.015582033433020115, + 0.02053690329194069, + 0.009122608229517937, + 0.011247961781919003, + 0.010463857091963291, + 0.020230066031217575, + 0.06586020439863205, + 0.02431306801736355, + 0.03746761754155159, + 0.04555991291999817, + 0.02691199816763401, + 0.04064938798546791, + 0.028978435322642326, + 0.016493143513798714, + 0.060559313744306564, + 0.006150959990918636, + 0.02251860313117504, + 0.0343293771147728, + 0.013883656822144985, + 0.008682718500494957, + 0.01652655564248562, + 0.006653452757745981, + 0.013449746184051037, + 0.004941868595778942, + 0.006402332801371813, + 0.009366165846586227, + 0.006526948884129524, + 0.006384878419339657, + 0.006048165261745453, + 0.006678411271423101, + 0.004087152425199747, + 0.0038300021551549435, + 0.003220162820070982, + 0.0029139237012714148, + 0.003700743429362774, + 0.002503414638340473, + 0.004856012295931578, + 0.004562539514154196, + 0.00351532525382936, + 0.0019482686184346676, + 0.0021430582273751497, + 0.004949785768985748, + 0.007892238907516003, + 0.003825690131634474, + 0.009787949733436108, + 0.0020924988202750683, + 0.0029005648102611303, + 0.0005478985258378088, + 0.0006031993543729186, + 0.0018246256513521075, + 0.0014072629855945706, + 0.0014248363440856338, + 0.00213880673982203, + 0.005068918690085411, + 0.0027647879905998707, + 0.002675876021385193, + 0.0024116116110235453, + 0.0045297215692698956, + 0.011520744301378727, + 0.00545274093747139, + 0.0091554569080472, + 0.006459261756390333, + 0.008600831031799316, + 0.006547438446432352, + 0.012747433967888355, + 0.012960267253220081, + 0.017619680613279343, + 0.011730538681149483, + 0.02019118145108223, + 0.00864261481910944, + 0.012063534930348396, + 0.009392943233251572, + 0.016540411859750748, + 0.017504224553704262, + 0.025587741285562515, + 0.012806002981960773, + 0.026927389204502106 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 2, + 9 + ], + "selected_reward": -0.16053573787212372 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a minimalist wireless earbud case?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the case made of matte space gray polymer with a fine-grained texture?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the background a pure, out-of-focus, deep black?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a cool sidelight shine from the upper left creating a sharp highlight on the rounded edge?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera slowly tilt downward while the case rotates and falls?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01821376010775566 + ], + "opposite_logprob_tokens": [ + -4.018213748931885 + ], + "logprob_reward": -0.01821376010775566, + "yes_no_margin_reward": 4.0, + "token_credit_reward": -0.01821376010775566, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.22441992163658142, + 0.031803518533706665, + 0.03407597169280052, + 0.08320736140012741, + 0.016892893239855766, + 0.05161447823047638, + 0.028344981372356415, + 0.05264715477824211, + 0.33251020312309265, + 0.14448361098766327 + ], + "credit_map": [ + 0.008312670513987541, + 0.08527093380689621, + 0.007445930037647486, + 0.02798539586365223, + 0.03754454478621483, + 0.01571814715862274, + 0.009096053428947926, + 0.033046215772628784, + 0.001533813076093793, + 0.00895508099347353, + 0.002631623763591051, + 0.004723696503788233, + 0.00455856416374445, + 0.0026995691005140543, + 0.0031078853644430637, + 0.0035932790488004684, + 0.0036490410566329956, + 0.007200727704912424, + 0.0023875536862760782, + 0.004871871788054705, + 0.005155840422958136, + 0.003318053437396884, + 0.003147262381389737, + 0.004345615394413471, + 0.008692879229784012, + 0.019509801641106606, + 0.006527129095047712, + 0.009042613208293915, + 0.012004283256828785, + 0.008930086158216, + 0.007868167944252491, + 0.010632392019033432, + 0.0028817325364798307, + 0.0016507181571796536, + 0.0015806038863956928, + 0.0021213367581367493, + 0.002422022633254528, + 0.0020148607436567545, + 0.0021071028895676136, + 0.0021145129576325417, + 0.0058874404057860374, + 0.012316099368035793, + 0.002397976815700531, + 0.0030676168389618397, + 0.005385595839470625, + 0.006172557827085257, + 0.00436168210580945, + 0.012025498785078526, + 0.0031702613923698664, + 0.00198492594063282, + 0.001068417332135141, + 0.0013024399522691965, + 0.0024560370948165655, + 0.008661986328661442, + 0.0027358513325452805, + 0.0069650569930672646, + 0.007927551865577698, + 0.0033152613323181868, + 0.005055075045675039, + 0.0036351624876260757, + 0.006229253485798836, + 0.00988644827157259, + 0.006956928875297308, + 0.009641469456255436, + 0.01884627155959606, + 0.03410043939948082, + 0.02325686812400818, + 0.046843815594911575, + 0.04237524792551994, + 0.06401986628770828, + 0.03244354948401451, + 0.07062407582998276, + 0.010330058634281158, + 0.011775223538279533, + 0.010666463524103165, + 0.022662071511149406, + 0.024482963606715202, + 0.025723475962877274, + 0.013580415397882462, + 0.025262925773859024 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 0, + 9 + ], + "selected_reward": -0.01821376010775566 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a minimalist wireless earbud case?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the case made of matte space gray polymer with a fine-grained texture?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the background a pure, out-of-focus, deep black?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a cool sidelight shine from the upper left creating a sharp highlight on the rounded edge?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera slowly tilt downward while the case rotates and falls?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.007639241870492697 + ], + "opposite_logprob_tokens": [ + -4.882639408111572 + ], + "logprob_reward": -0.007639241870492697, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.007639241870492697, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.16634517908096313, + 0.1304890513420105, + 0.03033347614109516, + 0.08807278424501419, + 0.03262548893690109, + 0.11498674005270004, + 0.12706370651721954, + 0.07681029289960861, + 0.06107135862112045, + 0.17220187187194824 + ], + "credit_map": [ + 0.006356685888022184, + 0.04895799607038498, + 0.005952440667897463, + 0.01980053074657917, + 0.020675810053944588, + 0.020659659057855606, + 0.008882313035428524, + 0.03505976125597954, + 0.007160969544202089, + 0.03788580000400543, + 0.007708802353590727, + 0.015512100420892239, + 0.02104548178613186, + 0.011974625289440155, + 0.015232139267027378, + 0.013969146646559238, + 0.003966711927205324, + 0.006623708177357912, + 0.002450522966682911, + 0.0032995049841701984, + 0.0052214753814041615, + 0.003054883563891053, + 0.0026802190113812685, + 0.0030364531558007, + 0.01124399621039629, + 0.013421832583844662, + 0.00826053973287344, + 0.009034208953380585, + 0.010217982344329357, + 0.01242354791611433, + 0.009797831065952778, + 0.013672847300767899, + 0.006969746667891741, + 0.0038093619514256716, + 0.0035854862071573734, + 0.0031569323036819696, + 0.004164823796600103, + 0.003591300453990698, + 0.0024355705827474594, + 0.0049122697673738, + 0.007941478863358498, + 0.016189711168408394, + 0.0037508844397962093, + 0.0035166938323527575, + 0.010518092662096024, + 0.026047514751553535, + 0.008263597264885902, + 0.0387587808072567, + 0.010581018403172493, + 0.0073814550414681435, + 0.007203489542007446, + 0.005609150975942612, + 0.011829717084765434, + 0.037476565688848495, + 0.013543480075895786, + 0.033438846468925476, + 0.007968953810632229, + 0.0040754517540335655, + 0.005206701345741749, + 0.003743142122402787, + 0.007869887165725231, + 0.020934933796525, + 0.009637297131121159, + 0.017373934388160706, + 0.005938258487731218, + 0.0056091854348778725, + 0.003847327549010515, + 0.0070343720726668835, + 0.009538127109408379, + 0.013053505681455135, + 0.0044420575723052025, + 0.011608530767261982, + 0.009322044439613819, + 0.0133324870839715, + 0.015748506411910057, + 0.03681960701942444, + 0.027358869090676308, + 0.027023838832974434, + 0.015942787751555443, + 0.026653751730918884 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 0, + 1 + ], + "selected_reward": -0.007639241870492697 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a minimalist wireless earbud case?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the case made of matte space gray polymer with a fine-grained texture?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the background a pure, out-of-focus, deep black?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a cool sidelight shine from the upper left creating a sharp highlight on the rounded edge?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera slowly tilt downward while the case rotates and falls?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.11296098679304123 + ], + "opposite_logprob_tokens": [ + -2.2379610538482666 + ], + "logprob_reward": -0.11296098679304123, + "yes_no_margin_reward": 2.125, + "token_credit_reward": -0.11296098679304123, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1961388736963272, + 0.04281489551067352, + 0.02907211147248745, + 0.030803551897406578, + 0.01957225799560547, + 0.1233506128191948, + 0.1692393720149994, + 0.12065128237009048, + 0.0898691788315773, + 0.17848777770996094 + ], + "credit_map": [ + 0.00651957793161273, + 0.07178979367017746, + 0.009385147131979465, + 0.024482134729623795, + 0.030993567779660225, + 0.01651419699192047, + 0.011277087964117527, + 0.025177402421832085, + 0.00522423442453146, + 0.005587778985500336, + 0.0063635967671871185, + 0.004955823998898268, + 0.005109622608870268, + 0.0048911781050264835, + 0.005956630688160658, + 0.004726035054773092, + 0.002935238881036639, + 0.007360421121120453, + 0.00201631267555058, + 0.004190734587609768, + 0.004469425417482853, + 0.002846058225259185, + 0.0021521581802517176, + 0.0031017661094665527, + 0.0033116922713816166, + 0.006708243861794472, + 0.002366557950153947, + 0.0033771030139178038, + 0.004893201868981123, + 0.003575512208044529, + 0.002530881902202964, + 0.004040362313389778, + 0.0024387333542108536, + 0.002553438302129507, + 0.001968459691852331, + 0.002127103041857481, + 0.0030831967014819384, + 0.002557233674451709, + 0.002182857133448124, + 0.0026612384244799614, + 0.006753197871148586, + 0.02025003544986248, + 0.003018926829099655, + 0.0025765420868992805, + 0.007801963482052088, + 0.029680799692869186, + 0.007938164286315441, + 0.045330990105867386, + 0.01239180751144886, + 0.012670018710196018, + 0.005067975725978613, + 0.005453644786030054, + 0.009911610744893551, + 0.06299809366464615, + 0.015262839384377003, + 0.04548338055610657, + 0.012904427014291286, + 0.008789644576609135, + 0.007835674099624157, + 0.00739422719925642, + 0.013967407867312431, + 0.03005712479352951, + 0.01681738719344139, + 0.022885404527187347, + 0.007594439201056957, + 0.007921313866972923, + 0.005880133248865604, + 0.009880087338387966, + 0.012718631885945797, + 0.023172227665781975, + 0.006130729801952839, + 0.016571611166000366, + 0.011143176816403866, + 0.014203556813299656, + 0.01757686212658882, + 0.05319305136799812, + 0.02211756631731987, + 0.024673109874129295, + 0.0167068000882864, + 0.01887367106974125 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 6 + ], + "selected_reward": -0.11296098679304123 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "In an extreme close-up, shot handheld for a subtle and natural shakiness, the camera slowly tilts downward, following the graceful, slow rotation and fall of a minimalist wireless earbud case. The case is crafted from a matte, space gray polymer with a nearly imperceptible fine-grained texture, set against a pure, out-of-focus, deep black background. A precise and dramatic, cool sidelight shines from the upper left, creating a sharp, bright highlight along the rounded edge. Each slow rotation of the light allows the light to flow along its smooth curves, highlighting the flawless industrial design. A soft, reflected light subtly illuminates the other side, revealing rich shadows and material details without detracting from the overall depth. The image, shrouded in a cool, luxurious monochrome, exudes a focused, refined, and technologically modern aesthetic, as if time has been slowed down to showcase the object's exquisite craftsmanship and pure form.", + "seed": 2517, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000034/group_6/index_281_seed_2517_dp_13_rank_26_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 6, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.14149143174290657, + "vlm_logprob": -0.14149143174290657, + "vlm_yes_no_margin": 2.699999988079071, + "vlm_token_credit": -0.14149143174290657, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a minimalist wireless earbud case?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the case made of matte space gray polymer with a fine-grained texture?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the background a pure, out-of-focus, deep black?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a cool sidelight shine from the upper left creating a sharp highlight on the rounded edge?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera slowly tilt downward while the case rotates and falls?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.018453583121299744 + ], + "opposite_logprob_tokens": [ + -4.768453598022461 + ], + "logprob_reward": -0.018453583121299744, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.018453583121299744, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2736702859401703, + 0.0868397206068039, + 0.06050334498286247, + 0.18726781010627747, + 0.09330931305885315, + 0.07603535056114197, + 0.03906808793544769, + 0.04484359920024872, + 0.060568321496248245, + 0.0778941810131073 + ], + "credit_map": [ + 0.016155174002051353, + 0.013490939512848854, + 0.06720364093780518, + 0.019842583686113358, + 0.017647936940193176, + 0.07732203602790833, + 0.040214523673057556, + 0.021793441846966743, + 0.004832902457565069, + 0.004123036749660969, + 0.02119436301290989, + 0.005861673504114151, + 0.004810736980289221, + 0.028900660574436188, + 0.007335812319070101, + 0.009780540131032467, + 0.007223103195428848, + 0.00478634238243103, + 0.009089406579732895, + 0.005924893077462912, + 0.006084860302507877, + 0.008222139440476894, + 0.009482464753091335, + 0.009690134786069393, + 0.022347159683704376, + 0.010885876603424549, + 0.014531794004142284, + 0.015146676450967789, + 0.010596348904073238, + 0.03315714746713638, + 0.024316763505339622, + 0.056286055594682693, + 0.010409142822027206, + 0.006085454486310482, + 0.0054769511334598064, + 0.010747834108769894, + 0.007090964820235968, + 0.021060792729258537, + 0.013094061985611916, + 0.019344117492437363, + 0.006634163670241833, + 0.0062234485521912575, + 0.0068604606203734875, + 0.012104632332921028, + 0.012431337498128414, + 0.01160131674259901, + 0.01245182566344738, + 0.00772815989330411, + 0.003315253648906946, + 0.0026035746559500694, + 0.0038675288669764996, + 0.006860092282295227, + 0.004992096684873104, + 0.005854458082467318, + 0.007407631725072861, + 0.004167451057583094, + 0.0036482971627265215, + 0.0022677164524793625, + 0.005565024446696043, + 0.00441114092245698, + 0.004346217028796673, + 0.010484782047569752, + 0.007919510826468468, + 0.0062009114772081375, + 0.005719176027923822, + 0.002727401675656438, + 0.01787731796503067, + 0.005349508486688137, + 0.003953916020691395, + 0.013025813736021519, + 0.00822929572314024, + 0.0036858911626040936, + 0.010688001289963722, + 0.005261886399239302, + 0.013273842632770538, + 0.010636521503329277, + 0.005994679871946573, + 0.01141731534153223, + 0.012772280722856522, + 0.007849650457501411 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 4 + ], + "selected_reward": -0.018453583121299744 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a minimalist wireless earbud case?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the case made of matte space gray polymer with a fine-grained texture?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the background a pure, out-of-focus, deep black?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a cool sidelight shine from the upper left creating a sharp highlight on the rounded edge?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera slowly tilt downward while the case rotates and falls?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.22579972445964813 + ], + "opposite_logprob_tokens": [ + -1.6007996797561646 + ], + "logprob_reward": -0.22579972445964813, + "yes_no_margin_reward": 1.375, + "token_credit_reward": -0.22579972445964813, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.20022279024124146, + 0.04722151905298233, + 0.128865048289299, + 0.07759172469377518, + 0.08135056495666504, + 0.07313785701990128, + 0.1481388509273529, + 0.09839877486228943, + 0.0618639811873436, + 0.08320897072553635 + ], + "credit_map": [ + 0.01609509065747261, + 0.013038781471550465, + 0.03071637451648712, + 0.01660517416894436, + 0.020068621262907982, + 0.02892155759036541, + 0.038669344037771225, + 0.03610781952738762, + 0.00419517420232296, + 0.004376962780952454, + 0.008412301540374756, + 0.004923551343381405, + 0.005939525552093983, + 0.006927837152034044, + 0.0072088600136339664, + 0.005237303674221039, + 0.014436048455536366, + 0.007923685014247894, + 0.022638294845819473, + 0.007545474451035261, + 0.008733597584068775, + 0.03602656349539757, + 0.014925440773367882, + 0.01663592830300331, + 0.007221561390906572, + 0.007397777866572142, + 0.005323567893356085, + 0.010258172638714314, + 0.006018775515258312, + 0.018504681065678596, + 0.010540354996919632, + 0.012326817028224468, + 0.007546514272689819, + 0.007067390717566013, + 0.005760118830949068, + 0.010629450902342796, + 0.009193319827318192, + 0.014430079609155655, + 0.011511803604662418, + 0.015211883932352066, + 0.005979808513075113, + 0.006408004555851221, + 0.0070110587403178215, + 0.011666834354400635, + 0.008229222148656845, + 0.012185823172330856, + 0.012329624965786934, + 0.009327472187578678, + 0.011648033745586872, + 0.006888005416840315, + 0.014015903696417809, + 0.02254405803978443, + 0.02217651531100273, + 0.022414207458496094, + 0.022604171186685562, + 0.02584793046116829, + 0.007098040077835321, + 0.004355407785624266, + 0.01203600037842989, + 0.01429971493780613, + 0.014414489269256592, + 0.016318459063768387, + 0.01588510163128376, + 0.013991550542414188, + 0.006195443216711283, + 0.002505382290109992, + 0.017508674412965775, + 0.006784544326364994, + 0.0036795074120163918, + 0.01326450053602457, + 0.007744185160845518, + 0.004181732889264822, + 0.00895986519753933, + 0.008971798233687878, + 0.012679689563810825, + 0.010878035798668861, + 0.007146731484681368, + 0.00991789996623993, + 0.01427500881254673, + 0.010379929095506668 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 6, + 2 + ], + "selected_reward": -0.22579972445964813 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a minimalist wireless earbud case?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the case made of matte space gray polymer with a fine-grained texture?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the background a pure, out-of-focus, deep black?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a cool sidelight shine from the upper left creating a sharp highlight on the rounded edge?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera slowly tilt downward while the case rotates and falls?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.3486705422401428 + ], + "opposite_logprob_tokens": [ + -1.223670482635498 + ], + "logprob_reward": -0.3486705422401428, + "yes_no_margin_reward": 0.8749999403953552, + "token_credit_reward": -0.3486705422401428, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2597433626651764, + 0.22208251059055328, + 0.13480904698371887, + 0.06260958313941956, + 0.035662051290273666, + 0.07193184643983841, + 0.04703487083315849, + 0.043043848127126694, + 0.030924415215849876, + 0.09215845167636871 + ], + "credit_map": [ + 0.01698151044547558, + 0.00785053800791502, + 0.06472185254096985, + 0.015733156353235245, + 0.01158072054386139, + 0.07628572732210159, + 0.042277369648218155, + 0.024312490597367287, + 0.014727533794939518, + 0.009083977900445461, + 0.044380687177181244, + 0.010962019674479961, + 0.009502114728093147, + 0.08066868036985397, + 0.011933689005672932, + 0.040823787450790405, + 0.01913742534816265, + 0.00766275729984045, + 0.018796518445014954, + 0.010878818109631538, + 0.00968480296432972, + 0.025888290256261826, + 0.02194175124168396, + 0.020818669348955154, + 0.007507639471441507, + 0.004815415013581514, + 0.005266217514872551, + 0.00539234047755599, + 0.00606453325599432, + 0.009555292315781116, + 0.008382655680179596, + 0.01562548615038395, + 0.004833603743463755, + 0.0025343941524624825, + 0.002700512297451496, + 0.005198249593377113, + 0.003932748455554247, + 0.006315217819064856, + 0.004925379529595375, + 0.005221943370997906, + 0.008902042172849178, + 0.0046087163500487804, + 0.006069767754524946, + 0.008757182396948338, + 0.007059337571263313, + 0.014104295521974564, + 0.007632515858858824, + 0.014797989279031754, + 0.004601494874805212, + 0.0025749006308615208, + 0.004876782186329365, + 0.006916276179254055, + 0.006706537213176489, + 0.00912247784435749, + 0.006295676343142986, + 0.005940727889537811, + 0.004642916843295097, + 0.002213789615780115, + 0.005007379688322544, + 0.005861890502274036, + 0.004802951589226723, + 0.007005341816693544, + 0.006433800794184208, + 0.007075777743011713, + 0.003203736850991845, + 0.0024228347465395927, + 0.0037748285103589296, + 0.005496702156960964, + 0.0029781442135572433, + 0.0038700245786458254, + 0.005648450925946236, + 0.0035296932328492403, + 0.00953463651239872, + 0.004467148333787918, + 0.017918992787599564, + 0.01350749097764492, + 0.00655071996152401, + 0.013933949172496796, + 0.019691303372383118, + 0.006554207298904657 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 2 + ], + "selected_reward": -0.3486705422401428 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a minimalist wireless earbud case?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the case made of matte space gray polymer with a fine-grained texture?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the background a pure, out-of-focus, deep black?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a cool sidelight shine from the upper left creating a sharp highlight on the rounded edge?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera slowly tilt downward while the case rotates and falls?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.014202821999788284 + ], + "opposite_logprob_tokens": [ + -4.26420259475708 + ], + "logprob_reward": -0.014202821999788284, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.014202821999788284, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.158151775598526, + 0.08155398070812225, + 0.07642281800508499, + 0.0509379543364048, + 0.15322653949260712, + 0.0724104717373848, + 0.07023166865110397, + 0.07412446290254593, + 0.13915666937828064, + 0.12378354370594025 + ], + "credit_map": [ + 0.012302469462156296, + 0.01477508433163166, + 0.03575843200087547, + 0.010693264193832874, + 0.013471611775457859, + 0.028799714520573616, + 0.025831041857600212, + 0.016520153731107712, + 0.006365069653838873, + 0.004969275556504726, + 0.010298584587872028, + 0.00660450104624033, + 0.006519084796309471, + 0.017403393983840942, + 0.00800193939357996, + 0.021392133086919785, + 0.007556087803095579, + 0.0063602132722735405, + 0.012886470183730125, + 0.007806089241057634, + 0.006286769639700651, + 0.014236225746572018, + 0.011960376054048538, + 0.009330582804977894, + 0.0054294574074447155, + 0.005698152352124453, + 0.004805757198482752, + 0.007973862811923027, + 0.004627653863281012, + 0.007793919648975134, + 0.007821645587682724, + 0.006787504535168409, + 0.017750192433595657, + 0.009878807701170444, + 0.010696007870137691, + 0.01625513844192028, + 0.012537417933344841, + 0.03655023127794266, + 0.026547616347670555, + 0.02301112376153469, + 0.00718644866719842, + 0.003957508597522974, + 0.006994873285293579, + 0.010297517292201519, + 0.010417361743748188, + 0.00997399166226387, + 0.00923810712993145, + 0.014344664290547371, + 0.0065077743493020535, + 0.005241568200290203, + 0.009431538172066212, + 0.010009623132646084, + 0.009339938871562481, + 0.0103498175740242, + 0.011402909643948078, + 0.007948498241603374, + 0.005450906231999397, + 0.003912883345037699, + 0.008456655777990818, + 0.014150326140224934, + 0.010188393294811249, + 0.011632194742560387, + 0.012255148962140083, + 0.00807795487344265, + 0.0168504286557436, + 0.008515378460288048, + 0.02676483988761902, + 0.015474282205104828, + 0.010465630330145359, + 0.021676216274499893, + 0.026632152497768402, + 0.012777759693562984, + 0.01341711264103651, + 0.011171897873282433, + 0.015280247665941715, + 0.025040768086910248, + 0.011543089523911476, + 0.013858373276889324, + 0.02338491939008236, + 0.010087150149047375 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 8 + ], + "selected_reward": -0.014202821999788284 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a minimalist wireless earbud case?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the case made of matte space gray polymer with a fine-grained texture?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the background a pure, out-of-focus, deep black?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a cool sidelight shine from the upper left creating a sharp highlight on the rounded edge?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera slowly tilt downward while the case rotates and falls?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.10033048689365387 + ], + "opposite_logprob_tokens": [ + -2.3503305912017822 + ], + "logprob_reward": -0.10033048689365387, + "yes_no_margin_reward": 2.25, + "token_credit_reward": -0.10033048689365387, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08490285277366638, + 0.0404154472053051, + 0.11567190289497375, + 0.08004116266965866, + 0.035669341683387756, + 0.09329212456941605, + 0.09753986448049545, + 0.13207952678203583, + 0.15808291733264923, + 0.1623048037290573 + ], + "credit_map": [ + 0.007224711589515209, + 0.009687833487987518, + 0.012124871835112572, + 0.007957947440445423, + 0.007556689437478781, + 0.013642700389027596, + 0.016467610374093056, + 0.01024049986153841, + 0.004021218977868557, + 0.0037747786846011877, + 0.005782778840512037, + 0.004598978906869888, + 0.006119545083492994, + 0.0054194023832678795, + 0.007376115769147873, + 0.0033226313535124063, + 0.013712388463318348, + 0.010343071073293686, + 0.014657609164714813, + 0.010377312079071999, + 0.011045482940971851, + 0.020577583461999893, + 0.01827825978398323, + 0.016680201515555382, + 0.008928047493100166, + 0.0077598197385668755, + 0.005577863659709692, + 0.010543076321482658, + 0.006245560012757778, + 0.017233598977327347, + 0.011880802921950817, + 0.01187239121645689, + 0.005113101098686457, + 0.00347247626632452, + 0.002610002411529422, + 0.0038939914666116238, + 0.0046792724169790745, + 0.006575976964086294, + 0.004671452101320028, + 0.004653073847293854, + 0.008568999357521534, + 0.004945017863065004, + 0.007495911791920662, + 0.011053129099309444, + 0.009038731455802917, + 0.020468469709157944, + 0.014328593388199806, + 0.017393281683325768, + 0.012293386273086071, + 0.004389305133372545, + 0.00981075968593359, + 0.009731973521411419, + 0.012369127944111824, + 0.014900379814207554, + 0.010860968381166458, + 0.023183975368738174, + 0.011407504789531231, + 0.006806741002947092, + 0.014490611851215363, + 0.021679092198610306, + 0.015233563259243965, + 0.02107352949678898, + 0.022879134863615036, + 0.018509352579712868, + 0.014750748872756958, + 0.010714136064052582, + 0.024850813671946526, + 0.01994161121547222, + 0.013625413179397583, + 0.022362684831023216, + 0.03528573736548424, + 0.016551803797483444, + 0.020822055637836456, + 0.01164104975759983, + 0.03714926168322563, + 0.01835070177912712, + 0.011698540300130844, + 0.027456827461719513, + 0.020842723548412323, + 0.014343656599521637 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 8, + 7 + ], + "selected_reward": -0.10033048689365387 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000034/group_8/index_tensor([1388, 1388])_group_8_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000034/group_8/index_tensor([1388, 1388])_group_8_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..b6bc960b8cd38ba530327aa0b434280cc90fe0c2 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000034/group_8/index_tensor([1388, 1388])_group_8_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "Told from an intimate first-person perspective, the camera captures a static close-up, tilting toward a small candle flickering on a cluttered, dark wooden table. The table is strewn with the charming clutter of a scholar or artist: a pile of old books with faded spines, a half-full ceramic mug, and scattered drawing tools. Initially, the small, teardrop-shaped flame casts a soft, flickering glow, its rays dancing gently around. Then, with a single, fluid motion, the flame suddenly and joyfully springs upward, instantly erupting into a magical, brilliant, warm glow, from which soft golden sparks drift lazily like ethereal motes. This stream of light sweeps across the scene, instantly revealing a cozy, meticulously detailed room. The entire visual effect is rendered in the beautiful hand-drawn style of Ghibli animation. The soft-edged objects, rich, warm color palette, and painterly textures imbue this sudden burst of light with a wondrous and sincere quality, creating a tranquil and magical atmosphere.", + "seed": 22575, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000034/group_8/index_1388_seed_22575_dp_16_rank_32_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 8, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.045011519989930096, + "vlm_logprob": -0.045011519989930096, + "vlm_yes_no_margin": 4.15, + "vlm_token_credit": -0.045011519989930096, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene shot from a static close-up perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a small candle with a teardrop-shaped flame visible on a dark wooden table?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are scholar or artist items like old books, a ceramic mug, and drawing tools visible on the table?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the flame suddenly spring upward and erupt into a brilliant glow with drifting golden sparks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the video rendered in the hand-drawn style of Ghibli animation with soft edges and painterly textures?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.10959494858980179 + ], + "opposite_logprob_tokens": [ + -2.8595950603485107 + ], + "logprob_reward": -0.10959494858980179, + "yes_no_margin_reward": 2.75, + "token_credit_reward": -0.10959494858980179, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.05806008353829384, + 0.026035113260149956, + 0.059147126972675323, + 0.10173039883375168, + 0.6051228046417236, + 0.018364187330007553, + 0.039721209555864334, + 0.02264435589313507, + 0.0461721196770668, + 0.023002585396170616 + ], + "credit_map": [ + 0.005965319462120533, + 0.006704634055495262, + 0.006513632833957672, + 0.005082968156784773, + 0.007759521249681711, + 0.008773724548518658, + 0.006526504643261433, + 0.010733778588473797, + 0.002827397082000971, + 0.003642565105110407, + 0.0031576172914355993, + 0.002885247813537717, + 0.0042380839586257935, + 0.004318331368267536, + 0.0027109577786177397, + 0.0022549121640622616, + 0.0055893403477966785, + 0.006144991144537926, + 0.006851222366094589, + 0.005321368109434843, + 0.014046331867575645, + 0.00818940531462431, + 0.007072144653648138, + 0.005932320840656757, + 0.017272530123591423, + 0.008963548578321934, + 0.012396621517837048, + 0.010910781100392342, + 0.01476211566478014, + 0.017073795199394226, + 0.007744105998426676, + 0.012606900185346603, + 0.03776704892516136, + 0.2208002805709839, + 0.02872365713119507, + 0.06133296713232994, + 0.05084080249071121, + 0.0364716574549675, + 0.1352129429578781, + 0.03397343307733536, + 0.002791681792587042, + 0.002661534119397402, + 0.002462885109707713, + 0.0019569157157093287, + 0.0025761171709746122, + 0.002076389966532588, + 0.0018705236725509167, + 0.0019681400153785944, + 0.006221896503120661, + 0.0034795666579157114, + 0.005810847971588373, + 0.0029213642701506615, + 0.0060980794951319695, + 0.006947279442101717, + 0.003237065626308322, + 0.00500510772690177, + 0.002689789514988661, + 0.0023839790374040604, + 0.003344343975186348, + 0.0017122996505349874, + 0.002179816598072648, + 0.003114215098321438, + 0.0020866997074335814, + 0.005133213009685278, + 0.00838303193449974, + 0.003991178702563047, + 0.0032709725201129913, + 0.0026427654083818197, + 0.005470723379403353, + 0.010821227915585041, + 0.0032627908512949944, + 0.008329433389008045, + 0.0029456999618560076, + 0.0026789384428411722, + 0.0024798575323075056, + 0.002734303707256913, + 0.0034466800279915333, + 0.0026864823885262012, + 0.0029150554910302162, + 0.003115567611530423 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 3, + 2 + ], + "selected_reward": -0.10959494858980179 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene shot from a static close-up perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a small candle with a teardrop-shaped flame visible on a dark wooden table?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are scholar or artist items like old books, a ceramic mug, and drawing tools visible on the table?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the flame suddenly spring upward and erupt into a brilliant glow with drifting golden sparks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the video rendered in the hand-drawn style of Ghibli animation with soft edges and painterly textures?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.003220373997464776 + ], + "opposite_logprob_tokens": [ + -5.753220558166504 + ], + "logprob_reward": -0.003220373997464776, + "yes_no_margin_reward": 5.75, + "token_credit_reward": -0.003220373997464776, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1531475931406021, + 0.20973560214042664, + 0.06438354402780533, + 0.15324915945529938, + 0.09914638847112656, + 0.05752895772457123, + 0.028499742969870567, + 0.028661783784627914, + 0.16683009266853333, + 0.038817182183265686 + ], + "credit_map": [ + 0.018769675865769386, + 0.018865467980504036, + 0.017221758142113686, + 0.015081989578902721, + 0.03131794556975365, + 0.026445183902978897, + 0.011767257004976273, + 0.013678298331797123, + 0.01886853016912937, + 0.026352711021900177, + 0.02013968862593174, + 0.01952357590198517, + 0.050169624388217926, + 0.02946413867175579, + 0.027824625372886658, + 0.017392709851264954, + 0.0075284396298229694, + 0.006722848396748304, + 0.008849339559674263, + 0.005329440347850323, + 0.01318279653787613, + 0.008658469654619694, + 0.007774726953357458, + 0.0063374838791787624, + 0.013698482885956764, + 0.013299733400344849, + 0.011896238662302494, + 0.027382802218198776, + 0.016564538702368736, + 0.02848796918988228, + 0.010970575734972954, + 0.030948827043175697, + 0.007438073866069317, + 0.03172887861728668, + 0.005848072934895754, + 0.01105500664561987, + 0.010478726588189602, + 0.005839738994836807, + 0.019920485094189644, + 0.006837413180619478, + 0.005657800938934088, + 0.008178884163498878, + 0.008035811595618725, + 0.004341072868555784, + 0.008636703714728355, + 0.009444467723369598, + 0.0054288869723677635, + 0.007805330213159323, + 0.00362822855822742, + 0.002592413453385234, + 0.0038926538545638323, + 0.002774331485852599, + 0.005014828406274319, + 0.0045973630622029305, + 0.0027305723633617163, + 0.0032693531829863787, + 0.002643593354150653, + 0.002511939499527216, + 0.0033907515462487936, + 0.0019098904449492693, + 0.004225665237754583, + 0.0037229659501463175, + 0.0031301858834922314, + 0.007126792799681425, + 0.012648983858525753, + 0.019274398684501648, + 0.006717161275446415, + 0.010640676133334637, + 0.07001399248838425, + 0.027681659907102585, + 0.008507036603987217, + 0.011346185579895973, + 0.0047811600379645824, + 0.005558720789849758, + 0.0044556874781847, + 0.004758845083415508, + 0.005374358035624027, + 0.004484690725803375, + 0.004317798651754856, + 0.005085917189717293 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 8, + 3 + ], + "selected_reward": -0.003220373997464776 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene shot from a static close-up perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a small candle with a teardrop-shaped flame visible on a dark wooden table?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are scholar or artist items like old books, a ceramic mug, and drawing tools visible on the table?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the flame suddenly spring upward and erupt into a brilliant glow with drifting golden sparks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the video rendered in the hand-drawn style of Ghibli animation with soft edges and painterly textures?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.048677146434783936 + ], + "opposite_logprob_tokens": [ + -3.0486772060394287 + ], + "logprob_reward": -0.048677146434783936, + "yes_no_margin_reward": 3.0, + "token_credit_reward": -0.048677146434783936, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15295130014419556, + 0.08385689556598663, + 0.08290927112102509, + 0.08491028845310211, + 0.2014874666929245, + 0.10287899523973465, + 0.033751361072063446, + 0.04267501458525658, + 0.15660738945007324, + 0.0579720214009285 + ], + "credit_map": [ + 0.02070804126560688, + 0.01755310781300068, + 0.022829687222838402, + 0.01057159248739481, + 0.016215702518820763, + 0.03770460560917854, + 0.012528123334050179, + 0.014840424060821533, + 0.009476297535002232, + 0.009674071334302425, + 0.017793580889701843, + 0.007328228559345007, + 0.01356393564492464, + 0.009992952458560467, + 0.008952299132943153, + 0.007075523026287556, + 0.01214470062404871, + 0.007593510206788778, + 0.015437188558280468, + 0.006241114344447851, + 0.012180878780782223, + 0.01212091650813818, + 0.007648773957043886, + 0.009542192332446575, + 0.007007839623838663, + 0.005610388237982988, + 0.007210617884993553, + 0.01664973981678486, + 0.007352008949965239, + 0.01618129573762417, + 0.005566027481108904, + 0.019332367926836014, + 0.020662866532802582, + 0.03898323327302933, + 0.018491540104150772, + 0.022165769711136818, + 0.028199374675750732, + 0.021671738475561142, + 0.029851846396923065, + 0.021461090072989464, + 0.011810090392827988, + 0.014906584285199642, + 0.013300791382789612, + 0.009096486493945122, + 0.016450420022010803, + 0.014361310750246048, + 0.011540177278220654, + 0.011413133703172207, + 0.004600659478455782, + 0.0029706889763474464, + 0.004941476508975029, + 0.00349514652043581, + 0.005067032761871815, + 0.00470567774027586, + 0.004040327854454517, + 0.003930350765585899, + 0.004906993359327316, + 0.004179860930889845, + 0.0060676103457808495, + 0.002761413808912039, + 0.00807455275207758, + 0.005205186083912849, + 0.004101021680980921, + 0.007378376089036465, + 0.023376721888780594, + 0.015217389911413193, + 0.010543838143348694, + 0.006385011598467827, + 0.026419315487146378, + 0.04207516089081764, + 0.008331838063895702, + 0.024258123710751534, + 0.007768566720187664, + 0.0057749818079173565, + 0.006411499809473753, + 0.008243996649980545, + 0.007911124266684055, + 0.006677649915218353, + 0.0075388019904494286, + 0.007645402103662491 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 8, + 0 + ], + "selected_reward": -0.048677146434783936 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene shot from a static close-up perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a small candle with a teardrop-shaped flame visible on a dark wooden table?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are scholar or artist items like old books, a ceramic mug, and drawing tools visible on the table?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the flame suddenly spring upward and erupt into a brilliant glow with drifting golden sparks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the video rendered in the hand-drawn style of Ghibli animation with soft edges and painterly textures?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0015147175872698426 + ], + "opposite_logprob_tokens": [ + -6.501514911651611 + ], + "logprob_reward": -0.0015147175872698426, + "yes_no_margin_reward": 6.5, + "token_credit_reward": -0.0015147175872698426, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09227193146944046, + 0.11348671466112137, + 0.06179056689143181, + 0.07207438349723816, + 0.1782802790403366, + 0.14389441907405853, + 0.08106774091720581, + 0.10339745134115219, + 0.11030866205692291, + 0.04342777654528618 + ], + "credit_map": [ + 0.011553194373846054, + 0.010357475839555264, + 0.009193235076963902, + 0.008985179476439953, + 0.016437195241451263, + 0.020160505548119545, + 0.006245785858482122, + 0.009339362382888794, + 0.017706329002976418, + 0.014195152558386326, + 0.012807092629373074, + 0.00978242140263319, + 0.023190489038825035, + 0.012945684604346752, + 0.013187459670007229, + 0.009672087617218494, + 0.006814565975219011, + 0.006412586197257042, + 0.007572140544652939, + 0.0049143158830702305, + 0.014206990599632263, + 0.008304333314299583, + 0.007124121766537428, + 0.006441513542085886, + 0.013445485383272171, + 0.008059963583946228, + 0.005139761138707399, + 0.006092301104217768, + 0.010056490078568459, + 0.013247164897620678, + 0.005430936347693205, + 0.01060227956622839, + 0.017192300409078598, + 0.050353989005088806, + 0.015669619664549828, + 0.021636340767145157, + 0.02110086753964424, + 0.015046386048197746, + 0.023110724985599518, + 0.01417005155235529, + 0.014903608709573746, + 0.01877775229513645, + 0.017337972298264503, + 0.01426935289055109, + 0.02643416076898575, + 0.019603388383984566, + 0.016458328813314438, + 0.016109853982925415, + 0.007337183691561222, + 0.0056692552752792835, + 0.010333674028515816, + 0.008142507635056973, + 0.013103850185871124, + 0.01745547540485859, + 0.007147199008613825, + 0.011878595687448978, + 0.012985588982701302, + 0.009287519380450249, + 0.012312258593738079, + 0.006671771872788668, + 0.01952958293259144, + 0.012998855672776699, + 0.010273786261677742, + 0.019338088110089302, + 0.01864572800695896, + 0.010965246707201004, + 0.00954099465161562, + 0.006145033985376358, + 0.014547240920364857, + 0.02639981359243393, + 0.006724934559315443, + 0.0173396747559309, + 0.004273247439414263, + 0.006475479807704687, + 0.004447323735803366, + 0.004726921208202839, + 0.006255013402551413, + 0.006088481284677982, + 0.005237945355474949, + 0.005923368502408266 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 5, + 1 + ], + "selected_reward": -0.0015147175872698426 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene shot from a static close-up perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a small candle with a teardrop-shaped flame visible on a dark wooden table?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are scholar or artist items like old books, a ceramic mug, and drawing tools visible on the table?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the flame suddenly spring upward and erupt into a brilliant glow with drifting golden sparks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the video rendered in the hand-drawn style of Ghibli animation with soft edges and painterly textures?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.062050413340330124 + ], + "opposite_logprob_tokens": [ + -2.8120503425598145 + ], + "logprob_reward": -0.062050413340330124, + "yes_no_margin_reward": 2.75, + "token_credit_reward": -0.062050413340330124, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07187461107969284, + 0.055556271225214005, + 0.023501193150877953, + 0.08687680959701538, + 0.5019629001617432, + 0.0647033080458641, + 0.031784966588020325, + 0.07144363224506378, + 0.02404325269162655, + 0.0682530328631401 + ], + "credit_map": [ + 0.009155203588306904, + 0.007899886928498745, + 0.007827337831258774, + 0.004894251003861427, + 0.008249145932495594, + 0.013346197083592415, + 0.007237576879560947, + 0.013265011832118034, + 0.008043733425438404, + 0.005415447521954775, + 0.007415360305458307, + 0.004855033941566944, + 0.010245878249406815, + 0.008738919161260128, + 0.006753643974661827, + 0.004088254179805517, + 0.002339037135243416, + 0.002626352710649371, + 0.0022233121562749147, + 0.0021964465267956257, + 0.005451692268252373, + 0.0038790826220065355, + 0.0028632781468331814, + 0.0019219889072701335, + 0.008667035028338432, + 0.005928478669375181, + 0.006423070095479488, + 0.014792267233133316, + 0.0085211843252182, + 0.01719667948782444, + 0.005982274655252695, + 0.019365821033716202, + 0.02387504279613495, + 0.19978904724121094, + 0.01683952286839485, + 0.04895469918847084, + 0.03192064166069031, + 0.019013207405805588, + 0.13148218393325806, + 0.030088558793067932, + 0.005558540113270283, + 0.009150877594947815, + 0.0075067258439958096, + 0.005510641261935234, + 0.011679412797093391, + 0.00985229667276144, + 0.008217791095376015, + 0.007227018009871244, + 0.002954475348815322, + 0.0025714051444083452, + 0.003289242275059223, + 0.0031056341249495745, + 0.0043469988740980625, + 0.0069441767409443855, + 0.0035143562126904726, + 0.005058677867054939, + 0.008836738765239716, + 0.007477832492440939, + 0.011348887346684933, + 0.004794907756149769, + 0.009300909005105495, + 0.011453601531684399, + 0.006397539284080267, + 0.01183321699500084, + 0.00340252835303545, + 0.003494365606456995, + 0.002656706143170595, + 0.0027476276736706495, + 0.0036121676675975323, + 0.0025173949543386698, + 0.0031292224302887917, + 0.0024832405615597963, + 0.007361257914453745, + 0.008705870248377323, + 0.008298002183437347, + 0.006722552236169577, + 0.009658814407885075, + 0.009801877662539482, + 0.009560280479490757, + 0.00814437959343195 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 3, + 0 + ], + "selected_reward": -0.062050413340330124 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "Told from an intimate first-person perspective, the camera captures a static close-up, tilting toward a small candle flickering on a cluttered, dark wooden table. The table is strewn with the charming clutter of a scholar or artist: a pile of old books with faded spines, a half-full ceramic mug, and scattered drawing tools. Initially, the small, teardrop-shaped flame casts a soft, flickering glow, its rays dancing gently around. Then, with a single, fluid motion, the flame suddenly and joyfully springs upward, instantly erupting into a magical, brilliant, warm glow, from which soft golden sparks drift lazily like ethereal motes. This stream of light sweeps across the scene, instantly revealing a cozy, meticulously detailed room. The entire visual effect is rendered in the beautiful hand-drawn style of Ghibli animation. The soft-edged objects, rich, warm color palette, and painterly textures imbue this sudden burst of light with a wondrous and sincere quality, creating a tranquil and magical atmosphere.", + "seed": 22576, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000034/group_8/index_1388_seed_22576_dp_16_rank_32_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 8, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.045260890782810745, + "vlm_logprob": -0.045260890782810745, + "vlm_yes_no_margin": 4.775, + "vlm_token_credit": -0.045260890782810745, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene shot from a static close-up perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a small candle with a teardrop-shaped flame visible on a dark wooden table?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are scholar or artist items like old books, a ceramic mug, and drawing tools visible on the table?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the flame suddenly spring upward and erupt into a brilliant glow with drifting golden sparks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the video rendered in the hand-drawn style of Ghibli animation with soft edges and painterly textures?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.12057414650917053 + ], + "opposite_logprob_tokens": [ + -2.6205742359161377 + ], + "logprob_reward": -0.12057414650917053, + "yes_no_margin_reward": 2.5, + "token_credit_reward": -0.12057414650917053, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1450754702091217, + 0.04555920511484146, + 0.06933266669511795, + 0.07765613496303558, + 0.08526846021413803, + 0.22341524064540863, + 0.12066441774368286, + 0.08785328269004822, + 0.09133424609899521, + 0.05384090542793274 + ], + "credit_map": [ + 0.018549736589193344, + 0.019112572073936462, + 0.02497982792556286, + 0.013426369987428188, + 0.014485195279121399, + 0.021520130336284637, + 0.01613473705947399, + 0.0168668981641531, + 0.003916624933481216, + 0.004112090915441513, + 0.0044058579951524734, + 0.006216786801815033, + 0.009221958927810192, + 0.004776462446898222, + 0.00646889628842473, + 0.0064405230805277824, + 0.005675340536981821, + 0.007200191728770733, + 0.006314822938293219, + 0.00801978912204504, + 0.01666755974292755, + 0.008401189930737019, + 0.00851251557469368, + 0.008541256189346313, + 0.007909147068858147, + 0.008344900794327259, + 0.006354469805955887, + 0.0063092028722167015, + 0.018652712926268578, + 0.012420174665749073, + 0.009185707196593285, + 0.00847981870174408, + 0.0067463102750480175, + 0.005528330337256193, + 0.00737798074260354, + 0.008638295345008373, + 0.009146726690232754, + 0.012817572802305222, + 0.01862640306353569, + 0.016386834904551506, + 0.019076688215136528, + 0.022341005504131317, + 0.02560211904346943, + 0.030573055148124695, + 0.047286469489336014, + 0.02263457141816616, + 0.03028419241309166, + 0.02561713010072708, + 0.013677945360541344, + 0.010292265564203262, + 0.024898720905184746, + 0.012707234360277653, + 0.0143800750374794, + 0.019301675260066986, + 0.013749418780207634, + 0.01165707502514124, + 0.006446486804634333, + 0.009158821776509285, + 0.007960802875459194, + 0.018569108098745346, + 0.01304643228650093, + 0.011145553551614285, + 0.012598797678947449, + 0.008927274495363235, + 0.006580608431249857, + 0.016492262482643127, + 0.005225569475442171, + 0.012652480974793434, + 0.009965975768864155, + 0.009007331915199757, + 0.021299198269844055, + 0.010110811330378056, + 0.004286689683794975, + 0.009532536379992962, + 0.004238130524754524, + 0.006235070526599884, + 0.007576209958642721, + 0.008014524355530739, + 0.007495338097214699, + 0.006462405901402235 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 0, + 6 + ], + "selected_reward": -0.12057414650917053 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene shot from a static close-up perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a small candle with a teardrop-shaped flame visible on a dark wooden table?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are scholar or artist items like old books, a ceramic mug, and drawing tools visible on the table?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the flame suddenly spring upward and erupt into a brilliant glow with drifting golden sparks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the video rendered in the hand-drawn style of Ghibli animation with soft edges and painterly textures?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.001209004782140255 + ], + "opposite_logprob_tokens": [ + -6.751208782196045 + ], + "logprob_reward": -0.001209004782140255, + "yes_no_margin_reward": 6.75, + "token_credit_reward": -0.001209004782140255, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1641850620508194, + 0.06272095441818237, + 0.20720236003398895, + 0.08041594922542572, + 0.06329081952571869, + 0.057618673890829086, + 0.21528886258602142, + 0.07109425961971283, + 0.03615265712141991, + 0.04203033819794655 + ], + "credit_map": [ + 0.012846504338085651, + 0.012709753587841988, + 0.016220882534980774, + 0.028270335868000984, + 0.01682443730533123, + 0.013088298961520195, + 0.0378699004650116, + 0.02635495364665985, + 0.005059048067778349, + 0.004746921826153994, + 0.005748269148170948, + 0.009774301201105118, + 0.010081718675792217, + 0.004442309960722923, + 0.013738925568759441, + 0.009129463694989681, + 0.01642487570643425, + 0.024111609905958176, + 0.014586052857339382, + 0.026746535673737526, + 0.05143824219703674, + 0.019770585000514984, + 0.03043406270444393, + 0.02369040623307228, + 0.0074697635136544704, + 0.007073208689689636, + 0.006420792080461979, + 0.009841456077992916, + 0.01739993505179882, + 0.008314035832881927, + 0.013964611105620861, + 0.009932151064276695, + 0.0037516909651458263, + 0.0037003392353653908, + 0.004288475029170513, + 0.008825640194118023, + 0.005166046321392059, + 0.005932377185672522, + 0.019199825823307037, + 0.012426427565515041, + 0.006111056078225374, + 0.005122028756886721, + 0.006508826278150082, + 0.006155720446258783, + 0.012468146160244942, + 0.007366926409304142, + 0.005930348299443722, + 0.007955627515912056, + 0.013948168605566025, + 0.02342146821320057, + 0.018331686034798622, + 0.03992950916290283, + 0.04593956097960472, + 0.016285205259919167, + 0.04149875417351723, + 0.015934528782963753, + 0.005370164755731821, + 0.013778209686279297, + 0.008557847701013088, + 0.011634772643446922, + 0.008806800469756126, + 0.007277771830558777, + 0.01004383061081171, + 0.0056248679757118225, + 0.004472939297556877, + 0.004873472265899181, + 0.0030769763980060816, + 0.004620774649083614, + 0.005632431246340275, + 0.0033803179394453764, + 0.007415701635181904, + 0.0026800474151968956, + 0.0026879259385168552, + 0.011479529552161694, + 0.00307707698084414, + 0.004168201703578234, + 0.0063956063240766525, + 0.0034401174634695053, + 0.007763352710753679, + 0.0030185291543602943 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 2, + 0 + ], + "selected_reward": -0.001209004782140255 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene shot from a static close-up perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a small candle with a teardrop-shaped flame visible on a dark wooden table?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are scholar or artist items like old books, a ceramic mug, and drawing tools visible on the table?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the flame suddenly spring upward and erupt into a brilliant glow with drifting golden sparks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the video rendered in the hand-drawn style of Ghibli animation with soft edges and painterly textures?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0024970562662929296 + ], + "opposite_logprob_tokens": [ + -6.00249719619751 + ], + "logprob_reward": -0.0024970562662929296, + "yes_no_margin_reward": 6.0, + "token_credit_reward": -0.0024970562662929296, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13707154989242554, + 0.056361354887485504, + 0.06207101047039032, + 0.04292337968945503, + 0.036580681800842285, + 0.15099619328975677, + 0.10272440314292908, + 0.1334909349679947, + 0.13472972810268402, + 0.14305071532726288 + ], + "credit_map": [ + 0.020074063912034035, + 0.012141507118940353, + 0.02177751064300537, + 0.015105944126844406, + 0.017102885991334915, + 0.014752577058970928, + 0.01690874621272087, + 0.019208315759897232, + 0.00488533778116107, + 0.0058519477024674416, + 0.0041543892584741116, + 0.008075381629168987, + 0.012693017721176147, + 0.005605501122772694, + 0.007490257732570171, + 0.007605522871017456, + 0.00579792121425271, + 0.005690063815563917, + 0.00932096317410469, + 0.0070996368303895, + 0.013493563048541546, + 0.006202847696840763, + 0.007149834185838699, + 0.00731618469581008, + 0.0037495929282158613, + 0.003614740213379264, + 0.00635205302387476, + 0.00596706522628665, + 0.0066836620680987835, + 0.005054868757724762, + 0.006427969317883253, + 0.005073429550975561, + 0.002818503649905324, + 0.003413920756429434, + 0.002333708107471466, + 0.004789867904037237, + 0.0036010034382343292, + 0.002717113820835948, + 0.009461247362196445, + 0.007445317693054676, + 0.015351560898125172, + 0.012664412148296833, + 0.02277771569788456, + 0.016359509900212288, + 0.022651368752121925, + 0.023849843069911003, + 0.01905573531985283, + 0.018286069855093956, + 0.012800772674381733, + 0.009741350077092648, + 0.02293293923139572, + 0.01207495853304863, + 0.009550081565976143, + 0.014400744810700417, + 0.011948004364967346, + 0.009275555610656738, + 0.010708590969443321, + 0.024753479287028313, + 0.016540557146072388, + 0.020390896126627922, + 0.017218593508005142, + 0.016647445037961006, + 0.01681278645992279, + 0.010418600402772427, + 0.01211559772491455, + 0.009326601400971413, + 0.010379275307059288, + 0.01755351573228836, + 0.034934237599372864, + 0.010933113284409046, + 0.030374901369214058, + 0.009112488478422165, + 0.011114200577139854, + 0.025509724393486977, + 0.015160412527620792, + 0.016025014221668243, + 0.020696744322776794, + 0.02157542295753956, + 0.01951008290052414, + 0.013459119014441967 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 9, + 0 + ], + "selected_reward": -0.0024970562662929296 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene shot from a static close-up perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a small candle with a teardrop-shaped flame visible on a dark wooden table?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are scholar or artist items like old books, a ceramic mug, and drawing tools visible on the table?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the flame suddenly spring upward and erupt into a brilliant glow with drifting golden sparks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the video rendered in the hand-drawn style of Ghibli animation with soft edges and painterly textures?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0017147850012406707 + ], + "opposite_logprob_tokens": [ + -6.376714706420898 + ], + "logprob_reward": -0.0017147850012406707, + "yes_no_margin_reward": 6.375, + "token_credit_reward": -0.0017147850012406707, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06893542408943176, + 0.03423016890883446, + 0.052493199706077576, + 0.07199594378471375, + 0.05228421837091446, + 0.0815265029668808, + 0.13069593906402588, + 0.11616644263267517, + 0.07903538644313812, + 0.31263676285743713 + ], + "credit_map": [ + 0.005238260608166456, + 0.006451432593166828, + 0.009396587498486042, + 0.008178766816854477, + 0.009071361273527145, + 0.008024269714951515, + 0.011302502825856209, + 0.011272240430116653, + 0.0024565146304666996, + 0.004742935765534639, + 0.00322720012627542, + 0.005135918967425823, + 0.0037699821405112743, + 0.004031399264931679, + 0.006093357689678669, + 0.004772858694195747, + 0.004968931898474693, + 0.004203454125672579, + 0.003915009554475546, + 0.006281387992203236, + 0.010637004859745502, + 0.006793908309191465, + 0.008857659064233303, + 0.006835845299065113, + 0.00797991268336773, + 0.0070687560364604, + 0.0060790590941905975, + 0.006953314878046513, + 0.015814069658517838, + 0.010018420405685902, + 0.009325498715043068, + 0.008756909519433975, + 0.004134632181376219, + 0.004263088572770357, + 0.004956305027008057, + 0.008015841245651245, + 0.004327452275902033, + 0.004898099694401026, + 0.0133518623188138, + 0.008336936123669147, + 0.007206369191408157, + 0.007288182619959116, + 0.010467637330293655, + 0.011064688675105572, + 0.016671039164066315, + 0.009509057737886906, + 0.00822749175131321, + 0.011092032305896282, + 0.012572428211569786, + 0.012190345674753189, + 0.023941276594996452, + 0.01784413866698742, + 0.01724659651517868, + 0.016963180154561996, + 0.018681950867176056, + 0.011256017722189426, + 0.010724129155278206, + 0.014336967840790749, + 0.011114547029137611, + 0.01569434069097042, + 0.02066384255886078, + 0.015858734026551247, + 0.01849149912595749, + 0.009282377548515797, + 0.008330891840159893, + 0.011394980363547802, + 0.006232963874936104, + 0.008894151076674461, + 0.01311844028532505, + 0.008114048279821873, + 0.01703055016696453, + 0.005919359624385834, + 0.02216726914048195, + 0.036473583430051804, + 0.019719872623682022, + 0.03633587807416916, + 0.04514262452721596, + 0.03973662853240967, + 0.07597523927688599, + 0.03708568215370178 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 6, + 7 + ], + "selected_reward": -0.0017147850012406707 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene shot from a static close-up perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a small candle with a teardrop-shaped flame visible on a dark wooden table?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are scholar or artist items like old books, a ceramic mug, and drawing tools visible on the table?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the flame suddenly spring upward and erupt into a brilliant glow with drifting golden sparks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the video rendered in the hand-drawn style of Ghibli animation with soft edges and painterly textures?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.10030946135520935 + ], + "opposite_logprob_tokens": [ + -2.350309371948242 + ], + "logprob_reward": -0.10030946135520935, + "yes_no_margin_reward": 2.25, + "token_credit_reward": -0.10030946135520935, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10987216979265213, + 0.025072971358895302, + 0.05677265301346779, + 0.02950100414454937, + 0.029899803921580315, + 0.0504143089056015, + 0.1957533210515976, + 0.16672033071517944, + 0.2235507071018219, + 0.11244269460439682 + ], + "credit_map": [ + 0.012345699593424797, + 0.011240718886256218, + 0.014252852648496628, + 0.015443122945725918, + 0.011543749831616879, + 0.010915784165263176, + 0.01579330302774906, + 0.01833694614470005, + 0.001872036955319345, + 0.0016978789353743196, + 0.0034965539816766977, + 0.004283125977963209, + 0.0030697640031576157, + 0.0021737811621278524, + 0.004423221107572317, + 0.004056611564010382, + 0.004144076257944107, + 0.00611621281132102, + 0.0035641035065054893, + 0.005374035332351923, + 0.01619495078921318, + 0.007730383425951004, + 0.006922569125890732, + 0.006726328749209642, + 0.0021057124249637127, + 0.003092538332566619, + 0.003178489627316594, + 0.0037914945278316736, + 0.006130039691925049, + 0.0034943106584250927, + 0.0043363310396671295, + 0.0033720892388373613, + 0.0026261545717716217, + 0.0016939429333433509, + 0.002054427983239293, + 0.003109496086835861, + 0.002506440505385399, + 0.00387401576153934, + 0.007695275824517012, + 0.006340053863823414, + 0.002877315506339073, + 0.003987474832683802, + 0.006928488612174988, + 0.005916672293096781, + 0.0076631964184343815, + 0.00586197292432189, + 0.009399558417499065, + 0.007779631298035383, + 0.02282872423529625, + 0.017541278153657913, + 0.03594961017370224, + 0.026540711522102356, + 0.020574839785695076, + 0.030263878405094147, + 0.021777791902422905, + 0.020276498049497604, + 0.016327867284417152, + 0.03009956330060959, + 0.017887920141220093, + 0.029947267845273018, + 0.01896536909043789, + 0.01958845928311348, + 0.018546054139733315, + 0.015357841737568378, + 0.02397557720541954, + 0.029364939779043198, + 0.014491434209048748, + 0.02868838608264923, + 0.04276520013809204, + 0.020817406475543976, + 0.04718377813696861, + 0.016264010220766068, + 0.009637675248086452, + 0.016434544697403908, + 0.012181347236037254, + 0.014570544473826885, + 0.011950625106692314, + 0.012075044214725494, + 0.01847868226468563, + 0.017114251852035522 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 6, + 7 + ], + "selected_reward": -0.10030946135520935 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "Told from an intimate first-person perspective, the camera captures a static close-up, tilting toward a small candle flickering on a cluttered, dark wooden table. The table is strewn with the charming clutter of a scholar or artist: a pile of old books with faded spines, a half-full ceramic mug, and scattered drawing tools. Initially, the small, teardrop-shaped flame casts a soft, flickering glow, its rays dancing gently around. Then, with a single, fluid motion, the flame suddenly and joyfully springs upward, instantly erupting into a magical, brilliant, warm glow, from which soft golden sparks drift lazily like ethereal motes. This stream of light sweeps across the scene, instantly revealing a cozy, meticulously detailed room. The entire visual effect is rendered in the beautiful hand-drawn style of Ghibli animation. The soft-edged objects, rich, warm color palette, and painterly textures imbue this sudden burst of light with a wondrous and sincere quality, creating a tranquil and magical atmosphere.", + "seed": 22577, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000034/group_8/index_1388_seed_22577_dp_17_rank_34_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 8, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.579669753741473, + "vlm_logprob": -0.579669753741473, + "vlm_yes_no_margin": 2.375, + "vlm_token_credit": -0.579669753741473, + "vlm_yes_ratio": 0.8, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene shot from a static close-up perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a small candle with a teardrop-shaped flame visible on a dark wooden table?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are scholar or artist items like old books, a ceramic mug, and drawing tools visible on the table?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the flame suddenly spring upward and erupt into a brilliant glow with drifting golden sparks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the video rendered in the hand-drawn style of Ghibli animation with soft edges and painterly textures?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -2.7179245948791504 + ], + "opposite_logprob_tokens": [ + -0.09292469173669815 + ], + "logprob_reward": -2.7179245948791504, + "yes_no_margin_reward": -2.625, + "token_credit_reward": -2.7179245948791504, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.19395017623901367, + 0.12024889141321182, + 0.027093496173620224, + 0.026949120685458183, + 0.0547463558614254, + 0.10655203461647034, + 0.1117691770195961, + 0.12255857139825821, + 0.09985533356666565, + 0.13627682626247406 + ], + "credit_map": [ + 0.018139738589525223, + 0.03350640833377838, + 0.029652222990989685, + 0.015020174905657768, + 0.02562563121318817, + 0.03263041377067566, + 0.01750614494085312, + 0.021869437769055367, + 0.010930415242910385, + 0.013945461250841618, + 0.026516931131482124, + 0.010033633559942245, + 0.018006475642323494, + 0.011020470410585403, + 0.017862264066934586, + 0.011933238245546818, + 0.0022324586752802134, + 0.003598035778850317, + 0.004108413588255644, + 0.0023900154046714306, + 0.004540690686553717, + 0.00357007491402328, + 0.0034797247499227524, + 0.0031740826088935137, + 0.002432472538203001, + 0.003695665393024683, + 0.005599473137408495, + 0.0025408032815903425, + 0.003804658306762576, + 0.003241129219532013, + 0.0027476074174046516, + 0.0028873111587017775, + 0.005874687805771828, + 0.008892209269106388, + 0.005602309945970774, + 0.004409405402839184, + 0.012150347232818604, + 0.006147316657006741, + 0.006192722357809544, + 0.005477355793118477, + 0.009854091331362724, + 0.014370176009833813, + 0.02514920011162758, + 0.01470138318836689, + 0.01498627383261919, + 0.01112582627683878, + 0.008498181588947773, + 0.007866909727454185, + 0.008353141136467457, + 0.01767602190375328, + 0.02155490778386593, + 0.011375273577868938, + 0.012682540342211723, + 0.011384588666260242, + 0.02068619430065155, + 0.008056508377194405, + 0.012682308442890644, + 0.017781727015972137, + 0.03585759550333023, + 0.00970854889601469, + 0.01082125585526228, + 0.011443255469202995, + 0.016745228320360184, + 0.007518660742789507, + 0.010786497965455055, + 0.010786822997033596, + 0.015944892540574074, + 0.016449378803372383, + 0.009942969307303429, + 0.014853045344352722, + 0.013131774961948395, + 0.007959950715303421, + 0.013251302763819695, + 0.017659883946180344, + 0.021741798147559166, + 0.021159561350941658, + 0.012155639939010143, + 0.016181977465748787, + 0.01136088278144598, + 0.022765781730413437 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 7 + ], + "selected_reward": -2.7179245948791504 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene shot from a static close-up perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a small candle with a teardrop-shaped flame visible on a dark wooden table?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are scholar or artist items like old books, a ceramic mug, and drawing tools visible on the table?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the flame suddenly spring upward and erupt into a brilliant glow with drifting golden sparks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the video rendered in the hand-drawn style of Ghibli animation with soft edges and painterly textures?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.011131919920444489 + ], + "opposite_logprob_tokens": [ + -4.511131763458252 + ], + "logprob_reward": -0.011131919920444489, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.011131919920444489, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13863548636436462, + 0.16491183638572693, + 0.10622287541627884, + 0.07056696712970734, + 0.02695232816040516, + 0.08548026531934738, + 0.11960615962743759, + 0.12166879326105118, + 0.08652430027723312, + 0.07943098247051239 + ], + "credit_map": [ + 0.015826506540179253, + 0.019713489338755608, + 0.018775636330246925, + 0.015334440395236015, + 0.01893341727554798, + 0.022096464410424232, + 0.013535190373659134, + 0.014420348219573498, + 0.0137484110891819, + 0.01827874779701233, + 0.041781336069107056, + 0.012265156954526901, + 0.022393841296434402, + 0.015908973291516304, + 0.022746983915567398, + 0.017788371071219444, + 0.01140997838228941, + 0.011499052867293358, + 0.012783830054104328, + 0.011678173206746578, + 0.021578887477517128, + 0.01228728611022234, + 0.014121291227638721, + 0.010864383541047573, + 0.00793805718421936, + 0.01005895622074604, + 0.007426549680531025, + 0.007221517618745565, + 0.014029128476977348, + 0.007773755118250847, + 0.009790838696062565, + 0.006328166928142309, + 0.002668835688382387, + 0.004268310032784939, + 0.0023397784680128098, + 0.002843928523361683, + 0.005262322258204222, + 0.0034770038910210133, + 0.003270344343036413, + 0.002821805188432336, + 0.008909779600799084, + 0.009404648095369339, + 0.01930585503578186, + 0.012001142837107182, + 0.009097529575228691, + 0.00813009962439537, + 0.011210797354578972, + 0.007420407608151436, + 0.008963433094322681, + 0.017523346468806267, + 0.022127287462353706, + 0.012644569389522076, + 0.01790260337293148, + 0.011444919742643833, + 0.020635036751627922, + 0.008364968001842499, + 0.014580285176634789, + 0.016557563096284866, + 0.03877552971243858, + 0.009272393770515919, + 0.0077205840498209, + 0.01100251730531454, + 0.014335418120026588, + 0.009424502961337566, + 0.00936285313218832, + 0.009342996403574944, + 0.017405955120921135, + 0.013458904810249805, + 0.007419091649353504, + 0.012055366300046444, + 0.01097544003278017, + 0.006503697484731674, + 0.007818293757736683, + 0.010594584979116917, + 0.01178277749568224, + 0.010909376665949821, + 0.007072109263390303, + 0.012814650312066078, + 0.005809666123241186, + 0.012629523873329163 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 7 + ], + "selected_reward": -0.011131919920444489 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene shot from a static close-up perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a small candle with a teardrop-shaped flame visible on a dark wooden table?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are scholar or artist items like old books, a ceramic mug, and drawing tools visible on the table?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the flame suddenly spring upward and erupt into a brilliant glow with drifting golden sparks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the video rendered in the hand-drawn style of Ghibli animation with soft edges and painterly textures?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04868827387690544 + ], + "opposite_logprob_tokens": [ + -3.0486881732940674 + ], + "logprob_reward": -0.04868827387690544, + "yes_no_margin_reward": 3.0, + "token_credit_reward": -0.04868827387690544, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.29756101965904236, + 0.26713401079177856, + 0.022179728373885155, + 0.03794117644429207, + 0.043690893799066544, + 0.053505364805459976, + 0.033626165241003036, + 0.1435375064611435, + 0.032128237187862396, + 0.06869594752788544 + ], + "credit_map": [ + 0.027535364031791687, + 0.055507417768239975, + 0.060663461685180664, + 0.02220691181719303, + 0.025164369493722916, + 0.046517618000507355, + 0.024304667487740517, + 0.03566122427582741, + 0.014533564448356628, + 0.031587060540914536, + 0.06225745379924774, + 0.019562045112252235, + 0.04344317689538002, + 0.02929406799376011, + 0.03607926145195961, + 0.030377374961972237, + 0.002589011797681451, + 0.002980825724080205, + 0.002532044192776084, + 0.0019398846197873354, + 0.003001890378072858, + 0.0037676403298974037, + 0.002686514984816313, + 0.0026819156482815742, + 0.004869458265602589, + 0.006251761689782143, + 0.006254369858652353, + 0.002766192192211747, + 0.0050798323936760426, + 0.0048181782476603985, + 0.003554978873580694, + 0.004346405155956745, + 0.003529737703502178, + 0.005760271567851305, + 0.005534450523555279, + 0.004478152841329575, + 0.00961567834019661, + 0.0046752565540373325, + 0.005528570152819157, + 0.004568778444081545, + 0.006226584780961275, + 0.008497724309563637, + 0.006239575333893299, + 0.00452060392126441, + 0.011237848550081253, + 0.005030810367316008, + 0.004584566690027714, + 0.007167648058384657, + 0.002308756345883012, + 0.004785037599503994, + 0.004740969743579626, + 0.003669834928587079, + 0.0039464375004172325, + 0.0035178610123693943, + 0.005723940208554268, + 0.004933327902108431, + 0.018357718363404274, + 0.018619108945131302, + 0.04397069662809372, + 0.012621955946087837, + 0.007111906539648771, + 0.016306912526488304, + 0.013558575883507729, + 0.012990632094442844, + 0.003206004621461034, + 0.005573707167059183, + 0.007663171738386154, + 0.0036820857785642147, + 0.0034139130730181932, + 0.0029818278271704912, + 0.0036925410386174917, + 0.0019149823347106576, + 0.007894598878920078, + 0.006577149033546448, + 0.008510930463671684, + 0.012594396248459816, + 0.005004245322197676, + 0.012440456077456474, + 0.006713102571666241, + 0.008961066603660583 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 7 + ], + "selected_reward": -0.04868827387690544 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene shot from a static close-up perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a small candle with a teardrop-shaped flame visible on a dark wooden table?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are scholar or artist items like old books, a ceramic mug, and drawing tools visible on the table?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the flame suddenly spring upward and erupt into a brilliant glow with drifting golden sparks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the video rendered in the hand-drawn style of Ghibli animation with soft edges and painterly textures?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.007630960550159216 + ], + "opposite_logprob_tokens": [ + -4.882630825042725 + ], + "logprob_reward": -0.007630960550159216, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.007630960550159216, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11426866054534912, + 0.1445370614528656, + 0.04800840839743614, + 0.02172892913222313, + 0.06151861324906349, + 0.05241870880126953, + 0.06404955685138702, + 0.23341572284698486, + 0.13213913142681122, + 0.1279151886701584 + ], + "credit_map": [ + 0.0128526845946908, + 0.015252321027219296, + 0.01802929677069187, + 0.010640505701303482, + 0.014953224919736385, + 0.0191354937851429, + 0.01110914722084999, + 0.01229599118232727, + 0.009783213958144188, + 0.018785616382956505, + 0.027741555124521255, + 0.011119351722300053, + 0.02411426045000553, + 0.01978892832994461, + 0.016900451853871346, + 0.016303684562444687, + 0.004309562966227531, + 0.006438160315155983, + 0.005171594209969044, + 0.003923418466001749, + 0.009842295199632645, + 0.006695399526506662, + 0.006028302013874054, + 0.005599673371762037, + 0.002536970656365156, + 0.00419343076646328, + 0.004020768217742443, + 0.0016689363401383162, + 0.002875551115721464, + 0.002501850947737694, + 0.0017006921116262674, + 0.002230730140581727, + 0.006098661106079817, + 0.010653499513864517, + 0.007000138517469168, + 0.004696367308497429, + 0.01320126187056303, + 0.006994379218667746, + 0.006381960120052099, + 0.006492347456514835, + 0.008031167089939117, + 0.007257139775902033, + 0.00812575314193964, + 0.007100946269929409, + 0.005262916442006826, + 0.005394508130848408, + 0.006358045618981123, + 0.0048882314004004, + 0.005304640159010887, + 0.011255544610321522, + 0.011786669492721558, + 0.007512614130973816, + 0.007336125709116459, + 0.005995051935315132, + 0.008151857182383537, + 0.006707053165882826, + 0.024728119373321533, + 0.03423244506120682, + 0.06101158261299133, + 0.02773302234709263, + 0.015750383958220482, + 0.02709752880036831, + 0.023181304335594177, + 0.019681347534060478, + 0.010956625454127789, + 0.014908385463058949, + 0.026906926184892654, + 0.016323363408446312, + 0.02159803919494152, + 0.011979871429502964, + 0.02011670544743538, + 0.009349209256470203, + 0.014538017101585865, + 0.012604111805558205, + 0.013157038949429989, + 0.022039970383048058, + 0.013338759541511536, + 0.0193257424980402, + 0.017385387793183327, + 0.015526155941188335 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 1, + 8 + ], + "selected_reward": -0.007630960550159216 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene shot from a static close-up perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a small candle with a teardrop-shaped flame visible on a dark wooden table?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are scholar or artist items like old books, a ceramic mug, and drawing tools visible on the table?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the flame suddenly spring upward and erupt into a brilliant glow with drifting golden sparks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the video rendered in the hand-drawn style of Ghibli animation with soft edges and painterly textures?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.11297301948070526 + ], + "opposite_logprob_tokens": [ + -2.2379729747772217 + ], + "logprob_reward": -0.11297301948070526, + "yes_no_margin_reward": 2.125, + "token_credit_reward": -0.11297301948070526, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1326991766691208, + 0.03170650079846382, + 0.03765399008989334, + 0.024325627833604813, + 0.04424626752734184, + 0.089295394718647, + 0.060076456516981125, + 0.34683677554130554, + 0.0938592478632927, + 0.13930052518844604 + ], + "credit_map": [ + 0.013607521541416645, + 0.015851054340600967, + 0.023249654099345207, + 0.016471590846776962, + 0.017933383584022522, + 0.016333099454641342, + 0.014087140560150146, + 0.01516574528068304, + 0.0032826997339725494, + 0.004195244517177343, + 0.006452745757997036, + 0.0039006234146654606, + 0.0035437974147498608, + 0.0030084201134741306, + 0.0035978665109723806, + 0.003725104033946991, + 0.003995983861386776, + 0.0051067364402115345, + 0.004679783713072538, + 0.003044807119295001, + 0.0074910870753228664, + 0.004151028115302324, + 0.005365312099456787, + 0.0038192500360310078, + 0.002712174551561475, + 0.0043571582064032555, + 0.0035252473317086697, + 0.002189922844991088, + 0.003869125619530678, + 0.0021991191897541285, + 0.0027528151404112577, + 0.002720064716413617, + 0.005253023933619261, + 0.007118879817426205, + 0.008274712599813938, + 0.004292793106287718, + 0.006797425914555788, + 0.0036580865271389484, + 0.003788834670558572, + 0.005062512122094631, + 0.01004146970808506, + 0.012634367682039738, + 0.01631321758031845, + 0.012672706507146358, + 0.01315535232424736, + 0.007566828280687332, + 0.010737033560872078, + 0.006174419075250626, + 0.004418608732521534, + 0.010758496820926666, + 0.008045925758779049, + 0.007643518503755331, + 0.007291130255907774, + 0.005241643637418747, + 0.009229585528373718, + 0.007447544485330582, + 0.04091239348053932, + 0.05289372056722641, + 0.09949369728565216, + 0.03095952235162258, + 0.020222315564751625, + 0.04072067141532898, + 0.03410859778523445, + 0.02752585895359516, + 0.011032753624022007, + 0.010425818152725697, + 0.027883460745215416, + 0.012289725244045258, + 0.006354090757668018, + 0.010293747298419476, + 0.009239575825631618, + 0.006340084131807089, + 0.015442798845469952, + 0.014227176085114479, + 0.01927659660577774, + 0.02534486912190914, + 0.00968858040869236, + 0.026043759658932686, + 0.011819533072412014, + 0.017457207664847374 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 9, + 0 + ], + "selected_reward": -0.11297301948070526 + } + ], + "vlm_decoded_outputs": [ + "No", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "Told from an intimate first-person perspective, the camera captures a static close-up, tilting toward a small candle flickering on a cluttered, dark wooden table. The table is strewn with the charming clutter of a scholar or artist: a pile of old books with faded spines, a half-full ceramic mug, and scattered drawing tools. Initially, the small, teardrop-shaped flame casts a soft, flickering glow, its rays dancing gently around. Then, with a single, fluid motion, the flame suddenly and joyfully springs upward, instantly erupting into a magical, brilliant, warm glow, from which soft golden sparks drift lazily like ethereal motes. This stream of light sweeps across the scene, instantly revealing a cozy, meticulously detailed room. The entire visual effect is rendered in the beautiful hand-drawn style of Ghibli animation. The soft-edged objects, rich, warm color palette, and painterly textures imbue this sudden burst of light with a wondrous and sincere quality, creating a tranquil and magical atmosphere.", + "seed": 22578, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000034/group_8/index_1388_seed_22578_dp_17_rank_34_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 8, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.1466101338621229, + "vlm_logprob": -0.1466101338621229, + "vlm_yes_no_margin": 3.775, + "vlm_token_credit": -0.1466101338621229, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene shot from a static close-up perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a small candle with a teardrop-shaped flame visible on a dark wooden table?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are scholar or artist items like old books, a ceramic mug, and drawing tools visible on the table?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the flame suddenly spring upward and erupt into a brilliant glow with drifting golden sparks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the video rendered in the hand-drawn style of Ghibli animation with soft edges and painterly textures?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.6710336208343506 + ], + "opposite_logprob_tokens": [ + -0.7960336208343506 + ], + "logprob_reward": -0.6710336208343506, + "yes_no_margin_reward": 0.125, + "token_credit_reward": -0.6710336208343506, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.16199971735477448, + 0.3731558918952942, + 0.30181294679641724, + 0.02109292335808277, + 0.019273201003670692, + 0.054002340883016586, + 0.030044348910450935, + 0.019905975088477135, + 0.010364961810410023, + 0.008347732946276665 + ], + "credit_map": [ + 0.026194028556346893, + 0.02864946238696575, + 0.013292812742292881, + 0.016761211678385735, + 0.013921529054641724, + 0.013375012204051018, + 0.03044535033404827, + 0.019360315054655075, + 0.04693140089511871, + 0.053275641053915024, + 0.03560561314225197, + 0.04185137152671814, + 0.02826041728258133, + 0.04003998264670372, + 0.07835249602794647, + 0.04883897304534912, + 0.027202289551496506, + 0.012648966163396835, + 0.037205543369054794, + 0.0629381537437439, + 0.022461554035544395, + 0.011499307118356228, + 0.10944025963544846, + 0.01841685175895691, + 0.0029739446472376585, + 0.0024588985834270716, + 0.003009367734193802, + 0.002629924798384309, + 0.003991654608398676, + 0.002026311121881008, + 0.0022107453551143408, + 0.0017920780228450894, + 0.003313428955152631, + 0.0023148872423917055, + 0.0034512996207922697, + 0.002093493938446045, + 0.0021221667993813753, + 0.002002670429646969, + 0.0017879045335575938, + 0.002187349135056138, + 0.008420988917350769, + 0.004559522029012442, + 0.007779664359986782, + 0.0034692822955548763, + 0.013132928870618343, + 0.003610658925026655, + 0.007458217907696962, + 0.005571079906076193, + 0.003756011836230755, + 0.00385093386285007, + 0.003229104448109865, + 0.00326116313226521, + 0.0032666183542460203, + 0.004085662309080362, + 0.005020759999752045, + 0.003574095666408539, + 0.0021666795946657658, + 0.0040481118485331535, + 0.0017455891938880086, + 0.0023960666730999947, + 0.002320248167961836, + 0.0038161962293088436, + 0.0015849926276132464, + 0.001828090287744999, + 0.001148789655417204, + 0.0019807966891676188, + 0.0010205445578321815, + 0.0011648713843896985, + 0.0012647777330130339, + 0.0010939047206193209, + 0.0012212768197059631, + 0.0014700005995109677, + 0.0012823677388951182, + 0.0011940238764509559, + 0.001041735871694982, + 0.0011883967090398073, + 0.0009467849158681929, + 0.0006429807399399579, + 0.0008254240965470672, + 0.0012260186485946178 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 2, + 0 + ], + "selected_reward": -0.6710336208343506 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene shot from a static close-up perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a small candle with a teardrop-shaped flame visible on a dark wooden table?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are scholar or artist items like old books, a ceramic mug, and drawing tools visible on the table?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the flame suddenly spring upward and erupt into a brilliant glow with drifting golden sparks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the video rendered in the hand-drawn style of Ghibli animation with soft edges and painterly textures?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0032263153698295355 + ], + "opposite_logprob_tokens": [ + -5.753226280212402 + ], + "logprob_reward": -0.0032263153698295355, + "yes_no_margin_reward": 5.75, + "token_credit_reward": -0.0032263153698295355, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13741524517536163, + 0.24727845191955566, + 0.15874892473220825, + 0.06583784520626068, + 0.11732304096221924, + 0.036005519330501556, + 0.11279351264238358, + 0.05796750262379646, + 0.04582982882857323, + 0.020800042897462845 + ], + "credit_map": [ + 0.015981946140527725, + 0.023439772427082062, + 0.011449755169451237, + 0.014290406368672848, + 0.011506311595439911, + 0.014375383965671062, + 0.029283059760928154, + 0.017088614404201508, + 0.028460144996643066, + 0.03332719951868057, + 0.023996083065867424, + 0.026937421411275864, + 0.01781555451452732, + 0.028267592191696167, + 0.05579306557774544, + 0.0326814167201519, + 0.013419675640761852, + 0.007799896411597729, + 0.0207776241004467, + 0.03187265992164612, + 0.013625796884298325, + 0.006916001904755831, + 0.05511973425745964, + 0.009217537939548492, + 0.008798191323876381, + 0.00792363565415144, + 0.009577498771250248, + 0.007476110942661762, + 0.012474973686039448, + 0.0046723373234272, + 0.009510839357972145, + 0.005404262337833643, + 0.013654695823788643, + 0.011738602072000504, + 0.013765351846814156, + 0.01887761801481247, + 0.014299817383289337, + 0.007178532890975475, + 0.020935580134391785, + 0.01687285490334034, + 0.004247863311320543, + 0.00412015663459897, + 0.005180359352380037, + 0.0037181968800723553, + 0.006394818890839815, + 0.0028040653560310602, + 0.004708945285528898, + 0.004831114783883095, + 0.011349284090101719, + 0.012460215017199516, + 0.011761930771172047, + 0.0156331118196249, + 0.014413424767553806, + 0.012931195087730885, + 0.024578630924224854, + 0.00966573040932417, + 0.005097553133964539, + 0.008682863786816597, + 0.006865214090794325, + 0.005429889541119337, + 0.007583319209516048, + 0.012131493538618088, + 0.005326106213033199, + 0.006851066369563341, + 0.005224169697612524, + 0.005209781229496002, + 0.0040662409737706184, + 0.004929936956614256, + 0.006131465081125498, + 0.008090191520750523, + 0.005851029884070158, + 0.006327017676085234, + 0.0026406487450003624, + 0.002578668063506484, + 0.0024457720573991537, + 0.003269066335633397, + 0.0024692844599485397, + 0.0015423831064254045, + 0.0030900579877197742, + 0.0027641626074910164 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 2, + 0 + ], + "selected_reward": -0.0032263153698295355 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene shot from a static close-up perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a small candle with a teardrop-shaped flame visible on a dark wooden table?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are scholar or artist items like old books, a ceramic mug, and drawing tools visible on the table?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the flame suddenly spring upward and erupt into a brilliant glow with drifting golden sparks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the video rendered in the hand-drawn style of Ghibli animation with soft edges and painterly textures?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.011087591759860516 + ], + "opposite_logprob_tokens": [ + -4.511087417602539 + ], + "logprob_reward": -0.011087591759860516, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.011087591759860516, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.17541688680648804, + 0.1388559639453888, + 0.13270628452301025, + 0.08994845300912857, + 0.07984067499637604, + 0.10082431137561798, + 0.18008548021316528, + 0.06136896461248398, + 0.019085410982370377, + 0.02186756581068039 + ], + "credit_map": [ + 0.0241546593606472, + 0.03282344341278076, + 0.01757689379155636, + 0.018446102738380432, + 0.0200189258903265, + 0.023537691682577133, + 0.022195428609848022, + 0.01666373386979103, + 0.015410831198096275, + 0.020538443699479103, + 0.013839265331625938, + 0.013326311483979225, + 0.01322667021304369, + 0.015542970038950443, + 0.02993454411625862, + 0.017036937177181244, + 0.020440788939595222, + 0.010780137963593006, + 0.025635967031121254, + 0.008849944919347763, + 0.014646557159721851, + 0.010693478398025036, + 0.030254337936639786, + 0.01140507496893406, + 0.01133885607123375, + 0.011786291375756264, + 0.008471754379570484, + 0.006707632914185524, + 0.02201942540705204, + 0.007134913466870785, + 0.01596558466553688, + 0.006523989140987396, + 0.011150590144097805, + 0.010030344128608704, + 0.010943035595119, + 0.011716826818883419, + 0.011207913048565388, + 0.007074711844325066, + 0.00779432849958539, + 0.009922920726239681, + 0.013594966381788254, + 0.007344108074903488, + 0.00888292770832777, + 0.006464933045208454, + 0.03173083811998367, + 0.008385362103581429, + 0.01728251576423645, + 0.007138666231185198, + 0.019495956599712372, + 0.029320336878299713, + 0.025099392980337143, + 0.022227827459573746, + 0.016278278082609177, + 0.023638013750314713, + 0.017036421224474907, + 0.02698925882577896, + 0.005683761090040207, + 0.013670242391526699, + 0.004372735973447561, + 0.006818030029535294, + 0.008053397759795189, + 0.011669419705867767, + 0.005679444409906864, + 0.0054219258017838, + 0.0021407040767371655, + 0.0022774075623601675, + 0.0021664255764335394, + 0.002278077183291316, + 0.002593272365629673, + 0.0025948730763047934, + 0.002149397972971201, + 0.0028852534014731646, + 0.00280828308314085, + 0.0035132516641169786, + 0.003727380186319351, + 0.0023238572757691145, + 0.0026907920837402344, + 0.002064269036054611, + 0.0025262944400310516, + 0.0022134389728307724 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 0, + 1 + ], + "selected_reward": -0.011087591759860516 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene shot from a static close-up perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a small candle with a teardrop-shaped flame visible on a dark wooden table?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are scholar or artist items like old books, a ceramic mug, and drawing tools visible on the table?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the flame suddenly spring upward and erupt into a brilliant glow with drifting golden sparks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the video rendered in the hand-drawn style of Ghibli animation with soft edges and painterly textures?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004637913778424263 + ], + "opposite_logprob_tokens": [ + -5.379637718200684 + ], + "logprob_reward": -0.004637913778424263, + "yes_no_margin_reward": 5.375, + "token_credit_reward": -0.004637913778424263, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07374406605958939, + 0.3044721484184265, + 0.04753310978412628, + 0.06578847765922546, + 0.07959730923175812, + 0.14008305966854095, + 0.08278273046016693, + 0.07370265573263168, + 0.08453527837991714, + 0.047761086374521255 + ], + "credit_map": [ + 0.01044555101543665, + 0.010015320964157581, + 0.00903416983783245, + 0.00714645953848958, + 0.008266809396445751, + 0.007089730817824602, + 0.012742957100272179, + 0.009003075771033764, + 0.03299718722701073, + 0.03935345262289047, + 0.030420340597629547, + 0.03485732898116112, + 0.02311503328382969, + 0.03489214926958084, + 0.06983984261751175, + 0.038996852934360504, + 0.006232281215488911, + 0.00652683898806572, + 0.0066905817948281765, + 0.006512309890240431, + 0.005134578328579664, + 0.0038244647439569235, + 0.008224544115364552, + 0.004387517459690571, + 0.006918114144355059, + 0.005639353767037392, + 0.005832220893353224, + 0.005756480619311333, + 0.01678316481411457, + 0.004309716168791056, + 0.01643192023038864, + 0.004117513075470924, + 0.009759600274264812, + 0.007663730066269636, + 0.009403550066053867, + 0.012075741775333881, + 0.009774532169103622, + 0.005227748304605484, + 0.012551219202578068, + 0.013141196221113205, + 0.016109826043248177, + 0.00979716144502163, + 0.012237206101417542, + 0.013948403298854828, + 0.04153173416852951, + 0.012578029185533524, + 0.020368855446577072, + 0.013511857949197292, + 0.008284703828394413, + 0.009021710604429245, + 0.008588557131588459, + 0.012500084936618805, + 0.010829715989530087, + 0.00993200484663248, + 0.014211217872798443, + 0.009414739906787872, + 0.00677367253229022, + 0.015471707098186016, + 0.008702578954398632, + 0.007649844512343407, + 0.010862025432288647, + 0.006754034198820591, + 0.011533157899975777, + 0.00595564441755414, + 0.007380388677120209, + 0.010333895683288574, + 0.0059540183283388615, + 0.009211721830070019, + 0.011318380944430828, + 0.016235165297985077, + 0.013797271996736526, + 0.010304447263479233, + 0.0064672124572098255, + 0.006311009172350168, + 0.004125954583287239, + 0.004749412648379803, + 0.007271498907357454, + 0.006905867252498865, + 0.006864701397716999, + 0.005065436474978924 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 5, + 8 + ], + "selected_reward": -0.004637913778424263 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene shot from a static close-up perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a small candle with a teardrop-shaped flame visible on a dark wooden table?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are scholar or artist items like old books, a ceramic mug, and drawing tools visible on the table?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the flame suddenly spring upward and erupt into a brilliant glow with drifting golden sparks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the video rendered in the hand-drawn style of Ghibli animation with soft edges and painterly textures?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04306522756814957 + ], + "opposite_logprob_tokens": [ + -3.168065309524536 + ], + "logprob_reward": -0.04306522756814957, + "yes_no_margin_reward": 3.125, + "token_credit_reward": -0.04306522756814957, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.17090675234794617, + 0.18333373963832855, + 0.15429940819740295, + 0.10049867630004883, + 0.07199424505233765, + 0.1395510584115982, + 0.06475841253995895, + 0.051100462675094604, + 0.031072698533535004, + 0.032484568655490875 + ], + "credit_map": [ + 0.0221512820571661, + 0.02637651190161705, + 0.016112223267555237, + 0.015452886931598186, + 0.014127169735729694, + 0.024732308462262154, + 0.03023504838347435, + 0.021719317883253098, + 0.01585964486002922, + 0.020316502079367638, + 0.01988881640136242, + 0.021639179438352585, + 0.01544104889035225, + 0.024680841714143753, + 0.04102499037981033, + 0.0244827251881361, + 0.015562169253826141, + 0.01442748587578535, + 0.016478946432471275, + 0.027441827580332756, + 0.012175925076007843, + 0.010459480807185173, + 0.04739212989807129, + 0.01036144234240055, + 0.009817599318921566, + 0.007399058435112238, + 0.008256398141384125, + 0.009685561992228031, + 0.027147691696882248, + 0.00631616311147809, + 0.0251217782497406, + 0.006754428148269653, + 0.011133146472275257, + 0.009602143429219723, + 0.009135148487985134, + 0.009259711019694805, + 0.007569893728941679, + 0.0074822003953158855, + 0.005956714507192373, + 0.0118552902713418, + 0.016440993174910545, + 0.011413183063268661, + 0.013533693738281727, + 0.011017391458153725, + 0.03750922530889511, + 0.01492120698094368, + 0.020920444279909134, + 0.013794934377074242, + 0.008399649523198605, + 0.009908227249979973, + 0.007143963593989611, + 0.006452036090195179, + 0.006595598068088293, + 0.007214054465293884, + 0.011669262312352657, + 0.007375624030828476, + 0.0045707314275205135, + 0.010101634077727795, + 0.004390406422317028, + 0.004397211596369743, + 0.006721148733049631, + 0.011517800390720367, + 0.005224788561463356, + 0.004176740534603596, + 0.0029248574282974005, + 0.004352256190031767, + 0.003561086254194379, + 0.004459654912352562, + 0.004583684261888266, + 0.004645527806133032, + 0.0038931560702621937, + 0.0026524753775447607, + 0.0036750691942870617, + 0.005440935958176851, + 0.004284607246518135, + 0.003228166839107871, + 0.004868877585977316, + 0.003652620827779174, + 0.004250229336321354, + 0.0030840595718473196 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 2 + ], + "selected_reward": -0.04306522756814957 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000034/group_9/index_tensor([4000, 4000])_group_9_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000034/group_9/index_tensor([4000, 4000])_group_9_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..fbb7a39eae36d46a6d811a0b7e93b51c3ea1403e --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000034/group_9/index_tensor([4000, 4000])_group_9_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "slow-motion, time-lapse medium shot, the camera held still, captures a vast, bustling train platform. Beneath a high, vaulted ceiling, a blur of ghostly figures, dragging suitcases and carrying backpacks, surges in a frantic, accelerated pattern, creating a chaotic, blurry mess of colors. Trains enter and depart the station at breakneck speed, their lights casting long trails of light across the frame. Amidst this surging sea, a young woman stands as the sole still anchor. She wears a long, dark trench coat, a beige scarf neatly wrapped around her neck. Industrial lights arrayed on the platform's ceiling cast a bright yet slightly cold light, framing her focused profile. Completely oblivious to the surrounding clamor, she clutches a thick, hardcover book in her hands, her long, dark brown hair casually tied back. In stark contrast to the accelerated movement around her, her fingers gently turn the yellowed pages, one by one, at a normal, leisurely pace. This tiny, steady and focused movement creates an oasis of tranquility in this silent and busy scene. The entire picture builds a strong dramatic tension between the cool background and the focused warmth emanating from her reading, creating a surreal and thought-provoking atmosphere.", + "seed": 12914, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000034/group_9/index_4000_seed_12914_dp_18_rank_36_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 9, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.16166969202458858, + "vlm_logprob": -0.16166969202458858, + "vlm_yes_no_margin": 2.7249999761581423, + "vlm_token_credit": -0.16166969202458858, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set on a vast, bustling train platform beneath a high, vaulted ceiling?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a young woman visible standing still in the center of the frame wearing a long dark trench coat and a beige scarf?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are ghostly figures dragging suitcases and carrying backpacks moving in a frantic, accelerated blur around the woman?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the woman clutching a thick hardcover book and turning its yellowed pages at a normal, leisurely pace?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera held still in a medium shot while industrial lights cast a bright yet slightly cold illumination on the scene?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.16754566133022308 + ], + "opposite_logprob_tokens": [ + -2.0425455570220947 + ], + "logprob_reward": -0.16754566133022308, + "yes_no_margin_reward": 1.8749998807907104, + "token_credit_reward": -0.16754566133022308, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3132052421569824, + 0.06852012872695923, + 0.07333946973085403, + 0.07321690022945404, + 0.04186238721013069, + 0.032503850758075714, + 0.292568564414978, + 0.0813206285238266, + 0.006806167308241129, + 0.01665661297738552 + ], + "credit_map": [ + 0.052234187722206116, + 0.042308662086725235, + 0.03856457397341728, + 0.04100467264652252, + 0.037130385637283325, + 0.045341283082962036, + 0.030554546043276787, + 0.026066934689879417, + 0.009230143390595913, + 0.01191230583935976, + 0.01437214482575655, + 0.007750009652227163, + 0.005203472450375557, + 0.006559415720403194, + 0.006740181241184473, + 0.006752460263669491, + 0.009435062296688557, + 0.009385605342686176, + 0.017155908048152924, + 0.012178105302155018, + 0.0036594350822269917, + 0.008055676706135273, + 0.0053720977157354355, + 0.008097575046122074, + 0.012297292239964008, + 0.00910258386284113, + 0.008487149141728878, + 0.0073058512061834335, + 0.00781635008752346, + 0.008859770372509956, + 0.009206115268170834, + 0.010141796432435513, + 0.006228576879948378, + 0.00704363314434886, + 0.006849402096122503, + 0.005249666050076485, + 0.0038833599537611008, + 0.002632764633744955, + 0.006370203569531441, + 0.0036047804169356823, + 0.0038915707264095545, + 0.0033963951282203197, + 0.004698410164564848, + 0.0034809717908501625, + 0.0033915690146386623, + 0.004896839614957571, + 0.0039001558907330036, + 0.004847938660532236, + 0.018771782517433167, + 0.014300871640443802, + 0.08739815652370453, + 0.11333940178155899, + 0.011645249091088772, + 0.008965742774307728, + 0.017524180933833122, + 0.02062319777905941, + 0.0070916712284088135, + 0.008461648598313332, + 0.017203280702233315, + 0.01738137938082218, + 0.0037316696252673864, + 0.007368091959506273, + 0.008224480785429478, + 0.011858414858579636, + 0.0010787799255922437, + 0.0010283373994752765, + 0.0008236704161390662, + 0.0006565229850821197, + 0.00045234421850182116, + 0.0007293944363482296, + 0.0009113503620028496, + 0.0011257677106186748, + 0.0016484892694279552, + 0.0016839917516335845, + 0.0032682407181710005, + 0.002244386123493314, + 0.0015368055319413543, + 0.001433823723345995, + 0.002459720242768526, + 0.0023811578284949064 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 6, + 7 + ], + "selected_reward": -0.16754566133022308 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set on a vast, bustling train platform beneath a high, vaulted ceiling?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a young woman visible standing still in the center of the frame wearing a long dark trench coat and a beige scarf?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are ghostly figures dragging suitcases and carrying backpacks moving in a frantic, accelerated blur around the woman?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the woman clutching a thick hardcover book and turning its yellowed pages at a normal, leisurely pace?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera held still in a medium shot while industrial lights cast a bright yet slightly cold illumination on the scene?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02061687409877777 + ], + "opposite_logprob_tokens": [ + -3.8956167697906494 + ], + "logprob_reward": -0.02061687409877777, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.02061687409877777, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.23726880550384521, + 0.1555979698896408, + 0.03198918327689171, + 0.06380341202020645, + 0.06143798306584358, + 0.022844351828098297, + 0.34197285771369934, + 0.03548040986061096, + 0.021652836352586746, + 0.02795213833451271 + ], + "credit_map": [ + 0.042536571621894836, + 0.03312182053923607, + 0.029890060424804688, + 0.032348182052373886, + 0.029483864083886147, + 0.02821340598165989, + 0.023203236982226372, + 0.018471697345376015, + 0.009780988097190857, + 0.010676228441298008, + 0.023073716089129448, + 0.019581042230129242, + 0.024351367726922035, + 0.01650278829038143, + 0.031159229576587677, + 0.020472627133131027, + 0.003594506997615099, + 0.00255276495590806, + 0.005105479154735804, + 0.004261006601154804, + 0.003608986735343933, + 0.0047209723852574825, + 0.0036668479442596436, + 0.004478622227907181, + 0.0038033321034163237, + 0.003982015885412693, + 0.00871254038065672, + 0.0074790893122553825, + 0.007889418862760067, + 0.005020508076995611, + 0.01979970932006836, + 0.007116804365068674, + 0.006669936235994101, + 0.006656567566096783, + 0.010115590877830982, + 0.007283920422196388, + 0.007272243965417147, + 0.0037512360140681267, + 0.013301935978233814, + 0.006386559456586838, + 0.002773741027340293, + 0.002425630809739232, + 0.003171093761920929, + 0.003220492973923683, + 0.002369082299992442, + 0.0020767308305948973, + 0.003458735765889287, + 0.0033488476183265448, + 0.02294584922492504, + 0.018486006185412407, + 0.0948169082403183, + 0.125698983669281, + 0.017036091536283493, + 0.015142923220992088, + 0.02324247919023037, + 0.024603672325611115, + 0.002253467682749033, + 0.00237968354485929, + 0.005329318344593048, + 0.005237319506704807, + 0.0029200438875705004, + 0.0025096703320741653, + 0.009372743777930737, + 0.0054781693033874035, + 0.0012960925232619047, + 0.0013581569073721766, + 0.006341199856251478, + 0.006009412929415703, + 0.0013120825169607997, + 0.0014659424778074026, + 0.002318582497537136, + 0.0015513701364398003, + 0.003452097997069359, + 0.0035784589126706123, + 0.003792980220168829, + 0.003918632864952087, + 0.002613001735880971, + 0.0028103715740144253, + 0.003207352478057146, + 0.004579243715852499 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 0, + 1 + ], + "selected_reward": -0.02061687409877777 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set on a vast, bustling train platform beneath a high, vaulted ceiling?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a young woman visible standing still in the center of the frame wearing a long dark trench coat and a beige scarf?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are ghostly figures dragging suitcases and carrying backpacks moving in a frantic, accelerated blur around the woman?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the woman clutching a thick hardcover book and turning its yellowed pages at a normal, leisurely pace?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera held still in a medium shot while industrial lights cast a bright yet slightly cold illumination on the scene?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.020609749481081963 + ], + "opposite_logprob_tokens": [ + -3.8956098556518555 + ], + "logprob_reward": -0.020609749481081963, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.020609749481081963, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12295419722795486, + 0.0719313845038414, + 0.060368988662958145, + 0.05377676710486412, + 0.1950002759695053, + 0.041072383522987366, + 0.31263601779937744, + 0.04049640893936157, + 0.041650447994470596, + 0.06011315435171127 + ], + "credit_map": [ + 0.015194590203464031, + 0.01968243718147278, + 0.021273739635944366, + 0.019704174250364304, + 0.011377965100109577, + 0.012349720112979412, + 0.01416437141597271, + 0.009207193739712238, + 0.008251255378127098, + 0.00854840874671936, + 0.0095003517344594, + 0.011125164106488228, + 0.006933204829692841, + 0.009794283658266068, + 0.0065175071358680725, + 0.01126120425760746, + 0.007042938377708197, + 0.004500323440879583, + 0.01534353569149971, + 0.009913910180330276, + 0.004670098423957825, + 0.0068032699637115, + 0.005058753304183483, + 0.007036157883703709, + 0.006569908931851387, + 0.0070470180362463, + 0.0055110570974648, + 0.006224391050636768, + 0.0054568639025092125, + 0.0058991932310163975, + 0.010522977448999882, + 0.0065453555434942245, + 0.032855816185474396, + 0.029803747311234474, + 0.032948095351457596, + 0.020997941493988037, + 0.01516659278422594, + 0.01299589779227972, + 0.031031327322125435, + 0.01920085772871971, + 0.003753217402845621, + 0.005957566201686859, + 0.004876129794865847, + 0.006008015479892492, + 0.004779938142746687, + 0.0038218642584979534, + 0.006749861873686314, + 0.005125789903104305, + 0.020922724157571793, + 0.015826977789402008, + 0.08911830931901932, + 0.1187392920255661, + 0.012722842395305634, + 0.013040482066571712, + 0.01975422166287899, + 0.022511160001158714, + 0.0035247267223894596, + 0.003172650933265686, + 0.0076504722237586975, + 0.005708420183509588, + 0.002814030973240733, + 0.004079204984009266, + 0.007237832061946392, + 0.006309071555733681, + 0.0025775989051908255, + 0.002841910580173135, + 0.01091661024838686, + 0.011271369643509388, + 0.0024118581786751747, + 0.0038583138957619667, + 0.003986509051173925, + 0.003786275628954172, + 0.0063200886361300945, + 0.005595252383500338, + 0.011901837773621082, + 0.007731114514172077, + 0.0048715886659920216, + 0.005489450413733721, + 0.007194746285676956, + 0.011009079404175282 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 4, + 0 + ], + "selected_reward": -0.020609749481081963 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set on a vast, bustling train platform beneath a high, vaulted ceiling?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a young woman visible standing still in the center of the frame wearing a long dark trench coat and a beige scarf?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are ghostly figures dragging suitcases and carrying backpacks moving in a frantic, accelerated blur around the woman?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the woman clutching a thick hardcover book and turning its yellowed pages at a normal, leisurely pace?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera held still in a medium shot while industrial lights cast a bright yet slightly cold illumination on the scene?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.5762702822685242 + ], + "opposite_logprob_tokens": [ + -0.8262702822685242 + ], + "logprob_reward": -0.5762702822685242, + "yes_no_margin_reward": 0.25, + "token_credit_reward": -0.5762702822685242, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.17575214803218842, + 0.028779402375221252, + 0.05751005932688713, + 0.03574413061141968, + 0.04844805970788002, + 0.026546627283096313, + 0.5778324007987976, + 0.01895718276500702, + 0.015094885602593422, + 0.015335130505263805 + ], + "credit_map": [ + 0.029539339244365692, + 0.026561852544546127, + 0.021940510720014572, + 0.023897269740700722, + 0.02198299951851368, + 0.022215044125914574, + 0.016217591241002083, + 0.013397532515227795, + 0.0036884876899421215, + 0.002621255349367857, + 0.004767121281474829, + 0.004950190894305706, + 0.0025361692532896996, + 0.003071171697229147, + 0.002928483998402953, + 0.004216519650071859, + 0.011852867901325226, + 0.007892084307968616, + 0.01218412071466446, + 0.009456834755837917, + 0.0029721150640398264, + 0.005081394221633673, + 0.004096733406186104, + 0.003973904997110367, + 0.0029820934869349003, + 0.0035302196629345417, + 0.004367447923868895, + 0.00466771237552166, + 0.0042229206301271915, + 0.003727658186107874, + 0.0054594650864601135, + 0.00678660674020648, + 0.007590968161821365, + 0.006632912904024124, + 0.007786287926137447, + 0.005298492964357138, + 0.005113536491990089, + 0.004167422652244568, + 0.00671748910099268, + 0.005140946712344885, + 0.0017645899206399918, + 0.0032113667111843824, + 0.00356568512506783, + 0.0040082489140331745, + 0.003623749827966094, + 0.0023062254767864943, + 0.004868804011493921, + 0.003197955433279276, + 0.04076109081506729, + 0.03216508403420448, + 0.15391501784324646, + 0.20827844738960266, + 0.029740363359451294, + 0.029369845986366272, + 0.04019025340676308, + 0.04341225326061249, + 0.0022496855817735195, + 0.002873919904232025, + 0.002211632439866662, + 0.0017651263624429703, + 0.001392447855323553, + 0.002274399157613516, + 0.0026532993651926517, + 0.003536670468747616, + 0.0016733042430132627, + 0.0011979499831795692, + 0.0037172746378928423, + 0.0037961152847856283, + 0.0011586342006921768, + 0.0008862524409778416, + 0.0013825640780851245, + 0.0012827898608520627, + 0.0017475482309237123, + 0.0015936152776703238, + 0.002600456587970257, + 0.00197256519459188, + 0.0015396022936329246, + 0.0017896884819492698, + 0.0022308293264359236, + 0.0018608237151056528 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 0, + 2 + ], + "selected_reward": -0.5762702822685242 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set on a vast, bustling train platform beneath a high, vaulted ceiling?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a young woman visible standing still in the center of the frame wearing a long dark trench coat and a beige scarf?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are ghostly figures dragging suitcases and carrying backpacks moving in a frantic, accelerated blur around the woman?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the woman clutching a thick hardcover book and turning its yellowed pages at a normal, leisurely pace?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera held still in a medium shot while industrial lights cast a bright yet slightly cold illumination on the scene?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.023305892944335938 + ], + "opposite_logprob_tokens": [ + -3.773305892944336 + ], + "logprob_reward": -0.023305892944335938, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.023305892944335938, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.043106164783239365, + 0.03777455911040306, + 0.03324607387185097, + 0.021258171647787094, + 0.029695337638258934, + 0.02627669647336006, + 0.7564668655395508, + 0.012018617242574692, + 0.012301184237003326, + 0.02785635180771351 + ], + "credit_map": [ + 0.004851913079619408, + 0.005412287078797817, + 0.0056532034650444984, + 0.0064010657370090485, + 0.004338852129876614, + 0.00566148804500699, + 0.004904264584183693, + 0.005883088335394859, + 0.0035194428637623787, + 0.002643461339175701, + 0.006167550105601549, + 0.00572350574657321, + 0.004641816485673189, + 0.0033545326441526413, + 0.006599958520382643, + 0.0051242895424366, + 0.005311970133334398, + 0.003990457393229008, + 0.0071872528642416, + 0.005087577737867832, + 0.002148327184841037, + 0.003127859439700842, + 0.0032300795428454876, + 0.003162553533911705, + 0.002803815994411707, + 0.0022965846583247185, + 0.002339188242331147, + 0.002742768730968237, + 0.001949426718056202, + 0.002147100865840912, + 0.003939519636332989, + 0.003039766801521182, + 0.005287615582346916, + 0.003918052650988102, + 0.0055053140968084335, + 0.004772642627358437, + 0.0020992334466427565, + 0.002453946741297841, + 0.002859660191461444, + 0.002798873232677579, + 0.0024222317151725292, + 0.0033137991558760405, + 0.002786323893815279, + 0.0031167655251920223, + 0.0031367267947643995, + 0.0036800687666982412, + 0.0033373874612152576, + 0.0044833943247795105, + 0.05082101747393608, + 0.040958087891340256, + 0.21183273196220398, + 0.2798295319080353, + 0.03554194048047066, + 0.033655472099781036, + 0.049670007079839706, + 0.05415809527039528, + 0.0011559802806004882, + 0.0011254707351326942, + 0.0013060065684840083, + 0.0016806842759251595, + 0.0010016625747084618, + 0.0008388683781959116, + 0.0029497856739908457, + 0.001960159046575427, + 0.001322686206549406, + 0.0017126757884398103, + 0.0016817405121400952, + 0.0015404813457280397, + 0.0018419168191030622, + 0.0012186586391180754, + 0.0013187918812036514, + 0.0016642329283058643, + 0.002739453222602606, + 0.002491001272574067, + 0.004497032146900892, + 0.003695859806612134, + 0.0027748579159379005, + 0.003256067866459489, + 0.004434156697243452, + 0.003967920783907175 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 0, + 1 + ], + "selected_reward": -0.023305892944335938 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "slow-motion, time-lapse medium shot, the camera held still, captures a vast, bustling train platform. Beneath a high, vaulted ceiling, a blur of ghostly figures, dragging suitcases and carrying backpacks, surges in a frantic, accelerated pattern, creating a chaotic, blurry mess of colors. Trains enter and depart the station at breakneck speed, their lights casting long trails of light across the frame. Amidst this surging sea, a young woman stands as the sole still anchor. She wears a long, dark trench coat, a beige scarf neatly wrapped around her neck. Industrial lights arrayed on the platform's ceiling cast a bright yet slightly cold light, framing her focused profile. Completely oblivious to the surrounding clamor, she clutches a thick, hardcover book in her hands, her long, dark brown hair casually tied back. In stark contrast to the accelerated movement around her, her fingers gently turn the yellowed pages, one by one, at a normal, leisurely pace. This tiny, steady and focused movement creates an oasis of tranquility in this silent and busy scene. The entire picture builds a strong dramatic tension between the cool background and the focused warmth emanating from her reading, creating a surreal and thought-provoking atmosphere.", + "seed": 12915, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000034/group_9/index_4000_seed_12915_dp_18_rank_36_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 9, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.1830480143427849, + "vlm_logprob": -0.1830480143427849, + "vlm_yes_no_margin": 3.325, + "vlm_token_credit": -0.1830480143427849, + "vlm_yes_ratio": 0.8, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set on a vast, bustling train platform beneath a high, vaulted ceiling?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a young woman visible standing still in the center of the frame wearing a long dark trench coat and a beige scarf?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are ghostly figures dragging suitcases and carrying backpacks moving in a frantic, accelerated blur around the woman?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the woman clutching a thick hardcover book and turning its yellowed pages at a normal, leisurely pace?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera held still in a medium shot while industrial lights cast a bright yet slightly cold illumination on the scene?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.020171504467725754 + ], + "opposite_logprob_tokens": [ + -5.520171642303467 + ], + "logprob_reward": -0.020171504467725754, + "yes_no_margin_reward": 5.5, + "token_credit_reward": -0.020171504467725754, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.18813516199588776, + 0.03430861607193947, + 0.050672050565481186, + 0.5025400519371033, + 0.017842689529061317, + 0.015083431266248226, + 0.013299594633281231, + 0.045529309660196304, + 0.02315778099000454, + 0.10943129658699036 + ], + "credit_map": [ + 0.022293949499726295, + 0.023559754714369774, + 0.028225917369127274, + 0.021278727799654007, + 0.019343048334121704, + 0.030895892530679703, + 0.019162621349096298, + 0.02337525598704815, + 0.0037281648255884647, + 0.0027704641688615084, + 0.007467326708137989, + 0.0031998655758798122, + 0.006855915766209364, + 0.0032961121760308743, + 0.003224120708182454, + 0.0037666456773877144, + 0.004959103651344776, + 0.0043748472817242146, + 0.011894283816218376, + 0.00514308363199234, + 0.0067871129140257835, + 0.003577613504603505, + 0.007583588361740112, + 0.006352418567985296, + 0.02215496450662613, + 0.02842637337744236, + 0.03326449915766716, + 0.18846675753593445, + 0.03389390558004379, + 0.01493894588202238, + 0.12548869848251343, + 0.05590587481856346, + 0.0025380931328982115, + 0.002241296460852027, + 0.002464354271069169, + 0.0020503904670476913, + 0.001981221605092287, + 0.0026045162230730057, + 0.0017331123817712069, + 0.0022297045215964317, + 0.0023285902570933104, + 0.002142483601346612, + 0.0029047185089439154, + 0.0015292993048205972, + 0.0014278854941949248, + 0.0013961512595415115, + 0.001700061489827931, + 0.001654242048971355, + 0.0018276679329574108, + 0.0016660583205521107, + 0.0018879969138652086, + 0.0011695241555571556, + 0.0018292769091203809, + 0.0014158131089061499, + 0.001351886778138578, + 0.002151370281353593, + 0.003364854259416461, + 0.0019379407167434692, + 0.0040871077217161655, + 0.002214340493083, + 0.004078458994626999, + 0.0037685157731175423, + 0.01283237338066101, + 0.013245715759694576, + 0.0018652444705367088, + 0.0022591787856072187, + 0.005123802460730076, + 0.002826458076015115, + 0.0032695478294044733, + 0.0016113526653498411, + 0.003656438086181879, + 0.002545759314671159, + 0.005529580172151327, + 0.007510531228035688, + 0.014354185201227665, + 0.04138860106468201, + 0.007674383465200663, + 0.004583501257002354, + 0.02176157757639885, + 0.00662893895059824 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 0, + 9 + ], + "selected_reward": -0.020171504467725754 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set on a vast, bustling train platform beneath a high, vaulted ceiling?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a young woman visible standing still in the center of the frame wearing a long dark trench coat and a beige scarf?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are ghostly figures dragging suitcases and carrying backpacks moving in a frantic, accelerated blur around the woman?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the woman clutching a thick hardcover book and turning its yellowed pages at a normal, leisurely pace?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera held still in a medium shot while industrial lights cast a bright yet slightly cold illumination on the scene?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.026367560029029846 + ], + "opposite_logprob_tokens": [ + -3.651367664337158 + ], + "logprob_reward": -0.026367560029029846, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.026367560029029846, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08810150623321533, + 0.12635047733783722, + 0.11941950768232346, + 0.34622296690940857, + 0.058854840695858, + 0.0583660826086998, + 0.038495033979415894, + 0.05811944231390953, + 0.05202822387218475, + 0.05404188856482506 + ], + "credit_map": [ + 0.01171081606298685, + 0.008725015446543694, + 0.01227613165974617, + 0.012849027290940285, + 0.012129503302276134, + 0.010346266441047192, + 0.01078527420759201, + 0.009279473684728146, + 0.009983561933040619, + 0.01054766122251749, + 0.02599864825606346, + 0.011696490459144115, + 0.033717960119247437, + 0.010053587146103382, + 0.013906565494835377, + 0.010446006432175636, + 0.011940766125917435, + 0.007280650548636913, + 0.03661154955625534, + 0.010194054804742336, + 0.00986527930945158, + 0.005797863472253084, + 0.027546828612685204, + 0.010182525031268597, + 0.016769809648394585, + 0.02008648030459881, + 0.025463271886110306, + 0.12136242538690567, + 0.02509358897805214, + 0.011824282817542553, + 0.08634927123785019, + 0.03927384316921234, + 0.005091528408229351, + 0.005761748179793358, + 0.009920423850417137, + 0.0072823092341423035, + 0.011677606962621212, + 0.004172177519649267, + 0.00862954556941986, + 0.006319501902908087, + 0.00545803876593709, + 0.0053380499593913555, + 0.007248516194522381, + 0.006227828562259674, + 0.012944743037223816, + 0.003370904130861163, + 0.01023324579000473, + 0.007544761057943106, + 0.0038812642451375723, + 0.0037213501054793596, + 0.005965627264231443, + 0.0036264066584408283, + 0.009124787524342537, + 0.0025946754030883312, + 0.004789402242749929, + 0.0047915223985910416, + 0.004288392141461372, + 0.003560819197446108, + 0.004695510026067495, + 0.005119367968291044, + 0.008756469003856182, + 0.0035102798137813807, + 0.01707516983151436, + 0.01111343502998352, + 0.0036306637339293957, + 0.004580517299473286, + 0.007866928353905678, + 0.0054700132459402084, + 0.01500751543790102, + 0.0028411326929926872, + 0.007196650840342045, + 0.005434802267700434, + 0.004934322088956833, + 0.005052684340626001, + 0.010139625519514084, + 0.005688359960913658, + 0.008513311855494976, + 0.004262639209628105, + 0.00908950250595808, + 0.006361445877701044 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 1, + 2 + ], + "selected_reward": -0.026367560029029846 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set on a vast, bustling train platform beneath a high, vaulted ceiling?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a young woman visible standing still in the center of the frame wearing a long dark trench coat and a beige scarf?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are ghostly figures dragging suitcases and carrying backpacks moving in a frantic, accelerated blur around the woman?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the woman clutching a thick hardcover book and turning its yellowed pages at a normal, leisurely pace?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera held still in a medium shot while industrial lights cast a bright yet slightly cold illumination on the scene?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.026377543807029724 + ], + "opposite_logprob_tokens": [ + -3.6513774394989014 + ], + "logprob_reward": -0.026377543807029724, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.026377543807029724, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.03489776700735092, + 0.05246961489319801, + 0.11846008896827698, + 0.6440935730934143, + 0.032098188996315, + 0.01075436919927597, + 0.012775501236319542, + 0.02810780704021454, + 0.021045690402388573, + 0.04529736563563347 + ], + "credit_map": [ + 0.003554057562723756, + 0.003077099099755287, + 0.004442861769348383, + 0.004932770039886236, + 0.0042541432194411755, + 0.00362210045568645, + 0.005177080165594816, + 0.005837657023221254, + 0.004705250728875399, + 0.004991848487406969, + 0.010939173400402069, + 0.003969759680330753, + 0.015617966651916504, + 0.004359952639788389, + 0.004185640253126621, + 0.0037000211887061596, + 0.010531369596719742, + 0.007102062460035086, + 0.04058130830526352, + 0.008348623290657997, + 0.007486003451049328, + 0.0050099617801606655, + 0.031218746677041054, + 0.00818201620131731, + 0.028611646965146065, + 0.03728949651122093, + 0.043651748448610306, + 0.23818644881248474, + 0.04538959637284279, + 0.018937209621071815, + 0.16007554531097412, + 0.07195186614990234, + 0.004840322304517031, + 0.0029253929387778044, + 0.006674539297819138, + 0.003194129094481468, + 0.003977987449616194, + 0.002648742403835058, + 0.004243752919137478, + 0.0035933235194534063, + 0.0012834842782467604, + 0.0009470601798966527, + 0.0014777958858758211, + 0.0014839923242107034, + 0.0016264314763247967, + 0.0009631082066334784, + 0.0017120613483712077, + 0.0012604352086782455, + 0.0016577279893681407, + 0.0017814490711316466, + 0.0019448469392955303, + 0.002014199970290065, + 0.0015566812362521887, + 0.0010869590332731605, + 0.001105701201595366, + 0.0016279362607747316, + 0.002073840703815222, + 0.0010811545653268695, + 0.0024224906228482723, + 0.001659295172430575, + 0.002475626766681671, + 0.0018941649468615651, + 0.009928101673722267, + 0.0065731327049434185, + 0.0017411388689652085, + 0.0023062212858349085, + 0.0029960772953927517, + 0.0019006491638720036, + 0.004103797022253275, + 0.0015837847022339702, + 0.0029019524808973074, + 0.003512068884447217, + 0.0028158549685031176, + 0.0038234887178987265, + 0.009895019233226776, + 0.01651103049516678, + 0.0020668143406510353, + 0.0016222121194005013, + 0.0063807968981564045, + 0.002182150725275278 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 2, + 1 + ], + "selected_reward": -0.026377543807029724 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set on a vast, bustling train platform beneath a high, vaulted ceiling?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a young woman visible standing still in the center of the frame wearing a long dark trench coat and a beige scarf?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are ghostly figures dragging suitcases and carrying backpacks moving in a frantic, accelerated blur around the woman?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the woman clutching a thick hardcover book and turning its yellowed pages at a normal, leisurely pace?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera held still in a medium shot while industrial lights cast a bright yet slightly cold illumination on the scene?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -0.8262374997138977 + ], + "opposite_logprob_tokens": [ + -0.5762374997138977 + ], + "logprob_reward": -0.8262374997138977, + "yes_no_margin_reward": -0.25, + "token_credit_reward": -0.8262374997138977, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.12489188462495804, + 0.06663545221090317, + 0.4338146150112152, + 0.10492412745952606, + 0.057299528270959854, + 0.017201917245984077, + 0.02776961587369442, + 0.029218383133411407, + 0.024790218099951744, + 0.11345425993204117 + ], + "credit_map": [ + 0.00929717905819416, + 0.013438490219414234, + 0.019250135868787766, + 0.01985098607838154, + 0.019320417195558548, + 0.013381706550717354, + 0.016085684299468994, + 0.01426728442311287, + 0.005467989016324282, + 0.0058539630845189095, + 0.01548545341938734, + 0.00534041877835989, + 0.014188441447913647, + 0.005750998388975859, + 0.007738924585282803, + 0.006809257436543703, + 0.027176354080438614, + 0.016234366223216057, + 0.18725688755512238, + 0.023447316139936447, + 0.022207288071513176, + 0.011808037757873535, + 0.11965253204107285, + 0.02603183686733246, + 0.006365559529513121, + 0.007838180288672447, + 0.012510815635323524, + 0.02961687557399273, + 0.012757625430822372, + 0.0052855294197797775, + 0.0167359821498394, + 0.013813553377985954, + 0.007473193109035492, + 0.0061673750169575214, + 0.01054896879941225, + 0.005594950634986162, + 0.009267715737223625, + 0.005242548417299986, + 0.00667624780908227, + 0.006328525487333536, + 0.002541319699957967, + 0.0018125454662367702, + 0.0022065455559641123, + 0.00201890105381608, + 0.002557337051257491, + 0.0016940669156610966, + 0.0020007118582725525, + 0.002370490226894617, + 0.002567120362073183, + 0.0037754483055323362, + 0.003983659669756889, + 0.0038865702226758003, + 0.004665028303861618, + 0.0026291818358004093, + 0.003039914881810546, + 0.0032226908951997757, + 0.0016519736964255571, + 0.002447644015774131, + 0.004133672919124365, + 0.0053262170404195786, + 0.005432541482150555, + 0.0028451671823859215, + 0.003539272816851735, + 0.003841893980279565, + 0.001594232628121972, + 0.0020743566565215588, + 0.0027326380368322134, + 0.004055532161146402, + 0.005910076666623354, + 0.0021806054282933474, + 0.00344555894844234, + 0.0027972166426479816, + 0.005785015411674976, + 0.008348763920366764, + 0.014381044544279575, + 0.04548276215791702, + 0.004901400301605463, + 0.004427799955010414, + 0.023319760337471962, + 0.006807713769376278 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 0, + 9 + ], + "selected_reward": -0.8262374997138977 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set on a vast, bustling train platform beneath a high, vaulted ceiling?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a young woman visible standing still in the center of the frame wearing a long dark trench coat and a beige scarf?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are ghostly figures dragging suitcases and carrying backpacks moving in a frantic, accelerated blur around the woman?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the woman clutching a thick hardcover book and turning its yellowed pages at a normal, leisurely pace?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera held still in a medium shot while industrial lights cast a bright yet slightly cold illumination on the scene?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01608596369624138 + ], + "opposite_logprob_tokens": [ + -4.141086101531982 + ], + "logprob_reward": -0.01608596369624138, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.01608596369624138, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0794597938656807, + 0.04244630038738251, + 0.24373877048492432, + 0.4725235402584076, + 0.03535677492618561, + 0.02704780548810959, + 0.03374464064836502, + 0.033435188233852386, + 0.01680798828601837, + 0.015439189970493317 + ], + "credit_map": [ + 0.00699192238971591, + 0.009131275117397308, + 0.013514603488147259, + 0.008603336289525032, + 0.009342849254608154, + 0.009555421769618988, + 0.012116885744035244, + 0.010203496553003788, + 0.004714548587799072, + 0.005224341060966253, + 0.00695947278290987, + 0.004221349023282528, + 0.005394316278398037, + 0.0036979075521230698, + 0.006234617903828621, + 0.005999748595058918, + 0.01688672974705696, + 0.01001423317939043, + 0.1010034829378128, + 0.012902618385851383, + 0.015460304915904999, + 0.008219372481107712, + 0.06004031002521515, + 0.019211700186133385, + 0.0199237372726202, + 0.02808733470737934, + 0.03208162263035774, + 0.17970527708530426, + 0.029891250655055046, + 0.014163974672555923, + 0.11410722136497498, + 0.05456315353512764, + 0.0045663039200007915, + 0.00349046615883708, + 0.0062922146171331406, + 0.00432377215474844, + 0.004289049655199051, + 0.004277318716049194, + 0.003808899549767375, + 0.00430874852463603, + 0.003383777802810073, + 0.0021713385358452797, + 0.005078545305877924, + 0.0024771320167928934, + 0.005042603705078363, + 0.0018906797049567103, + 0.003751791547983885, + 0.0032519344240427017, + 0.0032055878546088934, + 0.0051129707135260105, + 0.00513426773250103, + 0.004231196828186512, + 0.005145334172993898, + 0.0029177728574723005, + 0.004327266477048397, + 0.0036702428478747606, + 0.0023941805120557547, + 0.00213563721626997, + 0.003027980914339423, + 0.0023162320721894503, + 0.0030361555982381105, + 0.0034008158836513758, + 0.008875356987118721, + 0.008248827420175076, + 0.001657848246395588, + 0.0016655671643093228, + 0.0020203187596052885, + 0.0020704553462564945, + 0.0030773852486163378, + 0.001319437869824469, + 0.0023749147076159716, + 0.0026220621075481176, + 0.0013072133297100663, + 0.001688759308308363, + 0.002978766104206443, + 0.002081566257402301, + 0.0018902517622336745, + 0.0016481542261317372, + 0.0019516079919412732, + 0.0018928711069747806 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 2, + 0 + ], + "selected_reward": -0.01608596369624138 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "No", + "Yes" + ] + }, + { + "prompt": "slow-motion, time-lapse medium shot, the camera held still, captures a vast, bustling train platform. Beneath a high, vaulted ceiling, a blur of ghostly figures, dragging suitcases and carrying backpacks, surges in a frantic, accelerated pattern, creating a chaotic, blurry mess of colors. Trains enter and depart the station at breakneck speed, their lights casting long trails of light across the frame. Amidst this surging sea, a young woman stands as the sole still anchor. She wears a long, dark trench coat, a beige scarf neatly wrapped around her neck. Industrial lights arrayed on the platform's ceiling cast a bright yet slightly cold light, framing her focused profile. Completely oblivious to the surrounding clamor, she clutches a thick, hardcover book in her hands, her long, dark brown hair casually tied back. In stark contrast to the accelerated movement around her, her fingers gently turn the yellowed pages, one by one, at a normal, leisurely pace. This tiny, steady and focused movement creates an oasis of tranquility in this silent and busy scene. The entire picture builds a strong dramatic tension between the cool background and the focused warmth emanating from her reading, creating a surreal and thought-provoking atmosphere.", + "seed": 12916, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000034/group_9/index_4000_seed_12916_dp_19_rank_38_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 9, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.19742219783365728, + "vlm_logprob": -0.19742219783365728, + "vlm_yes_no_margin": 2.25, + "vlm_token_credit": -0.19742219783365728, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set on a vast, bustling train platform beneath a high, vaulted ceiling?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a young woman visible standing still in the center of the frame wearing a long dark trench coat and a beige scarf?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are ghostly figures dragging suitcases and carrying backpacks moving in a frantic, accelerated blur around the woman?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the woman clutching a thick hardcover book and turning its yellowed pages at a normal, leisurely pace?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera held still in a medium shot while industrial lights cast a bright yet slightly cold illumination on the scene?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.3456677496433258 + ], + "opposite_logprob_tokens": [ + -1.3456677198410034 + ], + "logprob_reward": -0.3456677496433258, + "yes_no_margin_reward": 1.0, + "token_credit_reward": -0.3456677496433258, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1737951785326004, + 0.101678766310215, + 0.07664866000413895, + 0.25624433159828186, + 0.030898593366146088, + 0.07128263264894485, + 0.09004347026348114, + 0.026408398523926735, + 0.03349650651216507, + 0.13950350880622864 + ], + "credit_map": [ + 0.03594331815838814, + 0.01896149106323719, + 0.02181142382323742, + 0.021732032299041748, + 0.02889423258602619, + 0.014634341932833195, + 0.015836065635085106, + 0.0159822516143322, + 0.007902942597866058, + 0.010459464974701405, + 0.010348595678806305, + 0.027173347771167755, + 0.00480262516066432, + 0.007996750995516777, + 0.008380318991839886, + 0.024614715948700905, + 0.004192702937871218, + 0.006412551272660494, + 0.008128885179758072, + 0.015533020719885826, + 0.0054028877057135105, + 0.006418179254978895, + 0.006854359060525894, + 0.02370607480406761, + 0.017880506813526154, + 0.01787377893924713, + 0.03385162353515625, + 0.037914346903562546, + 0.028130177408456802, + 0.015488688834011555, + 0.08459204435348511, + 0.020513152703642845, + 0.0034192255698144436, + 0.003562362166121602, + 0.0028033319395035505, + 0.005774331744760275, + 0.003141341730952263, + 0.002642706036567688, + 0.005080178380012512, + 0.004475113470107317, + 0.005650094710290432, + 0.006000935565680265, + 0.00730056781321764, + 0.017350049689412117, + 0.005668532103300095, + 0.009080230258405209, + 0.010008005425333977, + 0.010224215686321259, + 0.005254015326499939, + 0.008705785498023033, + 0.009107548743486404, + 0.02113076113164425, + 0.004684946499764919, + 0.006956887897104025, + 0.005263575352728367, + 0.02893993817269802, + 0.0019983863458037376, + 0.0025126601103693247, + 0.0035330839455127716, + 0.004940268583595753, + 0.0013886189553886652, + 0.0020108954049646854, + 0.001517494791187346, + 0.008506988175213337, + 0.00296412012539804, + 0.00344251305796206, + 0.0030441407579928637, + 0.007639798801392317, + 0.002745340345427394, + 0.002325829118490219, + 0.00778950983658433, + 0.0035452537704259157, + 0.017787272110581398, + 0.01042004581540823, + 0.019014982506632805, + 0.010719668120145798, + 0.019521445035934448, + 0.02377266436815262, + 0.013556298799812794, + 0.0247111264616251 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 0, + 9 + ], + "selected_reward": -0.3456677496433258 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set on a vast, bustling train platform beneath a high, vaulted ceiling?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a young woman visible standing still in the center of the frame wearing a long dark trench coat and a beige scarf?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are ghostly figures dragging suitcases and carrying backpacks moving in a frantic, accelerated blur around the woman?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the woman clutching a thick hardcover book and turning its yellowed pages at a normal, leisurely pace?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera held still in a medium shot while industrial lights cast a bright yet slightly cold illumination on the scene?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.048696789890527725 + ], + "opposite_logprob_tokens": [ + -3.048696756362915 + ], + "logprob_reward": -0.048696789890527725, + "yes_no_margin_reward": 3.0, + "token_credit_reward": -0.048696789890527725, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15582171082496643, + 0.1836203634738922, + 0.08933907747268677, + 0.10422872751951218, + 0.15694740414619446, + 0.09412848949432373, + 0.07063566148281097, + 0.02327285148203373, + 0.04223628714680672, + 0.07976934313774109 + ], + "credit_map": [ + 0.02129259891808033, + 0.03242199122905731, + 0.01593882031738758, + 0.025607971474528313, + 0.02127709984779358, + 0.011169725097715855, + 0.01551386434584856, + 0.012599648907780647, + 0.013333741575479507, + 0.020873980596661568, + 0.01835744082927704, + 0.046351585537195206, + 0.006673175375908613, + 0.013610525988042355, + 0.0131731778383255, + 0.051246773451566696, + 0.006509056780487299, + 0.009397048503160477, + 0.01012251153588295, + 0.015620511025190353, + 0.00805630348622799, + 0.006696860771626234, + 0.012080108746886253, + 0.020856689661741257, + 0.007554023526608944, + 0.00931055098772049, + 0.011999184265732765, + 0.019895315170288086, + 0.007493242155760527, + 0.006682897452265024, + 0.03485570475459099, + 0.006437824107706547, + 0.01135835237801075, + 0.04424453526735306, + 0.009832845069468021, + 0.023269593715667725, + 0.01263711042702198, + 0.00705820182338357, + 0.03732852265238762, + 0.01121826097369194, + 0.00644397409632802, + 0.008011018857359886, + 0.007240147329866886, + 0.01010427251458168, + 0.015978537499904633, + 0.02057676762342453, + 0.008770077489316463, + 0.017003701999783516, + 0.0049178642220795155, + 0.008550528436899185, + 0.008854120038449764, + 0.017546415328979492, + 0.003928885329514742, + 0.004354986362159252, + 0.005263540428131819, + 0.01721932739019394, + 0.0033642337657511234, + 0.002258985536172986, + 0.003326478647068143, + 0.0024797660298645496, + 0.0016904682852327824, + 0.0015543432673439384, + 0.002251021796837449, + 0.006347557995468378, + 0.0032224534079432487, + 0.004526755306869745, + 0.003845764556899667, + 0.008615645579993725, + 0.003451431868597865, + 0.0033779211807996035, + 0.009273719973862171, + 0.005922600161284208, + 0.007349306717514992, + 0.0094126146286726, + 0.014707370661199093, + 0.007467301096767187, + 0.011149068363010883, + 0.006557016633450985, + 0.010563788004219532, + 0.012562890537083149 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 4, + 0 + ], + "selected_reward": -0.048696789890527725 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set on a vast, bustling train platform beneath a high, vaulted ceiling?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a young woman visible standing still in the center of the frame wearing a long dark trench coat and a beige scarf?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are ghostly figures dragging suitcases and carrying backpacks moving in a frantic, accelerated blur around the woman?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the woman clutching a thick hardcover book and turning its yellowed pages at a normal, leisurely pace?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera held still in a medium shot while industrial lights cast a bright yet slightly cold illumination on the scene?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.048706669360399246 + ], + "opposite_logprob_tokens": [ + -3.0487067699432373 + ], + "logprob_reward": -0.048706669360399246, + "yes_no_margin_reward": 3.0, + "token_credit_reward": -0.048706669360399246, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.22757036983966827, + 0.13862936198711395, + 0.07816223055124283, + 0.08927249163389206, + 0.08930738270282745, + 0.07473400980234146, + 0.08367403596639633, + 0.042563002556562424, + 0.07892801612615585, + 0.09715904295444489 + ], + "credit_map": [ + 0.04606376588344574, + 0.02747390978038311, + 0.025769514963030815, + 0.0323026180267334, + 0.01940988190472126, + 0.01571381650865078, + 0.039337046444416046, + 0.02149982377886772, + 0.011009388603270054, + 0.013862953521311283, + 0.01579163409769535, + 0.036981817334890366, + 0.005879259202629328, + 0.010560153983533382, + 0.009545117616653442, + 0.034999046474695206, + 0.009706133045256138, + 0.010129581205546856, + 0.016144413501024246, + 0.011945054866373539, + 0.008065073750913143, + 0.006138873752206564, + 0.00838382262736559, + 0.007649281527847052, + 0.007450788281857967, + 0.010165329091250896, + 0.013454778119921684, + 0.01029579620808363, + 0.009673201479017735, + 0.010218409821391106, + 0.015062250196933746, + 0.012951934710144997, + 0.008448113687336445, + 0.020294027402997017, + 0.012173804454505444, + 0.012583401054143906, + 0.006216532085090876, + 0.0047796485014259815, + 0.018567252904176712, + 0.00624459981918335, + 0.005237105302512646, + 0.005731901153922081, + 0.005248392466455698, + 0.009444162249565125, + 0.009400096721947193, + 0.01628182828426361, + 0.015558621846139431, + 0.00783189944922924, + 0.0056538768112659454, + 0.008187930099666119, + 0.00794096477329731, + 0.015584499575197697, + 0.0054912082850933075, + 0.007131156977266073, + 0.005278476979583502, + 0.028405915945768356, + 0.003496574703603983, + 0.004055839031934738, + 0.008532458916306496, + 0.0030092899687588215, + 0.0040548378601670265, + 0.0036543149035423994, + 0.0032622208818793297, + 0.01249746698886156, + 0.006860456429421902, + 0.010154902003705502, + 0.0072311656549572945, + 0.01603494957089424, + 0.005953832995146513, + 0.007428967393934727, + 0.013779688626527786, + 0.011484055779874325, + 0.008389388211071491, + 0.011286468245089054, + 0.019024688750505447, + 0.011959033086895943, + 0.010881904512643814, + 0.01276322454214096, + 0.010784756392240524, + 0.012069580145180225 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 9 + ], + "selected_reward": -0.048706669360399246 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set on a vast, bustling train platform beneath a high, vaulted ceiling?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a young woman visible standing still in the center of the frame wearing a long dark trench coat and a beige scarf?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are ghostly figures dragging suitcases and carrying backpacks moving in a frantic, accelerated blur around the woman?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the woman clutching a thick hardcover book and turning its yellowed pages at a normal, leisurely pace?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera held still in a medium shot while industrial lights cast a bright yet slightly cold illumination on the scene?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.5234405398368835 + ], + "opposite_logprob_tokens": [ + -0.8984405398368835 + ], + "logprob_reward": -0.5234405398368835, + "yes_no_margin_reward": 0.375, + "token_credit_reward": -0.5234405398368835, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1319621503353119, + 0.09818445146083832, + 0.16743910312652588, + 0.06395111232995987, + 0.07895856350660324, + 0.0633019208908081, + 0.03954433649778366, + 0.12448655068874359, + 0.1643296331167221, + 0.06784215569496155 + ], + "credit_map": [ + 0.02727022022008896, + 0.019339727237820625, + 0.018469005823135376, + 0.015544281341135502, + 0.014309408143162727, + 0.010283486917614937, + 0.012914841994643211, + 0.013831191696226597, + 0.006888378411531448, + 0.010904490016400814, + 0.012190357781946659, + 0.021063921973109245, + 0.004312431905418634, + 0.009202762506902218, + 0.009269209578633308, + 0.02435290440917015, + 0.008146844804286957, + 0.013641005381941795, + 0.01485469937324524, + 0.030501026660203934, + 0.011540408246219158, + 0.014913035556674004, + 0.020855998620390892, + 0.05298609286546707, + 0.004163057543337345, + 0.005836957134306431, + 0.00948947574943304, + 0.009342067874968052, + 0.005329250358045101, + 0.007066999562084675, + 0.017120273783802986, + 0.0056030405685305595, + 0.006504491902887821, + 0.019245782867074013, + 0.007860813289880753, + 0.013472666963934898, + 0.006025351583957672, + 0.003843204816803336, + 0.014784598723053932, + 0.007221655920147896, + 0.004449836444109678, + 0.005209306254982948, + 0.0051304916851222515, + 0.006452537607401609, + 0.010137408971786499, + 0.014575906097888947, + 0.007956642657518387, + 0.009389791637659073, + 0.0027231101412326097, + 0.003539368975907564, + 0.004244325682520866, + 0.0060967872850596905, + 0.004188390448689461, + 0.0036991641391068697, + 0.003365397220477462, + 0.01168779656291008, + 0.007436578162014484, + 0.008919418789446354, + 0.011882313527166843, + 0.023358086124062538, + 0.007340430747717619, + 0.009098226204514503, + 0.00867252517491579, + 0.04777897894382477, + 0.014163237065076828, + 0.015361993573606014, + 0.014839746057987213, + 0.03516693413257599, + 0.01192773599177599, + 0.01213501300662756, + 0.036873869597911835, + 0.023861121386289597, + 0.008864699862897396, + 0.004595964681357145, + 0.009924736805260181, + 0.00551212951540947, + 0.009312429465353489, + 0.00980503112077713, + 0.007351474370807409, + 0.012475697323679924 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 8, + 0 + ], + "selected_reward": -0.5234405398368835 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set on a vast, bustling train platform beneath a high, vaulted ceiling?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a young woman visible standing still in the center of the frame wearing a long dark trench coat and a beige scarf?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are ghostly figures dragging suitcases and carrying backpacks moving in a frantic, accelerated blur around the woman?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the woman clutching a thick hardcover book and turning its yellowed pages at a normal, leisurely pace?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera held still in a medium shot while industrial lights cast a bright yet slightly cold illumination on the scene?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02059924043715 + ], + "opposite_logprob_tokens": [ + -3.895599126815796 + ], + "logprob_reward": -0.02059924043715, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.02059924043715, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1315789371728897, + 0.0773615837097168, + 0.09328582137823105, + 0.08372977375984192, + 0.07024326175451279, + 0.19609609246253967, + 0.09085088223218918, + 0.08752627670764923, + 0.11498446017503738, + 0.054342858493328094 + ], + "credit_map": [ + 0.014425341971218586, + 0.02468283101916313, + 0.010844001546502113, + 0.024639632552862167, + 0.0155700808390975, + 0.011142830364406109, + 0.01557261124253273, + 0.014701608568429947, + 0.005889946594834328, + 0.013711586594581604, + 0.015809640288352966, + 0.010971431620419025, + 0.006266050040721893, + 0.010578121058642864, + 0.006537232547998428, + 0.007597572170197964, + 0.005847179796546698, + 0.007472914177924395, + 0.011265167966485023, + 0.018687041476368904, + 0.004956081043928862, + 0.006816417910158634, + 0.008984845131635666, + 0.029256165027618408, + 0.006952063180506229, + 0.010884031653404236, + 0.013969243504106998, + 0.009164097718894482, + 0.012139487080276012, + 0.011760874651372433, + 0.00784459337592125, + 0.011015383526682854, + 0.007089283317327499, + 0.011304891668260098, + 0.008238347247242928, + 0.009078770875930786, + 0.008713225834071636, + 0.0061094630509614944, + 0.009902077727019787, + 0.009807197377085686, + 0.012406324036419392, + 0.0159944836050272, + 0.013554207049310207, + 0.03071719966828823, + 0.02750449627637863, + 0.0409427136182785, + 0.02831457555294037, + 0.026662103831768036, + 0.007341809570789337, + 0.01300075463950634, + 0.012104233726859093, + 0.01255195401608944, + 0.0073751406744122505, + 0.006874017417430878, + 0.00567397428676486, + 0.025929002091288567, + 0.0058409683406353, + 0.00819357018917799, + 0.010467865504324436, + 0.015741495415568352, + 0.005253371316939592, + 0.006364638917148113, + 0.004872895777225494, + 0.030791470780968666, + 0.009283475577831268, + 0.010314533486962318, + 0.009142766706645489, + 0.026475008577108383, + 0.008887768723070621, + 0.008876251056790352, + 0.030660290271043777, + 0.011344367638230324, + 0.0047174789942801, + 0.008103604428470135, + 0.01176042202860117, + 0.006770191714167595, + 0.005593423265963793, + 0.004951086826622486, + 0.006359538529068232, + 0.0060871136374771595 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 0, + 8 + ], + "selected_reward": -0.02059924043715 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "slow-motion, time-lapse medium shot, the camera held still, captures a vast, bustling train platform. Beneath a high, vaulted ceiling, a blur of ghostly figures, dragging suitcases and carrying backpacks, surges in a frantic, accelerated pattern, creating a chaotic, blurry mess of colors. Trains enter and depart the station at breakneck speed, their lights casting long trails of light across the frame. Amidst this surging sea, a young woman stands as the sole still anchor. She wears a long, dark trench coat, a beige scarf neatly wrapped around her neck. Industrial lights arrayed on the platform's ceiling cast a bright yet slightly cold light, framing her focused profile. Completely oblivious to the surrounding clamor, she clutches a thick, hardcover book in her hands, her long, dark brown hair casually tied back. In stark contrast to the accelerated movement around her, her fingers gently turn the yellowed pages, one by one, at a normal, leisurely pace. This tiny, steady and focused movement creates an oasis of tranquility in this silent and busy scene. The entire picture builds a strong dramatic tension between the cool background and the focused warmth emanating from her reading, creating a surreal and thought-provoking atmosphere.", + "seed": 12917, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000034/group_9/index_4000_seed_12917_dp_19_rank_38_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 9, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.17632595933973788, + "vlm_logprob": -0.17632595933973788, + "vlm_yes_no_margin": 2.55, + "vlm_token_credit": -0.17632595933973788, + "vlm_yes_ratio": 0.8, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set on a vast, bustling train platform beneath a high, vaulted ceiling?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a young woman visible standing still in the center of the frame wearing a long dark trench coat and a beige scarf?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are ghostly figures dragging suitcases and carrying backpacks moving in a frantic, accelerated blur around the woman?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the woman clutching a thick hardcover book and turning its yellowed pages at a normal, leisurely pace?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera held still in a medium shot while industrial lights cast a bright yet slightly cold illumination on the scene?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.050489846616983414 + ], + "opposite_logprob_tokens": [ + -3.300489902496338 + ], + "logprob_reward": -0.050489846616983414, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.050489846616983414, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11230888217687607, + 0.05709316208958626, + 0.04401809722185135, + 0.04358941316604614, + 0.07256828248500824, + 0.04259481653571129, + 0.13083499670028687, + 0.4782409369945526, + 0.006907729897648096, + 0.011843698099255562 + ], + "credit_map": [ + 0.018978232517838478, + 0.014598341658711433, + 0.014726078137755394, + 0.017263783141970634, + 0.014614028856158257, + 0.008629836142063141, + 0.01222074031829834, + 0.011277836747467518, + 0.011050119064748287, + 0.0058280183002352715, + 0.002927235560491681, + 0.016727950423955917, + 0.00418271915987134, + 0.003874895628541708, + 0.005841753911226988, + 0.006660472135990858, + 0.005399428773671389, + 0.007804931607097387, + 0.007263707462698221, + 0.0048915911465883255, + 0.004093424417078495, + 0.004287784919142723, + 0.0055039129219949245, + 0.004773311782628298, + 0.005530130118131638, + 0.004520098678767681, + 0.005389289930462837, + 0.0031434448901563883, + 0.004746765363961458, + 0.003796827979385853, + 0.006165824364870787, + 0.010297032073140144, + 0.00850182306021452, + 0.008029629476368427, + 0.00936106126755476, + 0.006329281255602837, + 0.007298292592167854, + 0.007489145267754793, + 0.012229779735207558, + 0.013329267501831055, + 0.0034114301670342684, + 0.003263105172663927, + 0.010537369176745415, + 0.0026720205787569284, + 0.013642331585288048, + 0.003105544950813055, + 0.003158383071422577, + 0.0028046313673257828, + 0.01234285719692707, + 0.010861890390515327, + 0.016930043697357178, + 0.013245717622339725, + 0.021791718900203705, + 0.012373083271086216, + 0.02195832133293152, + 0.02133137546479702, + 0.0741422027349472, + 0.05908651649951935, + 0.04758797585964203, + 0.03983551636338234, + 0.050397686660289764, + 0.04261070117354393, + 0.0402434840798378, + 0.1243368536233902, + 0.0009283750550821424, + 0.0009228759445250034, + 0.0008702066843397915, + 0.0009609964909031987, + 0.0008251787512563169, + 0.0007190554752014577, + 0.0009244409739039838, + 0.0007566005806438625, + 0.0016242790734395385, + 0.0014407321577891707, + 0.0019753926899284124, + 0.0016182997496798635, + 0.0012461382430046797, + 0.0008780204225331545, + 0.0013623156119138002, + 0.0016985206166282296 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 6, + 0 + ], + "selected_reward": -0.050489846616983414 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set on a vast, bustling train platform beneath a high, vaulted ceiling?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a young woman visible standing still in the center of the frame wearing a long dark trench coat and a beige scarf?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are ghostly figures dragging suitcases and carrying backpacks moving in a frantic, accelerated blur around the woman?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the woman clutching a thick hardcover book and turning its yellowed pages at a normal, leisurely pace?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera held still in a medium shot while industrial lights cast a bright yet slightly cold illumination on the scene?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.06207539886236191 + ], + "opposite_logprob_tokens": [ + -2.81207537651062 + ], + "logprob_reward": -0.06207539886236191, + "yes_no_margin_reward": 2.75, + "token_credit_reward": -0.06207539886236191, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15686005353927612, + 0.13533584773540497, + 0.1421220302581787, + 0.15198275446891785, + 0.09232683479785919, + 0.07041153311729431, + 0.040851447731256485, + 0.13451027870178223, + 0.028613340109586716, + 0.04698590561747551 + ], + "credit_map": [ + 0.02194172888994217, + 0.03401554375886917, + 0.018000712618231773, + 0.01976410299539566, + 0.01331970002502203, + 0.011302724480628967, + 0.024069732055068016, + 0.014445791952311993, + 0.025319905951619148, + 0.05133967474102974, + 0.009240483865141869, + 0.011742943897843361, + 0.010639146901667118, + 0.015504259616136551, + 0.004783899523317814, + 0.0067655411548912525, + 0.014094077982008457, + 0.04911321401596069, + 0.023675182834267616, + 0.012173007242381573, + 0.011528411880135536, + 0.010017740540206432, + 0.011872846633195877, + 0.009647542610764503, + 0.024594014510512352, + 0.025763826444745064, + 0.019305957481265068, + 0.01468496024608612, + 0.01625220477581024, + 0.014260289259254932, + 0.012655651196837425, + 0.02446584217250347, + 0.011817255057394505, + 0.012813878245651722, + 0.012540665455162525, + 0.01331672165542841, + 0.005552347749471664, + 0.005621103569865227, + 0.011023102328181267, + 0.01964176818728447, + 0.005678202025592327, + 0.01316510047763586, + 0.006635942496359348, + 0.010458254255354404, + 0.01550140418112278, + 0.0038332173135131598, + 0.00944659486413002, + 0.005692813079804182, + 0.004275060724467039, + 0.005557541735470295, + 0.005718722008168697, + 0.007148307282477617, + 0.004560746718198061, + 0.0026840439531952143, + 0.006667869631201029, + 0.004239152651280165, + 0.021036678925156593, + 0.016935046762228012, + 0.015236209146678448, + 0.013530266471207142, + 0.011699272319674492, + 0.009350062347948551, + 0.010275105014443398, + 0.036447636783123016, + 0.003996944520622492, + 0.00396378431469202, + 0.0034040268510580063, + 0.0047407494857907295, + 0.002558173844590783, + 0.0022488266695290804, + 0.003369260812178254, + 0.0043315766379237175, + 0.008035056293010712, + 0.007890923880040646, + 0.007275101263076067, + 0.008870015852153301, + 0.0038086757995188236, + 0.0022551624570041895, + 0.004434514325112104, + 0.0044164517894387245 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 2 + ], + "selected_reward": -0.06207539886236191 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set on a vast, bustling train platform beneath a high, vaulted ceiling?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a young woman visible standing still in the center of the frame wearing a long dark trench coat and a beige scarf?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are ghostly figures dragging suitcases and carrying backpacks moving in a frantic, accelerated blur around the woman?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the woman clutching a thick hardcover book and turning its yellowed pages at a normal, leisurely pace?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera held still in a medium shot while industrial lights cast a bright yet slightly cold illumination on the scene?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.054996609687805176 + ], + "opposite_logprob_tokens": [ + -2.9299964904785156 + ], + "logprob_reward": -0.054996609687805176, + "yes_no_margin_reward": 2.875, + "token_credit_reward": -0.054996609687805176, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.17623542249202728, + 0.15104340016841888, + 0.060962092131376266, + 0.11092159897089005, + 0.03939399495720863, + 0.03495415672659874, + 0.0935828760266304, + 0.23439301550388336, + 0.04187951982021332, + 0.05663396418094635 + ], + "credit_map": [ + 0.020374009385704994, + 0.03977200388908386, + 0.03031156212091446, + 0.029546424746513367, + 0.015602448023855686, + 0.011567494831979275, + 0.014941760338842869, + 0.014119730331003666, + 0.020465221256017685, + 0.036073893308639526, + 0.017329715192317963, + 0.04560832679271698, + 0.007946153171360493, + 0.008488940075039864, + 0.006373289506882429, + 0.00875786691904068, + 0.007344091311097145, + 0.00535869924351573, + 0.006913396064192057, + 0.005568896885961294, + 0.007123942952603102, + 0.008495194837450981, + 0.011077938601374626, + 0.009079930372536182, + 0.016536762937903404, + 0.012108770199120045, + 0.018316395580768585, + 0.00711951544508338, + 0.014004330150783062, + 0.011625399813055992, + 0.010838544927537441, + 0.02037188783288002, + 0.005482811946421862, + 0.0056507219560444355, + 0.006308646872639656, + 0.004058389458805323, + 0.0029546027071774006, + 0.0028905326034873724, + 0.0052530658431351185, + 0.00679522380232811, + 0.004736251663416624, + 0.005522575229406357, + 0.002692319918423891, + 0.004783795215189457, + 0.005389743950217962, + 0.002389943227171898, + 0.005695250350981951, + 0.0037442781031131744, + 0.01107847411185503, + 0.007415791507810354, + 0.01351126004010439, + 0.012338654138147831, + 0.01401479821652174, + 0.00815447699278593, + 0.013956068083643913, + 0.013113357126712799, + 0.03479939326643944, + 0.02989749237895012, + 0.02295004017651081, + 0.02196061797440052, + 0.02134288102388382, + 0.01921108365058899, + 0.02006746456027031, + 0.06416404247283936, + 0.005262309219688177, + 0.005473667755723, + 0.004417816177010536, + 0.008440688252449036, + 0.0053366972133517265, + 0.003280405420809984, + 0.005144280847162008, + 0.004523655865341425, + 0.009135245345532894, + 0.007473968900740147, + 0.010557477362453938, + 0.010258497670292854, + 0.004105162341147661, + 0.003353007836267352, + 0.00553353875875473, + 0.006217067129909992 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 0, + 1 + ], + "selected_reward": -0.054996609687805176 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set on a vast, bustling train platform beneath a high, vaulted ceiling?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a young woman visible standing still in the center of the frame wearing a long dark trench coat and a beige scarf?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are ghostly figures dragging suitcases and carrying backpacks moving in a frantic, accelerated blur around the woman?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the woman clutching a thick hardcover book and turning its yellowed pages at a normal, leisurely pace?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera held still in a medium shot while industrial lights cast a bright yet slightly cold illumination on the scene?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -0.6934663653373718 + ], + "opposite_logprob_tokens": [ + -0.6934663653373718 + ], + "logprob_reward": -0.6934663653373718, + "yes_no_margin_reward": 0.0, + "token_credit_reward": -0.6934663653373718, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.18845705687999725, + 0.24579785764217377, + 0.07966692745685577, + 0.09293007850646973, + 0.04980212077498436, + 0.08493377268314362, + 0.057293470948934555, + 0.10146833956241608, + 0.036704372614622116, + 0.06294602900743484 + ], + "credit_map": [ + 0.03654912859201431, + 0.054769787937402725, + 0.018670378252863884, + 0.022301290184259415, + 0.01442804280668497, + 0.01390095055103302, + 0.013374879024922848, + 0.014462590217590332, + 0.03963282331824303, + 0.02577257715165615, + 0.02057022787630558, + 0.10689317435026169, + 0.010938762687146664, + 0.010936349630355835, + 0.013191409409046173, + 0.017862502485513687, + 0.01010835263878107, + 0.01666140928864479, + 0.015553933568298817, + 0.010345456190407276, + 0.007409867364913225, + 0.005147940944880247, + 0.009112954139709473, + 0.005327004007995129, + 0.020849628373980522, + 0.01438525877892971, + 0.016769250854849815, + 0.009011668153107166, + 0.00756941270083189, + 0.00535004073753953, + 0.008713684976100922, + 0.010281125083565712, + 0.006338943727314472, + 0.010803434066474438, + 0.007259502541273832, + 0.006866872310638428, + 0.004020772408694029, + 0.0036128670908510685, + 0.005330742336809635, + 0.00556897884234786, + 0.01210858952254057, + 0.013114569708704948, + 0.011178137734532356, + 0.012163831852376461, + 0.015350928530097008, + 0.0059269932098686695, + 0.008482350967824459, + 0.006608360446989536, + 0.006916487589478493, + 0.006347534712404013, + 0.009777572005987167, + 0.0075232801027596, + 0.005976864602416754, + 0.0036547754425555468, + 0.008618880063295364, + 0.008478071540594101, + 0.023064786568284035, + 0.00861972477287054, + 0.008850251324474812, + 0.010272610932588577, + 0.008599626831710339, + 0.006915336009114981, + 0.00783553533256054, + 0.027310457080602646, + 0.003423212328925729, + 0.004856728948652744, + 0.005800806917250156, + 0.007398418616503477, + 0.0036419089883565903, + 0.003100145375356078, + 0.004990428686141968, + 0.0034927213564515114, + 0.007920282892882824, + 0.009787321090698242, + 0.010573923587799072, + 0.015844421461224556, + 0.003678529057651758, + 0.0035125890281051397, + 0.006612327881157398, + 0.005016632843762636 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 7 + ], + "selected_reward": -0.6934663653373718 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set on a vast, bustling train platform beneath a high, vaulted ceiling?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a young woman visible standing still in the center of the frame wearing a long dark trench coat and a beige scarf?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are ghostly figures dragging suitcases and carrying backpacks moving in a frantic, accelerated blur around the woman?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the woman clutching a thick hardcover book and turning its yellowed pages at a normal, leisurely pace?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera held still in a medium shot while industrial lights cast a bright yet slightly cold illumination on the scene?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.020601576194167137 + ], + "opposite_logprob_tokens": [ + -3.895601511001587 + ], + "logprob_reward": -0.020601576194167137, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.020601576194167137, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.21688270568847656, + 0.11449084430932999, + 0.07697569578886032, + 0.051388032734394073, + 0.053952716290950775, + 0.06088913604617119, + 0.055472481995821, + 0.26945653557777405, + 0.03557616472244263, + 0.06491570919752121 + ], + "credit_map": [ + 0.031255025416612625, + 0.06774169206619263, + 0.01370155904442072, + 0.02476918324828148, + 0.017576737329363823, + 0.022120527923107147, + 0.021158883348107338, + 0.018559109419584274, + 0.009615923278033733, + 0.036206506192684174, + 0.016868965700268745, + 0.0232686810195446, + 0.007258811965584755, + 0.009702884592115879, + 0.0056974017061293125, + 0.005871662404388189, + 0.006937155965715647, + 0.014589875005185604, + 0.010756203904747963, + 0.008314423263072968, + 0.0066451095044612885, + 0.006558428518474102, + 0.015770863741636276, + 0.0074036321602761745, + 0.009239096194505692, + 0.010583044961094856, + 0.006612113676965237, + 0.006173793226480484, + 0.003343824530020356, + 0.0037306607700884342, + 0.004639598075300455, + 0.007065901067107916, + 0.00593278044834733, + 0.004603590350598097, + 0.008055949583649635, + 0.006862692069262266, + 0.004846328869462013, + 0.004770885221660137, + 0.010180074721574783, + 0.008700416423380375, + 0.00602798443287611, + 0.013117285445332527, + 0.009566525928676128, + 0.010073577053844929, + 0.007713894359767437, + 0.0041696783155202866, + 0.006268615368753672, + 0.003951573744416237, + 0.007179179694503546, + 0.006648209411650896, + 0.007662989664822817, + 0.011344711296260357, + 0.004053974989801645, + 0.0029686237685382366, + 0.008970140479505062, + 0.0066446526907384396, + 0.04952923581004143, + 0.032137684524059296, + 0.028106944635510445, + 0.034098152071237564, + 0.01630055531859398, + 0.014352260157465935, + 0.021073028445243835, + 0.07385867834091187, + 0.004708544816821814, + 0.00757947051897645, + 0.0037038892041891813, + 0.005582571029663086, + 0.003953751642256975, + 0.0026862826198339462, + 0.004409340675920248, + 0.0029523142147809267, + 0.01075529120862484, + 0.009635640308260918, + 0.010152816772460938, + 0.011104827746748924, + 0.0035502593964338303, + 0.0037821114528924227, + 0.007082679774612188, + 0.008852078579366207 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 0, + 1 + ], + "selected_reward": -0.020601576194167137 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "No", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000050/group_0/index_tensor([8040, 8040])_group_0_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000050/group_0/index_tensor([8040, 8040])_group_0_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..5c212aa849e0cce84bc4fbfa0e3ece9cb37ca9dd --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000050/group_0/index_tensor([8040, 8040])_group_0_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "medium, slow-motion shot, shot from a slightly upward angle, focuses on a muscular, middle-aged woodcutter, in the midst of a powerful downward swing with his axe. He has short, unkempt dark brown hair and a thick beard. Beads of sweat shimmer on his forehead in the sunlight. His facial expression is focused and powerful, his lips tightly pursed. He wears a frayed red and black flannel shirt, sleeves rolled up to the elbows, revealing forearms bulging with veins from exertion. He wears dark blue overalls stained with dirt. He grips a long-handled axe with both hands, the blade gleaming with a cool metallic sheen in the sunlight. He leans back, his legs firmly planted in the soft earth strewn with fallen leaves and broken branches. He is deep in a dense, primeval forest, surrounded by towering, ancient trees. The afternoon sun penetrates the lush canopy, creating dappled shafts of light (a Tyndall effect) that illuminate the dust particles floating in the air. As the axe blade struck the massive stump before him with thunderous precision, captured at an extremely high frame rate, countless sawdust of varying sizes exploded in all directions in elegant, crisp arcs, as if enchanted. The swirling trajectory and texture of each splinter were clearly visible against the backdrop of golden beams. The dominant tones of the image were deep forest green and warm earthy brown, accented by golden sunlight and the woodcutter's vibrant shirt, creating a stark contrast between a sense of primal power, the beauty of focused labor, and the tranquility of the deep forest.", + "seed": 7961, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000050/group_0/index_8040_seed_7961_dp_0_rank_0_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 0, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.0071909110061824325, + "vlm_logprob": -0.0071909110061824325, + "vlm_yes_no_margin": 5.15, + "vlm_token_credit": -0.0071909110061824325, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a muscular, middle-aged woodcutter with short dark brown hair and a thick beard?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the woodcutter wearing a frayed red and black flannel shirt with sleeves rolled up to the elbows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the woodcutter performing a powerful downward swing with an axe while leaning back with legs planted?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene set in a dense, primeval forest with towering ancient trees and dappled sunlight creating a Tyndall effect?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot captured in slow-motion showing sawdust exploding in crisp arcs against golden beams of light?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.007897106930613518 + ], + "opposite_logprob_tokens": [ + -5.132896900177002 + ], + "logprob_reward": -0.007897106930613518, + "yes_no_margin_reward": 5.125, + "token_credit_reward": -0.007897106930613518, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.29473093152046204, + 0.06784170866012573, + 0.046165626496076584, + 0.048463374376297, + 0.1807226985692978, + 0.04453780874609947, + 0.08728018403053284, + 0.06136494129896164, + 0.05277692899107933, + 0.11611578613519669 + ], + "credit_map": [ + 0.032024264335632324, + 0.0439806692302227, + 0.0276931244879961, + 0.0607779324054718, + 0.02750956267118454, + 0.0321827232837677, + 0.04066457971930504, + 0.029898062348365784, + 0.0055647604167461395, + 0.008969618938863277, + 0.008534746244549751, + 0.009753457270562649, + 0.006487179547548294, + 0.008991456590592861, + 0.009277956560254097, + 0.010262535884976387, + 0.005103346426039934, + 0.006905954331159592, + 0.006798041518777609, + 0.007747530471533537, + 0.004810252692550421, + 0.006030732300132513, + 0.005017642397433519, + 0.0037521261256188154, + 0.004984365310519934, + 0.006313132122159004, + 0.006893649697303772, + 0.004565747920423746, + 0.00821776408702135, + 0.007793230004608631, + 0.0035384066868573427, + 0.006157075520604849, + 0.00953654944896698, + 0.060487840324640274, + 0.06700284034013748, + 0.008913537487387657, + 0.008289654739201069, + 0.01344447024166584, + 0.005542891565710306, + 0.007504925597459078, + 0.005912716966122389, + 0.00836176984012127, + 0.005573797971010208, + 0.004684679210186005, + 0.005691747646778822, + 0.006810264196246862, + 0.003521599341183901, + 0.003981233574450016, + 0.01714497059583664, + 0.013809441588819027, + 0.009830165654420853, + 0.008802296593785286, + 0.014295301400125027, + 0.007498760707676411, + 0.007569149136543274, + 0.008330103009939194, + 0.0070429458282887936, + 0.011215033940970898, + 0.005674486514180899, + 0.006474622990936041, + 0.005186222027987242, + 0.011319407261908054, + 0.004628327675163746, + 0.009823895990848541, + 0.005950325168669224, + 0.007646157871931791, + 0.00529174879193306, + 0.006903734989464283, + 0.0065224310383200645, + 0.004731405060738325, + 0.00789539236575365, + 0.007835733704268932, + 0.010458760894834995, + 0.011685863137245178, + 0.013785100542008877, + 0.01941683143377304, + 0.01465723104774952, + 0.012583288364112377, + 0.01565026119351387, + 0.017878450453281403 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 9 + ], + "selected_reward": -0.007897106930613518 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a muscular, middle-aged woodcutter with short dark brown hair and a thick beard?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the woodcutter wearing a frayed red and black flannel shirt with sleeves rolled up to the elbows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the woodcutter performing a powerful downward swing with an axe while leaning back with legs planted?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene set in a dense, primeval forest with towering ancient trees and dappled sunlight creating a Tyndall effect?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot captured in slow-motion showing sawdust exploding in crisp arcs against golden beams of light?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004646338522434235 + ], + "opposite_logprob_tokens": [ + -5.379646301269531 + ], + "logprob_reward": -0.004646338522434235, + "yes_no_margin_reward": 5.375, + "token_credit_reward": -0.004646338522434235, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.247123122215271, + 0.07579287886619568, + 0.0733722373843193, + 0.09749004989862442, + 0.17838646471500397, + 0.030180780217051506, + 0.06743457913398743, + 0.06817403435707092, + 0.0827452689409256, + 0.07930051535367966 + ], + "credit_map": [ + 0.028542598709464073, + 0.029766565188765526, + 0.02670232579112053, + 0.047501079738140106, + 0.027332762256264687, + 0.027453279122710228, + 0.03129520267248154, + 0.028529353439807892, + 0.008736084215342999, + 0.012041950598359108, + 0.009893916547298431, + 0.01015770435333252, + 0.008269048295915127, + 0.01253625750541687, + 0.008641008287668228, + 0.005516916047781706, + 0.005211955867707729, + 0.00947954598814249, + 0.01242472417652607, + 0.009992996230721474, + 0.010151886381208897, + 0.01050553098320961, + 0.008783369325101376, + 0.006822236813604832, + 0.007633181754499674, + 0.020220013335347176, + 0.01393438596278429, + 0.009409224614501, + 0.008369145914912224, + 0.02085728570818901, + 0.007616231683641672, + 0.009450593963265419, + 0.013772860169410706, + 0.03446965292096138, + 0.0408833846449852, + 0.011117775924503803, + 0.023914940655231476, + 0.026876557618379593, + 0.013104088604450226, + 0.014247228391468525, + 0.0027609970420598984, + 0.005156766623258591, + 0.0031750695779919624, + 0.003044307231903076, + 0.003434948856011033, + 0.005794378463178873, + 0.0029034330509603024, + 0.0039108810015022755, + 0.009166085161268711, + 0.014215003699064255, + 0.008551624603569508, + 0.009399344213306904, + 0.006996745243668556, + 0.006532739382237196, + 0.005771853029727936, + 0.006801190320402384, + 0.0071053337305784225, + 0.009632869623601437, + 0.006403189152479172, + 0.009854543954133987, + 0.00774396164342761, + 0.01272659283131361, + 0.006271080579608679, + 0.0084364740177989, + 0.006013609003275633, + 0.00925561785697937, + 0.009587143547832966, + 0.009947363287210464, + 0.013165540993213654, + 0.0070862723514437675, + 0.012915395200252533, + 0.014774339273571968, + 0.005874800030142069, + 0.00863784458488226, + 0.009855731390416622, + 0.015157414600253105, + 0.010535525158047676, + 0.008206143975257874, + 0.008839461952447891, + 0.012193603441119194 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 3 + ], + "selected_reward": -0.004646338522434235 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a muscular, middle-aged woodcutter with short dark brown hair and a thick beard?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the woodcutter wearing a frayed red and black flannel shirt with sleeves rolled up to the elbows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the woodcutter performing a powerful downward swing with an axe while leaning back with legs planted?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene set in a dense, primeval forest with towering ancient trees and dappled sunlight creating a Tyndall effect?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot captured in slow-motion showing sawdust exploding in crisp arcs against golden beams of light?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.014193654991686344 + ], + "opposite_logprob_tokens": [ + -4.264193534851074 + ], + "logprob_reward": -0.014193654991686344, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.014193654991686344, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1419852077960968, + 0.07676786929368973, + 0.031674426048994064, + 0.03564630448818207, + 0.1504243165254593, + 0.031488969922065735, + 0.12841244041919708, + 0.20004025101661682, + 0.13999216258525848, + 0.0635681226849556 + ], + "credit_map": [ + 0.011643094010651112, + 0.017549555748701096, + 0.016918936744332314, + 0.03126204013824463, + 0.013695934787392616, + 0.01699727773666382, + 0.020013704895973206, + 0.013904647901654243, + 0.006219316739588976, + 0.012105586007237434, + 0.007489708252251148, + 0.010066420771181583, + 0.012001597322523594, + 0.011144513264298439, + 0.008955856785178185, + 0.00878486130386591, + 0.002556830644607544, + 0.004133967682719231, + 0.0051513235084712505, + 0.00347072328440845, + 0.00476142019033432, + 0.003730060765519738, + 0.00398884154856205, + 0.0038812588900327682, + 0.0033460184931755066, + 0.006585108116269112, + 0.00529458886012435, + 0.003883583704009652, + 0.0041213915683329105, + 0.006418109871447086, + 0.002683341968804598, + 0.0033141591120511293, + 0.01273644994944334, + 0.02481416054069996, + 0.022011741995811462, + 0.010389572940766811, + 0.022364798933267593, + 0.025400597602128983, + 0.015173213556408882, + 0.01753377355635166, + 0.003666312899440527, + 0.004301224835216999, + 0.00311618042178452, + 0.002779758535325527, + 0.004474833607673645, + 0.005491066724061966, + 0.004085752181708813, + 0.0035738369915634394, + 0.015014746226370335, + 0.02728910930454731, + 0.016699085012078285, + 0.01218001451343298, + 0.014672498218715191, + 0.012739953584969044, + 0.013841343112289906, + 0.01597568765282631, + 0.013653283007442951, + 0.03146781027317047, + 0.019676489755511284, + 0.01855672337114811, + 0.020029820501804352, + 0.047662924975156784, + 0.014598843641579151, + 0.03439434990286827, + 0.00867492239922285, + 0.011763905175030231, + 0.016128789633512497, + 0.01882198266685009, + 0.026960017159581184, + 0.012051462195813656, + 0.020679950714111328, + 0.02491111494600773, + 0.007466242182999849, + 0.0064420332200825214, + 0.007576348725706339, + 0.009643971920013428, + 0.010321589186787605, + 0.006703643128275871, + 0.006865844130516052, + 0.008548443205654621 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 4, + 0 + ], + "selected_reward": -0.014193654991686344 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a muscular, middle-aged woodcutter with short dark brown hair and a thick beard?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the woodcutter wearing a frayed red and black flannel shirt with sleeves rolled up to the elbows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the woodcutter performing a powerful downward swing with an axe while leaning back with legs planted?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene set in a dense, primeval forest with towering ancient trees and dappled sunlight creating a Tyndall effect?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot captured in slow-motion showing sawdust exploding in crisp arcs against golden beams of light?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.006730386987328529 + ], + "opposite_logprob_tokens": [ + -5.006730556488037 + ], + "logprob_reward": -0.006730386987328529, + "yes_no_margin_reward": 5.0, + "token_credit_reward": -0.006730386987328529, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1977156102657318, + 0.06616319715976715, + 0.03907700255513191, + 0.10556092113256454, + 0.15798969566822052, + 0.05497002601623535, + 0.18338137865066528, + 0.07307536154985428, + 0.049392666667699814, + 0.07267415523529053 + ], + "credit_map": [ + 0.01959196850657463, + 0.025103392079472542, + 0.026663433760404587, + 0.040289200842380524, + 0.017854992300271988, + 0.023859195411205292, + 0.01809590496122837, + 0.026257505640387535, + 0.007667375262826681, + 0.012228570878505707, + 0.008210917934775352, + 0.010297042317688465, + 0.006779538933187723, + 0.009527887217700481, + 0.006992727052420378, + 0.004459143150597811, + 0.004106330219656229, + 0.006573663558810949, + 0.008128668181598186, + 0.006491007748991251, + 0.0022884057834744453, + 0.0053064399398863316, + 0.0027192162815481424, + 0.0034632692113518715, + 0.01070167776197195, + 0.024182967841625214, + 0.015151066705584526, + 0.014459402300417423, + 0.008425774052739143, + 0.014171333983540535, + 0.0062728151679039, + 0.012195880524814129, + 0.010952330194413662, + 0.0501244030892849, + 0.05317750200629234, + 0.009148946963250637, + 0.009708050638437271, + 0.012284868396818638, + 0.005381297320127487, + 0.0072122919373214245, + 0.006593508645892143, + 0.01089451089501381, + 0.011737440712749958, + 0.007466795388609171, + 0.004062803927809, + 0.004677731078118086, + 0.004063929431140423, + 0.005473311059176922, + 0.037330131977796555, + 0.04231226071715355, + 0.023297825828194618, + 0.023327521979808807, + 0.016983039677143097, + 0.013727536424994469, + 0.012081271037459373, + 0.014321796596050262, + 0.01022592093795538, + 0.014991428703069687, + 0.0072515918873250484, + 0.015635306015610695, + 0.005622466094791889, + 0.00628424808382988, + 0.005000339820981026, + 0.008064058609306812, + 0.006929701194167137, + 0.005585186183452606, + 0.006946933455765247, + 0.007693941704928875, + 0.006973819807171822, + 0.004936167970299721, + 0.0047192685306072235, + 0.005607648752629757, + 0.0060089100152254105, + 0.01001643855124712, + 0.009456242434680462, + 0.01489683985710144, + 0.00722736120223999, + 0.00737919844686985, + 0.0077238529920578, + 0.009965308010578156 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 6, + 4 + ], + "selected_reward": -0.006730386987328529 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a muscular, middle-aged woodcutter with short dark brown hair and a thick beard?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the woodcutter wearing a frayed red and black flannel shirt with sleeves rolled up to the elbows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the woodcutter performing a powerful downward swing with an axe while leaning back with legs planted?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene set in a dense, primeval forest with towering ancient trees and dappled sunlight creating a Tyndall effect?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot captured in slow-motion showing sawdust exploding in crisp arcs against golden beams of light?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.002487067598849535 + ], + "opposite_logprob_tokens": [ + -6.0024871826171875 + ], + "logprob_reward": -0.002487067598849535, + "yes_no_margin_reward": 6.0, + "token_credit_reward": -0.002487067598849535, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.33885109424591064, + 0.1314900666475296, + 0.045309703797101974, + 0.048313796520233154, + 0.09093205630779266, + 0.024801528081297874, + 0.061646729707717896, + 0.09560012817382812, + 0.07788511365652084, + 0.08516969531774521 + ], + "credit_map": [ + 0.028123604133725166, + 0.04650973528623581, + 0.04669218137860298, + 0.08324054628610611, + 0.030216528102755547, + 0.04090088978409767, + 0.0369362011551857, + 0.02623146027326584, + 0.010562588460743427, + 0.020938469097018242, + 0.013929908163845539, + 0.02406962588429451, + 0.013978879898786545, + 0.021764792501926422, + 0.01473374292254448, + 0.011512060649693012, + 0.006731598172336817, + 0.009006715379655361, + 0.005502385552972555, + 0.003628895618021488, + 0.005154075101017952, + 0.006783365737646818, + 0.003799416357651353, + 0.0047032590955495834, + 0.0038264405447989702, + 0.008763615041971207, + 0.007865114137530327, + 0.005428538657724857, + 0.0038847976829856634, + 0.008655405603349209, + 0.004192353691905737, + 0.005697533488273621, + 0.005700300447642803, + 0.022021465003490448, + 0.022691160440444946, + 0.0057131205685436726, + 0.00971208792179823, + 0.01210295595228672, + 0.005329204723238945, + 0.007661767303943634, + 0.001959302229806781, + 0.003136202460154891, + 0.002738501178100705, + 0.0021562394686043262, + 0.003319657640531659, + 0.005165028851479292, + 0.0028950104024261236, + 0.003431587480008602, + 0.009000674821436405, + 0.008649230003356934, + 0.007090133614838123, + 0.005263217259198427, + 0.010313608683645725, + 0.005840879864990711, + 0.0077984025701880455, + 0.007690585218369961, + 0.007489482406526804, + 0.015610748901963234, + 0.0105445496737957, + 0.010846676304936409, + 0.009921996854245663, + 0.01913224346935749, + 0.007479602005332708, + 0.014574832282960415, + 0.0061486270278692245, + 0.009041638113558292, + 0.007142919115722179, + 0.010645076632499695, + 0.01210815366357565, + 0.006865495350211859, + 0.012522137723863125, + 0.013411075808107853, + 0.005379891023039818, + 0.010129786096513271, + 0.010048513300716877, + 0.01599022187292576, + 0.00864732638001442, + 0.011798182502388954, + 0.007620501331984997, + 0.015555277466773987 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 7 + ], + "selected_reward": -0.002487067598849535 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "medium, slow-motion shot, shot from a slightly upward angle, focuses on a muscular, middle-aged woodcutter, in the midst of a powerful downward swing with his axe. He has short, unkempt dark brown hair and a thick beard. Beads of sweat shimmer on his forehead in the sunlight. His facial expression is focused and powerful, his lips tightly pursed. He wears a frayed red and black flannel shirt, sleeves rolled up to the elbows, revealing forearms bulging with veins from exertion. He wears dark blue overalls stained with dirt. He grips a long-handled axe with both hands, the blade gleaming with a cool metallic sheen in the sunlight. He leans back, his legs firmly planted in the soft earth strewn with fallen leaves and broken branches. He is deep in a dense, primeval forest, surrounded by towering, ancient trees. The afternoon sun penetrates the lush canopy, creating dappled shafts of light (a Tyndall effect) that illuminate the dust particles floating in the air. As the axe blade struck the massive stump before him with thunderous precision, captured at an extremely high frame rate, countless sawdust of varying sizes exploded in all directions in elegant, crisp arcs, as if enchanted. The swirling trajectory and texture of each splinter were clearly visible against the backdrop of golden beams. The dominant tones of the image were deep forest green and warm earthy brown, accented by golden sunlight and the woodcutter's vibrant shirt, creating a stark contrast between a sense of primal power, the beauty of focused labor, and the tranquility of the deep forest.", + "seed": 7962, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000050/group_0/index_8040_seed_7962_dp_0_rank_0_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 0, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.007340689701959491, + "vlm_logprob": -0.007340689701959491, + "vlm_yes_no_margin": 5.125, + "vlm_token_credit": -0.007340689701959491, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a muscular, middle-aged woodcutter with short dark brown hair and a thick beard?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the woodcutter wearing a frayed red and black flannel shirt with sleeves rolled up to the elbows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the woodcutter performing a powerful downward swing with an axe while leaning back with legs planted?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene set in a dense, primeval forest with towering ancient trees and dappled sunlight creating a Tyndall effect?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot captured in slow-motion showing sawdust exploding in crisp arcs against golden beams of light?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01778746210038662 + ], + "opposite_logprob_tokens": [ + -4.142787456512451 + ], + "logprob_reward": -0.01778746210038662, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.01778746210038662, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1404762864112854, + 0.10700147598981857, + 0.09312888979911804, + 0.07443010061979294, + 0.04200800508260727, + 0.03324802592396736, + 0.13231827318668365, + 0.23612257838249207, + 0.0831451416015625, + 0.05812118202447891 + ], + "credit_map": [ + 0.008749963715672493, + 0.017560485750436783, + 0.01789521612226963, + 0.009044475853443146, + 0.018915893509984016, + 0.020674994215369225, + 0.029179083183407784, + 0.018456172198057175, + 0.007729986682534218, + 0.007471680641174316, + 0.01568528637290001, + 0.007514758501201868, + 0.01648983173072338, + 0.01795901171863079, + 0.014353587292134762, + 0.019797345623373985, + 0.007688620127737522, + 0.012687715701758862, + 0.013202204369008541, + 0.010829620063304901, + 0.008217941038310528, + 0.012058013118803501, + 0.01087825559079647, + 0.017566515132784843, + 0.005772903561592102, + 0.007802304811775684, + 0.012003117240965366, + 0.01250129658728838, + 0.007611035369336605, + 0.011227442882955074, + 0.006528851576149464, + 0.010983145795762539, + 0.0037283594720065594, + 0.005735255777835846, + 0.004538507666438818, + 0.0036884238943457603, + 0.003816380864009261, + 0.008694165386259556, + 0.005084853153675795, + 0.00672206049785018, + 0.002947400091215968, + 0.004033807199448347, + 0.006006143055856228, + 0.0030541543383151293, + 0.004329463001340628, + 0.004051343537867069, + 0.004531113430857658, + 0.0042945994064211845, + 0.010369050316512585, + 0.013256175443530083, + 0.020152252167463303, + 0.022786831483244896, + 0.012802505865693092, + 0.01842401921749115, + 0.0192206259816885, + 0.015306806191802025, + 0.015673574060201645, + 0.019512929022312164, + 0.021751612424850464, + 0.02631932683289051, + 0.041283756494522095, + 0.028760526329278946, + 0.04986639320850372, + 0.03295447304844856, + 0.004939818289130926, + 0.0059297578409314156, + 0.008430794812738895, + 0.007033859845250845, + 0.009956789202988148, + 0.011886567808687687, + 0.021643219515681267, + 0.013324333354830742, + 0.004221597220748663, + 0.005234114360064268, + 0.006666283588856459, + 0.0059896656312048435, + 0.0076957340352237225, + 0.0069885775446891785, + 0.013548183254897594, + 0.007777026854455471 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 0, + 6 + ], + "selected_reward": -0.01778746210038662 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a muscular, middle-aged woodcutter with short dark brown hair and a thick beard?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the woodcutter wearing a frayed red and black flannel shirt with sleeves rolled up to the elbows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the woodcutter performing a powerful downward swing with an axe while leaning back with legs planted?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene set in a dense, primeval forest with towering ancient trees and dappled sunlight creating a Tyndall effect?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot captured in slow-motion showing sawdust exploding in crisp arcs against golden beams of light?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005260791163891554 + ], + "opposite_logprob_tokens": [ + -5.255260944366455 + ], + "logprob_reward": -0.005260791163891554, + "yes_no_margin_reward": 5.25, + "token_credit_reward": -0.005260791163891554, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12325163185596466, + 0.15594206750392914, + 0.15402410924434662, + 0.06038454547524452, + 0.09003262221813202, + 0.034591324627399445, + 0.20450975000858307, + 0.06798683106899261, + 0.06046675145626068, + 0.048810381442308426 + ], + "credit_map": [ + 0.007811501156538725, + 0.01782589592039585, + 0.01357419602572918, + 0.006285795941948891, + 0.014794380404055119, + 0.02097763679921627, + 0.012625877745449543, + 0.029356349259614944, + 0.007355635054409504, + 0.009074621833860874, + 0.009917150251567364, + 0.005957398563623428, + 0.022613851353526115, + 0.04584794491529465, + 0.012508207000792027, + 0.04266727343201637, + 0.008805479854345322, + 0.010547557845711708, + 0.017073754221200943, + 0.01445473451167345, + 0.01920514740049839, + 0.03356683626770973, + 0.024303289130330086, + 0.026067323982715607, + 0.005434578284621239, + 0.007403240539133549, + 0.008290444500744343, + 0.005666246637701988, + 0.010259621776640415, + 0.008536950685083866, + 0.007657463662326336, + 0.007136005442589521, + 0.007508012466132641, + 0.013488673605024815, + 0.008597839623689651, + 0.007525348104536533, + 0.013881424441933632, + 0.01347704604268074, + 0.013892007991671562, + 0.011662280187010765, + 0.0024368532467633486, + 0.0040380097925662994, + 0.006107080262154341, + 0.0033396307844668627, + 0.003931443206965923, + 0.004270473495125771, + 0.00491004204377532, + 0.005557794589549303, + 0.01928119733929634, + 0.019195953384041786, + 0.03219420462846756, + 0.039615485817193985, + 0.01994726061820984, + 0.02744118869304657, + 0.024809682741761208, + 0.02202477678656578, + 0.007272293791174889, + 0.00693977577611804, + 0.007139664143323898, + 0.007144350092858076, + 0.011161481030285358, + 0.00948096439242363, + 0.009756750427186489, + 0.009091557003557682, + 0.0041204714216291904, + 0.004216608591377735, + 0.007524037268012762, + 0.006531600374728441, + 0.005865863524377346, + 0.007244046311825514, + 0.015537912026047707, + 0.009426213800907135, + 0.004732557572424412, + 0.006571378093212843, + 0.005141361150890589, + 0.005406447686254978, + 0.007040128111839294, + 0.006180427502840757, + 0.007389463018625975, + 0.006348624359816313 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 1, + 2 + ], + "selected_reward": -0.005260791163891554 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a muscular, middle-aged woodcutter with short dark brown hair and a thick beard?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the woodcutter wearing a frayed red and black flannel shirt with sleeves rolled up to the elbows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the woodcutter performing a powerful downward swing with an axe while leaning back with legs planted?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene set in a dense, primeval forest with towering ancient trees and dappled sunlight creating a Tyndall effect?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot captured in slow-motion showing sawdust exploding in crisp arcs against golden beams of light?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.00594940735027194 + ], + "opposite_logprob_tokens": [ + -5.1309494972229 + ], + "logprob_reward": -0.00594940735027194, + "yes_no_margin_reward": 5.125, + "token_credit_reward": -0.00594940735027194, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14446961879730225, + 0.11293117702007294, + 0.20964032411575317, + 0.08615493029356003, + 0.06671883165836334, + 0.04619869589805603, + 0.10153666883707047, + 0.12534983456134796, + 0.055296026170253754, + 0.05170391499996185 + ], + "credit_map": [ + 0.006945795379579067, + 0.012791309505701065, + 0.02014843188226223, + 0.008668357506394386, + 0.017241256311535835, + 0.023097779601812363, + 0.017566189169883728, + 0.03801051154732704, + 0.009015010669827461, + 0.007684545125812292, + 0.012570828199386597, + 0.00862005166709423, + 0.01838802359998226, + 0.023256752640008926, + 0.014127462171018124, + 0.019268494099378586, + 0.011138410307466984, + 0.025279739871621132, + 0.02263391576707363, + 0.024776963517069817, + 0.01628568395972252, + 0.04751494154334068, + 0.022891083732247353, + 0.03911958262324333, + 0.009291963651776314, + 0.008414803072810173, + 0.010935665108263493, + 0.014212943613529205, + 0.007073909975588322, + 0.015613613650202751, + 0.006213054992258549, + 0.014398978091776371, + 0.005086350720375776, + 0.006077764090150595, + 0.006594551261514425, + 0.0044206297025084496, + 0.004272340331226587, + 0.018806330859661102, + 0.004453971516340971, + 0.017006894573569298, + 0.009013645350933075, + 0.004970118869096041, + 0.006879185326397419, + 0.0033548276405781507, + 0.005120835732668638, + 0.005373803898692131, + 0.005754027049988508, + 0.005732246674597263, + 0.006065994966775179, + 0.007271086797118187, + 0.012538738548755646, + 0.010920761153101921, + 0.01991097256541252, + 0.012987377122044563, + 0.01840994507074356, + 0.01343179028481245, + 0.00727192685008049, + 0.010463301092386246, + 0.011948623694479465, + 0.015607655048370361, + 0.02137586660683155, + 0.01452796533703804, + 0.02683132514357567, + 0.017323166131973267, + 0.006440259981900454, + 0.005835087038576603, + 0.0078405337408185, + 0.0055366139858961105, + 0.00788255874067545, + 0.006636760663241148, + 0.008251143619418144, + 0.006873068865388632, + 0.0042691221460700035, + 0.006467156112194061, + 0.006437224801629782, + 0.009279453195631504, + 0.003204087261110544, + 0.00873657502233982, + 0.005802068393677473, + 0.007508226670324802 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 0, + 7 + ], + "selected_reward": -0.00594940735027194 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a muscular, middle-aged woodcutter with short dark brown hair and a thick beard?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the woodcutter wearing a frayed red and black flannel shirt with sleeves rolled up to the elbows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the woodcutter performing a powerful downward swing with an axe while leaning back with legs planted?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene set in a dense, primeval forest with towering ancient trees and dappled sunlight creating a Tyndall effect?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot captured in slow-motion showing sawdust exploding in crisp arcs against golden beams of light?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004091701935976744 + ], + "opposite_logprob_tokens": [ + -5.504091739654541 + ], + "logprob_reward": -0.004091701935976744, + "yes_no_margin_reward": 5.5, + "token_credit_reward": -0.004091701935976744, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15110141038894653, + 0.1005077213048935, + 0.06917732954025269, + 0.0626370906829834, + 0.041521355509757996, + 0.04856251925230026, + 0.2027362585067749, + 0.21935968101024628, + 0.05801203101873398, + 0.04638461768627167 + ], + "credit_map": [ + 0.00824157427996397, + 0.00979524850845337, + 0.016744939610362053, + 0.009731709957122803, + 0.01684054546058178, + 0.03279561921954155, + 0.014569914899766445, + 0.04238184913992882, + 0.00780075928196311, + 0.009388910606503487, + 0.010551664978265762, + 0.009457724168896675, + 0.010876970365643501, + 0.021462948992848396, + 0.010012336075305939, + 0.02095640078186989, + 0.008785880170762539, + 0.00791503582149744, + 0.008071296848356724, + 0.006921024993062019, + 0.009226877242326736, + 0.00963064655661583, + 0.009961643256247044, + 0.008664922788739204, + 0.005971885751932859, + 0.00834653526544571, + 0.008614304475486279, + 0.005379294510930777, + 0.0071997796185314655, + 0.011423139832913876, + 0.0054391538724303246, + 0.01026298850774765, + 0.003943894058465958, + 0.002852515783160925, + 0.004648707341402769, + 0.0037290446925908327, + 0.0034923776984214783, + 0.010739225894212723, + 0.003527451539412141, + 0.008588137105107307, + 0.005513590294867754, + 0.004725652746856213, + 0.008889902383089066, + 0.005235446151345968, + 0.004605392459779978, + 0.006666019093245268, + 0.006272834725677967, + 0.006653679069131613, + 0.018459230661392212, + 0.02199219912290573, + 0.031212346628308296, + 0.03804939240217209, + 0.020521484315395355, + 0.026219762861728668, + 0.025563837960362434, + 0.020717989653348923, + 0.031171686947345734, + 0.04427658021450043, + 0.03303701803088188, + 0.025365164503455162, + 0.013716713525354862, + 0.029001226648688316, + 0.01819046586751938, + 0.024600809440016747, + 0.004619016777724028, + 0.009824239648878574, + 0.006416617427021265, + 0.007249805610626936, + 0.005412640515714884, + 0.008267266675829887, + 0.008961058221757412, + 0.007261380553245544, + 0.0035336739383637905, + 0.005846213549375534, + 0.005923135671764612, + 0.009596042335033417, + 0.0035638087429106236, + 0.006126922555267811, + 0.006029046140611172, + 0.005765773821622133 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 6, + 0 + ], + "selected_reward": -0.004091701935976744 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a muscular, middle-aged woodcutter with short dark brown hair and a thick beard?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the woodcutter wearing a frayed red and black flannel shirt with sleeves rolled up to the elbows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the woodcutter performing a powerful downward swing with an axe while leaning back with legs planted?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene set in a dense, primeval forest with towering ancient trees and dappled sunlight creating a Tyndall effect?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot captured in slow-motion showing sawdust exploding in crisp arcs against golden beams of light?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0036140859592705965 + ], + "opposite_logprob_tokens": [ + -5.6286139488220215 + ], + "logprob_reward": -0.0036140859592705965, + "yes_no_margin_reward": 5.625, + "token_credit_reward": -0.0036140859592705965, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09031149744987488, + 0.15195569396018982, + 0.3102947473526001, + 0.05443406477570534, + 0.1139325350522995, + 0.0591513067483902, + 0.06396801769733429, + 0.041512202471494675, + 0.05377832055091858, + 0.06066161021590233 + ], + "credit_map": [ + 0.0077726044692099094, + 0.008437879383563995, + 0.01089487411081791, + 0.007000298239290714, + 0.01344249676913023, + 0.012078617699444294, + 0.016342539340257645, + 0.01434218417853117, + 0.008493146859109402, + 0.00952232163399458, + 0.01124174427241087, + 0.008244004100561142, + 0.019682712852954865, + 0.04122502729296684, + 0.01639518328011036, + 0.03715156018733978, + 0.02354174107313156, + 0.04692094773054123, + 0.040579307824373245, + 0.04524221271276474, + 0.03986474871635437, + 0.037372976541519165, + 0.040885284543037415, + 0.035887520760297775, + 0.004379300866276026, + 0.006539518479257822, + 0.007195084821432829, + 0.004853877704590559, + 0.004621550440788269, + 0.009164081886410713, + 0.005178595893085003, + 0.012502051889896393, + 0.010693552903831005, + 0.008037582971155643, + 0.01330603752285242, + 0.010249670594930649, + 0.0058319843374192715, + 0.03094971925020218, + 0.007537908852100372, + 0.027326088398694992, + 0.00612822687253356, + 0.005346005316823721, + 0.008537368848919868, + 0.005256150849163532, + 0.006917188409715891, + 0.010604741983115673, + 0.006335808429867029, + 0.010025816038250923, + 0.005840417463332415, + 0.007551449816673994, + 0.010395217686891556, + 0.010587368160486221, + 0.005880007520318031, + 0.008124686777591705, + 0.009083941578865051, + 0.006504926830530167, + 0.0042951698414981365, + 0.0041800253093242645, + 0.004414539318531752, + 0.004440824966877699, + 0.006322804838418961, + 0.0052716382779181, + 0.006145148072391748, + 0.006442050449550152, + 0.006281739100813866, + 0.005835295654833317, + 0.009282935410737991, + 0.006929304450750351, + 0.00569912837818265, + 0.006628783885389566, + 0.006672299467027187, + 0.006448832340538502, + 0.005173655226826668, + 0.008098375983536243, + 0.009290033020079136, + 0.011087087914347649, + 0.003951139748096466, + 0.008565240539610386, + 0.006251174956560135, + 0.008244904689490795 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 1, + 4 + ], + "selected_reward": -0.0036140859592705965 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "medium, slow-motion shot, shot from a slightly upward angle, focuses on a muscular, middle-aged woodcutter, in the midst of a powerful downward swing with his axe. He has short, unkempt dark brown hair and a thick beard. Beads of sweat shimmer on his forehead in the sunlight. His facial expression is focused and powerful, his lips tightly pursed. He wears a frayed red and black flannel shirt, sleeves rolled up to the elbows, revealing forearms bulging with veins from exertion. He wears dark blue overalls stained with dirt. He grips a long-handled axe with both hands, the blade gleaming with a cool metallic sheen in the sunlight. He leans back, his legs firmly planted in the soft earth strewn with fallen leaves and broken branches. He is deep in a dense, primeval forest, surrounded by towering, ancient trees. The afternoon sun penetrates the lush canopy, creating dappled shafts of light (a Tyndall effect) that illuminate the dust particles floating in the air. As the axe blade struck the massive stump before him with thunderous precision, captured at an extremely high frame rate, countless sawdust of varying sizes exploded in all directions in elegant, crisp arcs, as if enchanted. The swirling trajectory and texture of each splinter were clearly visible against the backdrop of golden beams. The dominant tones of the image were deep forest green and warm earthy brown, accented by golden sunlight and the woodcutter's vibrant shirt, creating a stark contrast between a sense of primal power, the beauty of focused labor, and the tranquility of the deep forest.", + "seed": 7963, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000050/group_0/index_8040_seed_7963_dp_1_rank_2_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 0, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.010982627142220735, + "vlm_logprob": -0.010982627142220735, + "vlm_yes_no_margin": 4.7, + "vlm_token_credit": -0.010982627142220735, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a muscular, middle-aged woodcutter with short dark brown hair and a thick beard?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the woodcutter wearing a frayed red and black flannel shirt with sleeves rolled up to the elbows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the woodcutter performing a powerful downward swing with an axe while leaning back with legs planted?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene set in a dense, primeval forest with towering ancient trees and dappled sunlight creating a Tyndall effect?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot captured in slow-motion showing sawdust exploding in crisp arcs against golden beams of light?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01882718876004219 + ], + "opposite_logprob_tokens": [ + -4.268826961517334 + ], + "logprob_reward": -0.01882718876004219, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.01882718876004219, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12114487588405609, + 0.08795642107725143, + 0.1188596859574318, + 0.04020768776535988, + 0.12840525805950165, + 0.11508291214704514, + 0.11236466467380524, + 0.07981342822313309, + 0.059802424162626266, + 0.13636264204978943 + ], + "credit_map": [ + 0.016971193253993988, + 0.021928811445832253, + 0.01231839694082737, + 0.014064854942262173, + 0.01627388224005699, + 0.012196460738778114, + 0.008940248750150204, + 0.018451029434800148, + 0.010911534540355206, + 0.014283413998782635, + 0.00981693621724844, + 0.009239587932825089, + 0.016667790710926056, + 0.01473467517644167, + 0.0036902197171002626, + 0.00861226487904787, + 0.011853338219225407, + 0.021493593230843544, + 0.017553817480802536, + 0.01322336494922638, + 0.014036955311894417, + 0.015294769778847694, + 0.010699444450438023, + 0.01470439787954092, + 0.004105714615434408, + 0.008949010632932186, + 0.0061014448292553425, + 0.004124814178794622, + 0.0039199641905725, + 0.00548516632989049, + 0.002140246331691742, + 0.005381326656788588, + 0.0126200495287776, + 0.020873894914984703, + 0.012949218042194843, + 0.011348378844559193, + 0.03381544351577759, + 0.01669117622077465, + 0.012346534989774227, + 0.0077605703845620155, + 0.01749739795923233, + 0.014848805963993073, + 0.014338323846459389, + 0.016016483306884766, + 0.014736592769622803, + 0.019483476877212524, + 0.009232032112777233, + 0.008929796516895294, + 0.010613967664539814, + 0.01608061045408249, + 0.012413370423018932, + 0.011879605241119862, + 0.020557643845677376, + 0.017721816897392273, + 0.013981438241899014, + 0.009116213768720627, + 0.010762250050902367, + 0.006643639877438545, + 0.012152284383773804, + 0.009273961186408997, + 0.00835205428302288, + 0.00686127133667469, + 0.013248506002128124, + 0.012519462034106255, + 0.0066263070330023766, + 0.008167307823896408, + 0.010080208070576191, + 0.007896577939391136, + 0.006909689866006374, + 0.008312324993312359, + 0.006543094292283058, + 0.005266916938126087, + 0.023309634998440742, + 0.01465606689453125, + 0.012096002697944641, + 0.01718055084347725, + 0.01673908717930317, + 0.012866691686213017, + 0.023309161886572838, + 0.01620544120669365 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 4, + 0 + ], + "selected_reward": -0.01882718876004219 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a muscular, middle-aged woodcutter with short dark brown hair and a thick beard?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the woodcutter wearing a frayed red and black flannel shirt with sleeves rolled up to the elbows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the woodcutter performing a powerful downward swing with an axe while leaning back with legs planted?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene set in a dense, primeval forest with towering ancient trees and dappled sunlight creating a Tyndall effect?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot captured in slow-motion showing sawdust exploding in crisp arcs against golden beams of light?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.006746609229594469 + ], + "opposite_logprob_tokens": [ + -5.006746768951416 + ], + "logprob_reward": -0.006746609229594469, + "yes_no_margin_reward": 5.0, + "token_credit_reward": -0.006746609229594469, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15170584619045258, + 0.14188824594020844, + 0.14648456871509552, + 0.062114179134368896, + 0.06947936117649078, + 0.08946594595909119, + 0.05556471645832062, + 0.05972687155008316, + 0.08211882412433624, + 0.1414514034986496 + ], + "credit_map": [ + 0.01588817499577999, + 0.035097386687994, + 0.0120917409658432, + 0.019266629591584206, + 0.024071207270026207, + 0.017918100580573082, + 0.009032166562974453, + 0.018340453505516052, + 0.017020082101225853, + 0.03854742273688316, + 0.013024196028709412, + 0.0152264554053545, + 0.016327477991580963, + 0.017348790541291237, + 0.006659748964011669, + 0.01773408241569996, + 0.013093420304358006, + 0.026770800352096558, + 0.01954544708132744, + 0.014039273373782635, + 0.024454571306705475, + 0.01648063212633133, + 0.014368605799973011, + 0.01773183047771454, + 0.00546792708337307, + 0.01872328296303749, + 0.007850732654333115, + 0.007448870223015547, + 0.004487467464059591, + 0.00690573500469327, + 0.003479466773569584, + 0.007750697899609804, + 0.007175060920417309, + 0.008067583665251732, + 0.009551421739161015, + 0.008617997169494629, + 0.010140825062990189, + 0.009522652253508568, + 0.008644002489745617, + 0.007759824860841036, + 0.011034594848752022, + 0.012657059356570244, + 0.010909617878496647, + 0.014557613991200924, + 0.014672868885099888, + 0.012017359957098961, + 0.0077042002230882645, + 0.0059126452542841434, + 0.006381579674780369, + 0.00777784688398242, + 0.006521405186504126, + 0.007895632646977901, + 0.006877311505377293, + 0.007115471176803112, + 0.0057515427470207214, + 0.00724392756819725, + 0.008812260814011097, + 0.008712082169950008, + 0.00847233273088932, + 0.0060823229141533375, + 0.0069970800541341305, + 0.008163215592503548, + 0.007028078660368919, + 0.005459502339363098, + 0.011240523308515549, + 0.013147387653589249, + 0.011116125620901585, + 0.009737889282405376, + 0.010159633122384548, + 0.011932464316487312, + 0.007212658412754536, + 0.007572151720523834, + 0.013595305383205414, + 0.01898363046348095, + 0.020329222083091736, + 0.013398378156125546, + 0.015084152109920979, + 0.029009925201535225, + 0.012869050726294518, + 0.018181752413511276 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 1 + ], + "selected_reward": -0.006746609229594469 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a muscular, middle-aged woodcutter with short dark brown hair and a thick beard?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the woodcutter wearing a frayed red and black flannel shirt with sleeves rolled up to the elbows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the woodcutter performing a powerful downward swing with an axe while leaning back with legs planted?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene set in a dense, primeval forest with towering ancient trees and dappled sunlight creating a Tyndall effect?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot captured in slow-motion showing sawdust exploding in crisp arcs against golden beams of light?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.016069892793893814 + ], + "opposite_logprob_tokens": [ + -4.1410698890686035 + ], + "logprob_reward": -0.016069892793893814, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.016069892793893814, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07941114902496338, + 0.05941363051533699, + 0.15821844339370728, + 0.03955085203051567, + 0.1948561668395996, + 0.06792990863323212, + 0.0547899566590786, + 0.049474820494651794, + 0.21599625051021576, + 0.0803588405251503 + ], + "credit_map": [ + 0.00861731730401516, + 0.012521483935415745, + 0.006758420262485743, + 0.007552335038781166, + 0.012831020168960094, + 0.009445182047784328, + 0.009743183851242065, + 0.01194220781326294, + 0.00674921041354537, + 0.008991901762783527, + 0.004003572277724743, + 0.003725445130839944, + 0.016309581696987152, + 0.009315668605268002, + 0.003981332760304213, + 0.00633692042902112, + 0.015037836506962776, + 0.03340182453393936, + 0.019303331151604652, + 0.021182671189308167, + 0.021873975172638893, + 0.01825014129281044, + 0.014850572682917118, + 0.014318090863525867, + 0.00427607586607337, + 0.010562443174421787, + 0.004641020204871893, + 0.004357142373919487, + 0.0031085091177374125, + 0.004103187471628189, + 0.003057207213714719, + 0.005445265211164951, + 0.016493376344442368, + 0.023677267134189606, + 0.019952338188886642, + 0.017456768080592155, + 0.04708932340145111, + 0.027220774441957474, + 0.02723720856010914, + 0.015729116275906563, + 0.006768824066966772, + 0.007814333774149418, + 0.006793153937906027, + 0.011610232293605804, + 0.012119783088564873, + 0.011887806467711926, + 0.00540702510625124, + 0.005528744775801897, + 0.004813975654542446, + 0.00746090617030859, + 0.008926046080887318, + 0.008387740701436996, + 0.007656823378056288, + 0.007972072809934616, + 0.005565222352743149, + 0.004007167648524046, + 0.007288674358278513, + 0.005178410094231367, + 0.006035518832504749, + 0.004841703921556473, + 0.006049894727766514, + 0.005571491550654173, + 0.009117554873228073, + 0.005391569342464209, + 0.025685805827379227, + 0.032055918127298355, + 0.033736277371644974, + 0.026589615270495415, + 0.02125038206577301, + 0.04244384542107582, + 0.01803065650165081, + 0.016203738749027252, + 0.0066999029368162155, + 0.011435646563768387, + 0.00932372733950615, + 0.010221333242952824, + 0.010548229329288006, + 0.010637455619871616, + 0.011818889528512955, + 0.009673653170466423 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 4, + 2 + ], + "selected_reward": -0.016069892793893814 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a muscular, middle-aged woodcutter with short dark brown hair and a thick beard?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the woodcutter wearing a frayed red and black flannel shirt with sleeves rolled up to the elbows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the woodcutter performing a powerful downward swing with an axe while leaning back with legs planted?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene set in a dense, primeval forest with towering ancient trees and dappled sunlight creating a Tyndall effect?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot captured in slow-motion showing sawdust exploding in crisp arcs against golden beams of light?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.008633785881102085 + ], + "opposite_logprob_tokens": [ + -4.758633613586426 + ], + "logprob_reward": -0.008633785881102085, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.008633785881102085, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09275388717651367, + 0.043402548879384995, + 0.07509098947048187, + 0.09304845333099365, + 0.11706354469060898, + 0.21231171488761902, + 0.13849957287311554, + 0.05748520791530609, + 0.08151140809059143, + 0.08883263170719147 + ], + "credit_map": [ + 0.012620040215551853, + 0.014616355299949646, + 0.00928373634815216, + 0.009241145104169846, + 0.012632418423891068, + 0.008671622723340988, + 0.010126101784408092, + 0.015562470071017742, + 0.0053215245716273785, + 0.006514303851872683, + 0.004562560003250837, + 0.003791842143982649, + 0.00881398469209671, + 0.005749777890741825, + 0.003394906409084797, + 0.00525365024805069, + 0.010325784794986248, + 0.012321566231548786, + 0.00950659066438675, + 0.009423411451280117, + 0.010999527759850025, + 0.006865853443741798, + 0.008175675757229328, + 0.007472584024071693, + 0.0116279236972332, + 0.019343307241797447, + 0.009975442662835121, + 0.012948337011039257, + 0.012153367511928082, + 0.009267076849937439, + 0.006803234573453665, + 0.010929767973721027, + 0.010285759344696999, + 0.017855359241366386, + 0.012518296018242836, + 0.011989347636699677, + 0.03166214004158974, + 0.014316282235085964, + 0.010883794166147709, + 0.007552571594715118, + 0.018660537898540497, + 0.026907477527856827, + 0.02233748510479927, + 0.023151658475399017, + 0.03116273134946823, + 0.041939713060855865, + 0.027656445279717445, + 0.020495666190981865, + 0.014804606325924397, + 0.01666932925581932, + 0.027681611478328705, + 0.01523605640977621, + 0.012850362807512283, + 0.02815018780529499, + 0.009236479178071022, + 0.013870933093130589, + 0.009617025963962078, + 0.006728345062583685, + 0.00929305050522089, + 0.006895433180034161, + 0.006473342422395945, + 0.007835394702851772, + 0.006551518104970455, + 0.004091099835932255, + 0.01245030015707016, + 0.01199648529291153, + 0.013144164346158504, + 0.007759030908346176, + 0.009290508925914764, + 0.012651818804442883, + 0.0075193848460912704, + 0.006699720397591591, + 0.010121366009116173, + 0.009297813288867474, + 0.00887016486376524, + 0.0116460295394063, + 0.009592628106474876, + 0.013394424691796303, + 0.017549922689795494, + 0.008360277861356735 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 6, + 4 + ], + "selected_reward": -0.008633785881102085 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a muscular, middle-aged woodcutter with short dark brown hair and a thick beard?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the woodcutter wearing a frayed red and black flannel shirt with sleeves rolled up to the elbows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the woodcutter performing a powerful downward swing with an axe while leaning back with legs planted?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene set in a dense, primeval forest with towering ancient trees and dappled sunlight creating a Tyndall effect?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot captured in slow-motion showing sawdust exploding in crisp arcs against golden beams of light?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004635659046471119 + ], + "opposite_logprob_tokens": [ + -5.379635810852051 + ], + "logprob_reward": -0.004635659046471119, + "yes_no_margin_reward": 5.375, + "token_credit_reward": -0.004635659046471119, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1928546279668808, + 0.08098641782999039, + 0.1379348784685135, + 0.060630518943071365, + 0.10119889676570892, + 0.08288495987653732, + 0.043793801218271255, + 0.0954509973526001, + 0.11714662611484528, + 0.08711826801300049 + ], + "credit_map": [ + 0.025173287838697433, + 0.026984011754393578, + 0.015944210812449455, + 0.016013771295547485, + 0.05174896866083145, + 0.01842731051146984, + 0.014719662256538868, + 0.02384338155388832, + 0.010514471679925919, + 0.0128560122102499, + 0.008120091632008553, + 0.006086756009608507, + 0.011211322620511055, + 0.019401755183935165, + 0.003976037260144949, + 0.008819968439638615, + 0.01089890394359827, + 0.018240541219711304, + 0.018524007871747017, + 0.01387122180312872, + 0.023826811462640762, + 0.024976510554552078, + 0.014490768313407898, + 0.01310611329972744, + 0.005720220040529966, + 0.019110992550849915, + 0.006293750833719969, + 0.00706776836887002, + 0.003973855637013912, + 0.006206423044204712, + 0.004585561342537403, + 0.007671944331377745, + 0.007358253002166748, + 0.011682264506816864, + 0.011610775254666805, + 0.010886544361710548, + 0.021916279569268227, + 0.01585111953318119, + 0.013789120130240917, + 0.0081045376136899, + 0.010097470134496689, + 0.012887636199593544, + 0.009563922882080078, + 0.011730507016181946, + 0.009810433723032475, + 0.010715547017753124, + 0.007929840125143528, + 0.010149593465030193, + 0.0032808424439281225, + 0.004681530874222517, + 0.00521159078925848, + 0.006418012548238039, + 0.007564855739474297, + 0.005823496729135513, + 0.005783497355878353, + 0.005029970780014992, + 0.015344381332397461, + 0.012955539859831333, + 0.015855420380830765, + 0.009848837740719318, + 0.007571895606815815, + 0.011588581837713718, + 0.01399149652570486, + 0.008294835686683655, + 0.01476968638598919, + 0.013660622760653496, + 0.027005035430192947, + 0.01118794921785593, + 0.006861263420432806, + 0.024747980758547783, + 0.008112827315926552, + 0.01080125942826271, + 0.011340686120092869, + 0.012724778614938259, + 0.011707819066941738, + 0.009515102952718735, + 0.008513789623975754, + 0.014870727434754372, + 0.0085312370210886, + 0.009914121590554714 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 8 + ], + "selected_reward": -0.004635659046471119 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "medium, slow-motion shot, shot from a slightly upward angle, focuses on a muscular, middle-aged woodcutter, in the midst of a powerful downward swing with his axe. He has short, unkempt dark brown hair and a thick beard. Beads of sweat shimmer on his forehead in the sunlight. His facial expression is focused and powerful, his lips tightly pursed. He wears a frayed red and black flannel shirt, sleeves rolled up to the elbows, revealing forearms bulging with veins from exertion. He wears dark blue overalls stained with dirt. He grips a long-handled axe with both hands, the blade gleaming with a cool metallic sheen in the sunlight. He leans back, his legs firmly planted in the soft earth strewn with fallen leaves and broken branches. He is deep in a dense, primeval forest, surrounded by towering, ancient trees. The afternoon sun penetrates the lush canopy, creating dappled shafts of light (a Tyndall effect) that illuminate the dust particles floating in the air. As the axe blade struck the massive stump before him with thunderous precision, captured at an extremely high frame rate, countless sawdust of varying sizes exploded in all directions in elegant, crisp arcs, as if enchanted. The swirling trajectory and texture of each splinter were clearly visible against the backdrop of golden beams. The dominant tones of the image were deep forest green and warm earthy brown, accented by golden sunlight and the woodcutter's vibrant shirt, creating a stark contrast between a sense of primal power, the beauty of focused labor, and the tranquility of the deep forest.", + "seed": 7964, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000050/group_0/index_8040_seed_7964_dp_1_rank_2_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 0, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.009417069237679243, + "vlm_logprob": -0.009417069237679243, + "vlm_yes_no_margin": 4.875, + "vlm_token_credit": -0.009417069237679243, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a muscular, middle-aged woodcutter with short dark brown hair and a thick beard?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the woodcutter wearing a frayed red and black flannel shirt with sleeves rolled up to the elbows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the woodcutter performing a powerful downward swing with an axe while leaning back with legs planted?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene set in a dense, primeval forest with towering ancient trees and dappled sunlight creating a Tyndall effect?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot captured in slow-motion showing sawdust exploding in crisp arcs against golden beams of light?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.016409561038017273 + ], + "opposite_logprob_tokens": [ + -4.266409397125244 + ], + "logprob_reward": -0.016409561038017273, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.016409561038017273, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.134943425655365, + 0.07420999556779861, + 0.0880037397146225, + 0.0676397904753685, + 0.05843505635857582, + 0.06049592047929764, + 0.0918671041727066, + 0.09601978957653046, + 0.26207855343818665, + 0.06630662828683853 + ], + "credit_map": [ + 0.01261779572814703, + 0.016659477725625038, + 0.02221408113837242, + 0.033328305929899216, + 0.008852098137140274, + 0.01156444288790226, + 0.017485881224274635, + 0.01222133170813322, + 0.006999487988650799, + 0.008027633652091026, + 0.014386001974344254, + 0.018012188374996185, + 0.007006766740232706, + 0.005204058717936277, + 0.004941279534250498, + 0.009632580913603306, + 0.0071355742402374744, + 0.007706755772233009, + 0.017341863363981247, + 0.026461025699973106, + 0.005785138346254826, + 0.004970256704837084, + 0.010991105809807777, + 0.007612013258039951, + 0.005753078497946262, + 0.00704398937523365, + 0.009958191774785519, + 0.015598061494529247, + 0.00826257560402155, + 0.006826816126704216, + 0.008618672378361225, + 0.005578401032835245, + 0.005702512338757515, + 0.004828848410397768, + 0.008354336954653263, + 0.009261012077331543, + 0.005226633045822382, + 0.01233561523258686, + 0.004476373083889484, + 0.008249722421169281, + 0.004654666408896446, + 0.006793247535824776, + 0.012588420882821083, + 0.011194439604878426, + 0.007536442019045353, + 0.005525325890630484, + 0.005984095390886068, + 0.0062192827463150024, + 0.006114240735769272, + 0.007389930076897144, + 0.012802882120013237, + 0.012163511477410793, + 0.010981947183609009, + 0.014953218400478363, + 0.009503346867859364, + 0.01795802265405655, + 0.009167773649096489, + 0.0061890119686722755, + 0.015204685740172863, + 0.016274549067020416, + 0.011775470338761806, + 0.016646461561322212, + 0.0077631608583033085, + 0.012998665682971478, + 0.017263425514101982, + 0.03673867881298065, + 0.04825635626912117, + 0.03217544034123421, + 0.02157878689467907, + 0.015860658138990402, + 0.022849027067422867, + 0.06735619902610779, + 0.0046783750876784325, + 0.004766584374010563, + 0.013593648560345173, + 0.01180843822658062, + 0.0073881326243281364, + 0.003177575534209609, + 0.00904041901230812, + 0.011853460222482681 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 0, + 7 + ], + "selected_reward": -0.016409561038017273 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a muscular, middle-aged woodcutter with short dark brown hair and a thick beard?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the woodcutter wearing a frayed red and black flannel shirt with sleeves rolled up to the elbows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the woodcutter performing a powerful downward swing with an axe while leaning back with legs planted?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene set in a dense, primeval forest with towering ancient trees and dappled sunlight creating a Tyndall effect?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot captured in slow-motion showing sawdust exploding in crisp arcs against golden beams of light?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.008650331757962704 + ], + "opposite_logprob_tokens": [ + -4.758650302886963 + ], + "logprob_reward": -0.008650331757962704, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.008650331757962704, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15186715126037598, + 0.06008662283420563, + 0.08350006490945816, + 0.19151176512241364, + 0.2229848951101303, + 0.10502517968416214, + 0.04962913691997528, + 0.05869424715638161, + 0.030353544279932976, + 0.04634735733270645 + ], + "credit_map": [ + 0.01144573837518692, + 0.019150255247950554, + 0.021456969901919365, + 0.026223354041576385, + 0.0214679017663002, + 0.020217139273881912, + 0.018901091068983078, + 0.013004690408706665, + 0.0055279843509197235, + 0.006121442653238773, + 0.013294309377670288, + 0.015369248576462269, + 0.004120358731597662, + 0.004868186544626951, + 0.0045583476312458515, + 0.006226749159395695, + 0.007367745973169804, + 0.008940685540437698, + 0.012052393518388271, + 0.012678112834692001, + 0.00830373726785183, + 0.014338362030684948, + 0.011165694333612919, + 0.008653338067233562, + 0.008893474005162716, + 0.01212251652032137, + 0.046489182859659195, + 0.06946403533220291, + 0.012581361457705498, + 0.01647190749645233, + 0.017966635525226593, + 0.007522652391344309, + 0.0151603352278471, + 0.015457579866051674, + 0.030380668118596077, + 0.029024258255958557, + 0.01706795021891594, + 0.06679481267929077, + 0.018685143440961838, + 0.030414139851927757, + 0.008969987742602825, + 0.012504169717431068, + 0.02022429369390011, + 0.017740966752171516, + 0.010342296212911606, + 0.011501515284180641, + 0.013109433464705944, + 0.010632512159645557, + 0.004127068445086479, + 0.0054632509127259254, + 0.00814323965460062, + 0.007469277363270521, + 0.005515919532626867, + 0.0070550646632909775, + 0.005309366155415773, + 0.00654594786465168, + 0.007074616849422455, + 0.004775508772581816, + 0.009655299596488476, + 0.009914625436067581, + 0.005782071501016617, + 0.009150306694209576, + 0.004873082973062992, + 0.007468735799193382, + 0.003168117720633745, + 0.0026078615337610245, + 0.005178405903279781, + 0.004503226839005947, + 0.0040620132349431515, + 0.0026246586348861456, + 0.00383355631493032, + 0.004375704098492861, + 0.003685247153043747, + 0.004354123491793871, + 0.008079547435045242, + 0.0076997773721814156, + 0.006951218470931053, + 0.0032756649889051914, + 0.0057123638689517975, + 0.006589414551854134 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 3, + 0 + ], + "selected_reward": -0.008650331757962704 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a muscular, middle-aged woodcutter with short dark brown hair and a thick beard?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the woodcutter wearing a frayed red and black flannel shirt with sleeves rolled up to the elbows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the woodcutter performing a powerful downward swing with an axe while leaning back with legs planted?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene set in a dense, primeval forest with towering ancient trees and dappled sunlight creating a Tyndall effect?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot captured in slow-motion showing sawdust exploding in crisp arcs against golden beams of light?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004640643019229174 + ], + "opposite_logprob_tokens": [ + -5.379640579223633 + ], + "logprob_reward": -0.004640643019229174, + "yes_no_margin_reward": 5.375, + "token_credit_reward": -0.004640643019229174, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.17897573113441467, + 0.07298707216978073, + 0.03820520266890526, + 0.07086401432752609, + 0.09418822824954987, + 0.14289718866348267, + 0.08887507766485214, + 0.11187553405761719, + 0.10501179844141006, + 0.096120186150074 + ], + "credit_map": [ + 0.018410228192806244, + 0.03663500398397446, + 0.022703761234879494, + 0.03038380667567253, + 0.016060180962085724, + 0.016077736392617226, + 0.020794369280338287, + 0.017910638824105263, + 0.0087700504809618, + 0.009556235745549202, + 0.009877804666757584, + 0.012641958892345428, + 0.006964573636651039, + 0.008602248504757881, + 0.008475025184452534, + 0.008099176920950413, + 0.006915890146046877, + 0.00555583368986845, + 0.0055992258712649345, + 0.006525689270347357, + 0.0027800779789686203, + 0.0026684857439249754, + 0.004303911700844765, + 0.0038560896646231413, + 0.012793095782399178, + 0.006845274940133095, + 0.010509849525988102, + 0.012174071744084358, + 0.004900418221950531, + 0.012913312762975693, + 0.004801202565431595, + 0.005926788318902254, + 0.006695046555250883, + 0.009352779015898705, + 0.012949332594871521, + 0.013060489669442177, + 0.006386704742908478, + 0.022830942645668983, + 0.008073913864791393, + 0.01483901683241129, + 0.017695937305688858, + 0.009809925220906734, + 0.03201742470264435, + 0.02251594513654709, + 0.016735238954424858, + 0.015957403928041458, + 0.012549242004752159, + 0.015616078861057758, + 0.006044208072125912, + 0.008540195412933826, + 0.012122130021452904, + 0.01596107706427574, + 0.009679073467850685, + 0.007392892614006996, + 0.014524475671350956, + 0.014611029997467995, + 0.011364046484231949, + 0.007749598938971758, + 0.017246507108211517, + 0.016034260392189026, + 0.013576899655163288, + 0.01085892878472805, + 0.01087587233632803, + 0.02416941151022911, + 0.007128696423023939, + 0.01919473148882389, + 0.017200404778122902, + 0.013871564529836178, + 0.009573685005307198, + 0.006430388428270817, + 0.010619964450597763, + 0.020992375910282135, + 0.00747573422268033, + 0.0066065313294529915, + 0.018054155632853508, + 0.02132188156247139, + 0.008719705045223236, + 0.005804633256047964, + 0.01319698803126812, + 0.01494055986404419 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 5, + 7 + ], + "selected_reward": -0.004640643019229174 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a muscular, middle-aged woodcutter with short dark brown hair and a thick beard?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the woodcutter wearing a frayed red and black flannel shirt with sleeves rolled up to the elbows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the woodcutter performing a powerful downward swing with an axe while leaning back with legs planted?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene set in a dense, primeval forest with towering ancient trees and dappled sunlight creating a Tyndall effect?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot captured in slow-motion showing sawdust exploding in crisp arcs against golden beams of light?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01419247966259718 + ], + "opposite_logprob_tokens": [ + -4.264192581176758 + ], + "logprob_reward": -0.01419247966259718, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.01419247966259718, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10958247631788254, + 0.05593454837799072, + 0.0805085152387619, + 0.20190398395061493, + 0.09362645447254181, + 0.23202908039093018, + 0.08751321583986282, + 0.03635065630078316, + 0.04999856650829315, + 0.0525524877011776 + ], + "credit_map": [ + 0.013111932203173637, + 0.011564143002033234, + 0.026909388601779938, + 0.026283403858542442, + 0.005114628933370113, + 0.007094116881489754, + 0.011885400861501694, + 0.007619467563927174, + 0.005807592533528805, + 0.0051062991842627525, + 0.011915347538888454, + 0.016112379729747772, + 0.0037410687655210495, + 0.00495342118665576, + 0.004261976573616266, + 0.0040364619344472885, + 0.013267895206809044, + 0.0051516201347112656, + 0.019921498373150826, + 0.025643983855843544, + 0.0037722515407949686, + 0.0036503009032458067, + 0.0051016840152442455, + 0.0039992788806557655, + 0.014613903127610683, + 0.014485449530184269, + 0.05737772583961487, + 0.08086744695901871, + 0.005421365145593882, + 0.008817491121590137, + 0.013471894897520542, + 0.006848710123449564, + 0.013946901075541973, + 0.012199005112051964, + 0.015090221539139748, + 0.019579097628593445, + 0.005310370586812496, + 0.007276201155036688, + 0.010511055588722229, + 0.009713607840240002, + 0.0398450642824173, + 0.04082074761390686, + 0.034088317304849625, + 0.03491407260298729, + 0.022266985848546028, + 0.016427019611001015, + 0.026586255058646202, + 0.01708061434328556, + 0.011288131587207317, + 0.015792204067111015, + 0.012100608088076115, + 0.013988723047077656, + 0.007027871906757355, + 0.006322480272501707, + 0.011790364980697632, + 0.009202836081385612, + 0.004007875453680754, + 0.004956588614732027, + 0.004553187172859907, + 0.007728809956461191, + 0.00402431795373559, + 0.003625525627285242, + 0.002742941491305828, + 0.0047114090994000435, + 0.004973789677023888, + 0.005330412648618221, + 0.010954543016850948, + 0.004946287255734205, + 0.002552759600803256, + 0.002493324689567089, + 0.0027871434576809406, + 0.01596030779182911, + 0.005029714200645685, + 0.003190030809491873, + 0.013625364750623703, + 0.006930673029273748, + 0.004668002482503653, + 0.0033892225474119186, + 0.0046416777186095715, + 0.011077800765633583 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 3, + 0 + ], + "selected_reward": -0.01419247966259718 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a muscular, middle-aged woodcutter with short dark brown hair and a thick beard?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the woodcutter wearing a frayed red and black flannel shirt with sleeves rolled up to the elbows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the woodcutter performing a powerful downward swing with an axe while leaning back with legs planted?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene set in a dense, primeval forest with towering ancient trees and dappled sunlight creating a Tyndall effect?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot captured in slow-motion showing sawdust exploding in crisp arcs against golden beams of light?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0031923307105898857 + ], + "opposite_logprob_tokens": [ + -5.75319242477417 + ], + "logprob_reward": -0.0031923307105898857, + "yes_no_margin_reward": 5.75, + "token_credit_reward": -0.0031923307105898857, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1046118512749672, + 0.4337312877178192, + 0.03885219618678093, + 0.06821942329406738, + 0.034096598625183105, + 0.04248206689953804, + 0.08434024453163147, + 0.08561033010482788, + 0.050960540771484375, + 0.057095494121313095 + ], + "credit_map": [ + 0.015708258375525475, + 0.021282343193888664, + 0.015443876385688782, + 0.013266265392303467, + 0.008929816074669361, + 0.009719928726553917, + 0.008928174152970314, + 0.011333180591464043, + 0.037199582904577255, + 0.03876369073987007, + 0.11975016444921494, + 0.13714589178562164, + 0.021584533154964447, + 0.025395046919584274, + 0.03367244452238083, + 0.020219910889863968, + 0.0037669390439987183, + 0.003821871941909194, + 0.008580131456255913, + 0.01140047051012516, + 0.0023753917776048183, + 0.002886944217607379, + 0.0036538124550133944, + 0.0023666340857744217, + 0.0034592230804264545, + 0.0039265514351427555, + 0.018975764513015747, + 0.026003679260611534, + 0.003946339711546898, + 0.004330735187977552, + 0.004324628971517086, + 0.003252496477216482, + 0.0036967971827834845, + 0.0031627200078219175, + 0.005116929765790701, + 0.004891660530120134, + 0.0019353125244379044, + 0.008269055746495724, + 0.0024408679455518723, + 0.004583256784826517, + 0.003643689677119255, + 0.004661658778786659, + 0.00653873011469841, + 0.006457914132624865, + 0.00477872509509325, + 0.0044394894503057, + 0.007543601095676422, + 0.004418258089572191, + 0.0060106604360044, + 0.010734311304986477, + 0.011480353772640228, + 0.017723890021443367, + 0.007962115108966827, + 0.006640831008553505, + 0.013017568737268448, + 0.010770518332719803, + 0.00767322676256299, + 0.006485129706561565, + 0.012339649721980095, + 0.012114387936890125, + 0.01073763519525528, + 0.01544290129095316, + 0.006738247349858284, + 0.01407915074378252, + 0.003484179498627782, + 0.007351851090788841, + 0.008384451270103455, + 0.007333073299378157, + 0.005871572531759739, + 0.004462722223252058, + 0.007467338349670172, + 0.006605351343750954, + 0.006173834204673767, + 0.0033931017387658358, + 0.011911831796169281, + 0.009205878712236881, + 0.006889825686812401, + 0.004321702290326357, + 0.007734898943454027, + 0.007464422844350338 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 7 + ], + "selected_reward": -0.0031923307105898857 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000050/group_1/index_tensor([7573, 7573])_group_1_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000050/group_1/index_tensor([7573, 7573])_group_1_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..d88b2dd8d40926f54f717a131398d9a8e63c661d --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000050/group_1/index_tensor([7573, 7573])_group_1_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "dramatic, high-angle drone shot tracks steadily alongside a herd of camels as they descend perilously down the steep face of a massive sand dune. The intense midday sun casts long, deep blue shadows that stretch behind the group and highlight the golden-orange textures of the sand. Each camel is laden with towering masses of vibrant textiles; rich crimson wool, deep indigo silk, and saffron-yellow cotton, woven in intricate geometric patterns, are securely strapped to their backs, creating vivid colors amidst the monochrome desert. The camels swoop down not in chaos but with a controlled, rhythmic gait, their powerful legs sinking deeply into the moving surface, raising soft plumes of sand that catch the light like a golden mist. The sheer scale of the sun-scorched dune dominates the frame, its crest forming a sharp line against the vast, cloudless azure sky, creating an atmosphere of epic struggle and eternal majesty.", + "seed": 14176, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000050/group_1/index_7573_seed_14176_dp_2_rank_4_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 1, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.35030168294906616, + "vlm_logprob": -0.35030168294906616, + "vlm_yes_no_margin": 0.949999988079071, + "vlm_token_credit": -0.35030168294906616, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot taken from a high-angle drone perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera tracking steadily alongside a herd of camels?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the camels descending down the steep face of a massive sand dune?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the camels carry vibrant textiles including crimson wool, indigo silk, and saffron-yellow cotton?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the scene feature long, deep blue shadows cast by an intense midday sun?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.3390573263168335 + ], + "opposite_logprob_tokens": [ + -1.3390573263168335 + ], + "logprob_reward": -0.3390573263168335, + "yes_no_margin_reward": 1.0, + "token_credit_reward": -0.3390573263168335, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10800540447235107, + 0.10623546689748764, + 0.08725167065858841, + 0.21000659465789795, + 0.07182127237319946, + 0.16695182025432587, + 0.11767052859067917, + 0.04825517162680626, + 0.030205659568309784, + 0.05359639227390289 + ], + "credit_map": [ + 0.014865496195852757, + 0.01753263548016548, + 0.012302589602768421, + 0.014386037364602089, + 0.010460762307047844, + 0.007031397428363562, + 0.013541234657168388, + 0.017885247245430946, + 0.012680862098932266, + 0.013659538701176643, + 0.016779154539108276, + 0.01373231876641512, + 0.009532314725220203, + 0.008074580691754818, + 0.015568187460303307, + 0.01620851829648018, + 0.010345686227083206, + 0.011501424945890903, + 0.008810564875602722, + 0.012593387626111507, + 0.01044329721480608, + 0.009112372994422913, + 0.012217641808092594, + 0.012227300554513931, + 0.015733903273940086, + 0.01521978247910738, + 0.03165624663233757, + 0.03754604607820511, + 0.019580531865358353, + 0.022226152941584587, + 0.03809412568807602, + 0.029949812218546867, + 0.009409224614501, + 0.011249684728682041, + 0.00814736820757389, + 0.013646744191646576, + 0.007072872016578913, + 0.006284418050199747, + 0.008121672086417675, + 0.007889294996857643, + 0.01594218984246254, + 0.01939086616039276, + 0.031038010492920876, + 0.021592140197753906, + 0.015764135867357254, + 0.016629181802272797, + 0.023762034252285957, + 0.022833257913589478, + 0.014956390485167503, + 0.01963619515299797, + 0.019769147038459778, + 0.01767374761402607, + 0.013410612940788269, + 0.008246680721640587, + 0.012757431715726852, + 0.011220329441130161, + 0.005127649754285812, + 0.005999722983688116, + 0.007827717810869217, + 0.010363283567130566, + 0.003914843779057264, + 0.0029357511084526777, + 0.006659563165158033, + 0.0054266429506242275, + 0.0026996631640940905, + 0.004147353116422892, + 0.003074727952480316, + 0.0037361388094723225, + 0.0030289862770587206, + 0.002790977945551276, + 0.005218125879764557, + 0.005509685259312391, + 0.0040489924140274525, + 0.005177510902285576, + 0.01192875113338232, + 0.007675532251596451, + 0.004562527406960726, + 0.003456525271758437, + 0.008781897835433483, + 0.007964656688272953 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 5, + 6 + ], + "selected_reward": -0.3390573263168335 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot taken from a high-angle drone perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera tracking steadily alongside a herd of camels?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the camels descending down the steep face of a massive sand dune?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the camels carry vibrant textiles including crimson wool, indigo silk, and saffron-yellow cotton?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the scene feature long, deep blue shadows cast by an intense midday sun?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.3487251400947571 + ], + "opposite_logprob_tokens": [ + -1.2237250804901123 + ], + "logprob_reward": -0.3487251400947571, + "yes_no_margin_reward": 0.8749999403953552, + "token_credit_reward": -0.3487251400947571, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06228142976760864, + 0.096495620906353, + 0.029161185026168823, + 0.41974708437919617, + 0.03293994069099426, + 0.1195773258805275, + 0.06648650020360947, + 0.14666175842285156, + 0.017049560323357582, + 0.009599671699106693 + ], + "credit_map": [ + 0.006628415547311306, + 0.007462772075086832, + 0.008024989627301693, + 0.01333020068705082, + 0.006688706576824188, + 0.003949508536607027, + 0.008024166338145733, + 0.00817266758531332, + 0.012617751024663448, + 0.01417126040905714, + 0.010391859337687492, + 0.011082104407250881, + 0.0076005118899047375, + 0.008053391240537167, + 0.016933757811784744, + 0.0156449805945158, + 0.0030410645995289087, + 0.003219994017854333, + 0.005846237763762474, + 0.005031554028391838, + 0.0028817164711654186, + 0.002041187835857272, + 0.003289648564532399, + 0.003809782210737467, + 0.034405481070280075, + 0.03686859831213951, + 0.047985341399908066, + 0.07995723187923431, + 0.03602868691086769, + 0.0435037836432457, + 0.0765138790011406, + 0.06448403745889664, + 0.004394682124257088, + 0.004528854042291641, + 0.004042796790599823, + 0.004954751580953598, + 0.0029032581951469183, + 0.003250923939049244, + 0.0048532928340137005, + 0.004011379089206457, + 0.011906499043107033, + 0.013232615776360035, + 0.017595866695046425, + 0.018294068053364754, + 0.014086369425058365, + 0.011353674344718456, + 0.018034150823950768, + 0.015074068680405617, + 0.007243243046104908, + 0.009743263944983482, + 0.008276905864477158, + 0.011381329968571663, + 0.005859061144292355, + 0.006055472884327173, + 0.00973760150372982, + 0.008189614862203598, + 0.013010203838348389, + 0.019003545865416527, + 0.02397426776587963, + 0.03085239604115486, + 0.009728152304887772, + 0.01116665918380022, + 0.021266886964440346, + 0.017659630626440048, + 0.0014063108246773481, + 0.002102319151163101, + 0.001815277966670692, + 0.0018525253981351852, + 0.0016892425483092666, + 0.0011119147529825568, + 0.0033432452473789454, + 0.003728723619133234, + 0.0008034417987801135, + 0.0012763807317242026, + 0.0016463404754176736, + 0.001018137321807444, + 0.0012037677224725485, + 0.000875583675224334, + 0.0014928021701052785, + 0.0012832169886678457 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 7, + 5 + ], + "selected_reward": -0.3487251400947571 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot taken from a high-angle drone perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera tracking steadily alongside a herd of camels?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the camels descending down the steep face of a massive sand dune?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the camels carry vibrant textiles including crimson wool, indigo silk, and saffron-yellow cotton?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the scene feature long, deep blue shadows cast by an intense midday sun?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.42895445227622986 + ], + "opposite_logprob_tokens": [ + -1.0539544820785522 + ], + "logprob_reward": -0.42895445227622986, + "yes_no_margin_reward": 0.625, + "token_credit_reward": -0.42895445227622986, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08113587647676468, + 0.04023507237434387, + 0.08781583607196808, + 0.44759076833724976, + 0.03595612198114395, + 0.1052146777510643, + 0.10756786167621613, + 0.030785515904426575, + 0.04261630028486252, + 0.02108195424079895 + ], + "credit_map": [ + 0.01057919580489397, + 0.014275613240897655, + 0.008117091841995716, + 0.01337972842156887, + 0.011350689455866814, + 0.005239761434495449, + 0.009213188663125038, + 0.00898061040788889, + 0.004586901981383562, + 0.005168842151761055, + 0.007175130769610405, + 0.003516509896144271, + 0.003304991638287902, + 0.004427788779139519, + 0.005715361796319485, + 0.006339545827358961, + 0.009816356003284454, + 0.01170960534363985, + 0.012445622123777866, + 0.014568791724741459, + 0.010953821241855621, + 0.007139512803405523, + 0.009954790584743023, + 0.011227340437471867, + 0.049501631408929825, + 0.05916275084018707, + 0.04644358530640602, + 0.06184692308306694, + 0.04256517067551613, + 0.047074832022190094, + 0.07576106488704681, + 0.06523479521274567, + 0.004203025717288256, + 0.004635373130440712, + 0.005400747060775757, + 0.005116606596857309, + 0.004337582271546125, + 0.004162599798291922, + 0.0040153986774384975, + 0.004084790125489235, + 0.01024913415312767, + 0.013574758544564247, + 0.016627006232738495, + 0.017070753499865532, + 0.011243934743106365, + 0.007907385006546974, + 0.015458912588655949, + 0.013082795776426792, + 0.014690092764794827, + 0.02067519538104534, + 0.016463959589600563, + 0.01380927860736847, + 0.010708015412092209, + 0.008293894119560719, + 0.012464632280170918, + 0.010462791658937931, + 0.0032415189780294895, + 0.004902272019535303, + 0.00451238127425313, + 0.004487097728997469, + 0.002665516221895814, + 0.0028851788956671953, + 0.004414776340126991, + 0.003676772816106677, + 0.004117870237678289, + 0.006202293559908867, + 0.004194289445877075, + 0.003834553062915802, + 0.0037851398810744286, + 0.003967247437685728, + 0.008687775582075119, + 0.007827128283679485, + 0.001884656143374741, + 0.0022716261446475983, + 0.0028376656118780375, + 0.0038875627797096968, + 0.002302246866747737, + 0.00132459937594831, + 0.0035870119463652372, + 0.0029865833930671215 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 6, + 5 + ], + "selected_reward": -0.42895445227622986 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot taken from a high-angle drone perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera tracking steadily alongside a herd of camels?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the camels descending down the steep face of a massive sand dune?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the camels carry vibrant textiles including crimson wool, indigo silk, and saffron-yellow cotton?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the scene feature long, deep blue shadows cast by an intense midday sun?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.16046792268753052 + ], + "opposite_logprob_tokens": [ + -1.9104678630828857 + ], + "logprob_reward": -0.16046792268753052, + "yes_no_margin_reward": 1.75, + "token_credit_reward": -0.16046792268753052, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07043134421110153, + 0.021673373878002167, + 0.05459652468562126, + 0.3211400508880615, + 0.0261884406208992, + 0.3064107298851013, + 0.05239980295300484, + 0.04584572836756706, + 0.058701205998659134, + 0.042612843215465546 + ], + "credit_map": [ + 0.01000133529305458, + 0.011777276173233986, + 0.009885260835289955, + 0.011553755030035973, + 0.007568002678453922, + 0.004780557472258806, + 0.007033215370029211, + 0.007831936702132225, + 0.002797371009364724, + 0.0034056154545396566, + 0.002600465202704072, + 0.001888635684736073, + 0.002614827360957861, + 0.002486716490238905, + 0.0026776331942528486, + 0.003202110528945923, + 0.005834956653416157, + 0.006520276889204979, + 0.006581338122487068, + 0.009681861847639084, + 0.005664945114403963, + 0.00435677170753479, + 0.007954679429531097, + 0.008001696318387985, + 0.027971260249614716, + 0.027339423075318336, + 0.04486066848039627, + 0.06410406529903412, + 0.03263320028781891, + 0.029521703720092773, + 0.0506712831556797, + 0.04403847083449364, + 0.00327559607103467, + 0.0034101135097444057, + 0.0038152129855006933, + 0.003103054827079177, + 0.003585363505408168, + 0.0025727867614477873, + 0.003675442421808839, + 0.0027508700732141733, + 0.04099436476826668, + 0.0552327036857605, + 0.04374496638774872, + 0.04336307570338249, + 0.028556790202856064, + 0.01922176405787468, + 0.038376107811927795, + 0.036920949816703796, + 0.0036675515584647655, + 0.0041403272189199924, + 0.006555987522006035, + 0.006484173703938723, + 0.004814956337213516, + 0.004304327070713043, + 0.012175184674561024, + 0.010257294401526451, + 0.005162861663848162, + 0.007923256605863571, + 0.00647496385499835, + 0.008083543740212917, + 0.004690946079790592, + 0.0031840060837566853, + 0.0054905298165977, + 0.004835615865886211, + 0.004202838521450758, + 0.005505152978003025, + 0.007208235561847687, + 0.009240190498530865, + 0.004258609376847744, + 0.004522526171058416, + 0.01152887661010027, + 0.012234778143465519, + 0.004147470463067293, + 0.005060294643044472, + 0.0067645288072526455, + 0.004458662588149309, + 0.006439013406634331, + 0.0036689864937216043, + 0.006062878295779228, + 0.0060110073536634445 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 5, + 0 + ], + "selected_reward": -0.16046792268753052 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot taken from a high-angle drone perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera tracking steadily alongside a herd of camels?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the camels descending down the steep face of a massive sand dune?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the camels carry vibrant textiles including crimson wool, indigo silk, and saffron-yellow cotton?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the scene feature long, deep blue shadows cast by an intense midday sun?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.47430357336997986 + ], + "opposite_logprob_tokens": [ + -0.9743036031723022 + ], + "logprob_reward": -0.47430357336997986, + "yes_no_margin_reward": 0.5, + "token_credit_reward": -0.47430357336997986, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07095356285572052, + 0.20294055342674255, + 0.11299004405736923, + 0.1913139373064041, + 0.07325078547000885, + 0.07507196068763733, + 0.0684717521071434, + 0.08134618401527405, + 0.060456518083810806, + 0.06320464611053467 + ], + "credit_map": [ + 0.008417756296694279, + 0.008305475115776062, + 0.008028787560760975, + 0.01588807813823223, + 0.007344000972807407, + 0.004678379278630018, + 0.008852646686136723, + 0.009438439272344112, + 0.02810797654092312, + 0.031565066426992416, + 0.020243214443325996, + 0.021529890596866608, + 0.01765579916536808, + 0.01813618279993534, + 0.03546632453799248, + 0.030236106365919113, + 0.011691193096339703, + 0.013231073506176472, + 0.01887642592191696, + 0.017213962972164154, + 0.012116062454879284, + 0.008530671708285809, + 0.015327964909374714, + 0.016002686694264412, + 0.015390045009553432, + 0.01417765486985445, + 0.026292581111192703, + 0.032887402921915054, + 0.017786746844649315, + 0.021832704544067383, + 0.03368895873427391, + 0.029257841408252716, + 0.008000249043107033, + 0.00914083607494831, + 0.009860859252512455, + 0.0136625487357378, + 0.006502174772322178, + 0.007331592962145805, + 0.009590382687747478, + 0.00916214194148779, + 0.00759865390136838, + 0.00688517140224576, + 0.010014059953391552, + 0.008799182251095772, + 0.010346216149628162, + 0.007244471926242113, + 0.014864792115986347, + 0.009319409728050232, + 0.006425538565963507, + 0.007514952681958675, + 0.006693695671856403, + 0.009044552221894264, + 0.008817251771688461, + 0.006495778914541006, + 0.012710848823189735, + 0.010769132524728775, + 0.008557409048080444, + 0.011915076524019241, + 0.008973128162324429, + 0.009433146566152573, + 0.007397002540528774, + 0.00605363305658102, + 0.01686323806643486, + 0.012153550051152706, + 0.004890060052275658, + 0.006523735821247101, + 0.005982124246656895, + 0.006325740832835436, + 0.008431650698184967, + 0.004609625320881605, + 0.011856097728013992, + 0.011837484315037727, + 0.005983440671116114, + 0.006036844570189714, + 0.012221770361065865, + 0.007929431274533272, + 0.006209774408489466, + 0.004735852591693401, + 0.010474059730768204, + 0.009613477624952793 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 3, + 2 + ], + "selected_reward": -0.47430357336997986 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "dramatic, high-angle drone shot tracks steadily alongside a herd of camels as they descend perilously down the steep face of a massive sand dune. The intense midday sun casts long, deep blue shadows that stretch behind the group and highlight the golden-orange textures of the sand. Each camel is laden with towering masses of vibrant textiles; rich crimson wool, deep indigo silk, and saffron-yellow cotton, woven in intricate geometric patterns, are securely strapped to their backs, creating vivid colors amidst the monochrome desert. The camels swoop down not in chaos but with a controlled, rhythmic gait, their powerful legs sinking deeply into the moving surface, raising soft plumes of sand that catch the light like a golden mist. The sheer scale of the sun-scorched dune dominates the frame, its crest forming a sharp line against the vast, cloudless azure sky, creating an atmosphere of epic struggle and eternal majesty.", + "seed": 14177, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000050/group_1/index_7573_seed_14177_dp_2_rank_4_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 1, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.2866948634386063, + "vlm_logprob": -0.2866948634386063, + "vlm_yes_no_margin": 1.3, + "vlm_token_credit": -0.2866948634386063, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot taken from a high-angle drone perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera tracking steadily alongside a herd of camels?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the camels descending down the steep face of a massive sand dune?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the camels carry vibrant textiles including crimson wool, indigo silk, and saffron-yellow cotton?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the scene feature long, deep blue shadows cast by an intense midday sun?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.278423935174942 + ], + "opposite_logprob_tokens": [ + -1.5284239053726196 + ], + "logprob_reward": -0.278423935174942, + "yes_no_margin_reward": 1.25, + "token_credit_reward": -0.278423935174942, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.18145407736301422, + 0.24706879258155823, + 0.06615377217531204, + 0.1477946639060974, + 0.04972673952579498, + 0.026660297065973282, + 0.03636708855628967, + 0.17135773599147797, + 0.02782294526696205, + 0.04559389874339104 + ], + "credit_map": [ + 0.019708048552274704, + 0.01959323137998581, + 0.023097869008779526, + 0.024825388565659523, + 0.014703309163451195, + 0.031238067895174026, + 0.02340831607580185, + 0.024879854172468185, + 0.019500868394970894, + 0.020740119740366936, + 0.042058803141117096, + 0.04616890102624893, + 0.024764271453022957, + 0.028427381068468094, + 0.03466658294200897, + 0.030741870403289795, + 0.005071340128779411, + 0.003629503306001425, + 0.00872872769832611, + 0.014384178444743156, + 0.006452625151723623, + 0.009479518979787827, + 0.008306511677801609, + 0.01010136865079403, + 0.01619567908346653, + 0.008917063474655151, + 0.014373992569744587, + 0.01538936235010624, + 0.014428991824388504, + 0.0382789708673954, + 0.017455216497182846, + 0.02275538444519043, + 0.00638157594949007, + 0.0033532173838466406, + 0.007832559756934643, + 0.007504291366785765, + 0.005394055508077145, + 0.00592742720618844, + 0.00644515547901392, + 0.006888458505272865, + 0.0023571045603603125, + 0.0024331414606422186, + 0.005540556740015745, + 0.0032885672990232706, + 0.001856391434557736, + 0.004272748250514269, + 0.0030236458405852318, + 0.0038881392683833838, + 0.004120417404919863, + 0.005219900049269199, + 0.005610649939626455, + 0.005603835918009281, + 0.0034037174191325903, + 0.0034643698018044233, + 0.004291995894163847, + 0.004652203526347876, + 0.02944532223045826, + 0.01791330985724926, + 0.018313229084014893, + 0.01761700212955475, + 0.02155652642250061, + 0.016563279554247856, + 0.020482733845710754, + 0.02946631982922554, + 0.003503113053739071, + 0.003392742946743965, + 0.0032487711869180202, + 0.00266320095397532, + 0.0030512050725519657, + 0.0033925664611160755, + 0.003391861217096448, + 0.005179485306143761, + 0.0062722438015043736, + 0.005140085704624653, + 0.008442898280918598, + 0.005793651565909386, + 0.003971057012677193, + 0.0058169979602098465, + 0.005066685378551483, + 0.00509027810767293 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 7 + ], + "selected_reward": -0.278423935174942 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot taken from a high-angle drone perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera tracking steadily alongside a herd of camels?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the camels descending down the steep face of a massive sand dune?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the camels carry vibrant textiles including crimson wool, indigo silk, and saffron-yellow cotton?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the scene feature long, deep blue shadows cast by an intense midday sun?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.14289572834968567 + ], + "opposite_logprob_tokens": [ + -2.0178956985473633 + ], + "logprob_reward": -0.14289572834968567, + "yes_no_margin_reward": 1.875, + "token_credit_reward": -0.14289572834968567, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.17113007605075836, + 0.22368696331977844, + 0.08372090011835098, + 0.06532390415668488, + 0.044333167374134064, + 0.030426140874624252, + 0.03268226981163025, + 0.24185976386070251, + 0.0761069506406784, + 0.030729876831173897 + ], + "credit_map": [ + 0.020933032035827637, + 0.021390151232481003, + 0.02124685049057007, + 0.017757097259163857, + 0.01631888933479786, + 0.028850458562374115, + 0.023388508707284927, + 0.021245073527097702, + 0.020572494715452194, + 0.016284335404634476, + 0.03558334335684776, + 0.0421738475561142, + 0.019455192610621452, + 0.029483672231435776, + 0.03240593150258064, + 0.027728121727705002, + 0.011179864406585693, + 0.004683420993387699, + 0.010172022506594658, + 0.009378383867442608, + 0.007189915515482426, + 0.0198119655251503, + 0.008447143249213696, + 0.012858176603913307, + 0.006834660191088915, + 0.0037225070409476757, + 0.007262700237333775, + 0.006705749314278364, + 0.007066579535603523, + 0.015127158723771572, + 0.008968543261289597, + 0.009636001661419868, + 0.004529144614934921, + 0.002638221951201558, + 0.006026660557836294, + 0.006591715384274721, + 0.0033939944114536047, + 0.010406779125332832, + 0.0042174882255494595, + 0.006529158912599087, + 0.0023477505892515182, + 0.002264267997816205, + 0.004998334217816591, + 0.004361255560070276, + 0.0023591932840645313, + 0.005419882480055094, + 0.004372735973447561, + 0.004302716813981533, + 0.0025782163720577955, + 0.0032868573907762766, + 0.005118943750858307, + 0.0035285840276628733, + 0.004040309693664312, + 0.004849613644182682, + 0.005183449946343899, + 0.004096292424947023, + 0.03349298983812332, + 0.02530692145228386, + 0.03683393448591232, + 0.03128812834620476, + 0.03605518117547035, + 0.02678653411567211, + 0.02412291243672371, + 0.027973156422376633, + 0.010053244419395924, + 0.008317523635923862, + 0.010974816046655178, + 0.008545459248125553, + 0.006893702317029238, + 0.007665704470127821, + 0.010022185742855072, + 0.013634308241307735, + 0.005672474857419729, + 0.0031463156919926405, + 0.0038470569998025894, + 0.0037898197770118713, + 0.0032948607113212347, + 0.002754608169198036, + 0.004105961415916681, + 0.004118776880204678 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 1, + 0 + ], + "selected_reward": -0.14289572834968567 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot taken from a high-angle drone perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera tracking steadily alongside a herd of camels?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the camels descending down the steep face of a massive sand dune?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the camels carry vibrant textiles including crimson wool, indigo silk, and saffron-yellow cotton?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the scene feature long, deep blue shadows cast by an intense midday sun?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.12708018720149994 + ], + "opposite_logprob_tokens": [ + -2.127080202102661 + ], + "logprob_reward": -0.12708018720149994, + "yes_no_margin_reward": 2.0, + "token_credit_reward": -0.12708018720149994, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08233330398797989, + 0.08225731551647186, + 0.043128881603479385, + 0.03336476534605026, + 0.06378322839736938, + 0.03980577364563942, + 0.020219383761286736, + 0.5870318412780762, + 0.024902313947677612, + 0.02317321114242077 + ], + "credit_map": [ + 0.008681273087859154, + 0.0108408872038126, + 0.0070473263040184975, + 0.007079225033521652, + 0.007088109385222197, + 0.01893557980656624, + 0.01100871805101633, + 0.011652180925011635, + 0.007656360976397991, + 0.00689004547894001, + 0.013419944792985916, + 0.014613301493227482, + 0.007745953276753426, + 0.01073259674012661, + 0.010873517021536827, + 0.01032558735460043, + 0.005411916878074408, + 0.002536159474402666, + 0.004654069431126118, + 0.004743671510368586, + 0.003434441052377224, + 0.010752585716545582, + 0.004539650399237871, + 0.007056386210024357, + 0.0033888500183820724, + 0.0019156730268150568, + 0.003256184747442603, + 0.0035778118763118982, + 0.0033625634387135506, + 0.009054900147020817, + 0.0034798255655914545, + 0.0053289541974663734, + 0.005151779856532812, + 0.003808023640885949, + 0.009935217909514904, + 0.01078016310930252, + 0.006762043572962284, + 0.009408039972186089, + 0.008828935213387012, + 0.009109023958444595, + 0.0032282364554703236, + 0.0031792004592716694, + 0.006333993747830391, + 0.003586494829505682, + 0.0032061163801699877, + 0.011021455749869347, + 0.003947259392589331, + 0.005303016863763332, + 0.001783651765435934, + 0.002133852569386363, + 0.003466310678049922, + 0.002831028075888753, + 0.0017148705665022135, + 0.0023604214657098055, + 0.0033427912276238203, + 0.002586458344012499, + 0.09581859409809113, + 0.058442749083042145, + 0.07870317250490189, + 0.06663612276315689, + 0.07394316047430038, + 0.06168510392308235, + 0.06062432378530502, + 0.09117860347032547, + 0.003378987079486251, + 0.002850891090929508, + 0.003901589196175337, + 0.002959644654765725, + 0.0022118398919701576, + 0.0026708226650953293, + 0.003125807037577033, + 0.0038027316331863403, + 0.0031364434398710728, + 0.0022810976952314377, + 0.003941086586564779, + 0.0031970140989869833, + 0.0025519828777760267, + 0.0029520944226533175, + 0.0026509230956435204, + 0.0024625686928629875 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 0, + 1 + ], + "selected_reward": -0.12708018720149994 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot taken from a high-angle drone perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera tracking steadily alongside a herd of camels?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the camels descending down the steep face of a massive sand dune?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the camels carry vibrant textiles including crimson wool, indigo silk, and saffron-yellow cotton?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the scene feature long, deep blue shadows cast by an intense midday sun?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.25223562121391296 + ], + "opposite_logprob_tokens": [ + -1.5022356510162354 + ], + "logprob_reward": -0.25223562121391296, + "yes_no_margin_reward": 1.25, + "token_credit_reward": -0.25223562121391296, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1600746214389801, + 0.12012632191181183, + 0.0661110207438469, + 0.10526705533266068, + 0.06791244447231293, + 0.042635537683963776, + 0.058248911052942276, + 0.2818340063095093, + 0.03982022404670715, + 0.057969897985458374 + ], + "credit_map": [ + 0.013417661190032959, + 0.015124259516596794, + 0.0275503508746624, + 0.027513764798641205, + 0.012957471422851086, + 0.017092019319534302, + 0.02239076793193817, + 0.024028325453400612, + 0.012450022622942924, + 0.008679317310452461, + 0.017707526683807373, + 0.02122517302632332, + 0.010493635199964046, + 0.015827061608433723, + 0.018816228955984116, + 0.014927363023161888, + 0.005770677234977484, + 0.003585232887417078, + 0.0129670025780797, + 0.010393135249614716, + 0.005334473215043545, + 0.011142521165311337, + 0.007458755746483803, + 0.00945922639220953, + 0.0089850639924407, + 0.006272772792726755, + 0.009993841871619225, + 0.011764395982027054, + 0.01041379850357771, + 0.03127165883779526, + 0.009318700060248375, + 0.017246820032596588, + 0.005595004186034203, + 0.0044207945466041565, + 0.010105503723025322, + 0.011523404158651829, + 0.0076979235745966434, + 0.00851371418684721, + 0.010271136648952961, + 0.00978496391326189, + 0.0034043644554913044, + 0.003022236982360482, + 0.007597444579005241, + 0.004908207803964615, + 0.0028728372417390347, + 0.009532998315989971, + 0.005093383602797985, + 0.006204067729413509, + 0.005674779415130615, + 0.006374158896505833, + 0.007614317815750837, + 0.007593964692205191, + 0.0060198865830898285, + 0.009097716771066189, + 0.008932256139814854, + 0.006941831670701504, + 0.042367592453956604, + 0.025979921221733093, + 0.044658564031124115, + 0.03492284566164017, + 0.03165543079376221, + 0.03167053312063217, + 0.030274726450443268, + 0.040304385125637054, + 0.0035030220169574022, + 0.0031702944543212652, + 0.0055075399577617645, + 0.004744111560285091, + 0.006824079900979996, + 0.006636513862758875, + 0.004806847777217627, + 0.004627816844731569, + 0.0075371707789599895, + 0.005944536998867989, + 0.009965356439352036, + 0.008010849356651306, + 0.0060506840236485004, + 0.00651653204113245, + 0.007155065890401602, + 0.006789701525121927 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 0, + 1 + ], + "selected_reward": -0.25223562121391296 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot taken from a high-angle drone perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera tracking steadily alongside a herd of camels?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the camels descending down the steep face of a massive sand dune?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the camels carry vibrant textiles including crimson wool, indigo silk, and saffron-yellow cotton?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the scene feature long, deep blue shadows cast by an intense midday sun?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.6328388452529907 + ], + "opposite_logprob_tokens": [ + -0.7578388452529907 + ], + "logprob_reward": -0.6328388452529907, + "yes_no_margin_reward": 0.125, + "token_credit_reward": -0.6328388452529907, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09176851063966751, + 0.1256455034017563, + 0.07705027610063553, + 0.058525700122117996, + 0.02224522829055786, + 0.03403960168361664, + 0.03833358734846115, + 0.3927324116230011, + 0.06675349175930023, + 0.0929056778550148 + ], + "credit_map": [ + 0.010812857188284397, + 0.007265610620379448, + 0.013514681719243526, + 0.012898817658424377, + 0.010318350978195667, + 0.012617101892828941, + 0.011714484542608261, + 0.012626607902348042, + 0.012784281745553017, + 0.008450287394225597, + 0.02143041417002678, + 0.022275948897004128, + 0.010608240962028503, + 0.018683195114135742, + 0.015937376767396927, + 0.015475760214030743, + 0.010012832470238209, + 0.0055418433621525764, + 0.007372711785137653, + 0.007933862507343292, + 0.00602786336094141, + 0.021105531603097916, + 0.007539359387010336, + 0.011516274884343147, + 0.005320773459970951, + 0.0037495484575629234, + 0.007719581015408039, + 0.007882327772676945, + 0.006357881240546703, + 0.008397020399570465, + 0.010441435500979424, + 0.00865713506937027, + 0.002272192155942321, + 0.001393945887684822, + 0.004037755541503429, + 0.0036725362297147512, + 0.0015254283789545298, + 0.004129436332732439, + 0.0019982471130788326, + 0.003215687582269311, + 0.004220142960548401, + 0.0033626961521804333, + 0.005560111720114946, + 0.0029440554790198803, + 0.0020721822511404753, + 0.007661090698093176, + 0.0033376659266650677, + 0.004881654866039753, + 0.003748225048184395, + 0.005134352017194033, + 0.005583725869655609, + 0.005334766581654549, + 0.0035663899034261703, + 0.005475916434079409, + 0.004537724424153566, + 0.0049524870701134205, + 0.06463927775621414, + 0.036925531923770905, + 0.03853048011660576, + 0.0394047312438488, + 0.05196278542280197, + 0.04037901386618614, + 0.05191734805703163, + 0.06897322088479996, + 0.005707929376512766, + 0.005162432324141264, + 0.008521303534507751, + 0.005203642416745424, + 0.007766633294522762, + 0.009756837040185928, + 0.010680057108402252, + 0.013954659923911095, + 0.0115277748554945, + 0.014165205880999565, + 0.013032508082687855, + 0.008020849898457527, + 0.010156657546758652, + 0.010981707833707333, + 0.011457025073468685, + 0.013563951477408409 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 1, + 9 + ], + "selected_reward": -0.6328388452529907 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "dramatic, high-angle drone shot tracks steadily alongside a herd of camels as they descend perilously down the steep face of a massive sand dune. The intense midday sun casts long, deep blue shadows that stretch behind the group and highlight the golden-orange textures of the sand. Each camel is laden with towering masses of vibrant textiles; rich crimson wool, deep indigo silk, and saffron-yellow cotton, woven in intricate geometric patterns, are securely strapped to their backs, creating vivid colors amidst the monochrome desert. The camels swoop down not in chaos but with a controlled, rhythmic gait, their powerful legs sinking deeply into the moving surface, raising soft plumes of sand that catch the light like a golden mist. The sheer scale of the sun-scorched dune dominates the frame, its crest forming a sharp line against the vast, cloudless azure sky, creating an atmosphere of epic struggle and eternal majesty.", + "seed": 14178, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000050/group_1/index_7573_seed_14178_dp_3_rank_6_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 1, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.3172062009572983, + "vlm_logprob": -0.3172062009572983, + "vlm_yes_no_margin": 1.100000023841858, + "vlm_token_credit": -0.3172062009572983, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot taken from a high-angle drone perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera tracking steadily alongside a herd of camels?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the camels descending down the steep face of a massive sand dune?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the camels carry vibrant textiles including crimson wool, indigo silk, and saffron-yellow cotton?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the scene feature long, deep blue shadows cast by an intense midday sun?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.3334314227104187 + ], + "opposite_logprob_tokens": [ + -1.3334314823150635 + ], + "logprob_reward": -0.3334314227104187, + "yes_no_margin_reward": 1.0, + "token_credit_reward": -0.3334314227104187, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10290840268135071, + 0.09232600778341293, + 0.06355270743370056, + 0.1236468032002449, + 0.15046817064285278, + 0.2070574164390564, + 0.04594433680176735, + 0.11164648830890656, + 0.061592865735292435, + 0.04085681959986687 + ], + "credit_map": [ + 0.012388696894049644, + 0.013452741317451, + 0.015336859971284866, + 0.012436932884156704, + 0.012416649609804153, + 0.012741368263959885, + 0.011449954472482204, + 0.01268518902361393, + 0.009019470773637295, + 0.008855601772665977, + 0.01173061691224575, + 0.017339292913675308, + 0.012957789935171604, + 0.010535884648561478, + 0.009194571524858475, + 0.012692767195403576, + 0.005925117060542107, + 0.00833370815962553, + 0.009210947901010513, + 0.008479153737425804, + 0.005939149763435125, + 0.007361843250691891, + 0.008381997235119343, + 0.009920787997543812, + 0.011043074540793896, + 0.017944611608982086, + 0.014928213320672512, + 0.015039426274597645, + 0.01764513924717903, + 0.01017038431018591, + 0.018140679225325584, + 0.018735267221927643, + 0.011695208959281445, + 0.017366845160722733, + 0.02869018167257309, + 0.023948566988110542, + 0.012014167383313179, + 0.017164500430226326, + 0.019401241093873978, + 0.02018745057284832, + 0.019598087295889854, + 0.03199305385351181, + 0.02520258165895939, + 0.025473646819591522, + 0.03404182940721512, + 0.016430571675300598, + 0.026257123798131943, + 0.02806050516664982, + 0.004392133094370365, + 0.006724025122821331, + 0.006262528710067272, + 0.004986400716006756, + 0.0036852904595434666, + 0.003708470147103071, + 0.006188231986016035, + 0.009997253306210041, + 0.008425239473581314, + 0.007825253531336784, + 0.015141583047807217, + 0.012662457302212715, + 0.021736953407526016, + 0.011093278415501118, + 0.019367950037121773, + 0.015393756330013275, + 0.006524646654725075, + 0.006569383200258017, + 0.008251755498349667, + 0.009576584212481976, + 0.005312415771186352, + 0.005983063019812107, + 0.009573323652148247, + 0.009801688604056835, + 0.0035972422920167446, + 0.005471251904964447, + 0.006160976830869913, + 0.006709555629640818, + 0.0026895443443208933, + 0.006580251269042492, + 0.00522656599059701, + 0.00442142877727747 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 4, + 3 + ], + "selected_reward": -0.3334314227104187 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot taken from a high-angle drone perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera tracking steadily alongside a herd of camels?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the camels descending down the steep face of a massive sand dune?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the camels carry vibrant textiles including crimson wool, indigo silk, and saffron-yellow cotton?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the scene feature long, deep blue shadows cast by an intense midday sun?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.14287103712558746 + ], + "opposite_logprob_tokens": [ + -2.017871141433716 + ], + "logprob_reward": -0.14287103712558746, + "yes_no_margin_reward": 1.8750001192092896, + "token_credit_reward": -0.14287103712558746, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08911366760730743, + 0.15570609271526337, + 0.145009383559227, + 0.06306836009025574, + 0.13730405271053314, + 0.06760433316230774, + 0.08975771069526672, + 0.12275008112192154, + 0.09757262468338013, + 0.03211364895105362 + ], + "credit_map": [ + 0.008206759579479694, + 0.01259849313646555, + 0.010947788134217262, + 0.009411652572453022, + 0.00967463944107294, + 0.008919059298932552, + 0.018815290182828903, + 0.010539984330534935, + 0.008049088530242443, + 0.00932138878852129, + 0.02448383718729019, + 0.05698877573013306, + 0.00956596340984106, + 0.00745514826849103, + 0.021061699837446213, + 0.018780190497636795, + 0.013932882808148861, + 0.025549884885549545, + 0.016408981755375862, + 0.022599495947360992, + 0.013553022406995296, + 0.01870596595108509, + 0.01683012582361698, + 0.017429029569029808, + 0.0053070406429469585, + 0.006233914755284786, + 0.007294448092579842, + 0.008009612560272217, + 0.012496046721935272, + 0.006035545375198126, + 0.009056306444108486, + 0.008635441772639751, + 0.011343387886881828, + 0.017650248482823372, + 0.028739120811223984, + 0.020937694236636162, + 0.008562237955629826, + 0.011917399242520332, + 0.02201695367693901, + 0.016137000173330307, + 0.0059639872051775455, + 0.009387539699673653, + 0.007784872781485319, + 0.007466692477464676, + 0.007317639887332916, + 0.011542958207428455, + 0.0074172248132526875, + 0.010723412968218327, + 0.010586867108941078, + 0.021243924275040627, + 0.011019980534911156, + 0.009246832691133022, + 0.009266485460102558, + 0.008699444122612476, + 0.008447210304439068, + 0.011246968992054462, + 0.008320900611579418, + 0.010413923300802708, + 0.017366992309689522, + 0.013159501366317272, + 0.01615038327872753, + 0.01702849008142948, + 0.021263984963297844, + 0.01904589869081974, + 0.008860254660248756, + 0.008428629487752914, + 0.012872360646724701, + 0.01518231350928545, + 0.011191987432539463, + 0.01382097415626049, + 0.013681476004421711, + 0.01353463064879179, + 0.002257757354527712, + 0.0034467121586203575, + 0.004582637920975685, + 0.003710570279508829, + 0.0032808103132992983, + 0.005918669514358044, + 0.0037518679164350033, + 0.005164625588804483 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 2, + 4 + ], + "selected_reward": -0.14287103712558746 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot taken from a high-angle drone perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera tracking steadily alongside a herd of camels?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the camels descending down the steep face of a massive sand dune?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the camels carry vibrant textiles including crimson wool, indigo silk, and saffron-yellow cotton?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the scene feature long, deep blue shadows cast by an intense midday sun?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.576187252998352 + ], + "opposite_logprob_tokens": [ + -0.826187252998352 + ], + "logprob_reward": -0.576187252998352, + "yes_no_margin_reward": 0.25, + "token_credit_reward": -0.576187252998352, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11500728875398636, + 0.21171638369560242, + 0.3051826059818268, + 0.10260258615016937, + 0.034883081912994385, + 0.05442728102207184, + 0.03945675864815712, + 0.0684821605682373, + 0.04407801106572151, + 0.02416379190981388 + ], + "credit_map": [ + 0.011645995080471039, + 0.010590394958853722, + 0.02265247330069542, + 0.013300562277436256, + 0.009902006015181541, + 0.01707185059785843, + 0.011863485909998417, + 0.01798052340745926, + 0.009768128395080566, + 0.012962394393980503, + 0.050191719084978104, + 0.07762105762958527, + 0.009987237863242626, + 0.01044731680303812, + 0.019498756155371666, + 0.021239763125777245, + 0.020772945135831833, + 0.02897241897881031, + 0.03142959251999855, + 0.02551957592368126, + 0.03778383880853653, + 0.06716218590736389, + 0.03472196310758591, + 0.05882011353969574, + 0.008021730929613113, + 0.014564531855285168, + 0.017946697771549225, + 0.009154743514955044, + 0.022575750946998596, + 0.006786603946238756, + 0.013767015188932419, + 0.009785523638129234, + 0.003104614093899727, + 0.0048551433719694614, + 0.004502265714108944, + 0.005052670836448669, + 0.0037318039685487747, + 0.004076268058270216, + 0.00571069261059165, + 0.003849626751616597, + 0.003299793926998973, + 0.004712424706667662, + 0.006864546332508326, + 0.0045529017224907875, + 0.011336390860378742, + 0.006642729509621859, + 0.008222654461860657, + 0.008795843459665775, + 0.003068116959184408, + 0.004112840164452791, + 0.005410925485193729, + 0.004844951909035444, + 0.004243248142302036, + 0.0046927714720368385, + 0.005678495857864618, + 0.007405411917716265, + 0.006328292191028595, + 0.006544336676597595, + 0.01024842169135809, + 0.007539444603025913, + 0.005782227963209152, + 0.0072047594003379345, + 0.012060885317623615, + 0.012773795984685421, + 0.004565387032926083, + 0.003963287454098463, + 0.006002919282764196, + 0.007546038832515478, + 0.004461313597857952, + 0.004965787287801504, + 0.006194826681166887, + 0.006378456484526396, + 0.0017630907241255045, + 0.0021671466529369354, + 0.0038788928650319576, + 0.003741572145372629, + 0.0014637068379670382, + 0.0054438249208033085, + 0.0028518682811409235, + 0.002853691577911377 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 1, + 0 + ], + "selected_reward": -0.576187252998352 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot taken from a high-angle drone perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera tracking steadily alongside a herd of camels?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the camels descending down the steep face of a massive sand dune?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the camels carry vibrant textiles including crimson wool, indigo silk, and saffron-yellow cotton?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the scene feature long, deep blue shadows cast by an intense midday sun?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.25218790769577026 + ], + "opposite_logprob_tokens": [ + -1.502187967300415 + ], + "logprob_reward": -0.25218790769577026, + "yes_no_margin_reward": 1.25, + "token_credit_reward": -0.25218790769577026, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10002722591161728, + 0.15205004811286926, + 0.08843532204627991, + 0.11227203160524368, + 0.11693514883518219, + 0.17028583586215973, + 0.06105397641658783, + 0.0861525759100914, + 0.06347130984067917, + 0.04931648448109627 + ], + "credit_map": [ + 0.007797927595674992, + 0.012148886919021606, + 0.014041350223124027, + 0.016075629740953445, + 0.009233561344444752, + 0.012961634434759617, + 0.012843230739235878, + 0.014925002120435238, + 0.0076754107140004635, + 0.0084073506295681, + 0.02623511478304863, + 0.05240719020366669, + 0.00640076445415616, + 0.011421051807701588, + 0.01982712745666504, + 0.019676046445965767, + 0.007651939522475004, + 0.006860296707600355, + 0.012769228778779507, + 0.008097258396446705, + 0.00980443973094225, + 0.01781582087278366, + 0.008806848898530006, + 0.016629492864012718, + 0.010274315252900124, + 0.018823616206645966, + 0.013954241760075092, + 0.010824629105627537, + 0.013779092580080032, + 0.012521610595285892, + 0.012477305717766285, + 0.019617222249507904, + 0.010113045573234558, + 0.012621086090803146, + 0.01524022314697504, + 0.012882564216852188, + 0.01867527700960636, + 0.012638485990464687, + 0.015023174695670605, + 0.01974128931760788, + 0.011104055680334568, + 0.013664846308529377, + 0.022844864055514336, + 0.014370924793183804, + 0.04212023317813873, + 0.014196768403053284, + 0.030793197453022003, + 0.021190937608480453, + 0.0072907498106360435, + 0.016123773530125618, + 0.008663887158036232, + 0.007742701098322868, + 0.005778355989605188, + 0.004767834674566984, + 0.0056713419035077095, + 0.005015332717448473, + 0.008258884772658348, + 0.012029715813696384, + 0.011535796336829662, + 0.010242889635264874, + 0.010063459165394306, + 0.006380647420883179, + 0.0148634547367692, + 0.012777731753885746, + 0.006416773423552513, + 0.006200908683240414, + 0.008214853703975677, + 0.007457701489329338, + 0.006428638007491827, + 0.010036632418632507, + 0.008857129141688347, + 0.009858669713139534, + 0.004910788964480162, + 0.008525119163095951, + 0.007012423127889633, + 0.00731277372688055, + 0.0036346563138067722, + 0.006854048930108547, + 0.005169789772480726, + 0.0058968854136765 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 1, + 4 + ], + "selected_reward": -0.25218790769577026 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot taken from a high-angle drone perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera tracking steadily alongside a herd of camels?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the camels descending down the steep face of a massive sand dune?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the camels carry vibrant textiles including crimson wool, indigo silk, and saffron-yellow cotton?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the scene feature long, deep blue shadows cast by an intense midday sun?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.2813533842563629 + ], + "opposite_logprob_tokens": [ + -1.4063533544540405 + ], + "logprob_reward": -0.2813533842563629, + "yes_no_margin_reward": 1.125, + "token_credit_reward": -0.2813533842563629, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08534660935401917, + 0.06991144269704819, + 0.07111606746912003, + 0.1689169704914093, + 0.0800953358411789, + 0.08932854235172272, + 0.21549448370933533, + 0.15376482903957367, + 0.03627737984061241, + 0.029748359695076942 + ], + "credit_map": [ + 0.006971083115786314, + 0.011963406577706337, + 0.009182053618133068, + 0.010188414715230465, + 0.008683757856488228, + 0.009817817248404026, + 0.017620906233787537, + 0.010919160209596157, + 0.005785367917269468, + 0.00514192832633853, + 0.01101522333920002, + 0.008261932991445065, + 0.012344803661108017, + 0.008360618725419044, + 0.008236904628574848, + 0.010764667764306068, + 0.005917584057897329, + 0.006865502335131168, + 0.013537205755710602, + 0.009215828031301498, + 0.006697410251945257, + 0.011182439513504505, + 0.008719381876289845, + 0.008980710990726948, + 0.017516210675239563, + 0.030822109431028366, + 0.020801516249775887, + 0.015972718596458435, + 0.025469694286584854, + 0.01699964702129364, + 0.018339499831199646, + 0.02299557626247406, + 0.009407514706254005, + 0.012498200871050358, + 0.010036412626504898, + 0.007875379174947739, + 0.006875793915241957, + 0.011788892559707165, + 0.011002857238054276, + 0.01061027217656374, + 0.009092486463487148, + 0.012442145496606827, + 0.011179499328136444, + 0.010121566243469715, + 0.00873037613928318, + 0.012745805084705353, + 0.01057535968720913, + 0.014441302977502346, + 0.015289130620658398, + 0.013195729814469814, + 0.03841130807995796, + 0.01929416134953499, + 0.02164917066693306, + 0.025934848934412003, + 0.03358989581465721, + 0.04813023656606674, + 0.012989130802452564, + 0.021308545023202896, + 0.02191798761487007, + 0.02601807191967964, + 0.014268421567976475, + 0.013445037417113781, + 0.023398052901029587, + 0.020419558510184288, + 0.003562769154086709, + 0.004077861551195383, + 0.005198448896408081, + 0.004182394128292799, + 0.002944955136626959, + 0.0035343815106898546, + 0.0052252295427024364, + 0.007551335263997316, + 0.002004635985940695, + 0.00428367592394352, + 0.0045300815254449844, + 0.004331778734922409, + 0.00240904837846756, + 0.0034075547009706497, + 0.0035610119812190533, + 0.005220569204539061 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 3, + 7 + ], + "selected_reward": -0.2813533842563629 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "dramatic, high-angle drone shot tracks steadily alongside a herd of camels as they descend perilously down the steep face of a massive sand dune. The intense midday sun casts long, deep blue shadows that stretch behind the group and highlight the golden-orange textures of the sand. Each camel is laden with towering masses of vibrant textiles; rich crimson wool, deep indigo silk, and saffron-yellow cotton, woven in intricate geometric patterns, are securely strapped to their backs, creating vivid colors amidst the monochrome desert. The camels swoop down not in chaos but with a controlled, rhythmic gait, their powerful legs sinking deeply into the moving surface, raising soft plumes of sand that catch the light like a golden mist. The sheer scale of the sun-scorched dune dominates the frame, its crest forming a sharp line against the vast, cloudless azure sky, creating an atmosphere of epic struggle and eternal majesty.", + "seed": 14179, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000050/group_1/index_7573_seed_14179_dp_3_rank_6_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 1, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.2687161058187485, + "vlm_logprob": -0.2687161058187485, + "vlm_yes_no_margin": 1.249999988079071, + "vlm_token_credit": -0.2687161058187485, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot taken from a high-angle drone perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera tracking steadily alongside a herd of camels?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the camels descending down the steep face of a massive sand dune?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the camels carry vibrant textiles including crimson wool, indigo silk, and saffron-yellow cotton?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the scene feature long, deep blue shadows cast by an intense midday sun?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.20743142068386078 + ], + "opposite_logprob_tokens": [ + -1.832431435585022 + ], + "logprob_reward": -0.20743142068386078, + "yes_no_margin_reward": 1.625, + "token_credit_reward": -0.20743142068386078, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11903100460767746, + 0.12794412672519684, + 0.04731759428977966, + 0.05313307046890259, + 0.08771855384111404, + 0.08016514778137207, + 0.1655646413564682, + 0.0883411318063736, + 0.07105698436498642, + 0.1597277969121933 + ], + "credit_map": [ + 0.016345346346497536, + 0.014010973274707794, + 0.01534993015229702, + 0.017657829448580742, + 0.009055525064468384, + 0.007346032653003931, + 0.02416149154305458, + 0.015103872865438461, + 0.0166210625320673, + 0.009313543327152729, + 0.010983261279761791, + 0.02095913328230381, + 0.006297335494309664, + 0.005626894533634186, + 0.04444851726293564, + 0.013694372028112411, + 0.00687473826110363, + 0.004827302880585194, + 0.00874199066311121, + 0.005018054973334074, + 0.005003317724913359, + 0.00697374576702714, + 0.005081975366920233, + 0.004796468652784824, + 0.008445369079709053, + 0.005870691034942865, + 0.007658089976757765, + 0.005379027687013149, + 0.004185080993920565, + 0.005193841643631458, + 0.0070767090655863285, + 0.009324256330728531, + 0.010910485871136189, + 0.007585559040307999, + 0.008881098590791225, + 0.010296589694917202, + 0.003589947707951069, + 0.0047730873338878155, + 0.010098890401422977, + 0.03158288821578026, + 0.011057311668992043, + 0.015480861067771912, + 0.008535673841834068, + 0.013193102553486824, + 0.007199745159596205, + 0.00628423597663641, + 0.0072372122667729855, + 0.011176997795701027, + 0.017233433201909065, + 0.025690019130706787, + 0.018456080928444862, + 0.02275322750210762, + 0.016237888485193253, + 0.01890193484723568, + 0.024747539311647415, + 0.02154451236128807, + 0.007729179225862026, + 0.008764865808188915, + 0.008904374204576015, + 0.012095422483980656, + 0.01054136361926794, + 0.013053582981228828, + 0.013502335175871849, + 0.01374999899417162, + 0.011186701245605946, + 0.006170161068439484, + 0.014061512425541878, + 0.009910279884934425, + 0.004349006339907646, + 0.0072324383072555065, + 0.008137560449540615, + 0.010009318590164185, + 0.020121514797210693, + 0.016010750085115433, + 0.01949804089963436, + 0.020816119387745857, + 0.018192440271377563, + 0.024662699550390244, + 0.018662119284272194, + 0.021764088422060013 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 9, + 1 + ], + "selected_reward": -0.20743142068386078 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot taken from a high-angle drone perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera tracking steadily alongside a herd of camels?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the camels descending down the steep face of a massive sand dune?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the camels carry vibrant textiles including crimson wool, indigo silk, and saffron-yellow cotton?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the scene feature long, deep blue shadows cast by an intense midday sun?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.16044679284095764 + ], + "opposite_logprob_tokens": [ + -1.9104467630386353 + ], + "logprob_reward": -0.16044679284095764, + "yes_no_margin_reward": 1.75, + "token_credit_reward": -0.16044679284095764, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11870730668306351, + 0.1812620759010315, + 0.0477762296795845, + 0.10434477031230927, + 0.0445793941617012, + 0.03606494888663292, + 0.09056150168180466, + 0.13376986980438232, + 0.15271322429180145, + 0.09022068977355957 + ], + "credit_map": [ + 0.019033757969737053, + 0.020027825608849525, + 0.016721049323678017, + 0.012894674204289913, + 0.015180734917521477, + 0.011332276277244091, + 0.01242173183709383, + 0.011095255613327026, + 0.02323872596025467, + 0.00997710507363081, + 0.011236150749027729, + 0.03489658236503601, + 0.007225276902318001, + 0.004964321386069059, + 0.06952450424432755, + 0.020199405029416084, + 0.007487038150429726, + 0.0039295232854783535, + 0.00507329311221838, + 0.005641896743327379, + 0.006570345722138882, + 0.00754232844337821, + 0.005711918231099844, + 0.005819888319820166, + 0.016419177874922752, + 0.009656543843448162, + 0.012582780793309212, + 0.011220423504710197, + 0.0059354472905397415, + 0.009323148988187313, + 0.01912534423172474, + 0.020081892609596252, + 0.005321941804140806, + 0.0034764434676617384, + 0.005514299962669611, + 0.005042863544076681, + 0.0024635500740259886, + 0.005528281442821026, + 0.004551699385046959, + 0.012680313549935818, + 0.003283487632870674, + 0.003834577975794673, + 0.004650432150810957, + 0.004691867623478174, + 0.0037150951102375984, + 0.00520357908681035, + 0.005353092215955257, + 0.005332818254828453, + 0.009688490070402622, + 0.011402235366404057, + 0.01805960014462471, + 0.012553573586046696, + 0.008152465336024761, + 0.010783355683088303, + 0.010275444947183132, + 0.009646330028772354, + 0.011080309748649597, + 0.009936120361089706, + 0.014500205405056477, + 0.013521222397685051, + 0.014117917977273464, + 0.028282513841986656, + 0.019773276522755623, + 0.02255830727517605, + 0.016952110454440117, + 0.01824512518942356, + 0.03657723218202591, + 0.022300349548459053, + 0.010953988879919052, + 0.015758652240037918, + 0.017655113711953163, + 0.014270653016865253, + 0.011121517047286034, + 0.011999730952084064, + 0.012358981184661388, + 0.012423504143953323, + 0.00918647926300764, + 0.010261964984238148, + 0.01016306784003973, + 0.012705446220934391 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 8, + 7 + ], + "selected_reward": -0.16044679284095764 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot taken from a high-angle drone perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera tracking steadily alongside a herd of camels?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the camels descending down the steep face of a massive sand dune?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the camels carry vibrant textiles including crimson wool, indigo silk, and saffron-yellow cotton?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the scene feature long, deep blue shadows cast by an intense midday sun?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.348667174577713 + ], + "opposite_logprob_tokens": [ + -1.2236671447753906 + ], + "logprob_reward": -0.348667174577713, + "yes_no_margin_reward": 0.875, + "token_credit_reward": -0.348667174577713, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14314201474189758, + 0.137080118060112, + 0.031803157180547714, + 0.09859739243984222, + 0.1192551702260971, + 0.03515711799263954, + 0.2034355103969574, + 0.06287316977977753, + 0.09873586148023605, + 0.06992042809724808 + ], + "credit_map": [ + 0.023977408185601234, + 0.02344677597284317, + 0.016673021018505096, + 0.02071075700223446, + 0.017878837883472443, + 0.012310893274843693, + 0.015478119254112244, + 0.012666217982769012, + 0.016007566824555397, + 0.008397798053920269, + 0.010714150033891201, + 0.021898509934544563, + 0.00842985324561596, + 0.011368966661393642, + 0.04463809356093407, + 0.015625186264514923, + 0.0047881449572741985, + 0.0032820170745253563, + 0.005726700648665428, + 0.0027366122230887413, + 0.0038969607558101416, + 0.0029202483128756285, + 0.004508385434746742, + 0.003944093827158213, + 0.015362090431153774, + 0.008413859643042088, + 0.009544259868562222, + 0.009728580713272095, + 0.007401542738080025, + 0.01013870257884264, + 0.020818648859858513, + 0.017189715057611465, + 0.015321310609579086, + 0.01017740834504366, + 0.012487757951021194, + 0.012356694787740707, + 0.0064791059121489525, + 0.008007857948541641, + 0.012962576933205128, + 0.04146246984601021, + 0.004711626563221216, + 0.004805568140000105, + 0.005152632016688585, + 0.005312490276992321, + 0.00314965914003551, + 0.00411541573703289, + 0.0032549449242651463, + 0.004654784686863422, + 0.01870162971317768, + 0.02438010834157467, + 0.030714502558112144, + 0.029744789004325867, + 0.017017370089888573, + 0.022164927795529366, + 0.030046453699469566, + 0.03066573292016983, + 0.005090351682156324, + 0.004884418100118637, + 0.007325778715312481, + 0.007853797636926174, + 0.006943020969629288, + 0.010978557169437408, + 0.00878080539405346, + 0.01101644802838564, + 0.012708651833236217, + 0.012598099187016487, + 0.01778719574213028, + 0.011409076862037182, + 0.00853000022470951, + 0.013032576069235802, + 0.010867267847061157, + 0.011803006753325462, + 0.007784449495375156, + 0.007821274921298027, + 0.006347282789647579, + 0.0074922265484929085, + 0.007885618135333061, + 0.012961991131305695, + 0.00934393797069788, + 0.01028365083038807 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 0, + 1 + ], + "selected_reward": -0.348667174577713 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot taken from a high-angle drone perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera tracking steadily alongside a herd of camels?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the camels descending down the steep face of a massive sand dune?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the camels carry vibrant textiles including crimson wool, indigo silk, and saffron-yellow cotton?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the scene feature long, deep blue shadows cast by an intense midday sun?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.3135598301887512 + ], + "opposite_logprob_tokens": [ + -1.3135597705841064 + ], + "logprob_reward": -0.3135598301887512, + "yes_no_margin_reward": 0.9999999403953552, + "token_credit_reward": -0.3135598301887512, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1572173833847046, + 0.10289418697357178, + 0.06723590195178986, + 0.07207965850830078, + 0.07557035237550735, + 0.06687493622303009, + 0.11304310709238052, + 0.12388283759355545, + 0.07110261172056198, + 0.1500989943742752 + ], + "credit_map": [ + 0.026410164311528206, + 0.028003813698887825, + 0.01581723429262638, + 0.017766976729035378, + 0.02166205458343029, + 0.01604732684791088, + 0.012822790071368217, + 0.018687020987272263, + 0.013038909994065762, + 0.006035217083990574, + 0.008825958706438541, + 0.01830708421766758, + 0.004786871373653412, + 0.006654059048742056, + 0.03394169360399246, + 0.011304396204650402, + 0.009357220493257046, + 0.007557761389762163, + 0.01002941932529211, + 0.006508791819214821, + 0.010201052762567997, + 0.007598914671689272, + 0.00876506045460701, + 0.007217680104076862, + 0.008936303667724133, + 0.005624642129987478, + 0.008954600431025028, + 0.0087202787399292, + 0.008277509361505508, + 0.011150565929710865, + 0.009941188618540764, + 0.010474571026861668, + 0.011594031937420368, + 0.006431903690099716, + 0.008262334391474724, + 0.006934578064829111, + 0.007216907572001219, + 0.019178474321961403, + 0.007230805698782206, + 0.008721313439309597, + 0.009402821771800518, + 0.008641167543828487, + 0.0061386036686599255, + 0.009253452531993389, + 0.006005267612636089, + 0.013291766867041588, + 0.005268994253128767, + 0.0088728629052639, + 0.01113122422248125, + 0.014994649216532707, + 0.01587497442960739, + 0.01770719885826111, + 0.009302031248807907, + 0.010972708463668823, + 0.01200859621167183, + 0.021051717922091484, + 0.010005226358771324, + 0.011296485550701618, + 0.01352638192474842, + 0.017545504495501518, + 0.014070207253098488, + 0.020802684128284454, + 0.016752570867538452, + 0.019883764907717705, + 0.007770185358822346, + 0.008230477571487427, + 0.010421647690236568, + 0.009315622970461845, + 0.008663994260132313, + 0.008657379075884819, + 0.008554188534617424, + 0.009489115327596664, + 0.01502370648086071, + 0.017401328310370445, + 0.012452739290893078, + 0.017862781882286072, + 0.024503352120518684, + 0.03280920907855034, + 0.012618034146726131, + 0.0174278412014246 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 7 + ], + "selected_reward": -0.3135598301887512 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot taken from a high-angle drone perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera tracking steadily alongside a herd of camels?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the camels descending down the steep face of a massive sand dune?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the camels carry vibrant textiles including crimson wool, indigo silk, and saffron-yellow cotton?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the scene feature long, deep blue shadows cast by an intense midday sun?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.3134753108024597 + ], + "opposite_logprob_tokens": [ + -1.3134753704071045 + ], + "logprob_reward": -0.3134753108024597, + "yes_no_margin_reward": 1.0, + "token_credit_reward": -0.3134753108024597, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.30361607670783997, + 0.1598622351884842, + 0.03051106072962284, + 0.039182763546705246, + 0.05422971397638321, + 0.0588509775698185, + 0.03101629577577114, + 0.045224227011203766, + 0.09886037558317184, + 0.17864619195461273 + ], + "credit_map": [ + 0.04350954294204712, + 0.04883745312690735, + 0.030961908400058746, + 0.03586135432124138, + 0.042761947959661484, + 0.025697214528918266, + 0.042147815227508545, + 0.03383887559175491, + 0.022798685356974602, + 0.022934608161449432, + 0.015237336978316307, + 0.02426951192319393, + 0.01178772747516632, + 0.014528054744005203, + 0.03310602903366089, + 0.015200302004814148, + 0.0038070445880293846, + 0.0027941642329096794, + 0.005042462144047022, + 0.0037321278359740973, + 0.003920581191778183, + 0.003488839603960514, + 0.004476551897823811, + 0.003249292029067874, + 0.006887019611895084, + 0.004886799957603216, + 0.004494878463447094, + 0.004157509654760361, + 0.0028412325773388147, + 0.0046140896156430244, + 0.005354898516088724, + 0.005946339573711157, + 0.0071526686660945415, + 0.003330113599076867, + 0.006132051348686218, + 0.004868028219789267, + 0.003856039373204112, + 0.006346044596284628, + 0.0058542764745652676, + 0.01669050008058548, + 0.007750377058982849, + 0.010993335396051407, + 0.006570232100784779, + 0.01146222185343504, + 0.0039214021526277065, + 0.003662058850750327, + 0.006547030992805958, + 0.007944325916469097, + 0.003088030032813549, + 0.003579930867999792, + 0.0047263773158192635, + 0.004860870540142059, + 0.0033135900739580393, + 0.00397372804582119, + 0.0038331360556185246, + 0.0036406368017196655, + 0.003952265251427889, + 0.006144307553768158, + 0.005844487342983484, + 0.006136677693575621, + 0.00518053350970149, + 0.003113821381703019, + 0.00859329104423523, + 0.006258849520236254, + 0.015371615998446941, + 0.008358817547559738, + 0.021008672192692757, + 0.012770434841513634, + 0.007645360194146633, + 0.008952125906944275, + 0.010326615534722805, + 0.014426745474338531, + 0.016878187656402588, + 0.01815141923725605, + 0.013384120538830757, + 0.018849410116672516, + 0.030930539593100548, + 0.03946759179234505, + 0.018170766532421112, + 0.022814176976680756 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 1 + ], + "selected_reward": -0.3134753108024597 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000050/group_10/index_tensor([8252, 8252])_group_10_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000050/group_10/index_tensor([8252, 8252])_group_10_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..15df57c97ffa7b3ebd9f54edceca2a5d6e954956 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000050/group_10/index_tensor([8252, 8252])_group_10_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "close-up shot from a slightly overhead angle, slowly but steadily panning forward, focuses on a three-dimensional holographic weather map suspended in a dimly lit futuristic control room. Within this holographic projection, countless slender, brilliant blue and white energy data streams intertwine and flow like a complex neural network. Between them, vast, irregular crimson danger zones pulse ominously, emitting a sinister shimmer. As the camera approaches, as if forcibly pulled by an invisible force, all the glowing energy lines and crimson zones begin to spiral and contract, rapidly and irreversibly converging towards the center, ultimately converging into a singular point of immense density and blinding light. At this moment of convergence, a dazzling, rainbow-colored anamorphic lens flare, bursting forth from the concentrated energy, pierces the screen, bathing the entire scene in Makoto Shinkai's signature aesthetic light. The light penetrates even the finest particles suspended in the air, creating an atmosphere both majestic and unsettling, culminating in a visual climax of catastrophic aesthetics.", + "seed": 27571, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000050/group_10/index_8252_seed_27571_dp_20_rank_40_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 10, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.08999795808922499, + "vlm_logprob": -0.08999795808922499, + "vlm_yes_no_margin": 3.875000023841858, + "vlm_token_credit": -0.08999795808922499, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a dimly lit futuristic control room featuring a three-dimensional holographic weather map?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera start with a close-up from a slightly overhead angle and slowly pan forward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do slender blue and white energy data streams intertwine with vast, irregular crimson danger zones that pulse ominously?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the energy lines and crimson zones spiral and contract towards a central point of blinding light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does a rainbow-colored anamorphic lens flare burst from the convergence, creating Makoto Shinkai's signature aesthetic light?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.15097129344940186 + ], + "opposite_logprob_tokens": [ + -2.0259714126586914 + ], + "logprob_reward": -0.15097129344940186, + "yes_no_margin_reward": 1.8750001192092896, + "token_credit_reward": -0.15097129344940186, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.43187737464904785, + 0.05269015580415726, + 0.125345379114151, + 0.056558866053819656, + 0.03477003052830696, + 0.04279521107673645, + 0.012432445771992207, + 0.11293681710958481, + 0.0186211746186018, + 0.11197248101234436 + ], + "credit_map": [ + 0.06818831712007523, + 0.06504976749420166, + 0.047832027077674866, + 0.046532344073057175, + 0.08579949289560318, + 0.04551522806286812, + 0.03574621304869652, + 0.0372140035033226, + 0.00855889730155468, + 0.00460538687184453, + 0.008091584779322147, + 0.00443249149248004, + 0.004862108267843723, + 0.008228575810790062, + 0.003312391461804509, + 0.010598726570606232, + 0.01902136020362377, + 0.005566562060266733, + 0.019584788009524345, + 0.007465047296136618, + 0.010231602005660534, + 0.03441806137561798, + 0.007110336795449257, + 0.02194763533771038, + 0.009137291461229324, + 0.002423429163172841, + 0.00907022412866354, + 0.005156295839697123, + 0.004324325360357761, + 0.014268477447330952, + 0.0030921553261578083, + 0.009086676873266697, + 0.006353417877107859, + 0.00235486077144742, + 0.006271183490753174, + 0.002071202965453267, + 0.002622251631692052, + 0.007519051432609558, + 0.0016697306418791413, + 0.005908336956053972, + 0.005591188557446003, + 0.002827996853739023, + 0.006371625699102879, + 0.004153157584369183, + 0.0034783179871737957, + 0.00974374357610941, + 0.004029113333672285, + 0.006600070279091597, + 0.0015348636079579592, + 0.0010113050229847431, + 0.0017533822683617473, + 0.0013352397363632917, + 0.001302898395806551, + 0.0026069371961057186, + 0.0016022641211748123, + 0.001285556354559958, + 0.01206931658089161, + 0.03056614100933075, + 0.004164029378443956, + 0.011312956921756268, + 0.006682155653834343, + 0.006652699317783117, + 0.034939851611852646, + 0.00654968386515975, + 0.002166221383959055, + 0.0019306415924802423, + 0.0033063204027712345, + 0.002515939064323902, + 0.0019957076292485, + 0.0032084151171147823, + 0.001693696016445756, + 0.0018042338779196143, + 0.005520143546164036, + 0.0064114476554095745, + 0.0071018049493432045, + 0.03907265514135361, + 0.010817019268870354, + 0.00577507121488452, + 0.030674763023853302, + 0.006599589716643095 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 7 + ], + "selected_reward": -0.15097129344940186 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a dimly lit futuristic control room featuring a three-dimensional holographic weather map?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera start with a close-up from a slightly overhead angle and slowly pan forward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do slender blue and white energy data streams intertwine with vast, irregular crimson danger zones that pulse ominously?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the energy lines and crimson zones spiral and contract towards a central point of blinding light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does a rainbow-colored anamorphic lens flare burst from the convergence, creating Makoto Shinkai's signature aesthetic light?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.2813376486301422 + ], + "opposite_logprob_tokens": [ + -1.4063376188278198 + ], + "logprob_reward": -0.2813376486301422, + "yes_no_margin_reward": 1.125, + "token_credit_reward": -0.2813376486301422, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3503982722759247, + 0.06647970527410507, + 0.08729878813028336, + 0.029255559667944908, + 0.027343647554516792, + 0.041253652423620224, + 0.028337273746728897, + 0.2840510308742523, + 0.01276350673288107, + 0.0728185698390007 + ], + "credit_map": [ + 0.04936186596751213, + 0.06363077461719513, + 0.026156779378652573, + 0.03361310064792633, + 0.08351205289363861, + 0.037685658782720566, + 0.03134315833449364, + 0.0250948928296566, + 0.012791831977665424, + 0.0036524443421512842, + 0.016152357682585716, + 0.006507982965558767, + 0.003720252774655819, + 0.010596566833555698, + 0.004473282024264336, + 0.00858498178422451, + 0.018614768981933594, + 0.004041410982608795, + 0.01783403940498829, + 0.005311590153723955, + 0.004875709768384695, + 0.01816510781645775, + 0.004212256986647844, + 0.014243905432522297, + 0.004240577574819326, + 0.0026840618811547756, + 0.004981968551874161, + 0.0035838340409100056, + 0.002213843632489443, + 0.005023552570492029, + 0.002097143791615963, + 0.004430579021573067, + 0.004566836170852184, + 0.00312623823992908, + 0.004407849628478289, + 0.003640831680968404, + 0.002386139938607812, + 0.003944450989365578, + 0.0024477182887494564, + 0.0028235826175659895, + 0.006141853053122759, + 0.003780824365094304, + 0.003899060422554612, + 0.004961457569152117, + 0.003066645236685872, + 0.008831506595015526, + 0.005173292942345142, + 0.00539901340380311, + 0.003523818217217922, + 0.0023201564326882362, + 0.0029720973689109087, + 0.003342192154377699, + 0.0038219348061829805, + 0.0041804504580795765, + 0.005735544487833977, + 0.0024410798214375973, + 0.032162636518478394, + 0.08060972392559052, + 0.010115709155797958, + 0.02667456865310669, + 0.017011374235153198, + 0.018181271851062775, + 0.08322076499462128, + 0.016074981540441513, + 0.0013154020998626947, + 0.0012966827489435673, + 0.0015829256735742092, + 0.0021331380121409893, + 0.0015704275574535131, + 0.0016290561761707067, + 0.0016699327388778329, + 0.001565941609442234, + 0.004841234534978867, + 0.0070172580890357494, + 0.007825174368917942, + 0.01333677675575018, + 0.014269726350903511, + 0.006967899855226278, + 0.011948988772928715, + 0.006611511576920748 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 7, + 2 + ], + "selected_reward": -0.2813376486301422 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a dimly lit futuristic control room featuring a three-dimensional holographic weather map?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera start with a close-up from a slightly overhead angle and slowly pan forward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do slender blue and white energy data streams intertwine with vast, irregular crimson danger zones that pulse ominously?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the energy lines and crimson zones spiral and contract towards a central point of blinding light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does a rainbow-colored anamorphic lens flare burst from the convergence, creating Makoto Shinkai's signature aesthetic light?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0019471748964861035 + ], + "opposite_logprob_tokens": [ + -6.251947402954102 + ], + "logprob_reward": -0.0019471748964861035, + "yes_no_margin_reward": 6.25, + "token_credit_reward": -0.0019471748964861035, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.17634159326553345, + 0.10968919843435287, + 0.17722168564796448, + 0.052433449774980545, + 0.06030001491308212, + 0.03599785640835762, + 0.02544410340487957, + 0.24449966847896576, + 0.016320474445819855, + 0.1017519161105156 + ], + "credit_map": [ + 0.030360864475369453, + 0.016709186136722565, + 0.0287562757730484, + 0.017271392047405243, + 0.027907272800803185, + 0.019483789801597595, + 0.01435311883687973, + 0.021499712020158768, + 0.016335168853402138, + 0.005241722799837589, + 0.015258225612342358, + 0.009827114641666412, + 0.00802314467728138, + 0.022708961740136147, + 0.006467111874371767, + 0.02582775615155697, + 0.027623891830444336, + 0.008759379386901855, + 0.02069964073598385, + 0.011315139941871166, + 0.013401772826910019, + 0.04987403377890587, + 0.009716245345771313, + 0.035831596702337265, + 0.008448800072073936, + 0.0027576095890253782, + 0.008664214052259922, + 0.005125841125845909, + 0.0039501432329416275, + 0.008796798065304756, + 0.0033021848648786545, + 0.01138786319643259, + 0.010274909436702728, + 0.005786487367004156, + 0.0119788171723485, + 0.004916398320347071, + 0.0038036752957850695, + 0.011441451497375965, + 0.003309931606054306, + 0.00878834817558527, + 0.006697263102978468, + 0.0021095105912536383, + 0.004751431755721569, + 0.00276533723808825, + 0.0030751328449696302, + 0.007480583153665066, + 0.0029466687701642513, + 0.006171932443976402, + 0.002477051690220833, + 0.0019529229030013084, + 0.0033554998226463795, + 0.0027540300507098436, + 0.0030511433724313974, + 0.006207965314388275, + 0.003509629052132368, + 0.0021358614321798086, + 0.0339561402797699, + 0.08259475976228714, + 0.007094957400113344, + 0.016647132113575935, + 0.014268367551267147, + 0.008658962324261665, + 0.06789170205593109, + 0.013387659564614296, + 0.0017826127586886287, + 0.0020174868404865265, + 0.0019646245054900646, + 0.0025813088286668062, + 0.001600607531145215, + 0.002272567944601178, + 0.0022004249040037394, + 0.0019008418312296271, + 0.005509179085493088, + 0.007910731248557568, + 0.006324287503957748, + 0.036028582602739334, + 0.0093134930357337, + 0.00571473641321063, + 0.025626476854085922, + 0.005324434954673052 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 2, + 0 + ], + "selected_reward": -0.0019471748964861035 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a dimly lit futuristic control room featuring a three-dimensional holographic weather map?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera start with a close-up from a slightly overhead angle and slowly pan forward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do slender blue and white energy data streams intertwine with vast, irregular crimson danger zones that pulse ominously?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the energy lines and crimson zones spiral and contract towards a central point of blinding light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does a rainbow-colored anamorphic lens flare burst from the convergence, creating Makoto Shinkai's signature aesthetic light?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0031907858792692423 + ], + "opposite_logprob_tokens": [ + -5.753190994262695 + ], + "logprob_reward": -0.0031907858792692423, + "yes_no_margin_reward": 5.75, + "token_credit_reward": -0.0031907858792692423, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2107498198747635, + 0.14472676813602448, + 0.1030317023396492, + 0.06159330531954765, + 0.033604659140110016, + 0.036910027265548706, + 0.05005457252264023, + 0.22488877177238464, + 0.017226433381438255, + 0.11721395701169968 + ], + "credit_map": [ + 0.031169123947620392, + 0.035970285534858704, + 0.01777489297091961, + 0.022958068177103996, + 0.04434957355260849, + 0.024605732411146164, + 0.017645472660660744, + 0.01627667248249054, + 0.022720810025930405, + 0.005523852538317442, + 0.02030789665877819, + 0.009219779632985592, + 0.010775457136332989, + 0.031507521867752075, + 0.008359335362911224, + 0.03631210699677467, + 0.02544296346604824, + 0.006520065013319254, + 0.03024139069020748, + 0.008901298977434635, + 0.007336333394050598, + 0.00970784667879343, + 0.005239250138401985, + 0.00964254792779684, + 0.008366499096155167, + 0.006947781890630722, + 0.009276539087295532, + 0.00842337403446436, + 0.00586454663425684, + 0.007640313822776079, + 0.0048744454979896545, + 0.01019980851560831, + 0.004990092944353819, + 0.004602306988090277, + 0.004008702468127012, + 0.0057334816083312035, + 0.0034691093023866415, + 0.0034499666653573513, + 0.0038558251690119505, + 0.0034951737616211176, + 0.00868512224406004, + 0.002480827970430255, + 0.005089464597404003, + 0.0032520340755581856, + 0.003324837191030383, + 0.00433823186904192, + 0.003378341207280755, + 0.0063611664809286594, + 0.005281048361212015, + 0.004257719498127699, + 0.005744887515902519, + 0.006775645073503256, + 0.0062681385315954685, + 0.010058981366455555, + 0.006830016151070595, + 0.004838138353079557, + 0.029719870537519455, + 0.0690225139260292, + 0.010338397696614265, + 0.019325127825140953, + 0.014222916215658188, + 0.013461818918585777, + 0.05694245547056198, + 0.011855662800371647, + 0.0018452940275892615, + 0.0018273151945322752, + 0.0021143117919564247, + 0.003262116340920329, + 0.001895730383694172, + 0.002532124985009432, + 0.0018003275617957115, + 0.0019492123974487185, + 0.014196566306054592, + 0.01329581905156374, + 0.010385232046246529, + 0.027004467323422432, + 0.014261133037507534, + 0.008601490408182144, + 0.019335197284817696, + 0.010134045965969563 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 0, + 1 + ], + "selected_reward": -0.0031907858792692423 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a dimly lit futuristic control room featuring a three-dimensional holographic weather map?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera start with a close-up from a slightly overhead angle and slowly pan forward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do slender blue and white energy data streams intertwine with vast, irregular crimson danger zones that pulse ominously?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the energy lines and crimson zones spiral and contract towards a central point of blinding light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does a rainbow-colored anamorphic lens flare burst from the convergence, creating Makoto Shinkai's signature aesthetic light?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.012542887590825558 + ], + "opposite_logprob_tokens": [ + -4.387542724609375 + ], + "logprob_reward": -0.012542887590825558, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.012542887590825558, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15178023278713226, + 0.03685612976551056, + 0.07099378108978271, + 0.035206686705350876, + 0.030549755319952965, + 0.023088544607162476, + 0.018524471670389175, + 0.30678319931030273, + 0.014578460715711117, + 0.31163880228996277 + ], + "credit_map": [ + 0.02364106848835945, + 0.022259747609496117, + 0.01801237463951111, + 0.015517442487180233, + 0.03186829015612602, + 0.014686433598399162, + 0.01271137222647667, + 0.013083474710583687, + 0.006087963003665209, + 0.003252114402130246, + 0.005911371670663357, + 0.0033769363071769476, + 0.003382850904017687, + 0.005279603414237499, + 0.0028855029959231615, + 0.006679781712591648, + 0.011949536390602589, + 0.003316525835543871, + 0.013146767392754555, + 0.0049143098294734955, + 0.005524147301912308, + 0.01682906784117222, + 0.003575908951461315, + 0.01173750963062048, + 0.004652179311960936, + 0.002908730413764715, + 0.006098721642047167, + 0.0036965759936720133, + 0.0033871433697640896, + 0.006948322989046574, + 0.0021965147461742163, + 0.005318493582308292, + 0.004898533690720797, + 0.003269846085458994, + 0.005313807167112827, + 0.004046279471367598, + 0.0027685314416885376, + 0.003950261510908604, + 0.0023784609511494637, + 0.003924032207578421, + 0.005189590621739626, + 0.0024600671604275703, + 0.003980157431215048, + 0.0019964154344052076, + 0.0017979411641135812, + 0.002558075822889805, + 0.0022731164935976267, + 0.002833176637068391, + 0.002181330695748329, + 0.0019466083031147718, + 0.002356533659622073, + 0.002259582746773958, + 0.0018186415545642376, + 0.0021268846467137337, + 0.004149682354182005, + 0.0016852066619321704, + 0.0471041239798069, + 0.10629434138536453, + 0.009420543909072876, + 0.02079351246356964, + 0.017552748322486877, + 0.013985734432935715, + 0.07586804032325745, + 0.015764130279421806, + 0.0012659996282309294, + 0.001418829197064042, + 0.001828285283409059, + 0.0020334823057055473, + 0.0021017412655055523, + 0.002740327036008239, + 0.0015098806470632553, + 0.0016799136064946651, + 0.021582020446658134, + 0.017054783180356026, + 0.027005981653928757, + 0.09531038254499435, + 0.02621498331427574, + 0.01630767434835434, + 0.08406011760234833, + 0.02410285733640194 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 7, + 0 + ], + "selected_reward": -0.012542887590825558 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "close-up shot from a slightly overhead angle, slowly but steadily panning forward, focuses on a three-dimensional holographic weather map suspended in a dimly lit futuristic control room. Within this holographic projection, countless slender, brilliant blue and white energy data streams intertwine and flow like a complex neural network. Between them, vast, irregular crimson danger zones pulse ominously, emitting a sinister shimmer. As the camera approaches, as if forcibly pulled by an invisible force, all the glowing energy lines and crimson zones begin to spiral and contract, rapidly and irreversibly converging towards the center, ultimately converging into a singular point of immense density and blinding light. At this moment of convergence, a dazzling, rainbow-colored anamorphic lens flare, bursting forth from the concentrated energy, pierces the screen, bathing the entire scene in Makoto Shinkai's signature aesthetic light. The light penetrates even the finest particles suspended in the air, creating an atmosphere both majestic and unsettling, culminating in a visual climax of catastrophic aesthetics.", + "seed": 27572, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000050/group_10/index_8252_seed_27572_dp_20_rank_40_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 10, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.051502558100037274, + "vlm_logprob": -0.051502558100037274, + "vlm_yes_no_margin": 4.125, + "vlm_token_credit": -0.051502558100037274, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a dimly lit futuristic control room featuring a three-dimensional holographic weather map?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera start with a close-up from a slightly overhead angle and slowly pan forward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do slender blue and white energy data streams intertwine with vast, irregular crimson danger zones that pulse ominously?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the energy lines and crimson zones spiral and contract towards a central point of blinding light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does a rainbow-colored anamorphic lens flare burst from the convergence, creating Makoto Shinkai's signature aesthetic light?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0941036194562912 + ], + "opposite_logprob_tokens": [ + -2.4691035747528076 + ], + "logprob_reward": -0.0941036194562912, + "yes_no_margin_reward": 2.375, + "token_credit_reward": -0.0941036194562912, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.18046124279499054, + 0.1494656801223755, + 0.021141696721315384, + 0.05541517958045006, + 0.223783940076828, + 0.07500244677066803, + 0.0413241982460022, + 0.04241742566227913, + 0.10690166801214218, + 0.10408651828765869 + ], + "credit_map": [ + 0.020025502890348434, + 0.02724221721291542, + 0.02234007976949215, + 0.03099764697253704, + 0.023006921634078026, + 0.017175396904349327, + 0.02481820620596409, + 0.014855281449854374, + 0.014291143976151943, + 0.01593434065580368, + 0.02702634036540985, + 0.03274916857481003, + 0.010513383895158768, + 0.017224401235580444, + 0.015930382534861565, + 0.01579652912914753, + 0.002272011712193489, + 0.0024713617749512196, + 0.0027189331594854593, + 0.003152577206492424, + 0.0021634779404848814, + 0.003114188788458705, + 0.0026299678720533848, + 0.002619178267195821, + 0.005279757548123598, + 0.0038163908757269382, + 0.007594444323331118, + 0.008133531548082829, + 0.004325229674577713, + 0.008500737138092518, + 0.01082625426352024, + 0.006938837002962828, + 0.014076999388635159, + 0.009363992139697075, + 0.07886708527803421, + 0.08908455818891525, + 0.0035517150536179543, + 0.009702758863568306, + 0.014047152362763882, + 0.005089688580483198, + 0.0059739151038229465, + 0.004253873135894537, + 0.020851492881774902, + 0.009397490881383419, + 0.0036301668733358383, + 0.005380850285291672, + 0.021148722618818283, + 0.004365930333733559, + 0.0054235137067735195, + 0.005809633061289787, + 0.005834186915308237, + 0.005292451940476894, + 0.005099799949675798, + 0.003561317455023527, + 0.005785307846963406, + 0.004517987370491028, + 0.004349707625806332, + 0.0045642307959496975, + 0.003671567188575864, + 0.0061751496978104115, + 0.006252630613744259, + 0.00563465990126133, + 0.005948339123278856, + 0.005821141414344311, + 0.01213060412555933, + 0.009408914484083652, + 0.007192838937044144, + 0.016537697985768318, + 0.01744125597178936, + 0.012448376975953579, + 0.01753087341785431, + 0.01421110425144434, + 0.0128529854118824, + 0.021896950900554657, + 0.008392888121306896, + 0.011071842163801193, + 0.01799400895833969, + 0.012698019854724407, + 0.010218084789812565, + 0.008961735293269157 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 0, + 1 + ], + "selected_reward": -0.0941036194562912 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a dimly lit futuristic control room featuring a three-dimensional holographic weather map?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera start with a close-up from a slightly overhead angle and slowly pan forward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do slender blue and white energy data streams intertwine with vast, irregular crimson danger zones that pulse ominously?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the energy lines and crimson zones spiral and contract towards a central point of blinding light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does a rainbow-colored anamorphic lens flare burst from the convergence, creating Makoto Shinkai's signature aesthetic light?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.1428024172782898 + ], + "opposite_logprob_tokens": [ + -2.0178024768829346 + ], + "logprob_reward": -0.1428024172782898, + "yes_no_margin_reward": 1.875, + "token_credit_reward": -0.1428024172782898, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1493380218744278, + 0.0632210522890091, + 0.025802578777074814, + 0.03464413806796074, + 0.12666121125221252, + 0.12261611223220825, + 0.25393912196159363, + 0.05402492359280586, + 0.11285651475191116, + 0.056896332651376724 + ], + "credit_map": [ + 0.01546652615070343, + 0.02091452293097973, + 0.0158842820674181, + 0.01979869417846203, + 0.0158903356641531, + 0.031711645424366, + 0.013511902652680874, + 0.016160106286406517, + 0.006605240516364574, + 0.006591536570340395, + 0.01005073543637991, + 0.01061847060918808, + 0.006180279888212681, + 0.010455884039402008, + 0.005944882519543171, + 0.006774024106562138, + 0.004560538101941347, + 0.0027483515441417694, + 0.0031701019033789635, + 0.004307587631046772, + 0.002533749910071492, + 0.0024236368481069803, + 0.002716118935495615, + 0.0033424929715692997, + 0.004077025689184666, + 0.003682235023006797, + 0.007913683541119099, + 0.005168033763766289, + 0.00303895166143775, + 0.003388312179595232, + 0.004112510941922665, + 0.00326338573358953, + 0.008325707167387009, + 0.005363539792597294, + 0.04130832850933075, + 0.049199339002370834, + 0.003937427420169115, + 0.005404795054346323, + 0.009300688281655312, + 0.0038213664665818214, + 0.011565716937184334, + 0.010280942544341087, + 0.027581797912716866, + 0.021666662767529488, + 0.009267517365515232, + 0.010493914596736431, + 0.02574092335999012, + 0.006018634885549545, + 0.02997429296374321, + 0.019852403551340103, + 0.047463033348321915, + 0.03692455589771271, + 0.025778386741876602, + 0.018334360793232918, + 0.04042843356728554, + 0.03518364951014519, + 0.005648269318044186, + 0.0047865635715425014, + 0.005077243782579899, + 0.008707167580723763, + 0.005443526431918144, + 0.005337913520634174, + 0.009232869371771812, + 0.00979137048125267, + 0.011997579596936703, + 0.008556519635021687, + 0.007048840168863535, + 0.014975209720432758, + 0.019823811948299408, + 0.01266240794211626, + 0.02173452638089657, + 0.01605762168765068, + 0.0052120923064649105, + 0.0067282854579389095, + 0.005212921183556318, + 0.010578793473541737, + 0.007041275035589933, + 0.006843259558081627, + 0.006734853144735098, + 0.00854485109448433 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 0, + 4 + ], + "selected_reward": -0.1428024172782898 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a dimly lit futuristic control room featuring a three-dimensional holographic weather map?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera start with a close-up from a slightly overhead angle and slowly pan forward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do slender blue and white energy data streams intertwine with vast, irregular crimson danger zones that pulse ominously?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the energy lines and crimson zones spiral and contract towards a central point of blinding light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does a rainbow-colored anamorphic lens flare burst from the convergence, creating Makoto Shinkai's signature aesthetic light?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.001718236249871552 + ], + "opposite_logprob_tokens": [ + -6.376718044281006 + ], + "logprob_reward": -0.001718236249871552, + "yes_no_margin_reward": 6.375, + "token_credit_reward": -0.001718236249871552, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11133059114217758, + 0.04832848906517029, + 0.029659440740942955, + 0.0589166134595871, + 0.26887035369873047, + 0.03308611735701561, + 0.07800839841365814, + 0.14749100804328918, + 0.14990398287773132, + 0.07440496236085892 + ], + "credit_map": [ + 0.012486460618674755, + 0.016359545290470123, + 0.01916838064789772, + 0.013861125335097313, + 0.012133716605603695, + 0.012309594079852104, + 0.01414668932557106, + 0.01086508110165596, + 0.0058599249459803104, + 0.005832700524479151, + 0.006052094046026468, + 0.010478162206709385, + 0.004492979496717453, + 0.0054581016302108765, + 0.004990528337657452, + 0.0051639932207763195, + 0.004042020067572594, + 0.0035945558920502663, + 0.004071068484336138, + 0.005305896978825331, + 0.002762335352599621, + 0.00414392352104187, + 0.0029808301478624344, + 0.0027588095981627703, + 0.006558588705956936, + 0.004432024899870157, + 0.01163493562489748, + 0.012140531092882156, + 0.004817452747374773, + 0.004772108048200607, + 0.008350052870810032, + 0.006210921797901392, + 0.018849221989512444, + 0.011883695609867573, + 0.09148261696100235, + 0.10741794854402542, + 0.005131123121827841, + 0.011620108038187027, + 0.016154572367668152, + 0.006331070326268673, + 0.0029672866221517324, + 0.002890432020649314, + 0.00707530090585351, + 0.004099289420992136, + 0.003041405463591218, + 0.004406374879181385, + 0.004334594588726759, + 0.004271435551345348, + 0.007807193789631128, + 0.014340183697640896, + 0.008262905292212963, + 0.008651336655020714, + 0.014174354262650013, + 0.006629789713770151, + 0.010861683636903763, + 0.0072809504345059395, + 0.018302662298083305, + 0.013184910640120506, + 0.017211193218827248, + 0.01799974963068962, + 0.017536574974656105, + 0.0157109834253788, + 0.02160719595849514, + 0.025937741622328758, + 0.016283342614769936, + 0.012854540720582008, + 0.009785334579646587, + 0.018256232142448425, + 0.029520926997065544, + 0.019439006224274635, + 0.025531865656375885, + 0.01823273114860058, + 0.0073624555952847, + 0.01764472760260105, + 0.004962127655744553, + 0.008410216309130192, + 0.016495978459715843, + 0.0070258090272545815, + 0.0075963204726576805, + 0.004907327704131603 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 8, + 7 + ], + "selected_reward": -0.001718236249871552 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a dimly lit futuristic control room featuring a three-dimensional holographic weather map?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera start with a close-up from a slightly overhead angle and slowly pan forward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do slender blue and white energy data streams intertwine with vast, irregular crimson danger zones that pulse ominously?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the energy lines and crimson zones spiral and contract towards a central point of blinding light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does a rainbow-colored anamorphic lens flare burst from the convergence, creating Makoto Shinkai's signature aesthetic light?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.002815808402374387 + ], + "opposite_logprob_tokens": [ + -5.8778157234191895 + ], + "logprob_reward": -0.002815808402374387, + "yes_no_margin_reward": 5.875, + "token_credit_reward": -0.002815808402374387, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07726039737462997, + 0.06581956893205643, + 0.02059968002140522, + 0.04447902739048004, + 0.13999710977077484, + 0.14015191793441772, + 0.10580035299062729, + 0.22608627378940582, + 0.10573706775903702, + 0.07406866550445557 + ], + "credit_map": [ + 0.010307994671165943, + 0.009799283929169178, + 0.008192313835024834, + 0.013621782884001732, + 0.010336719453334808, + 0.011803824454545975, + 0.006790583487600088, + 0.006407890468835831, + 0.005623604170978069, + 0.0074451700784265995, + 0.009974165819585323, + 0.01580059342086315, + 0.005309112835675478, + 0.006995655596256256, + 0.008391398936510086, + 0.0062798731960356236, + 0.004810225684195757, + 0.002387497341260314, + 0.0025852008257061243, + 0.003708860371261835, + 0.001495748176239431, + 0.0016781665617600083, + 0.0018495803233236074, + 0.002084399340674281, + 0.004075711127370596, + 0.0034223003312945366, + 0.005910343490540981, + 0.00998524110764265, + 0.0042900084517896175, + 0.004763358272612095, + 0.007975000888109207, + 0.00405706325545907, + 0.008313905447721481, + 0.007942277938127518, + 0.04508011415600777, + 0.04809974506497383, + 0.006076877471059561, + 0.008926833979785442, + 0.010984248481690884, + 0.004573115613311529, + 0.01083720475435257, + 0.008999824523925781, + 0.03573111817240715, + 0.02190639078617096, + 0.00937525276094675, + 0.01650802232325077, + 0.027368806302547455, + 0.009425297379493713, + 0.015466046519577503, + 0.012077433057129383, + 0.015623446553945541, + 0.012685742229223251, + 0.013265189714729786, + 0.008487897925078869, + 0.017672674730420113, + 0.010521916672587395, + 0.023380085825920105, + 0.021288540214300156, + 0.023389561101794243, + 0.03990631178021431, + 0.024368561804294586, + 0.02214151993393898, + 0.031769149005413055, + 0.03984254598617554, + 0.009884629398584366, + 0.008997654542326927, + 0.007431709207594395, + 0.014622175134718418, + 0.018192851915955544, + 0.01128037367016077, + 0.022084711119532585, + 0.013242963701486588, + 0.008211793377995491, + 0.008671448566019535, + 0.006859519053250551, + 0.011498963460326195, + 0.007361057214438915, + 0.009441884234547615, + 0.00978853926062584, + 0.01223546452820301 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 5, + 4 + ], + "selected_reward": -0.002815808402374387 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a dimly lit futuristic control room featuring a three-dimensional holographic weather map?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera start with a close-up from a slightly overhead angle and slowly pan forward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do slender blue and white energy data streams intertwine with vast, irregular crimson danger zones that pulse ominously?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the energy lines and crimson zones spiral and contract towards a central point of blinding light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does a rainbow-colored anamorphic lens flare burst from the convergence, creating Makoto Shinkai's signature aesthetic light?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01607270911335945 + ], + "opposite_logprob_tokens": [ + -4.141072750091553 + ], + "logprob_reward": -0.01607270911335945, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.01607270911335945, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07583580911159515, + 0.04121079668402672, + 0.021704716607928276, + 0.030667079612612724, + 0.2836529016494751, + 0.09907428175210953, + 0.07979180663824081, + 0.12700331211090088, + 0.03790345415472984, + 0.20315586030483246 + ], + "credit_map": [ + 0.007099988870322704, + 0.008833889849483967, + 0.01069902628660202, + 0.014163886196911335, + 0.006641429383307695, + 0.013632893562316895, + 0.007933113723993301, + 0.0068315789103507996, + 0.005551338195800781, + 0.0042847879230976105, + 0.006302028428763151, + 0.010445705614984035, + 0.0031599404755979776, + 0.004191029351204634, + 0.003853109898045659, + 0.0034228574950248003, + 0.005972309969365597, + 0.0018290007719770074, + 0.0029536511283367872, + 0.002954451134428382, + 0.0017805964453145862, + 0.0021955640986561775, + 0.0019665441941469908, + 0.0020525993313640356, + 0.003907442092895508, + 0.002418114338070154, + 0.005283083766698837, + 0.005094025284051895, + 0.002495350083336234, + 0.004171144682914019, + 0.003702217247337103, + 0.003595703514292836, + 0.01901061460375786, + 0.012393425218760967, + 0.09805510193109512, + 0.11311675608158112, + 0.004806643817573786, + 0.011846678331494331, + 0.017673663794994354, + 0.006750004831701517, + 0.00986290443688631, + 0.008625970222055912, + 0.017207346856594086, + 0.015540072694420815, + 0.007375616580247879, + 0.008290264755487442, + 0.021304655820131302, + 0.010867448523640633, + 0.008441345766186714, + 0.010854051448404789, + 0.0063640642911195755, + 0.0107235973700881, + 0.013790672644972801, + 0.006110924296081066, + 0.01718045398592949, + 0.006326694507151842, + 0.014071512967348099, + 0.01388473529368639, + 0.009626956656575203, + 0.027604220435023308, + 0.01779508963227272, + 0.010775599628686905, + 0.022885948419570923, + 0.010359239764511585, + 0.005974491126835346, + 0.0032078176736831665, + 0.003925482276827097, + 0.0054028648883104324, + 0.005027106963098049, + 0.0038429719861596823, + 0.0051786634139716625, + 0.005344056524336338, + 0.01657089963555336, + 0.04620829597115517, + 0.011070596054196358, + 0.035966906696558, + 0.03485945239663124, + 0.015699477866292, + 0.02103610523045063, + 0.02174413576722145 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 9, + 7 + ], + "selected_reward": -0.01607270911335945 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "close-up shot from a slightly overhead angle, slowly but steadily panning forward, focuses on a three-dimensional holographic weather map suspended in a dimly lit futuristic control room. Within this holographic projection, countless slender, brilliant blue and white energy data streams intertwine and flow like a complex neural network. Between them, vast, irregular crimson danger zones pulse ominously, emitting a sinister shimmer. As the camera approaches, as if forcibly pulled by an invisible force, all the glowing energy lines and crimson zones begin to spiral and contract, rapidly and irreversibly converging towards the center, ultimately converging into a singular point of immense density and blinding light. At this moment of convergence, a dazzling, rainbow-colored anamorphic lens flare, bursting forth from the concentrated energy, pierces the screen, bathing the entire scene in Makoto Shinkai's signature aesthetic light. The light penetrates even the finest particles suspended in the air, creating an atmosphere both majestic and unsettling, culminating in a visual climax of catastrophic aesthetics.", + "seed": 27573, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000050/group_10/index_8252_seed_27573_dp_21_rank_42_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 10, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.19467761740088463, + "vlm_logprob": -0.19467761740088463, + "vlm_yes_no_margin": 3.475, + "vlm_token_credit": -0.19467761740088463, + "vlm_yes_ratio": 0.8, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a dimly lit futuristic control room featuring a three-dimensional holographic weather map?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera start with a close-up from a slightly overhead angle and slowly pan forward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do slender blue and white energy data streams intertwine with vast, irregular crimson danger zones that pulse ominously?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the energy lines and crimson zones spiral and contract towards a central point of blinding light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does a rainbow-colored anamorphic lens flare burst from the convergence, creating Makoto Shinkai's signature aesthetic light?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -0.8286967277526855 + ], + "opposite_logprob_tokens": [ + -0.5786967277526855 + ], + "logprob_reward": -0.8286967277526855, + "yes_no_margin_reward": -0.25, + "token_credit_reward": -0.8286967277526855, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.25257745385169983, + 0.1254776418209076, + 0.05287883058190346, + 0.0618903711438179, + 0.05688067525625229, + 0.024600600823760033, + 0.05745762214064598, + 0.2673293948173523, + 0.031306784600019455, + 0.06960062682628632 + ], + "credit_map": [ + 0.037111882120370865, + 0.03313227742910385, + 0.020724469795823097, + 0.02409415692090988, + 0.029443824663758278, + 0.0420110821723938, + 0.02938731200993061, + 0.0366724394261837, + 0.011474190279841423, + 0.016054194420576096, + 0.01749533601105213, + 0.023110194131731987, + 0.01304886769503355, + 0.013660923577845097, + 0.011494900099933147, + 0.01913902349770069, + 0.008050353266298771, + 0.00659238500520587, + 0.005956446751952171, + 0.00447846669703722, + 0.006747250910848379, + 0.007694799918681383, + 0.006026454735547304, + 0.007332672365009785, + 0.012374560348689556, + 0.009189925156533718, + 0.006347691640257835, + 0.004772720392793417, + 0.004066609777510166, + 0.006737862713634968, + 0.005261801183223724, + 0.013139202259480953, + 0.005500900559127331, + 0.006449616048485041, + 0.007553115487098694, + 0.007328640203922987, + 0.007283365353941917, + 0.0076013049110770226, + 0.007201335858553648, + 0.007962397299706936, + 0.0025673515629023314, + 0.003908428829163313, + 0.0014587779296562076, + 0.0034414860419929028, + 0.0026683444157242775, + 0.0018702810630202293, + 0.0029596546664834023, + 0.005726277362555265, + 0.006480240263044834, + 0.006785987410694361, + 0.008595412597060204, + 0.008601858280599117, + 0.004789846949279308, + 0.010135763324797153, + 0.0049597956240177155, + 0.007108716759830713, + 0.03772173076868057, + 0.011878501623868942, + 0.11475344002246857, + 0.04011455550789833, + 0.010425078682601452, + 0.025199778378009796, + 0.015741370618343353, + 0.011494963429868221, + 0.003925176803022623, + 0.0047849067486822605, + 0.0035739478189498186, + 0.005523674190044403, + 0.004074269440025091, + 0.0020407915581017733, + 0.0031435510609298944, + 0.004240464419126511, + 0.010239087976515293, + 0.00987065490335226, + 0.010542190633714199, + 0.012336469255387783, + 0.00732674403116107, + 0.006235505919903517, + 0.006849254015833139, + 0.006200714968144894 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 0, + 1 + ], + "selected_reward": -0.8286967277526855 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a dimly lit futuristic control room featuring a three-dimensional holographic weather map?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera start with a close-up from a slightly overhead angle and slowly pan forward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do slender blue and white energy data streams intertwine with vast, irregular crimson danger zones that pulse ominously?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the energy lines and crimson zones spiral and contract towards a central point of blinding light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does a rainbow-colored anamorphic lens flare burst from the convergence, creating Makoto Shinkai's signature aesthetic light?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.11295172572135925 + ], + "opposite_logprob_tokens": [ + -2.2379517555236816 + ], + "logprob_reward": -0.11295172572135925, + "yes_no_margin_reward": 2.125, + "token_credit_reward": -0.11295172572135925, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.22362709045410156, + 0.09272897243499756, + 0.03958723694086075, + 0.042824968695640564, + 0.060996659100055695, + 0.0747290551662445, + 0.07594558596611023, + 0.26477041840553284, + 0.06739544123411179, + 0.057394497096538544 + ], + "credit_map": [ + 0.029756944626569748, + 0.032302357256412506, + 0.017593709751963615, + 0.021873202174901962, + 0.023189278319478035, + 0.03168604150414467, + 0.032774653285741806, + 0.03445093333721161, + 0.009304611012339592, + 0.012154364958405495, + 0.014507834799587727, + 0.013380022719502449, + 0.01172383688390255, + 0.010371011681854725, + 0.007265040650963783, + 0.014022253453731537, + 0.006145477760583162, + 0.003957204520702362, + 0.00411797733977437, + 0.0031820861622691154, + 0.0041165160946547985, + 0.008866412565112114, + 0.003674023551866412, + 0.005527544766664505, + 0.004575661849230528, + 0.0064439014531672, + 0.0036550224758684635, + 0.0038470507133752108, + 0.003862390061840415, + 0.005316169001162052, + 0.006966971326619387, + 0.008157805539667606, + 0.007107030134648085, + 0.007218535523861647, + 0.005676248576492071, + 0.005997487343847752, + 0.009630871936678886, + 0.00882666278630495, + 0.007754688151180744, + 0.008785135112702847, + 0.007057576905936003, + 0.010138274170458317, + 0.005492182914167643, + 0.011178513988852501, + 0.0085157984867692, + 0.005742114968597889, + 0.011168994009494781, + 0.015435605309903622, + 0.01311024371534586, + 0.009009245783090591, + 0.015026535838842392, + 0.009036585688591003, + 0.005184634122997522, + 0.006455465219914913, + 0.007921374402940273, + 0.010201514698565006, + 0.0368376262485981, + 0.010526865720748901, + 0.11040620505809784, + 0.03721701726317406, + 0.012746504507958889, + 0.028084905818104744, + 0.014562618918716908, + 0.014388701878488064, + 0.008102248422801495, + 0.008537809364497662, + 0.006881800480186939, + 0.013499976135790348, + 0.008100821636617184, + 0.005456768441945314, + 0.006864805705845356, + 0.009951215237379074, + 0.006861490197479725, + 0.006446253042668104, + 0.008678634651005268, + 0.010581952519714832, + 0.004914760589599609, + 0.006403251551091671, + 0.006863836664706469, + 0.0066443257965147495 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 0, + 1 + ], + "selected_reward": -0.11295172572135925 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a dimly lit futuristic control room featuring a three-dimensional holographic weather map?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera start with a close-up from a slightly overhead angle and slowly pan forward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do slender blue and white energy data streams intertwine with vast, irregular crimson danger zones that pulse ominously?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the energy lines and crimson zones spiral and contract towards a central point of blinding light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does a rainbow-colored anamorphic lens flare burst from the convergence, creating Makoto Shinkai's signature aesthetic light?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.002201516181230545 + ], + "opposite_logprob_tokens": [ + -6.127201557159424 + ], + "logprob_reward": -0.002201516181230545, + "yes_no_margin_reward": 6.125, + "token_credit_reward": -0.002201516181230545, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2700563967227936, + 0.07027386128902435, + 0.17821213603019714, + 0.06403268873691559, + 0.1169753298163414, + 0.0359218455851078, + 0.0802774503827095, + 0.06918195635080338, + 0.059863679111003876, + 0.05520465224981308 + ], + "credit_map": [ + 0.038809604942798615, + 0.03622117638587952, + 0.03341786935925484, + 0.023567916825413704, + 0.028219932690262794, + 0.03656613826751709, + 0.03353413939476013, + 0.039719581604003906, + 0.00683227926492691, + 0.009958014823496342, + 0.006025990936905146, + 0.008682903833687305, + 0.01004844531416893, + 0.009492685087025166, + 0.006140928249806166, + 0.01309260819107294, + 0.012746848165988922, + 0.016505448147654533, + 0.016546057537198067, + 0.014182629995048046, + 0.01741241104900837, + 0.05597735196352005, + 0.009828424081206322, + 0.03501295670866966, + 0.006782970391213894, + 0.01056612841784954, + 0.005138398613780737, + 0.004683081526309252, + 0.005045298486948013, + 0.007541600149124861, + 0.007600372191518545, + 0.016674833372235298, + 0.012181207537651062, + 0.013075866736471653, + 0.008958360180258751, + 0.016643468290567398, + 0.013163362629711628, + 0.017934488132596016, + 0.015964066609740257, + 0.019054517149925232, + 0.004022887907922268, + 0.006673108320683241, + 0.0026797899045050144, + 0.004709720611572266, + 0.003971015103161335, + 0.003653236199170351, + 0.004707393702119589, + 0.005504695698618889, + 0.011773700825870037, + 0.014827086590230465, + 0.009862307459115982, + 0.016462085768580437, + 0.005247682798653841, + 0.008449526503682137, + 0.00522241136059165, + 0.00843264814466238, + 0.005810994654893875, + 0.011098108254373074, + 0.008917349390685558, + 0.009324458427727222, + 0.009076230227947235, + 0.0072774337604641914, + 0.007593131624162197, + 0.010084250941872597, + 0.0062737734988331795, + 0.01065873820334673, + 0.005607222206890583, + 0.012307560071349144, + 0.006674888078123331, + 0.0035676362458616495, + 0.007562051527202129, + 0.007211808115243912, + 0.006439395248889923, + 0.00809638574719429, + 0.007137846667319536, + 0.008594110608100891, + 0.008613895624876022, + 0.005638424307107925, + 0.00506895687431097, + 0.0056156376376748085 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 4 + ], + "selected_reward": -0.002201516181230545 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a dimly lit futuristic control room featuring a three-dimensional holographic weather map?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera start with a close-up from a slightly overhead angle and slowly pan forward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do slender blue and white energy data streams intertwine with vast, irregular crimson danger zones that pulse ominously?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the energy lines and crimson zones spiral and contract towards a central point of blinding light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does a rainbow-colored anamorphic lens flare burst from the convergence, creating Makoto Shinkai's signature aesthetic light?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.00318971648812294 + ], + "opposite_logprob_tokens": [ + -5.753189563751221 + ], + "logprob_reward": -0.00318971648812294, + "yes_no_margin_reward": 5.75, + "token_credit_reward": -0.00318971648812294, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1782447099685669, + 0.04545818641781807, + 0.03823261335492134, + 0.024434547871351242, + 0.05577990040183067, + 0.030628245323896408, + 0.04695993661880493, + 0.361756294965744, + 0.10238952189683914, + 0.11611603945493698 + ], + "credit_map": [ + 0.023905787616968155, + 0.02027931809425354, + 0.014138026162981987, + 0.013332447037100792, + 0.018761584535241127, + 0.026832235977053642, + 0.020643172785639763, + 0.040352147072553635, + 0.004079424310475588, + 0.005622718948870897, + 0.008415035903453827, + 0.00759607320651412, + 0.007348434533923864, + 0.00413000350818038, + 0.0032748563680797815, + 0.004991641268134117, + 0.0039574033580720425, + 0.005273018963634968, + 0.004009381402283907, + 0.004590539261698723, + 0.004198754206299782, + 0.007929407060146332, + 0.0037930363323539495, + 0.004481069277971983, + 0.003728598589077592, + 0.003876130562275648, + 0.0021863081492483616, + 0.002312916796654463, + 0.002239104127511382, + 0.0023456458002328873, + 0.0036176848225295544, + 0.004128159489482641, + 0.005222686566412449, + 0.004839497152715921, + 0.005866828374564648, + 0.00788882840424776, + 0.006198713555932045, + 0.007770469877868891, + 0.007014947477728128, + 0.01097792573273182, + 0.003032707842066884, + 0.00398235023021698, + 0.002886576810851693, + 0.0054645417258143425, + 0.0033152687828987837, + 0.0025521963834762573, + 0.004192342050373554, + 0.0052022626623511314, + 0.005880436860024929, + 0.008094406686723232, + 0.006613587494939566, + 0.00765504315495491, + 0.002819495741277933, + 0.004987697582691908, + 0.004934615921229124, + 0.0059746550396084785, + 0.052734144032001495, + 0.01554877869784832, + 0.15582521259784698, + 0.049006883054971695, + 0.020504357293248177, + 0.03284242004156113, + 0.020780662074685097, + 0.014513872563838959, + 0.012293864972889423, + 0.020389171317219734, + 0.00915590301156044, + 0.01989549770951271, + 0.01198878325521946, + 0.007569096516817808, + 0.008550631813704967, + 0.012546577490866184, + 0.011605612002313137, + 0.016600418835878372, + 0.01161948498338461, + 0.017997518181800842, + 0.017211703583598137, + 0.015516962856054306, + 0.014027925208210945, + 0.011536415666341782 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 0, + 9 + ], + "selected_reward": -0.00318971648812294 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a dimly lit futuristic control room featuring a three-dimensional holographic weather map?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera start with a close-up from a slightly overhead angle and slowly pan forward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do slender blue and white energy data streams intertwine with vast, irregular crimson danger zones that pulse ominously?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the energy lines and crimson zones spiral and contract towards a central point of blinding light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does a rainbow-colored anamorphic lens flare burst from the convergence, creating Makoto Shinkai's signature aesthetic light?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.026348400861024857 + ], + "opposite_logprob_tokens": [ + -3.651348352432251 + ], + "logprob_reward": -0.026348400861024857, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.026348400861024857, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06723669171333313, + 0.03751455992460251, + 0.02297324128448963, + 0.05664417892694473, + 0.03800160437822342, + 0.0226670540869236, + 0.03413783386349678, + 0.5193625688552856, + 0.07515218108892441, + 0.12631013989448547 + ], + "credit_map": [ + 0.00604788213968277, + 0.010073092766106129, + 0.003484579501673579, + 0.005733417812734842, + 0.010006037540733814, + 0.008125890046358109, + 0.008201699703931808, + 0.015564090572297573, + 0.003795830998569727, + 0.006678654812276363, + 0.006065533962100744, + 0.005278643686324358, + 0.004606207367032766, + 0.003962332848459482, + 0.003702355083078146, + 0.0034250004682689905, + 0.002631435403600335, + 0.0023722106125205755, + 0.0020560009870678186, + 0.002000577747821808, + 0.0018445906462147832, + 0.007190281059592962, + 0.0016581672243773937, + 0.00321997725404799, + 0.013351923786103725, + 0.006896903272718191, + 0.006703061982989311, + 0.004308137111365795, + 0.005296270363032818, + 0.005628092214465141, + 0.005076935514807701, + 0.009382856078445911, + 0.003938593901693821, + 0.0047433041036129, + 0.0028999042697250843, + 0.005724338348954916, + 0.004743610043078661, + 0.004966736771166325, + 0.006344749592244625, + 0.004640366416424513, + 0.002239921595901251, + 0.0037590041756629944, + 0.0018339415546506643, + 0.002870123367756605, + 0.003336598165333271, + 0.002127828309312463, + 0.001956101506948471, + 0.0045435368083417416, + 0.004426604136824608, + 0.0035762693732976913, + 0.0035427839029580355, + 0.008239682763814926, + 0.0021827637683600187, + 0.004222122486680746, + 0.004338840488344431, + 0.0036087704356759787, + 0.07856351882219315, + 0.016098326072096825, + 0.23207613825798035, + 0.07526223361492157, + 0.02040746994316578, + 0.04829298332333565, + 0.03106265887618065, + 0.0175991989672184, + 0.008559021167457104, + 0.014446971006691456, + 0.006010888144373894, + 0.016075115650892258, + 0.008582108654081821, + 0.0035057533532381058, + 0.007837709039449692, + 0.010134613141417503, + 0.01137122604995966, + 0.03223620727658272, + 0.011556334793567657, + 0.016735468059778214, + 0.023048335686326027, + 0.007757091894745827, + 0.011585365980863571, + 0.012020106427371502 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 9, + 8 + ], + "selected_reward": -0.026348400861024857 + } + ], + "vlm_decoded_outputs": [ + "No", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "close-up shot from a slightly overhead angle, slowly but steadily panning forward, focuses on a three-dimensional holographic weather map suspended in a dimly lit futuristic control room. Within this holographic projection, countless slender, brilliant blue and white energy data streams intertwine and flow like a complex neural network. Between them, vast, irregular crimson danger zones pulse ominously, emitting a sinister shimmer. As the camera approaches, as if forcibly pulled by an invisible force, all the glowing energy lines and crimson zones begin to spiral and contract, rapidly and irreversibly converging towards the center, ultimately converging into a singular point of immense density and blinding light. At this moment of convergence, a dazzling, rainbow-colored anamorphic lens flare, bursting forth from the concentrated energy, pierces the screen, bathing the entire scene in Makoto Shinkai's signature aesthetic light. The light penetrates even the finest particles suspended in the air, creating an atmosphere both majestic and unsettling, culminating in a visual climax of catastrophic aesthetics.", + "seed": 27574, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000050/group_10/index_8252_seed_27574_dp_21_rank_42_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 10, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.2358641677070409, + "vlm_logprob": -0.2358641677070409, + "vlm_yes_no_margin": 3.650000011920929, + "vlm_token_credit": -0.2358641677070409, + "vlm_yes_ratio": 0.8, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a dimly lit futuristic control room featuring a three-dimensional holographic weather map?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera start with a close-up from a slightly overhead angle and slowly pan forward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do slender blue and white energy data streams intertwine with vast, irregular crimson danger zones that pulse ominously?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the energy lines and crimson zones spiral and contract towards a central point of blinding light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does a rainbow-colored anamorphic lens flare burst from the convergence, creating Makoto Shinkai's signature aesthetic light?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -1.057140827178955 + ], + "opposite_logprob_tokens": [ + -0.43214088678359985 + ], + "logprob_reward": -1.057140827178955, + "yes_no_margin_reward": -0.6249999403953552, + "token_credit_reward": -1.057140827178955, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.16413052380084991, + 0.03681279346346855, + 0.02051139622926712, + 0.03430831432342529, + 0.044552031904459, + 0.11279428750276566, + 0.03858255222439766, + 0.15493784844875336, + 0.04066174477338791, + 0.35270848870277405 + ], + "credit_map": [ + 0.050134897232055664, + 0.009504474699497223, + 0.012613067403435707, + 0.011293984949588776, + 0.007246539928019047, + 0.04349203407764435, + 0.010281111113727093, + 0.01956441067159176, + 0.004403000231832266, + 0.005366341210901737, + 0.004486162215471268, + 0.004823422525078058, + 0.0036132249515503645, + 0.005015623290091753, + 0.00474828714504838, + 0.00435673538595438, + 0.0030183393973857164, + 0.0028976111207157373, + 0.002340917708352208, + 0.001640285481698811, + 0.0029590516351163387, + 0.0026012284215539694, + 0.0022856167051941156, + 0.0027683465741574764, + 0.004391457419842482, + 0.003263662802055478, + 0.004372050985693932, + 0.0034418818540871143, + 0.0034733873326331377, + 0.0043657380156219006, + 0.005122021771967411, + 0.0058781178668141365, + 0.00612086895853281, + 0.004194807261228561, + 0.0070526981726288795, + 0.005027758423238993, + 0.005070214159786701, + 0.0033462143037468195, + 0.009933996945619583, + 0.003805474378168583, + 0.016470545902848244, + 0.014017625711858273, + 0.014880535192787647, + 0.012463039718568325, + 0.014238706789910793, + 0.015149395912885666, + 0.016513938084244728, + 0.009060505777597427, + 0.007264069747179747, + 0.004663112573325634, + 0.006120867561548948, + 0.002695803064852953, + 0.004856886342167854, + 0.005131545476615429, + 0.0037163393571972847, + 0.004133928567171097, + 0.008382417261600494, + 0.025734199211001396, + 0.0054853931069374084, + 0.019863709807395935, + 0.06121526658535004, + 0.014215584844350815, + 0.013079260475933552, + 0.006962025072425604, + 0.003587364684790373, + 0.0031902617774903774, + 0.0062604486010968685, + 0.009065522812306881, + 0.00407258328050375, + 0.003935499582439661, + 0.007182275876402855, + 0.0033677900210022926, + 0.007361069321632385, + 0.022715510800480843, + 0.03268839046359062, + 0.15963418781757355, + 0.009359699673950672, + 0.005800125654786825, + 0.10432539135217667, + 0.010824164375662804 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 0, + 7 + ], + "selected_reward": -1.057140827178955 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a dimly lit futuristic control room featuring a three-dimensional holographic weather map?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera start with a close-up from a slightly overhead angle and slowly pan forward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do slender blue and white energy data streams intertwine with vast, irregular crimson danger zones that pulse ominously?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the energy lines and crimson zones spiral and contract towards a central point of blinding light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does a rainbow-colored anamorphic lens flare burst from the convergence, creating Makoto Shinkai's signature aesthetic light?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.10033480077981949 + ], + "opposite_logprob_tokens": [ + -2.350334882736206 + ], + "logprob_reward": -0.10033480077981949, + "yes_no_margin_reward": 2.25, + "token_credit_reward": -0.10033480077981949, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.162960946559906, + 0.024109195917844772, + 0.020396899431943893, + 0.030134227126836777, + 0.03736235573887825, + 0.04334260895848274, + 0.049397245049476624, + 0.13672682642936707, + 0.2555232346057892, + 0.24004636704921722 + ], + "credit_map": [ + 0.04568662494421005, + 0.013101417571306229, + 0.01359854731708765, + 0.011541550047695637, + 0.009929593652486801, + 0.034379687160253525, + 0.011983501724898815, + 0.022740038111805916, + 0.0022159579675644636, + 0.0030484104063361883, + 0.003104252042248845, + 0.002697910415008664, + 0.002689101966097951, + 0.0034412830136716366, + 0.003377101384103298, + 0.0035351798869669437, + 0.0033780422527343035, + 0.002653507748618722, + 0.0022291101049631834, + 0.001971885794773698, + 0.002658701501786709, + 0.0026894949842244387, + 0.002235189313068986, + 0.0025809681974351406, + 0.0045925830490887165, + 0.0031083147041499615, + 0.004025839734822512, + 0.00310756778344512, + 0.00332014006562531, + 0.0025523442309349775, + 0.003089685458689928, + 0.00633775582537055, + 0.004887014627456665, + 0.0036765385884791613, + 0.003622762393206358, + 0.004041946493089199, + 0.004969420842826366, + 0.003060863818973303, + 0.009454533457756042, + 0.0036492757499217987, + 0.005990348756313324, + 0.003495726501569152, + 0.007646429818123579, + 0.004764951299875975, + 0.004360683262348175, + 0.006449935957789421, + 0.006779799237847328, + 0.0038547369185835123, + 0.008434111252427101, + 0.004936110228300095, + 0.004175262525677681, + 0.003985214978456497, + 0.0042735072784125805, + 0.016295453533530235, + 0.0035121457185596228, + 0.003785443492233753, + 0.02210353873670101, + 0.02129930630326271, + 0.00981878861784935, + 0.01819755509495735, + 0.025582952424883842, + 0.010816407389938831, + 0.020245570689439774, + 0.008662729524075985, + 0.00806910078972578, + 0.011160746216773987, + 0.01773337833583355, + 0.06333044916391373, + 0.03506805747747421, + 0.007793989963829517, + 0.10232528299093246, + 0.010042245499789715, + 0.011991644278168678, + 0.01781638339161873, + 0.015403243713080883, + 0.10914625227451324, + 0.014585208147764206, + 0.0062919133342802525, + 0.05719870328903198, + 0.00761305121704936 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 9, + 0 + ], + "selected_reward": -0.10033480077981949 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a dimly lit futuristic control room featuring a three-dimensional holographic weather map?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera start with a close-up from a slightly overhead angle and slowly pan forward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do slender blue and white energy data streams intertwine with vast, irregular crimson danger zones that pulse ominously?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the energy lines and crimson zones spiral and contract towards a central point of blinding light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does a rainbow-colored anamorphic lens flare burst from the convergence, creating Makoto Shinkai's signature aesthetic light?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0017187122721225023 + ], + "opposite_logprob_tokens": [ + -6.376718521118164 + ], + "logprob_reward": -0.0017187122721225023, + "yes_no_margin_reward": 6.375, + "token_credit_reward": -0.0017187122721225023, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07200904190540314, + 0.02774702198803425, + 0.014728737995028496, + 0.01660434901714325, + 0.022872887551784515, + 0.09041957557201385, + 0.024386374279856682, + 0.142159104347229, + 0.07801273465156555, + 0.5110601782798767 + ], + "credit_map": [ + 0.020939065143465996, + 0.004655718803405762, + 0.005398943554610014, + 0.004095064010471106, + 0.003694351762533188, + 0.01534517202526331, + 0.005881721619516611, + 0.011998993344604969, + 0.0027259020134806633, + 0.002442939905449748, + 0.0035310559906065464, + 0.005034483037889004, + 0.0025147756095975637, + 0.0033891613129526377, + 0.004210865590721369, + 0.003897834802046418, + 0.0020541988778859377, + 0.0022025934886187315, + 0.0018827830208465457, + 0.0014349455013871193, + 0.0018942143069580197, + 0.0014703632332384586, + 0.001861895783804357, + 0.0019277435494586825, + 0.0021543712355196476, + 0.001620201743207872, + 0.002521035959944129, + 0.0013617369113489985, + 0.0015699139330536127, + 0.0012827719328925014, + 0.002566870301961899, + 0.003527446184307337, + 0.0030993730761110783, + 0.002244085306301713, + 0.00569746270775795, + 0.0024818771053105593, + 0.002303150249645114, + 0.0023821440991014242, + 0.0021064833272248507, + 0.0025583074893802404, + 0.013301524333655834, + 0.011710082180798054, + 0.00960156787186861, + 0.009718898683786392, + 0.012266715988516808, + 0.012938953936100006, + 0.013468370772898197, + 0.007413456216454506, + 0.003984543494880199, + 0.002927492605522275, + 0.0032038295175880194, + 0.001985659124329686, + 0.0031008475925773382, + 0.004530693404376507, + 0.0023742932826280594, + 0.0022790119983255863, + 0.009654573164880276, + 0.02308811992406845, + 0.005499690305441618, + 0.017241261899471283, + 0.05650164932012558, + 0.010241073556244373, + 0.01299133151769638, + 0.006941409315913916, + 0.004015232436358929, + 0.004203116055577993, + 0.00839606299996376, + 0.014531904831528664, + 0.010020486079156399, + 0.0044115460477769375, + 0.028583252802491188, + 0.0038511212915182114, + 0.007490986492484808, + 0.026003358885645866, + 0.0364353284239769, + 0.2160908728837967, + 0.020248839631676674, + 0.00571427633985877, + 0.1812298595905304, + 0.017846638336777687 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 7, + 5 + ], + "selected_reward": -0.0017187122721225023 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a dimly lit futuristic control room featuring a three-dimensional holographic weather map?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera start with a close-up from a slightly overhead angle and slowly pan forward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do slender blue and white energy data streams intertwine with vast, irregular crimson danger zones that pulse ominously?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the energy lines and crimson zones spiral and contract towards a central point of blinding light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does a rainbow-colored anamorphic lens flare burst from the convergence, creating Makoto Shinkai's signature aesthetic light?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0019386084750294685 + ], + "opposite_logprob_tokens": [ + -6.251938819885254 + ], + "logprob_reward": -0.0019386084750294685, + "yes_no_margin_reward": 6.25, + "token_credit_reward": -0.0019386084750294685, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0747600644826889, + 0.02296648547053337, + 0.017020272091031075, + 0.01768738403916359, + 0.03378475829958916, + 0.05025007575750351, + 0.028614168986678123, + 0.12062984704971313, + 0.2111070305109024, + 0.4231799244880676 + ], + "credit_map": [ + 0.011275029741227627, + 0.006828653160482645, + 0.00750485435128212, + 0.007127909921109676, + 0.005821832921355963, + 0.007751428987830877, + 0.008899956941604614, + 0.019550397992134094, + 0.0030216448940336704, + 0.0028127748519182205, + 0.0028059016913175583, + 0.003292314475402236, + 0.002249953569844365, + 0.00289256707765162, + 0.003080789465457201, + 0.002810540609061718, + 0.0019323724554851651, + 0.002237428445369005, + 0.002253012964501977, + 0.0015199503395706415, + 0.0028068304527550936, + 0.0019237602828070521, + 0.002360208658501506, + 0.001986709190532565, + 0.0027139633893966675, + 0.001466417801566422, + 0.0018300587544217706, + 0.0013940910575911403, + 0.0023899818770587444, + 0.0012268312275409698, + 0.002712870715186, + 0.0039531695656478405, + 0.0051768007688224316, + 0.0028724120929837227, + 0.006243253592401743, + 0.004023877903819084, + 0.0037607657723128796, + 0.002547360258176923, + 0.0062296148389577866, + 0.0029306744690984488, + 0.005803592037409544, + 0.0041214642114937305, + 0.004823668394237757, + 0.0056683155708014965, + 0.007279715035110712, + 0.0051185679621994495, + 0.01314359437674284, + 0.004291156306862831, + 0.005041096825152636, + 0.00244081299751997, + 0.0040251826867461205, + 0.0027639088220894337, + 0.0023679626174271107, + 0.00657035643234849, + 0.0023657670244574547, + 0.003039080183953047, + 0.01332206279039383, + 0.017783451825380325, + 0.00643495935946703, + 0.018954357132315636, + 0.026460163295269012, + 0.011542427353560925, + 0.01831590011715889, + 0.007816535420715809, + 0.011712979525327682, + 0.023494722321629524, + 0.013611938804388046, + 0.03695473447442055, + 0.06452815979719162, + 0.016888467594981194, + 0.032567400485277176, + 0.011348611675202847, + 0.0074790408834815025, + 0.022836418822407722, + 0.02337788976728916, + 0.1812480241060257, + 0.01610083319246769, + 0.0046570696868002415, + 0.15352573990821838, + 0.01395491138100624 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 8, + 7 + ], + "selected_reward": -0.0019386084750294685 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a dimly lit futuristic control room featuring a three-dimensional holographic weather map?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera start with a close-up from a slightly overhead angle and slowly pan forward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do slender blue and white energy data streams intertwine with vast, irregular crimson danger zones that pulse ominously?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the energy lines and crimson zones spiral and contract towards a central point of blinding light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does a rainbow-colored anamorphic lens flare burst from the convergence, creating Makoto Shinkai's signature aesthetic light?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.018187889829277992 + ], + "opposite_logprob_tokens": [ + -4.018187999725342 + ], + "logprob_reward": -0.018187889829277992, + "yes_no_margin_reward": 4.0, + "token_credit_reward": -0.018187889829277992, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13362883031368256, + 0.020391952246427536, + 0.016207050532102585, + 0.02497662417590618, + 0.023858841508626938, + 0.02862907201051712, + 0.05490059033036232, + 0.10470249503850937, + 0.24941596388816833, + 0.3432885408401489 + ], + "credit_map": [ + 0.04180936887860298, + 0.009022779762744904, + 0.009187735617160797, + 0.008154884912073612, + 0.0071492125280201435, + 0.032398831099271774, + 0.008421167731285095, + 0.01748484931886196, + 0.00223086541518569, + 0.0024456894025206566, + 0.0019354074029251933, + 0.002088319743052125, + 0.003039696253836155, + 0.002994425827637315, + 0.002711367094889283, + 0.002946181921288371, + 0.0015596102457493544, + 0.00195250881370157, + 0.00205525872297585, + 0.0019012404372915626, + 0.002011755248531699, + 0.002098867204040289, + 0.002689182059839368, + 0.0019386277999728918, + 0.002821202389895916, + 0.003238715697079897, + 0.0026347741950303316, + 0.003905929159373045, + 0.0036218210589140654, + 0.002089856192469597, + 0.0038344417698681355, + 0.002829883713275194, + 0.0031777338590472937, + 0.002464634831994772, + 0.0034109572879970074, + 0.0026005408726632595, + 0.0028277586679905653, + 0.0024866291787475348, + 0.004468874540179968, + 0.0024217122700065374, + 0.0031730474438518286, + 0.003183764172717929, + 0.003282444551587105, + 0.003444492118433118, + 0.004725188948214054, + 0.003516154130920768, + 0.00415057735517621, + 0.003153405385091901, + 0.00893198698759079, + 0.006545673124492168, + 0.004440590273588896, + 0.004104091785848141, + 0.007773126941174269, + 0.013934030197560787, + 0.005114808212965727, + 0.00405628327280283, + 0.01128664705902338, + 0.012552885338664055, + 0.006997300777584314, + 0.02409268543124199, + 0.026113765314221382, + 0.008644879795610905, + 0.008828422985970974, + 0.006185903213918209, + 0.009727824479341507, + 0.016428934410214424, + 0.023010291159152985, + 0.014994320459663868, + 0.07003045082092285, + 0.010955027304589748, + 0.09293584525585175, + 0.011333273723721504, + 0.01868237741291523, + 0.010732188820838928, + 0.01181156188249588, + 0.09436601400375366, + 0.04044531658291817, + 0.009821590967476368, + 0.14023038744926453, + 0.01719910465180874 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 8, + 0 + ], + "selected_reward": -0.018187889829277992 + } + ], + "vlm_decoded_outputs": [ + "No", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000050/group_11/index_tensor([3100, 3100])_group_11_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000050/group_11/index_tensor([3100, 3100])_group_11_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..4a3a2f185faa77643b2e39d091ffecf92ef4f664 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000050/group_11/index_tensor([3100, 3100])_group_11_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "handheld camera, shot from a slightly elevated angle, follows a ragged elderly man as he walks down a muddy, rain-soaked country road. His sparse, short, graying hair is ruffled by the cold wind. His bronzed face is etched with deep, knife-like wrinkles, and his sunken eyes hold a gaze that is a mixture of fatigue and steeliness. He wears a faded, frayed dark blue Mao suit, the hems and cuffs ripped. He wears a gray shirt underneath, the collar yellowed, and his equally worn, dark trousers are splattered with damp mud. Under the gloomy evening sky, the thick clouds cast a faint, cold, scattered light, casting a chilling gray-brown hue on the entire scene. His steps were heavy and unsteady, each one sinking deeply into the sticky mud. Pulling out his stick was incredibly laborious, his body swaying from side to side, struggling to maintain balance with a crude crutch carved from a tree branch, tightly clenched in one hand. The road flanked by bleak winter fields, bare branches trembling slightly in the wind. The entire scene is imbued with a raw documentary quality, devoid of embellishment, focused solely on capturing every step of the old man's struggle against adversity, conveying a poignant, authentic, and resilient atmosphere.", + "seed": 11386, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000050/group_11/index_3100_seed_11386_dp_22_rank_44_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 11, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.2973839769139886, + "vlm_logprob": -0.2973839769139886, + "vlm_yes_no_margin": 2.325, + "vlm_token_credit": -0.2973839769139886, + "vlm_yes_ratio": 0.8, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot handheld?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera angle slightly elevated?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the man wearing a faded, frayed dark blue Mao suit with ripped hems and cuffs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the man using a crude crutch carved from a tree branch to support himself?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the environment a muddy, rain-soaked country road flanked by bleak winter fields?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -1.1667017936706543 + ], + "opposite_logprob_tokens": [ + -0.4167018234729767 + ], + "logprob_reward": -1.1667017936706543, + "yes_no_margin_reward": -0.75, + "token_credit_reward": -1.1667017936706543, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.06902237981557846, + 0.03717643767595291, + 0.025658385828137398, + 0.049333613365888596, + 0.025914806872606277, + 0.6999296545982361, + 0.017557507380843163, + 0.04259426146745682, + 0.024413224309682846, + 0.008399805054068565 + ], + "credit_map": [ + 0.011678186245262623, + 0.009700573049485683, + 0.006576938088983297, + 0.007929079234600067, + 0.01150421891361475, + 0.007957050576806068, + 0.007770896889269352, + 0.005905440542846918, + 0.004246972501277924, + 0.0037726869340986013, + 0.004638099577277899, + 0.005347643978893757, + 0.0033228944521397352, + 0.006116758566349745, + 0.005245383363217115, + 0.00448599923402071, + 0.0033391332253813744, + 0.0036517370026558638, + 0.004484409000724554, + 0.0042007919400930405, + 0.00255009182728827, + 0.0019236605148762465, + 0.0032601437997072935, + 0.0022484189830720425, + 0.012798367068171501, + 0.003117568325251341, + 0.003939460963010788, + 0.003409529570490122, + 0.0035659626591950655, + 0.014492066577076912, + 0.0037566989194601774, + 0.004253958351910114, + 0.0034621029626578093, + 0.0029017312917858362, + 0.004173200111836195, + 0.002864095149561763, + 0.0028844554908573627, + 0.0039030001498758793, + 0.0025831281673163176, + 0.003143095411360264, + 0.2695160508155823, + 0.028875047340989113, + 0.01629466563463211, + 0.021206775680184364, + 0.026086153462529182, + 0.292741984128952, + 0.015139477327466011, + 0.030069468542933464, + 0.001872014719992876, + 0.0025525609962642193, + 0.0020293963607400656, + 0.002037202473729849, + 0.003123806556686759, + 0.0017041111132130027, + 0.001852461020462215, + 0.002385954139754176, + 0.00753010855987668, + 0.00605013919994235, + 0.004508072976022959, + 0.00454587722197175, + 0.005401840899139643, + 0.00459152041003108, + 0.004998357966542244, + 0.004968340042978525, + 0.0020980997942388058, + 0.0030570970848202705, + 0.003699467284604907, + 0.004023609217256308, + 0.002285770373418927, + 0.003894681576639414, + 0.002750210929661989, + 0.0026042861863970757, + 0.0009697945206426084, + 0.0009187421528622508, + 0.0009556124568916857, + 0.0009691976010799408, + 0.001389350974932313, + 0.001043190946802497, + 0.0013202971313148737, + 0.0008336196187883615 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 0, + 3 + ], + "selected_reward": -1.1667017936706543 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot handheld?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera angle slightly elevated?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the man wearing a faded, frayed dark blue Mao suit with ripped hems and cuffs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the man using a crude crutch carved from a tree branch to support himself?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the environment a muddy, rain-soaked country road flanked by bleak winter fields?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.22565656900405884 + ], + "opposite_logprob_tokens": [ + -1.600656509399414 + ], + "logprob_reward": -0.22565656900405884, + "yes_no_margin_reward": 1.375, + "token_credit_reward": -0.22565656900405884, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08765335381031036, + 0.02986626885831356, + 0.012882411479949951, + 0.0322989746928215, + 0.13560357689857483, + 0.5231272578239441, + 0.026527991518378258, + 0.049747396260499954, + 0.05164492502808571, + 0.0506478026509285 + ], + "credit_map": [ + 0.015166381374001503, + 0.015311995521187782, + 0.007664118427783251, + 0.010170510038733482, + 0.010833072476089, + 0.00973882619291544, + 0.010220065712928772, + 0.008548391982913017, + 0.0035311351530253887, + 0.0029965867288410664, + 0.0037982971407473087, + 0.0051374128088355064, + 0.002770643215626478, + 0.004392304923385382, + 0.004707457963377237, + 0.0025324297603219748, + 0.0019261231645941734, + 0.0018742267275229096, + 0.0015005627647042274, + 0.001469859154894948, + 0.001601704629138112, + 0.0015729235019534826, + 0.0016250930493697524, + 0.001311917556449771, + 0.005461728200316429, + 0.0036065957974642515, + 0.0031579181086272, + 0.0028704642318189144, + 0.00451526278629899, + 0.0056800623424351215, + 0.003746289759874344, + 0.003260654164478183, + 0.04561859369277954, + 0.005806786008179188, + 0.006042814347893, + 0.006446509622037411, + 0.007395664229989052, + 0.051297999918460846, + 0.004532033111900091, + 0.008463184349238873, + 0.20213332772254944, + 0.020699063315987587, + 0.012193641625344753, + 0.01550878956913948, + 0.019978679716587067, + 0.21969614923000336, + 0.011134465225040913, + 0.021783120930194855, + 0.0051149469800293446, + 0.00532517721876502, + 0.002576886909082532, + 0.002227550372481346, + 0.0035226759500801563, + 0.0022259505931288004, + 0.0030145125929266214, + 0.0025202929973602295, + 0.009966021403670311, + 0.00821667816489935, + 0.005458852741867304, + 0.0037829687353223562, + 0.005052907392382622, + 0.005944619420915842, + 0.004935882519930601, + 0.006389467045664787, + 0.004174510482698679, + 0.005970033351331949, + 0.009693050757050514, + 0.009904618375003338, + 0.0043784151785075665, + 0.007901403121650219, + 0.0055482336319983006, + 0.00407466059550643, + 0.00845347624272108, + 0.007961270399391651, + 0.0024146027863025665, + 0.0040230038575828075, + 0.010549010708928108, + 0.007658438291400671, + 0.005918646231293678, + 0.0036693550646305084 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 4, + 0 + ], + "selected_reward": -0.22565656900405884 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot handheld?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera angle slightly elevated?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the man wearing a faded, frayed dark blue Mao suit with ripped hems and cuffs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the man using a crude crutch carved from a tree branch to support himself?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the environment a muddy, rain-soaked country road flanked by bleak winter fields?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.05498374626040459 + ], + "opposite_logprob_tokens": [ + -2.9299838542938232 + ], + "logprob_reward": -0.05498374626040459, + "yes_no_margin_reward": 2.875, + "token_credit_reward": -0.05498374626040459, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.04674245044589043, + 0.03260583430528641, + 0.01816197670996189, + 0.036952074617147446, + 0.09289474040269852, + 0.6884306073188782, + 0.022601600736379623, + 0.029018167406320572, + 0.018483376130461693, + 0.014109152369201183 + ], + "credit_map": [ + 0.00621327618137002, + 0.005506092682480812, + 0.005704070907086134, + 0.004571419674903154, + 0.009947706013917923, + 0.006166404578834772, + 0.005100148264318705, + 0.0035333349369466305, + 0.003357545007020235, + 0.003355122869834304, + 0.0036243100184947252, + 0.004259493201971054, + 0.0039009435568004847, + 0.005596253555268049, + 0.005025441292673349, + 0.003486722707748413, + 0.0019458612659946084, + 0.0020326075609773397, + 0.0024762521497905254, + 0.002083502011373639, + 0.002559299347922206, + 0.0030241666827350855, + 0.0020046618301421404, + 0.002035626210272312, + 0.006392260082066059, + 0.002791373524814844, + 0.00265705119818449, + 0.003759530605748296, + 0.004823518451303244, + 0.0070787519216537476, + 0.004675646312534809, + 0.004773943219333887, + 0.03336211293935776, + 0.0029726694338023663, + 0.004419212695211172, + 0.0032015731558203697, + 0.004288761876523495, + 0.03743622824549675, + 0.0024212163407355547, + 0.004792966414242983, + 0.26785150170326233, + 0.026662256568670273, + 0.01616617664694786, + 0.019712453708052635, + 0.025175374001264572, + 0.2904626131057739, + 0.014217198826372623, + 0.02818308025598526, + 0.003337140427902341, + 0.0036689280532300472, + 0.0036432344932109118, + 0.002621900988742709, + 0.0031458812300115824, + 0.0021470095962285995, + 0.0019329861970618367, + 0.002104521030560136, + 0.003102437127381563, + 0.0033150871749967337, + 0.0037663548719137907, + 0.0026724583003669977, + 0.004570958204567432, + 0.004074732307344675, + 0.003138671861961484, + 0.004377467557787895, + 0.0015061498852446675, + 0.0021336572244763374, + 0.0035146859008818865, + 0.003801798913627863, + 0.00118898821529001, + 0.002665000269189477, + 0.002006187569350004, + 0.0016669082688167691, + 0.0016893466236069798, + 0.001847522216849029, + 0.0008648679358884692, + 0.0010598916560411453, + 0.0032990293111652136, + 0.002041799249127507, + 0.001777445781044662, + 0.0015292501775547862 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 4, + 0 + ], + "selected_reward": -0.05498374626040459 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot handheld?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera angle slightly elevated?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the man wearing a faded, frayed dark blue Mao suit with ripped hems and cuffs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the man using a crude crutch carved from a tree branch to support himself?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the environment a muddy, rain-soaked country road flanked by bleak winter fields?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.029801562428474426 + ], + "opposite_logprob_tokens": [ + -3.529801607131958 + ], + "logprob_reward": -0.029801562428474426, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.029801562428474426, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.04326998069882393, + 0.03910031169652939, + 0.022797556594014168, + 0.017760522663593292, + 0.02864595502614975, + 0.7476269602775574, + 0.033009957522153854, + 0.03296905383467674, + 0.022882379591464996, + 0.011937251314520836 + ], + "credit_map": [ + 0.007251928560435772, + 0.005593045614659786, + 0.0037729765754193068, + 0.0036379913799464703, + 0.008282371796667576, + 0.005732018034905195, + 0.005119376815855503, + 0.003880272386595607, + 0.004167716484516859, + 0.0041563441045582294, + 0.004034995101392269, + 0.005548444576561451, + 0.004153045359998941, + 0.00688948156312108, + 0.005986230913549662, + 0.004164055921137333, + 0.0023456939961761236, + 0.002261912217363715, + 0.002278485568240285, + 0.0023329402320086956, + 0.004389449954032898, + 0.002801134018227458, + 0.0037896684370934963, + 0.0025982733350247145, + 0.0017712092958390713, + 0.0021451811771839857, + 0.0021920939907431602, + 0.0022391011007130146, + 0.002203593496233225, + 0.0026223447639495134, + 0.002075391123071313, + 0.002511608647182584, + 0.0074254549108445644, + 0.0017212809761986136, + 0.003194966819137335, + 0.002017434686422348, + 0.0022085884120315313, + 0.008336111903190613, + 0.001762464176863432, + 0.0019796511624008417, + 0.2886347472667694, + 0.030347514897584915, + 0.017818806692957878, + 0.022170813754200935, + 0.027559399604797363, + 0.3137145936489105, + 0.015512404032051563, + 0.031868673861026764, + 0.0049547916278243065, + 0.005725614260882139, + 0.00579325994476676, + 0.004141617566347122, + 0.0031386851333081722, + 0.002862296998500824, + 0.002656215336173773, + 0.0037374754901975393, + 0.0037839417345821857, + 0.003990145865827799, + 0.00408545508980751, + 0.00307381059974432, + 0.00462412228807807, + 0.004875363782048225, + 0.003610832616686821, + 0.004925382323563099, + 0.0017396421171724796, + 0.002284345217049122, + 0.004510543774813414, + 0.004857551772147417, + 0.0017447328427806497, + 0.0021756740752607584, + 0.0028079035691916943, + 0.0027619872707873583, + 0.001293681561946869, + 0.0014695842983201146, + 0.0006391727947629988, + 0.0009747813455760479, + 0.002612595446407795, + 0.001966958399862051, + 0.0019487456884235144, + 0.0010317320702597499 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 0, + 1 + ], + "selected_reward": -0.029801562428474426 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot handheld?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera angle slightly elevated?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the man wearing a faded, frayed dark blue Mao suit with ripped hems and cuffs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the man using a crude crutch carved from a tree branch to support himself?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the environment a muddy, rain-soaked country road flanked by bleak winter fields?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.009776213206350803 + ], + "opposite_logprob_tokens": [ + -4.6347761154174805 + ], + "logprob_reward": -0.009776213206350803, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.009776213206350803, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.040382202714681625, + 0.011954735964536667, + 0.013996911235153675, + 0.01968383975327015, + 0.09616947919130325, + 0.6985238790512085, + 0.022044170647859573, + 0.025463372468948364, + 0.06409114599227905, + 0.00769029650837183 + ], + "credit_map": [ + 0.008407378569245338, + 0.006995042320340872, + 0.003895652014762163, + 0.0026684794574975967, + 0.006029604934155941, + 0.005316917784512043, + 0.004077285062521696, + 0.002991842105984688, + 0.0011742680799216032, + 0.001081403810530901, + 0.0018722981913015246, + 0.0024215062148869038, + 0.0010406097862869501, + 0.0012652702862396836, + 0.0017699110321700573, + 0.0013294684467837214, + 0.0015521295135840774, + 0.0016172403702512383, + 0.0023235150147229433, + 0.0020291858818382025, + 0.0010793170658871531, + 0.002471490064635873, + 0.0013524601235985756, + 0.0015715720364823937, + 0.00562633341178298, + 0.0015455465763807297, + 0.0010598101653158665, + 0.0010761955054476857, + 0.0014180615544319153, + 0.00617259880527854, + 0.0013190075987949967, + 0.0014662867179140449, + 0.034349434077739716, + 0.002489423146471381, + 0.005824659019708633, + 0.0030912933871150017, + 0.004314026329666376, + 0.03882545605301857, + 0.0024878159165382385, + 0.004787383135408163, + 0.27298328280448914, + 0.02709418162703514, + 0.015677817165851593, + 0.020397428423166275, + 0.024312753230333328, + 0.29572564363479614, + 0.014803196303546429, + 0.027529554441571236, + 0.0056746359914541245, + 0.006203817669302225, + 0.0020746563095599413, + 0.0016988038551062346, + 0.0016396405408158898, + 0.001741433166898787, + 0.0015358497621491551, + 0.0014753325376659632, + 0.002457799855619669, + 0.0027792470064014196, + 0.004184023942798376, + 0.003635935252532363, + 0.003301463322713971, + 0.00324082700535655, + 0.0024995259009301662, + 0.0033645497169345617, + 0.00537642277777195, + 0.008677706122398376, + 0.009904691018164158, + 0.010454918257892132, + 0.004419007338583469, + 0.013173924759030342, + 0.005833683535456657, + 0.006250790320336819, + 0.001544595812447369, + 0.0010038502514362335, + 0.0005853246548213065, + 0.0006378605030477047, + 0.0013569785514846444, + 0.001040023984387517, + 0.0008261596667580307, + 0.000695502501912415 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 4, + 8 + ], + "selected_reward": -0.009776213206350803 + } + ], + "vlm_decoded_outputs": [ + "No", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "handheld camera, shot from a slightly elevated angle, follows a ragged elderly man as he walks down a muddy, rain-soaked country road. His sparse, short, graying hair is ruffled by the cold wind. His bronzed face is etched with deep, knife-like wrinkles, and his sunken eyes hold a gaze that is a mixture of fatigue and steeliness. He wears a faded, frayed dark blue Mao suit, the hems and cuffs ripped. He wears a gray shirt underneath, the collar yellowed, and his equally worn, dark trousers are splattered with damp mud. Under the gloomy evening sky, the thick clouds cast a faint, cold, scattered light, casting a chilling gray-brown hue on the entire scene. His steps were heavy and unsteady, each one sinking deeply into the sticky mud. Pulling out his stick was incredibly laborious, his body swaying from side to side, struggling to maintain balance with a crude crutch carved from a tree branch, tightly clenched in one hand. The road flanked by bleak winter fields, bare branches trembling slightly in the wind. The entire scene is imbued with a raw documentary quality, devoid of embellishment, focused solely on capturing every step of the old man's struggle against adversity, conveying a poignant, authentic, and resilient atmosphere.", + "seed": 11387, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000050/group_11/index_3100_seed_11387_dp_22_rank_44_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 11, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.26463526003062726, + "vlm_logprob": -0.26463526003062726, + "vlm_yes_no_margin": 2.3749999523162844, + "vlm_token_credit": -0.26463526003062726, + "vlm_yes_ratio": 0.8, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot handheld?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera angle slightly elevated?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the man wearing a faded, frayed dark blue Mao suit with ripped hems and cuffs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the man using a crude crutch carved from a tree branch to support himself?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the environment a muddy, rain-soaked country road flanked by bleak winter fields?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -1.0902893543243408 + ], + "opposite_logprob_tokens": [ + -0.4652893543243408 + ], + "logprob_reward": -1.0902893543243408, + "yes_no_margin_reward": -0.625, + "token_credit_reward": -1.0902893543243408, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.11232471466064453, + 0.15299151837825775, + 0.08461969345808029, + 0.04247725382447243, + 0.07454660534858704, + 0.02191903255879879, + 0.07448957860469818, + 0.0890892893075943, + 0.31027692556381226, + 0.03726537525653839 + ], + "credit_map": [ + 0.016029492020606995, + 0.013672999106347561, + 0.017267761752009392, + 0.00990288145840168, + 0.013935834169387817, + 0.01230406854301691, + 0.012181725353002548, + 0.01702995039522648, + 0.01914318837225437, + 0.015655113384127617, + 0.023276785388588905, + 0.019360912963747978, + 0.011781280860304832, + 0.01856081746518612, + 0.027903705835342407, + 0.017309708520770073, + 0.013621820136904716, + 0.01412933599203825, + 0.009090708568692207, + 0.009329400025308132, + 0.008707623928785324, + 0.013782518915832043, + 0.00834727007895708, + 0.007611012551933527, + 0.0035188808105885983, + 0.005413026548922062, + 0.007242174819111824, + 0.005745126400142908, + 0.004723128862679005, + 0.006310903932899237, + 0.005841691046953201, + 0.0036823186092078686, + 0.003776739351451397, + 0.0027021332643926144, + 0.0217786505818367, + 0.007370262872427702, + 0.003945766482502222, + 0.005561197176575661, + 0.008238724432885647, + 0.021173128858208656, + 0.0022166618146002293, + 0.0020165094174444675, + 0.003112038830295205, + 0.0025593750178813934, + 0.0022416329011321068, + 0.003973356913775206, + 0.0028500098269432783, + 0.0029494480695575476, + 0.009160913527011871, + 0.006823977455496788, + 0.013729199767112732, + 0.008029792457818985, + 0.007845292799174786, + 0.01215255819261074, + 0.00899492483586073, + 0.007752917241305113, + 0.013305403292179108, + 0.008401082828640938, + 0.026870429515838623, + 0.006946536246687174, + 0.003721862565726042, + 0.00887748971581459, + 0.010680444538593292, + 0.010286038741469383, + 0.033540256321430206, + 0.05259039252996445, + 0.05026029795408249, + 0.03405701741576195, + 0.025871003046631813, + 0.039948757737874985, + 0.03951534628868103, + 0.03449385613203049, + 0.0027978771831840277, + 0.001790520385839045, + 0.01114975567907095, + 0.003971816506236792, + 0.0025198988150805235, + 0.00423181289806962, + 0.005772254895418882, + 0.005031437613070011 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 1, + 0 + ], + "selected_reward": -1.0902893543243408 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot handheld?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera angle slightly elevated?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the man wearing a faded, frayed dark blue Mao suit with ripped hems and cuffs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the man using a crude crutch carved from a tree branch to support himself?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the environment a muddy, rain-soaked country road flanked by bleak winter fields?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.1428808569908142 + ], + "opposite_logprob_tokens": [ + -2.017880916595459 + ], + "logprob_reward": -0.1428808569908142, + "yes_no_margin_reward": 1.875, + "token_credit_reward": -0.1428808569908142, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09165846556425095, + 0.15128573775291443, + 0.07141715288162231, + 0.059878572821617126, + 0.18348173797130585, + 0.04912599176168442, + 0.18408025801181793, + 0.05332560837268829, + 0.08666584640741348, + 0.06908057630062103 + ], + "credit_map": [ + 0.015542363747954369, + 0.011854974552989006, + 0.012668185867369175, + 0.008378468453884125, + 0.009941058233380318, + 0.010630220174789429, + 0.01179674081504345, + 0.010846451856195927, + 0.019267817959189415, + 0.014243465848267078, + 0.026128241792321205, + 0.018556300550699234, + 0.012098322622478008, + 0.01747177541255951, + 0.026441778987646103, + 0.017078038305044174, + 0.010322240181267262, + 0.0055263531394302845, + 0.011829624883830547, + 0.005662182811647654, + 0.0043458095751702785, + 0.02203819528222084, + 0.005116392858326435, + 0.006576347630470991, + 0.007154493127018213, + 0.014132173731923103, + 0.0075346059165894985, + 0.0067295776680111885, + 0.004544203635305166, + 0.008190409280359745, + 0.007198637817054987, + 0.0043944744393229485, + 0.00958847627043724, + 0.01576833985745907, + 0.061101801693439484, + 0.01333257183432579, + 0.006164059042930603, + 0.008153888396918774, + 0.011036992073059082, + 0.05833561345934868, + 0.004147185944020748, + 0.005018498282879591, + 0.007768465671688318, + 0.00821612123399973, + 0.0058314925990998745, + 0.006687191314995289, + 0.0063063581474125385, + 0.005150677170604467, + 0.023337440565228462, + 0.01760118082165718, + 0.033104535192251205, + 0.019335227087140083, + 0.018622854724526405, + 0.02992641180753708, + 0.022126847878098488, + 0.020025771111249924, + 0.007635829504579306, + 0.009873210452497005, + 0.008908528834581375, + 0.0040709576569497585, + 0.005083664320409298, + 0.008735202252864838, + 0.004367835819721222, + 0.004650379531085491, + 0.007916472852230072, + 0.01276266947388649, + 0.015648113563656807, + 0.009490923024713993, + 0.00732285063713789, + 0.012089863419532776, + 0.010784704238176346, + 0.010650244541466236, + 0.006904406938701868, + 0.006190122105181217, + 0.02275564707815647, + 0.0052797128446400166, + 0.0029709667433053255, + 0.0059806834906339645, + 0.010911068879067898, + 0.008087972179055214 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 4, + 1 + ], + "selected_reward": -0.1428808569908142 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot handheld?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera angle slightly elevated?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the man wearing a faded, frayed dark blue Mao suit with ripped hems and cuffs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the man using a crude crutch carved from a tree branch to support himself?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the environment a muddy, rain-soaked country road flanked by bleak winter fields?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03812369331717491 + ], + "opposite_logprob_tokens": [ + -3.288123607635498 + ], + "logprob_reward": -0.03812369331717491, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.03812369331717491, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06984451413154602, + 0.03407808020710945, + 0.3531385362148285, + 0.021647276356816292, + 0.1854611188173294, + 0.020367279648780823, + 0.06057818979024887, + 0.09722074866294861, + 0.0782126784324646, + 0.07945165038108826 + ], + "credit_map": [ + 0.012714767828583717, + 0.008933153003454208, + 0.009040174074470997, + 0.004316368605941534, + 0.008115548640489578, + 0.00817064382135868, + 0.008389316499233246, + 0.010164541192352772, + 0.0031337114050984383, + 0.003453565528616309, + 0.004245288670063019, + 0.005555728450417519, + 0.004820151254534721, + 0.004091647919267416, + 0.005107259843498468, + 0.003670725040137768, + 0.05299489200115204, + 0.04490362107753754, + 0.0505286268889904, + 0.02756272256374359, + 0.02668062038719654, + 0.10105272382497787, + 0.02572144940495491, + 0.023693876340985298, + 0.0022706009913235903, + 0.005833406932651997, + 0.0022700990084558725, + 0.0026027378626167774, + 0.0017673899419605732, + 0.002777198562398553, + 0.0021146342623978853, + 0.002011209260672331, + 0.01028903666883707, + 0.005233032628893852, + 0.0655696839094162, + 0.009132667444646358, + 0.0053846039809286594, + 0.013005007989704609, + 0.011515560559928417, + 0.06533154845237732, + 0.0024544831831008196, + 0.0024906648322939873, + 0.003644016571342945, + 0.0021612464915961027, + 0.0028312942013144493, + 0.00216320320032537, + 0.0025148186832666397, + 0.0021075510885566473, + 0.008377605117857456, + 0.004074154421687126, + 0.015118780545890331, + 0.006318111438304186, + 0.004373034927994013, + 0.0069817365147173405, + 0.006139005534350872, + 0.009195759892463684, + 0.017680419608950615, + 0.013020377606153488, + 0.025193333625793457, + 0.007360961753875017, + 0.00586639903485775, + 0.008613476529717445, + 0.010229624807834625, + 0.009256157092750072, + 0.007572480011731386, + 0.0115664042532444, + 0.013972907327115536, + 0.008454916998744011, + 0.00585714727640152, + 0.01076639723032713, + 0.010033497586846352, + 0.009988931939005852, + 0.006483510136604309, + 0.002467042300850153, + 0.027166780084371567, + 0.006338599603623152, + 0.004105580039322376, + 0.007993790321052074, + 0.012936185114085674, + 0.011960159987211227 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 4, + 7 + ], + "selected_reward": -0.03812369331717491 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot handheld?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera angle slightly elevated?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the man wearing a faded, frayed dark blue Mao suit with ripped hems and cuffs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the man using a crude crutch carved from a tree branch to support himself?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the environment a muddy, rain-soaked country road flanked by bleak winter fields?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.033703286200761795 + ], + "opposite_logprob_tokens": [ + -3.408703327178955 + ], + "logprob_reward": -0.033703286200761795, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.033703286200761795, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12301196902990341, + 0.11926131695508957, + 0.05301624536514282, + 0.04724553972482681, + 0.15147413313388824, + 0.028866268694400787, + 0.05033281445503235, + 0.09878828376531601, + 0.21442684531211853, + 0.11357657611370087 + ], + "credit_map": [ + 0.024715978652238846, + 0.013955731876194477, + 0.016757356002926826, + 0.008168709464371204, + 0.013965239748358727, + 0.01265786588191986, + 0.012301256880164146, + 0.020489824935793877, + 0.01825081743299961, + 0.012595470994710922, + 0.01962187886238098, + 0.01338149607181549, + 0.008813886903226376, + 0.014301806688308716, + 0.01893998682498932, + 0.013355966657400131, + 0.00770487729460001, + 0.009742447175085545, + 0.005385758820921183, + 0.005655090324580669, + 0.005715025123208761, + 0.008253423497080803, + 0.0056656738743186, + 0.004893946927040815, + 0.004501239862293005, + 0.008346768096089363, + 0.006203965283930302, + 0.005565331783145666, + 0.00465977331623435, + 0.006175421178340912, + 0.006582377012819052, + 0.0052106622606515884, + 0.028587574139237404, + 0.00840790569782257, + 0.026382988318800926, + 0.01018991693854332, + 0.006209626328200102, + 0.0201099906116724, + 0.011047258973121643, + 0.04053886979818344, + 0.0035084541887044907, + 0.0019741614814847708, + 0.008023185655474663, + 0.002665124135091901, + 0.0021995941642671824, + 0.0029842841904610395, + 0.0034812421072274446, + 0.004030222073197365, + 0.006610308773815632, + 0.00320740370079875, + 0.010842092335224152, + 0.005445949733257294, + 0.00444769486784935, + 0.0065694693475961685, + 0.006396417040377855, + 0.006813481915742159, + 0.017489369958639145, + 0.01103421300649643, + 0.025821572169661522, + 0.007900063879787922, + 0.006314281839877367, + 0.009232726879417896, + 0.0110552329570055, + 0.009940831921994686, + 0.02187296748161316, + 0.03576737642288208, + 0.035998132079839706, + 0.022842474281787872, + 0.017444180324673653, + 0.0276817474514246, + 0.02642771787941456, + 0.026392264291644096, + 0.00898652896285057, + 0.0033967141062021255, + 0.03944046422839165, + 0.009597081691026688, + 0.005003131460398436, + 0.010651602409780025, + 0.019001876935362816, + 0.01749916933476925 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 4, + 0 + ], + "selected_reward": -0.033703286200761795 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot handheld?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera angle slightly elevated?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the man wearing a faded, frayed dark blue Mao suit with ripped hems and cuffs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the man using a crude crutch carved from a tree branch to support himself?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the environment a muddy, rain-soaked country road flanked by bleak winter fields?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.018179109320044518 + ], + "opposite_logprob_tokens": [ + -4.018178939819336 + ], + "logprob_reward": -0.018179109320044518, + "yes_no_margin_reward": 3.999999761581421, + "token_credit_reward": -0.018179109320044518, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07880149781703949, + 0.050237398594617844, + 0.031301338225603104, + 0.029067574068903923, + 0.0680864006280899, + 0.03185158222913742, + 0.04340880736708641, + 0.28599750995635986, + 0.3549460172653198, + 0.026301832869648933 + ], + "credit_map": [ + 0.009545770473778248, + 0.00974696408957243, + 0.013378026895225048, + 0.006413003895431757, + 0.00846501812338829, + 0.01085885800421238, + 0.007183996960520744, + 0.013209858909249306, + 0.004764862824231386, + 0.004515153355896473, + 0.00849637296050787, + 0.007184996269643307, + 0.004732951056212187, + 0.005844056606292725, + 0.009196890518069267, + 0.005502115935087204, + 0.003540276549756527, + 0.004508458077907562, + 0.003498292062431574, + 0.00312213902361691, + 0.002989300526678562, + 0.004783932585269213, + 0.0038889891002327204, + 0.004969949834048748, + 0.0035107028670608997, + 0.008234313689172268, + 0.002893040655180812, + 0.0033482539001852274, + 0.002077843062579632, + 0.004514534492045641, + 0.001979280263185501, + 0.0025096035096794367, + 0.0038794847205281258, + 0.001765023567713797, + 0.021992256864905357, + 0.0038447538390755653, + 0.0029877128545194864, + 0.006390292197465897, + 0.004874758422374725, + 0.022352123633027077, + 0.003262526588514447, + 0.002393991220742464, + 0.006254799664020538, + 0.002909950679168105, + 0.00235877581872046, + 0.006605604663491249, + 0.0026872812304645777, + 0.0053786528296768665, + 0.006214479450136423, + 0.002974682953208685, + 0.0103079192340374, + 0.0044105518609285355, + 0.003023858880624175, + 0.005452982150018215, + 0.004919208586215973, + 0.006105124019086361, + 0.05290903151035309, + 0.0390169583261013, + 0.07520551234483719, + 0.02236933447420597, + 0.013828625902533531, + 0.0262004341930151, + 0.028817154467105865, + 0.027650464326143265, + 0.0420226976275444, + 0.06573368608951569, + 0.052158333361148834, + 0.04074377194046974, + 0.030840441584587097, + 0.04238884523510933, + 0.045467790216207504, + 0.03559046611189842, + 0.003686662996187806, + 0.003987033851444721, + 0.0055606784299016, + 0.002314555225893855, + 0.0016339237336069345, + 0.003354107029736042, + 0.0027322284877300262, + 0.0030326417181640863 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 7, + 0 + ], + "selected_reward": -0.018179109320044518 + } + ], + "vlm_decoded_outputs": [ + "No", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "handheld camera, shot from a slightly elevated angle, follows a ragged elderly man as he walks down a muddy, rain-soaked country road. His sparse, short, graying hair is ruffled by the cold wind. His bronzed face is etched with deep, knife-like wrinkles, and his sunken eyes hold a gaze that is a mixture of fatigue and steeliness. He wears a faded, frayed dark blue Mao suit, the hems and cuffs ripped. He wears a gray shirt underneath, the collar yellowed, and his equally worn, dark trousers are splattered with damp mud. Under the gloomy evening sky, the thick clouds cast a faint, cold, scattered light, casting a chilling gray-brown hue on the entire scene. His steps were heavy and unsteady, each one sinking deeply into the sticky mud. Pulling out his stick was incredibly laborious, his body swaying from side to side, struggling to maintain balance with a crude crutch carved from a tree branch, tightly clenched in one hand. The road flanked by bleak winter fields, bare branches trembling slightly in the wind. The entire scene is imbued with a raw documentary quality, devoid of embellishment, focused solely on capturing every step of the old man's struggle against adversity, conveying a poignant, authentic, and resilient atmosphere.", + "seed": 11388, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000050/group_11/index_3100_seed_11388_dp_23_rank_46_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 11, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.28360764905810354, + "vlm_logprob": -0.28360764905810354, + "vlm_yes_no_margin": 1.925, + "vlm_token_credit": -0.28360764905810354, + "vlm_yes_ratio": 0.8, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot handheld?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera angle slightly elevated?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the man wearing a faded, frayed dark blue Mao suit with ripped hems and cuffs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the man using a crude crutch carved from a tree branch to support himself?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the environment a muddy, rain-soaked country road flanked by bleak winter fields?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -1.007197380065918 + ], + "opposite_logprob_tokens": [ + -0.507197380065918 + ], + "logprob_reward": -1.007197380065918, + "yes_no_margin_reward": -0.5, + "token_credit_reward": -1.007197380065918, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.11354798823595047, + 0.09236355125904083, + 0.02687820978462696, + 0.5462923049926758, + 0.019529322162270546, + 0.04581689089536667, + 0.04131747782230377, + 0.029762592166662216, + 0.05550912395119667, + 0.028982432559132576 + ], + "credit_map": [ + 0.015062381513416767, + 0.01996784843504429, + 0.02756204456090927, + 0.00915196631103754, + 0.008798860013484955, + 0.011533728800714016, + 0.010505563579499722, + 0.010965591296553612, + 0.031926725059747696, + 0.00991854164749384, + 0.012079467065632343, + 0.009509885683655739, + 0.0076532308012247086, + 0.005987330339848995, + 0.008848423138260841, + 0.006439949385821819, + 0.0024763874243944883, + 0.0020694672130048275, + 0.004116868134588003, + 0.005303127225488424, + 0.0033152734395116568, + 0.0017728053499013186, + 0.004386935383081436, + 0.003437344916164875, + 0.09092464298009872, + 0.09836363792419434, + 0.06933914124965668, + 0.10142862051725388, + 0.04422157630324364, + 0.044282976537942886, + 0.058565884828567505, + 0.03916585445404053, + 0.0017569728661328554, + 0.002455092268064618, + 0.0030508970376104116, + 0.0037554404698312283, + 0.0024218885228037834, + 0.0016896791057661176, + 0.002458114642649889, + 0.0019412358524277806, + 0.006286420859396458, + 0.0046262661926448345, + 0.007490441203117371, + 0.005655340384691954, + 0.005464601330459118, + 0.004530719947069883, + 0.006534123793244362, + 0.005228976719081402, + 0.004218791611492634, + 0.006040338892489672, + 0.008107469417154789, + 0.00920465961098671, + 0.002745406236499548, + 0.002628502668812871, + 0.005780031438916922, + 0.0025922805070877075, + 0.003861604956910014, + 0.004063138272613287, + 0.0036774701438844204, + 0.004389399196952581, + 0.004079598933458328, + 0.002916119759902358, + 0.004042874556034803, + 0.002732387976720929, + 0.005479668732732534, + 0.005687979515641928, + 0.010374943725764751, + 0.012051649391651154, + 0.0040076719596982, + 0.0038450188003480434, + 0.008854133076965809, + 0.0052080596797168255, + 0.0029155854135751724, + 0.004268813878297806, + 0.005277642980217934, + 0.005737795028835535, + 0.0018670611316338181, + 0.0017111749621108174, + 0.004049032460898161, + 0.0031553246080875397 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 0, + 1 + ], + "selected_reward": -1.007197380065918 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot handheld?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera angle slightly elevated?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the man wearing a faded, frayed dark blue Mao suit with ripped hems and cuffs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the man using a crude crutch carved from a tree branch to support himself?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the environment a muddy, rain-soaked country road flanked by bleak winter fields?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.2521447241306305 + ], + "opposite_logprob_tokens": [ + -1.502144694328308 + ], + "logprob_reward": -0.2521447241306305, + "yes_no_margin_reward": 1.25, + "token_credit_reward": -0.2521447241306305, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1405901461839676, + 0.07694224268198013, + 0.04110897704958916, + 0.08497565239667892, + 0.34803688526153564, + 0.044740475714206696, + 0.11539436131715775, + 0.1160568818449974, + 0.021655235439538956, + 0.010499104857444763 + ], + "credit_map": [ + 0.00871129147708416, + 0.02979610674083233, + 0.045711226761341095, + 0.008915597572922707, + 0.008962550200521946, + 0.013111905194818974, + 0.01578187569975853, + 0.00959958415478468, + 0.02177765779197216, + 0.007574550341814756, + 0.00991507526487112, + 0.011381582356989384, + 0.006831645965576172, + 0.0055380770936608315, + 0.007818673737347126, + 0.006104984320700169, + 0.004104951396584511, + 0.004722789395600557, + 0.010628538206219673, + 0.005354573018848896, + 0.004190171603113413, + 0.002698327647522092, + 0.0059968335554003716, + 0.003412790596485138, + 0.01122382190078497, + 0.014086621813476086, + 0.009378417395055294, + 0.013231183402240276, + 0.010264134034514427, + 0.009731869213283062, + 0.00950491614639759, + 0.007554689422249794, + 0.028246872127056122, + 0.07220029085874557, + 0.12277102470397949, + 0.04707594960927963, + 0.015050511807203293, + 0.012198302894830704, + 0.036877214908599854, + 0.01361672580242157, + 0.003912113606929779, + 0.003034060588106513, + 0.006856876891106367, + 0.007112039718776941, + 0.003499670885503292, + 0.004437888041138649, + 0.006944750901311636, + 0.008943069726228714, + 0.012259134091436863, + 0.010864506475627422, + 0.013918058946728706, + 0.025441734120249748, + 0.009249435737729073, + 0.007325157057493925, + 0.017188778147101402, + 0.01914754882454872, + 0.015960566699504852, + 0.01635137014091015, + 0.013048230670392513, + 0.01583750732243061, + 0.018746353685855865, + 0.009279945865273476, + 0.014979315921664238, + 0.011853588744997978, + 0.002166189718991518, + 0.00244948361068964, + 0.004043707624077797, + 0.004636158235371113, + 0.001616911729797721, + 0.0013519866624847054, + 0.0033802983816713095, + 0.002010500989854336, + 0.0010813307017087936, + 0.0016328321071341634, + 0.0013465903466567397, + 0.0018378272652626038, + 0.0009941714815795422, + 0.0007011210545897484, + 0.0015375526854768395, + 0.0013676793314516544 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 0, + 7 + ], + "selected_reward": -0.2521447241306305 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot handheld?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera angle slightly elevated?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the man wearing a faded, frayed dark blue Mao suit with ripped hems and cuffs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the man using a crude crutch carved from a tree branch to support himself?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the environment a muddy, rain-soaked country road flanked by bleak winter fields?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07004415988922119 + ], + "opposite_logprob_tokens": [ + -2.6950440406799316 + ], + "logprob_reward": -0.07004415988922119, + "yes_no_margin_reward": 2.625, + "token_credit_reward": -0.07004415988922119, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.33566367626190186, + 0.06700599938631058, + 0.08178943395614624, + 0.0889713242650032, + 0.167348712682724, + 0.03230715170502663, + 0.0354725681245327, + 0.0812380462884903, + 0.07947643101215363, + 0.030726667493581772 + ], + "credit_map": [ + 0.01674056053161621, + 0.07016003131866455, + 0.11786258965730667, + 0.027241166681051254, + 0.019064484164118767, + 0.023275502026081085, + 0.0406484454870224, + 0.020670881494879723, + 0.007532912772148848, + 0.005473453551530838, + 0.008490610867738724, + 0.011465341784060001, + 0.010990549810230732, + 0.006718970835208893, + 0.008999928832054138, + 0.007334230933338404, + 0.009350776672363281, + 0.009779784828424454, + 0.0147200096398592, + 0.012936120852828026, + 0.009672471322119236, + 0.00538899190723896, + 0.012737526558339596, + 0.007203750777989626, + 0.012878164649009705, + 0.01669028028845787, + 0.010673373937606812, + 0.016381075605750084, + 0.007992574013769627, + 0.008993381634354591, + 0.00971168465912342, + 0.005650787614285946, + 0.015109016560018063, + 0.03739432245492935, + 0.05271416902542114, + 0.02122040092945099, + 0.009208139032125473, + 0.007756352424621582, + 0.016764914616942406, + 0.007181390654295683, + 0.0034162404481321573, + 0.0033156576100736856, + 0.0032311794348061085, + 0.004942362196743488, + 0.00396057078614831, + 0.0036144328769296408, + 0.005286070518195629, + 0.004540635272860527, + 0.00452674413099885, + 0.005542968865483999, + 0.005432560574263334, + 0.0063470941968262196, + 0.0030260293278843164, + 0.0020688059739768505, + 0.004501291085034609, + 0.004027071874588728, + 0.010901181027293205, + 0.015643134713172913, + 0.012354844249784946, + 0.010266339406371117, + 0.008021614514291286, + 0.007713616825640202, + 0.00883975438773632, + 0.007497555110603571, + 0.00896076112985611, + 0.013984917663037777, + 0.012385173700749874, + 0.015063205733895302, + 0.005689444951713085, + 0.006263652816414833, + 0.009821881540119648, + 0.007307389285415411, + 0.003586822422221303, + 0.0055618067272007465, + 0.004448247142136097, + 0.004140061791986227, + 0.0026906721759587526, + 0.0019922354258596897, + 0.004663374274969101, + 0.003643446136265993 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 3 + ], + "selected_reward": -0.07004415988922119 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot handheld?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera angle slightly elevated?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the man wearing a faded, frayed dark blue Mao suit with ripped hems and cuffs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the man using a crude crutch carved from a tree branch to support himself?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the environment a muddy, rain-soaked country road flanked by bleak winter fields?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.05496298521757126 + ], + "opposite_logprob_tokens": [ + -2.9299628734588623 + ], + "logprob_reward": -0.05496298521757126, + "yes_no_margin_reward": 2.875, + "token_credit_reward": -0.05496298521757126, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3651449978351593, + 0.07411644607782364, + 0.06669489294290543, + 0.06447640806436539, + 0.16863185167312622, + 0.04128677397966385, + 0.03129298985004425, + 0.06052979826927185, + 0.09884563088417053, + 0.02898022159934044 + ], + "credit_map": [ + 0.023847196251153946, + 0.08196928352117538, + 0.11926525086164474, + 0.027968259528279305, + 0.018817350268363953, + 0.029002025723457336, + 0.04160058870911598, + 0.02267501875758171, + 0.008753273636102676, + 0.0074575054459273815, + 0.008998113684356213, + 0.011959903873503208, + 0.012562007643282413, + 0.007853981107473373, + 0.009941024705767632, + 0.006590632256120443, + 0.007486524526029825, + 0.008496236056089401, + 0.010414162650704384, + 0.01232922077178955, + 0.00762877007946372, + 0.004892105236649513, + 0.010086902417242527, + 0.0053609698079526424, + 0.00633361516520381, + 0.005590346641838551, + 0.008598974905908108, + 0.015224874950945377, + 0.007031196728348732, + 0.00757839297875762, + 0.008051694370806217, + 0.006067310925573111, + 0.014170759357511997, + 0.03659157454967499, + 0.05556482821702957, + 0.022324029356241226, + 0.008948521688580513, + 0.00682376092299819, + 0.017112972214818, + 0.00709539232775569, + 0.004381862003356218, + 0.004391932860016823, + 0.004321492277085781, + 0.006005208473652601, + 0.004189028404653072, + 0.005293001886457205, + 0.006209858693182468, + 0.0064943875186145306, + 0.0034118525218218565, + 0.0031067021191120148, + 0.005989750847220421, + 0.004848067183047533, + 0.0026579913683235645, + 0.002249633427709341, + 0.004295071586966515, + 0.004733920097351074, + 0.008288564160466194, + 0.011395891197025776, + 0.009487305767834187, + 0.0072839148342609406, + 0.005594293586909771, + 0.005472741089761257, + 0.006813923362642527, + 0.006193157751113176, + 0.010891328565776348, + 0.01643957756459713, + 0.01814848929643631, + 0.0197137463837862, + 0.006989629007875919, + 0.007449593394994736, + 0.012261484749615192, + 0.006951778195798397, + 0.0024981924798339605, + 0.0028786782640963793, + 0.0028326630126684904, + 0.0034430809319019318, + 0.0025196210481226444, + 0.002363261766731739, + 0.004872036166489124, + 0.0075726862996816635 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 8 + ], + "selected_reward": -0.05496298521757126 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot handheld?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera angle slightly elevated?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the man wearing a faded, frayed dark blue Mao suit with ripped hems and cuffs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the man using a crude crutch carved from a tree branch to support himself?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the environment a muddy, rain-soaked country road flanked by bleak winter fields?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.033688995987176895 + ], + "opposite_logprob_tokens": [ + -3.408689022064209 + ], + "logprob_reward": -0.033688995987176895, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.033688995987176895, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1072930321097374, + 0.0730794370174408, + 0.0667889192700386, + 0.0495375320315361, + 0.1450185924768448, + 0.031002113595604897, + 0.0489947609603405, + 0.3112149238586426, + 0.13993728160858154, + 0.02713337354362011 + ], + "credit_map": [ + 0.010168618522584438, + 0.020244380459189415, + 0.017841607332229614, + 0.011205127462744713, + 0.011143580079078674, + 0.014743134379386902, + 0.010863141156733036, + 0.011083442717790604, + 0.018836526200175285, + 0.009081578813493252, + 0.007801502011716366, + 0.009171958081424236, + 0.00819883868098259, + 0.005736186169087887, + 0.008071385324001312, + 0.006181471515446901, + 0.006272573955357075, + 0.006187160033732653, + 0.014878024347126484, + 0.011442522518336773, + 0.005558862816542387, + 0.004216797184199095, + 0.010198836214840412, + 0.008034148253500462, + 0.0034355991519987583, + 0.004225796554237604, + 0.007286870386451483, + 0.012666990980505943, + 0.0066832490265369415, + 0.004838673397898674, + 0.00663980795070529, + 0.0037605492398142815, + 0.012920832261443138, + 0.032056424766778946, + 0.048361167311668396, + 0.019965751096606255, + 0.0065817018039524555, + 0.00480695441365242, + 0.014328796416521072, + 0.005996973719447851, + 0.0031242044642567635, + 0.0029582679271698, + 0.005009328480809927, + 0.006000263150781393, + 0.003114428836852312, + 0.0031182318925857544, + 0.004030844662338495, + 0.003646550001576543, + 0.005433711688965559, + 0.0059926146641373634, + 0.010555315762758255, + 0.008183007128536701, + 0.003813741961494088, + 0.003966102842241526, + 0.006428242195397615, + 0.004622024949640036, + 0.04050469398498535, + 0.06056564301252365, + 0.0500793531537056, + 0.03994198516011238, + 0.02738972380757332, + 0.028549322858452797, + 0.03456365317106247, + 0.029620572924613953, + 0.01916164718568325, + 0.03162603825330734, + 0.02160954289138317, + 0.019415779039263725, + 0.008677476085722446, + 0.0103353476151824, + 0.015681805089116096, + 0.013429665938019753, + 0.0025485947262495756, + 0.002439944539219141, + 0.005353838205337524, + 0.006184098776429892, + 0.0014858981594443321, + 0.0015078486176207662, + 0.004271921701729298, + 0.0033412303309887648 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 4, + 8 + ], + "selected_reward": -0.033688995987176895 + } + ], + "vlm_decoded_outputs": [ + "No", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "handheld camera, shot from a slightly elevated angle, follows a ragged elderly man as he walks down a muddy, rain-soaked country road. His sparse, short, graying hair is ruffled by the cold wind. His bronzed face is etched with deep, knife-like wrinkles, and his sunken eyes hold a gaze that is a mixture of fatigue and steeliness. He wears a faded, frayed dark blue Mao suit, the hems and cuffs ripped. He wears a gray shirt underneath, the collar yellowed, and his equally worn, dark trousers are splattered with damp mud. Under the gloomy evening sky, the thick clouds cast a faint, cold, scattered light, casting a chilling gray-brown hue on the entire scene. His steps were heavy and unsteady, each one sinking deeply into the sticky mud. Pulling out his stick was incredibly laborious, his body swaying from side to side, struggling to maintain balance with a crude crutch carved from a tree branch, tightly clenched in one hand. The road flanked by bleak winter fields, bare branches trembling slightly in the wind. The entire scene is imbued with a raw documentary quality, devoid of embellishment, focused solely on capturing every step of the old man's struggle against adversity, conveying a poignant, authentic, and resilient atmosphere.", + "seed": 11389, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000050/group_11/index_3100_seed_11389_dp_23_rank_46_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 11, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.24877372980117798, + "vlm_logprob": -0.24877372980117798, + "vlm_yes_no_margin": 2.575, + "vlm_token_credit": -0.24877372980117798, + "vlm_yes_ratio": 0.8, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot handheld?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera angle slightly elevated?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the man wearing a faded, frayed dark blue Mao suit with ripped hems and cuffs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the man using a crude crutch carved from a tree branch to support himself?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the environment a muddy, rain-soaked country road flanked by bleak winter fields?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -1.0136171579360962 + ], + "opposite_logprob_tokens": [ + -0.5136171579360962 + ], + "logprob_reward": -1.0136171579360962, + "yes_no_margin_reward": -0.5, + "token_credit_reward": -1.0136171579360962, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.06145260110497475, + 0.0343577042222023, + 0.021042611449956894, + 0.015213433653116226, + 0.03888804465532303, + 0.01615285500884056, + 0.14878657460212708, + 0.014233390800654888, + 0.6324973702430725, + 0.017375420778989792 + ], + "credit_map": [ + 0.006781008560210466, + 0.00988674908876419, + 0.006362579297274351, + 0.0071190171875059605, + 0.009874368086457253, + 0.006639839615672827, + 0.006873889360576868, + 0.007915153168141842, + 0.005140900146216154, + 0.004650917835533619, + 0.0039079878479242325, + 0.00473219808191061, + 0.002998586278408766, + 0.0041123609989881516, + 0.004046202637255192, + 0.004768548998981714, + 0.0032087890431284904, + 0.0019932363647967577, + 0.0037182860542088747, + 0.0031765878666192293, + 0.003047192469239235, + 0.0019127367995679379, + 0.0017890891758725047, + 0.002196692395955324, + 0.002238776534795761, + 0.0016477532917633653, + 0.0025089343544095755, + 0.0022180990781635046, + 0.001662547467276454, + 0.0015333176124840975, + 0.0016161918174475431, + 0.0017878132639452815, + 0.011087746359407902, + 0.004021979868412018, + 0.0032647191546857357, + 0.001521010883152485, + 0.0017662571044638753, + 0.013688700273633003, + 0.0017390860011801124, + 0.0017985421000048518, + 0.0018282162491232157, + 0.0013029451947659254, + 0.00143929582554847, + 0.0021237123291939497, + 0.0028575861360877752, + 0.0015517655992880464, + 0.002756608184427023, + 0.002292725956067443, + 0.019170818850398064, + 0.01794368214905262, + 0.016756299883127213, + 0.018044598400592804, + 0.02346038818359375, + 0.021318722516298294, + 0.018542492762207985, + 0.013549563474953175, + 0.0014920954126864672, + 0.001129064941778779, + 0.002136270748451352, + 0.0019181236857548356, + 0.0025645035784691572, + 0.001555239432491362, + 0.0017708237282931805, + 0.0016672698548063636, + 0.17800244688987732, + 0.025864433497190475, + 0.04825742170214653, + 0.03345973789691925, + 0.040657274425029755, + 0.20231585204601288, + 0.04083646461367607, + 0.06310375034809113, + 0.0015960106393322349, + 0.0016927864635363221, + 0.0028250610921531916, + 0.0023588340263813734, + 0.0027211536653339863, + 0.0019481915514916182, + 0.00209444179199636, + 0.002138940617442131 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 6, + 0 + ], + "selected_reward": -1.0136171579360962 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot handheld?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera angle slightly elevated?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the man wearing a faded, frayed dark blue Mao suit with ripped hems and cuffs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the man using a crude crutch carved from a tree branch to support himself?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the environment a muddy, rain-soaked country road flanked by bleak winter fields?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.160408616065979 + ], + "opposite_logprob_tokens": [ + -1.910408616065979 + ], + "logprob_reward": -0.160408616065979, + "yes_no_margin_reward": 1.75, + "token_credit_reward": -0.160408616065979, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09814746677875519, + 0.028107130900025368, + 0.0294511541724205, + 0.044100116938352585, + 0.024743782356381416, + 0.021199746057391167, + 0.08496686816215515, + 0.015510418452322483, + 0.6223630309104919, + 0.03141026571393013 + ], + "credit_map": [ + 0.014860186725854874, + 0.012528442777693272, + 0.009890156798064709, + 0.010494046844542027, + 0.011649535968899727, + 0.011649529449641705, + 0.009180727414786816, + 0.017894839867949486, + 0.003268864471465349, + 0.004894566256552935, + 0.0035756367724388838, + 0.0026238032151013613, + 0.003860668744891882, + 0.002616531914100051, + 0.0028509406838566065, + 0.004416117910295725, + 0.0037336554378271103, + 0.0032433359883725643, + 0.0033396901562809944, + 0.00421806750819087, + 0.005014220252633095, + 0.0026306970976293087, + 0.0035031840670853853, + 0.0037683025002479553, + 0.007421600632369518, + 0.004616767633706331, + 0.006347618997097015, + 0.005433517973870039, + 0.0041268509812653065, + 0.005617067217826843, + 0.004611257929354906, + 0.005925433710217476, + 0.0035819278564304113, + 0.0031540615018457174, + 0.0037908582016825676, + 0.0024384355638176203, + 0.002474820474162698, + 0.0030445712618529797, + 0.0033623240888118744, + 0.002896784571930766, + 0.003044445998966694, + 0.001572518376633525, + 0.0025511030107736588, + 0.0022348975762724876, + 0.002656305907294154, + 0.0030886484310030937, + 0.003028734354302287, + 0.0030230912379920483, + 0.00762081379070878, + 0.006529416423290968, + 0.016625704243779182, + 0.011164399795234203, + 0.009224864654242992, + 0.01082881074398756, + 0.01124664954841137, + 0.011726204305887222, + 0.0016644465504214168, + 0.0009041840676218271, + 0.0029013887979090214, + 0.0022916258312761784, + 0.0017661361489444971, + 0.0015014377422630787, + 0.001935575157403946, + 0.002545624505728483, + 0.18484854698181152, + 0.026024289429187775, + 0.03537436202168465, + 0.03193827345967293, + 0.03874633088707924, + 0.2149164378643036, + 0.0365394651889801, + 0.053975291550159454, + 0.002262091962620616, + 0.00210928893648088, + 0.006187337450683117, + 0.004125582054257393, + 0.0031594180036336184, + 0.002339929109439254, + 0.0031579581554979086, + 0.008068657480180264 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 0, + 6 + ], + "selected_reward": -0.160408616065979 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot handheld?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera angle slightly elevated?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the man wearing a faded, frayed dark blue Mao suit with ripped hems and cuffs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the man using a crude crutch carved from a tree branch to support himself?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the environment a muddy, rain-soaked country road flanked by bleak winter fields?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.026369880884885788 + ], + "opposite_logprob_tokens": [ + -3.65136981010437 + ], + "logprob_reward": -0.026369880884885788, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.026369880884885788, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0830361470580101, + 0.09453046321868896, + 0.04759743809700012, + 0.04794846102595329, + 0.03682779520750046, + 0.060189638286828995, + 0.03928174823522568, + 0.03597796708345413, + 0.5044006705284119, + 0.05020967498421669 + ], + "credit_map": [ + 0.008300374262034893, + 0.013581559993326664, + 0.00925259105861187, + 0.010294586420059204, + 0.008353992365300655, + 0.0076143029145896435, + 0.007656577974557877, + 0.01798216626048088, + 0.01110140047967434, + 0.015830151736736298, + 0.011425672098994255, + 0.012335256673395634, + 0.009930864907801151, + 0.01000774186104536, + 0.012567398138344288, + 0.011331981979310513, + 0.005001184996217489, + 0.003574591362848878, + 0.013720519840717316, + 0.004777324851602316, + 0.006190174259245396, + 0.003705100854858756, + 0.0052580395713448524, + 0.005370506085455418, + 0.0057701533660292625, + 0.003506228793412447, + 0.008272807113826275, + 0.005022065714001656, + 0.007034933660179377, + 0.004998863209038973, + 0.005488129332661629, + 0.007855279371142387, + 0.009011122398078442, + 0.0025558758061379194, + 0.004359133541584015, + 0.0029460894875228405, + 0.0024670599959790707, + 0.009540684521198273, + 0.003567079547792673, + 0.0023807475809007883, + 0.007400611415505409, + 0.004849803168326616, + 0.009221181273460388, + 0.007412857376039028, + 0.006964696571230888, + 0.007730627432465553, + 0.008388170972466469, + 0.008221692405641079, + 0.003947568591684103, + 0.0031232547480612993, + 0.00483830226585269, + 0.007142957765609026, + 0.0063764736987650394, + 0.003664045361801982, + 0.00585758313536644, + 0.0043315589427948, + 0.003488334594294429, + 0.0031581528019160032, + 0.0042865704745054245, + 0.006240141112357378, + 0.005417341832071543, + 0.003633701940998435, + 0.005405051168054342, + 0.004348672926425934, + 0.14885105192661285, + 0.020871274173259735, + 0.028905140236020088, + 0.0271566454321146, + 0.03367720916867256, + 0.1733381301164627, + 0.028175299987196922, + 0.04342592507600784, + 0.004075820557773113, + 0.007108782883733511, + 0.007257123477756977, + 0.008472703397274017, + 0.005675503518432379, + 0.00487576425075531, + 0.004771376959979534, + 0.007972599938511848 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 1, + 0 + ], + "selected_reward": -0.026369880884885788 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot handheld?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera angle slightly elevated?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the man wearing a faded, frayed dark blue Mao suit with ripped hems and cuffs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the man using a crude crutch carved from a tree branch to support himself?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the environment a muddy, rain-soaked country road flanked by bleak winter fields?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.033698905259370804 + ], + "opposite_logprob_tokens": [ + -3.408698797225952 + ], + "logprob_reward": -0.033698905259370804, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.033698905259370804, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0472649410367012, + 0.07965006679296494, + 0.027050592005252838, + 0.03201138228178024, + 0.01623729057610035, + 0.029727540910243988, + 0.024408338591456413, + 0.0495733916759491, + 0.6768344044685364, + 0.01724201999604702 + ], + "credit_map": [ + 0.004826199263334274, + 0.005953535437583923, + 0.0045280312187969685, + 0.006583390291780233, + 0.005321463104337454, + 0.004220783710479736, + 0.0048158722929656506, + 0.011015667580068111, + 0.009279905818402767, + 0.012153299525380135, + 0.01006148662418127, + 0.009498132392764091, + 0.008670607581734657, + 0.0076280501671135426, + 0.012717577628791332, + 0.009641007520258427, + 0.002706527942791581, + 0.002518806606531143, + 0.006547247059643269, + 0.0038804432842880487, + 0.003397915745154023, + 0.0021063589956611395, + 0.0027553876861929893, + 0.003137908410280943, + 0.004551176447421312, + 0.002822755603119731, + 0.004751700907945633, + 0.004470157437026501, + 0.004085048567503691, + 0.003989186137914658, + 0.0037150646094232798, + 0.0036262939684093, + 0.0031044676434248686, + 0.0012629716657102108, + 0.0019748718477785587, + 0.0018241774523630738, + 0.0014356455067172647, + 0.0033170124515891075, + 0.0016861080657690763, + 0.0016320364084094763, + 0.0038221953436732292, + 0.0032027713023126125, + 0.004467937164008617, + 0.0032101504039019346, + 0.004314733203500509, + 0.003773060627281666, + 0.0031484896317124367, + 0.003788204863667488, + 0.002160565694794059, + 0.002313718432560563, + 0.0035718234721571207, + 0.003333866596221924, + 0.0030512765515595675, + 0.003233451396226883, + 0.002811593934893608, + 0.003932042978703976, + 0.005514263641089201, + 0.0035619859118014574, + 0.0050160824321210384, + 0.009080386720597744, + 0.007302506826817989, + 0.005421403329819441, + 0.008064753375947475, + 0.0056120119988918304, + 0.19607117772102356, + 0.02898186631500721, + 0.03789525106549263, + 0.03481126204133034, + 0.04381304234266281, + 0.22716189920902252, + 0.0411858893930912, + 0.06691403687000275, + 0.0012780301040038466, + 0.0024838922545313835, + 0.0020579807460308075, + 0.003009611973538995, + 0.002207936020568013, + 0.0019269549520686269, + 0.002146056853234768, + 0.002131558023393154 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 1, + 7 + ], + "selected_reward": -0.033698905259370804 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot handheld?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera angle slightly elevated?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the man wearing a faded, frayed dark blue Mao suit with ripped hems and cuffs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the man using a crude crutch carved from a tree branch to support himself?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the environment a muddy, rain-soaked country road flanked by bleak winter fields?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.009774088859558105 + ], + "opposite_logprob_tokens": [ + -4.634774208068848 + ], + "logprob_reward": -0.009774088859558105, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.009774088859558105, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.026375658810138702, + 0.015905793756246567, + 0.023503480479121208, + 0.017454486340284348, + 0.017981408163905144, + 0.01588357426226139, + 0.018292702734470367, + 0.018839318305253983, + 0.8344765305519104, + 0.01128700003027916 + ], + "credit_map": [ + 0.003682712558656931, + 0.00362570327706635, + 0.0035951449535787106, + 0.0031102090142667294, + 0.002630286617204547, + 0.0029103998094797134, + 0.002413557143881917, + 0.0044076452031731606, + 0.0023589155171066523, + 0.002049614442512393, + 0.001932975952513516, + 0.0019920358899980783, + 0.0016010949620977044, + 0.0018279992509633303, + 0.0016920927446335554, + 0.002451063599437475, + 0.0044738841243088245, + 0.002389251720160246, + 0.0043828957714140415, + 0.002433237386867404, + 0.0022860937751829624, + 0.003555961651727557, + 0.0015498200664296746, + 0.002432335866615176, + 0.0038634883239865303, + 0.0018412591889500618, + 0.003371843136847019, + 0.0014404484536498785, + 0.0011921764817088842, + 0.0022606244310736656, + 0.0014426399720832705, + 0.0020420062355697155, + 0.00422794371843338, + 0.002324327128008008, + 0.001499033416621387, + 0.0009870553622022271, + 0.0007352493121288717, + 0.005891967564821243, + 0.00107431726064533, + 0.0012415152741596103, + 0.0019506344106048346, + 0.0012998791644349694, + 0.0030523098539561033, + 0.001797487842850387, + 0.0016570446314290166, + 0.0021265510004013777, + 0.0018865286838263273, + 0.002113139256834984, + 0.0019272065255790949, + 0.0015562737826257944, + 0.004569195676594973, + 0.0021060581784695387, + 0.0018725825939327478, + 0.0019240091787651181, + 0.0020585868041962385, + 0.002278788946568966, + 0.002875648206099868, + 0.001349105965346098, + 0.0024624154902994633, + 0.0022530413698405027, + 0.0015988359227776527, + 0.0036507511977106333, + 0.002178444294258952, + 0.0024710772559046745, + 0.23641349375247955, + 0.034796830266714096, + 0.05944017693400383, + 0.045501958578825, + 0.05187726765871048, + 0.27115052938461304, + 0.052836984395980835, + 0.08245929330587387, + 0.0007444875664077699, + 0.0011401331285014749, + 0.002058160724118352, + 0.0015233753947541118, + 0.0012202983489260077, + 0.0009147952077910304, + 0.00119765754789114, + 0.0024880922865122557 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 0, + 2 + ], + "selected_reward": -0.009774088859558105 + } + ], + "vlm_decoded_outputs": [ + "No", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000050/group_12/index_tensor([8904, 8904])_group_12_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000050/group_12/index_tensor([8904, 8904])_group_12_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..4143f1b1ee7a40060fbd063b1c9a50c7aa8f42af --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000050/group_12/index_tensor([8904, 8904])_group_12_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "In a dynamic, low-angle, long shot, the camera pans steadily, following a young, athletic male with short, wind-twisted hair as he rides a mountain bike through a vast, grassy park at night. He wears a thick, dark wool sweater and black, full-finger performance gloves for a secure grip on the handlebars. He pedals at a steady, moderate pace, but his body and bike are subjected to constant, sharp, and erratic motion as he expertly navigates the uneven ground. The bike itself is a sturdy model with a dark frame and thick, knobby tires that bounce and jostle over the terrain. The entire scene is bathed in a dramatic, high-contrast golden glow cast by tall, unseen park lights, which carve the rider's silhouette out of the darkness and cast long, twisting shadows that stretch and race behind him. A strong wind whips through the tall grass, subtly tugging at the fabric of his sweater, heightening the sense of solitary effort and creating an atmosphere of tension, focus, and visually arresting action against the inky night sky.", + "seed": 5949, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000050/group_12/index_8904_seed_5949_dp_24_rank_48_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 12, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.03528119544498622, + "vlm_logprob": -0.03528119544498622, + "vlm_yes_no_margin": 3.775, + "vlm_token_credit": -0.03528119544498622, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a young, athletic male riding a mountain bike?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene set in a vast, grassy park at night?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the rider wearing a thick, dark wool sweater and black, full-finger performance gloves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene illuminated by a high-contrast golden glow from unseen park lights?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera pan steadily while following the rider in a low-angle, long shot?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07742787152528763 + ], + "opposite_logprob_tokens": [ + -2.702427864074707 + ], + "logprob_reward": -0.07742787152528763, + "yes_no_margin_reward": 2.625, + "token_credit_reward": -0.07742787152528763, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.6654517650604248, + 0.06739600002765656, + 0.057868558913469315, + 0.017218420282006264, + 0.020753873512148857, + 0.05624628812074661, + 0.02580609731376171, + 0.03502461686730385, + 0.02805302105844021, + 0.02618134208023548 + ], + "credit_map": [ + 0.06553247570991516, + 0.037181247025728226, + 0.02980673499405384, + 0.021597104147076607, + 0.25205668807029724, + 0.20225977897644043, + 0.028802962973713875, + 0.028214728459715843, + 0.006206423044204712, + 0.009269037283957005, + 0.009276357479393482, + 0.004171435721218586, + 0.011939778923988342, + 0.015164355747401714, + 0.006053525023162365, + 0.005315085873007774, + 0.0030100056901574135, + 0.006368571426719427, + 0.006878263782709837, + 0.0025573247112333775, + 0.017391713336110115, + 0.011978385969996452, + 0.00452367402613163, + 0.0051606204360723495, + 0.0016854786081239581, + 0.0026809496339410543, + 0.0016801090678200126, + 0.002405311679467559, + 0.002853917423635721, + 0.002043413231149316, + 0.0020392807200551033, + 0.0018299599178135395, + 0.0024548214860260487, + 0.0024201262276619673, + 0.0029970204923301935, + 0.002013924065977335, + 0.003202660707756877, + 0.0033794823102653027, + 0.00174056482501328, + 0.002545273397117853, + 0.0032189853955060244, + 0.010203152894973755, + 0.00496572395786643, + 0.006804646458476782, + 0.012737068347632885, + 0.010206623002886772, + 0.004267487209290266, + 0.003842598758637905, + 0.001905428827740252, + 0.003253164002671838, + 0.0030460888519883156, + 0.0028718316461890936, + 0.0029813176952302456, + 0.002812504069879651, + 0.004090484697371721, + 0.004845275077968836, + 0.003067602403461933, + 0.00584619864821434, + 0.004563214257359505, + 0.004379403777420521, + 0.003907416015863419, + 0.0031664480920881033, + 0.005099312402307987, + 0.004995022434741259, + 0.0031181888189166784, + 0.004808854777365923, + 0.0031433922704309225, + 0.0027676925528794527, + 0.004482018295675516, + 0.003276639385148883, + 0.0036749811843037605, + 0.002781255403533578, + 0.0033851901534944773, + 0.003462101798504591, + 0.00289241224527359, + 0.0050504570826888084, + 0.0019452456617727876, + 0.00447137700393796, + 0.003281742101535201, + 0.00169281754642725 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 2 + ], + "selected_reward": -0.07742787152528763 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a young, athletic male riding a mountain bike?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene set in a vast, grassy park at night?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the rider wearing a thick, dark wool sweater and black, full-finger performance gloves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene illuminated by a high-contrast golden glow from unseen park lights?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera pan steadily while following the rider in a low-angle, long shot?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04307401925325394 + ], + "opposite_logprob_tokens": [ + -3.168074131011963 + ], + "logprob_reward": -0.04307401925325394, + "yes_no_margin_reward": 3.125, + "token_credit_reward": -0.04307401925325394, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.31685692071914673, + 0.10089575499296188, + 0.06744935363531113, + 0.052558597177267075, + 0.08725523203611374, + 0.17335575819015503, + 0.05886613950133324, + 0.042205337435007095, + 0.04135821387171745, + 0.059198737144470215 + ], + "credit_map": [ + 0.02877875789999962, + 0.01780337654054165, + 0.05264988914132118, + 0.018899988383054733, + 0.040848951786756516, + 0.05279727652668953, + 0.021434815600514412, + 0.0836438462138176, + 0.009562162682414055, + 0.011030446738004684, + 0.019602147862315178, + 0.013021474704146385, + 0.011828839778900146, + 0.012065202929079533, + 0.011429266072809696, + 0.012356210500001907, + 0.004140281118452549, + 0.011301384307444096, + 0.006421583704650402, + 0.003536827862262726, + 0.01712542399764061, + 0.011426379904150963, + 0.004296046681702137, + 0.009201430715620518, + 0.004998404067009687, + 0.007219817955046892, + 0.00673916470259428, + 0.005135387182235718, + 0.006583992391824722, + 0.006074367091059685, + 0.00570523738861084, + 0.010102223604917526, + 0.0063309697434306145, + 0.00662585161626339, + 0.011501247994601727, + 0.006768937688320875, + 0.012768525630235672, + 0.013393300585448742, + 0.008692721836268902, + 0.02117367461323738, + 0.007623131386935711, + 0.009173094294965267, + 0.012341996654868126, + 0.008106080815196037, + 0.056890856474637985, + 0.052648257464170456, + 0.006596743129193783, + 0.01997559331357479, + 0.00494185509160161, + 0.0062521821819245815, + 0.00789057370275259, + 0.007395583670586348, + 0.006621263455599546, + 0.00686341617256403, + 0.008842862211167812, + 0.01005840115249157, + 0.0037772648502141237, + 0.00659562274813652, + 0.0068632615730166435, + 0.007004464510828257, + 0.005878876429051161, + 0.0051169125363230705, + 0.0030042279977351427, + 0.0039647105149924755, + 0.004237503744661808, + 0.006774228531867266, + 0.004179714247584343, + 0.004617735743522644, + 0.00763741135597229, + 0.005245732609182596, + 0.004899509251117706, + 0.0037663793191313744, + 0.006666023749858141, + 0.007781699299812317, + 0.009105293080210686, + 0.005265267100185156, + 0.009523182176053524, + 0.007034622598439455, + 0.005989418365061283, + 0.007833234034478664 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 5, + 1 + ], + "selected_reward": -0.04307401925325394 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a young, athletic male riding a mountain bike?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene set in a vast, grassy park at night?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the rider wearing a thick, dark wool sweater and black, full-finger performance gloves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene illuminated by a high-contrast golden glow from unseen park lights?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera pan steadily while following the rider in a low-angle, long shot?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.023303447291254997 + ], + "opposite_logprob_tokens": [ + -3.773303508758545 + ], + "logprob_reward": -0.023303447291254997, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.023303447291254997, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.5926648378372192, + 0.05991075932979584, + 0.05407777428627014, + 0.023044779896736145, + 0.02258501946926117, + 0.046530868858098984, + 0.03720066696405411, + 0.07564924657344818, + 0.04287106916308403, + 0.04546498879790306 + ], + "credit_map": [ + 0.0636720284819603, + 0.02888268418610096, + 0.05301262438297272, + 0.02733238972723484, + 0.20044924318790436, + 0.1548902541399002, + 0.03166457265615463, + 0.032761044800281525, + 0.00584949878975749, + 0.009947900660336018, + 0.009756731800734997, + 0.006273055914789438, + 0.00846205372363329, + 0.00712570920586586, + 0.0072421240620315075, + 0.005253684241324663, + 0.005088131874799728, + 0.012590707279741764, + 0.006537492386996746, + 0.004139124881476164, + 0.008094871416687965, + 0.005487451329827309, + 0.004274711012840271, + 0.007865282706916332, + 0.002346630673855543, + 0.0031425897032022476, + 0.0027004280127584934, + 0.002839582273736596, + 0.0033819277305155993, + 0.002927954774349928, + 0.003301423043012619, + 0.002404243452474475, + 0.0016603590920567513, + 0.0019219971727579832, + 0.001889113336801529, + 0.001129211625084281, + 0.006012053694576025, + 0.00536491209641099, + 0.0018180635524913669, + 0.0027893097139894962, + 0.0026115041691809893, + 0.003769206814467907, + 0.003744852729141712, + 0.002592104487121105, + 0.01367436908185482, + 0.01359703578054905, + 0.0024121287278831005, + 0.004129666835069656, + 0.004023958463221788, + 0.0055419509299099445, + 0.0059222071431577206, + 0.004446601495146751, + 0.005179196130484343, + 0.005158907268196344, + 0.003161417320370674, + 0.003766426583752036, + 0.007107919547706842, + 0.015078125521540642, + 0.006630132906138897, + 0.01637941226363182, + 0.008952340111136436, + 0.008705311454832554, + 0.006459283176809549, + 0.006336722988635302, + 0.005240655038505793, + 0.006769396830350161, + 0.004894568584859371, + 0.007429222576320171, + 0.003853682428598404, + 0.00688457814976573, + 0.004959973506629467, + 0.0028389880899339914, + 0.005660464521497488, + 0.005125593394041061, + 0.007609140127897263, + 0.005759617313742638, + 0.00442251842468977, + 0.006474296562373638, + 0.005775408819317818, + 0.004637952893972397 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 7, + 1 + ], + "selected_reward": -0.023303447291254997 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a young, athletic male riding a mountain bike?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene set in a vast, grassy park at night?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the rider wearing a thick, dark wool sweater and black, full-finger performance gloves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene illuminated by a high-contrast golden glow from unseen park lights?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera pan steadily while following the rider in a low-angle, long shot?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.002817591419443488 + ], + "opposite_logprob_tokens": [ + -5.877817630767822 + ], + "logprob_reward": -0.002817591419443488, + "yes_no_margin_reward": 5.875, + "token_credit_reward": -0.002817591419443488, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.18217085301876068, + 0.09817736595869064, + 0.19803275167942047, + 0.03665157034993172, + 0.03883593901991844, + 0.03836715221405029, + 0.03307533636689186, + 0.16276031732559204, + 0.0627937763929367, + 0.14913499355316162 + ], + "credit_map": [ + 0.013747992925345898, + 0.013164279982447624, + 0.019720030948519707, + 0.010302562266588211, + 0.04732728376984596, + 0.03988292068243027, + 0.010775633156299591, + 0.0272501390427351, + 0.008603042922914028, + 0.013286236673593521, + 0.015623311512172222, + 0.008454867638647556, + 0.014814085327088833, + 0.01672280952334404, + 0.009343027137219906, + 0.011329978704452515, + 0.01099669374525547, + 0.014398382976651192, + 0.02536146342754364, + 0.006469381041824818, + 0.07154393196105957, + 0.04994167760014534, + 0.01158918533474207, + 0.007732020225375891, + 0.005009057465940714, + 0.004845889750868082, + 0.004116926807910204, + 0.0024153352715075016, + 0.0058259074576199055, + 0.0054445043206214905, + 0.0030369546730071306, + 0.005956992972642183, + 0.005108482204377651, + 0.004871033597737551, + 0.005614018999040127, + 0.0035043854732066393, + 0.0045592994429171085, + 0.005576977040618658, + 0.0037904754281044006, + 0.005811267998069525, + 0.0029806755483150482, + 0.0049982210621237755, + 0.004834732972085476, + 0.0044515072368085384, + 0.0068107955157756805, + 0.006685764528810978, + 0.0031627521384507418, + 0.004442701116204262, + 0.002809695666655898, + 0.004182375036180019, + 0.004675721749663353, + 0.004640924744307995, + 0.0038734637200832367, + 0.0039494880475103855, + 0.00432297820225358, + 0.004620688501745462, + 0.015517782419919968, + 0.029989901930093765, + 0.01738932356238365, + 0.0327645018696785, + 0.02012665569782257, + 0.017395881935954094, + 0.015597834251821041, + 0.013978434726595879, + 0.006930800620466471, + 0.009699317626655102, + 0.008477484807372093, + 0.005988468416035175, + 0.011463458649814129, + 0.007905036211013794, + 0.007444772869348526, + 0.004884439520537853, + 0.018289627507328987, + 0.019508332014083862, + 0.028673797845840454, + 0.011974138207733631, + 0.01812601089477539, + 0.016885748133063316, + 0.015188571065664291, + 0.020488763228058815 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 0, + 7 + ], + "selected_reward": -0.002817591419443488 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a young, athletic male riding a mountain bike?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene set in a vast, grassy park at night?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the rider wearing a thick, dark wool sweater and black, full-finger performance gloves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene illuminated by a high-contrast golden glow from unseen park lights?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera pan steadily while following the rider in a low-angle, long shot?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02978304773569107 + ], + "opposite_logprob_tokens": [ + -3.529783010482788 + ], + "logprob_reward": -0.02978304773569107, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.02978304773569107, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3074757158756256, + 0.06267175078392029, + 0.053145669400691986, + 0.05176135525107384, + 0.055043332278728485, + 0.12968921661376953, + 0.03807707875967026, + 0.08928295224905014, + 0.14037297666072845, + 0.07248001545667648 + ], + "credit_map": [ + 0.029684359207749367, + 0.018514538183808327, + 0.02670823596417904, + 0.02144416607916355, + 0.091743603348732, + 0.07182873040437698, + 0.023485885933041573, + 0.024066168814897537, + 0.007981909438967705, + 0.006611617282032967, + 0.009859798476099968, + 0.005658190697431564, + 0.0072726779617369175, + 0.007953491061925888, + 0.009589675813913345, + 0.0077443802729249, + 0.004090563394129276, + 0.00945237372070551, + 0.006371178664267063, + 0.0035907940473407507, + 0.01172793097794056, + 0.008628797717392445, + 0.003971919417381287, + 0.0053121112287044525, + 0.0031404257752001286, + 0.008254040963947773, + 0.006016361992806196, + 0.005700174253433943, + 0.00754450261592865, + 0.0055037070997059345, + 0.008292723447084427, + 0.007309416774660349, + 0.004443416837602854, + 0.007208141963928938, + 0.006250347010791302, + 0.005454525351524353, + 0.006228569895029068, + 0.006451720371842384, + 0.00574397761374712, + 0.013262630440294743, + 0.004337403457611799, + 0.006345592904835939, + 0.004594501107931137, + 0.005673842038959265, + 0.05225026234984398, + 0.046735696494579315, + 0.002917892299592495, + 0.006834012921899557, + 0.0029722172766923904, + 0.006559507455676794, + 0.00493861548602581, + 0.004717924632132053, + 0.005257836543023586, + 0.004968364257365465, + 0.0037010638043284416, + 0.004961547441780567, + 0.008036164566874504, + 0.018314754590392113, + 0.01269756443798542, + 0.014307389967143536, + 0.007786816451698542, + 0.008658562786877155, + 0.010725011117756367, + 0.008756685070693493, + 0.01318467129021883, + 0.024707630276679993, + 0.010833624750375748, + 0.015813753008842468, + 0.03243007883429527, + 0.01810375601053238, + 0.01564641483128071, + 0.009653046727180481, + 0.011845661327242851, + 0.00772838992998004, + 0.009731398895382881, + 0.009705799631774426, + 0.005735600367188454, + 0.015228062868118286, + 0.0068354192189872265, + 0.005669678095728159 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 5 + ], + "selected_reward": -0.02978304773569107 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "In a dynamic, low-angle, long shot, the camera pans steadily, following a young, athletic male with short, wind-twisted hair as he rides a mountain bike through a vast, grassy park at night. He wears a thick, dark wool sweater and black, full-finger performance gloves for a secure grip on the handlebars. He pedals at a steady, moderate pace, but his body and bike are subjected to constant, sharp, and erratic motion as he expertly navigates the uneven ground. The bike itself is a sturdy model with a dark frame and thick, knobby tires that bounce and jostle over the terrain. The entire scene is bathed in a dramatic, high-contrast golden glow cast by tall, unseen park lights, which carve the rider's silhouette out of the darkness and cast long, twisting shadows that stretch and race behind him. A strong wind whips through the tall grass, subtly tugging at the fabric of his sweater, heightening the sense of solitary effort and creating an atmosphere of tension, focus, and visually arresting action against the inky night sky.", + "seed": 5950, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000050/group_12/index_8904_seed_5950_dp_24_rank_48_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 12, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.03394414195790887, + "vlm_logprob": -0.03394414195790887, + "vlm_yes_no_margin": 3.975, + "vlm_token_credit": -0.03394414195790887, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a young, athletic male riding a mountain bike?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene set in a vast, grassy park at night?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the rider wearing a thick, dark wool sweater and black, full-finger performance gloves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene illuminated by a high-contrast golden glow from unseen park lights?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera pan steadily while following the rider in a low-angle, long shot?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.06560641527175903 + ], + "opposite_logprob_tokens": [ + -2.8156063556671143 + ], + "logprob_reward": -0.06560641527175903, + "yes_no_margin_reward": 2.75, + "token_credit_reward": -0.06560641527175903, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0519447959959507, + 0.04623240977525711, + 0.07943618297576904, + 0.1398734450340271, + 0.047040458768606186, + 0.0711885318160057, + 0.14288252592086792, + 0.3175743520259857, + 0.072773277759552, + 0.031054005026817322 + ], + "credit_map": [ + 0.00491005415096879, + 0.0042871651239693165, + 0.007188412826508284, + 0.0062219807878136635, + 0.006521143484860659, + 0.006914039142429829, + 0.01058216392993927, + 0.005319833289831877, + 0.007695995736867189, + 0.008985497057437897, + 0.004145332612097263, + 0.005276236683130264, + 0.003988419659435749, + 0.003602184820920229, + 0.004459508694708347, + 0.008079233579337597, + 0.010722104460000992, + 0.016493823379278183, + 0.012477741576731205, + 0.008511577732861042, + 0.009700177237391472, + 0.008790558204054832, + 0.0052534001879394054, + 0.007486792281270027, + 0.035586096346378326, + 0.05805313587188721, + 0.004027761984616518, + 0.0027888233307749033, + 0.022460659965872765, + 0.008457867428660393, + 0.003703803289681673, + 0.004795292392373085, + 0.006578992586582899, + 0.008570166304707527, + 0.003808575449511409, + 0.0037510315887629986, + 0.006146157626062632, + 0.004097122699022293, + 0.006068108137696981, + 0.008020305074751377, + 0.008134744130074978, + 0.01094190776348114, + 0.0084100142121315, + 0.007384553551673889, + 0.008911543525755405, + 0.0069686309434473515, + 0.009434090927243233, + 0.01100304163992405, + 0.034520503133535385, + 0.042745210230350494, + 0.008842301554977894, + 0.009003219194710255, + 0.013019503094255924, + 0.014083168469369411, + 0.011926296167075634, + 0.008742325939238071, + 0.02801278419792652, + 0.07853499054908752, + 0.09265150874853134, + 0.038956161588430405, + 0.022147804498672485, + 0.022055434063076973, + 0.01643087901175022, + 0.018784793093800545, + 0.008360381238162518, + 0.010457003489136696, + 0.008858469314873219, + 0.0063492911867797375, + 0.007382901851087809, + 0.01071999967098236, + 0.012713424861431122, + 0.007931803353130817, + 0.0038422702345997095, + 0.005653160624206066, + 0.00457036541774869, + 0.003583332756534219, + 0.0026618693955242634, + 0.0035202258732169867, + 0.003173194359987974, + 0.004049586597830057 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 6, + 3 + ], + "selected_reward": -0.06560641527175903 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a young, athletic male riding a mountain bike?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene set in a vast, grassy park at night?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the rider wearing a thick, dark wool sweater and black, full-finger performance gloves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene illuminated by a high-contrast golden glow from unseen park lights?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera pan steadily while following the rider in a low-angle, long shot?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0181918703019619 + ], + "opposite_logprob_tokens": [ + -4.018191814422607 + ], + "logprob_reward": -0.0181918703019619, + "yes_no_margin_reward": 4.0, + "token_credit_reward": -0.0181918703019619, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10074476152658463, + 0.10340026021003723, + 0.2521074116230011, + 0.17640958726406097, + 0.03134662285447121, + 0.04412539303302765, + 0.022113926708698273, + 0.08091158419847488, + 0.14120930433273315, + 0.04763113334774971 + ], + "credit_map": [ + 0.012906786985695362, + 0.012324990704655647, + 0.009359300136566162, + 0.008637899532914162, + 0.018796555697917938, + 0.011476602405309677, + 0.017107054591178894, + 0.010135568678379059, + 0.013456965796649456, + 0.012791186571121216, + 0.009678120724856853, + 0.019203992560505867, + 0.0073833283968269825, + 0.007975994609296322, + 0.011682571843266487, + 0.021228089928627014, + 0.04602726548910141, + 0.07072917371988297, + 0.03403734043240547, + 0.030038004741072655, + 0.027126645669341087, + 0.012776181101799011, + 0.01531156525015831, + 0.016061238944530487, + 0.042918942868709564, + 0.0747189074754715, + 0.005083448253571987, + 0.0036297340411692858, + 0.03192375972867012, + 0.008226998150348663, + 0.00410595303401351, + 0.005801829509437084, + 0.00491533987224102, + 0.0059606293216347694, + 0.0041033802554011345, + 0.00234862114302814, + 0.0038349423557519913, + 0.0023114557843655348, + 0.0038866472896188498, + 0.003985605202615261, + 0.00946102850139141, + 0.012600407935678959, + 0.0041764117777347565, + 0.0031694171484559774, + 0.0032829558476805687, + 0.004280377645045519, + 0.0036106824409216642, + 0.0035441045183688402, + 0.004558970686048269, + 0.004996593110263348, + 0.0022178625222295523, + 0.0020202076993882656, + 0.002977317664772272, + 0.0020243250764906406, + 0.0019082332728430629, + 0.0014104163274168968, + 0.012157288379967213, + 0.025611722841858864, + 0.019410183653235435, + 0.005068245809525251, + 0.007384759373962879, + 0.0038629404734820127, + 0.003601707750931382, + 0.0038147333543747663, + 0.01903364062309265, + 0.02325838804244995, + 0.01513820793479681, + 0.012295981869101524, + 0.015949340537190437, + 0.021811550483107567, + 0.020428113639354706, + 0.013294064439833164, + 0.0053346469067037106, + 0.008466798812150955, + 0.007614857982844114, + 0.005458781495690346, + 0.004524386953562498, + 0.004923135042190552, + 0.004749319516122341, + 0.006559202913194895 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 3, + 8 + ], + "selected_reward": -0.0181918703019619 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a young, athletic male riding a mountain bike?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene set in a vast, grassy park at night?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the rider wearing a thick, dark wool sweater and black, full-finger performance gloves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene illuminated by a high-contrast golden glow from unseen park lights?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera pan steadily while following the rider in a low-angle, long shot?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.05495993793010712 + ], + "opposite_logprob_tokens": [ + -2.929960012435913 + ], + "logprob_reward": -0.05495993793010712, + "yes_no_margin_reward": 2.875, + "token_credit_reward": -0.05495993793010712, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08724961429834366, + 0.02514176443219185, + 0.22845587134361267, + 0.13921743631362915, + 0.04820261895656586, + 0.19167256355285645, + 0.08374415338039398, + 0.09437961131334305, + 0.05997235327959061, + 0.04196398705244064 + ], + "credit_map": [ + 0.010394039563834667, + 0.010513662360608578, + 0.014921817928552628, + 0.009230277501046658, + 0.009735225699841976, + 0.009406664408743382, + 0.013337462209165096, + 0.0097104636952281, + 0.002828404074534774, + 0.002244654344394803, + 0.003172700759023428, + 0.0041268011555075645, + 0.0020407482516020536, + 0.002073079813271761, + 0.003425918286666274, + 0.005229457747191191, + 0.0356137752532959, + 0.05285456031560898, + 0.008836892433464527, + 0.011758830398321152, + 0.03939583897590637, + 0.04108426347374916, + 0.01358445268124342, + 0.025327250361442566, + 0.0324675478041172, + 0.059285927563905716, + 0.003314289031550288, + 0.002857369836419821, + 0.0267763901501894, + 0.006754639092832804, + 0.003482118947431445, + 0.004279156681150198, + 0.007422860711812973, + 0.008992484770715237, + 0.005610637366771698, + 0.0038908522110432386, + 0.005512901581823826, + 0.003603558288887143, + 0.005756214261054993, + 0.007413111161440611, + 0.024235494434833527, + 0.03270848095417023, + 0.013421710580587387, + 0.014449512585997581, + 0.029393499717116356, + 0.011725240387022495, + 0.03633715957403183, + 0.029401468113064766, + 0.011040536686778069, + 0.01395840011537075, + 0.00802089273929596, + 0.008486210368573666, + 0.019348019734025, + 0.006940840277820826, + 0.010403822176158428, + 0.005545436404645443, + 0.022907232865691185, + 0.01996038854122162, + 0.01149532850831747, + 0.007055789697915316, + 0.011595852673053741, + 0.0071266647428274155, + 0.008509895764291286, + 0.005728460848331451, + 0.00850516464561224, + 0.010514146648347378, + 0.006428265944123268, + 0.004911547992378473, + 0.008721983060240746, + 0.009090033359825611, + 0.008124117739498615, + 0.0036770880687981844, + 0.004069576971232891, + 0.007005722261965275, + 0.0057736956514418125, + 0.00685102678835392, + 0.0034532214049249887, + 0.004719400312751532, + 0.004174820613116026, + 0.005916521418839693 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 5, + 3 + ], + "selected_reward": -0.05495993793010712 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a young, athletic male riding a mountain bike?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene set in a vast, grassy park at night?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the rider wearing a thick, dark wool sweater and black, full-finger performance gloves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene illuminated by a high-contrast golden glow from unseen park lights?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera pan steadily while following the rider in a low-angle, long shot?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0011781663633883 + ], + "opposite_logprob_tokens": [ + -6.75117826461792 + ], + "logprob_reward": -0.0011781663633883, + "yes_no_margin_reward": 6.75, + "token_credit_reward": -0.0011781663633883, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14849212765693665, + 0.04849269241094589, + 0.2180170863866806, + 0.21947383880615234, + 0.05164453387260437, + 0.06562401354312897, + 0.027688292786478996, + 0.05575094372034073, + 0.13396042585372925, + 0.03085603378713131 + ], + "credit_map": [ + 0.01641736924648285, + 0.01586218550801277, + 0.019407447427511215, + 0.022222483530640602, + 0.014520061202347279, + 0.018434014171361923, + 0.020485475659370422, + 0.02114308625459671, + 0.006453399080783129, + 0.007270400878041983, + 0.007777318824082613, + 0.007806885521858931, + 0.004440860822796822, + 0.0036452426575124264, + 0.005662205163389444, + 0.00543638551607728, + 0.05111817643046379, + 0.08173856139183044, + 0.009926420636475086, + 0.011089109815657139, + 0.029733870178461075, + 0.014130836352705956, + 0.008332224562764168, + 0.011947883293032646, + 0.050736553966999054, + 0.0808601826429367, + 0.007944951765239239, + 0.006079451646655798, + 0.033900048583745956, + 0.01836566999554634, + 0.00951722264289856, + 0.012069746851921082, + 0.008889670483767986, + 0.011544457636773586, + 0.004864897113293409, + 0.002626231173053384, + 0.006153133697807789, + 0.003415679559111595, + 0.006903888192027807, + 0.007246573455631733, + 0.01549385767430067, + 0.02044963650405407, + 0.00475722923874855, + 0.002964842366054654, + 0.006044404581189156, + 0.004759073723107576, + 0.005784094333648682, + 0.005370876751840115, + 0.004720612894743681, + 0.004917313810437918, + 0.003944132011383772, + 0.0027590373065322638, + 0.004912544973194599, + 0.001909850281663239, + 0.002569419564679265, + 0.0019553822930902243, + 0.008826189674437046, + 0.013641144149005413, + 0.007908293977379799, + 0.004713321104645729, + 0.007354239467531443, + 0.00408977922052145, + 0.005397310480475426, + 0.0038206649478524923, + 0.02080301195383072, + 0.02715870551764965, + 0.01601344905793667, + 0.01022354606539011, + 0.017474744468927383, + 0.023498212918639183, + 0.01248224452137947, + 0.006306508090347052, + 0.003790933871641755, + 0.004835882224142551, + 0.004188202321529388, + 0.004277617670595646, + 0.002737379865720868, + 0.005509212613105774, + 0.003099508583545685, + 0.0024173001293092966 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 2, + 0 + ], + "selected_reward": -0.0011781663633883 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a young, athletic male riding a mountain bike?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene set in a vast, grassy park at night?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the rider wearing a thick, dark wool sweater and black, full-finger performance gloves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene illuminated by a high-contrast golden glow from unseen park lights?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera pan steadily while following the rider in a low-angle, long shot?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.029784319922327995 + ], + "opposite_logprob_tokens": [ + -3.5297842025756836 + ], + "logprob_reward": -0.029784319922327995, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.029784319922327995, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11469846218824387, + 0.04979987069964409, + 0.13382503390312195, + 0.11246400326490402, + 0.036369770765304565, + 0.1749497950077057, + 0.09122283011674881, + 0.13809305429458618, + 0.08952169865369797, + 0.05905543267726898 + ], + "credit_map": [ + 0.010244574397802353, + 0.008187820203602314, + 0.01744595542550087, + 0.012794861570000648, + 0.013211788609623909, + 0.013517406769096851, + 0.02650674246251583, + 0.012789317406713963, + 0.005158627405762672, + 0.0070318360812962055, + 0.006169911008328199, + 0.007252897601574659, + 0.004362129606306553, + 0.004009670577943325, + 0.007792195770889521, + 0.008022602647542953, + 0.0327000692486763, + 0.05123578757047653, + 0.004764301236718893, + 0.007632901892066002, + 0.017390748485922813, + 0.00666266493499279, + 0.0058853505179286, + 0.007553223054856062, + 0.02334170788526535, + 0.0436084121465683, + 0.0024225804954767227, + 0.004011761862784624, + 0.02246074005961418, + 0.006808031816035509, + 0.004667297471314669, + 0.005143469199538231, + 0.004516619257628918, + 0.005488325376063585, + 0.004708176478743553, + 0.004571867175400257, + 0.004083164501935244, + 0.0026762268971651793, + 0.005840623285621405, + 0.0044847712852060795, + 0.05394327640533447, + 0.07130862772464752, + 0.0071926843374967575, + 0.004728685598820448, + 0.011234729550778866, + 0.012219156138598919, + 0.008940295316278934, + 0.0053823539055883884, + 0.015858488157391548, + 0.022588100284337997, + 0.009122657589614391, + 0.00952752586454153, + 0.013150593265891075, + 0.009071257896721363, + 0.006148362532258034, + 0.005755851976573467, + 0.04308217018842697, + 0.04060531035065651, + 0.01006034854799509, + 0.008113640360534191, + 0.00982572790235281, + 0.01434888131916523, + 0.007387241814285517, + 0.004669740796089172, + 0.00981053989380598, + 0.011959681287407875, + 0.014275532215833664, + 0.011065098457038403, + 0.010730724781751633, + 0.012790165841579437, + 0.01241551898419857, + 0.006474439520388842, + 0.007506945636123419, + 0.01029323972761631, + 0.0065466309897601604, + 0.007359410636126995, + 0.0062773204408586025, + 0.010863402858376503, + 0.0049837990663945675, + 0.005224684253334999 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 7, + 2 + ], + "selected_reward": -0.029784319922327995 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "In a dynamic, low-angle, long shot, the camera pans steadily, following a young, athletic male with short, wind-twisted hair as he rides a mountain bike through a vast, grassy park at night. He wears a thick, dark wool sweater and black, full-finger performance gloves for a secure grip on the handlebars. He pedals at a steady, moderate pace, but his body and bike are subjected to constant, sharp, and erratic motion as he expertly navigates the uneven ground. The bike itself is a sturdy model with a dark frame and thick, knobby tires that bounce and jostle over the terrain. The entire scene is bathed in a dramatic, high-contrast golden glow cast by tall, unseen park lights, which carve the rider's silhouette out of the darkness and cast long, twisting shadows that stretch and race behind him. A strong wind whips through the tall grass, subtly tugging at the fabric of his sweater, heightening the sense of solitary effort and creating an atmosphere of tension, focus, and visually arresting action against the inky night sky.", + "seed": 5951, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000050/group_12/index_8904_seed_5951_dp_25_rank_50_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 12, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.1013404055032879, + "vlm_logprob": -0.1013404055032879, + "vlm_yes_no_margin": 3.175, + "vlm_token_credit": -0.1013404055032879, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a young, athletic male riding a mountain bike?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene set in a vast, grassy park at night?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the rider wearing a thick, dark wool sweater and black, full-finger performance gloves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene illuminated by a high-contrast golden glow from unseen park lights?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera pan steadily while following the rider in a low-angle, long shot?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.12336507439613342 + ], + "opposite_logprob_tokens": [ + -2.2483651638031006 + ], + "logprob_reward": -0.12336507439613342, + "yes_no_margin_reward": 2.125, + "token_credit_reward": -0.12336507439613342, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08408202975988388, + 0.05986189842224121, + 0.09105903655290604, + 0.13378649950027466, + 0.17000949382781982, + 0.10806532949209213, + 0.048685673624277115, + 0.10120954364538193, + 0.16911356151103973, + 0.03412690758705139 + ], + "credit_map": [ + 0.01154076773673296, + 0.011592664755880833, + 0.012692896649241447, + 0.009152952581644058, + 0.013558541424572468, + 0.009683583863079548, + 0.007263468578457832, + 0.008597147651016712, + 0.008176741190254688, + 0.007382823154330254, + 0.011540438048541546, + 0.0069280387833714485, + 0.0050495583564043045, + 0.008196528069674969, + 0.005322588607668877, + 0.007265178952366114, + 0.014058495871722698, + 0.011507000774145126, + 0.014688893221318722, + 0.007728613447397947, + 0.014926298521459103, + 0.010167517699301243, + 0.007653695065528154, + 0.010328522883355618, + 0.014756639488041401, + 0.012006684206426144, + 0.02758486568927765, + 0.01431384589523077, + 0.024391964077949524, + 0.01701849140226841, + 0.01853219047188759, + 0.005181828513741493, + 0.014676927588880062, + 0.01481782365590334, + 0.04494079202413559, + 0.016497060656547546, + 0.018354980275034904, + 0.03247838839888573, + 0.01797918975353241, + 0.010264325886964798, + 0.013546125032007694, + 0.014491884037852287, + 0.01079399511218071, + 0.011728397570550442, + 0.016524726524949074, + 0.014686976559460163, + 0.013342372141778469, + 0.012950855307281017, + 0.006710444577038288, + 0.00503283878788352, + 0.0029848499689251184, + 0.0029696801211684942, + 0.01041962020099163, + 0.011262254789471626, + 0.005310037173330784, + 0.003995946142822504, + 0.019829437136650085, + 0.00854644924402237, + 0.014924230054020882, + 0.008037557825446129, + 0.010703176259994507, + 0.020272275432944298, + 0.01207614503800869, + 0.0068202693946659565, + 0.015742488205432892, + 0.025819364935159683, + 0.02359870634973049, + 0.020383276045322418, + 0.013606368564069271, + 0.03120090439915657, + 0.028996627777814865, + 0.009765827096998692, + 0.007027394603937864, + 0.00566864525899291, + 0.004617399536073208, + 0.0031613693572580814, + 0.002855730475857854, + 0.0024639901239424944, + 0.003898743074387312, + 0.004433630965650082 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 8, + 3 + ], + "selected_reward": -0.12336507439613342 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a young, athletic male riding a mountain bike?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene set in a vast, grassy park at night?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the rider wearing a thick, dark wool sweater and black, full-finger performance gloves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene illuminated by a high-contrast golden glow from unseen park lights?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera pan steadily while following the rider in a low-angle, long shot?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.3134453296661377 + ], + "opposite_logprob_tokens": [ + -1.3134453296661377 + ], + "logprob_reward": -0.3134453296661377, + "yes_no_margin_reward": 1.0, + "token_credit_reward": -0.3134453296661377, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09949817508459091, + 0.15710504353046417, + 0.09552823007106781, + 0.0751129612326622, + 0.26875853538513184, + 0.07719869911670685, + 0.0898628756403923, + 0.020366841927170753, + 0.04818953201174736, + 0.06837908923625946 + ], + "credit_map": [ + 0.017568545415997505, + 0.015783026814460754, + 0.01596711203455925, + 0.009588619694113731, + 0.011534444987773895, + 0.012345938012003899, + 0.008694186806678772, + 0.008016307838261127, + 0.025300217792391777, + 0.02066630870103836, + 0.03125043585896492, + 0.012458188459277153, + 0.01245258841663599, + 0.03302610293030739, + 0.011883178725838661, + 0.010068015195429325, + 0.024087751284241676, + 0.025322435423731804, + 0.01570488139986992, + 0.006344472989439964, + 0.0067474436946213245, + 0.008767745457589626, + 0.004196496680378914, + 0.004357002209872007, + 0.012157564051449299, + 0.008784099481999874, + 0.006634625606238842, + 0.006626021582633257, + 0.013243584893643856, + 0.016313903033733368, + 0.007329973392188549, + 0.004023195710033178, + 0.02314385585486889, + 0.01788202114403248, + 0.07531654834747314, + 0.015920694917440414, + 0.027878951281309128, + 0.08539628982543945, + 0.011140755377709866, + 0.012079406529664993, + 0.007622839882969856, + 0.00717563321813941, + 0.010416379198431969, + 0.005788303446024656, + 0.014457440935075283, + 0.015619891695678234, + 0.0051076365634799, + 0.011010571382939816, + 0.010254085063934326, + 0.006817762274295092, + 0.006600112654268742, + 0.004418391268700361, + 0.020981887355446815, + 0.026965616270899773, + 0.006554631050676107, + 0.007270390167832375, + 0.0029592986684292555, + 0.0016819959273561835, + 0.005272941663861275, + 0.0016866984078660607, + 0.0020036452915519476, + 0.0030629721004515886, + 0.0015638149343430996, + 0.0021354767959564924, + 0.005517021287232637, + 0.0067536053247749805, + 0.00829069223254919, + 0.006356694269925356, + 0.004202078562229872, + 0.0063991122879087925, + 0.007833962328732014, + 0.002836365718394518, + 0.011704345233738422, + 0.01083688996732235, + 0.01162730623036623, + 0.011616630479693413, + 0.004851586185395718, + 0.004939593840390444, + 0.006221006158739328, + 0.006581730209290981 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 1, + 0 + ], + "selected_reward": -0.3134453296661377 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a young, athletic male riding a mountain bike?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene set in a vast, grassy park at night?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the rider wearing a thick, dark wool sweater and black, full-finger performance gloves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene illuminated by a high-contrast golden glow from unseen park lights?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera pan steadily while following the rider in a low-angle, long shot?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03371838480234146 + ], + "opposite_logprob_tokens": [ + -3.4087183475494385 + ], + "logprob_reward": -0.03371838480234146, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.03371838480234146, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07220623642206192, + 0.07438605278730392, + 0.030043238773941994, + 0.09666076302528381, + 0.2212732583284378, + 0.10899464786052704, + 0.06259416043758392, + 0.12449382245540619, + 0.16688841581344604, + 0.04245934635400772 + ], + "credit_map": [ + 0.010998500511050224, + 0.01200917735695839, + 0.008196311071515083, + 0.007780004292726517, + 0.011204247362911701, + 0.007948263548314571, + 0.008485226891934872, + 0.0055845025926828384, + 0.012940223328769207, + 0.007524545770138502, + 0.006897334475070238, + 0.006849865894764662, + 0.010635490529239178, + 0.012950446456670761, + 0.008905933238565922, + 0.0076822093687951565, + 0.005109408870339394, + 0.004303548950701952, + 0.0073830746114254, + 0.002972460351884365, + 0.003517907578498125, + 0.002513656159862876, + 0.0019026327645406127, + 0.002340548438951373, + 0.01800416223704815, + 0.010334452614188194, + 0.014555339701473713, + 0.010888389311730862, + 0.009991922415792942, + 0.015436412766575813, + 0.012218480929732323, + 0.005231607239693403, + 0.017860589548945427, + 0.01685466803610325, + 0.06523146480321884, + 0.016513044014573097, + 0.01990577206015587, + 0.05661429092288017, + 0.013649286702275276, + 0.014644138514995575, + 0.012792174704372883, + 0.015153899788856506, + 0.009373171254992485, + 0.012826823629438877, + 0.015528982505202293, + 0.015340053476393223, + 0.015506780706346035, + 0.012472759932279587, + 0.007582359481602907, + 0.009515874087810516, + 0.0050578732043504715, + 0.003948307130485773, + 0.011308497749269009, + 0.013066999614238739, + 0.005549944005906582, + 0.0065643079578876495, + 0.028023498132824898, + 0.00981197226792574, + 0.01100163534283638, + 0.011137779802083969, + 0.013211208395659924, + 0.02932887151837349, + 0.014024057425558567, + 0.00795479491353035, + 0.018260471522808075, + 0.022119371220469475, + 0.024050099775195122, + 0.018327629193663597, + 0.014327571727335453, + 0.030880950391292572, + 0.026600923389196396, + 0.01232139766216278, + 0.008099092170596123, + 0.007636190392076969, + 0.008774799294769764, + 0.004850778263062239, + 0.0035179248079657555, + 0.0033867182210087776, + 0.0030425579752773046, + 0.0031512815039604902 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 8, + 7 + ], + "selected_reward": -0.03371838480234146 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a young, athletic male riding a mountain bike?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene set in a vast, grassy park at night?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the rider wearing a thick, dark wool sweater and black, full-finger performance gloves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene illuminated by a high-contrast golden glow from unseen park lights?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera pan steadily while following the rider in a low-angle, long shot?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0024889700580388308 + ], + "opposite_logprob_tokens": [ + -6.00248908996582 + ], + "logprob_reward": -0.0024889700580388308, + "yes_no_margin_reward": 6.0, + "token_credit_reward": -0.0024889700580388308, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08605869114398956, + 0.23590584099292755, + 0.03243307024240494, + 0.10002841055393219, + 0.21678420901298523, + 0.10694482922554016, + 0.07767630368471146, + 0.049690842628479004, + 0.06103062629699707, + 0.03344707190990448 + ], + "credit_map": [ + 0.01437580306082964, + 0.011898997239768505, + 0.018135543912649155, + 0.008798996917903423, + 0.00906627718359232, + 0.010613798163831234, + 0.006510710809379816, + 0.006658569443970919, + 0.037908948957920074, + 0.03251039981842041, + 0.04868334159255028, + 0.019712962210178375, + 0.02095722034573555, + 0.043251894414424896, + 0.015128113329410553, + 0.01775301620364189, + 0.006364933680742979, + 0.006267962045967579, + 0.007814595475792885, + 0.002834347076714039, + 0.00242619845084846, + 0.0027128453366458416, + 0.001583956298418343, + 0.0024282352533191442, + 0.01720244064927101, + 0.008463410660624504, + 0.03378915414214134, + 0.008065713569521904, + 0.008694621734321117, + 0.013147036544978619, + 0.007092623971402645, + 0.003573417430743575, + 0.01803700067102909, + 0.009690579026937485, + 0.06020941957831383, + 0.011057685129344463, + 0.025792952626943588, + 0.07466230541467667, + 0.008972722105681896, + 0.008361558429896832, + 0.011418008245527744, + 0.008117547258734703, + 0.014864196069538593, + 0.008573146536946297, + 0.019416065886616707, + 0.02250656485557556, + 0.008516204543411732, + 0.013533107005059719, + 0.009338942356407642, + 0.009575990960001945, + 0.008981906808912754, + 0.005153501871973276, + 0.014988712966442108, + 0.014798075892031193, + 0.006307880859822035, + 0.00853129755705595, + 0.009810242801904678, + 0.0038940932136029005, + 0.00968943815678358, + 0.0033534520771354437, + 0.004075947217643261, + 0.008786443620920181, + 0.0036411143373697996, + 0.006440119352191687, + 0.005936576519161463, + 0.00862948689609766, + 0.009248272515833378, + 0.007814044132828712, + 0.004995593335479498, + 0.009947706945240498, + 0.010140960104763508, + 0.004317997023463249, + 0.00894275214523077, + 0.006264288444072008, + 0.006286916323006153, + 0.003208615817129612, + 0.0025802934542298317, + 0.00239186710678041, + 0.001830340246669948, + 0.001942004426382482 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 4, + 5 + ], + "selected_reward": -0.0024889700580388308 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a young, athletic male riding a mountain bike?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene set in a vast, grassy park at night?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the rider wearing a thick, dark wool sweater and black, full-finger performance gloves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene illuminated by a high-contrast golden glow from unseen park lights?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera pan steadily while following the rider in a low-angle, long shot?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03368426859378815 + ], + "opposite_logprob_tokens": [ + -3.408684253692627 + ], + "logprob_reward": -0.03368426859378815, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.03368426859378815, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.046650078147649765, + 0.049972984939813614, + 0.04515061527490616, + 0.11678163707256317, + 0.37868791818618774, + 0.0863889679312706, + 0.11056341975927353, + 0.047024600207805634, + 0.04177183657884598, + 0.07700799405574799 + ], + "credit_map": [ + 0.005459878593683243, + 0.008299684152007103, + 0.007233891636133194, + 0.0038623055443167686, + 0.007930011488497257, + 0.0057212780229747295, + 0.004580391570925713, + 0.0035626341123133898, + 0.008640872314572334, + 0.0032859027851372957, + 0.010234483517706394, + 0.0032801961060613394, + 0.004272831603884697, + 0.013716059736907482, + 0.002820572117343545, + 0.003722064197063446, + 0.006805753335356712, + 0.006762640550732613, + 0.00607417244464159, + 0.005205234047025442, + 0.004534303676337004, + 0.0035771916154772043, + 0.003650573082268238, + 0.008540746755897999, + 0.014512711204588413, + 0.010840180329978466, + 0.03929466754198074, + 0.009467153809964657, + 0.01131721306592226, + 0.021188847720623016, + 0.006368144880980253, + 0.0037927082739770412, + 0.040588125586509705, + 0.023862037807703018, + 0.0781015083193779, + 0.02238803543150425, + 0.050427235662937164, + 0.12809231877326965, + 0.02123212441802025, + 0.013996507041156292, + 0.00883247796446085, + 0.009811637923121452, + 0.011469044722616673, + 0.006660629063844681, + 0.010991877876222134, + 0.011648531071841717, + 0.010412609204649925, + 0.01656215824186802, + 0.01298864372074604, + 0.009072494693100452, + 0.009628807194530964, + 0.007843916304409504, + 0.021620484068989754, + 0.02810610458254814, + 0.009020482189953327, + 0.012282480485737324, + 0.009134230203926563, + 0.006546817719936371, + 0.005637085065245628, + 0.0038178875111043453, + 0.005613496992737055, + 0.008212005719542503, + 0.0050253053195774555, + 0.003037769813090563, + 0.003934765700250864, + 0.00528680020943284, + 0.007152568083256483, + 0.005455539561808109, + 0.002345950109884143, + 0.0077333953231573105, + 0.007055744528770447, + 0.002807067008689046, + 0.010388512164354324, + 0.01163241546601057, + 0.017216546460986137, + 0.013414870947599411, + 0.005380951799452305, + 0.004396477714180946, + 0.009752878919243813, + 0.0048253340646624565 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 3, + 6 + ], + "selected_reward": -0.03368426859378815 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "In a dynamic, low-angle, long shot, the camera pans steadily, following a young, athletic male with short, wind-twisted hair as he rides a mountain bike through a vast, grassy park at night. He wears a thick, dark wool sweater and black, full-finger performance gloves for a secure grip on the handlebars. He pedals at a steady, moderate pace, but his body and bike are subjected to constant, sharp, and erratic motion as he expertly navigates the uneven ground. The bike itself is a sturdy model with a dark frame and thick, knobby tires that bounce and jostle over the terrain. The entire scene is bathed in a dramatic, high-contrast golden glow cast by tall, unseen park lights, which carve the rider's silhouette out of the darkness and cast long, twisting shadows that stretch and race behind him. A strong wind whips through the tall grass, subtly tugging at the fabric of his sweater, heightening the sense of solitary effort and creating an atmosphere of tension, focus, and visually arresting action against the inky night sky.", + "seed": 5952, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000050/group_12/index_8904_seed_5952_dp_25_rank_50_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 12, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.05104082529433072, + "vlm_logprob": -0.05104082529433072, + "vlm_yes_no_margin": 3.6, + "vlm_token_credit": -0.05104082529433072, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a young, athletic male riding a mountain bike?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene set in a vast, grassy park at night?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the rider wearing a thick, dark wool sweater and black, full-finger performance gloves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene illuminated by a high-contrast golden glow from unseen park lights?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera pan steadily while following the rider in a low-angle, long shot?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.11098390817642212 + ], + "opposite_logprob_tokens": [ + -2.3609838485717773 + ], + "logprob_reward": -0.11098390817642212, + "yes_no_margin_reward": 2.25, + "token_credit_reward": -0.11098390817642212, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11952269822359085, + 0.08162988722324371, + 0.03401612117886543, + 0.057563748210668564, + 0.06505563110113144, + 0.18737289309501648, + 0.03149888664484024, + 0.03947463259100914, + 0.2951473891735077, + 0.08871811628341675 + ], + "credit_map": [ + 0.012851729989051819, + 0.02764219604432583, + 0.015975836664438248, + 0.009918343275785446, + 0.017968429252505302, + 0.016853831708431244, + 0.009609601460397243, + 0.008702726103365421, + 0.009736268781125546, + 0.018804779276251793, + 0.00707516074180603, + 0.005673454608768225, + 0.007688213139772415, + 0.023325767368078232, + 0.005450911819934845, + 0.003875330789014697, + 0.005107945762574673, + 0.0066117532551288605, + 0.00209143478423357, + 0.004931204952299595, + 0.004193687811493874, + 0.0036745949182659388, + 0.004152290057390928, + 0.0032532084733247757, + 0.005266927648335695, + 0.00791841745376587, + 0.005721502471715212, + 0.009051342494785786, + 0.007707561366260052, + 0.0056623718701303005, + 0.010797870345413685, + 0.005437755491584539, + 0.006885826587677002, + 0.010028298944234848, + 0.011555535718798637, + 0.007909937761723995, + 0.008131610229611397, + 0.009100374765694141, + 0.00867172796279192, + 0.002772325649857521, + 0.04346805810928345, + 0.013757838867604733, + 0.013103527016937733, + 0.013242468237876892, + 0.012137326411902905, + 0.0618799552321434, + 0.01911807991564274, + 0.010665654204785824, + 0.0043458533473312855, + 0.003872082568705082, + 0.005108975805342197, + 0.0044925762340426445, + 0.003099776804447174, + 0.0038804823998361826, + 0.0034073893912136555, + 0.0032917484641075134, + 0.0044770147651433945, + 0.005914181005209684, + 0.005932336673140526, + 0.005530079361051321, + 0.0037280465476214886, + 0.005506807938218117, + 0.004336599726229906, + 0.0040495689027011395, + 0.02922535501420498, + 0.023006927222013474, + 0.08352213352918625, + 0.023877713829278946, + 0.029402388259768486, + 0.04563377797603607, + 0.03736482933163643, + 0.02311427891254425, + 0.007515666075050831, + 0.01922047883272171, + 0.01338975504040718, + 0.011422297917306423, + 0.01171416137367487, + 0.008242182433605194, + 0.01057205256074667, + 0.0066415127366781235 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 5, + 0 + ], + "selected_reward": -0.11098390817642212 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a young, athletic male riding a mountain bike?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene set in a vast, grassy park at night?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the rider wearing a thick, dark wool sweater and black, full-finger performance gloves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene illuminated by a high-contrast golden glow from unseen park lights?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera pan steadily while following the rider in a low-angle, long shot?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07899407297372818 + ], + "opposite_logprob_tokens": [ + -2.578994035720825 + ], + "logprob_reward": -0.07899407297372818, + "yes_no_margin_reward": 2.5, + "token_credit_reward": -0.07899407297372818, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.19684743881225586, + 0.3379635214805603, + 0.0635327622294426, + 0.018243156373500824, + 0.032252199947834015, + 0.1633836328983307, + 0.08975347131490707, + 0.02263663522899151, + 0.03300178423523903, + 0.042385391891002655 + ], + "credit_map": [ + 0.024524103850126266, + 0.0440782867372036, + 0.03320016339421272, + 0.017310170456767082, + 0.019023744389414787, + 0.02734644152224064, + 0.014863685704767704, + 0.016500862315297127, + 0.07868346571922302, + 0.05074228718876839, + 0.016076017171144485, + 0.014022436924278736, + 0.026998616755008698, + 0.12293779850006104, + 0.017773965373635292, + 0.010728977620601654, + 0.005655307322740555, + 0.010946007445454597, + 0.010872677899897099, + 0.013426088728010654, + 0.00838056206703186, + 0.005752402823418379, + 0.004444553982466459, + 0.0040551661513745785, + 0.002081760438159108, + 0.002821451285853982, + 0.0024242957588285208, + 0.0021851679775863886, + 0.0022286272142082453, + 0.002405210630968213, + 0.0023453731555491686, + 0.0017512724734842777, + 0.002723839832469821, + 0.0053048827685415745, + 0.003699852619320154, + 0.0031856605783104897, + 0.006414846051484346, + 0.005159424152225256, + 0.00428408570587635, + 0.0014796126633882523, + 0.042805805802345276, + 0.010171428322792053, + 0.01383235678076744, + 0.008850906044244766, + 0.010038841515779495, + 0.05887006223201752, + 0.010258692316710949, + 0.008555560372769833, + 0.010077860206365585, + 0.01211266964673996, + 0.02404286339879036, + 0.012744540348649025, + 0.007028624415397644, + 0.009112486615777016, + 0.007424740120768547, + 0.00720969308167696, + 0.002263743197545409, + 0.0020876412745565176, + 0.0034545499365776777, + 0.0033783402759581804, + 0.0023523960262537003, + 0.003638235153630376, + 0.0031837550923228264, + 0.002277976367622614, + 0.0033947741612792015, + 0.003695765044540167, + 0.00521860783919692, + 0.003496520919725299, + 0.00456763431429863, + 0.004623120650649071, + 0.004163397941738367, + 0.0038419649936258793, + 0.0036542732268571854, + 0.007431478705257177, + 0.006446674931794405, + 0.005413369741290808, + 0.004975161049515009, + 0.0048948898911476135, + 0.006472604814916849, + 0.00309694348834455 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 5 + ], + "selected_reward": -0.07899407297372818 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a young, athletic male riding a mountain bike?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene set in a vast, grassy park at night?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the rider wearing a thick, dark wool sweater and black, full-finger performance gloves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene illuminated by a high-contrast golden glow from unseen park lights?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera pan steadily while following the rider in a low-angle, long shot?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.029824240133166313 + ], + "opposite_logprob_tokens": [ + -3.5298242568969727 + ], + "logprob_reward": -0.029824240133166313, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.029824240133166313, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.088405542075634, + 0.18085196614265442, + 0.08998316526412964, + 0.05557741969823837, + 0.1037425547838211, + 0.3101675808429718, + 0.05265093967318535, + 0.047488730400800705, + 0.04245280101895332, + 0.028679275885224342 + ], + "credit_map": [ + 0.012746446765959263, + 0.017273008823394775, + 0.009590080939233303, + 0.007601388264447451, + 0.010628960095345974, + 0.01293451339006424, + 0.00752598512917757, + 0.010105159133672714, + 0.03901228308677673, + 0.02934405207633972, + 0.00579554820433259, + 0.004804251715540886, + 0.011450941674411297, + 0.07452365756034851, + 0.008066949434578419, + 0.007854287512600422, + 0.01377080474048853, + 0.016735613346099854, + 0.004009590949863195, + 0.009351578541100025, + 0.01589975319802761, + 0.008424936793744564, + 0.009966185316443443, + 0.011824699118733406, + 0.004905165638774633, + 0.008521596901118755, + 0.006238909903913736, + 0.00838707946240902, + 0.006076486315578222, + 0.0070190830156207085, + 0.008843054994940758, + 0.005586044862866402, + 0.007525249384343624, + 0.02254052273929119, + 0.009345718659460545, + 0.008404183201491833, + 0.019448325037956238, + 0.017078625038266182, + 0.013606352731585503, + 0.005793575197458267, + 0.09098978340625763, + 0.019599856808781624, + 0.019392037764191628, + 0.010789125226438046, + 0.015073266811668873, + 0.12155993282794952, + 0.01538079883903265, + 0.01738276705145836, + 0.006484863348305225, + 0.005814716219902039, + 0.012720288708806038, + 0.00787307694554329, + 0.004509761463850737, + 0.006369610317051411, + 0.004748832900077105, + 0.004129788838326931, + 0.006662759929895401, + 0.005907483398914337, + 0.0066833654418587685, + 0.00521086435765028, + 0.00531860301271081, + 0.006769201252609491, + 0.005374044179916382, + 0.005562407895922661, + 0.0029499325901269913, + 0.004740525037050247, + 0.004687619861215353, + 0.0065328567288815975, + 0.00596791785210371, + 0.00395559798926115, + 0.0072904955595731735, + 0.006327854469418526, + 0.0023711244575679302, + 0.006253077182918787, + 0.004201728850603104, + 0.003934146836400032, + 0.003701632609590888, + 0.0022671984042972326, + 0.002953267190605402, + 0.0029971012845635414 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 1, + 4 + ], + "selected_reward": -0.029824240133166313 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a young, athletic male riding a mountain bike?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene set in a vast, grassy park at night?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the rider wearing a thick, dark wool sweater and black, full-finger performance gloves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene illuminated by a high-contrast golden glow from unseen park lights?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera pan steadily while following the rider in a low-angle, long shot?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0017138330731540918 + ], + "opposite_logprob_tokens": [ + -6.376713752746582 + ], + "logprob_reward": -0.0017138330731540918, + "yes_no_margin_reward": 6.375, + "token_credit_reward": -0.0017138330731540918, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.20185871422290802, + 0.1281261295080185, + 0.03396907076239586, + 0.04269516095519066, + 0.08197971433401108, + 0.13075675070285797, + 0.24332326650619507, + 0.06901133805513382, + 0.021810179576277733, + 0.04646965488791466 + ], + "credit_map": [ + 0.025737103074789047, + 0.037083011120557785, + 0.03600894659757614, + 0.01965939626097679, + 0.027387235313653946, + 0.027774695307016373, + 0.015321802347898483, + 0.012886535376310349, + 0.027124755084514618, + 0.02205483429133892, + 0.0048310356214642525, + 0.004491647705435753, + 0.008338875137269497, + 0.05112258717417717, + 0.004785758443176746, + 0.005376636981964111, + 0.003215472213923931, + 0.005570441484451294, + 0.0021085115149617195, + 0.002578436629846692, + 0.006225114222615957, + 0.006276212632656097, + 0.004563644994050264, + 0.0034312386997044086, + 0.003930982667952776, + 0.006734746042639017, + 0.006136263720691204, + 0.006967747583985329, + 0.004368380643427372, + 0.005989537574350834, + 0.004586541559547186, + 0.003980966284871101, + 0.01337739173322916, + 0.009185681119561195, + 0.02257360704243183, + 0.00739506259560585, + 0.007643557153642178, + 0.012100208550691605, + 0.005706045310944319, + 0.003998172469437122, + 0.017483007162809372, + 0.029778409749269485, + 0.01515429001301527, + 0.012826522812247276, + 0.010786649771034718, + 0.02233525924384594, + 0.01344381645321846, + 0.008948804810643196, + 0.027871180325746536, + 0.025330910459160805, + 0.05016276240348816, + 0.04099012166261673, + 0.02093691937625408, + 0.027191925793886185, + 0.03519570082426071, + 0.015643764287233353, + 0.0073677049949765205, + 0.007629154250025749, + 0.010806404985487461, + 0.008812051266431808, + 0.0060469359159469604, + 0.009714733809232712, + 0.010191402398049831, + 0.008442959748208523, + 0.0021764514967799187, + 0.002725900150835514, + 0.0024687957484275103, + 0.002862523542717099, + 0.0035555644426494837, + 0.002733759582042694, + 0.002538014203310013, + 0.0027491713408380747, + 0.004422792233526707, + 0.007256489247083664, + 0.007082476746290922, + 0.006054920144379139, + 0.004187039099633694, + 0.006261554080992937, + 0.00553973950445652, + 0.005664643365889788 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 0, + 5 + ], + "selected_reward": -0.0017138330731540918 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a young, athletic male riding a mountain bike?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene set in a vast, grassy park at night?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the rider wearing a thick, dark wool sweater and black, full-finger performance gloves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene illuminated by a high-contrast golden glow from unseen park lights?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera pan steadily while following the rider in a low-angle, long shot?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03368807211518288 + ], + "opposite_logprob_tokens": [ + -3.4086880683898926 + ], + "logprob_reward": -0.03368807211518288, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.03368807211518288, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1449865847826004, + 0.05760284885764122, + 0.054565832018852234, + 0.14737248420715332, + 0.06539870798587799, + 0.07727374881505966, + 0.2817484140396118, + 0.0706067606806755, + 0.060716137290000916, + 0.039728496223688126 + ], + "credit_map": [ + 0.01743888296186924, + 0.022688692435622215, + 0.020971836522221565, + 0.01133181992918253, + 0.02615549974143505, + 0.02233307622373104, + 0.011632665060460567, + 0.012434097006917, + 0.007052612956613302, + 0.00944901630282402, + 0.00404576538130641, + 0.004400185775011778, + 0.005672608036547899, + 0.014711618423461914, + 0.004576720297336578, + 0.0076943146996200085, + 0.004967539105564356, + 0.008002061396837234, + 0.005527818109840155, + 0.004736107774078846, + 0.010686032474040985, + 0.006471945904195309, + 0.007535399403423071, + 0.006638925988227129, + 0.016932794824242592, + 0.02323862724006176, + 0.015590834431350231, + 0.023690013214945793, + 0.015268467366695404, + 0.01453887578099966, + 0.024380100890994072, + 0.013732768595218658, + 0.00408211350440979, + 0.013102849945425987, + 0.005554624367505312, + 0.006526716519147158, + 0.011808788403868675, + 0.010672853328287601, + 0.009897871874272823, + 0.0037528853863477707, + 0.014915821142494678, + 0.008066745474934578, + 0.0074194190092384815, + 0.007995122112333775, + 0.004906344693154097, + 0.02045820839703083, + 0.008221968077123165, + 0.005290116649121046, + 0.03983374685049057, + 0.03189675882458687, + 0.08883234113454819, + 0.0374753400683403, + 0.017060745507478714, + 0.01517406664788723, + 0.03748598322272301, + 0.01398941408842802, + 0.006734867114573717, + 0.008768616244196892, + 0.012170602567493916, + 0.008498492650687695, + 0.007521616294980049, + 0.009272211231291294, + 0.010585838928818703, + 0.007054513785988092, + 0.006316282320767641, + 0.006548731587827206, + 0.01177317276597023, + 0.010312463156878948, + 0.00432851817458868, + 0.009122294373810291, + 0.007914448156952858, + 0.004400226753205061, + 0.0036411499604582787, + 0.009589840658009052, + 0.005205548834055662, + 0.0044752564281225204, + 0.004206447396427393, + 0.003790014423429966, + 0.004617260303348303, + 0.0042029754258692265 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 3, + 0 + ], + "selected_reward": -0.03368807211518288 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000050/group_13/index_tensor([6668, 6668])_group_13_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000050/group_13/index_tensor([6668, 6668])_group_13_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..ad2217480b7660086418c36d21dfd2d54116b0df --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000050/group_13/index_tensor([6668, 6668])_group_13_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "static, eye-level, wide-angle shot captures a humorous and playful scene in a sun-drenched suburban driveway in front of a two-story house. A young man with tousled light brown hair, dressed in a bright orange hoodie, faded blue jeans, and worn sneakers, perches precariously on a sleek, metal, jet-propelled hoverboard. With their eyes wide open in fear and a determined expression, they awkwardly lean back, shifting their weight and clumsily attempting to twist the device. Instead of moving backward, the hoverboard's side jets splash erratically, sending small puffs of white smoke that cause the board to begin a slow, uncontrolled clockwise spin. The person's arms flail frantically for balance as they spin in a comical circle, their trajectory bringing them dangerously close to a jovial, classic garden gnome in a red, pointy hat, standing stoically at the edge of a manicured green lawn. The edge of the spinning board narrowly missed hitting the unsuspecting dwarf by inches, and their clumsy movements and splashing board created an atmosphere of lighthearted fun before the man continued his dizzying, circular journey across the asphalt.", + "seed": 24572, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000050/group_13/index_6668_seed_24572_dp_26_rank_52_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 13, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.049914642702788115, + "vlm_logprob": -0.049914642702788115, + "vlm_yes_no_margin": 3.75, + "vlm_token_credit": -0.049914642702788115, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot captured from a static, eye-level, wide-angle perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a young man wearing a bright orange hoodie, faded blue jeans, and worn sneakers perched on a metal hoverboard?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the hoverboard emit small puffs of white smoke from its side jets while spinning clockwise?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a jovial garden gnome with a red pointy hat visible standing near the edge of a manicured green lawn?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the scene take place in a sun-drenched suburban driveway in front of a two-story house?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.1486867219209671 + ], + "opposite_logprob_tokens": [ + -2.523686647415161 + ], + "logprob_reward": -0.1486867219209671, + "yes_no_margin_reward": 2.375, + "token_credit_reward": -0.1486867219209671, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.21542277932167053, + 0.16012431681156158, + 0.09664428234100342, + 0.11142728477716446, + 0.10662063211202621, + 0.09100167453289032, + 0.047573283314704895, + 0.06433163583278656, + 0.0599389523267746, + 0.046915147453546524 + ], + "credit_map": [ + 0.024982130154967308, + 0.0348409079015255, + 0.032458335161209106, + 0.04091932997107506, + 0.02161954529583454, + 0.019567906856536865, + 0.02156664989888668, + 0.019467968493700027, + 0.016175689175724983, + 0.015613272786140442, + 0.037889111787080765, + 0.012635520659387112, + 0.012822582386434078, + 0.03232095018029213, + 0.02137780748307705, + 0.011289387010037899, + 0.014803874306380749, + 0.01476225070655346, + 0.01810520887374878, + 0.010782375931739807, + 0.007951385341584682, + 0.011904732324182987, + 0.008830496110022068, + 0.009503953158855438, + 0.010452553629875183, + 0.010924200527369976, + 0.012934624217450619, + 0.01878543011844158, + 0.012096426449716091, + 0.014107449911534786, + 0.014550544321537018, + 0.017576048150658607, + 0.015238520689308643, + 0.010443457402288914, + 0.007898109033703804, + 0.01780889183282852, + 0.01953786425292492, + 0.011453576385974884, + 0.01459487434476614, + 0.009645340032875538, + 0.007060330826789141, + 0.0209781713783741, + 0.006923549808561802, + 0.01144677959382534, + 0.013960275799036026, + 0.009250659495592117, + 0.011295489966869354, + 0.010086415335536003, + 0.0061840880662202835, + 0.006151882000267506, + 0.005388460122048855, + 0.005543088540434837, + 0.005213736090809107, + 0.005066416691988707, + 0.006950990296900272, + 0.007074624300003052, + 0.005410217680037022, + 0.01115504465997219, + 0.007107011508196592, + 0.013166980817914009, + 0.007935826666653156, + 0.0060243261978030205, + 0.007166940253227949, + 0.006365292239934206, + 0.008399340324103832, + 0.010459150187671185, + 0.005270124413073063, + 0.005976757500320673, + 0.00765648391097784, + 0.006980251520872116, + 0.00597120588645339, + 0.009225640445947647, + 0.009657502174377441, + 0.006261443253606558, + 0.004564962815493345, + 0.006048547104001045, + 0.0057566603645682335, + 0.004682895261794329, + 0.006104571744799614, + 0.00383856357075274 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 3 + ], + "selected_reward": -0.1486867219209671 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot captured from a static, eye-level, wide-angle perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a young man wearing a bright orange hoodie, faded blue jeans, and worn sneakers perched on a metal hoverboard?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the hoverboard emit small puffs of white smoke from its side jets while spinning clockwise?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a jovial garden gnome with a red pointy hat visible standing near the edge of a manicured green lawn?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the scene take place in a sun-drenched suburban driveway in front of a two-story house?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.014205172657966614 + ], + "opposite_logprob_tokens": [ + -4.264204978942871 + ], + "logprob_reward": -0.014205172657966614, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.014205172657966614, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07589370757341385, + 0.08183086663484573, + 0.055312469601631165, + 0.07574769109487534, + 0.10012031346559525, + 0.15900827944278717, + 0.0948253646492958, + 0.05289609730243683, + 0.1634753942489624, + 0.14088977873325348 + ], + "credit_map": [ + 0.006721680983901024, + 0.012727498076856136, + 0.010864581912755966, + 0.01265688519924879, + 0.00794108398258686, + 0.007011469919234514, + 0.009803247638046741, + 0.008167260326445103, + 0.009640851058065891, + 0.011304713785648346, + 0.016989750787615776, + 0.0061515443958342075, + 0.007052502129226923, + 0.01116238534450531, + 0.0093883341178298, + 0.010140789672732353, + 0.005873079877346754, + 0.006251052021980286, + 0.00700465589761734, + 0.006843977607786655, + 0.004904735833406448, + 0.007211298681795597, + 0.010804475285112858, + 0.006419199053198099, + 0.007005840074270964, + 0.009904248639941216, + 0.006467065773904324, + 0.014810318127274513, + 0.011885846965014935, + 0.007903283461928368, + 0.008570543490350246, + 0.00920055154711008, + 0.018844425678253174, + 0.009416972286999226, + 0.010622309520840645, + 0.016372397541999817, + 0.01272347941994667, + 0.010714433155953884, + 0.011631879024207592, + 0.009794419631361961, + 0.01031962875276804, + 0.015778647735714912, + 0.015330481342971325, + 0.03211323171854019, + 0.03280942142009735, + 0.01939525082707405, + 0.018270475789904594, + 0.014991141855716705, + 0.007050991058349609, + 0.017995016649365425, + 0.008042103610932827, + 0.02233893983066082, + 0.011421105824410915, + 0.008597885258495808, + 0.011406990699470043, + 0.00797233171761036, + 0.006766419392079115, + 0.008176121860742569, + 0.004455766174942255, + 0.009593467228114605, + 0.006124191917479038, + 0.004899635445326567, + 0.008001272566616535, + 0.004879227373749018, + 0.015215574763715267, + 0.03035730868577957, + 0.04122805967926979, + 0.026371976360678673, + 0.01069612056016922, + 0.016293929889798164, + 0.013808591291308403, + 0.009503846988081932, + 0.012071351520717144, + 0.02987649478018284, + 0.01401526853442192, + 0.021488824859261513, + 0.012684622779488564, + 0.013218237087130547, + 0.027807235717773438, + 0.00972775835543871 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 5, + 9 + ], + "selected_reward": -0.014205172657966614 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot captured from a static, eye-level, wide-angle perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a young man wearing a bright orange hoodie, faded blue jeans, and worn sneakers perched on a metal hoverboard?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the hoverboard emit small puffs of white smoke from its side jets while spinning clockwise?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a jovial garden gnome with a red pointy hat visible standing near the edge of a manicured green lawn?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the scene take place in a sun-drenched suburban driveway in front of a two-story house?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.062024082988500595 + ], + "opposite_logprob_tokens": [ + -2.8120241165161133 + ], + "logprob_reward": -0.062024082988500595, + "yes_no_margin_reward": 2.75, + "token_credit_reward": -0.062024082988500595, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.19894665479660034, + 0.0695914626121521, + 0.08453433960676193, + 0.13279426097869873, + 0.05429957062005997, + 0.07477585971355438, + 0.0773361548781395, + 0.055439699441194534, + 0.17108124494552612, + 0.0812007337808609 + ], + "credit_map": [ + 0.022060386836528778, + 0.029302088543772697, + 0.026235532015562057, + 0.0353962779045105, + 0.02714693173766136, + 0.019206976518034935, + 0.02105839177966118, + 0.018540050834417343, + 0.005383842624723911, + 0.007472643628716469, + 0.007053624372929335, + 0.009489760734140873, + 0.008579661138355732, + 0.0073113529942929745, + 0.015669014304876328, + 0.00863157119601965, + 0.011018010787665844, + 0.01124877855181694, + 0.013517004437744617, + 0.006553086452186108, + 0.010374785400927067, + 0.009451014921069145, + 0.01506068091839552, + 0.007310979068279266, + 0.011176317930221558, + 0.021902263164520264, + 0.011388732120394707, + 0.028991570696234703, + 0.009415101259946823, + 0.015013744123280048, + 0.01565951108932495, + 0.0192470271140337, + 0.003679603338241577, + 0.006114509422332048, + 0.006669275928288698, + 0.009584249928593636, + 0.011027366854250431, + 0.006170021370053291, + 0.006205673795193434, + 0.004848872777074575, + 0.004242233466356993, + 0.022452402859926224, + 0.006787530612200499, + 0.009779361076653004, + 0.008578935638070107, + 0.007067538797855377, + 0.009183473885059357, + 0.006684385240077972, + 0.006389761809259653, + 0.012968158349394798, + 0.006204976700246334, + 0.019080907106399536, + 0.009445210918784142, + 0.008129185996949673, + 0.009551793336868286, + 0.0055661629885435104, + 0.0041979458183050156, + 0.011316132731735706, + 0.007704724557697773, + 0.008899437263607979, + 0.005818867590278387, + 0.005890958942472935, + 0.00547171151265502, + 0.00613992428407073, + 0.01705235429108143, + 0.030043885111808777, + 0.028633005917072296, + 0.03249995782971382, + 0.015846990048885345, + 0.01690107397735119, + 0.018436813727021217, + 0.01166716031730175, + 0.00838661752641201, + 0.01675994321703911, + 0.009397123008966446, + 0.01444141287356615, + 0.006865084171295166, + 0.007933585904538631, + 0.01128390897065401, + 0.006133057177066803 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 3 + ], + "selected_reward": -0.062024082988500595 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot captured from a static, eye-level, wide-angle perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a young man wearing a bright orange hoodie, faded blue jeans, and worn sneakers perched on a metal hoverboard?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the hoverboard emit small puffs of white smoke from its side jets while spinning clockwise?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a jovial garden gnome with a red pointy hat visible standing near the edge of a manicured green lawn?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the scene take place in a sun-drenched suburban driveway in front of a two-story house?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.020570045337080956 + ], + "opposite_logprob_tokens": [ + -3.8955700397491455 + ], + "logprob_reward": -0.020570045337080956, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.020570045337080956, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2318543642759323, + 0.0804053544998169, + 0.08665543794631958, + 0.282686710357666, + 0.056161537766456604, + 0.022213958203792572, + 0.0808986946940422, + 0.05307436361908913, + 0.0585114024579525, + 0.04753812775015831 + ], + "credit_map": [ + 0.02350551448762417, + 0.037147168070077896, + 0.024642420932650566, + 0.040513891726732254, + 0.027575897052884102, + 0.031236259266734123, + 0.02519240230321884, + 0.022040830925107002, + 0.010492525063455105, + 0.013511422090232372, + 0.010838325135409832, + 0.0080558555200696, + 0.011172354221343994, + 0.008203472010791302, + 0.011901909485459328, + 0.006229485850781202, + 0.009695248678326607, + 0.011813356541097164, + 0.00871792621910572, + 0.01258445717394352, + 0.008844994939863682, + 0.012047595344483852, + 0.011338175274431705, + 0.01161368191242218, + 0.020519021898508072, + 0.03765471279621124, + 0.024114137515425682, + 0.06414902955293655, + 0.030614547431468964, + 0.029664013534784317, + 0.037012964487075806, + 0.038958288729190826, + 0.004356682766228914, + 0.007161792833358049, + 0.004697819240391254, + 0.008660998195409775, + 0.009871029295027256, + 0.006821859162300825, + 0.008072566241025925, + 0.006518792826682329, + 0.0022211717441678047, + 0.0030127037316560745, + 0.003399707842618227, + 0.0037093099672347307, + 0.0026436536572873592, + 0.0022030603140592575, + 0.0026527829468250275, + 0.0023715675342828035, + 0.006715265102684498, + 0.01356852799654007, + 0.00828615389764309, + 0.01826825551688671, + 0.008358008228242397, + 0.007517549674957991, + 0.010467552579939365, + 0.0077173844911158085, + 0.004225211683660746, + 0.007179567124694586, + 0.006407823879271746, + 0.01277706678956747, + 0.00501131359487772, + 0.004945254418998957, + 0.007890132255852222, + 0.004637992475181818, + 0.005130557343363762, + 0.004984342493116856, + 0.009889315813779831, + 0.014602643437683582, + 0.00430934876203537, + 0.005740957334637642, + 0.008740570396184921, + 0.005113663617521524, + 0.0063882251270115376, + 0.008307969197630882, + 0.0067887380719184875, + 0.009396795183420181, + 0.004474471788853407, + 0.004894905257970095, + 0.004073077347129583, + 0.003213943913578987 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 0, + 2 + ], + "selected_reward": -0.020570045337080956 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot captured from a static, eye-level, wide-angle perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a young man wearing a bright orange hoodie, faded blue jeans, and worn sneakers perched on a metal hoverboard?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the hoverboard emit small puffs of white smoke from its side jets while spinning clockwise?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a jovial garden gnome with a red pointy hat visible standing near the edge of a manicured green lawn?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the scene take place in a sun-drenched suburban driveway in front of a two-story house?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004087190609425306 + ], + "opposite_logprob_tokens": [ + -5.504086971282959 + ], + "logprob_reward": -0.004087190609425306, + "yes_no_margin_reward": 5.5, + "token_credit_reward": -0.004087190609425306, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2641805410385132, + 0.23435939848423004, + 0.05645003169775009, + 0.08067349344491959, + 0.06940644979476929, + 0.1190660372376442, + 0.04816345125436783, + 0.033535949885845184, + 0.050613753497600555, + 0.04355090484023094 + ], + "credit_map": [ + 0.02682882361114025, + 0.044975243508815765, + 0.03459636867046356, + 0.0507497638463974, + 0.02874644100666046, + 0.021974002942442894, + 0.03299153223633766, + 0.02331836335361004, + 0.01978379115462303, + 0.05147957429289818, + 0.05564654991030693, + 0.022598428651690483, + 0.018567757681012154, + 0.03490257263183594, + 0.01672947220504284, + 0.014651268720626831, + 0.008781922981142998, + 0.0071990094147622585, + 0.006691330578178167, + 0.0053779007866978645, + 0.005336226429790258, + 0.008211063221096992, + 0.009128831326961517, + 0.00572374789044261, + 0.005386741366237402, + 0.00869463849812746, + 0.010835282504558563, + 0.018817702308297157, + 0.007256767712533474, + 0.008708754554390907, + 0.01150721125304699, + 0.009466391988098621, + 0.006422411184757948, + 0.008680964820086956, + 0.005305762402713299, + 0.007117736153304577, + 0.02233666554093361, + 0.00825206283479929, + 0.005665328353643417, + 0.005625512450933456, + 0.007380691356956959, + 0.012822397984564304, + 0.011442787945270538, + 0.021752696484327316, + 0.027022577822208405, + 0.014712116681039333, + 0.013296606950461864, + 0.0106361610814929, + 0.002743388758972287, + 0.005425193812698126, + 0.0041579692624509335, + 0.011609243229031563, + 0.004679771140217781, + 0.005603203549981117, + 0.009366774931550026, + 0.004577907267957926, + 0.003036180045455694, + 0.0048442138358950615, + 0.004802170675247908, + 0.005820154212415218, + 0.0034697777591645718, + 0.0037909122183918953, + 0.004750628024339676, + 0.003021910786628723, + 0.0049130236729979515, + 0.0067741828970611095, + 0.004483070690184832, + 0.005815942771732807, + 0.004357649013400078, + 0.0053542060777544975, + 0.00616797199472785, + 0.012747704982757568, + 0.0051830830052495, + 0.007956835441291332, + 0.005837202072143555, + 0.0060103111900389194, + 0.003721859073266387, + 0.004340638872236013, + 0.0048452867195010185, + 0.0056556896306574345 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 5 + ], + "selected_reward": -0.004087190609425306 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "static, eye-level, wide-angle shot captures a humorous and playful scene in a sun-drenched suburban driveway in front of a two-story house. A young man with tousled light brown hair, dressed in a bright orange hoodie, faded blue jeans, and worn sneakers, perches precariously on a sleek, metal, jet-propelled hoverboard. With their eyes wide open in fear and a determined expression, they awkwardly lean back, shifting their weight and clumsily attempting to twist the device. Instead of moving backward, the hoverboard's side jets splash erratically, sending small puffs of white smoke that cause the board to begin a slow, uncontrolled clockwise spin. The person's arms flail frantically for balance as they spin in a comical circle, their trajectory bringing them dangerously close to a jovial, classic garden gnome in a red, pointy hat, standing stoically at the edge of a manicured green lawn. The edge of the spinning board narrowly missed hitting the unsuspecting dwarf by inches, and their clumsy movements and splashing board created an atmosphere of lighthearted fun before the man continued his dizzying, circular journey across the asphalt.", + "seed": 24573, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000050/group_13/index_6668_seed_24573_dp_26_rank_52_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 13, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.04754730202257633, + "vlm_logprob": -0.04754730202257633, + "vlm_yes_no_margin": 3.75, + "vlm_token_credit": -0.04754730202257633, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot captured from a static, eye-level, wide-angle perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a young man wearing a bright orange hoodie, faded blue jeans, and worn sneakers perched on a metal hoverboard?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the hoverboard emit small puffs of white smoke from its side jets while spinning clockwise?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a jovial garden gnome with a red pointy hat visible standing near the edge of a manicured green lawn?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the scene take place in a sun-drenched suburban driveway in front of a two-story house?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.14663970470428467 + ], + "opposite_logprob_tokens": [ + -2.771639823913574 + ], + "logprob_reward": -0.14663970470428467, + "yes_no_margin_reward": 2.625, + "token_credit_reward": -0.14663970470428467, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13518178462982178, + 0.08582059293985367, + 0.1015424057841301, + 0.12092798948287964, + 0.033400021493434906, + 0.08142159134149551, + 0.07626564055681229, + 0.11851483583450317, + 0.09320839494466782, + 0.15371674299240112 + ], + "credit_map": [ + 0.010429737158119678, + 0.02226966992020607, + 0.011483459733426571, + 0.02635146863758564, + 0.013196667656302452, + 0.011887853033840656, + 0.02293752133846283, + 0.016625406220555305, + 0.004888689611107111, + 0.009828433394432068, + 0.006919453386217356, + 0.01450724434107542, + 0.018023381009697914, + 0.00926294457167387, + 0.01086087990552187, + 0.011529562063515186, + 0.0064740488305687904, + 0.013642246834933758, + 0.01764867454767227, + 0.014937575906515121, + 0.008980624377727509, + 0.008531917817890644, + 0.01723611354827881, + 0.014091203920543194, + 0.014788988046348095, + 0.021183421835303307, + 0.016350245103240013, + 0.016326846554875374, + 0.013293812982738018, + 0.010465506464242935, + 0.01242944784462452, + 0.01608971506357193, + 0.0032225637696683407, + 0.002781691262498498, + 0.00715438649058342, + 0.003720240667462349, + 0.002453617984429002, + 0.00359556102193892, + 0.003159554908052087, + 0.007312405854463577, + 0.008096438832581043, + 0.009250917471945286, + 0.010220831260085106, + 0.009781911037862301, + 0.006585058756172657, + 0.006505484692752361, + 0.012825416401028633, + 0.018155531957745552, + 0.006423790007829666, + 0.007343500852584839, + 0.010880524292588234, + 0.009714494459331036, + 0.008329811505973339, + 0.009839012287557125, + 0.009959682822227478, + 0.013774830847978592, + 0.014306114055216312, + 0.013831560499966145, + 0.016023702919483185, + 0.013911176472902298, + 0.011010954156517982, + 0.009119613096117973, + 0.010322893038392067, + 0.02998882718384266, + 0.009632878936827183, + 0.011178894899785519, + 0.012298361398279667, + 0.016531363129615784, + 0.00837857648730278, + 0.00644616549834609, + 0.013548304326832294, + 0.015193844214081764, + 0.019923632964491844, + 0.025517120957374573, + 0.025270545855164528, + 0.021440746262669563, + 0.010146834887564182, + 0.01145486906170845, + 0.014349035918712616, + 0.025613967329263687 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 0, + 3 + ], + "selected_reward": -0.14663970470428467 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot captured from a static, eye-level, wide-angle perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a young man wearing a bright orange hoodie, faded blue jeans, and worn sneakers perched on a metal hoverboard?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the hoverboard emit small puffs of white smoke from its side jets while spinning clockwise?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a jovial garden gnome with a red pointy hat visible standing near the edge of a manicured green lawn?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the scene take place in a sun-drenched suburban driveway in front of a two-story house?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.011087710037827492 + ], + "opposite_logprob_tokens": [ + -4.511087894439697 + ], + "logprob_reward": -0.011087710037827492, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.011087710037827492, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09885771572589874, + 0.10968445241451263, + 0.20197127759456635, + 0.1772860735654831, + 0.036403581500053406, + 0.06056293845176697, + 0.04512394964694977, + 0.1222088411450386, + 0.058502912521362305, + 0.08939822763204575 + ], + "credit_map": [ + 0.01047136913985014, + 0.012295682914555073, + 0.008988608606159687, + 0.015644626691937447, + 0.008728371933102608, + 0.017401909455657005, + 0.009959333576261997, + 0.015367807820439339, + 0.0067108566872775555, + 0.013346926309168339, + 0.013230742886662483, + 0.02406708151102066, + 0.011529787443578243, + 0.008517473004758358, + 0.018747037276625633, + 0.013534551486372948, + 0.018581906333565712, + 0.030287833884358406, + 0.02367296814918518, + 0.024538327008485794, + 0.01865234412252903, + 0.01236459519714117, + 0.047820691019296646, + 0.026052627712488174, + 0.011954009532928467, + 0.03319167345762253, + 0.03853987157344818, + 0.01329150889068842, + 0.01525522768497467, + 0.012265753000974655, + 0.010224536061286926, + 0.042563509196043015, + 0.004673469346016645, + 0.007371171377599239, + 0.004706898704171181, + 0.005249912850558758, + 0.0033470792695879936, + 0.003566176164895296, + 0.0036557128187268972, + 0.0038331616669893265, + 0.004715082701295614, + 0.005737775471061468, + 0.007530448492616415, + 0.012313547544181347, + 0.008240751922130585, + 0.0040914868004620075, + 0.010279552079737186, + 0.007654294371604919, + 0.005006113555282354, + 0.00594905111938715, + 0.005547435954213142, + 0.006195408292114735, + 0.005665177013725042, + 0.005209842696785927, + 0.004948128014802933, + 0.006602794863283634, + 0.014009237289428711, + 0.016791900619864464, + 0.010691729374229908, + 0.020312251523137093, + 0.016713440418243408, + 0.010553227737545967, + 0.02065170928835869, + 0.01248534768819809, + 0.0054521034471690655, + 0.00815136544406414, + 0.005281362682580948, + 0.00866568461060524, + 0.0065386644564569, + 0.00402039336040616, + 0.01083685178309679, + 0.009556489065289497, + 0.01213495060801506, + 0.008780308999121189, + 0.016522785648703575, + 0.011646769009530544, + 0.010013995692133904, + 0.008979021571576595, + 0.009963561780750751, + 0.011356836184859276 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 3, + 7 + ], + "selected_reward": -0.011087710037827492 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot captured from a static, eye-level, wide-angle perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a young man wearing a bright orange hoodie, faded blue jeans, and worn sneakers perched on a metal hoverboard?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the hoverboard emit small puffs of white smoke from its side jets while spinning clockwise?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a jovial garden gnome with a red pointy hat visible standing near the edge of a manicured green lawn?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the scene take place in a sun-drenched suburban driveway in front of a two-story house?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.05494651198387146 + ], + "opposite_logprob_tokens": [ + -2.9299464225769043 + ], + "logprob_reward": -0.05494651198387146, + "yes_no_margin_reward": 2.875, + "token_credit_reward": -0.05494651198387146, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.16516757011413574, + 0.08370216190814972, + 0.08488350361585617, + 0.1595345139503479, + 0.03323088213801384, + 0.07941751927137375, + 0.03746752440929413, + 0.050936780869960785, + 0.25207728147506714, + 0.05358224734663963 + ], + "credit_map": [ + 0.022106528282165527, + 0.02370673231780529, + 0.019047150388360023, + 0.02909974567592144, + 0.01616065576672554, + 0.013560001738369465, + 0.01949724182486534, + 0.02198951132595539, + 0.005262451712042093, + 0.011022373102605343, + 0.010198792442679405, + 0.01664339005947113, + 0.010212119668722153, + 0.0054778228513896465, + 0.013350507244467735, + 0.01153471041470766, + 0.0070043946616351604, + 0.009542247280478477, + 0.01974904164671898, + 0.01471656747162342, + 0.006124754901975393, + 0.00550460908561945, + 0.010277138091623783, + 0.011964754201471806, + 0.01824590563774109, + 0.020282933488488197, + 0.03417820483446121, + 0.01748582161962986, + 0.011841809377074242, + 0.012907208874821663, + 0.01446189358830452, + 0.03013072907924652, + 0.0034861727617681026, + 0.005805586464703083, + 0.0055751195177435875, + 0.004297043662518263, + 0.0046216826885938644, + 0.002477851929143071, + 0.003711638506501913, + 0.0032557863742113113, + 0.006253935396671295, + 0.007218310609459877, + 0.011367611587047577, + 0.017271339893341064, + 0.004829993937164545, + 0.005117329303175211, + 0.01245373860001564, + 0.014905255287885666, + 0.0029496909119188786, + 0.00330742122605443, + 0.006427248008549213, + 0.003976297099143267, + 0.00366387446410954, + 0.003792728763073683, + 0.006749230902642012, + 0.0066010355949401855, + 0.0046845246106386185, + 0.007791406940668821, + 0.00565527006983757, + 0.0044465623795986176, + 0.008707660250365734, + 0.004928491078317165, + 0.007060202304273844, + 0.007662665098905563, + 0.03202425688505173, + 0.05925720930099487, + 0.022834638133645058, + 0.0301507581025362, + 0.021144771948456764, + 0.018984556198120117, + 0.03545961529016495, + 0.032221466302871704, + 0.007132250815629959, + 0.0082884905859828, + 0.006435364484786987, + 0.007169246673583984, + 0.0058853053487837315, + 0.005449329502880573, + 0.00652202358469367, + 0.006700239144265652 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 0, + 3 + ], + "selected_reward": -0.05494651198387146 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot captured from a static, eye-level, wide-angle perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a young man wearing a bright orange hoodie, faded blue jeans, and worn sneakers perched on a metal hoverboard?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the hoverboard emit small puffs of white smoke from its side jets while spinning clockwise?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a jovial garden gnome with a red pointy hat visible standing near the edge of a manicured green lawn?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the scene take place in a sun-drenched suburban driveway in front of a two-story house?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.012533234432339668 + ], + "opposite_logprob_tokens": [ + -4.387533187866211 + ], + "logprob_reward": -0.012533234432339668, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.012533234432339668, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14814935624599457, + 0.07261648029088974, + 0.0925406962633133, + 0.3921016752719879, + 0.05043930932879448, + 0.07457797229290009, + 0.019918816164135933, + 0.035083357244729996, + 0.05653808265924454, + 0.05803424119949341 + ], + "credit_map": [ + 0.012073107063770294, + 0.023312462493777275, + 0.014806948602199554, + 0.03204193338751793, + 0.016082774847745895, + 0.017051924020051956, + 0.014647319912910461, + 0.018132882192730904, + 0.007206107955425978, + 0.006985198240727186, + 0.010830305516719818, + 0.01283092238008976, + 0.0064338515512645245, + 0.007203775458037853, + 0.011997675523161888, + 0.009128640405833721, + 0.00668426463380456, + 0.01447032019495964, + 0.012092160992324352, + 0.00873884093016386, + 0.009495895355939865, + 0.006191651336848736, + 0.02011081762611866, + 0.014756737276911736, + 0.02439858764410019, + 0.027532538399100304, + 0.11205252259969711, + 0.027643980458378792, + 0.02702181600034237, + 0.02852802351117134, + 0.020960470661520958, + 0.12396372854709625, + 0.00700016412883997, + 0.006046667229384184, + 0.007075341884046793, + 0.0052825165912508965, + 0.005114842671900988, + 0.006643639411777258, + 0.00479374174028635, + 0.008482393808662891, + 0.006308017298579216, + 0.005607471335679293, + 0.012723565101623535, + 0.014495453797280788, + 0.00603064801543951, + 0.004568855743855238, + 0.015116220340132713, + 0.009727728553116322, + 0.0019607548601925373, + 0.002748719183728099, + 0.002666423562914133, + 0.002294303849339485, + 0.002296819118782878, + 0.0020562231075018644, + 0.0025469937827438116, + 0.0033485775347799063, + 0.0044634719379246235, + 0.004852491430938244, + 0.0034558535553514957, + 0.004024686757475138, + 0.00333177438005805, + 0.003521401435136795, + 0.004378342069685459, + 0.007055334281176329, + 0.0059878150932490826, + 0.008296750485897064, + 0.004765134304761887, + 0.007953510619699955, + 0.00415429100394249, + 0.0051839821971952915, + 0.010938199236989021, + 0.009258395060896873, + 0.007587692234665155, + 0.010159039869904518, + 0.006481855642050505, + 0.006331970449537039, + 0.007210531737655401, + 0.00635994877666235, + 0.006427973508834839, + 0.007475226651877165 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 0, + 2 + ], + "selected_reward": -0.012533234432339668 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot captured from a static, eye-level, wide-angle perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a young man wearing a bright orange hoodie, faded blue jeans, and worn sneakers perched on a metal hoverboard?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the hoverboard emit small puffs of white smoke from its side jets while spinning clockwise?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a jovial garden gnome with a red pointy hat visible standing near the edge of a manicured green lawn?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the scene take place in a sun-drenched suburban driveway in front of a two-story house?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.012529348954558372 + ], + "opposite_logprob_tokens": [ + -4.387529373168945 + ], + "logprob_reward": -0.012529348954558372, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.012529348954558372, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1919565200805664, + 0.10826165974140167, + 0.14529021084308624, + 0.1489301174879074, + 0.02523043565452099, + 0.08161665499210358, + 0.040778495371341705, + 0.13158434629440308, + 0.05224881321191788, + 0.07410270720720291 + ], + "credit_map": [ + 0.01893802173435688, + 0.022493064403533936, + 0.01747741363942623, + 0.023094698786735535, + 0.0218735970556736, + 0.03424517810344696, + 0.026404686272144318, + 0.027429845184087753, + 0.016737937927246094, + 0.01129897776991129, + 0.013081631623208523, + 0.011555220931768417, + 0.010498205199837685, + 0.018988575786352158, + 0.012178849428892136, + 0.013922257348895073, + 0.01437487080693245, + 0.01773223839700222, + 0.025497091934084892, + 0.014600628055632114, + 0.01103324443101883, + 0.012276430614292622, + 0.027116920799016953, + 0.022658778354525566, + 0.012212068773806095, + 0.02189008891582489, + 0.03407945856451988, + 0.010949989780783653, + 0.008683833293616772, + 0.012234844267368317, + 0.008731554262340069, + 0.040148280560970306, + 0.004251530393958092, + 0.0035214563831686974, + 0.004056504461914301, + 0.0021159809548407793, + 0.0024429159238934517, + 0.003646265948191285, + 0.0017503531416878104, + 0.0034454273991286755, + 0.009053713642060757, + 0.009676416404545307, + 0.00938352383673191, + 0.009329861961305141, + 0.015867220237851143, + 0.008957984857261181, + 0.009893002919852734, + 0.009454927407205105, + 0.004096865653991699, + 0.003744221990928054, + 0.006179175339639187, + 0.004733189940452576, + 0.003334331326186657, + 0.0047371042892336845, + 0.004346068482846022, + 0.00960754044353962, + 0.009181015193462372, + 0.01473528053611517, + 0.011970814317464828, + 0.01752963662147522, + 0.020848046988248825, + 0.008889894001185894, + 0.026726286858320236, + 0.021703360602259636, + 0.00459621986374259, + 0.010419463738799095, + 0.007254771888256073, + 0.006342506967484951, + 0.00418204115703702, + 0.0031508582178503275, + 0.007003935519605875, + 0.009299016557633877, + 0.008503522723913193, + 0.01349006686359644, + 0.007816102355718613, + 0.008624117821455002, + 0.008090936578810215, + 0.007247292902320623, + 0.00909387692809105, + 0.011236798018217087 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 2 + ], + "selected_reward": -0.012529348954558372 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "static, eye-level, wide-angle shot captures a humorous and playful scene in a sun-drenched suburban driveway in front of a two-story house. A young man with tousled light brown hair, dressed in a bright orange hoodie, faded blue jeans, and worn sneakers, perches precariously on a sleek, metal, jet-propelled hoverboard. With their eyes wide open in fear and a determined expression, they awkwardly lean back, shifting their weight and clumsily attempting to twist the device. Instead of moving backward, the hoverboard's side jets splash erratically, sending small puffs of white smoke that cause the board to begin a slow, uncontrolled clockwise spin. The person's arms flail frantically for balance as they spin in a comical circle, their trajectory bringing them dangerously close to a jovial, classic garden gnome in a red, pointy hat, standing stoically at the edge of a manicured green lawn. The edge of the spinning board narrowly missed hitting the unsuspecting dwarf by inches, and their clumsy movements and splashing board created an atmosphere of lighthearted fun before the man continued his dizzying, circular journey across the asphalt.", + "seed": 24574, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000050/group_13/index_6668_seed_24574_dp_27_rank_54_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 13, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.04189128093421459, + "vlm_logprob": -0.04189128093421459, + "vlm_yes_no_margin": 4.125, + "vlm_token_credit": -0.04189128093421459, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot captured from a static, eye-level, wide-angle perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a young man wearing a bright orange hoodie, faded blue jeans, and worn sneakers perched on a metal hoverboard?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the hoverboard emit small puffs of white smoke from its side jets while spinning clockwise?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a jovial garden gnome with a red pointy hat visible standing near the edge of a manicured green lawn?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the scene take place in a sun-drenched suburban driveway in front of a two-story house?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.11434953659772873 + ], + "opposite_logprob_tokens": [ + -3.739349603652954 + ], + "logprob_reward": -0.11434953659772873, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.11434953659772873, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2381339818239212, + 0.11040519922971725, + 0.14210256934165955, + 0.07038936764001846, + 0.08011138439178467, + 0.07842899858951569, + 0.07652659714221954, + 0.05002680420875549, + 0.11204932630062103, + 0.04182576760649681 + ], + "credit_map": [ + 0.027017870917916298, + 0.025222033262252808, + 0.020416025072336197, + 0.02270806021988392, + 0.04819050058722496, + 0.029589563608169556, + 0.02975994162261486, + 0.03522998467087746, + 0.017687316983938217, + 0.013571633957326412, + 0.012770397588610649, + 0.01805727556347847, + 0.014326261356472969, + 0.01348353736102581, + 0.012758093886077404, + 0.007750684395432472, + 0.007549145258963108, + 0.008103913627564907, + 0.014592813327908516, + 0.015601824969053268, + 0.02213284745812416, + 0.0188261978328228, + 0.04284810647368431, + 0.012447724118828773, + 0.008196982555091381, + 0.009940028190612793, + 0.006028654985129833, + 0.007800316903740168, + 0.010137451812624931, + 0.008913111872971058, + 0.010787448845803738, + 0.00858536921441555, + 0.008890921249985695, + 0.006944053340703249, + 0.01047926489263773, + 0.010860070586204529, + 0.009930984117090702, + 0.011483381502330303, + 0.011736663989722729, + 0.009786045178771019, + 0.007348296698182821, + 0.008434993214905262, + 0.007510035298764706, + 0.010535810142755508, + 0.015506996773183346, + 0.008641351014375687, + 0.01262877881526947, + 0.007822735235095024, + 0.008726600557565689, + 0.01174239907413721, + 0.00833697896450758, + 0.01024648267775774, + 0.009536061435937881, + 0.007518234197050333, + 0.014321734197437763, + 0.006098106037825346, + 0.005569003522396088, + 0.0048443106934428215, + 0.004882544279098511, + 0.005498400889337063, + 0.00532889598980546, + 0.005428073462098837, + 0.009562606923282146, + 0.008912969380617142, + 0.012993448413908482, + 0.011500716209411621, + 0.012708118185400963, + 0.016873866319656372, + 0.01807570271193981, + 0.011857625097036362, + 0.016428044065833092, + 0.011611801572144032, + 0.00489615835249424, + 0.004472797270864248, + 0.005205747205764055, + 0.006552449427545071, + 0.005439225118607283, + 0.00443602679297328, + 0.006858350243419409, + 0.003965011797845364 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 8 + ], + "selected_reward": -0.11434953659772873 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot captured from a static, eye-level, wide-angle perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a young man wearing a bright orange hoodie, faded blue jeans, and worn sneakers perched on a metal hoverboard?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the hoverboard emit small puffs of white smoke from its side jets while spinning clockwise?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a jovial garden gnome with a red pointy hat visible standing near the edge of a manicured green lawn?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the scene take place in a sun-drenched suburban driveway in front of a two-story house?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.016081271693110466 + ], + "opposite_logprob_tokens": [ + -4.1410813331604 + ], + "logprob_reward": -0.016081271693110466, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.016081271693110466, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.24076281487941742, + 0.07780211418867111, + 0.16923099756240845, + 0.10025954246520996, + 0.07997439801692963, + 0.0346575602889061, + 0.1427725851535797, + 0.06376020610332489, + 0.0548749677836895, + 0.035904835909605026 + ], + "credit_map": [ + 0.025580082088708878, + 0.020176755264401436, + 0.028165362775325775, + 0.02224457636475563, + 0.04560576006770134, + 0.030646912753582, + 0.039365362375974655, + 0.0289780143648386, + 0.008327419869601727, + 0.006148096174001694, + 0.005959449801594019, + 0.008335948921740055, + 0.013379781506955624, + 0.00853126123547554, + 0.017349954694509506, + 0.009770200587809086, + 0.009685910306870937, + 0.010215366259217262, + 0.011978021822869778, + 0.01672566495835781, + 0.02517755702137947, + 0.02428347058594227, + 0.05222728103399277, + 0.018937725573778152, + 0.006379588972777128, + 0.007672049570828676, + 0.007528300862759352, + 0.01150510087609291, + 0.009864017367362976, + 0.013326720334589481, + 0.03326302766799927, + 0.010720732621848583, + 0.00596193503588438, + 0.005749890580773354, + 0.007478066720068455, + 0.012337720021605492, + 0.011079161427915096, + 0.010432212613523006, + 0.012879092246294022, + 0.01405632309615612, + 0.0035109189338982105, + 0.00507228821516037, + 0.002806832315400243, + 0.003843056270852685, + 0.004034403711557388, + 0.0037928938400000334, + 0.007931133732199669, + 0.003666036296635866, + 0.012565730139613152, + 0.007845276966691017, + 0.011134468019008636, + 0.010769856162369251, + 0.029654648154973984, + 0.022030925378203392, + 0.032531507313251495, + 0.01624017022550106, + 0.006742298137396574, + 0.005315595306456089, + 0.004927252419292927, + 0.00810379907488823, + 0.007855681702494621, + 0.00811294186860323, + 0.01593961752951145, + 0.006763016805052757, + 0.007097635418176651, + 0.010063672438263893, + 0.0037885354831814766, + 0.007136994507163763, + 0.00673010665923357, + 0.005894016940146685, + 0.008264941163361073, + 0.005899067036807537, + 0.0047112321481108665, + 0.00234665023162961, + 0.003860654542222619, + 0.006372617091983557, + 0.0041441298089921474, + 0.004089908208698034, + 0.006198350805789232, + 0.004181292839348316 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 6 + ], + "selected_reward": -0.016081271693110466 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot captured from a static, eye-level, wide-angle perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a young man wearing a bright orange hoodie, faded blue jeans, and worn sneakers perched on a metal hoverboard?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the hoverboard emit small puffs of white smoke from its side jets while spinning clockwise?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a jovial garden gnome with a red pointy hat visible standing near the edge of a manicured green lawn?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the scene take place in a sun-drenched suburban driveway in front of a two-story house?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.06202968582510948 + ], + "opposite_logprob_tokens": [ + -2.8120296001434326 + ], + "logprob_reward": -0.06202968582510948, + "yes_no_margin_reward": 2.75, + "token_credit_reward": -0.06202968582510948, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1104695275425911, + 0.19283345341682434, + 0.1511974334716797, + 0.17642773687839508, + 0.05545828118920326, + 0.05753553658723831, + 0.13774621486663818, + 0.05479513108730316, + 0.03910532966256142, + 0.02443131059408188 + ], + "credit_map": [ + 0.01580139994621277, + 0.01325216144323349, + 0.011749060824513435, + 0.011918243020772934, + 0.018069850280880928, + 0.013666690327227116, + 0.011765505187213421, + 0.014246620237827301, + 0.010064179077744484, + 0.011583641171455383, + 0.034406453371047974, + 0.05849282070994377, + 0.015414727851748466, + 0.020873557776212692, + 0.020492617040872574, + 0.02150544710457325, + 0.012624134309589863, + 0.008556785993278027, + 0.021326567977666855, + 0.025741059333086014, + 0.02046762965619564, + 0.018459070473909378, + 0.028453117236495018, + 0.015569066628813744, + 0.01510826125741005, + 0.021361807361245155, + 0.015203801915049553, + 0.01590963825583458, + 0.029984911903738976, + 0.020047707483172417, + 0.045825548470020294, + 0.01298606488853693, + 0.006087176967412233, + 0.006186317652463913, + 0.004602136090397835, + 0.006329039577394724, + 0.008037161082029343, + 0.006483016069978476, + 0.011165179312229156, + 0.006568255368620157, + 0.005636528134346008, + 0.0037607294507324696, + 0.004868405405431986, + 0.008325867354869843, + 0.008687538094818592, + 0.0062351590022444725, + 0.01411386113613844, + 0.005907450802624226, + 0.01095426082611084, + 0.007973017171025276, + 0.010871507227420807, + 0.017998211085796356, + 0.020478371530771255, + 0.014367656782269478, + 0.03991848602890968, + 0.015184707939624786, + 0.004732203669846058, + 0.0036850643809884787, + 0.0046574813313782215, + 0.007951084524393082, + 0.0051663280464708805, + 0.00632527656853199, + 0.016112305223941803, + 0.0061653912998735905, + 0.005007236730307341, + 0.0070106456987559795, + 0.0030010677874088287, + 0.00362571538425982, + 0.004242800176143646, + 0.0049283006228506565, + 0.005908946506679058, + 0.005380612798035145, + 0.0020265872590243816, + 0.002413348527625203, + 0.002314985264092684, + 0.00443460326641798, + 0.0033355988562107086, + 0.0028806531336158514, + 0.004735854919999838, + 0.0022896795999258757 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 3, + 2 + ], + "selected_reward": -0.06202968582510948 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot captured from a static, eye-level, wide-angle perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a young man wearing a bright orange hoodie, faded blue jeans, and worn sneakers perched on a metal hoverboard?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the hoverboard emit small puffs of white smoke from its side jets while spinning clockwise?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a jovial garden gnome with a red pointy hat visible standing near the edge of a manicured green lawn?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the scene take place in a sun-drenched suburban driveway in front of a two-story house?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.014183311723172665 + ], + "opposite_logprob_tokens": [ + -4.264183521270752 + ], + "logprob_reward": -0.014183311723172665, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.014183311723172665, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1899208426475525, + 0.1296011507511139, + 0.10653500258922577, + 0.20156380534172058, + 0.112083800137043, + 0.09516210854053497, + 0.045303571969270706, + 0.04677855595946312, + 0.03973500058054924, + 0.03331611305475235 + ], + "credit_map": [ + 0.023164672777056694, + 0.01611771248281002, + 0.0166521817445755, + 0.01598571240901947, + 0.03282018378376961, + 0.02573760785162449, + 0.029069826006889343, + 0.030372949317097664, + 0.011609885841608047, + 0.010594877414405346, + 0.021433861926198006, + 0.02737485058605671, + 0.016614697873592377, + 0.013941848650574684, + 0.017185287550091743, + 0.010845839977264404, + 0.007891989313066006, + 0.007532581686973572, + 0.017103059217333794, + 0.017706114798784256, + 0.012736665084958076, + 0.01267394982278347, + 0.01931193098425865, + 0.011578720062971115, + 0.017110103741288185, + 0.027553793042898178, + 0.019721340388059616, + 0.021318696439266205, + 0.02579476311802864, + 0.024042507633566856, + 0.04247652366757393, + 0.02354607917368412, + 0.01018140185624361, + 0.010096927173435688, + 0.01076616533100605, + 0.020047690719366074, + 0.013977417722344398, + 0.0147487698122859, + 0.017905035987496376, + 0.014360391534864902, + 0.013084501959383488, + 0.007327646482735872, + 0.00848987977951765, + 0.008979633450508118, + 0.02010326273739338, + 0.011429423466324806, + 0.016993297263979912, + 0.008754459209740162, + 0.003962776158004999, + 0.0026646845508366823, + 0.003509491914883256, + 0.0039418358355760574, + 0.010031210258603096, + 0.005943045020103455, + 0.009724493138492107, + 0.005526033695787191, + 0.0054169828072190285, + 0.004889524541795254, + 0.004003807902336121, + 0.004933925345540047, + 0.005137008149176836, + 0.005384400021284819, + 0.01101963222026825, + 0.005993273109197617, + 0.0041810073889791965, + 0.006267951801419258, + 0.0025793337263166904, + 0.0036948653869330883, + 0.004526936914771795, + 0.004827093333005905, + 0.008586538955569267, + 0.005071271676570177, + 0.003701217472553253, + 0.003084353404119611, + 0.004795518238097429, + 0.006197740323841572, + 0.0046003046445548534, + 0.0033917634282261133, + 0.0047078803181648254, + 0.0028373347595334053 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 0, + 1 + ], + "selected_reward": -0.014183311723172665 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot captured from a static, eye-level, wide-angle perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a young man wearing a bright orange hoodie, faded blue jeans, and worn sneakers perched on a metal hoverboard?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the hoverboard emit small puffs of white smoke from its side jets while spinning clockwise?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a jovial garden gnome with a red pointy hat visible standing near the edge of a manicured green lawn?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the scene take place in a sun-drenched suburban driveway in front of a two-story house?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.002812598831951618 + ], + "opposite_logprob_tokens": [ + -5.877812385559082 + ], + "logprob_reward": -0.002812598831951618, + "yes_no_margin_reward": 5.875, + "token_credit_reward": -0.002812598831951618, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1765103042125702, + 0.10601785778999329, + 0.2976926267147064, + 0.061124566942453384, + 0.07765120267868042, + 0.03834410756826401, + 0.07674947381019592, + 0.09400831162929535, + 0.037955500185489655, + 0.033946070820093155 + ], + "credit_map": [ + 0.029475178569555283, + 0.03220902010798454, + 0.013844217173755169, + 0.02079370990395546, + 0.022290803492069244, + 0.021007150411605835, + 0.01945374347269535, + 0.01743645966053009, + 0.01892305165529251, + 0.01364908181130886, + 0.020058056339621544, + 0.011637968011200428, + 0.008107036352157593, + 0.0160337146371603, + 0.01079511921852827, + 0.006813816726207733, + 0.029951920732855797, + 0.03762935847043991, + 0.05054495856165886, + 0.0475546196103096, + 0.046399712562561035, + 0.03566323220729828, + 0.03041401319205761, + 0.01953478343784809, + 0.005695707630366087, + 0.009407797828316689, + 0.005980798043310642, + 0.006662499159574509, + 0.007637097500264645, + 0.007121332921087742, + 0.012356195598840714, + 0.006263131741434336, + 0.00841169897466898, + 0.008344807662069798, + 0.008419578894972801, + 0.013899918645620346, + 0.006020902190357447, + 0.009078294970095158, + 0.00833490677177906, + 0.015141087584197521, + 0.004731372464448214, + 0.005795803386718035, + 0.0034772087819874287, + 0.004584802780300379, + 0.005503559950739145, + 0.0052893455140292645, + 0.004924253560602665, + 0.004037754610180855, + 0.0076745422556996346, + 0.00467816274613142, + 0.006866152863949537, + 0.006309571210294962, + 0.014770920388400555, + 0.013684829697012901, + 0.013556071557104588, + 0.009209216572344303, + 0.01131983008235693, + 0.015562761574983597, + 0.007795340847223997, + 0.009433591738343239, + 0.01284679677337408, + 0.012600940652191639, + 0.014239474199712276, + 0.010209567844867706, + 0.005479849874973297, + 0.004569602198898792, + 0.0037305494770407677, + 0.004889535252004862, + 0.005173513200134039, + 0.004333076067268848, + 0.006484717130661011, + 0.0032946504652500153, + 0.004538584966212511, + 0.0035324471537023783, + 0.0037559112533926964, + 0.004510565660893917, + 0.004196593072265387, + 0.003741310443729162, + 0.005923944991081953, + 0.003746707458049059 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 0, + 1 + ], + "selected_reward": -0.002812598831951618 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "static, eye-level, wide-angle shot captures a humorous and playful scene in a sun-drenched suburban driveway in front of a two-story house. A young man with tousled light brown hair, dressed in a bright orange hoodie, faded blue jeans, and worn sneakers, perches precariously on a sleek, metal, jet-propelled hoverboard. With their eyes wide open in fear and a determined expression, they awkwardly lean back, shifting their weight and clumsily attempting to twist the device. Instead of moving backward, the hoverboard's side jets splash erratically, sending small puffs of white smoke that cause the board to begin a slow, uncontrolled clockwise spin. The person's arms flail frantically for balance as they spin in a comical circle, their trajectory bringing them dangerously close to a jovial, classic garden gnome in a red, pointy hat, standing stoically at the edge of a manicured green lawn. The edge of the spinning board narrowly missed hitting the unsuspecting dwarf by inches, and their clumsy movements and splashing board created an atmosphere of lighthearted fun before the man continued his dizzying, circular journey across the asphalt.", + "seed": 24575, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000050/group_13/index_6668_seed_24575_dp_27_rank_54_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 13, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.05995869347825646, + "vlm_logprob": -0.05995869347825646, + "vlm_yes_no_margin": 3.674999976158142, + "vlm_token_credit": -0.05995869347825646, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot captured from a static, eye-level, wide-angle perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a young man wearing a bright orange hoodie, faded blue jeans, and worn sneakers perched on a metal hoverboard?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the hoverboard emit small puffs of white smoke from its side jets while spinning clockwise?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a jovial garden gnome with a red pointy hat visible standing near the edge of a manicured green lawn?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the scene take place in a sun-drenched suburban driveway in front of a two-story house?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.1933179795742035 + ], + "opposite_logprob_tokens": [ + -2.1933178901672363 + ], + "logprob_reward": -0.1933179795742035, + "yes_no_margin_reward": 1.9999998807907104, + "token_credit_reward": -0.1933179795742035, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10525312274694443, + 0.04184432327747345, + 0.10685347020626068, + 0.04904374107718468, + 0.0635688453912735, + 0.08764152228832245, + 0.12234175205230713, + 0.16574841737747192, + 0.2269466072320938, + 0.03075815737247467 + ], + "credit_map": [ + 0.009949494153261185, + 0.018721582368016243, + 0.011400151066482067, + 0.01366871315985918, + 0.011413575150072575, + 0.01566300541162491, + 0.013532334938645363, + 0.010904256254434586, + 0.00369553011842072, + 0.005070011131465435, + 0.008009642362594604, + 0.009398658759891987, + 0.004258987493813038, + 0.0033059429842978716, + 0.004533584229648113, + 0.003571966663002968, + 0.024999532848596573, + 0.01596534624695778, + 0.011697875335812569, + 0.012050493620336056, + 0.007163494825363159, + 0.012423312291502953, + 0.008827334269881248, + 0.013726075179874897, + 0.005438254214823246, + 0.006933015771210194, + 0.005727988667786121, + 0.004907434340566397, + 0.00622995849698782, + 0.0048410966992378235, + 0.0086461640894413, + 0.006319829262793064, + 0.006248705089092255, + 0.008202367462217808, + 0.007896386086940765, + 0.014937523752450943, + 0.007446457631886005, + 0.0035941442474722862, + 0.00795062631368637, + 0.007292634341865778, + 0.007383154239505529, + 0.009547588415443897, + 0.012661963701248169, + 0.020579868927598, + 0.010646630078554153, + 0.005440548062324524, + 0.011744611896574497, + 0.009637157432734966, + 0.011231941170990467, + 0.013777196407318115, + 0.017573894932866096, + 0.013770765624940395, + 0.02473723329603672, + 0.012561012990772724, + 0.013820961117744446, + 0.014868752099573612, + 0.01302183698862791, + 0.024227580055594444, + 0.02053075097501278, + 0.02534264326095581, + 0.027603065595030785, + 0.011009732261300087, + 0.022247284650802612, + 0.02176552452147007, + 0.05858714506030083, + 0.012997561134397984, + 0.024867907166481018, + 0.008429734967648983, + 0.006222250405699015, + 0.009108695201575756, + 0.0265374556183815, + 0.080195851624012, + 0.002521023852750659, + 0.0031939735636115074, + 0.002935220720246434, + 0.004040845669806004, + 0.005492968484759331, + 0.002888219431042671, + 0.005967243108898401, + 0.003718664636835456 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 7, + 6 + ], + "selected_reward": -0.1933179795742035 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot captured from a static, eye-level, wide-angle perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a young man wearing a bright orange hoodie, faded blue jeans, and worn sneakers perched on a metal hoverboard?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the hoverboard emit small puffs of white smoke from its side jets while spinning clockwise?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a jovial garden gnome with a red pointy hat visible standing near the edge of a manicured green lawn?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the scene take place in a sun-drenched suburban driveway in front of a two-story house?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.014204702340066433 + ], + "opposite_logprob_tokens": [ + -4.264204502105713 + ], + "logprob_reward": -0.014204702340066433, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.014204702340066433, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07317975908517838, + 0.03717329725623131, + 0.1170826256275177, + 0.031356751918792725, + 0.06492052227258682, + 0.10692425817251205, + 0.0795196145772934, + 0.16181740164756775, + 0.26069214940071106, + 0.0673336535692215 + ], + "credit_map": [ + 0.008780013769865036, + 0.011084591038525105, + 0.007725188508629799, + 0.013199293985962868, + 0.008903821930289268, + 0.00822189636528492, + 0.00920337438583374, + 0.006061574909836054, + 0.003490397008135915, + 0.005318220239132643, + 0.004955912008881569, + 0.005536674056202173, + 0.006068172864615917, + 0.0033664219081401825, + 0.004673684481531382, + 0.0037638128269463778, + 0.025587113574147224, + 0.014483303762972355, + 0.014464404433965683, + 0.013343737460672855, + 0.014536497183144093, + 0.011063645593822002, + 0.013560028746724129, + 0.01004389114677906, + 0.0038907008711248636, + 0.004268145654350519, + 0.002875610487535596, + 0.003332599764689803, + 0.004756423179060221, + 0.002924009459093213, + 0.00571243604645133, + 0.0035968266893178225, + 0.005680501461029053, + 0.007757068146020174, + 0.007195362821221352, + 0.012934674508869648, + 0.010762347839772701, + 0.003899663919582963, + 0.009506821632385254, + 0.007184086833149195, + 0.008006159216165543, + 0.008435238152742386, + 0.020598189905285835, + 0.025005239993333817, + 0.008178571239113808, + 0.006906541995704174, + 0.014866904355585575, + 0.014927412383258343, + 0.008142787963151932, + 0.01107565313577652, + 0.01267436146736145, + 0.007877441123127937, + 0.013518806546926498, + 0.008337928913533688, + 0.008158204145729542, + 0.009734425693750381, + 0.013009046204388142, + 0.021253148093819618, + 0.020993642508983612, + 0.022356685250997543, + 0.026790479198098183, + 0.012573350220918655, + 0.02217109501361847, + 0.022669965401291847, + 0.020444152876734734, + 0.014496367424726486, + 0.06713486462831497, + 0.020632373169064522, + 0.007080213166773319, + 0.012211545370519161, + 0.024076731875538826, + 0.09461592137813568, + 0.0048431712202727795, + 0.004360666498541832, + 0.011837305501103401, + 0.00774407060816884, + 0.006897752173244953, + 0.004416909534484148, + 0.011006524786353111, + 0.01622726209461689 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 7, + 2 + ], + "selected_reward": -0.014204702340066433 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot captured from a static, eye-level, wide-angle perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a young man wearing a bright orange hoodie, faded blue jeans, and worn sneakers perched on a metal hoverboard?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the hoverboard emit small puffs of white smoke from its side jets while spinning clockwise?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a jovial garden gnome with a red pointy hat visible standing near the edge of a manicured green lawn?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the scene take place in a sun-drenched suburban driveway in front of a two-story house?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07000669836997986 + ], + "opposite_logprob_tokens": [ + -2.6950066089630127 + ], + "logprob_reward": -0.07000669836997986, + "yes_no_margin_reward": 2.625, + "token_credit_reward": -0.07000669836997986, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0790037140250206, + 0.05658024176955223, + 0.0902840718626976, + 0.040721595287323, + 0.07217579334974289, + 0.06865625828504562, + 0.08122137188911438, + 0.04390931874513626, + 0.4369569718837738, + 0.03049059957265854 + ], + "credit_map": [ + 0.007027898915112019, + 0.011524601839482784, + 0.00735128577798605, + 0.01498886477202177, + 0.008252536877989769, + 0.009830432012677193, + 0.012437628582119942, + 0.007590474095195532, + 0.008277815766632557, + 0.006489339284598827, + 0.008047596551477909, + 0.008429820649325848, + 0.006416182033717632, + 0.0052106683142483234, + 0.00816995557397604, + 0.00553886778652668, + 0.019669577479362488, + 0.012830262072384357, + 0.011468330398201942, + 0.010556857101619244, + 0.009041963145136833, + 0.00846910011023283, + 0.009397418238222599, + 0.008850563317537308, + 0.00583027396351099, + 0.0045153177343308926, + 0.004066652152687311, + 0.0028177592903375626, + 0.006112866569310427, + 0.005494725424796343, + 0.007152526173740625, + 0.0047314781695604324, + 0.006159694865345955, + 0.007495177444070578, + 0.009611752815544605, + 0.01479496993124485, + 0.009883902035653591, + 0.007368206977844238, + 0.009864488616585732, + 0.00699760764837265, + 0.005837979726493359, + 0.008304156363010406, + 0.008587602525949478, + 0.013999974355101585, + 0.0072498987428843975, + 0.00806032307446003, + 0.008924862369894981, + 0.007691464852541685, + 0.007728265132755041, + 0.012730566784739494, + 0.01589766889810562, + 0.007545371539890766, + 0.008745557628571987, + 0.0066867731511592865, + 0.010130823589861393, + 0.0117563521489501, + 0.003779628546908498, + 0.005450474563986063, + 0.006143222562968731, + 0.008344740606844425, + 0.007051866035908461, + 0.00339379720389843, + 0.005209978204220533, + 0.004535616375505924, + 0.09593275934457779, + 0.020992575213313103, + 0.03353145718574524, + 0.016912803053855896, + 0.010082058608531952, + 0.020468059927225113, + 0.05194968357682228, + 0.18708761036396027, + 0.0026464834809303284, + 0.003208297537639737, + 0.004470623563975096, + 0.0043882462196052074, + 0.0034578125923871994, + 0.00304573867470026, + 0.004584955517202616, + 0.004688443150371313 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 2, + 6 + ], + "selected_reward": -0.07000669836997986 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot captured from a static, eye-level, wide-angle perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a young man wearing a bright orange hoodie, faded blue jeans, and worn sneakers perched on a metal hoverboard?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the hoverboard emit small puffs of white smoke from its side jets while spinning clockwise?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a jovial garden gnome with a red pointy hat visible standing near the edge of a manicured green lawn?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the scene take place in a sun-drenched suburban driveway in front of a two-story house?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.018176065757870674 + ], + "opposite_logprob_tokens": [ + -4.018176078796387 + ], + "logprob_reward": -0.018176065757870674, + "yes_no_margin_reward": 4.0, + "token_credit_reward": -0.018176065757870674, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08153348416090012, + 0.09840860962867737, + 0.06666035205125809, + 0.10923050343990326, + 0.0742102786898613, + 0.09282771497964859, + 0.06441016495227814, + 0.1142612174153328, + 0.2560649514198303, + 0.0423927865922451 + ], + "credit_map": [ + 0.006030659191310406, + 0.01068735122680664, + 0.010780071839690208, + 0.012688100337982178, + 0.009707004763185978, + 0.009202074259519577, + 0.012469872832298279, + 0.009968341328203678, + 0.012547262944281101, + 0.00866029504686594, + 0.012613768689334393, + 0.015476120635867119, + 0.016706261783838272, + 0.008661726489663124, + 0.0134971234947443, + 0.010246042162179947, + 0.013462786562740803, + 0.008124380372464657, + 0.00992871355265379, + 0.008773700334131718, + 0.007655822206288576, + 0.005550588946789503, + 0.008262068964540958, + 0.0049022831954061985, + 0.015550601296126842, + 0.012355074286460876, + 0.011621471494436264, + 0.00963984988629818, + 0.010124530643224716, + 0.013755513355135918, + 0.02187349647283554, + 0.014309952966868877, + 0.004760486073791981, + 0.012651782482862473, + 0.00889388844370842, + 0.015561188571155071, + 0.008962629362940788, + 0.004638249520212412, + 0.011665815487504005, + 0.00707622617483139, + 0.00952491071075201, + 0.01245957426726818, + 0.012241004966199398, + 0.017889540642499924, + 0.01148628443479538, + 0.0062119304202497005, + 0.011172082275152206, + 0.011842375621199608, + 0.007095627952367067, + 0.011301537975668907, + 0.010481161996722221, + 0.008835275657474995, + 0.007019610609859228, + 0.004937153309583664, + 0.006668561138212681, + 0.008071225136518478, + 0.008355630561709404, + 0.018651608377695084, + 0.01573658548295498, + 0.023804714903235435, + 0.01364810299128294, + 0.006248759105801582, + 0.015883153304457664, + 0.011932652443647385, + 0.06190422177314758, + 0.013805244117975235, + 0.016844792291522026, + 0.010458928532898426, + 0.006236021872609854, + 0.011329306289553642, + 0.0313662551343441, + 0.1041201502084732, + 0.003333145519718528, + 0.004180127289146185, + 0.004272129852324724, + 0.005899193696677685, + 0.00659521808847785, + 0.002997762057930231, + 0.008009780198335648, + 0.007105427328497171 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 7, + 3 + ], + "selected_reward": -0.018176065757870674 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot captured from a static, eye-level, wide-angle perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a young man wearing a bright orange hoodie, faded blue jeans, and worn sneakers perched on a metal hoverboard?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the hoverboard emit small puffs of white smoke from its side jets while spinning clockwise?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a jovial garden gnome with a red pointy hat visible standing near the edge of a manicured green lawn?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the scene take place in a sun-drenched suburban driveway in front of a two-story house?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004088021349161863 + ], + "opposite_logprob_tokens": [ + -5.504087924957275 + ], + "logprob_reward": -0.004088021349161863, + "yes_no_margin_reward": 5.5, + "token_credit_reward": -0.004088021349161863, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10098469257354736, + 0.12491842359304428, + 0.10693047940731049, + 0.06966721266508102, + 0.11807263642549515, + 0.1416890174150467, + 0.10676351189613342, + 0.060059115290641785, + 0.10134059935808182, + 0.06957431882619858 + ], + "credit_map": [ + 0.0136874970048666, + 0.011562272906303406, + 0.01155396644026041, + 0.013617105782032013, + 0.013985330238938332, + 0.010791170410811901, + 0.013523410074412823, + 0.012263928540050983, + 0.014873557724058628, + 0.02321682497859001, + 0.01787114143371582, + 0.01914067566394806, + 0.020303253084421158, + 0.008736172690987587, + 0.011857224628329277, + 0.008919568732380867, + 0.02513815276324749, + 0.015506897121667862, + 0.01327019277960062, + 0.012234239839017391, + 0.007044615689665079, + 0.01059529185295105, + 0.01040633860975504, + 0.012734740041196346, + 0.005923615302890539, + 0.00801042839884758, + 0.006691420450806618, + 0.007291221991181374, + 0.00936675351113081, + 0.00867034774273634, + 0.015843793749809265, + 0.007869627326726913, + 0.010791696608066559, + 0.01266368106007576, + 0.015275491401553154, + 0.016792992129921913, + 0.02357296273112297, + 0.015686118975281715, + 0.012462764047086239, + 0.01082692388445139, + 0.019963189959526062, + 0.016518929973244667, + 0.028757115826010704, + 0.011593958362936974, + 0.016287602484226227, + 0.015093349851667881, + 0.019162477925419807, + 0.014312384650111198, + 0.00946041475981474, + 0.00915686134248972, + 0.02155498042702675, + 0.015333746559917927, + 0.016108546406030655, + 0.007538365665823221, + 0.013630847446620464, + 0.013979749754071236, + 0.005396439228206873, + 0.005304386373609304, + 0.00756378099322319, + 0.00727364094927907, + 0.014113160781562328, + 0.006605303846299648, + 0.006795272696763277, + 0.00700712576508522, + 0.020432407036423683, + 0.008526709862053394, + 0.03614410012960434, + 0.008545366115868092, + 0.006667619105428457, + 0.0039911409839987755, + 0.008191381581127644, + 0.008841870352625847, + 0.007419389206916094, + 0.005893191322684288, + 0.012543611228466034, + 0.008762144483625889, + 0.006751340348273516, + 0.006390477064996958, + 0.008382238447666168, + 0.013431921601295471 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 1, + 4 + ], + "selected_reward": -0.004088021349161863 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000050/group_14/index_tensor([2691, 2691])_group_14_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000050/group_14/index_tensor([2691, 2691])_group_14_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..99a92ba0b3dbcdfb0259371193d32448b55287da --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000050/group_14/index_tensor([2691, 2691])_group_14_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "medium shot, shot from a slightly upward angle, slowly but steadily advances, drawing the viewer's attention into a spacious military tent. The only illumination comes from the flickering flame of a candle in an antique bronze candlestick at the corner of the table, its unsteady glow casting large, oscillating shadows on the thick canvas walls. A general with a ruddy complexion and a thick, black beard leans over a rough wooden marching table. He wears a heavy, worn suit of dark red metal armor, the beastly patterns carved on his shoulder plates reflecting a dull luster in the firelight. His furrowed brow carves a deep, crisscross pattern on his forehead, and his hawk-like eyes stare intently at a yellowed parchment map spread out on the table. The map's intricate, ink-lined outlines of mountains, rivers, and military markings appear intricate. His gaze wasn't static, but moved slowly and carefully between several key red markers on the map. His hands, braced on the edge of the table, their knuckles white from exertion. His entire posture exuded strength and focus. The flickering candlelight cast bright highlights on his high cheekbones and nose, while deepening the shadows of his eye sockets and cheeks, adding depth to his resolute expression. The entire scene was warm and deep, imbued with ochre, brown, and dark gold, creating an atmosphere of solemnity and intense tension as a decision-maker grappled alone on the eve of a crucial battle.", + "seed": 20352, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000050/group_14/index_2691_seed_20352_dp_28_rank_56_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 14, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.07813309952616691, + "vlm_logprob": -0.07813309952616691, + "vlm_yes_no_margin": 3.2, + "vlm_token_credit": -0.07813309952616691, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set inside a spacious military tent with thick canvas walls?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the lighting provided solely by a flickering candle in an antique bronze candlestick at a table corner?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a general wearing a heavy, worn suit of dark red metal armor with beastly patterns visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the general leaning over a rough wooden table examining a yellowed parchment map with red markers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera slowly and steadily advance from a slightly upward angle into the scene?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03423437103629112 + ], + "opposite_logprob_tokens": [ + -4.034234523773193 + ], + "logprob_reward": -0.03423437103629112, + "yes_no_margin_reward": 4.0, + "token_credit_reward": -0.03423437103629112, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.18509018421173096, + 0.1646612137556076, + 0.041064485907554626, + 0.04456160590052605, + 0.07301734387874603, + 0.060561973601579666, + 0.04399881511926651, + 0.04054246097803116, + 0.166162371635437, + 0.18033960461616516 + ], + "credit_map": [ + 0.010930418036878109, + 0.013663453981280327, + 0.01801368035376072, + 0.026808569207787514, + 0.03549180179834366, + 0.01983022131025791, + 0.037589311599731445, + 0.022762710228562355, + 0.004547423683106899, + 0.0071338508278131485, + 0.00711102457717061, + 0.059594422578811646, + 0.011532539501786232, + 0.007677691988646984, + 0.05819832161068916, + 0.008865924552083015, + 0.0031974781304597855, + 0.003387268865481019, + 0.003304439829662442, + 0.0056150746531784534, + 0.008255058899521828, + 0.004161215852946043, + 0.006172411609441042, + 0.006971536669880152, + 0.0045270901173353195, + 0.005972651299089193, + 0.004825076088309288, + 0.0059017641469836235, + 0.006515227258205414, + 0.0068785762414336205, + 0.005258458200842142, + 0.004682755563408136, + 0.00590139115229249, + 0.01000542938709259, + 0.009395909495651722, + 0.009882189333438873, + 0.008529448881745338, + 0.009505566209554672, + 0.008579523302614689, + 0.011217880994081497, + 0.006535921711474657, + 0.006464436184614897, + 0.0068075996823608875, + 0.008078718557953835, + 0.010357280261814594, + 0.00785274151712656, + 0.007268688175827265, + 0.007196581456810236, + 0.005070852115750313, + 0.004038851708173752, + 0.0047910925932228565, + 0.007163349539041519, + 0.006921599619090557, + 0.003725076327100396, + 0.0073623741045594215, + 0.0049256132915616035, + 0.0028332271613180637, + 0.0024743988178670406, + 0.0036552140954881907, + 0.009930668398737907, + 0.005543472710996866, + 0.003171422053128481, + 0.007554600015282631, + 0.005379452370107174, + 0.014611855149269104, + 0.013147381134331226, + 0.025589941069483757, + 0.025022873654961586, + 0.01631883531808853, + 0.017603913322091103, + 0.03757086023688316, + 0.016296694055199623, + 0.0097217857837677, + 0.006765431724488735, + 0.012362449429929256, + 0.04386262223124504, + 0.012195561081171036, + 0.012127093970775604, + 0.06333018094301224, + 0.01997445523738861 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 8 + ], + "selected_reward": -0.03423437103629112 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set inside a spacious military tent with thick canvas walls?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the lighting provided solely by a flickering candle in an antique bronze candlestick at a table corner?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a general wearing a heavy, worn suit of dark red metal armor with beastly patterns visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the general leaning over a rough wooden table examining a yellowed parchment map with red markers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera slowly and steadily advance from a slightly upward angle into the scene?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07007627934217453 + ], + "opposite_logprob_tokens": [ + -2.6950762271881104 + ], + "logprob_reward": -0.07007627934217453, + "yes_no_margin_reward": 2.625, + "token_credit_reward": -0.07007627934217453, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.18350033462047577, + 0.2332814484834671, + 0.04546372592449188, + 0.11151954531669617, + 0.06094473972916603, + 0.060659490525722504, + 0.06692405790090561, + 0.08710534870624542, + 0.07026825845241547, + 0.08033312857151031 + ], + "credit_map": [ + 0.012616058811545372, + 0.020249569788575172, + 0.013216345570981503, + 0.036353450268507004, + 0.018421409651637077, + 0.02690546214580536, + 0.03013165481388569, + 0.025606384500861168, + 0.011621372774243355, + 0.008344083093106747, + 0.012445325963199139, + 0.0842861607670784, + 0.009351731278002262, + 0.008050769567489624, + 0.08594860881567001, + 0.013233400881290436, + 0.004219825845211744, + 0.0065619745291769505, + 0.0038281353190541267, + 0.005857191514223814, + 0.005948635749518871, + 0.0054402281530201435, + 0.007273775991052389, + 0.006333957426249981, + 0.011390120722353458, + 0.015480592846870422, + 0.013111730106174946, + 0.014868548139929771, + 0.01626233570277691, + 0.01691720075905323, + 0.011325548402965069, + 0.012163459323346615, + 0.006160997785627842, + 0.009447506628930569, + 0.008160244673490524, + 0.005985877942293882, + 0.00981727335602045, + 0.007568607106804848, + 0.006259104236960411, + 0.007545127999037504, + 0.005963221192359924, + 0.006497988477349281, + 0.007226784713566303, + 0.009122761897742748, + 0.011030296795070171, + 0.006617006380110979, + 0.007529037538915873, + 0.00667239585891366, + 0.00597494188696146, + 0.007361986208707094, + 0.009488587267696857, + 0.010652105323970318, + 0.008979292586445808, + 0.006598632782697678, + 0.009462503716349602, + 0.008405998349189758, + 0.007515719160437584, + 0.0070035625249147415, + 0.010055948048830032, + 0.018766043707728386, + 0.009666782803833485, + 0.0070060440339148045, + 0.014465320855379105, + 0.012625925242900848, + 0.005214114207774401, + 0.004089865833520889, + 0.011988267302513123, + 0.009288263507187366, + 0.009225735440850258, + 0.010789202526211739, + 0.010542713105678558, + 0.009130100719630718, + 0.004256397485733032, + 0.0038064797408878803, + 0.010110154747962952, + 0.01732192188501358, + 0.005700332578271627, + 0.00559689337387681, + 0.023402057588100433, + 0.010138895362615585 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 3 + ], + "selected_reward": -0.07007627934217453 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set inside a spacious military tent with thick canvas walls?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the lighting provided solely by a flickering candle in an antique bronze candlestick at a table corner?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a general wearing a heavy, worn suit of dark red metal armor with beastly patterns visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the general leaning over a rough wooden table examining a yellowed parchment map with red markers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera slowly and steadily advance from a slightly upward angle into the scene?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.2520959973335266 + ], + "opposite_logprob_tokens": [ + -1.5020959377288818 + ], + "logprob_reward": -0.2520959973335266, + "yes_no_margin_reward": 1.25, + "token_credit_reward": -0.2520959973335266, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14186258614063263, + 0.3392453193664551, + 0.03743106126785278, + 0.13733692467212677, + 0.031807225197553635, + 0.045748595148324966, + 0.053869009017944336, + 0.05819424241781235, + 0.08529377728700638, + 0.06921127438545227 + ], + "credit_map": [ + 0.009658589959144592, + 0.014546982944011688, + 0.013882464729249477, + 0.018753716722130775, + 0.02047489769756794, + 0.030001971870660782, + 0.018108269199728966, + 0.016435693949460983, + 0.008074050769209862, + 0.01521244179457426, + 0.014167285524308681, + 0.1251910924911499, + 0.023382015526294708, + 0.017073003575205803, + 0.11795881390571594, + 0.018186619505286217, + 0.0025386898778378963, + 0.0043518734164536, + 0.0028573963791131973, + 0.004771044477820396, + 0.006704016588628292, + 0.0037779835984110832, + 0.006015043240040541, + 0.006415013689547777, + 0.01460785698145628, + 0.01938902772963047, + 0.015732085332274437, + 0.018596405163407326, + 0.018236150965094566, + 0.020504821091890335, + 0.01432111021131277, + 0.015949467197060585, + 0.0030241035856306553, + 0.003458509687334299, + 0.004654649645090103, + 0.0034737226087599993, + 0.00493277981877327, + 0.004274727310985327, + 0.003674098988994956, + 0.0043146321550011635, + 0.004198495298624039, + 0.004394490737468004, + 0.004613213706761599, + 0.005031152628362179, + 0.00602608360350132, + 0.007922614924609661, + 0.007331451866775751, + 0.006231090519577265, + 0.005050061736255884, + 0.004208747297525406, + 0.005583202466368675, + 0.008838326670229435, + 0.010365012101829052, + 0.0053105708211660385, + 0.007814967073500156, + 0.006698118057101965, + 0.0038322662003338337, + 0.006202255841344595, + 0.005473138298839331, + 0.0126895597204566, + 0.007033194415271282, + 0.005226312670856714, + 0.010617963969707489, + 0.007119547110050917, + 0.0071598077192902565, + 0.008688990958034992, + 0.01213549729436636, + 0.013168147765100002, + 0.0075389184057712555, + 0.008152451366186142, + 0.020323699340224266, + 0.008126262575387955, + 0.004029072821140289, + 0.0027136881835758686, + 0.006032528355717659, + 0.0162825845181942, + 0.004296576138585806, + 0.004526475910097361, + 0.022823456674814224, + 0.008506895042955875 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 3 + ], + "selected_reward": -0.2520959973335266 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set inside a spacious military tent with thick canvas walls?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the lighting provided solely by a flickering candle in an antique bronze candlestick at a table corner?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a general wearing a heavy, worn suit of dark red metal armor with beastly patterns visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the general leaning over a rough wooden table examining a yellowed parchment map with red markers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera slowly and steadily advance from a slightly upward angle into the scene?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.016076110303401947 + ], + "opposite_logprob_tokens": [ + -4.14107608795166 + ], + "logprob_reward": -0.016076110303401947, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.016076110303401947, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09959184378385544, + 0.21275457739830017, + 0.10989085584878922, + 0.241228386759758, + 0.1262136697769165, + 0.06362303346395493, + 0.028155868873000145, + 0.06905999034643173, + 0.030094848945736885, + 0.01938692107796669 + ], + "credit_map": [ + 0.007080447394400835, + 0.010266191326081753, + 0.008546887896955013, + 0.01462877169251442, + 0.011592874303460121, + 0.013990305364131927, + 0.017520641908049583, + 0.015965720638632774, + 0.01003305334597826, + 0.009759818203747272, + 0.010312735103070736, + 0.07141710072755814, + 0.014291164465248585, + 0.012369262985885143, + 0.07127834856510162, + 0.013293098658323288, + 0.007985529489815235, + 0.010109739378094673, + 0.009420693852007389, + 0.01818825677037239, + 0.014763548970222473, + 0.012089614756405354, + 0.019098490476608276, + 0.0182349793612957, + 0.02977592684328556, + 0.029100783169269562, + 0.03559543564915657, + 0.03258788585662842, + 0.031485337764024734, + 0.03265231102705002, + 0.023813586682081223, + 0.026217130944132805, + 0.014316050335764885, + 0.012138906866312027, + 0.022093603387475014, + 0.014678537845611572, + 0.013837700709700584, + 0.012956889346241951, + 0.016762781888246536, + 0.019429195672273636, + 0.005146663635969162, + 0.007052399218082428, + 0.006906687747687101, + 0.007105830125510693, + 0.010006026364862919, + 0.011227698996663094, + 0.007727885618805885, + 0.008449846878647804, + 0.002382678212597966, + 0.0022601052187383175, + 0.0026528420858085155, + 0.0054255761206150055, + 0.005050464067608118, + 0.002495185937732458, + 0.00419679656624794, + 0.0036922176368534565, + 0.0046142591163516045, + 0.0050007011741399765, + 0.008058776147663593, + 0.016372976824641228, + 0.008767365477979183, + 0.004906290210783482, + 0.013653970323503017, + 0.0076856594532728195, + 0.0025246317964047194, + 0.002642101375386119, + 0.004321145359426737, + 0.0046790349297225475, + 0.004224840085953474, + 0.0026841936632990837, + 0.005164359230548143, + 0.0038545436691492796, + 0.0013593776384368539, + 0.0014931282494217157, + 0.0022821517195552588, + 0.0034587488044053316, + 0.0024781590327620506, + 0.0017113328212872148, + 0.004148886073380709, + 0.0024551376700401306 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 1, + 4 + ], + "selected_reward": -0.016076110303401947 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set inside a spacious military tent with thick canvas walls?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the lighting provided solely by a flickering candle in an antique bronze candlestick at a table corner?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a general wearing a heavy, worn suit of dark red metal armor with beastly patterns visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the general leaning over a rough wooden table examining a yellowed parchment map with red markers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera slowly and steadily advance from a slightly upward angle into the scene?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01818273961544037 + ], + "opposite_logprob_tokens": [ + -4.018182754516602 + ], + "logprob_reward": -0.01818273961544037, + "yes_no_margin_reward": 4.0, + "token_credit_reward": -0.01818273961544037, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08861282467842102, + 0.33543768525123596, + 0.048004478216171265, + 0.040728580206632614, + 0.043737608939409256, + 0.050213735550642014, + 0.055386073887348175, + 0.1252664178609848, + 0.17637158930301666, + 0.03624102473258972 + ], + "credit_map": [ + 0.008532557636499405, + 0.010829628445208073, + 0.008585453033447266, + 0.010630745440721512, + 0.011017934419214725, + 0.01952100172638893, + 0.009102617390453815, + 0.010392885655164719, + 0.0070743439719080925, + 0.012504582293331623, + 0.01355800312012434, + 0.13226284086704254, + 0.015282525680959225, + 0.0105424914509058, + 0.1279861181974411, + 0.01622673310339451, + 0.003060045652091503, + 0.004401002079248428, + 0.003955695312470198, + 0.007497792132198811, + 0.005587266758084297, + 0.005469293799251318, + 0.007945256307721138, + 0.01008811965584755, + 0.0031481462065130472, + 0.00416059885174036, + 0.004067155998200178, + 0.0059680575504899025, + 0.006119914352893829, + 0.005758417770266533, + 0.004833984188735485, + 0.006672303657978773, + 0.0034838481806218624, + 0.006260790396481752, + 0.004433125723153353, + 0.006834708619862795, + 0.004359137266874313, + 0.006382170133292675, + 0.006172937341034412, + 0.005810885690152645, + 0.004314168356359005, + 0.00454716058447957, + 0.004930218681693077, + 0.008145913481712341, + 0.00593307800590992, + 0.008224752731621265, + 0.007402217946946621, + 0.006716218777000904, + 0.005114823114126921, + 0.005122488364577293, + 0.004939514212310314, + 0.00814082846045494, + 0.007846026681363583, + 0.00833180919289589, + 0.00895144697278738, + 0.00693913409486413, + 0.007153333630412817, + 0.004886031150817871, + 0.012361547909677029, + 0.03636413812637329, + 0.011796742677688599, + 0.006179625168442726, + 0.028000827878713608, + 0.0185241661965847, + 0.012239360250532627, + 0.014001806266605854, + 0.028730416670441628, + 0.029585251584649086, + 0.024162964895367622, + 0.014531711116433144, + 0.02804708108305931, + 0.02507300116121769, + 0.00232724379748106, + 0.0018238413613289595, + 0.004704403690993786, + 0.006871618330478668, + 0.003598941722884774, + 0.002983334008604288, + 0.009608070366084576, + 0.004323570057749748 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 8, + 7 + ], + "selected_reward": -0.01818273961544037 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "medium shot, shot from a slightly upward angle, slowly but steadily advances, drawing the viewer's attention into a spacious military tent. The only illumination comes from the flickering flame of a candle in an antique bronze candlestick at the corner of the table, its unsteady glow casting large, oscillating shadows on the thick canvas walls. A general with a ruddy complexion and a thick, black beard leans over a rough wooden marching table. He wears a heavy, worn suit of dark red metal armor, the beastly patterns carved on his shoulder plates reflecting a dull luster in the firelight. His furrowed brow carves a deep, crisscross pattern on his forehead, and his hawk-like eyes stare intently at a yellowed parchment map spread out on the table. The map's intricate, ink-lined outlines of mountains, rivers, and military markings appear intricate. His gaze wasn't static, but moved slowly and carefully between several key red markers on the map. His hands, braced on the edge of the table, their knuckles white from exertion. His entire posture exuded strength and focus. The flickering candlelight cast bright highlights on his high cheekbones and nose, while deepening the shadows of his eye sockets and cheeks, adding depth to his resolute expression. The entire scene was warm and deep, imbued with ochre, brown, and dark gold, creating an atmosphere of solemnity and intense tension as a decision-maker grappled alone on the eve of a crucial battle.", + "seed": 20353, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000050/group_14/index_2691_seed_20353_dp_28_rank_56_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 14, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.04153681602329016, + "vlm_logprob": -0.04153681602329016, + "vlm_yes_no_margin": 3.4999999523162844, + "vlm_token_credit": -0.04153681602329016, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set inside a spacious military tent with thick canvas walls?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the lighting provided solely by a flickering candle in an antique bronze candlestick at a table corner?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a general wearing a heavy, worn suit of dark red metal armor with beastly patterns visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the general leaning over a rough wooden table examining a yellowed parchment map with red markers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera slowly and steadily advance from a slightly upward angle into the scene?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.025842268019914627 + ], + "opposite_logprob_tokens": [ + -4.025842189788818 + ], + "logprob_reward": -0.025842268019914627, + "yes_no_margin_reward": 4.0, + "token_credit_reward": -0.025842268019914627, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3952866792678833, + 0.029539842158555984, + 0.03916043043136597, + 0.14957837760448456, + 0.015615774318575859, + 0.0335364006459713, + 0.16168969869613647, + 0.02976524829864502, + 0.10843203216791153, + 0.037395525723695755 + ], + "credit_map": [ + 0.021780570968985558, + 0.01453486829996109, + 0.14404267072677612, + 0.02890855446457863, + 0.01866736076772213, + 0.010711777955293655, + 0.1338411569595337, + 0.022799693048000336, + 0.002583131892606616, + 0.002520462265238166, + 0.006501074880361557, + 0.0038306100759655237, + 0.003292602254077792, + 0.0017019875813275576, + 0.006572416517883539, + 0.002537552732974291, + 0.0038276633713394403, + 0.004658563528209925, + 0.004927908536046743, + 0.0053834556601941586, + 0.005667507648468018, + 0.0056629544124007225, + 0.0047263638116419315, + 0.004306010436266661, + 0.01862628012895584, + 0.016118928790092468, + 0.01277631800621748, + 0.01626594178378582, + 0.02335289493203163, + 0.021887734532356262, + 0.02334015816450119, + 0.01721012033522129, + 0.002545496914535761, + 0.0014244526391848922, + 0.0020140993874520063, + 0.0019769645296037197, + 0.0023595893289893866, + 0.001660783658735454, + 0.0020754095166921616, + 0.0015589772956445813, + 0.003309218678623438, + 0.0031280643306672573, + 0.0033135011326521635, + 0.006716273725032806, + 0.004220771603286266, + 0.0052342964336276054, + 0.004864479880779982, + 0.002749789971858263, + 0.042474307119846344, + 0.010130730457603931, + 0.01417290885001421, + 0.012223047204315662, + 0.01319029089063406, + 0.007662560790777206, + 0.015075908042490482, + 0.04675993323326111, + 0.004018150269985199, + 0.0034628394059836864, + 0.003281249199062586, + 0.0038036354817450047, + 0.002968910848721862, + 0.0028734933584928513, + 0.006605562288314104, + 0.0027514025568962097, + 0.018113646656274796, + 0.011724178679287434, + 0.017007406800985336, + 0.012401568703353405, + 0.012126955203711987, + 0.011397485621273518, + 0.015513910911977291, + 0.010146863758563995, + 0.0026655548717826605, + 0.003666800446808338, + 0.004350693896412849, + 0.008056125603616238, + 0.0058032735250890255, + 0.00288595468737185, + 0.005264881066977978, + 0.004702238366007805 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 6, + 3 + ], + "selected_reward": -0.025842268019914627 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set inside a spacious military tent with thick canvas walls?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the lighting provided solely by a flickering candle in an antique bronze candlestick at a table corner?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a general wearing a heavy, worn suit of dark red metal armor with beastly patterns visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the general leaning over a rough wooden table examining a yellowed parchment map with red markers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera slowly and steadily advance from a slightly upward angle into the scene?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.06208794564008713 + ], + "opposite_logprob_tokens": [ + -2.8120880126953125 + ], + "logprob_reward": -0.06208794564008713, + "yes_no_margin_reward": 2.75, + "token_credit_reward": -0.06208794564008713, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2702620327472687, + 0.11253374069929123, + 0.04132102429866791, + 0.09727852791547775, + 0.12634044885635376, + 0.11667917668819427, + 0.07530979067087173, + 0.06144403666257858, + 0.036223966628313065, + 0.0626072809100151 + ], + "credit_map": [ + 0.01645476557314396, + 0.010044189170002937, + 0.08727647364139557, + 0.02294214256107807, + 0.01647825539112091, + 0.007748739328235388, + 0.09037714451551437, + 0.018940316513180733, + 0.012013047002255917, + 0.011626201681792736, + 0.015613820403814316, + 0.014581073075532913, + 0.0106678931042552, + 0.009505675174295902, + 0.02244790829718113, + 0.016078125685453415, + 0.00278187682852149, + 0.0054467045702040195, + 0.005195540841668844, + 0.0088213374838233, + 0.004973442293703556, + 0.003149277763441205, + 0.006934420205652714, + 0.004018421284854412, + 0.009685661643743515, + 0.009335783310234547, + 0.011512498371303082, + 0.019976405426859856, + 0.01089491881430149, + 0.008668805472552776, + 0.015019219368696213, + 0.012185237370431423, + 0.029455924406647682, + 0.0068376013077795506, + 0.009780174121260643, + 0.014469354413449764, + 0.014437258243560791, + 0.006999423261731863, + 0.03390798345208168, + 0.01045272871851921, + 0.01273104827851057, + 0.01158694177865982, + 0.011977740563452244, + 0.0212418045848608, + 0.018449164927005768, + 0.011676616966724396, + 0.015345197170972824, + 0.013670666143298149, + 0.006394437979906797, + 0.0034016843419522047, + 0.00430496409535408, + 0.004930751398205757, + 0.007256946060806513, + 0.00468277744948864, + 0.01940954476594925, + 0.024928683415055275, + 0.009032475762069225, + 0.00819434318691492, + 0.008075923658907413, + 0.006773530039936304, + 0.00742230424657464, + 0.005241767503321171, + 0.009752761572599411, + 0.006950938142836094, + 0.003884495934471488, + 0.003594975220039487, + 0.0036152778193354607, + 0.006360429804772139, + 0.006887244526296854, + 0.002467031590640545, + 0.005198386497795582, + 0.0042161233723163605, + 0.005573768634349108, + 0.007298119366168976, + 0.006741863675415516, + 0.01570524089038372, + 0.007885444909334183, + 0.003200199455022812, + 0.008694984950125217, + 0.007507655769586563 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 5 + ], + "selected_reward": -0.06208794564008713 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set inside a spacious military tent with thick canvas walls?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the lighting provided solely by a flickering candle in an antique bronze candlestick at a table corner?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a general wearing a heavy, worn suit of dark red metal armor with beastly patterns visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the general leaning over a rough wooden table examining a yellowed parchment map with red markers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera slowly and steadily advance from a slightly upward angle into the scene?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.08902964740991592 + ], + "opposite_logprob_tokens": [ + -2.464029550552368 + ], + "logprob_reward": -0.08902964740991592, + "yes_no_margin_reward": 2.375, + "token_credit_reward": -0.08902964740991592, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.195156529545784, + 0.08139418810606003, + 0.0895218774676323, + 0.04011786729097366, + 0.24681898951530457, + 0.05469471588730812, + 0.04126670956611633, + 0.06123081222176552, + 0.039231233298778534, + 0.15056709945201874 + ], + "credit_map": [ + 0.027696121484041214, + 0.01612986996769905, + 0.04264916852116585, + 0.01938783749938011, + 0.02345101721584797, + 0.016628853976726532, + 0.027852607890963554, + 0.021361056715250015, + 0.010714113712310791, + 0.00797201320528984, + 0.018483059480786324, + 0.009762498550117016, + 0.00890087429434061, + 0.00756538612768054, + 0.00985060166567564, + 0.008145648054778576, + 0.00907795038074255, + 0.011312798596918583, + 0.011325789615511894, + 0.015728643164038658, + 0.010812344960868359, + 0.007269177585840225, + 0.014263528399169445, + 0.009731645695865154, + 0.004915183410048485, + 0.003718659980222583, + 0.004953516181558371, + 0.00821707770228386, + 0.00574436504393816, + 0.0031034734565764666, + 0.0053930990397930145, + 0.0040724934078752995, + 0.05452233552932739, + 0.01112767867743969, + 0.015561430715024471, + 0.031627971678972244, + 0.017327576875686646, + 0.01103891059756279, + 0.08344385027885437, + 0.02216923236846924, + 0.005968816112726927, + 0.006837405264377594, + 0.007149191573262215, + 0.007309762295335531, + 0.008036097511649132, + 0.004237494897097349, + 0.008699390105903149, + 0.006456557661294937, + 0.01071178074926138, + 0.003060535993427038, + 0.0036208906676620245, + 0.0036490147467702627, + 0.003374506253749132, + 0.0019080648198723793, + 0.007147098425775766, + 0.007794820703566074, + 0.008029524236917496, + 0.004693258553743362, + 0.00708412379026413, + 0.008585596457123756, + 0.00535898981615901, + 0.0049322606064379215, + 0.017451314255595207, + 0.005095750093460083, + 0.005428365897387266, + 0.003764441702514887, + 0.004024779889732599, + 0.007903250865638256, + 0.0061566452495753765, + 0.0015651998110115528, + 0.00580430356785655, + 0.004584246315062046, + 0.009931601583957672, + 0.015082759782671928, + 0.018943680450320244, + 0.0387396439909935, + 0.016290919855237007, + 0.007830103859305382, + 0.023482264950871468, + 0.020266132429242134 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 0, + 9 + ], + "selected_reward": -0.08902964740991592 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set inside a spacious military tent with thick canvas walls?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the lighting provided solely by a flickering candle in an antique bronze candlestick at a table corner?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a general wearing a heavy, worn suit of dark red metal armor with beastly patterns visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the general leaning over a rough wooden table examining a yellowed parchment map with red markers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera slowly and steadily advance from a slightly upward angle into the scene?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.012542181648314 + ], + "opposite_logprob_tokens": [ + -4.387542247772217 + ], + "logprob_reward": -0.012542181648314, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.012542181648314, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.5585674047470093, + 0.05337205156683922, + 0.021014081314206123, + 0.022500168532133102, + 0.1203618049621582, + 0.07627642154693604, + 0.03909013047814369, + 0.016947250813245773, + 0.03754393756389618, + 0.05432671308517456 + ], + "credit_map": [ + 0.028999794274568558, + 0.019936101511120796, + 0.20258931815624237, + 0.04626397788524628, + 0.024421347305178642, + 0.012237410992383957, + 0.19016899168491364, + 0.03395049273967743, + 0.0033970684744417667, + 0.004381954204291105, + 0.011457063257694244, + 0.005198619794100523, + 0.010685646906495094, + 0.004540730733424425, + 0.007540337275713682, + 0.006170631852000952, + 0.001522436155937612, + 0.0028046078514307737, + 0.002299122978001833, + 0.005350286141037941, + 0.0016396467108279467, + 0.0011564681772142649, + 0.003828630782663822, + 0.0024128814693540335, + 0.0022420298773795366, + 0.002041275380179286, + 0.002730026375502348, + 0.00283009628765285, + 0.0038156588561832905, + 0.0030626130755990744, + 0.0037082205526530743, + 0.002070247195661068, + 0.02720213495194912, + 0.004812999162822962, + 0.005806344095617533, + 0.016507679596543312, + 0.006716681178659201, + 0.003699786728248, + 0.04512012377381325, + 0.010496051982045174, + 0.012750130146741867, + 0.009523414075374603, + 0.009026052430272102, + 0.01339710596948862, + 0.005990355275571346, + 0.006848704535514116, + 0.01003718376159668, + 0.008703476749360561, + 0.01029272936284542, + 0.002549896715208888, + 0.003389466553926468, + 0.0031645672861486673, + 0.003093250095844269, + 0.002044334774836898, + 0.0035863686352968216, + 0.010969514958560467, + 0.0019192753825336695, + 0.0018302146345376968, + 0.002686191350221634, + 0.0021703802049160004, + 0.0025074395816773176, + 0.0017402480589225888, + 0.002398093231022358, + 0.0016954102320596576, + 0.006328945513814688, + 0.0037556930910795927, + 0.00492046819999814, + 0.005726636853069067, + 0.004374763462692499, + 0.0025829747319221497, + 0.005545216612517834, + 0.004309241659939289, + 0.003339000977575779, + 0.005172010976821184, + 0.005954232066869736, + 0.014653129503130913, + 0.008357854560017586, + 0.0023380895145237446, + 0.008107803761959076, + 0.006404591258615255 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 5 + ], + "selected_reward": -0.012542181648314 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set inside a spacious military tent with thick canvas walls?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the lighting provided solely by a flickering candle in an antique bronze candlestick at a table corner?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a general wearing a heavy, worn suit of dark red metal armor with beastly patterns visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the general leaning over a rough wooden table examining a yellowed parchment map with red markers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera slowly and steadily advance from a slightly upward angle into the scene?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01818203739821911 + ], + "opposite_logprob_tokens": [ + -4.018181800842285 + ], + "logprob_reward": -0.01818203739821911, + "yes_no_margin_reward": 3.999999761581421, + "token_credit_reward": -0.01818203739821911, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.41656962037086487, + 0.052654508501291275, + 0.05208328366279602, + 0.04148721322417259, + 0.07869067788124084, + 0.07744278013706207, + 0.07957305014133453, + 0.05074780434370041, + 0.08323440700769424, + 0.06751668453216553 + ], + "credit_map": [ + 0.023064371198415756, + 0.01691371761262417, + 0.1398020088672638, + 0.04114135354757309, + 0.02068556845188141, + 0.012125005945563316, + 0.133915975689888, + 0.02892160601913929, + 0.007083434611558914, + 0.0044639320112764835, + 0.005944415926933289, + 0.0063050552271306515, + 0.01073339395225048, + 0.003938283305615187, + 0.007037887815386057, + 0.007148106582462788, + 0.004570935387164354, + 0.005566360894590616, + 0.007886162027716637, + 0.010711310431361198, + 0.006441069301217794, + 0.0036116288974881172, + 0.00815320760011673, + 0.005142606794834137, + 0.0035752879921346903, + 0.0052172034047544, + 0.004725118167698383, + 0.008888708427548409, + 0.00481560779735446, + 0.0034611292649060488, + 0.0058987922966480255, + 0.004905364941805601, + 0.013248053379356861, + 0.0040349154733121395, + 0.005330952815711498, + 0.011917388066649437, + 0.005089593585580587, + 0.0033879808615893126, + 0.02785726822912693, + 0.007824525237083435, + 0.015035402029752731, + 0.009100581519305706, + 0.008753159083425999, + 0.011582070961594582, + 0.0063719600439071655, + 0.005802564788609743, + 0.010128036141395569, + 0.01066900696605444, + 0.006393203977495432, + 0.006075307261198759, + 0.005997917149215937, + 0.006818526890128851, + 0.01170614454895258, + 0.004322963301092386, + 0.019141003489494324, + 0.01911798119544983, + 0.006833397317677736, + 0.0047453478910028934, + 0.0059661162085831165, + 0.004913005046546459, + 0.008463716134428978, + 0.005821205209940672, + 0.00923419650644064, + 0.004770822357386351, + 0.009320956654846668, + 0.009383535012602806, + 0.009509734809398651, + 0.016147669404745102, + 0.012750943191349506, + 0.004671057220548391, + 0.011272233910858631, + 0.0101782763376832, + 0.008148806169629097, + 0.00782431848347187, + 0.007378201000392437, + 0.01104624755680561, + 0.011331532150506973, + 0.0043600620701909065, + 0.009876098483800888, + 0.007551416754722595 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 6 + ], + "selected_reward": -0.01818203739821911 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "medium shot, shot from a slightly upward angle, slowly but steadily advances, drawing the viewer's attention into a spacious military tent. The only illumination comes from the flickering flame of a candle in an antique bronze candlestick at the corner of the table, its unsteady glow casting large, oscillating shadows on the thick canvas walls. A general with a ruddy complexion and a thick, black beard leans over a rough wooden marching table. He wears a heavy, worn suit of dark red metal armor, the beastly patterns carved on his shoulder plates reflecting a dull luster in the firelight. His furrowed brow carves a deep, crisscross pattern on his forehead, and his hawk-like eyes stare intently at a yellowed parchment map spread out on the table. The map's intricate, ink-lined outlines of mountains, rivers, and military markings appear intricate. His gaze wasn't static, but moved slowly and carefully between several key red markers on the map. His hands, braced on the edge of the table, their knuckles white from exertion. His entire posture exuded strength and focus. The flickering candlelight cast bright highlights on his high cheekbones and nose, while deepening the shadows of his eye sockets and cheeks, adding depth to his resolute expression. The entire scene was warm and deep, imbued with ochre, brown, and dark gold, creating an atmosphere of solemnity and intense tension as a decision-maker grappled alone on the eve of a crucial battle.", + "seed": 20354, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000050/group_14/index_2691_seed_20354_dp_29_rank_58_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 14, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.11648594010621309, + "vlm_logprob": -0.11648594010621309, + "vlm_yes_no_margin": 2.6, + "vlm_token_credit": -0.11648594010621309, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set inside a spacious military tent with thick canvas walls?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the lighting provided solely by a flickering candle in an antique bronze candlestick at a table corner?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a general wearing a heavy, worn suit of dark red metal armor with beastly patterns visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the general leaning over a rough wooden table examining a yellowed parchment map with red markers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera slowly and steadily advance from a slightly upward angle into the scene?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.2388831228017807 + ], + "opposite_logprob_tokens": [ + -1.613883137702942 + ], + "logprob_reward": -0.2388831228017807, + "yes_no_margin_reward": 1.375, + "token_credit_reward": -0.2388831228017807, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.30839771032333374, + 0.044200725853443146, + 0.05473104119300842, + 0.11669008433818817, + 0.05425648391246796, + 0.042658060789108276, + 0.06495093554258347, + 0.1424681544303894, + 0.023040704429149628, + 0.14860613644123077 + ], + "credit_map": [ + 0.035933881998062134, + 0.047097235918045044, + 0.036525290459394455, + 0.043066415935754776, + 0.03261273354291916, + 0.039401713758707047, + 0.04667704924941063, + 0.027083396911621094, + 0.004135150462388992, + 0.008278345689177513, + 0.004327032715082169, + 0.006819142494350672, + 0.004676239565014839, + 0.004851199686527252, + 0.0066872648894786835, + 0.00442635640501976, + 0.008674832992255688, + 0.007850541733205318, + 0.007065830752253532, + 0.006183230318129063, + 0.00770328426733613, + 0.005888994317501783, + 0.00493863457813859, + 0.006425695028156042, + 0.02096080780029297, + 0.016148153692483902, + 0.019099444150924683, + 0.01130447443574667, + 0.014847319573163986, + 0.013902882114052773, + 0.009259706363081932, + 0.011167304590344429, + 0.006825527176260948, + 0.00610629515722394, + 0.005585478153079748, + 0.007583686150610447, + 0.004737410228699446, + 0.006532952655106783, + 0.007532852236181498, + 0.009352285414934158, + 0.0053972820751369, + 0.006341319531202316, + 0.0054708439856767654, + 0.004854816012084484, + 0.004181420896202326, + 0.0050672125071287155, + 0.004450860898941755, + 0.006894308608025312, + 0.011411968618631363, + 0.007293795235455036, + 0.010274188593029976, + 0.007056251168251038, + 0.008043814450502396, + 0.008365470916032791, + 0.006386495195329189, + 0.006118957418948412, + 0.01828913949429989, + 0.011957282200455666, + 0.022225603461265564, + 0.01895320415496826, + 0.013837086036801338, + 0.018690109252929688, + 0.01186077855527401, + 0.026654953137040138, + 0.002943194005638361, + 0.002803374780341983, + 0.0037053623236715794, + 0.0029119185637682676, + 0.0020857814233750105, + 0.0030236237216740847, + 0.002861961955204606, + 0.0027054883539676666, + 0.016351331025362015, + 0.02017819508910179, + 0.018833449110388756, + 0.023607999086380005, + 0.016789674758911133, + 0.016986219212412834, + 0.01445972640067339, + 0.02139953523874283 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 7 + ], + "selected_reward": -0.2388831228017807 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set inside a spacious military tent with thick canvas walls?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the lighting provided solely by a flickering candle in an antique bronze candlestick at a table corner?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a general wearing a heavy, worn suit of dark red metal armor with beastly patterns visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the general leaning over a rough wooden table examining a yellowed parchment map with red markers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera slowly and steadily advance from a slightly upward angle into the scene?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07008717209100723 + ], + "opposite_logprob_tokens": [ + -2.695087194442749 + ], + "logprob_reward": -0.07008717209100723, + "yes_no_margin_reward": 2.625, + "token_credit_reward": -0.07008717209100723, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.19991908967494965, + 0.09265138953924179, + 0.22805118560791016, + 0.047657813876867294, + 0.17681634426116943, + 0.06993132829666138, + 0.058576930314302444, + 0.05681895092129707, + 0.023301806300878525, + 0.04627522453665733 + ], + "credit_map": [ + 0.028999129310250282, + 0.025639668107032776, + 0.02790999598801136, + 0.025903692469000816, + 0.022958576679229736, + 0.02659543603658676, + 0.01925385184586048, + 0.022658726200461388, + 0.0181569866836071, + 0.012407112866640091, + 0.014523429796099663, + 0.011377796530723572, + 0.005691148340702057, + 0.011237671598792076, + 0.010649902746081352, + 0.008607341907918453, + 0.03144145756959915, + 0.02522120624780655, + 0.023369502276182175, + 0.02933737076818943, + 0.02032337337732315, + 0.024989059194922447, + 0.0301479771733284, + 0.043221235275268555, + 0.005360770504921675, + 0.006379993166774511, + 0.005089780315756798, + 0.008123615756630898, + 0.004676913842558861, + 0.005208924412727356, + 0.0052198669873178005, + 0.007597945164889097, + 0.016702326014637947, + 0.022157562896609306, + 0.019883759319782257, + 0.022112848237156868, + 0.014563622884452343, + 0.02115434594452381, + 0.025019684806466103, + 0.03522220253944397, + 0.00795079953968525, + 0.00803239643573761, + 0.009611532092094421, + 0.008946374990046024, + 0.00664642546325922, + 0.011456768028438091, + 0.008237257599830627, + 0.009049774147570133, + 0.008459105156362057, + 0.007602234836667776, + 0.0068010566756129265, + 0.006422179751098156, + 0.006241939961910248, + 0.007048537954688072, + 0.005180907901376486, + 0.010820965282619, + 0.011621110141277313, + 0.005091414321213961, + 0.007701762020587921, + 0.006483457051217556, + 0.005427833180874586, + 0.0075522372499108315, + 0.003375810571014881, + 0.0095653235912323, + 0.003279062919318676, + 0.0026912419125437737, + 0.0033611205872148275, + 0.004373891279101372, + 0.0016186353750526905, + 0.0027685295790433884, + 0.002361980499699712, + 0.0028473420534282923, + 0.004128170199692249, + 0.006842236965894699, + 0.006203971803188324, + 0.008855321444571018, + 0.004176451358944178, + 0.00442093051970005, + 0.0071867890655994415, + 0.004461354576051235 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 0, + 4 + ], + "selected_reward": -0.07008717209100723 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set inside a spacious military tent with thick canvas walls?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the lighting provided solely by a flickering candle in an antique bronze candlestick at a table corner?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a general wearing a heavy, worn suit of dark red metal armor with beastly patterns visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the general leaning over a rough wooden table examining a yellowed parchment map with red markers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera slowly and steadily advance from a slightly upward angle into the scene?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.2255621999502182 + ], + "opposite_logprob_tokens": [ + -1.6005622148513794 + ], + "logprob_reward": -0.2255621999502182, + "yes_no_margin_reward": 1.375, + "token_credit_reward": -0.2255621999502182, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1925310492515564, + 0.04726092889904976, + 0.10400263220071793, + 0.06885126233100891, + 0.0687568336725235, + 0.08472064882516861, + 0.056687962263822556, + 0.1308259218931198, + 0.045377206057310104, + 0.20098555088043213 + ], + "credit_map": [ + 0.03273612633347511, + 0.028899498283863068, + 0.020381968468427658, + 0.02564811334013939, + 0.017189525067806244, + 0.02861665189266205, + 0.018754787743091583, + 0.020304374396800995, + 0.004746104124933481, + 0.008212274871766567, + 0.004990539513528347, + 0.005411190912127495, + 0.005466946866363287, + 0.006193895824253559, + 0.007641639560461044, + 0.0045983376912772655, + 0.01348987128585577, + 0.010904744267463684, + 0.010585913434624672, + 0.013155774213373661, + 0.01331288367509842, + 0.01202026940882206, + 0.013646955601871014, + 0.01688622683286667, + 0.008025627583265305, + 0.00755182234570384, + 0.008617106825113297, + 0.01062092650681734, + 0.006216508336365223, + 0.007604959420859814, + 0.00884441938251257, + 0.011369890533387661, + 0.007969630882143974, + 0.009943537414073944, + 0.007312614005059004, + 0.0086115263402462, + 0.005244414322078228, + 0.008564124815165997, + 0.008661244995892048, + 0.01244974322617054, + 0.00828450359404087, + 0.011551911011338234, + 0.01165346335619688, + 0.010499048046767712, + 0.010497557930648327, + 0.012371637858450413, + 0.009105692617595196, + 0.010756829753518105, + 0.006483979988843203, + 0.0059707798063755035, + 0.007804263383150101, + 0.00902778934687376, + 0.006453592795878649, + 0.0060827420093119144, + 0.00705216359347105, + 0.00781265739351511, + 0.013297424651682377, + 0.013591039925813675, + 0.018083946779370308, + 0.027410225942730904, + 0.016234418377280235, + 0.01399233192205429, + 0.01363768707960844, + 0.01457884255796671, + 0.00639542518183589, + 0.005486981011927128, + 0.005411047954112291, + 0.007723112590610981, + 0.0034560724161565304, + 0.004866386763751507, + 0.006115281023085117, + 0.005922897718846798, + 0.019453270360827446, + 0.02038615569472313, + 0.02455565147101879, + 0.029281780123710632, + 0.013812192715704441, + 0.023739315569400787, + 0.021141337230801582, + 0.04861583560705185 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 0, + 7 + ], + "selected_reward": -0.2255621999502182 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set inside a spacious military tent with thick canvas walls?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the lighting provided solely by a flickering candle in an antique bronze candlestick at a table corner?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a general wearing a heavy, worn suit of dark red metal armor with beastly patterns visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the general leaning over a rough wooden table examining a yellowed parchment map with red markers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera slowly and steadily advance from a slightly upward angle into the scene?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.014201411046087742 + ], + "opposite_logprob_tokens": [ + -4.264201641082764 + ], + "logprob_reward": -0.014201411046087742, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.014201411046087742, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1676105558872223, + 0.4358711242675781, + 0.0401974618434906, + 0.024213192984461784, + 0.09975898265838623, + 0.02816598303616047, + 0.03764600679278374, + 0.024609755724668503, + 0.05470038205385208, + 0.08722655475139618 + ], + "credit_map": [ + 0.016881346702575684, + 0.022560792043805122, + 0.018197519704699516, + 0.028308598324656487, + 0.020975027233362198, + 0.026104990392923355, + 0.01742277480661869, + 0.01715949736535549, + 0.0348849855363369, + 0.047573529183864594, + 0.05387117713689804, + 0.07988642901182175, + 0.03889179229736328, + 0.05661415681242943, + 0.06260322034358978, + 0.06154579296708107, + 0.004506185185164213, + 0.0058693247847259045, + 0.003764804219827056, + 0.0050628636963665485, + 0.007209132891148329, + 0.0045972783118486404, + 0.004825939424335957, + 0.004361928440630436, + 0.0023522558622062206, + 0.004528900142759085, + 0.0025646137073636055, + 0.0038097198121249676, + 0.0028757539112120867, + 0.0030438133981078863, + 0.002767072292044759, + 0.0022710622288286686, + 0.009182827547192574, + 0.012052563019096851, + 0.010022582486271858, + 0.013819432817399502, + 0.009922054596245289, + 0.011163122020661831, + 0.014109104871749878, + 0.019487295299768448, + 0.003956962376832962, + 0.0036017734091728926, + 0.003247061977162957, + 0.0030011911876499653, + 0.0025971641298383474, + 0.005584903992712498, + 0.0026502578984946012, + 0.003526665037497878, + 0.003696585074067116, + 0.0055908141657710075, + 0.005558367352932692, + 0.005516222212463617, + 0.0044226269237697124, + 0.004471559543162584, + 0.004864600952714682, + 0.0035252301022410393, + 0.0030450131744146347, + 0.0032245260663330555, + 0.0034039125312119722, + 0.002997948555275798, + 0.002369357505813241, + 0.001987047726288438, + 0.0035758130252361298, + 0.004006136208772659, + 0.006320225540548563, + 0.006016869097948074, + 0.010104639455676079, + 0.0064423405565321445, + 0.004378989804536104, + 0.00569011690095067, + 0.009439332410693169, + 0.006307860370725393, + 0.006070427130907774, + 0.01172273512929678, + 0.012760579586029053, + 0.0198847483843565, + 0.007104027550667524, + 0.008803101256489754, + 0.01156627107411623, + 0.00931465532630682 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 4 + ], + "selected_reward": -0.014201411046087742 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set inside a spacious military tent with thick canvas walls?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the lighting provided solely by a flickering candle in an antique bronze candlestick at a table corner?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a general wearing a heavy, worn suit of dark red metal armor with beastly patterns visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the general leaning over a rough wooden table examining a yellowed parchment map with red markers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera slowly and steadily advance from a slightly upward angle into the scene?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03369579464197159 + ], + "opposite_logprob_tokens": [ + -3.408695697784424 + ], + "logprob_reward": -0.03369579464197159, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.03369579464197159, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1591847985982895, + 0.1893174648284912, + 0.051666922867298126, + 0.06501460075378418, + 0.038689035922288895, + 0.08528430014848709, + 0.08862452208995819, + 0.04397991672158241, + 0.07343603670597076, + 0.20480242371559143 + ], + "credit_map": [ + 0.023241505026817322, + 0.024332299828529358, + 0.021484652534127235, + 0.019490007311105728, + 0.013913823291659355, + 0.021834317594766617, + 0.016020726412534714, + 0.01886746473610401, + 0.020404089242219925, + 0.02201719768345356, + 0.021661553531885147, + 0.03226654976606369, + 0.01743585243821144, + 0.02334713190793991, + 0.024543389678001404, + 0.027641713619232178, + 0.004830647725611925, + 0.0065126265399158, + 0.006589264143258333, + 0.006579638458788395, + 0.006528631318360567, + 0.00626418087631464, + 0.008271213620901108, + 0.006090723443776369, + 0.0069434563629329205, + 0.0065470789559185505, + 0.007948903366923332, + 0.010573765262961388, + 0.006314230151474476, + 0.007218768820166588, + 0.008845306932926178, + 0.01062309555709362, + 0.00485303346067667, + 0.004102759063243866, + 0.003947888500988483, + 0.0045519326813519, + 0.0031800754368305206, + 0.005255354568362236, + 0.005714066792279482, + 0.007083925884217024, + 0.008001729846000671, + 0.008810903877019882, + 0.009944116696715355, + 0.008593865670263767, + 0.006263669114559889, + 0.016313208267092705, + 0.007467367686331272, + 0.01988944038748741, + 0.007133105304092169, + 0.009156289510428905, + 0.00928186159580946, + 0.014181267470121384, + 0.006718614604324102, + 0.011064548045396805, + 0.009140747599303722, + 0.021948091685771942, + 0.0046340483240783215, + 0.004970539826899767, + 0.005671587307006121, + 0.006527346558868885, + 0.004313340876251459, + 0.004872963298112154, + 0.004269643686711788, + 0.008720450103282928, + 0.006942966487258673, + 0.008657724596560001, + 0.009502191096544266, + 0.008299500681459904, + 0.006026540882885456, + 0.010731149464845657, + 0.0100459149107337, + 0.013230051845312119, + 0.02150871604681015, + 0.0221034474670887, + 0.034637950360774994, + 0.03522239997982979, + 0.013627138920128345, + 0.0232307817786932, + 0.02664337120950222, + 0.027828626334667206 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 1, + 0 + ], + "selected_reward": -0.03369579464197159 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "medium shot, shot from a slightly upward angle, slowly but steadily advances, drawing the viewer's attention into a spacious military tent. The only illumination comes from the flickering flame of a candle in an antique bronze candlestick at the corner of the table, its unsteady glow casting large, oscillating shadows on the thick canvas walls. A general with a ruddy complexion and a thick, black beard leans over a rough wooden marching table. He wears a heavy, worn suit of dark red metal armor, the beastly patterns carved on his shoulder plates reflecting a dull luster in the firelight. His furrowed brow carves a deep, crisscross pattern on his forehead, and his hawk-like eyes stare intently at a yellowed parchment map spread out on the table. The map's intricate, ink-lined outlines of mountains, rivers, and military markings appear intricate. His gaze wasn't static, but moved slowly and carefully between several key red markers on the map. His hands, braced on the edge of the table, their knuckles white from exertion. His entire posture exuded strength and focus. The flickering candlelight cast bright highlights on his high cheekbones and nose, while deepening the shadows of his eye sockets and cheeks, adding depth to his resolute expression. The entire scene was warm and deep, imbued with ochre, brown, and dark gold, creating an atmosphere of solemnity and intense tension as a decision-maker grappled alone on the eve of a crucial battle.", + "seed": 20355, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000050/group_14/index_2691_seed_20355_dp_29_rank_58_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 14, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.17961057499051095, + "vlm_logprob": -0.17961057499051095, + "vlm_yes_no_margin": 2.2, + "vlm_token_credit": -0.17961057499051095, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set inside a spacious military tent with thick canvas walls?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the lighting provided solely by a flickering candle in an antique bronze candlestick at a table corner?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a general wearing a heavy, worn suit of dark red metal armor with beastly patterns visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the general leaning over a rough wooden table examining a yellowed parchment map with red markers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera slowly and steadily advance from a slightly upward angle into the scene?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.48565027117729187 + ], + "opposite_logprob_tokens": [ + -0.9856503009796143 + ], + "logprob_reward": -0.48565027117729187, + "yes_no_margin_reward": 0.5, + "token_credit_reward": -0.48565027117729187, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1704479604959488, + 0.062064167112112045, + 0.1171267181634903, + 0.056281689554452896, + 0.12765821814537048, + 0.18758611381053925, + 0.13609908521175385, + 0.06838814169168472, + 0.045176148414611816, + 0.029171716421842575 + ], + "credit_map": [ + 0.024236803874373436, + 0.020127924159169197, + 0.02071862854063511, + 0.025044292211532593, + 0.03481702134013176, + 0.013642966747283936, + 0.0154929393902421, + 0.016367394477128983, + 0.007684031035751104, + 0.010874464176595211, + 0.007374580483883619, + 0.006575512699782848, + 0.01035219244658947, + 0.006234683096408844, + 0.0054984064772725105, + 0.007470303680747747, + 0.01933247782289982, + 0.01503752265125513, + 0.030932234600186348, + 0.009044691920280457, + 0.007444293703883886, + 0.010577062144875526, + 0.010078559629619122, + 0.014679882675409317, + 0.009651967324316502, + 0.005854190792888403, + 0.012652000412344933, + 0.005980098154395819, + 0.004796290770173073, + 0.0046165031380951405, + 0.0036292916629463434, + 0.009101349860429764, + 0.015518037602305412, + 0.021808605641126633, + 0.015675418078899384, + 0.019579334184527397, + 0.01957029290497303, + 0.014667348004877567, + 0.00914021022617817, + 0.011698991060256958, + 0.03181363269686699, + 0.017382090911269188, + 0.03447525575757027, + 0.023717723786830902, + 0.02269638702273369, + 0.01844344474375248, + 0.014016150496900082, + 0.025041446089744568, + 0.01938381791114807, + 0.00946858897805214, + 0.02199370786547661, + 0.01322999969124794, + 0.015879211947321892, + 0.020721733570098877, + 0.011205408722162247, + 0.02421661838889122, + 0.006757041905075312, + 0.006358680315315723, + 0.013102135621011257, + 0.0072733392007648945, + 0.0074204569682478905, + 0.013102762401103973, + 0.005850335583090782, + 0.008523394353687763, + 0.00466376356780529, + 0.00403201300650835, + 0.006586219649761915, + 0.004352737218141556, + 0.007494233548641205, + 0.005067662801593542, + 0.006339467596262693, + 0.006640060339123011, + 0.00468778982758522, + 0.0029970768373459578, + 0.003784832777455449, + 0.002997193019837141, + 0.004182737786322832, + 0.0025946360547095537, + 0.003945920150727034, + 0.003981532994657755 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 0, + 6 + ], + "selected_reward": -0.48565027117729187 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set inside a spacious military tent with thick canvas walls?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the lighting provided solely by a flickering candle in an antique bronze candlestick at a table corner?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a general wearing a heavy, worn suit of dark red metal armor with beastly patterns visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the general leaning over a rough wooden table examining a yellowed parchment map with red markers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera slowly and steadily advance from a slightly upward angle into the scene?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04871722683310509 + ], + "opposite_logprob_tokens": [ + -3.0487172603607178 + ], + "logprob_reward": -0.04871722683310509, + "yes_no_margin_reward": 3.0, + "token_credit_reward": -0.04871722683310509, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.20384283363819122, + 0.13605713844299316, + 0.08146879822015762, + 0.13210569322109222, + 0.0913253203034401, + 0.16391469538211823, + 0.04566439613699913, + 0.04801768437027931, + 0.049999989569187164, + 0.04760349541902542 + ], + "credit_map": [ + 0.030750375241041183, + 0.03022075444459915, + 0.049698907881975174, + 0.01869571954011917, + 0.02339848503470421, + 0.019104793667793274, + 0.01535197626799345, + 0.01662181317806244, + 0.03041614405810833, + 0.021280109882354736, + 0.01678050123155117, + 0.012100597843527794, + 0.022311720997095108, + 0.011951618827879429, + 0.009772525168955326, + 0.01144392043352127, + 0.009130300022661686, + 0.02067604474723339, + 0.010661128908395767, + 0.008647102862596512, + 0.007698459085077047, + 0.010800999589264393, + 0.006417917553335428, + 0.007436837535351515, + 0.020850472152233124, + 0.011801688000559807, + 0.03628666326403618, + 0.011988682672381401, + 0.011305802501738071, + 0.012400214560329914, + 0.010395493358373642, + 0.017076659947633743, + 0.012322794646024704, + 0.012062298133969307, + 0.011237474158406258, + 0.013183197006583214, + 0.015147275291383266, + 0.00949685275554657, + 0.008817760273814201, + 0.009057658724486828, + 0.01803341880440712, + 0.012187143787741661, + 0.02894020639359951, + 0.021964633837342262, + 0.017535775899887085, + 0.01166609302163124, + 0.016212355345487595, + 0.0373750664293766, + 0.005353348795324564, + 0.0037949327379465103, + 0.007847011089324951, + 0.003800863865762949, + 0.005432652775198221, + 0.008707270957529545, + 0.003659042762592435, + 0.007069271523505449, + 0.005134047009050846, + 0.004357608966529369, + 0.007896038703620434, + 0.00502963038161397, + 0.006055076140910387, + 0.008813725784420967, + 0.004869058728218079, + 0.005862492602318525, + 0.0073809013701975346, + 0.004530230071395636, + 0.005033321678638458, + 0.00517246313393116, + 0.01064828597009182, + 0.00537956552579999, + 0.006917379330843687, + 0.004937836900353432, + 0.005198380909860134, + 0.0050322688184678555, + 0.004522943869233131, + 0.005412175320088863, + 0.012520600110292435, + 0.0035756665747612715, + 0.006494401954114437, + 0.004847054835408926 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 5, + 1 + ], + "selected_reward": -0.04871722683310509 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set inside a spacious military tent with thick canvas walls?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the lighting provided solely by a flickering candle in an antique bronze candlestick at a table corner?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a general wearing a heavy, worn suit of dark red metal armor with beastly patterns visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the general leaning over a rough wooden table examining a yellowed parchment map with red markers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera slowly and steadily advance from a slightly upward angle into the scene?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.28131991624832153 + ], + "opposite_logprob_tokens": [ + -1.4063198566436768 + ], + "logprob_reward": -0.28131991624832153, + "yes_no_margin_reward": 1.125, + "token_credit_reward": -0.28131991624832153, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2438744157552719, + 0.0649409294128418, + 0.11367326974868774, + 0.049537256360054016, + 0.07490234822034836, + 0.14374752342700958, + 0.0753006786108017, + 0.07533786445856094, + 0.06778597086668015, + 0.0908997505903244 + ], + "credit_map": [ + 0.019159406423568726, + 0.030344272032380104, + 0.04670005291700363, + 0.022523725405335426, + 0.049471665173769, + 0.021776635199785233, + 0.024432579055428505, + 0.02946610003709793, + 0.007342956960201263, + 0.009862972423434258, + 0.008156487718224525, + 0.006130517460405827, + 0.012868937104940414, + 0.006240542512387037, + 0.008449383080005646, + 0.005889142397791147, + 0.007201874628663063, + 0.011518505401909351, + 0.01589486002922058, + 0.012844173237681389, + 0.03472308814525604, + 0.00876551866531372, + 0.01305942703038454, + 0.00966582726687193, + 0.0031875737477093935, + 0.00412822887301445, + 0.008637798950076103, + 0.005662410520017147, + 0.009008158929646015, + 0.003915603272616863, + 0.00577771058306098, + 0.009219774045050144, + 0.006294891703873873, + 0.008684196509420872, + 0.017059773206710815, + 0.009401146322488785, + 0.00876811146736145, + 0.0056999316439032555, + 0.006518291775137186, + 0.01247600931674242, + 0.00995681993663311, + 0.013783188536763191, + 0.01665276288986206, + 0.03035132586956024, + 0.030609074980020523, + 0.008843256160616875, + 0.013706286437809467, + 0.019844816997647285, + 0.01154019683599472, + 0.007135518826544285, + 0.0085994778200984, + 0.009983062744140625, + 0.014636609703302383, + 0.005768404342234135, + 0.00822121649980545, + 0.009416195563971996, + 0.010550746694207191, + 0.0065416935831308365, + 0.007078635971993208, + 0.009408138692378998, + 0.016118571162223816, + 0.006630257703363895, + 0.011588898487389088, + 0.0074209305457770824, + 0.00982572976499796, + 0.00647104112431407, + 0.006316178012639284, + 0.0069209132343530655, + 0.01454704999923706, + 0.005637892056256533, + 0.009409131482243538, + 0.008658038452267647, + 0.013668953441083431, + 0.010927431285381317, + 0.01017522718757391, + 0.009985202923417091, + 0.019866587594151497, + 0.00554721849039197, + 0.011739471927285194, + 0.008989664725959301 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 5, + 2 + ], + "selected_reward": -0.28131991624832153 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set inside a spacious military tent with thick canvas walls?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the lighting provided solely by a flickering candle in an antique bronze candlestick at a table corner?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a general wearing a heavy, worn suit of dark red metal armor with beastly patterns visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the general leaning over a rough wooden table examining a yellowed parchment map with red markers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera slowly and steadily advance from a slightly upward angle into the scene?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04866897314786911 + ], + "opposite_logprob_tokens": [ + -3.04866886138916 + ], + "logprob_reward": -0.04866897314786911, + "yes_no_margin_reward": 3.0, + "token_credit_reward": -0.04866897314786911, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.16628210246562958, + 0.20655065774917603, + 0.08441732078790665, + 0.04583664983510971, + 0.1312955766916275, + 0.20064957439899445, + 0.0815010741353035, + 0.029614128172397614, + 0.02957463264465332, + 0.024278325960040092 + ], + "credit_map": [ + 0.0182416383177042, + 0.020120013505220413, + 0.016513755545020103, + 0.028879031538963318, + 0.03335591033101082, + 0.013888990506529808, + 0.019627222791314125, + 0.015655528753995895, + 0.019163979217410088, + 0.01953839138150215, + 0.024219267070293427, + 0.02528618648648262, + 0.05566041171550751, + 0.015150107443332672, + 0.02878452092409134, + 0.018747784197330475, + 0.012381626293063164, + 0.011793172918260098, + 0.013033639639616013, + 0.007828830741345882, + 0.014402211643755436, + 0.0074792830273509026, + 0.010845128446817398, + 0.006653424818068743, + 0.003745999652892351, + 0.004369072150439024, + 0.010278640314936638, + 0.003971748519688845, + 0.008948208764195442, + 0.003480315674096346, + 0.004952470306307077, + 0.006090196315199137, + 0.013542858883738518, + 0.013642236590385437, + 0.02258475311100483, + 0.014857958070933819, + 0.032547760754823685, + 0.009458191692829132, + 0.012340391986072063, + 0.012321436777710915, + 0.015531012788414955, + 0.014675932005047798, + 0.028275154531002045, + 0.04156561195850372, + 0.0237527284771204, + 0.012908645905554295, + 0.02179166115820408, + 0.042148832231760025, + 0.01189711969345808, + 0.006918347906321287, + 0.010622052475810051, + 0.010147668421268463, + 0.009930994361639023, + 0.010411200113594532, + 0.006875522434711456, + 0.01469816267490387, + 0.003559795441105962, + 0.0030267878901213408, + 0.005037606228142977, + 0.003947796765714884, + 0.0028423815965652466, + 0.0045315963216125965, + 0.0031642066314816475, + 0.003503957996144891, + 0.00477331830188632, + 0.0025162738747894764, + 0.004166753962635994, + 0.002722486387938261, + 0.0032267242204397917, + 0.005087708588689566, + 0.0033347979187965393, + 0.0037465698551386595, + 0.0029969485476613045, + 0.002600614447146654, + 0.002469971776008606, + 0.003469135146588087, + 0.00414588488638401, + 0.002152407541871071, + 0.0032938357908278704, + 0.0031495282892137766 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 5, + 0 + ], + "selected_reward": -0.04866897314786911 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set inside a spacious military tent with thick canvas walls?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the lighting provided solely by a flickering candle in an antique bronze candlestick at a table corner?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a general wearing a heavy, worn suit of dark red metal armor with beastly patterns visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the general leaning over a rough wooden table examining a yellowed parchment map with red markers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera slowly and steadily advance from a slightly upward angle into the scene?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0336964875459671 + ], + "opposite_logprob_tokens": [ + -3.408696413040161 + ], + "logprob_reward": -0.0336964875459671, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.0336964875459671, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14212411642074585, + 0.07281754910945892, + 0.10065936297178268, + 0.062212344259023666, + 0.1407298445701599, + 0.2539573609828949, + 0.06998424977064133, + 0.049644604325294495, + 0.051696937531232834, + 0.05617361143231392 + ], + "credit_map": [ + 0.017688244581222534, + 0.021940208971500397, + 0.021476350724697113, + 0.014752261340618134, + 0.02092237025499344, + 0.014984894543886185, + 0.014968471601605415, + 0.015391320921480656, + 0.007694825995713472, + 0.009517916478216648, + 0.00781463086605072, + 0.009309901855885983, + 0.01615898311138153, + 0.008159894496202469, + 0.006989746820181608, + 0.007171645760536194, + 0.02432600036263466, + 0.013633490540087223, + 0.01574975624680519, + 0.008738444186747074, + 0.012619750574231148, + 0.010555573739111423, + 0.006183896213769913, + 0.008852456696331501, + 0.006920452229678631, + 0.006474159192293882, + 0.011238424107432365, + 0.009642906486988068, + 0.007979413494467735, + 0.008380676619708538, + 0.005585868842899799, + 0.0059904479421675205, + 0.023643983528017998, + 0.017057212069630623, + 0.020232412964105606, + 0.027351105585694313, + 0.016223464161157608, + 0.012213684618473053, + 0.010327237658202648, + 0.013680757023394108, + 0.02955700270831585, + 0.024273864924907684, + 0.04573862999677658, + 0.031483594328165054, + 0.024409577250480652, + 0.025117574259638786, + 0.023565366864204407, + 0.049811773002147675, + 0.007428099866956472, + 0.007688642479479313, + 0.009513221681118011, + 0.008412241004407406, + 0.014417432248592377, + 0.007543099578469992, + 0.007206919603049755, + 0.007774601224809885, + 0.006439444608986378, + 0.005105447955429554, + 0.005343297962099314, + 0.005894283298403025, + 0.010524758137762547, + 0.00446046469733119, + 0.007291342131793499, + 0.004585565999150276, + 0.00602208124473691, + 0.0045128013007342815, + 0.0057008517906069756, + 0.004249121993780136, + 0.010563070885837078, + 0.005774311255663633, + 0.008027992211282253, + 0.006846710108220577, + 0.009046487510204315, + 0.0056470949202775955, + 0.00768181262537837, + 0.007131215650588274, + 0.008969251997768879, + 0.004454292822629213, + 0.007873233407735825, + 0.0053702229633927345 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 0, + 4 + ], + "selected_reward": -0.0336964875459671 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000050/group_15/index_tensor([4495, 4495])_group_15_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000050/group_15/index_tensor([4495, 4495])_group_15_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..5e4718677654302a89149e50f86161a6a325a850 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000050/group_15/index_tensor([4495, 4495])_group_15_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "frantic montage of lightning-fast pans and jump cuts plunges the viewer directly into the catastrophic battle at the heart of a swirling, inky black storm cloud. The camera switches dizzyingly between the claws and scales of two enormous mythical creatures, creating a sense of ultimate chaos and power. At one moment, the frame is filled with massive cyan dragon scales, each shimmering with iridescence like ancient jade, emitting dazzling sparks in a fierce collision with a ferocious tiger. Then, in an instant, the camera cuts to an extreme close-up, revealing a gleaming, scythe-like white tiger claw, ripping into the dragon's body with ferocious force, its claws digging deep into the cracks between its scales. The entire frenzied scene is illuminated only by intermittent, blinding flashes of lightning, each one freezing the brutal impact in high-contrast black and white silhouettes. We see the Azure Dragon's massive, serpentine form coiling and tightening amidst the howling wind and pounding rain, attempting to strangle its opponent; while the equally massive Tiger, its snow-white fur streaked with black lightning, dodges and strikes back with incredible agility, its amber eyes burning with pure fury. This disorienting, rapid-fire switching between gleaming scales, bloodied fur, snarling maws, and blazing dragon pupils creates a heart-stopping atmosphere of primal, mythical fury.", + "seed": 30379, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000050/group_15/index_4495_seed_30379_dp_30_rank_60_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 15, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.5145658927038312, + "vlm_logprob": -0.5145658927038312, + "vlm_yes_no_margin": 1.8, + "vlm_token_credit": -0.5145658927038312, + "vlm_yes_ratio": 0.8, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are two enormous mythical creatures, specifically a cyan dragon and a white tiger, visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the environment a swirling, inky black storm cloud illuminated by intermittent lightning flashes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the video feature frantic montage editing with lightning-fast pans and jump cuts?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the dragon depicted with massive cyan scales that shimmer with iridescence like ancient jade?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the tiger depicted with snow-white fur streaked with black lightning and amber eyes?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.5247058868408203 + ], + "opposite_logprob_tokens": [ + -0.8997058868408203 + ], + "logprob_reward": -0.5247058868408203, + "yes_no_margin_reward": 0.375, + "token_credit_reward": -0.5247058868408203, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.6454416513442993, + 0.029892532154917717, + 0.03713642805814743, + 0.08403255045413971, + 0.03172267973423004, + 0.014833993278443813, + 0.043294426053762436, + 0.04517213627696037, + 0.045027557760477066, + 0.023446017876267433 + ], + "credit_map": [ + 0.036113280802965164, + 0.016305310651659966, + 0.016908347606658936, + 0.0150736253708601, + 0.26855871081352234, + 0.04176729544997215, + 0.07086960971355438, + 0.17984548211097717, + 0.0029682423919439316, + 0.0029207533225417137, + 0.0036837265361100435, + 0.0031054862774908543, + 0.003494205651804805, + 0.0038062776438891888, + 0.0036423325072973967, + 0.0062715099193155766, + 0.00381494895555079, + 0.006150358356535435, + 0.0031060443725436926, + 0.0031279060058295727, + 0.006030889693647623, + 0.006968855857849121, + 0.0032174333464354277, + 0.0047199921682477, + 0.00833736453205347, + 0.01201429683715105, + 0.007368763908743858, + 0.011850139126181602, + 0.01803061179816723, + 0.005134581588208675, + 0.011103393509984016, + 0.010193399153649807, + 0.0034143466036766768, + 0.004948199260979891, + 0.0032829444389790297, + 0.004706673789769411, + 0.004400194156914949, + 0.0029196504037827253, + 0.003725434886291623, + 0.004325234796851873, + 0.001951440004631877, + 0.0021008087787777185, + 0.0020056478679180145, + 0.0018203151412308216, + 0.001696362509392202, + 0.0015632717404514551, + 0.001803834573365748, + 0.0018923132447525859, + 0.0036044924054294825, + 0.003159080632030964, + 0.005050739273428917, + 0.005751205142587423, + 0.008103037253022194, + 0.004449838772416115, + 0.007287710439413786, + 0.005888322368264198, + 0.002841064939275384, + 0.004560746718198061, + 0.007977803237736225, + 0.007270104717463255, + 0.0057053775526583195, + 0.0043348101899027824, + 0.005112427286803722, + 0.007369804196059704, + 0.003446504706516862, + 0.008131601847708225, + 0.004939322825521231, + 0.007610119879245758, + 0.006454722955822945, + 0.005330557469278574, + 0.004111460410058498, + 0.005003270227462053, + 0.005254062823951244, + 0.0020190412178635597, + 0.002602197462692857, + 0.0019543843809515238, + 0.0021872303914278746, + 0.002896778052672744, + 0.002175664296373725, + 0.0043566590175032616 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 7 + ], + "selected_reward": -0.5247058868408203 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are two enormous mythical creatures, specifically a cyan dragon and a white tiger, visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the environment a swirling, inky black storm cloud illuminated by intermittent lightning flashes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the video feature frantic montage editing with lightning-fast pans and jump cuts?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the dragon depicted with massive cyan scales that shimmer with iridescence like ancient jade?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the tiger depicted with snow-white fur streaked with black lightning and amber eyes?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.10032995045185089 + ], + "opposite_logprob_tokens": [ + -2.350329875946045 + ], + "logprob_reward": -0.10032995045185089, + "yes_no_margin_reward": 2.25, + "token_credit_reward": -0.10032995045185089, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.587604820728302, + 0.05434050410985947, + 0.11139614880084991, + 0.10403945297002792, + 0.025244295597076416, + 0.026499897241592407, + 0.047583118081092834, + 0.01374548114836216, + 0.016020892187952995, + 0.01352540124207735 + ], + "credit_map": [ + 0.03208578750491142, + 0.01715734787285328, + 0.017669733613729477, + 0.016124187037348747, + 0.23915879428386688, + 0.0387081541121006, + 0.06355296820402145, + 0.16314783692359924, + 0.005055873189121485, + 0.006668988149613142, + 0.010702681727707386, + 0.006485334597527981, + 0.006914576515555382, + 0.007742295507341623, + 0.004506214056164026, + 0.006264532450586557, + 0.01515064388513565, + 0.010445632971823215, + 0.00842717383056879, + 0.008285022340714931, + 0.015275252982974052, + 0.02862791158258915, + 0.010049092583358288, + 0.015135404653847218, + 0.008936099708080292, + 0.013824312016367912, + 0.00785080622881651, + 0.014912436716258526, + 0.023271238431334496, + 0.008629892021417618, + 0.014226755127310753, + 0.01238789688795805, + 0.0024621058255434036, + 0.0030182800255715847, + 0.0029092023614794016, + 0.004516984801739454, + 0.00276687229052186, + 0.002932317089289427, + 0.0033650118857622147, + 0.0032735196873545647, + 0.0029833836015313864, + 0.0042436192743480206, + 0.0033178937155753374, + 0.003459879895672202, + 0.0031778980046510696, + 0.0029528222512453794, + 0.003225793829187751, + 0.0031386022455990314, + 0.004138424061238766, + 0.004827491473406553, + 0.006446403451263905, + 0.007240424398332834, + 0.005543711129575968, + 0.00586263183504343, + 0.006431182846426964, + 0.007092844229191542, + 0.0011809960706159472, + 0.0015146820805966854, + 0.002180890180170536, + 0.0017448891885578632, + 0.0013533032033592463, + 0.0015340722166001797, + 0.0018284347606822848, + 0.0024082111194729805, + 0.0015291150193661451, + 0.0024487758055329323, + 0.001980302855372429, + 0.0022213042248040438, + 0.0022960002534091473, + 0.001925453543663025, + 0.0017580819549039006, + 0.0018618573667481542, + 0.0011938604293391109, + 0.0014435159973800182, + 0.0020582638680934906, + 0.002190931234508753, + 0.0012309856247156858, + 0.0010317506967112422, + 0.0029442692175507545, + 0.0014318212633952498 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 3 + ], + "selected_reward": -0.10032995045185089 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are two enormous mythical creatures, specifically a cyan dragon and a white tiger, visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the environment a swirling, inky black storm cloud illuminated by intermittent lightning flashes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the video feature frantic montage editing with lightning-fast pans and jump cuts?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the dragon depicted with massive cyan scales that shimmer with iridescence like ancient jade?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the tiger depicted with snow-white fur streaked with black lightning and amber eyes?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -1.9103368520736694 + ], + "opposite_logprob_tokens": [ + -0.16033682227134705 + ], + "logprob_reward": -1.9103368520736694, + "yes_no_margin_reward": -1.75, + "token_credit_reward": -1.9103368520736694, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.6275801658630371, + 0.03628573194146156, + 0.07713405787944794, + 0.025509724393486977, + 0.04068276286125183, + 0.039329010993242264, + 0.057706303894519806, + 0.01993812620639801, + 0.021339496597647667, + 0.054494649171829224 + ], + "credit_map": [ + 0.03468366712331772, + 0.016144659370183945, + 0.01669432409107685, + 0.015047467313706875, + 0.26157742738723755, + 0.04101485386490822, + 0.06768789887428284, + 0.17472980916500092, + 0.002569830510765314, + 0.0032206557225435972, + 0.005361844785511494, + 0.0036181786563247442, + 0.007139355875551701, + 0.004524604883044958, + 0.0037706303410232067, + 0.00608063256368041, + 0.009081704542040825, + 0.009037179872393608, + 0.006049571558833122, + 0.006328788585960865, + 0.011545549146831036, + 0.01901359111070633, + 0.006486925762146711, + 0.009590751491487026, + 0.003160501830279827, + 0.0027443247381597757, + 0.001756500219926238, + 0.0038944638799875975, + 0.003443206660449505, + 0.0035872231237590313, + 0.0035707526840269566, + 0.0033527531195431948, + 0.004484439734369516, + 0.005053919740021229, + 0.004173752851784229, + 0.0054825544357299805, + 0.0059039699845016, + 0.004548149649053812, + 0.004382861778140068, + 0.006653117015957832, + 0.004206124227494001, + 0.003331009764224291, + 0.004657819401472807, + 0.003911065403372049, + 0.005702704656869173, + 0.007031587418168783, + 0.005119642708450556, + 0.00536905974149704, + 0.0046745529398322105, + 0.006330499425530434, + 0.007814034819602966, + 0.010390120558440685, + 0.00802934356033802, + 0.00683604646474123, + 0.006051335483789444, + 0.007580372039228678, + 0.0013874087017029524, + 0.0021959368605166674, + 0.003390126395970583, + 0.003225408960133791, + 0.002854414051398635, + 0.0020294189453125, + 0.0022286190651357174, + 0.0026267932262271643, + 0.0018782176775857806, + 0.002835221588611603, + 0.00303472438827157, + 0.004526094067841768, + 0.002086237771436572, + 0.0026127570308744907, + 0.0021991345565766096, + 0.00216710870154202, + 0.014042059890925884, + 0.003771038493141532, + 0.0046100555919110775, + 0.00468593230471015, + 0.00462327990680933, + 0.006445656530559063, + 0.006777750328183174, + 0.009538876824080944 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 6 + ], + "selected_reward": -1.9103368520736694 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are two enormous mythical creatures, specifically a cyan dragon and a white tiger, visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the environment a swirling, inky black storm cloud illuminated by intermittent lightning flashes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the video feature frantic montage editing with lightning-fast pans and jump cuts?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the dragon depicted with massive cyan scales that shimmer with iridescence like ancient jade?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the tiger depicted with snow-white fur streaked with black lightning and amber eyes?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.011102682910859585 + ], + "opposite_logprob_tokens": [ + -4.511102676391602 + ], + "logprob_reward": -0.011102682910859585, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.011102682910859585, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3782500922679901, + 0.05677042528986931, + 0.060785602778196335, + 0.04352230951189995, + 0.06785814464092255, + 0.09894823282957077, + 0.06982485204935074, + 0.07045523077249527, + 0.0742817372083664, + 0.07930337637662888 + ], + "credit_map": [ + 0.023503251373767853, + 0.010938719846308231, + 0.014892375096678734, + 0.010949080809950829, + 0.14841322600841522, + 0.024700405076146126, + 0.042053282260894775, + 0.10279974341392517, + 0.007544718682765961, + 0.006661516614258289, + 0.008528177626430988, + 0.006469183601438999, + 0.006506488658487797, + 0.008209402672946453, + 0.004923494067043066, + 0.007927442900836468, + 0.007986380718648434, + 0.009252144955098629, + 0.0050206296145915985, + 0.004974423907697201, + 0.010588517412543297, + 0.008254746906459332, + 0.005695868283510208, + 0.009012890979647636, + 0.005619677249342203, + 0.005056958645582199, + 0.00348346633836627, + 0.008238189853727818, + 0.007836862467229366, + 0.0033478927798569202, + 0.004701996222138405, + 0.00523726548999548, + 0.008712515234947205, + 0.01033359207212925, + 0.008234153501689434, + 0.010175816714763641, + 0.009133479557931423, + 0.006257102359086275, + 0.006822384428232908, + 0.008189107291400433, + 0.010602820664644241, + 0.008048104122281075, + 0.011601105332374573, + 0.00932910107076168, + 0.01423158310353756, + 0.01836300641298294, + 0.012839295901358128, + 0.013933214358985424, + 0.00580515107139945, + 0.0052536544390022755, + 0.007501758635044098, + 0.01380896382033825, + 0.009570975787937641, + 0.009077299386262894, + 0.010779029689729214, + 0.008028018288314342, + 0.004391602240502834, + 0.007334709167480469, + 0.012323656119406223, + 0.011401805095374584, + 0.007826224900782108, + 0.007282606791704893, + 0.007310854736715555, + 0.012583769857883453, + 0.005341040436178446, + 0.013011330738663673, + 0.008832856081426144, + 0.012821144424378872, + 0.01057788822799921, + 0.009043918922543526, + 0.00675384933128953, + 0.00789970625191927, + 0.011021855287253857, + 0.007820786908268929, + 0.005494660697877407, + 0.011870325542986393, + 0.008329752832651138, + 0.006936564110219479, + 0.017826732248067856, + 0.010002701543271542 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 5, + 9 + ], + "selected_reward": -0.011102682910859585 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are two enormous mythical creatures, specifically a cyan dragon and a white tiger, visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the environment a swirling, inky black storm cloud illuminated by intermittent lightning flashes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the video feature frantic montage editing with lightning-fast pans and jump cuts?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the dragon depicted with massive cyan scales that shimmer with iridescence like ancient jade?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the tiger depicted with snow-white fur streaked with black lightning and amber eyes?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.026354091241955757 + ], + "opposite_logprob_tokens": [ + -3.6513540744781494 + ], + "logprob_reward": -0.026354091241955757, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.026354091241955757, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.4824119210243225, + 0.030482178553938866, + 0.04786023497581482, + 0.04744730144739151, + 0.06309446692466736, + 0.09218079596757889, + 0.03344086557626724, + 0.046125009655952454, + 0.042893752455711365, + 0.11406345665454865 + ], + "credit_map": [ + 0.027248423546552658, + 0.012862932868301868, + 0.013089685700833797, + 0.011349528096616268, + 0.200462207198143, + 0.031208772212266922, + 0.052719056606292725, + 0.13347133994102478, + 0.0036655094008892775, + 0.0029092393815517426, + 0.006662460044026375, + 0.003276816802099347, + 0.003512970870360732, + 0.003422298701480031, + 0.002096078358590603, + 0.00493680639192462, + 0.00675937719643116, + 0.006441912148147821, + 0.005398062523454428, + 0.004127299878746271, + 0.011257272213697433, + 0.005135886371135712, + 0.0033030416816473007, + 0.005437382496893406, + 0.006121817976236343, + 0.0069528319872915745, + 0.005013974383473396, + 0.0058442289009690285, + 0.00763459037989378, + 0.0030479610431939363, + 0.006955190096050501, + 0.005876710172742605, + 0.007452236022800207, + 0.009666521102190018, + 0.007119404152035713, + 0.008884363807737827, + 0.009182519279420376, + 0.0058549861423671246, + 0.006794492714107037, + 0.00813994836062193, + 0.012987006455659866, + 0.011109684593975544, + 0.01276356726884842, + 0.011066772043704987, + 0.01209790725260973, + 0.012769311666488647, + 0.009581031277775764, + 0.009805520996451378, + 0.00298682670108974, + 0.0022807724308222532, + 0.0030955220572650433, + 0.0038551981560885906, + 0.004265565890818834, + 0.0052304575219750404, + 0.005817665718495846, + 0.005908861290663481, + 0.003580851713195443, + 0.004493636544793844, + 0.0062059806659817696, + 0.00486211059615016, + 0.0034085093066096306, + 0.006683724466711283, + 0.004831024445593357, + 0.012059178203344345, + 0.0035371652338653803, + 0.006745146121829748, + 0.00407473323866725, + 0.006443415302783251, + 0.00656378036364913, + 0.0054041980765759945, + 0.004881683271378279, + 0.005243633408099413, + 0.012836620211601257, + 0.01009285356849432, + 0.009102662093937397, + 0.018385663628578186, + 0.011214330792427063, + 0.0073322877287864685, + 0.03222282603383064, + 0.012876227498054504 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 5 + ], + "selected_reward": -0.026354091241955757 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "No", + "Yes", + "Yes" + ] + }, + { + "prompt": "frantic montage of lightning-fast pans and jump cuts plunges the viewer directly into the catastrophic battle at the heart of a swirling, inky black storm cloud. The camera switches dizzyingly between the claws and scales of two enormous mythical creatures, creating a sense of ultimate chaos and power. At one moment, the frame is filled with massive cyan dragon scales, each shimmering with iridescence like ancient jade, emitting dazzling sparks in a fierce collision with a ferocious tiger. Then, in an instant, the camera cuts to an extreme close-up, revealing a gleaming, scythe-like white tiger claw, ripping into the dragon's body with ferocious force, its claws digging deep into the cracks between its scales. The entire frenzied scene is illuminated only by intermittent, blinding flashes of lightning, each one freezing the brutal impact in high-contrast black and white silhouettes. We see the Azure Dragon's massive, serpentine form coiling and tightening amidst the howling wind and pounding rain, attempting to strangle its opponent; while the equally massive Tiger, its snow-white fur streaked with black lightning, dodges and strikes back with incredible agility, its amber eyes burning with pure fury. This disorienting, rapid-fire switching between gleaming scales, bloodied fur, snarling maws, and blazing dragon pupils creates a heart-stopping atmosphere of primal, mythical fury.", + "seed": 30380, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000050/group_15/index_4495_seed_30380_dp_30_rank_60_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 15, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.34829972675070164, + "vlm_logprob": -0.34829972675070164, + "vlm_yes_no_margin": 3.5, + "vlm_token_credit": -0.34829972675070164, + "vlm_yes_ratio": 0.8, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are two enormous mythical creatures, specifically a cyan dragon and a white tiger, visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the environment a swirling, inky black storm cloud illuminated by intermittent lightning flashes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the video feature frantic montage editing with lightning-fast pans and jump cuts?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the dragon depicted with massive cyan scales that shimmer with iridescence like ancient jade?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the tiger depicted with snow-white fur streaked with black lightning and amber eyes?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.015135560184717178 + ], + "opposite_logprob_tokens": [ + -4.265135765075684 + ], + "logprob_reward": -0.015135560184717178, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.015135560184717178, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.057360708713531494, + 0.36093610525131226, + 0.04194477200508118, + 0.06289699673652649, + 0.02645699866116047, + 0.029555287212133408, + 0.06377547234296799, + 0.08696296811103821, + 0.1902286261320114, + 0.07988204807043076 + ], + "credit_map": [ + 0.004185810685157776, + 0.007196325343102217, + 0.007647344376891851, + 0.007751911412924528, + 0.006928703282028437, + 0.0073309303261339664, + 0.00759794469922781, + 0.008721740916371346, + 0.02730807475745678, + 0.026474205777049065, + 0.03068355657160282, + 0.020094718784093857, + 0.024644985795021057, + 0.1082569882273674, + 0.02959728054702282, + 0.09387630224227905, + 0.0043693785555660725, + 0.006650236900895834, + 0.005246427841484547, + 0.006020591128617525, + 0.0055655380710959435, + 0.004194352775812149, + 0.004649112932384014, + 0.005249137058854103, + 0.0074345506727695465, + 0.007055722177028656, + 0.007967918179929256, + 0.009444857947528362, + 0.007420806679874659, + 0.006742777768522501, + 0.008623084053397179, + 0.008207280188798904, + 0.0033688044641166925, + 0.003148080548271537, + 0.004373220726847649, + 0.003576823277398944, + 0.002515933010727167, + 0.002515265019610524, + 0.002548886463046074, + 0.004409985151141882, + 0.003601374104619026, + 0.004828816279768944, + 0.0032559309620410204, + 0.002514977240934968, + 0.0034930186811834574, + 0.0043393694795668125, + 0.0038835913874208927, + 0.0036382100079208612, + 0.008587340824306011, + 0.011059942655265331, + 0.01062076911330223, + 0.005266190972179174, + 0.0072573223151266575, + 0.004185325000435114, + 0.005026013124734163, + 0.01177256740629673, + 0.00789212342351675, + 0.019819848239421844, + 0.015584374777972698, + 0.006336466874927282, + 0.006889761425554752, + 0.004427565261721611, + 0.009355492889881134, + 0.016657330095767975, + 0.029237894341349602, + 0.04485585540533066, + 0.011615414172410965, + 0.011963428929448128, + 0.013657464645802975, + 0.016826698556542397, + 0.022033199667930603, + 0.04003867134451866, + 0.00938966404646635, + 0.01252829097211361, + 0.0101771866902709, + 0.0070745875127613544, + 0.00814322754740715, + 0.013227497227489948, + 0.01010924018919468, + 0.009232350625097752 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 8, + 7 + ], + "selected_reward": -0.015135560184717178 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are two enormous mythical creatures, specifically a cyan dragon and a white tiger, visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the environment a swirling, inky black storm cloud illuminated by intermittent lightning flashes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the video feature frantic montage editing with lightning-fast pans and jump cuts?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the dragon depicted with massive cyan scales that shimmer with iridescence like ancient jade?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the tiger depicted with snow-white fur streaked with black lightning and amber eyes?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.008648558519780636 + ], + "opposite_logprob_tokens": [ + -4.75864839553833 + ], + "logprob_reward": -0.008648558519780636, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.008648558519780636, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.03938544541597366, + 0.3429543077945709, + 0.04849492758512497, + 0.09598962962627411, + 0.04418163001537323, + 0.06186414137482643, + 0.04397466033697128, + 0.2311745584011078, + 0.04192951321601868, + 0.05005123093724251 + ], + "credit_map": [ + 0.0032783395145088434, + 0.004915266297757626, + 0.003957982640713453, + 0.004200759343802929, + 0.005649819038808346, + 0.005238268990069628, + 0.005292239598929882, + 0.0068527646362781525, + 0.027267523109912872, + 0.02122047357261181, + 0.02529842033982277, + 0.01709284447133541, + 0.021781593561172485, + 0.1103113517165184, + 0.024845752865076065, + 0.09513632953166962, + 0.0034587786067277193, + 0.005171710159629583, + 0.005420941859483719, + 0.007006901316344738, + 0.006812277249991894, + 0.007843672297894955, + 0.006455694790929556, + 0.006324950605630875, + 0.010342354886233807, + 0.00985423568636179, + 0.013626966625452042, + 0.012906655669212341, + 0.011540601961314678, + 0.012638161890208721, + 0.01397203840315342, + 0.01110861450433731, + 0.0060456180945038795, + 0.0046953316777944565, + 0.005713975988328457, + 0.007729401811957359, + 0.003859213087707758, + 0.004924381617456675, + 0.005607511382550001, + 0.005606191232800484, + 0.008858585730195045, + 0.005497966427356005, + 0.005261974409222603, + 0.007074015215039253, + 0.004440022632479668, + 0.009658063761889935, + 0.011032048612833023, + 0.010041458532214165, + 0.004206742160022259, + 0.005186155904084444, + 0.010056265629827976, + 0.0036246704403311014, + 0.0041849445551633835, + 0.0021254655439406633, + 0.003755046986043453, + 0.010835368186235428, + 0.017383744940161705, + 0.029734691604971886, + 0.08457648009061813, + 0.01091745775192976, + 0.016336165368556976, + 0.017371345311403275, + 0.017025042325258255, + 0.037829622626304626, + 0.007132778409868479, + 0.006339304149150848, + 0.0055801826529204845, + 0.004450883250683546, + 0.004246608819812536, + 0.004280569031834602, + 0.0048476047813892365, + 0.0050515783950686455, + 0.006165897008031607, + 0.008833130821585655, + 0.007080751936882734, + 0.0044723330065608025, + 0.005433744750916958, + 0.006506056059151888, + 0.005929823499172926, + 0.005629491992294788 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 7, + 3 + ], + "selected_reward": -0.008648558519780636 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are two enormous mythical creatures, specifically a cyan dragon and a white tiger, visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the environment a swirling, inky black storm cloud illuminated by intermittent lightning flashes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the video feature frantic montage editing with lightning-fast pans and jump cuts?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the dragon depicted with massive cyan scales that shimmer with iridescence like ancient jade?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the tiger depicted with snow-white fur streaked with black lightning and amber eyes?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -1.701536774635315 + ], + "opposite_logprob_tokens": [ + -0.20153680443763733 + ], + "logprob_reward": -1.701536774635315, + "yes_no_margin_reward": -1.5, + "token_credit_reward": -1.701536774635315, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.02730254828929901, + 0.32376572489738464, + 0.04789780452847481, + 0.10254256427288055, + 0.04672569781541824, + 0.0362144336104393, + 0.08791624009609222, + 0.1818435937166214, + 0.08170205354690552, + 0.06408930569887161 + ], + "credit_map": [ + 0.0024427163880318403, + 0.004777912981808186, + 0.00498846685513854, + 0.003138165920972824, + 0.0035324213095009327, + 0.002600719453766942, + 0.002956474432721734, + 0.002865672344341874, + 0.024499420076608658, + 0.017346587032079697, + 0.02165723219513893, + 0.014016741886734962, + 0.019693348556756973, + 0.10879676043987274, + 0.023039866238832474, + 0.09471576660871506, + 0.0030129877850413322, + 0.006729298736900091, + 0.00548073835670948, + 0.005199148319661617, + 0.006926725152879953, + 0.00709828594699502, + 0.005735759157687426, + 0.007714860141277313, + 0.009623010642826557, + 0.010635539889335632, + 0.014353390783071518, + 0.010328497737646103, + 0.017425542697310448, + 0.014314491301774979, + 0.014649804681539536, + 0.011212286539375782, + 0.004582779947668314, + 0.005228521768003702, + 0.006804305128753185, + 0.006780438125133514, + 0.0038349530659615993, + 0.0049399384297430515, + 0.005310336593538523, + 0.00924442708492279, + 0.006159430369734764, + 0.004245471209287643, + 0.004102044738829136, + 0.004683622159063816, + 0.0033778853248804808, + 0.004400600213557482, + 0.00533702177926898, + 0.003908359445631504, + 0.011151427403092384, + 0.013028530403971672, + 0.018290281295776367, + 0.006903682369738817, + 0.007979607209563255, + 0.005447609815746546, + 0.006773605942726135, + 0.018341494724154472, + 0.012953614816069603, + 0.06344310939311981, + 0.032801225781440735, + 0.010093708522617817, + 0.015238804742693901, + 0.009159567765891552, + 0.01934240013360977, + 0.018811162561178207, + 0.012129285372793674, + 0.011727997101843357, + 0.01756303571164608, + 0.006870817858725786, + 0.00658025499433279, + 0.00882274005562067, + 0.010371493175625801, + 0.007636427413672209, + 0.01074589416384697, + 0.007937227375805378, + 0.007057732902467251, + 0.005246592219918966, + 0.00517671275883913, + 0.011421001516282558, + 0.008654270321130753, + 0.007849876768887043 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 7, + 3 + ], + "selected_reward": -1.701536774635315 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are two enormous mythical creatures, specifically a cyan dragon and a white tiger, visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the environment a swirling, inky black storm cloud illuminated by intermittent lightning flashes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the video feature frantic montage editing with lightning-fast pans and jump cuts?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the dragon depicted with massive cyan scales that shimmer with iridescence like ancient jade?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the tiger depicted with snow-white fur streaked with black lightning and amber eyes?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.003632734064012766 + ], + "opposite_logprob_tokens": [ + -5.628632545471191 + ], + "logprob_reward": -0.003632734064012766, + "yes_no_margin_reward": 5.625, + "token_credit_reward": -0.003632734064012766, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.037049826234579086, + 0.32492417097091675, + 0.08719984441995621, + 0.05936158448457718, + 0.02965121529996395, + 0.04532254859805107, + 0.11654777079820633, + 0.14356383681297302, + 0.08872044086456299, + 0.06765876710414886 + ], + "credit_map": [ + 0.004258438944816589, + 0.00807820726186037, + 0.005984779447317123, + 0.0034052280243486166, + 0.003849145257845521, + 0.003983710426837206, + 0.003303515026345849, + 0.004186800215393305, + 0.029591389000415802, + 0.014666689559817314, + 0.02135464921593666, + 0.015092486515641212, + 0.017133696004748344, + 0.11236276477575302, + 0.016369134187698364, + 0.09835336357355118, + 0.009252862073481083, + 0.01319482084363699, + 0.011328441090881824, + 0.0112745501101017, + 0.011974882334470749, + 0.010427058674395084, + 0.008791180327534676, + 0.010956048965454102, + 0.006705977488309145, + 0.005711518228054047, + 0.008107936009764671, + 0.006053450983017683, + 0.010175444185733795, + 0.007694536354392767, + 0.008667880669236183, + 0.006244842428714037, + 0.0034960205666720867, + 0.0029929864685982466, + 0.004107926040887833, + 0.0044512622989714146, + 0.0025972817093133926, + 0.003113939892500639, + 0.003676745807752013, + 0.005215050652623177, + 0.009385203942656517, + 0.005209591239690781, + 0.004675769712775946, + 0.00436353450641036, + 0.003684758208692074, + 0.00507135596126318, + 0.005257858894765377, + 0.007674475200474262, + 0.009579828940331936, + 0.008248534984886646, + 0.03563626483082771, + 0.009301143698394299, + 0.004511828068643808, + 0.006436898373067379, + 0.00959197897464037, + 0.033241283148527145, + 0.009901385754346848, + 0.05354669317603111, + 0.0230628103017807, + 0.007559554185718298, + 0.011823199689388275, + 0.0062920693308115005, + 0.015367795713245869, + 0.0160103440284729, + 0.012930198572576046, + 0.01751505397260189, + 0.008718827739357948, + 0.006423091050237417, + 0.006142382975667715, + 0.008061413653194904, + 0.012981753796339035, + 0.01594771258533001, + 0.008210728876292706, + 0.010472203604876995, + 0.0066431132145226, + 0.005551231559365988, + 0.008983139880001545, + 0.011059503071010113, + 0.008758132345974445, + 0.007980708964169025 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 7, + 6 + ], + "selected_reward": -0.003632734064012766 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are two enormous mythical creatures, specifically a cyan dragon and a white tiger, visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the environment a swirling, inky black storm cloud illuminated by intermittent lightning flashes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the video feature frantic montage editing with lightning-fast pans and jump cuts?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the dragon depicted with massive cyan scales that shimmer with iridescence like ancient jade?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the tiger depicted with snow-white fur streaked with black lightning and amber eyes?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.012545006349682808 + ], + "opposite_logprob_tokens": [ + -4.387545108795166 + ], + "logprob_reward": -0.012545006349682808, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.012545006349682808, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.057672642171382904, + 0.14756928384304047, + 0.14121635258197784, + 0.16298188269138336, + 0.043057553470134735, + 0.05356134846806526, + 0.08828983455896378, + 0.1659029722213745, + 0.0638970136642456, + 0.07585115730762482 + ], + "credit_map": [ + 0.005457586608827114, + 0.008812406100332737, + 0.007448546122759581, + 0.007616767194122076, + 0.006976548582315445, + 0.005641220137476921, + 0.006067204289138317, + 0.009652361273765564, + 0.013390341773629189, + 0.008820492774248123, + 0.013521533459424973, + 0.011759240180253983, + 0.009492023847997189, + 0.04655034467577934, + 0.009385071694850922, + 0.03465023636817932, + 0.018400512635707855, + 0.021227926015853882, + 0.02410239167511463, + 0.019357912242412567, + 0.016565481200814247, + 0.013858098536729813, + 0.011841455474495888, + 0.015862571075558662, + 0.018551185727119446, + 0.01605619490146637, + 0.018197836354374886, + 0.016083091497421265, + 0.025666363537311554, + 0.024817273020744324, + 0.024066241458058357, + 0.019543679431080818, + 0.00375751918181777, + 0.005815977230668068, + 0.005909026600420475, + 0.005947762168943882, + 0.0044689360074698925, + 0.004613765049725771, + 0.0055961632169783115, + 0.006948399823158979, + 0.012813722714781761, + 0.004912852309644222, + 0.00516857672482729, + 0.005153880920261145, + 0.0029457861091941595, + 0.006014172453433275, + 0.007470914162695408, + 0.009081436321139336, + 0.01253350730985403, + 0.01908073201775551, + 0.01179635338485241, + 0.008152530528604984, + 0.01154365111142397, + 0.006993736606091261, + 0.006876192055642605, + 0.011313129216432571, + 0.011640960350632668, + 0.06296157836914062, + 0.025059709325432777, + 0.008434574119746685, + 0.013620901852846146, + 0.007086505182087421, + 0.017414024099707603, + 0.019684694707393646, + 0.009932867251336575, + 0.007849516347050667, + 0.00946117751300335, + 0.007351459003984928, + 0.005698685068637133, + 0.006914955098181963, + 0.009544584900140762, + 0.007143763825297356, + 0.010758803226053715, + 0.009434190578758717, + 0.011885767802596092, + 0.005868682637810707, + 0.008259562775492668, + 0.011803285218775272, + 0.009372173808515072, + 0.008468684740364552 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 3, + 1 + ], + "selected_reward": -0.012545006349682808 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "No", + "Yes", + "Yes" + ] + }, + { + "prompt": "frantic montage of lightning-fast pans and jump cuts plunges the viewer directly into the catastrophic battle at the heart of a swirling, inky black storm cloud. The camera switches dizzyingly between the claws and scales of two enormous mythical creatures, creating a sense of ultimate chaos and power. At one moment, the frame is filled with massive cyan dragon scales, each shimmering with iridescence like ancient jade, emitting dazzling sparks in a fierce collision with a ferocious tiger. Then, in an instant, the camera cuts to an extreme close-up, revealing a gleaming, scythe-like white tiger claw, ripping into the dragon's body with ferocious force, its claws digging deep into the cracks between its scales. The entire frenzied scene is illuminated only by intermittent, blinding flashes of lightning, each one freezing the brutal impact in high-contrast black and white silhouettes. We see the Azure Dragon's massive, serpentine form coiling and tightening amidst the howling wind and pounding rain, attempting to strangle its opponent; while the equally massive Tiger, its snow-white fur streaked with black lightning, dodges and strikes back with incredible agility, its amber eyes burning with pure fury. This disorienting, rapid-fire switching between gleaming scales, bloodied fur, snarling maws, and blazing dragon pupils creates a heart-stopping atmosphere of primal, mythical fury.", + "seed": 30381, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000050/group_15/index_4495_seed_30381_dp_31_rank_62_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 15, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.2783277120441198, + "vlm_logprob": -0.2783277120441198, + "vlm_yes_no_margin": 3.125, + "vlm_token_credit": -0.2783277120441198, + "vlm_yes_ratio": 0.8, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are two enormous mythical creatures, specifically a cyan dragon and a white tiger, visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the environment a swirling, inky black storm cloud illuminated by intermittent lightning flashes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the video feature frantic montage editing with lightning-fast pans and jump cuts?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the dragon depicted with massive cyan scales that shimmer with iridescence like ancient jade?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the tiger depicted with snow-white fur streaked with black lightning and amber eyes?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.031163109466433525 + ], + "opposite_logprob_tokens": [ + -3.531163215637207 + ], + "logprob_reward": -0.031163109466433525, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.031163109466433525, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1426614373922348, + 0.056492265313863754, + 0.027834469452500343, + 0.060481660068035126, + 0.056164927780628204, + 0.019813939929008484, + 0.07683903723955154, + 0.06988339871168137, + 0.10211324691772461, + 0.387715607881546 + ], + "credit_map": [ + 0.011529254727065563, + 0.024704985320568085, + 0.027667149901390076, + 0.016280798241496086, + 0.012153291143476963, + 0.016447357833385468, + 0.015138952061533928, + 0.018739670515060425, + 0.004146648570895195, + 0.0066923419944942, + 0.006630642805248499, + 0.007355331443250179, + 0.005745848175138235, + 0.010797956958413124, + 0.008577647618949413, + 0.006545850541442633, + 0.0034316638484597206, + 0.004086300730705261, + 0.0031337810214608908, + 0.004113869275897741, + 0.002541903406381607, + 0.003007207764312625, + 0.003775309771299362, + 0.00374443456530571, + 0.004866553004831076, + 0.006562754977494478, + 0.006947973743081093, + 0.007097259629517794, + 0.008107536472380161, + 0.00955995637923479, + 0.00913176964968443, + 0.00820786040276289, + 0.004151621833443642, + 0.0035830200649797916, + 0.007956458255648613, + 0.005394128151237965, + 0.008340971544384956, + 0.008393015712499619, + 0.012406038120388985, + 0.005939678754657507, + 0.0019946089014410973, + 0.0013441573828458786, + 0.0026972205378115177, + 0.001908907201141119, + 0.0028995051980018616, + 0.002170286374166608, + 0.0029628733173012733, + 0.003836381481960416, + 0.009967121295630932, + 0.005831953138113022, + 0.01029315497726202, + 0.006247533019632101, + 0.015967145562171936, + 0.005785316228866577, + 0.014914452098309994, + 0.00783236138522625, + 0.008119385689496994, + 0.0074324537999928, + 0.004829231183975935, + 0.006926219444721937, + 0.010811145417392254, + 0.010723980143666267, + 0.007257357239723206, + 0.013783635571599007, + 0.011954966932535172, + 0.010828537866473198, + 0.01565660908818245, + 0.014436135068535805, + 0.010942577384412289, + 0.009418410249054432, + 0.011162522248923779, + 0.01771349273622036, + 0.04185248166322708, + 0.043850626796483994, + 0.07267461717128754, + 0.056378260254859924, + 0.03744247928261757, + 0.042365364730358124, + 0.03610427305102348, + 0.057047516107559204 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 0, + 8 + ], + "selected_reward": -0.031163109466433525 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are two enormous mythical creatures, specifically a cyan dragon and a white tiger, visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the environment a swirling, inky black storm cloud illuminated by intermittent lightning flashes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the video feature frantic montage editing with lightning-fast pans and jump cuts?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the dragon depicted with massive cyan scales that shimmer with iridescence like ancient jade?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the tiger depicted with snow-white fur streaked with black lightning and amber eyes?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0263565294444561 + ], + "opposite_logprob_tokens": [ + -3.6513564586639404 + ], + "logprob_reward": -0.0263565294444561, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.0263565294444561, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13803787529468536, + 0.1425914317369461, + 0.04931119084358215, + 0.07436403632164001, + 0.1328364759683609, + 0.019895019009709358, + 0.04076305404305458, + 0.03137636557221413, + 0.2875933349132538, + 0.08323129266500473 + ], + "credit_map": [ + 0.010662754997611046, + 0.022081570699810982, + 0.02578141912817955, + 0.017109589651226997, + 0.012027241289615631, + 0.015936601907014847, + 0.015666071325540543, + 0.018772615119814873, + 0.013499231077730656, + 0.01184352021664381, + 0.018412524834275246, + 0.021436819806694984, + 0.014025203883647919, + 0.028468135744333267, + 0.01886894181370735, + 0.01603703200817108, + 0.006209871266037226, + 0.005482349544763565, + 0.006319132633507252, + 0.00831606239080429, + 0.003866499289870262, + 0.005794918164610863, + 0.004996776580810547, + 0.008325578644871712, + 0.007745735812932253, + 0.006523289252072573, + 0.010264693759381771, + 0.006905736401677132, + 0.010583075694739819, + 0.015348331071436405, + 0.008598962798714638, + 0.00839421059936285, + 0.010339473374187946, + 0.006124560255557299, + 0.015331769362092018, + 0.011022856459021568, + 0.016650494188070297, + 0.024201052263379097, + 0.03800526261329651, + 0.011161001399159431, + 0.003103294875472784, + 0.0018966816132888198, + 0.00288908532820642, + 0.0024272094015032053, + 0.002253750804811716, + 0.0025540953502058983, + 0.0025266457814723253, + 0.002244254108518362, + 0.003750327741727233, + 0.0023690375965088606, + 0.00759992515668273, + 0.0035339114256203175, + 0.006550585385411978, + 0.002869990188628435, + 0.008639177307486534, + 0.005450098775327206, + 0.003400311106815934, + 0.0021329328883439302, + 0.002570156706497073, + 0.0029520539101213217, + 0.003776513272896409, + 0.005552390590310097, + 0.004395664669573307, + 0.006596342194825411, + 0.03184721618890762, + 0.027444129809737206, + 0.053811948746442795, + 0.042970702052116394, + 0.030518967658281326, + 0.025302762165665627, + 0.030511684715747833, + 0.045185916125774384, + 0.007833440788090229, + 0.0070091537199914455, + 0.016030702739953995, + 0.01069585606455803, + 0.00956446398049593, + 0.009113408625125885, + 0.010198057629168034, + 0.012786200270056725 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 1, + 0 + ], + "selected_reward": -0.0263565294444561 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are two enormous mythical creatures, specifically a cyan dragon and a white tiger, visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the environment a swirling, inky black storm cloud illuminated by intermittent lightning flashes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the video feature frantic montage editing with lightning-fast pans and jump cuts?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the dragon depicted with massive cyan scales that shimmer with iridescence like ancient jade?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the tiger depicted with snow-white fur streaked with black lightning and amber eyes?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -1.31339430809021 + ], + "opposite_logprob_tokens": [ + -0.3133942782878876 + ], + "logprob_reward": -1.31339430809021, + "yes_no_margin_reward": -1.0, + "token_credit_reward": -1.31339430809021, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.1746959686279297, + 0.03171761706471443, + 0.04455195739865303, + 0.05430179461836815, + 0.22678865492343903, + 0.07390418648719788, + 0.10011961311101913, + 0.184773251414299, + 0.06884768605232239, + 0.04029931500554085 + ], + "credit_map": [ + 0.01507992297410965, + 0.03199164569377899, + 0.03417277708649635, + 0.02035127952694893, + 0.01365739293396473, + 0.021042585372924805, + 0.014667857438325882, + 0.023732520639896393, + 0.0033353068865835667, + 0.003398667089641094, + 0.0036737865302711725, + 0.0052786353044211864, + 0.0033926961477845907, + 0.003137835767120123, + 0.006133032031357288, + 0.003367658006027341, + 0.004193086177110672, + 0.005254671443253756, + 0.005034016445279121, + 0.008574371226131916, + 0.0034222432877868414, + 0.004353272262960672, + 0.008472052402794361, + 0.005248247645795345, + 0.00579713424667716, + 0.004713570699095726, + 0.006193859968334436, + 0.005906519014388323, + 0.006321106106042862, + 0.009891310706734657, + 0.00849818903952837, + 0.006980105768889189, + 0.017651978880167007, + 0.00820687785744667, + 0.01080030295997858, + 0.014029796235263348, + 0.026912398636341095, + 0.054337091743946075, + 0.0775938332080841, + 0.017256371676921844, + 0.010241521522402763, + 0.006659029517322779, + 0.008789240382611752, + 0.005864714737981558, + 0.008578713051974773, + 0.012277315370738506, + 0.007673902902752161, + 0.013819746673107147, + 0.011111337691545486, + 0.007324505131691694, + 0.017715901136398315, + 0.013827995397150517, + 0.017205798998475075, + 0.009100102819502354, + 0.01366489939391613, + 0.010169072076678276, + 0.035181015729904175, + 0.016870755702257156, + 0.011687138117849827, + 0.019992321729660034, + 0.019635513424873352, + 0.02968245930969715, + 0.020793002098798752, + 0.030931023880839348, + 0.00713532418012619, + 0.007087791804224253, + 0.01218441128730774, + 0.008628824725747108, + 0.011712099425494671, + 0.005979700479656458, + 0.006139399018138647, + 0.009980134665966034, + 0.0067245070822536945, + 0.004923831671476364, + 0.0048640575259923935, + 0.005214453674852848, + 0.004208953585475683, + 0.004457124974578619, + 0.0034619965590536594, + 0.0064443922601640224 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 7, + 0 + ], + "selected_reward": -1.31339430809021 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are two enormous mythical creatures, specifically a cyan dragon and a white tiger, visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the environment a swirling, inky black storm cloud illuminated by intermittent lightning flashes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the video feature frantic montage editing with lightning-fast pans and jump cuts?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the dragon depicted with massive cyan scales that shimmer with iridescence like ancient jade?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the tiger depicted with snow-white fur streaked with black lightning and amber eyes?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004651084542274475 + ], + "opposite_logprob_tokens": [ + -5.379651069641113 + ], + "logprob_reward": -0.004651084542274475, + "yes_no_margin_reward": 5.375, + "token_credit_reward": -0.004651084542274475, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.27769652009010315, + 0.041364170610904694, + 0.043033745139837265, + 0.06474519520998001, + 0.1939246952533722, + 0.06342950463294983, + 0.06755034625530243, + 0.06519125401973724, + 0.0855686217546463, + 0.0974959209561348 + ], + "credit_map": [ + 0.022532960399985313, + 0.05256592854857445, + 0.05456371605396271, + 0.032448261976242065, + 0.02234037220478058, + 0.032201141119003296, + 0.023872049525380135, + 0.037172093987464905, + 0.003626768011599779, + 0.00377646298147738, + 0.006276587024331093, + 0.006984181236475706, + 0.003522830316796899, + 0.005048692226409912, + 0.007531111594289541, + 0.004597536753863096, + 0.004681716673076153, + 0.005455585662275553, + 0.005238826386630535, + 0.008689094334840775, + 0.0034738220274448395, + 0.0035580347757786512, + 0.007083970587700605, + 0.004852692596614361, + 0.006056095007807016, + 0.006294201593846083, + 0.00758682144805789, + 0.007969019003212452, + 0.008045465685427189, + 0.00914278719574213, + 0.01046876609325409, + 0.009182036854326725, + 0.013933315873146057, + 0.008309287950396538, + 0.011592359282076359, + 0.013195190578699112, + 0.02415808103978634, + 0.04192464426159859, + 0.06394586712121964, + 0.01686594821512699, + 0.008962998166680336, + 0.004913754295557737, + 0.008125863038003445, + 0.005283852573484182, + 0.009499081410467625, + 0.009113259613513947, + 0.006563324015587568, + 0.010967373847961426, + 0.006560696754604578, + 0.004084999207407236, + 0.008961696177721024, + 0.009457869455218315, + 0.012941421940922737, + 0.006125595886260271, + 0.010863778181374073, + 0.008554286323487759, + 0.010586529038846493, + 0.00599147193133831, + 0.005019090138375759, + 0.008684706874191761, + 0.0074728126637637615, + 0.009409324266016483, + 0.008205515332520008, + 0.009821807965636253, + 0.010659935884177685, + 0.008818499743938446, + 0.01650598645210266, + 0.013010527938604355, + 0.01110144704580307, + 0.00714833801612258, + 0.008932526223361492, + 0.009391352534294128, + 0.010412199422717094, + 0.01205466128885746, + 0.017670471221208572, + 0.011704189702868462, + 0.011447914876043797, + 0.011960581876337528, + 0.010113703086972237, + 0.01213220227509737 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 9 + ], + "selected_reward": -0.004651084542274475 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are two enormous mythical creatures, specifically a cyan dragon and a white tiger, visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the environment a swirling, inky black storm cloud illuminated by intermittent lightning flashes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the video feature frantic montage editing with lightning-fast pans and jump cuts?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the dragon depicted with massive cyan scales that shimmer with iridescence like ancient jade?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the tiger depicted with snow-white fur streaked with black lightning and amber eyes?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.016073528677225113 + ], + "opposite_logprob_tokens": [ + -4.141073703765869 + ], + "logprob_reward": -0.016073528677225113, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.016073528677225113, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1939999908208847, + 0.04063353314995766, + 0.032136473804712296, + 0.05123194307088852, + 0.10316029191017151, + 0.03043275512754917, + 0.0481531135737896, + 0.09686598181724548, + 0.09307452291250229, + 0.3103114664554596 + ], + "credit_map": [ + 0.016388045623898506, + 0.03776015341281891, + 0.03876127675175667, + 0.022937076166272163, + 0.016616158187389374, + 0.020578280091285706, + 0.018328318372368813, + 0.02263064682483673, + 0.004045763052999973, + 0.003387273056432605, + 0.006547689903527498, + 0.006284307222813368, + 0.0036392826586961746, + 0.005667564459145069, + 0.006755162496119738, + 0.004306485410779715, + 0.0036763399839401245, + 0.00452820211648941, + 0.003368444973602891, + 0.006229123566299677, + 0.0027939972933381796, + 0.003337640082463622, + 0.004565556067973375, + 0.0036371664609760046, + 0.003867654362693429, + 0.004780799616128206, + 0.005866251420229673, + 0.0070166196674108505, + 0.006375311873853207, + 0.007923812605440617, + 0.007923879660665989, + 0.007477607112377882, + 0.008029014803469181, + 0.0054008616134524345, + 0.009101806208491325, + 0.007867265492677689, + 0.014908316545188427, + 0.02042064256966114, + 0.028386499732732773, + 0.009045871905982494, + 0.0029362887144088745, + 0.002533497055992484, + 0.004030332900583744, + 0.0024077920243144035, + 0.0044341436587274075, + 0.004355207551270723, + 0.003501720493659377, + 0.006233769468963146, + 0.0058357808738946915, + 0.004276525694876909, + 0.004922778345644474, + 0.005553250666707754, + 0.010586048476397991, + 0.003918046597391367, + 0.008423988707363605, + 0.004636690020561218, + 0.019157934933900833, + 0.0070503754541277885, + 0.007172563578933477, + 0.012237908318638802, + 0.009820730425417423, + 0.014067675918340683, + 0.013733415864408016, + 0.01362535823136568, + 0.011821559630334377, + 0.010569756850600243, + 0.015324732288718224, + 0.012559022754430771, + 0.012975341640412807, + 0.00786465872079134, + 0.011875513941049576, + 0.010083922185003757, + 0.03510652482509613, + 0.03424663841724396, + 0.05880143865942955, + 0.04594298452138901, + 0.030343245714902878, + 0.03490615636110306, + 0.02973911352455616, + 0.04122532531619072 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 0, + 4 + ], + "selected_reward": -0.016073528677225113 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "No", + "Yes", + "Yes" + ] + }, + { + "prompt": "frantic montage of lightning-fast pans and jump cuts plunges the viewer directly into the catastrophic battle at the heart of a swirling, inky black storm cloud. The camera switches dizzyingly between the claws and scales of two enormous mythical creatures, creating a sense of ultimate chaos and power. At one moment, the frame is filled with massive cyan dragon scales, each shimmering with iridescence like ancient jade, emitting dazzling sparks in a fierce collision with a ferocious tiger. Then, in an instant, the camera cuts to an extreme close-up, revealing a gleaming, scythe-like white tiger claw, ripping into the dragon's body with ferocious force, its claws digging deep into the cracks between its scales. The entire frenzied scene is illuminated only by intermittent, blinding flashes of lightning, each one freezing the brutal impact in high-contrast black and white silhouettes. We see the Azure Dragon's massive, serpentine form coiling and tightening amidst the howling wind and pounding rain, attempting to strangle its opponent; while the equally massive Tiger, its snow-white fur streaked with black lightning, dodges and strikes back with incredible agility, its amber eyes burning with pure fury. This disorienting, rapid-fire switching between gleaming scales, bloodied fur, snarling maws, and blazing dragon pupils creates a heart-stopping atmosphere of primal, mythical fury.", + "seed": 30382, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000050/group_15/index_4495_seed_30382_dp_31_rank_62_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 15, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.4316633786074817, + "vlm_logprob": -0.4316633786074817, + "vlm_yes_no_margin": 3.575, + "vlm_token_credit": -0.4316633786074817, + "vlm_yes_ratio": 0.8, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are two enormous mythical creatures, specifically a cyan dragon and a white tiger, visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the environment a swirling, inky black storm cloud illuminated by intermittent lightning flashes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the video feature frantic montage editing with lightning-fast pans and jump cuts?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the dragon depicted with massive cyan scales that shimmer with iridescence like ancient jade?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the tiger depicted with snow-white fur streaked with black lightning and amber eyes?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.008360269479453564 + ], + "opposite_logprob_tokens": [ + -4.883360385894775 + ], + "logprob_reward": -0.008360269479453564, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.008360269479453564, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1880115568637848, + 0.32222792506217957, + 0.11712384223937988, + 0.047377362847328186, + 0.027609845623373985, + 0.06472808867692947, + 0.08371756225824356, + 0.037913545966148376, + 0.022321460768580437, + 0.08896882086992264 + ], + "credit_map": [ + 0.019841672852635384, + 0.024283159524202347, + 0.029911747202277184, + 0.028218545019626617, + 0.01581844687461853, + 0.0239421296864748, + 0.014536228962242603, + 0.03145962208509445, + 0.014446111395955086, + 0.014635257422924042, + 0.01358908973634243, + 0.015698306262493134, + 0.024408195167779922, + 0.10526201874017715, + 0.016962194815278053, + 0.117226742208004, + 0.013174573890864849, + 0.01212884671986103, + 0.019054539501667023, + 0.02014428935945034, + 0.014654451981186867, + 0.01398212555795908, + 0.008934183977544308, + 0.015050826594233513, + 0.005768632050603628, + 0.006572823040187359, + 0.005338996648788452, + 0.005892791785299778, + 0.005856391508132219, + 0.006479579955339432, + 0.005588635802268982, + 0.005879510194063187, + 0.0029779754113405943, + 0.0032873451709747314, + 0.003567730076611042, + 0.0034146110992878675, + 0.0036974402610212564, + 0.003310251282528043, + 0.00302918441593647, + 0.004325307905673981, + 0.004925231449306011, + 0.00807147566229105, + 0.005014869384467602, + 0.010043110698461533, + 0.008143914863467216, + 0.009868847206234932, + 0.0057648527435958385, + 0.01289578340947628, + 0.00602604728192091, + 0.009222669526934624, + 0.009402725845575333, + 0.010652993805706501, + 0.01048225723206997, + 0.014054510742425919, + 0.014430086128413677, + 0.00944626796990633, + 0.004823366180062294, + 0.004408183507621288, + 0.00626745680347085, + 0.004659287165850401, + 0.004700885619968176, + 0.005279931239783764, + 0.0048980750143527985, + 0.0028763606678694487, + 0.0025809998624026775, + 0.0024798442609608173, + 0.0027033404912799597, + 0.0026258036959916353, + 0.003366743912920356, + 0.0034032040275633335, + 0.0022006856743246317, + 0.002960838144645095, + 0.011300934478640556, + 0.011369667947292328, + 0.009105977602303028, + 0.010795696638524532, + 0.01048042718321085, + 0.009740777313709259, + 0.012611265294253826, + 0.013564073480665684 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 2 + ], + "selected_reward": -0.008360269479453564 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are two enormous mythical creatures, specifically a cyan dragon and a white tiger, visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the environment a swirling, inky black storm cloud illuminated by intermittent lightning flashes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the video feature frantic montage editing with lightning-fast pans and jump cuts?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the dragon depicted with massive cyan scales that shimmer with iridescence like ancient jade?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the tiger depicted with snow-white fur streaked with black lightning and amber eyes?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.006743649020791054 + ], + "opposite_logprob_tokens": [ + -5.006743431091309 + ], + "logprob_reward": -0.006743649020791054, + "yes_no_margin_reward": 5.0, + "token_credit_reward": -0.006743649020791054, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11844820529222488, + 0.39950066804885864, + 0.07094556093215942, + 0.07173100113868713, + 0.044660769402980804, + 0.047562360763549805, + 0.06167731061577797, + 0.09079086035490036, + 0.05182215943932533, + 0.04286107420921326 + ], + "credit_map": [ + 0.007543784566223621, + 0.013947194442152977, + 0.017612842842936516, + 0.02013799361884594, + 0.011137320660054684, + 0.022792143747210503, + 0.00844937190413475, + 0.01682755909860134, + 0.03896741196513176, + 0.03287692368030548, + 0.03332791104912758, + 0.038747478276491165, + 0.07642094790935516, + 0.0603555329144001, + 0.03872302547097206, + 0.08008145540952682, + 0.0065870024263858795, + 0.008001710288226604, + 0.007837992161512375, + 0.006842931266874075, + 0.015966162085533142, + 0.008165078237652779, + 0.006601673550903797, + 0.010943006724119186, + 0.00887111946940422, + 0.013482474721968174, + 0.01305386982858181, + 0.007919885218143463, + 0.0074544912204146385, + 0.006135980132967234, + 0.006043447647243738, + 0.008769728243350983, + 0.005475717596709728, + 0.0038695549592375755, + 0.0032527309376746416, + 0.004371311515569687, + 0.004764107521623373, + 0.006441846955567598, + 0.00553118484094739, + 0.010954314842820168, + 0.0032013491727411747, + 0.005232200957834721, + 0.002615234348922968, + 0.004914086312055588, + 0.00444183312356472, + 0.0094855185598135, + 0.004246347583830357, + 0.013425793498754501, + 0.005901165306568146, + 0.006503541022539139, + 0.005052835680544376, + 0.006558175198733807, + 0.00794279482215643, + 0.009883321821689606, + 0.012253081426024437, + 0.007582396734505892, + 0.009235355071723461, + 0.008924119174480438, + 0.009464666247367859, + 0.010502384975552559, + 0.015073700807988644, + 0.00957239419221878, + 0.013282339088618755, + 0.014735904522240162, + 0.004796361085027456, + 0.004741312935948372, + 0.003977567423135042, + 0.004891176708042622, + 0.008851085789501667, + 0.004777004010975361, + 0.004428607411682606, + 0.015359041281044483, + 0.00467904657125473, + 0.005906150210648775, + 0.005584982223808765, + 0.006703732535243034, + 0.004902301821857691, + 0.004800374619662762, + 0.004876000341027975, + 0.0054084863513708115 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 7 + ], + "selected_reward": -0.006743649020791054 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are two enormous mythical creatures, specifically a cyan dragon and a white tiger, visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the environment a swirling, inky black storm cloud illuminated by intermittent lightning flashes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the video feature frantic montage editing with lightning-fast pans and jump cuts?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the dragon depicted with massive cyan scales that shimmer with iridescence like ancient jade?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the tiger depicted with snow-white fur streaked with black lightning and amber eyes?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -2.127030611038208 + ], + "opposite_logprob_tokens": [ + -0.12703052163124084 + ], + "logprob_reward": -2.127030611038208, + "yes_no_margin_reward": -2.0, + "token_credit_reward": -2.127030611038208, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.20011596381664276, + 0.11838727444410324, + 0.13699673116207123, + 0.08429467678070068, + 0.09400609135627747, + 0.10431808233261108, + 0.05427338555455208, + 0.1020524874329567, + 0.05766795575618744, + 0.047887373715639114 + ], + "credit_map": [ + 0.01989256963133812, + 0.029549311846494675, + 0.03433072939515114, + 0.04616742581129074, + 0.013450270518660545, + 0.017983965575695038, + 0.011693054810166359, + 0.027048643678426743, + 0.006859040819108486, + 0.007523460779339075, + 0.010496134869754314, + 0.008548679761588573, + 0.011557440273463726, + 0.02937968820333481, + 0.009675418958067894, + 0.03434741124510765, + 0.016002949327230453, + 0.013205403462052345, + 0.020978722721338272, + 0.0167236328125, + 0.017290327697992325, + 0.01938341185450554, + 0.015967359766364098, + 0.0174449160695076, + 0.011337862350046635, + 0.01149693038314581, + 0.010593822225928307, + 0.012200198136270046, + 0.009323809295892715, + 0.013554026372730732, + 0.00791207142174244, + 0.007875948213040829, + 0.007264275103807449, + 0.010234189219772816, + 0.009779051877558231, + 0.012426045723259449, + 0.00955221988260746, + 0.01141495630145073, + 0.011579175479710102, + 0.02175617404282093, + 0.009603426791727543, + 0.011259989812970161, + 0.006467402447015047, + 0.007592031266540289, + 0.008736330084502697, + 0.020349379628896713, + 0.009922940284013748, + 0.030386587604880333, + 0.005728205665946007, + 0.006116730626672506, + 0.004921285435557365, + 0.0057569085620343685, + 0.008392981253564358, + 0.009381267242133617, + 0.007188934367150068, + 0.0067870719358325005, + 0.014062827453017235, + 0.00833561085164547, + 0.016571398824453354, + 0.015341746620833874, + 0.011367866769433022, + 0.01189583819359541, + 0.012883851304650307, + 0.011593345552682877, + 0.00829091016203165, + 0.006195712368935347, + 0.0070506976917386055, + 0.008627042174339294, + 0.009540395811200142, + 0.006500380579382181, + 0.005606871098279953, + 0.005855940748006105, + 0.00748912850394845, + 0.005362904630601406, + 0.00519532710313797, + 0.005851960740983486, + 0.004142835736274719, + 0.005620916374027729, + 0.006339516956359148, + 0.007884783670306206 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 1 + ], + "selected_reward": -2.127030611038208 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are two enormous mythical creatures, specifically a cyan dragon and a white tiger, visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the environment a swirling, inky black storm cloud illuminated by intermittent lightning flashes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the video feature frantic montage editing with lightning-fast pans and jump cuts?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the dragon depicted with massive cyan scales that shimmer with iridescence like ancient jade?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the tiger depicted with snow-white fur streaked with black lightning and amber eyes?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0036362973041832447 + ], + "opposite_logprob_tokens": [ + -5.628636360168457 + ], + "logprob_reward": -0.0036362973041832447, + "yes_no_margin_reward": 5.625, + "token_credit_reward": -0.0036362973041832447, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.21384835243225098, + 0.22524817287921906, + 0.09736472368240356, + 0.0768442451953888, + 0.058649711310863495, + 0.07035510987043381, + 0.03510658070445061, + 0.12410033494234085, + 0.04351251572370529, + 0.05497024580836296 + ], + "credit_map": [ + 0.02054065465927124, + 0.03099868632853031, + 0.03440466523170471, + 0.05309942364692688, + 0.014599601738154888, + 0.01754334568977356, + 0.012173634953796864, + 0.03048834763467312, + 0.009831669740378857, + 0.009470882825553417, + 0.012190171517431736, + 0.010151483118534088, + 0.014106200076639652, + 0.07503272593021393, + 0.010527421720325947, + 0.0839376226067543, + 0.010936637409031391, + 0.011899062432348728, + 0.01228424347937107, + 0.011678013950586319, + 0.014260348863899708, + 0.012135189957916737, + 0.010170642286539078, + 0.014000598341226578, + 0.00903601199388504, + 0.011943499557673931, + 0.010785898193717003, + 0.010458766482770443, + 0.00894489698112011, + 0.010554253123700619, + 0.007994066923856735, + 0.007126860320568085, + 0.004785899538546801, + 0.005889921449124813, + 0.005713015794754028, + 0.007734590210020542, + 0.005053867120295763, + 0.007252437528222799, + 0.007489186245948076, + 0.014730804599821568, + 0.0061127180233597755, + 0.007939494214951992, + 0.00422290712594986, + 0.0058425189927220345, + 0.006464007776230574, + 0.013260236009955406, + 0.006558575201779604, + 0.019954657182097435, + 0.0036050297785550356, + 0.0036246476229280233, + 0.004428590182214975, + 0.00377184827812016, + 0.004960366990417242, + 0.006060267332941294, + 0.004111821297556162, + 0.004544013645499945, + 0.01570267044007778, + 0.010555931366980076, + 0.018270060420036316, + 0.017650848254561424, + 0.016423365101218224, + 0.013289734721183777, + 0.016533026471734047, + 0.015674717724323273, + 0.006348053924739361, + 0.005010075867176056, + 0.0052237072959542274, + 0.0065758139826357365, + 0.0059791323728859425, + 0.005344371311366558, + 0.004507528617978096, + 0.004523837007582188, + 0.006594720296561718, + 0.0078958785161376, + 0.008846224285662174, + 0.007842245511710644, + 0.005330925807356834, + 0.005645199213176966, + 0.005730185192078352, + 0.007084871642291546 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 7 + ], + "selected_reward": -0.0036362973041832447 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are two enormous mythical creatures, specifically a cyan dragon and a white tiger, visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the environment a swirling, inky black storm cloud illuminated by intermittent lightning flashes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the video feature frantic montage editing with lightning-fast pans and jump cuts?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the dragon depicted with massive cyan scales that shimmer with iridescence like ancient jade?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the tiger depicted with snow-white fur streaked with black lightning and amber eyes?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01254606619477272 + ], + "opposite_logprob_tokens": [ + -4.387546062469482 + ], + "logprob_reward": -0.01254606619477272, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.01254606619477272, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15059110522270203, + 0.21185657382011414, + 0.0722595676779747, + 0.07482141256332397, + 0.07671887427568436, + 0.08318088203668594, + 0.03883576765656471, + 0.16442303359508514, + 0.04056628420948982, + 0.086746446788311 + ], + "credit_map": [ + 0.013913251459598541, + 0.021582338958978653, + 0.02486012689769268, + 0.0368402898311615, + 0.01078448910266161, + 0.012502922676503658, + 0.008803123608231544, + 0.021304570138454437, + 0.008845029398798943, + 0.00846037082374096, + 0.011768289841711521, + 0.008932498283684254, + 0.0091081028804183, + 0.07518164813518524, + 0.008829441852867603, + 0.08073119074106216, + 0.008567809127271175, + 0.006395979784429073, + 0.011526434682309628, + 0.009589931927621365, + 0.0071018170565366745, + 0.011133117601275444, + 0.00979771837592125, + 0.008146765641868114, + 0.00890408456325531, + 0.012766468338668346, + 0.010290474630892277, + 0.013643941842019558, + 0.006749812047928572, + 0.00940634123980999, + 0.006791333202272654, + 0.006268960889428854, + 0.006315369624644518, + 0.007560045924037695, + 0.007510846946388483, + 0.010090561583638191, + 0.006232813000679016, + 0.009697682224214077, + 0.009944272227585316, + 0.0193672813475132, + 0.00738754915073514, + 0.008567649871110916, + 0.005140990950167179, + 0.00886939000338316, + 0.0115255918353796, + 0.014596624299883842, + 0.00785872433334589, + 0.019234366714954376, + 0.003343280404806137, + 0.006274609826505184, + 0.004184046760201454, + 0.005046388134360313, + 0.005020583048462868, + 0.005544030107557774, + 0.004694733303040266, + 0.00472809886559844, + 0.01812746189534664, + 0.014301037415862083, + 0.01849309355020523, + 0.020089896395802498, + 0.022611714899539948, + 0.020548218861222267, + 0.02502591349184513, + 0.025225702673196793, + 0.005891413893550634, + 0.0044772871769964695, + 0.004617130383849144, + 0.005700221285223961, + 0.006231905892491341, + 0.005017025396227837, + 0.004263773560523987, + 0.004367530345916748, + 0.009617993608117104, + 0.012651079334318638, + 0.012138538993895054, + 0.0096336230635643, + 0.011055737733840942, + 0.009187793359160423, + 0.010791270062327385, + 0.011670413427054882 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 7, + 0 + ], + "selected_reward": -0.01254606619477272 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "No", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000050/group_2/index_tensor([3304, 3304])_group_2_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000050/group_2/index_tensor([3304, 3304])_group_2_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..c7504a675780b8376733116be5250da29db5379d --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000050/group_2/index_tensor([3304, 3304])_group_2_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "In a breathtaking display of high-speed photography aesthetics, a dynamic low-angle tracking shot begins, the camera ascending smoothly in ultra-slow motion, following two majestic English setters. They are captured en route by a vast expanse of tall, golden grass, which sways gently around them. The two dogs, one slightly ahead, are rendered in sharp focus; their long, silky, feathered coats—bright white speckled with liver-brown spots—ripple and flow with an almost liquid grace. The lead dog arches its back, leaping powerfully, its mouth open to reveal a swell of joy, while the second dog presses forward, powerful muscles visible beneath its flowing fur. The entire scene is bathed in the warm, low-angle light of golden hour, capturing every hair in their coats and creating a brilliant rim-lit effect. As their paws stir the verdant undergrowth, wisps of morning dew are kicked into the air, each drop suspended in mid-air like a shimmering shower of diamonds, frozen in time and refracting the sunlight into tiny, brilliant prisms. The background features a lush green forest, creating a beautiful, soft-focus bokeh that isolates the dogs and emphasizes their dynamic movements, creating a stunning atmosphere of pure energy and natural elegance.", + "seed": 27065, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000050/group_2/index_3304_seed_27065_dp_4_rank_8_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 2, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.027907551173120736, + "vlm_logprob": -0.027907551173120736, + "vlm_yes_no_margin": 4.5, + "vlm_token_credit": -0.027907551173120736, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are two English setters visible in the video?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera positioned at a low angle and moving upward (ascending) while tracking the dogs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the dogs running through a vast expanse of tall, golden grass?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the lead dog arching its back to leap with its mouth open, while the second dog presses forward with visible muscles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene bathed in warm golden hour light with a lush green forest background featuring soft-focus bokeh?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.019472390413284302 + ], + "opposite_logprob_tokens": [ + -4.269472599029541 + ], + "logprob_reward": -0.019472390413284302, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.019472390413284302, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3416122496128082, + 0.16718293726444244, + 0.07162933796644211, + 0.05675329640507698, + 0.0220039002597332, + 0.09210753440856934, + 0.07185076922178268, + 0.030837010592222214, + 0.104251928627491, + 0.041771020740270615 + ], + "credit_map": [ + 0.012956826016306877, + 0.009764222428202629, + 0.07055526226758957, + 0.04412064328789711, + 0.013418986462056637, + 0.01207634899765253, + 0.14256881177425385, + 0.03615115210413933, + 0.009190915152430534, + 0.009569558314979076, + 0.018885228782892227, + 0.010194207541644573, + 0.010198794305324554, + 0.03044167160987854, + 0.024278922006487846, + 0.05442364886403084, + 0.005501375067979097, + 0.010626358911395073, + 0.01082657277584076, + 0.010132708586752415, + 0.008903025649487972, + 0.007808573078364134, + 0.007971680723130703, + 0.009859041310846806, + 0.005817726254463196, + 0.007222101092338562, + 0.00852598249912262, + 0.0069768610410392284, + 0.0073221605271101, + 0.006512735038995743, + 0.006902450695633888, + 0.007473277393728495, + 0.002237270586192608, + 0.002098572440445423, + 0.003304657293483615, + 0.0029578739777207375, + 0.0028908841777592897, + 0.003019876778125763, + 0.002946709282696247, + 0.0025480554904788733, + 0.010158965364098549, + 0.009946618229150772, + 0.007936756126582623, + 0.010766600258648396, + 0.01640360616147518, + 0.011505529284477234, + 0.015406477265059948, + 0.009982985444366932, + 0.00893158745020628, + 0.01344495452940464, + 0.006679674610495567, + 0.009110820479691029, + 0.006576525513082743, + 0.00914719793945551, + 0.011028626002371311, + 0.006931384559720755, + 0.002818885026499629, + 0.002934943651780486, + 0.003783335443586111, + 0.0029835612513124943, + 0.0052421241998672485, + 0.003740705316886306, + 0.006318566855043173, + 0.0030148879159241915, + 0.013187742792069912, + 0.011611105874180794, + 0.015742162242531776, + 0.009506060741841793, + 0.01333415973931551, + 0.01571948081254959, + 0.012882429175078869, + 0.012268790043890476, + 0.006851054262369871, + 0.006644264329224825, + 0.004977230913937092, + 0.009143372066318989, + 0.004225222393870354, + 0.0023722206242382526, + 0.0031887302175164223, + 0.004368928261101246 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 8 + ], + "selected_reward": -0.019472390413284302 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are two English setters visible in the video?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera positioned at a low angle and moving upward (ascending) while tracking the dogs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the dogs running through a vast expanse of tall, golden grass?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the lead dog arching its back to leap with its mouth open, while the second dog presses forward with visible muscles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene bathed in warm golden hour light with a lush green forest background featuring soft-focus bokeh?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.10031840950250626 + ], + "opposite_logprob_tokens": [ + -2.350318431854248 + ], + "logprob_reward": -0.10031840950250626, + "yes_no_margin_reward": 2.25, + "token_credit_reward": -0.10031840950250626, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.4856666028499603, + 0.1248544231057167, + 0.04247082397341728, + 0.04700988531112671, + 0.0722617655992508, + 0.05027743801474571, + 0.0632912665605545, + 0.04037657752633095, + 0.03424391523003578, + 0.03954729437828064 + ], + "credit_map": [ + 0.01874869130551815, + 0.015451975166797638, + 0.09802114963531494, + 0.0683404877781868, + 0.020592037588357925, + 0.018492231145501137, + 0.20090438425540924, + 0.04511560499668121, + 0.005943882744759321, + 0.012160955928266048, + 0.013574568554759026, + 0.01287799421697855, + 0.009620596654713154, + 0.018464235588908195, + 0.016758335754275322, + 0.03545383736491203, + 0.003943337593227625, + 0.006059208419173956, + 0.0061409794725477695, + 0.005664795637130737, + 0.006357293110340834, + 0.004972337745130062, + 0.004626908805221319, + 0.004705960396677256, + 0.0054843490943312645, + 0.004203815944492817, + 0.007008486893028021, + 0.005565695930272341, + 0.006659006699919701, + 0.006716631818562746, + 0.006135369185358286, + 0.005236529279500246, + 0.007825552485883236, + 0.00668034702539444, + 0.00822602491825819, + 0.01261474285274744, + 0.01029981579631567, + 0.007313732523471117, + 0.012897779233753681, + 0.006403767503798008, + 0.0052966163493692875, + 0.006763354409486055, + 0.009409942664206028, + 0.005085899494588375, + 0.005486358422785997, + 0.0060402583330869675, + 0.005229965317994356, + 0.006965041160583496, + 0.009385955519974232, + 0.010557858273386955, + 0.0044569410383701324, + 0.007410561200231314, + 0.006506799254566431, + 0.008742141537368298, + 0.009748920798301697, + 0.0064820824190974236, + 0.004357316065579653, + 0.003054990665987134, + 0.0039105527102947235, + 0.004507885314524174, + 0.009222082793712616, + 0.004978339187800884, + 0.005119189154356718, + 0.005226219538599253, + 0.0046321190893650055, + 0.0033136862330138683, + 0.0042879777029156685, + 0.004541337490081787, + 0.004525002557784319, + 0.004153029061853886, + 0.0043655335903167725, + 0.0044252267107367516, + 0.004582925233989954, + 0.005679840687662363, + 0.006198773626238108, + 0.006843494717031717, + 0.003383018309250474, + 0.0036359280347824097, + 0.003921170718967915, + 0.005302142351865768 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 4 + ], + "selected_reward": -0.10031840950250626 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are two English setters visible in the video?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera positioned at a low angle and moving upward (ascending) while tracking the dogs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the dogs running through a vast expanse of tall, golden grass?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the lead dog arching its back to leap with its mouth open, while the second dog presses forward with visible muscles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene bathed in warm golden hour light with a lush green forest background featuring soft-focus bokeh?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.009776685386896133 + ], + "opposite_logprob_tokens": [ + -4.634776592254639 + ], + "logprob_reward": -0.009776685386896133, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.009776685386896133, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12761537730693817, + 0.6318636536598206, + 0.03235483914613724, + 0.03184014931321144, + 0.021120121702551842, + 0.035172153264284134, + 0.06338482350111008, + 0.020583758130669594, + 0.023516224697232246, + 0.01254899613559246 + ], + "credit_map": [ + 0.00521064642816782, + 0.00428236136212945, + 0.02796364389359951, + 0.017257079482078552, + 0.005172906443476677, + 0.004371670074760914, + 0.05305834859609604, + 0.010298721492290497, + 0.03080638311803341, + 0.016321426257491112, + 0.07040107995271683, + 0.027400357648730278, + 0.03318255767226219, + 0.1246158629655838, + 0.09537170827388763, + 0.23376424610614777, + 0.002954490017145872, + 0.002387495245784521, + 0.009411654435098171, + 0.003505044849589467, + 0.0024152379482984543, + 0.0029587079770863056, + 0.0038471240550279617, + 0.004875082988291979, + 0.0037068419624119997, + 0.002168578328564763, + 0.007774279452860355, + 0.003021497279405594, + 0.002914148848503828, + 0.0046648927964270115, + 0.0030751717276871204, + 0.004514740314334631, + 0.002505206735804677, + 0.0019760963041335344, + 0.0037697528023272753, + 0.003455940168350935, + 0.0018248637206852436, + 0.002992704277858138, + 0.002787515986710787, + 0.0018080408917739987, + 0.00600072368979454, + 0.0022761868312954903, + 0.00582294212654233, + 0.003209500340744853, + 0.002645001746714115, + 0.004884450696408749, + 0.0029248360078781843, + 0.007408510893583298, + 0.008957916870713234, + 0.0065318867564201355, + 0.009278960525989532, + 0.0044739567674696445, + 0.005623956676572561, + 0.009393787942826748, + 0.009227178990840912, + 0.00989717710763216, + 0.004558206535875797, + 0.0012688509887084365, + 0.0043024891056120396, + 0.0017497926019132137, + 0.0017900853417813778, + 0.003535423893481493, + 0.0015628327382728457, + 0.0018160773906856775, + 0.0049914587289094925, + 0.0032478959765285254, + 0.0038095915224403143, + 0.0044557456858456135, + 0.0015539111336693168, + 0.0016769022913649678, + 0.0018411073833703995, + 0.0019396106945350766, + 0.002449839375913143, + 0.002234173007309437, + 0.0015084187034517527, + 0.0031655156053602695, + 0.0006985601503401995, + 0.0007073048618622124, + 0.0007448747055605054, + 0.0010403100168332458 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 6 + ], + "selected_reward": -0.009776685386896133 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are two English setters visible in the video?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera positioned at a low angle and moving upward (ascending) while tracking the dogs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the dogs running through a vast expanse of tall, golden grass?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the lead dog arching its back to leap with its mouth open, while the second dog presses forward with visible muscles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene bathed in warm golden hour light with a lush green forest background featuring soft-focus bokeh?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.008637921884655952 + ], + "opposite_logprob_tokens": [ + -4.75863790512085 + ], + "logprob_reward": -0.008637921884655952, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.008637921884655952, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.26617029309272766, + 0.38715943694114685, + 0.027122588828206062, + 0.048318538814783096, + 0.035098847001791, + 0.026087775826454163, + 0.059944842010736465, + 0.06473305821418762, + 0.03950151056051254, + 0.04586314037442207 + ], + "credit_map": [ + 0.012439815327525139, + 0.019527291879057884, + 0.052832476794719696, + 0.02862578257918358, + 0.024922948330640793, + 0.017996355891227722, + 0.08732162415981293, + 0.022503947839140892, + 0.01817513257265091, + 0.011931720189750195, + 0.042506128549575806, + 0.024494310840964317, + 0.020484022796154022, + 0.07192746549844742, + 0.0609881617128849, + 0.13665246963500977, + 0.002286097500473261, + 0.0029263272881507874, + 0.00545587670058012, + 0.004126843065023422, + 0.0027772856410592794, + 0.002392664086073637, + 0.003622455522418022, + 0.003535037161782384, + 0.0060919346287846565, + 0.005512439180165529, + 0.004484587348997593, + 0.006347776390612125, + 0.004785533994436264, + 0.009111604653298855, + 0.004462833050638437, + 0.007521822117269039, + 0.003741177497431636, + 0.0032907116692513227, + 0.00438027735799551, + 0.005814261734485626, + 0.00409591244533658, + 0.0040959930047392845, + 0.006650432478636503, + 0.0030300747603178024, + 0.004034190904349089, + 0.0027343800757080317, + 0.003005819860845804, + 0.0024197893217206, + 0.0031748968176543713, + 0.003445012727752328, + 0.0032418102491647005, + 0.004031873308122158, + 0.008186672814190388, + 0.008762354031205177, + 0.008311264216899872, + 0.005918098613619804, + 0.005640325136482716, + 0.007910268381237984, + 0.007150325924158096, + 0.008065525442361832, + 0.007532496005296707, + 0.0048864674754440784, + 0.006130258087068796, + 0.005753082223236561, + 0.01673908345401287, + 0.008979551494121552, + 0.008243612945079803, + 0.006468502804636955, + 0.005567504558712244, + 0.003438214771449566, + 0.005840274505317211, + 0.004059060011059046, + 0.004505774937570095, + 0.005108690820634365, + 0.005314914043992758, + 0.005667070392519236, + 0.0057090651243925095, + 0.006545716896653175, + 0.0045281825587153435, + 0.006915593054145575, + 0.006160009652376175, + 0.0038488467689603567, + 0.004468160681426525, + 0.007687559816986322 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 7 + ], + "selected_reward": -0.008637921884655952 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are two English setters visible in the video?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera positioned at a low angle and moving upward (ascending) while tracking the dogs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the dogs running through a vast expanse of tall, golden grass?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the lead dog arching its back to leap with its mouth open, while the second dog presses forward with visible muscles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene bathed in warm golden hour light with a lush green forest background featuring soft-focus bokeh?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0013323486782610416 + ], + "opposite_logprob_tokens": [ + -6.6263322830200195 + ], + "logprob_reward": -0.0013323486782610416, + "yes_no_margin_reward": 6.625, + "token_credit_reward": -0.0013323486782610416, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1374472677707672, + 0.3853561580181122, + 0.037681944668293, + 0.24162422120571136, + 0.018972689285874367, + 0.018455982208251953, + 0.04250631108880043, + 0.024457663297653198, + 0.03617323935031891, + 0.05732451006770134 + ], + "credit_map": [ + 0.006358944345265627, + 0.00784101989120245, + 0.023716524243354797, + 0.014909639023244381, + 0.01148009393364191, + 0.008042345754802227, + 0.047127608209848404, + 0.017971104010939598, + 0.018706947565078735, + 0.012489198707044125, + 0.04366593807935715, + 0.017667975276708603, + 0.020490199327468872, + 0.07419941574335098, + 0.057168446481227875, + 0.14096805453300476, + 0.00331958569586277, + 0.0030520674772560596, + 0.011059572920203209, + 0.004307452589273453, + 0.002886921865865588, + 0.0033824238926172256, + 0.004608623683452606, + 0.005065297242254019, + 0.026223011314868927, + 0.022820496931672096, + 0.049142781645059586, + 0.024425657466053963, + 0.025101805105805397, + 0.035452816635370255, + 0.028569519519805908, + 0.02988813817501068, + 0.001999937230721116, + 0.002681759186089039, + 0.0029030502773821354, + 0.0026399975176900625, + 0.002031691139563918, + 0.0024806917645037174, + 0.0024814491625875235, + 0.0017541133565828204, + 0.0029496264178305864, + 0.001995259430259466, + 0.0028865886852145195, + 0.0017133100191131234, + 0.0016909452388063073, + 0.0020764556247740984, + 0.0016246176091954112, + 0.003519179532304406, + 0.009735549800097942, + 0.0033114932011812925, + 0.005237456876784563, + 0.002670994959771633, + 0.004162841010838747, + 0.006517802365124226, + 0.0036788058932870626, + 0.007191367447376251, + 0.005190473049879074, + 0.0017413283931091428, + 0.0039597610011696815, + 0.0017561267595738173, + 0.0024836575612425804, + 0.003534238552674651, + 0.0026835771277546883, + 0.0031084981746971607, + 0.0063018216751515865, + 0.0050622993148863316, + 0.004978483077138662, + 0.008897214196622372, + 0.0022300032433122396, + 0.002630531322211027, + 0.002573701087385416, + 0.0034991854336112738, + 0.010356251150369644, + 0.011230012401938438, + 0.012914658524096012, + 0.010556498542428017, + 0.0026715677231550217, + 0.002530757337808609, + 0.0035984832793474197, + 0.0034662752877920866 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 3, + 0 + ], + "selected_reward": -0.0013323486782610416 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "In a breathtaking display of high-speed photography aesthetics, a dynamic low-angle tracking shot begins, the camera ascending smoothly in ultra-slow motion, following two majestic English setters. They are captured en route by a vast expanse of tall, golden grass, which sways gently around them. The two dogs, one slightly ahead, are rendered in sharp focus; their long, silky, feathered coats—bright white speckled with liver-brown spots—ripple and flow with an almost liquid grace. The lead dog arches its back, leaping powerfully, its mouth open to reveal a swell of joy, while the second dog presses forward, powerful muscles visible beneath its flowing fur. The entire scene is bathed in the warm, low-angle light of golden hour, capturing every hair in their coats and creating a brilliant rim-lit effect. As their paws stir the verdant undergrowth, wisps of morning dew are kicked into the air, each drop suspended in mid-air like a shimmering shower of diamonds, frozen in time and refracting the sunlight into tiny, brilliant prisms. The background features a lush green forest, creating a beautiful, soft-focus bokeh that isolates the dogs and emphasizes their dynamic movements, creating a stunning atmosphere of pure energy and natural elegance.", + "seed": 27066, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000050/group_2/index_3304_seed_27066_dp_4_rank_8_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 2, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.06882593077607453, + "vlm_logprob": -0.06882593077607453, + "vlm_yes_no_margin": 3.9, + "vlm_token_credit": -0.06882593077607453, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are two English setters visible in the video?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera positioned at a low angle and moving upward (ascending) while tracking the dogs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the dogs running through a vast expanse of tall, golden grass?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the lead dog arching its back to leap with its mouth open, while the second dog presses forward with visible muscles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene bathed in warm golden hour light with a lush green forest background featuring soft-focus bokeh?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.028047610074281693 + ], + "opposite_logprob_tokens": [ + -4.028047561645508 + ], + "logprob_reward": -0.028047610074281693, + "yes_no_margin_reward": 4.0, + "token_credit_reward": -0.028047610074281693, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15045443177223206, + 0.03856509178876877, + 0.13998432457447052, + 0.03161592036485672, + 0.2358267903327942, + 0.09559467434883118, + 0.05603843927383423, + 0.09071367233991623, + 0.060896407812833786, + 0.10031021386384964 + ], + "credit_map": [ + 0.010222242213785648, + 0.011209933087229729, + 0.011981232091784477, + 0.01332301925867796, + 0.009271856397390366, + 0.01198294572532177, + 0.04041847586631775, + 0.042044736444950104, + 0.004727236460894346, + 0.004370034672319889, + 0.004037305247038603, + 0.005217346828430891, + 0.004845461342483759, + 0.004907453898340464, + 0.005071267019957304, + 0.005388988181948662, + 0.006989581044763327, + 0.005559665150940418, + 0.017699768766760826, + 0.009993838146328926, + 0.010067429393529892, + 0.01305081881582737, + 0.03341735899448395, + 0.0432058721780777, + 0.004278190899640322, + 0.0023247457575052977, + 0.002227961318567395, + 0.0039346045814454556, + 0.00393739202991128, + 0.0031563821248710155, + 0.005131731741130352, + 0.006624914240092039, + 0.03157396614551544, + 0.015862038359045982, + 0.04908496513962746, + 0.02436540089547634, + 0.018396912142634392, + 0.028911206871271133, + 0.03499416261911392, + 0.03263816609978676, + 0.00839282851666212, + 0.008891015313565731, + 0.008238013833761215, + 0.010050658136606216, + 0.023386642336845398, + 0.012750234454870224, + 0.013309594243764877, + 0.010575691238045692, + 0.005376287270337343, + 0.005146737676113844, + 0.009010056033730507, + 0.006685537751764059, + 0.005608055740594864, + 0.008057381957769394, + 0.006951142102479935, + 0.009203244931995869, + 0.01156658586114645, + 0.010963977314531803, + 0.009958411566913128, + 0.013923434540629387, + 0.01155878882855177, + 0.010257112793624401, + 0.01228246372193098, + 0.010202907025814056, + 0.005612135864794254, + 0.0056242747232317924, + 0.008494526147842407, + 0.011533710174262524, + 0.005064613185822964, + 0.006239340640604496, + 0.010883642360568047, + 0.007444171234965324, + 0.00822949968278408, + 0.015335029922425747, + 0.007349674124270678, + 0.01817341148853302, + 0.00992108229547739, + 0.022294098511338234, + 0.011315764859318733, + 0.007691660430282354 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 0, + 2 + ], + "selected_reward": -0.028047610074281693 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are two English setters visible in the video?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera positioned at a low angle and moving upward (ascending) while tracking the dogs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the dogs running through a vast expanse of tall, golden grass?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the lead dog arching its back to leap with its mouth open, while the second dog presses forward with visible muscles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene bathed in warm golden hour light with a lush green forest background featuring soft-focus bokeh?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.2813486158847809 + ], + "opposite_logprob_tokens": [ + -1.4063485860824585 + ], + "logprob_reward": -0.2813486158847809, + "yes_no_margin_reward": 1.125, + "token_credit_reward": -0.2813486158847809, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14225342869758606, + 0.09457831084728241, + 0.04239600896835327, + 0.04836475849151611, + 0.10149448364973068, + 0.06114914268255234, + 0.07604434341192245, + 0.27552151679992676, + 0.09651417285203934, + 0.06168384850025177 + ], + "credit_map": [ + 0.007734306156635284, + 0.0077697280794382095, + 0.015835698693990707, + 0.013848960399627686, + 0.010641077533364296, + 0.01586475782096386, + 0.03358572721481323, + 0.03697318211197853, + 0.012292701750993729, + 0.015449770726263523, + 0.0112015251070261, + 0.014006122946739197, + 0.011118724010884762, + 0.008923368528485298, + 0.011523136869072914, + 0.010062959976494312, + 0.0026947257574647665, + 0.002789770718663931, + 0.0030798541847616434, + 0.004134789574891329, + 0.0034957046154886484, + 0.00303152808919549, + 0.011603133752942085, + 0.011566503904759884, + 0.005146006587892771, + 0.004938854835927486, + 0.00522336270660162, + 0.008826164528727531, + 0.0050707790069282055, + 0.005897286348044872, + 0.007685437332838774, + 0.005576869938522577, + 0.008813014253973961, + 0.00802833866328001, + 0.01657845824956894, + 0.014513472095131874, + 0.011645649559795856, + 0.011936415918171406, + 0.014189418405294418, + 0.015789711847901344, + 0.006324877962470055, + 0.0050733364187181, + 0.011122727766633034, + 0.007073854096233845, + 0.00716004241257906, + 0.005906605161726475, + 0.010273387655615807, + 0.008214314468204975, + 0.005735355895012617, + 0.007168253883719444, + 0.012804723344743252, + 0.006675445940345526, + 0.006710213143378496, + 0.01329889427870512, + 0.010338401421904564, + 0.01331306528300047, + 0.013769306242465973, + 0.026772186160087585, + 0.03685712069272995, + 0.014974797144532204, + 0.016485178843140602, + 0.06313945353031158, + 0.019286179915070534, + 0.08423730731010437, + 0.009921801276504993, + 0.007136758416891098, + 0.021863888949155807, + 0.0076238545589149, + 0.004072678275406361, + 0.011172525584697723, + 0.008868820033967495, + 0.025853849947452545, + 0.005236672703176737, + 0.01003974862396717, + 0.0067277150228619576, + 0.009388637728989124, + 0.007034298498183489, + 0.010490236803889275, + 0.007045212667435408, + 0.005721325520426035 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 0, + 4 + ], + "selected_reward": -0.2813486158847809 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are two English setters visible in the video?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera positioned at a low angle and moving upward (ascending) while tracking the dogs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the dogs running through a vast expanse of tall, golden grass?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the lead dog arching its back to leap with its mouth open, while the second dog presses forward with visible muscles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene bathed in warm golden hour light with a lush green forest background featuring soft-focus bokeh?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02328108623623848 + ], + "opposite_logprob_tokens": [ + -3.7732810974121094 + ], + "logprob_reward": -0.02328108623623848, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.02328108623623848, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.18089477717876434, + 0.02443680167198181, + 0.08444897085428238, + 0.03480801358819008, + 0.14391085505485535, + 0.03451114147901535, + 0.06685299426317215, + 0.0738762691617012, + 0.33916711807250977, + 0.017093006521463394 + ], + "credit_map": [ + 0.011660808697342873, + 0.010338173247873783, + 0.022838041186332703, + 0.015734782442450523, + 0.01842624321579933, + 0.025514837354421616, + 0.03782173991203308, + 0.03856014460325241, + 0.0022666603326797485, + 0.002807376440614462, + 0.0028411652892827988, + 0.003719047410413623, + 0.002456600544974208, + 0.0023078632075339556, + 0.003669297555461526, + 0.00436879089102149, + 0.004663517698645592, + 0.006570250261574984, + 0.01061864010989666, + 0.008538019843399525, + 0.0075032892636954784, + 0.009160256944596767, + 0.019001955166459084, + 0.01839303970336914, + 0.003957390785217285, + 0.002623350592330098, + 0.003097676672041416, + 0.004513496533036232, + 0.0033466252498328686, + 0.0044861831702291965, + 0.006191112566739321, + 0.0065921759232878685, + 0.014660315588116646, + 0.009312949143350124, + 0.030362609773874283, + 0.01842077448964119, + 0.011736325919628143, + 0.020131820812821388, + 0.01974838227033615, + 0.01953769102692604, + 0.0022217941004782915, + 0.0018726619891822338, + 0.010711190290749073, + 0.0025523302610963583, + 0.0031661810353398323, + 0.004333538934588432, + 0.004001097287982702, + 0.0056523471139371395, + 0.006767813116312027, + 0.006177074741572142, + 0.00828552432358265, + 0.006376602686941624, + 0.008310833014547825, + 0.01060426700860262, + 0.009960854426026344, + 0.010370027273893356, + 0.005241036880761385, + 0.0063349297270178795, + 0.010664029978215694, + 0.004672632552683353, + 0.004518368747085333, + 0.014714572578668594, + 0.005906246602535248, + 0.02182445302605629, + 0.014710312709212303, + 0.03829606622457504, + 0.06470996141433716, + 0.011260727420449257, + 0.009933610446751118, + 0.08097607642412186, + 0.012889898382127285, + 0.10639049112796783, + 0.0015246717957779765, + 0.0030013893265277147, + 0.0018821567064151168, + 0.0026865850668400526, + 0.0019508585100993514, + 0.0026431307196617126, + 0.0016670424956828356, + 0.0017371721332892776 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 0, + 4 + ], + "selected_reward": -0.02328108623623848 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are two English setters visible in the video?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera positioned at a low angle and moving upward (ascending) while tracking the dogs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the dogs running through a vast expanse of tall, golden grass?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the lead dog arching its back to leap with its mouth open, while the second dog presses forward with visible muscles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene bathed in warm golden hour light with a lush green forest background featuring soft-focus bokeh?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0028208012226969004 + ], + "opposite_logprob_tokens": [ + -5.87782096862793 + ], + "logprob_reward": -0.0028208012226969004, + "yes_no_margin_reward": 5.875, + "token_credit_reward": -0.0028208012226969004, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.18503274023532867, + 0.0801708847284317, + 0.04199594259262085, + 0.04460783302783966, + 0.051382940262556076, + 0.06907206028699875, + 0.10538928955793381, + 0.15704657137393951, + 0.2258232831954956, + 0.03947841376066208 + ], + "credit_map": [ + 0.010666314512491226, + 0.015546529553830624, + 0.020018085837364197, + 0.01935727149248123, + 0.01357276365160942, + 0.0168073158711195, + 0.04002658277750015, + 0.04903789609670639, + 0.011044166050851345, + 0.013349451124668121, + 0.009731020778417587, + 0.010075380094349384, + 0.009143281728029251, + 0.007214422337710857, + 0.010679231956601143, + 0.008933940902352333, + 0.004206002224236727, + 0.0054426370188593864, + 0.003360654693096876, + 0.005208964459598064, + 0.004568568896502256, + 0.005083882715553045, + 0.007167481817305088, + 0.0069577512331306934, + 0.004447834566235542, + 0.0032573698554188013, + 0.004719802178442478, + 0.009251458570361137, + 0.004245965275913477, + 0.0039477963000535965, + 0.006458533927798271, + 0.008279076777398586, + 0.005944457370787859, + 0.004787974059581757, + 0.008162400685250759, + 0.006207237485796213, + 0.005713434424251318, + 0.0051948982290923595, + 0.008054549805819988, + 0.0073179942555725574, + 0.006615716964006424, + 0.006059035658836365, + 0.007188279647380114, + 0.006999845616519451, + 0.012518316507339478, + 0.009828034788370132, + 0.011809661984443665, + 0.008053172379732132, + 0.008089725859463215, + 0.010587242431938648, + 0.015806404873728752, + 0.011111604049801826, + 0.013482918031513691, + 0.016152063384652138, + 0.016321716830134392, + 0.013837619684636593, + 0.009551686234772205, + 0.014532036148011684, + 0.021203475072979927, + 0.009044332429766655, + 0.010114411823451519, + 0.034622300416231155, + 0.010298228822648525, + 0.047680098563432693, + 0.015940766781568527, + 0.019661782309412956, + 0.03663695976138115, + 0.020510103553533554, + 0.015974029898643494, + 0.03588425740599632, + 0.0250955019146204, + 0.05611991509795189, + 0.003557114163413644, + 0.00501436460763216, + 0.004176386632025242, + 0.007533934898674488, + 0.003969232086092234, + 0.006140329409390688, + 0.005322573706507683, + 0.0037644763942807913 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 0, + 7 + ], + "selected_reward": -0.0028208012226969004 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are two English setters visible in the video?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera positioned at a low angle and moving upward (ascending) while tracking the dogs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the dogs running through a vast expanse of tall, golden grass?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the lead dog arching its back to leap with its mouth open, while the second dog presses forward with visible muscles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene bathed in warm golden hour light with a lush green forest background featuring soft-focus bokeh?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.008631540462374687 + ], + "opposite_logprob_tokens": [ + -4.758631706237793 + ], + "logprob_reward": -0.008631540462374687, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.008631540462374687, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0719258189201355, + 0.021394986659288406, + 0.02395014837384224, + 0.011137224733829498, + 0.1107645258307457, + 0.022156566381454468, + 0.05971597135066986, + 0.06124502047896385, + 0.6098523139953613, + 0.007857365533709526 + ], + "credit_map": [ + 0.005444547161459923, + 0.009753472171723843, + 0.009377124719321728, + 0.010732765309512615, + 0.006106849294155836, + 0.008852455765008926, + 0.009653202258050442, + 0.012005413882434368, + 0.0014574916567653418, + 0.0013271102216094732, + 0.0026302526239305735, + 0.0021791476756334305, + 0.0026377371978014708, + 0.004771131090819836, + 0.0032263894099742174, + 0.0031657300423830748, + 0.0016157381469383836, + 0.0022104978561401367, + 0.004753799177706242, + 0.002661315491423011, + 0.0025913994759321213, + 0.0029413856100291014, + 0.004068933427333832, + 0.003107082098722458, + 0.0015607508830726147, + 0.0011674618581309915, + 0.001174065051600337, + 0.0017305585788562894, + 0.0011262826155871153, + 0.001101315370760858, + 0.0016357994172722101, + 0.0016409916570410132, + 0.013033721596002579, + 0.007221340201795101, + 0.023566093295812607, + 0.011936582624912262, + 0.00854992400854826, + 0.015354624949395657, + 0.016067391261458397, + 0.015034861862659454, + 0.00218376936390996, + 0.0021881575230509043, + 0.003984977025538683, + 0.003057122230529785, + 0.002603826578706503, + 0.0024075109977275133, + 0.0027655453886836767, + 0.0029656614642590284, + 0.005904903635382652, + 0.005566196981817484, + 0.009375199675559998, + 0.005468958057463169, + 0.006961238104850054, + 0.008572251535952091, + 0.009254361502826214, + 0.008612868376076221, + 0.005083301570266485, + 0.008889200165867805, + 0.0073480261489748955, + 0.00844553392380476, + 0.006180042400956154, + 0.0072882822714746, + 0.008717821910977364, + 0.00929282046854496, + 0.024890318512916565, + 0.07047511637210846, + 0.11566529422998428, + 0.01880049705505371, + 0.018212899565696716, + 0.147969052195549, + 0.022025514394044876, + 0.1918136328458786, + 0.0009108781232498586, + 0.0012442675651982427, + 0.000805172196123749, + 0.000741174619179219, + 0.0009396147215738893, + 0.0014554322697222233, + 0.0008934595971368253, + 0.0008673671400174499 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 4, + 0 + ], + "selected_reward": -0.008631540462374687 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "In a breathtaking display of high-speed photography aesthetics, a dynamic low-angle tracking shot begins, the camera ascending smoothly in ultra-slow motion, following two majestic English setters. They are captured en route by a vast expanse of tall, golden grass, which sways gently around them. The two dogs, one slightly ahead, are rendered in sharp focus; their long, silky, feathered coats—bright white speckled with liver-brown spots—ripple and flow with an almost liquid grace. The lead dog arches its back, leaping powerfully, its mouth open to reveal a swell of joy, while the second dog presses forward, powerful muscles visible beneath its flowing fur. The entire scene is bathed in the warm, low-angle light of golden hour, capturing every hair in their coats and creating a brilliant rim-lit effect. As their paws stir the verdant undergrowth, wisps of morning dew are kicked into the air, each drop suspended in mid-air like a shimmering shower of diamonds, frozen in time and refracting the sunlight into tiny, brilliant prisms. The background features a lush green forest, creating a beautiful, soft-focus bokeh that isolates the dogs and emphasizes their dynamic movements, creating a stunning atmosphere of pure energy and natural elegance.", + "seed": 27067, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000050/group_2/index_3304_seed_27067_dp_5_rank_10_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 2, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.03183716915082187, + "vlm_logprob": -0.03183716915082187, + "vlm_yes_no_margin": 4.925, + "vlm_token_credit": -0.03183716915082187, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are two English setters visible in the video?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera positioned at a low angle and moving upward (ascending) while tracking the dogs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the dogs running through a vast expanse of tall, golden grass?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the lead dog arching its back to leap with its mouth open, while the second dog presses forward with visible muscles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene bathed in warm golden hour light with a lush green forest background featuring soft-focus bokeh?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.038990944623947144 + ], + "opposite_logprob_tokens": [ + -3.4139909744262695 + ], + "logprob_reward": -0.038990944623947144, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.038990944623947144, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1421666145324707, + 0.03471873328089714, + 0.15010346472263336, + 0.05285082384943962, + 0.0796927660703659, + 0.07245336472988129, + 0.16500408947467804, + 0.05120030418038368, + 0.14337508380413055, + 0.10843478888273239 + ], + "credit_map": [ + 0.013988121412694454, + 0.008454409427940845, + 0.02110743150115013, + 0.0321945920586586, + 0.007932005450129509, + 0.010180868208408356, + 0.014267592690885067, + 0.03404159098863602, + 0.002603778848424554, + 0.003250812878832221, + 0.0050307693891227245, + 0.006258237175643444, + 0.003794076619669795, + 0.002971583278849721, + 0.004844937939196825, + 0.005964536685496569, + 0.008745111525058746, + 0.006300490349531174, + 0.048891935497522354, + 0.018109112977981567, + 0.00797591544687748, + 0.006955181714147329, + 0.009839136153459549, + 0.04328657314181328, + 0.0043381089344620705, + 0.006423261482268572, + 0.010033813305199146, + 0.00856665801256895, + 0.0035744369961321354, + 0.005450087133795023, + 0.007742886431515217, + 0.006721571087837219, + 0.015239117667078972, + 0.011648591607809067, + 0.010075916536152363, + 0.010556607507169247, + 0.007451030425727367, + 0.00824659038335085, + 0.007925780490040779, + 0.00854913704097271, + 0.007525392808020115, + 0.014369003474712372, + 0.013787802308797836, + 0.01311742328107357, + 0.005822645500302315, + 0.007116962689906359, + 0.005340164992958307, + 0.005373972933739424, + 0.014597749337553978, + 0.021030718460679054, + 0.02213853970170021, + 0.02705463394522667, + 0.020496251061558723, + 0.014904930256307125, + 0.018667323514819145, + 0.02611394040286541, + 0.005508597008883953, + 0.005791880190372467, + 0.005878019612282515, + 0.006771152373403311, + 0.004903811030089855, + 0.0058834427036345005, + 0.010457014665007591, + 0.00600638659670949, + 0.02257189340889454, + 0.013798200525343418, + 0.019588442519307137, + 0.020591339096426964, + 0.011282301507890224, + 0.014996876940131187, + 0.018020207062363625, + 0.02252582088112831, + 0.01555425301194191, + 0.011898167431354523, + 0.011385586112737656, + 0.010516646318137646, + 0.023735104128718376, + 0.012151180766522884, + 0.012272427789866924, + 0.010921427048742771 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 2, + 8 + ], + "selected_reward": -0.038990944623947144 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are two English setters visible in the video?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera positioned at a low angle and moving upward (ascending) while tracking the dogs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the dogs running through a vast expanse of tall, golden grass?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the lead dog arching its back to leap with its mouth open, while the second dog presses forward with visible muscles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene bathed in warm golden hour light with a lush green forest background featuring soft-focus bokeh?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.11295406520366669 + ], + "opposite_logprob_tokens": [ + -2.2379541397094727 + ], + "logprob_reward": -0.11295406520366669, + "yes_no_margin_reward": 2.125, + "token_credit_reward": -0.11295406520366669, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08721242845058441, + 0.06410139054059982, + 0.18898814916610718, + 0.07360192388296127, + 0.2515925168991089, + 0.10528282076120377, + 0.04750810191035271, + 0.07898915559053421, + 0.060346879065036774, + 0.04237663373351097 + ], + "credit_map": [ + 0.009649388492107391, + 0.008138766512274742, + 0.010680163279175758, + 0.0184757299721241, + 0.010063840076327324, + 0.007922462187707424, + 0.010159584693610668, + 0.01212248858064413, + 0.008256758563220501, + 0.0056577748619019985, + 0.008020727895200253, + 0.009564482606947422, + 0.006260197144001722, + 0.008163661696016788, + 0.008424953557550907, + 0.009752830490469933, + 0.015128902159631252, + 0.008312918245792389, + 0.06061350554227829, + 0.021845297887921333, + 0.007447908632457256, + 0.009149310179054737, + 0.013153256848454475, + 0.05333704501390457, + 0.007919350638985634, + 0.007125711068511009, + 0.011689617298543453, + 0.011269365437328815, + 0.006123312283307314, + 0.010261277668178082, + 0.010260416194796562, + 0.008952873758971691, + 0.04000237211585045, + 0.033099524676799774, + 0.03663763776421547, + 0.04422128200531006, + 0.0187703724950552, + 0.0231636893004179, + 0.025134798139333725, + 0.030562831088900566, + 0.010763878934085369, + 0.017362700775265694, + 0.01164696179330349, + 0.0145309679210186, + 0.011011005379259586, + 0.011849352158606052, + 0.015182126313447952, + 0.01293581910431385, + 0.004289654083549976, + 0.005953735671937466, + 0.006847293581813574, + 0.00645323796197772, + 0.003987296484410763, + 0.004378836136311293, + 0.0054137068800628185, + 0.01018434390425682, + 0.010499885305762291, + 0.007032017223536968, + 0.010016435757279396, + 0.01350010558962822, + 0.008895517326891422, + 0.00688152713701129, + 0.013252345845103264, + 0.008911320939660072, + 0.007663238327950239, + 0.005256091710180044, + 0.0101854233071208, + 0.012377019040286541, + 0.0035942711401730776, + 0.004603585693985224, + 0.00681535666808486, + 0.009851895272731781, + 0.004687875043600798, + 0.003759783459827304, + 0.0061068423092365265, + 0.005540892947465181, + 0.003925215918570757, + 0.00372793385758996, + 0.006523179356008768, + 0.008104909211397171 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 2, + 5 + ], + "selected_reward": -0.11295406520366669 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are two English setters visible in the video?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera positioned at a low angle and moving upward (ascending) while tracking the dogs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the dogs running through a vast expanse of tall, golden grass?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the lead dog arching its back to leap with its mouth open, while the second dog presses forward with visible muscles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene bathed in warm golden hour light with a lush green forest background featuring soft-focus bokeh?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0008123196894302964 + ], + "opposite_logprob_tokens": [ + -7.125812530517578 + ], + "logprob_reward": -0.0008123196894302964, + "yes_no_margin_reward": 7.125, + "token_credit_reward": -0.0008123196894302964, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3731744885444641, + 0.0617988258600235, + 0.20910802483558655, + 0.05074234679341316, + 0.06708869338035583, + 0.030803533270955086, + 0.02840585820376873, + 0.04496690630912781, + 0.0996801108121872, + 0.03423114866018295 + ], + "credit_map": [ + 0.023712141439318657, + 0.017329098656773567, + 0.05764990299940109, + 0.09845343232154846, + 0.02218192256987095, + 0.021197713911533356, + 0.036108262836933136, + 0.09654208272695541, + 0.0055524627678096294, + 0.005859337281435728, + 0.009540600702166557, + 0.010636444203555584, + 0.006782382261008024, + 0.005652351770550013, + 0.00655166944488883, + 0.011223585344851017, + 0.012478411197662354, + 0.009010188281536102, + 0.06962310522794724, + 0.02234731987118721, + 0.011114311404526234, + 0.009287385269999504, + 0.014204632490873337, + 0.06104268506169319, + 0.00467744842171669, + 0.005612209904938936, + 0.009400663897395134, + 0.009164048358798027, + 0.0031667379662394524, + 0.005105467513203621, + 0.006525893695652485, + 0.007089884951710701, + 0.012315647676587105, + 0.009321197867393494, + 0.008718607947230339, + 0.008510210551321507, + 0.005747063551098108, + 0.006377102341502905, + 0.007806522771716118, + 0.008292345330119133, + 0.004819623660296202, + 0.004657722543925047, + 0.0035582929849624634, + 0.004461351782083511, + 0.0027192900888621807, + 0.003277981886640191, + 0.0037340521812438965, + 0.003575221635401249, + 0.0037833512760698795, + 0.0064127822406589985, + 0.0034637381322681904, + 0.003992755431681871, + 0.0021458761766552925, + 0.002408354077488184, + 0.0026445204857736826, + 0.003554483177140355, + 0.005456836894154549, + 0.0039010795298963785, + 0.00550832599401474, + 0.008593997918069363, + 0.0061146290972828865, + 0.004075523465871811, + 0.006303852424025536, + 0.0050126644782722, + 0.010281410068273544, + 0.006340376101434231, + 0.016769524663686752, + 0.01878366805613041, + 0.006277696695178747, + 0.008351870812475681, + 0.008749177679419518, + 0.02412639744579792, + 0.0056485640816390514, + 0.004258259665220976, + 0.003139851149171591, + 0.0036381161771714687, + 0.006497533526271582, + 0.0029324570205062628, + 0.004778758157044649, + 0.0033376128412783146 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 8 + ], + "selected_reward": -0.0008123196894302964 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are two English setters visible in the video?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera positioned at a low angle and moving upward (ascending) while tracking the dogs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the dogs running through a vast expanse of tall, golden grass?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the lead dog arching its back to leap with its mouth open, while the second dog presses forward with visible muscles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene bathed in warm golden hour light with a lush green forest background featuring soft-focus bokeh?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005251778755337 + ], + "opposite_logprob_tokens": [ + -5.255251884460449 + ], + "logprob_reward": -0.005251778755337, + "yes_no_margin_reward": 5.25, + "token_credit_reward": -0.005251778755337, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12762103974819183, + 0.05595286563038826, + 0.059944942593574524, + 0.07443708926439285, + 0.1298130452632904, + 0.06636381149291992, + 0.08701815456151962, + 0.21638430655002594, + 0.095232754945755, + 0.0872320681810379 + ], + "credit_map": [ + 0.014626869931817055, + 0.008033121936023235, + 0.016855059191584587, + 0.0235025305300951, + 0.01157633401453495, + 0.01089177094399929, + 0.014538013376295567, + 0.02759733609855175, + 0.004170721396803856, + 0.004988345317542553, + 0.008803247474133968, + 0.011517083272337914, + 0.0038812444545328617, + 0.0038676103577017784, + 0.007739027962088585, + 0.010985583066940308, + 0.0042539662681519985, + 0.00465051457285881, + 0.014521614648401737, + 0.008076940663158894, + 0.005485204979777336, + 0.004146585240960121, + 0.006235300097614527, + 0.012574818916618824, + 0.009502630680799484, + 0.00789930671453476, + 0.010316820815205574, + 0.009537453763186932, + 0.0065087564289569855, + 0.011097941547632217, + 0.01010112464427948, + 0.009473051875829697, + 0.019215712323784828, + 0.016206169500947, + 0.020292537286877632, + 0.021583689376711845, + 0.0107798483222723, + 0.011271888390183449, + 0.01428297907114029, + 0.016180209815502167, + 0.007082024589180946, + 0.009203173220157623, + 0.012474162504076958, + 0.011974552646279335, + 0.007168120238929987, + 0.0065580871887505054, + 0.006214526016265154, + 0.005689162760972977, + 0.008878584019839764, + 0.01129491813480854, + 0.011266166344285011, + 0.01303686574101448, + 0.009785514324903488, + 0.00876718107610941, + 0.009507822804152966, + 0.014481093734502792, + 0.019008958712220192, + 0.022064466029405594, + 0.027619820088148117, + 0.04275982454419136, + 0.020814113318920135, + 0.017945969477295876, + 0.042581088840961456, + 0.023590058088302612, + 0.0149746248498559, + 0.008986101485788822, + 0.010747503489255905, + 0.012150297872722149, + 0.008137300610542297, + 0.012338114902377129, + 0.013040347956120968, + 0.014858460985124111, + 0.012590117752552032, + 0.0074496944434940815, + 0.009834484197199345, + 0.007815894670784473, + 0.02084159106016159, + 0.007429158315062523, + 0.00856689177453518, + 0.012704232707619667 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 4, + 0 + ], + "selected_reward": -0.005251778755337 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are two English setters visible in the video?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera positioned at a low angle and moving upward (ascending) while tracking the dogs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the dogs running through a vast expanse of tall, golden grass?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the lead dog arching its back to leap with its mouth open, while the second dog presses forward with visible muscles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene bathed in warm golden hour light with a lush green forest background featuring soft-focus bokeh?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0011767374817281961 + ], + "opposite_logprob_tokens": [ + -6.751176834106445 + ], + "logprob_reward": -0.0011767374817281961, + "yes_no_margin_reward": 6.75, + "token_credit_reward": -0.0011767374817281961, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10304836183786392, + 0.03980446234345436, + 0.21411477029323578, + 0.07320361584424973, + 0.06635526567697525, + 0.030380714684724808, + 0.06916655600070953, + 0.0988667905330658, + 0.24610675871372223, + 0.0589526928961277 + ], + "credit_map": [ + 0.009230579249560833, + 0.009564436972141266, + 0.01326401624828577, + 0.017877832055091858, + 0.013440831564366817, + 0.009194580838084221, + 0.01670873537659645, + 0.013767349533736706, + 0.0046802544966340065, + 0.004594794474542141, + 0.005808081943541765, + 0.006087764631956816, + 0.0038541185203939676, + 0.004559638444334269, + 0.005154810845851898, + 0.005064999219030142, + 0.016004297882318497, + 0.032372839748859406, + 0.035823479294776917, + 0.036099307239055634, + 0.015782007947564125, + 0.022551286965608597, + 0.02044149488210678, + 0.035040054470300674, + 0.008311969228088856, + 0.004888491239398718, + 0.019762596115469933, + 0.011207913979887962, + 0.0038800351321697235, + 0.010288204066455364, + 0.005012524779886007, + 0.009851878508925438, + 0.011106566525995731, + 0.008379851467907429, + 0.00928440224379301, + 0.011635830625891685, + 0.003994286991655827, + 0.006105589680373669, + 0.0074865976348519325, + 0.00836214330047369, + 0.005242710467427969, + 0.003990828990936279, + 0.005074383225291967, + 0.005631172563880682, + 0.0023980310652405024, + 0.00320827541872859, + 0.0025568115524947643, + 0.0022785011678934097, + 0.007567914668470621, + 0.0121458750218153, + 0.008291441015899181, + 0.008136504329741001, + 0.005673978943377733, + 0.007264711894094944, + 0.00832480937242508, + 0.011761329136788845, + 0.01489326823502779, + 0.011151795275509357, + 0.009286514483392239, + 0.01128415111452341, + 0.018265407532453537, + 0.009764943271875381, + 0.015134919434785843, + 0.009085786528885365, + 0.028840992599725723, + 0.018427029252052307, + 0.03892356529831886, + 0.03941555693745613, + 0.017002880573272705, + 0.022388873621821404, + 0.022763365879654884, + 0.05834450200200081, + 0.010343946516513824, + 0.0056074089370667934, + 0.00504320627078414, + 0.006374555639922619, + 0.0149820726364851, + 0.005501955281943083, + 0.0069686975330114365, + 0.004130850546061993 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 2, + 0 + ], + "selected_reward": -0.0011767374817281961 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "In a breathtaking display of high-speed photography aesthetics, a dynamic low-angle tracking shot begins, the camera ascending smoothly in ultra-slow motion, following two majestic English setters. They are captured en route by a vast expanse of tall, golden grass, which sways gently around them. The two dogs, one slightly ahead, are rendered in sharp focus; their long, silky, feathered coats—bright white speckled with liver-brown spots—ripple and flow with an almost liquid grace. The lead dog arches its back, leaping powerfully, its mouth open to reveal a swell of joy, while the second dog presses forward, powerful muscles visible beneath its flowing fur. The entire scene is bathed in the warm, low-angle light of golden hour, capturing every hair in their coats and creating a brilliant rim-lit effect. As their paws stir the verdant undergrowth, wisps of morning dew are kicked into the air, each drop suspended in mid-air like a shimmering shower of diamonds, frozen in time and refracting the sunlight into tiny, brilliant prisms. The background features a lush green forest, creating a beautiful, soft-focus bokeh that isolates the dogs and emphasizes their dynamic movements, creating a stunning atmosphere of pure energy and natural elegance.", + "seed": 27068, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000050/group_2/index_3304_seed_27068_dp_5_rank_10_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 2, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.0718693190254271, + "vlm_logprob": -0.0718693190254271, + "vlm_yes_no_margin": 3.9, + "vlm_token_credit": -0.0718693190254271, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are two English setters visible in the video?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera positioned at a low angle and moving upward (ascending) while tracking the dogs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the dogs running through a vast expanse of tall, golden grass?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the lead dog arching its back to leap with its mouth open, while the second dog presses forward with visible muscles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene bathed in warm golden hour light with a lush green forest background featuring soft-focus bokeh?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.055638182908296585 + ], + "opposite_logprob_tokens": [ + -3.180638074874878 + ], + "logprob_reward": -0.055638182908296585, + "yes_no_margin_reward": 3.125, + "token_credit_reward": -0.055638182908296585, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14932586252689362, + 0.05877762287855148, + 0.05361658334732056, + 0.21403883397579193, + 0.04847458377480507, + 0.15936335921287537, + 0.03748888522386551, + 0.08287078887224197, + 0.14290620386600494, + 0.053137216717004776 + ], + "credit_map": [ + 0.012811757624149323, + 0.010072202421724796, + 0.01810552179813385, + 0.03696862608194351, + 0.014632467180490494, + 0.00951619353145361, + 0.02331128902733326, + 0.023907799273729324, + 0.0061043850146234035, + 0.006437630392611027, + 0.006466357968747616, + 0.010081819258630276, + 0.00677003338932991, + 0.005352955311536789, + 0.0081355394795537, + 0.009428899735212326, + 0.004511155188083649, + 0.004697985947132111, + 0.006338187493383884, + 0.00646363478153944, + 0.0045867436565458775, + 0.004285850096493959, + 0.014253145083785057, + 0.008479880169034004, + 0.014579277485609055, + 0.019835827872157097, + 0.03254224359989166, + 0.03781065344810486, + 0.03292646259069443, + 0.019958870485424995, + 0.028149167075753212, + 0.02823633886873722, + 0.005014920141547918, + 0.004059930797666311, + 0.005787207745015621, + 0.008130602538585663, + 0.004372242372483015, + 0.00649449834600091, + 0.00641458947211504, + 0.00820059236139059, + 0.01555642019957304, + 0.02638521045446396, + 0.020912697538733482, + 0.02708355523645878, + 0.016405409201979637, + 0.015375600196421146, + 0.020224671810865402, + 0.017419803887605667, + 0.0038476285990327597, + 0.0036336167249828577, + 0.004927204456180334, + 0.00348958489485085, + 0.004838863853365183, + 0.003837927244603634, + 0.007000989746302366, + 0.005913070868700743, + 0.01056664902716875, + 0.004518046975135803, + 0.010787767358124256, + 0.012254057452082634, + 0.010428238660097122, + 0.009454541839659214, + 0.013526449911296368, + 0.011335039511322975, + 0.01881876401603222, + 0.011421536095440388, + 0.014937514439225197, + 0.025758620351552963, + 0.013640502467751503, + 0.017317818477749825, + 0.02737530507147312, + 0.013636130839586258, + 0.007408599369227886, + 0.004740615375339985, + 0.010482666082680225, + 0.006285436451435089, + 0.004606219939887524, + 0.005430284421890974, + 0.008522645570337772, + 0.00566075136885047 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 5, + 0 + ], + "selected_reward": -0.055638182908296585 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are two English setters visible in the video?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera positioned at a low angle and moving upward (ascending) while tracking the dogs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the dogs running through a vast expanse of tall, golden grass?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the lead dog arching its back to leap with its mouth open, while the second dog presses forward with visible muscles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene bathed in warm golden hour light with a lush green forest background featuring soft-focus bokeh?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.28133025765419006 + ], + "opposite_logprob_tokens": [ + -1.4063302278518677 + ], + "logprob_reward": -0.28133025765419006, + "yes_no_margin_reward": 1.125, + "token_credit_reward": -0.28133025765419006, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.31010425090789795, + 0.061622630804777145, + 0.0574108250439167, + 0.07403779774904251, + 0.08744530379772186, + 0.06620973348617554, + 0.0908396989107132, + 0.1065688207745552, + 0.05055834725499153, + 0.09520255774259567 + ], + "credit_map": [ + 0.021330727264285088, + 0.019756190478801727, + 0.043506309390068054, + 0.07511202991008759, + 0.022295663133263588, + 0.017206426709890366, + 0.05768275633454323, + 0.05321415513753891, + 0.007099973503500223, + 0.005953633692115545, + 0.007783274631947279, + 0.01076535228639841, + 0.005641678348183632, + 0.005206263158470392, + 0.010249802842736244, + 0.008922655135393143, + 0.008166053332388401, + 0.008813384920358658, + 0.007035760674625635, + 0.007941526360809803, + 0.005407167132943869, + 0.005053770262748003, + 0.00765793165192008, + 0.007335230242460966, + 0.006318067666143179, + 0.0052863783203065395, + 0.01097800862044096, + 0.00931750237941742, + 0.006547186058014631, + 0.006172510329633951, + 0.018901845440268517, + 0.010516304522752762, + 0.008210464380681515, + 0.004857530817389488, + 0.01615152880549431, + 0.009212802164256573, + 0.00796013418585062, + 0.008651508018374443, + 0.01850079372525215, + 0.013900546357035637, + 0.004475433379411697, + 0.003222682047635317, + 0.014155011624097824, + 0.006964992731809616, + 0.004593716003000736, + 0.004157332703471184, + 0.0189973171800375, + 0.00964324176311493, + 0.009245771914720535, + 0.007801573257893324, + 0.01708488166332245, + 0.008851777762174606, + 0.011928465217351913, + 0.008060552179813385, + 0.017579777166247368, + 0.01028689369559288, + 0.01202754769474268, + 0.006577266845852137, + 0.015582063235342503, + 0.016126247122883797, + 0.006195097230374813, + 0.012174723669886589, + 0.02591390162706375, + 0.011971971951425076, + 0.0059767067432403564, + 0.004884588066488504, + 0.006665953900665045, + 0.007450311444699764, + 0.004807536955922842, + 0.006290701683610678, + 0.009216840378940105, + 0.005265707615762949, + 0.010690763592720032, + 0.012269251979887486, + 0.012458630837500095, + 0.011456849053502083, + 0.00963084027171135, + 0.008405612781643867, + 0.011387730948626995, + 0.01890287734568119 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 7, + 9 + ], + "selected_reward": -0.28133025765419006 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are two English setters visible in the video?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera positioned at a low angle and moving upward (ascending) while tracking the dogs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the dogs running through a vast expanse of tall, golden grass?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the lead dog arching its back to leap with its mouth open, while the second dog presses forward with visible muscles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene bathed in warm golden hour light with a lush green forest background featuring soft-focus bokeh?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.014190951362252235 + ], + "opposite_logprob_tokens": [ + -4.264191150665283 + ], + "logprob_reward": -0.014190951362252235, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.014190951362252235, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3048163056373596, + 0.030436048284173012, + 0.18165694177150726, + 0.06719310581684113, + 0.05467089265584946, + 0.05108879506587982, + 0.11513678729534149, + 0.14929543435573578, + 0.022332895547151566, + 0.02337275631725788 + ], + "credit_map": [ + 0.029950762167572975, + 0.01791158691048622, + 0.03782918304204941, + 0.07899995148181915, + 0.016885967925190926, + 0.016489867120981216, + 0.049115244299173355, + 0.057633768767118454, + 0.0026507750153541565, + 0.0036257707979530096, + 0.0025978905614465475, + 0.004371054004877806, + 0.003442180110141635, + 0.004841704852879047, + 0.004154385533183813, + 0.004752288106828928, + 0.01413056906312704, + 0.009286012500524521, + 0.01952890306711197, + 0.016243671998381615, + 0.012481477111577988, + 0.01741980016231537, + 0.05755123868584633, + 0.03501525893807411, + 0.00529750669375062, + 0.004917752929031849, + 0.009788813069462776, + 0.008239109069108963, + 0.007528951391577721, + 0.008749357424676418, + 0.013067482970654964, + 0.0096041364595294, + 0.004534010775387287, + 0.0031527322717010975, + 0.007031840272247791, + 0.007523654494434595, + 0.005054070148617029, + 0.006859079468995333, + 0.010969199240207672, + 0.009546306915581226, + 0.0038764216005802155, + 0.003322462784126401, + 0.008569454774260521, + 0.005786059889942408, + 0.004285061731934547, + 0.005827073007822037, + 0.012378724291920662, + 0.007043539546430111, + 0.01588149182498455, + 0.00887245126068592, + 0.012541363015770912, + 0.014134771190583706, + 0.016335198655724525, + 0.012695980258286, + 0.01811249554157257, + 0.016563037410378456, + 0.02560785599052906, + 0.007493715733289719, + 0.020580526441335678, + 0.023179564625024796, + 0.009559922851622105, + 0.02020905539393425, + 0.023395514115691185, + 0.019269293174147606, + 0.0024990884121507406, + 0.0014210134977474809, + 0.00397598696872592, + 0.0026017543859779835, + 0.0015281703090295196, + 0.0021846676245331764, + 0.0036652248818427324, + 0.004456990398466587, + 0.00317785725928843, + 0.004053076263517141, + 0.002877084771171212, + 0.0029854909516870975, + 0.0018690506694838405, + 0.0027158199809491634, + 0.002898892154917121, + 0.002795485081151128 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 7 + ], + "selected_reward": -0.014190951362252235 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are two English setters visible in the video?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera positioned at a low angle and moving upward (ascending) while tracking the dogs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the dogs running through a vast expanse of tall, golden grass?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the lead dog arching its back to leap with its mouth open, while the second dog presses forward with visible muscles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene bathed in warm golden hour light with a lush green forest background featuring soft-focus bokeh?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004097400698810816 + ], + "opposite_logprob_tokens": [ + -5.5040974617004395 + ], + "logprob_reward": -0.004097400698810816, + "yes_no_margin_reward": 5.5, + "token_credit_reward": -0.004097400698810816, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.16779205203056335, + 0.08829262107610703, + 0.050160493701696396, + 0.12613581120967865, + 0.12217190861701965, + 0.06934453547000885, + 0.1442769169807434, + 0.09324101358652115, + 0.042756468057632446, + 0.09582817554473877 + ], + "credit_map": [ + 0.020151428878307343, + 0.011741611175239086, + 0.02043902687728405, + 0.024808472022414207, + 0.012285420671105385, + 0.015545541420578957, + 0.020123496651649475, + 0.04269704595208168, + 0.00836083572357893, + 0.007645762525498867, + 0.009408051148056984, + 0.022216659039258957, + 0.006217180751264095, + 0.006623128894716501, + 0.015022705309092999, + 0.012798291631042957, + 0.010099267587065697, + 0.007501534651964903, + 0.006134140770882368, + 0.005434050224721432, + 0.003944921772927046, + 0.005429188720881939, + 0.005674529820680618, + 0.005942854564636946, + 0.011437151581048965, + 0.012141279876232147, + 0.023256797343492508, + 0.01730620488524437, + 0.01644201949238777, + 0.010640017688274384, + 0.017823562026023865, + 0.017088772729039192, + 0.013357695192098618, + 0.014100492000579834, + 0.019427409395575523, + 0.02048533782362938, + 0.011371422559022903, + 0.01300820056349039, + 0.012729122303426266, + 0.017692239955067635, + 0.006051407195627689, + 0.011440792120993137, + 0.007863669656217098, + 0.012679625302553177, + 0.008148143999278545, + 0.006884897593408823, + 0.009426374919712543, + 0.0068496279418468475, + 0.013273187913000584, + 0.013640533201396465, + 0.02452932298183441, + 0.01918826438486576, + 0.023255977779626846, + 0.016490187495946884, + 0.019154200330376625, + 0.014745229855179787, + 0.008945541456341743, + 0.00872977264225483, + 0.01122724637389183, + 0.018725918605923653, + 0.007610739208757877, + 0.010646195150911808, + 0.01900380104780197, + 0.008351805619895458, + 0.004300221800804138, + 0.004441690165549517, + 0.005596253555268049, + 0.0068937549367547035, + 0.004562892485409975, + 0.00458394642919302, + 0.005593671463429928, + 0.0067840334959328175, + 0.015894541516900063, + 0.012281716801226139, + 0.029694676399230957, + 0.008965750224888325, + 0.005714744795113802, + 0.00660734111443162, + 0.006972299423068762, + 0.009697106666862965 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 6, + 3 + ], + "selected_reward": -0.004097400698810816 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are two English setters visible in the video?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera positioned at a low angle and moving upward (ascending) while tracking the dogs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the dogs running through a vast expanse of tall, golden grass?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the lead dog arching its back to leap with its mouth open, while the second dog presses forward with visible muscles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene bathed in warm golden hour light with a lush green forest background featuring soft-focus bokeh?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004089802503585815 + ], + "opposite_logprob_tokens": [ + -5.504089832305908 + ], + "logprob_reward": -0.004089802503585815, + "yes_no_margin_reward": 5.5, + "token_credit_reward": -0.004089802503585815, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2317737638950348, + 0.10925335437059402, + 0.04830920696258545, + 0.06782560050487518, + 0.2073807269334793, + 0.1544346660375595, + 0.04273960366845131, + 0.06473243981599808, + 0.01626237854361534, + 0.0572882778942585 + ], + "credit_map": [ + 0.017493296414613724, + 0.014345544390380383, + 0.030909869819879532, + 0.058308523148298264, + 0.013791575096547604, + 0.013772633858025074, + 0.0433805026113987, + 0.03977181017398834, + 0.011922764591872692, + 0.012311180122196674, + 0.01345041487365961, + 0.016667239367961884, + 0.010942780412733555, + 0.013997891917824745, + 0.01578359864652157, + 0.014177481643855572, + 0.004904883913695812, + 0.003104862989857793, + 0.005732037127017975, + 0.00524294376373291, + 0.0048835063353180885, + 0.004664176143705845, + 0.011464790441095829, + 0.008312004618346691, + 0.005679521709680557, + 0.005244830623269081, + 0.010143878869712353, + 0.009209628216922283, + 0.006392193958163261, + 0.006060587242245674, + 0.014212915673851967, + 0.01088204886764288, + 0.0152052603662014, + 0.010759824886918068, + 0.0372886024415493, + 0.029566211625933647, + 0.015876157209277153, + 0.016850754618644714, + 0.04899507761001587, + 0.0328388474881649, + 0.011888114735484123, + 0.00903516635298729, + 0.03789680451154709, + 0.01736207865178585, + 0.007646958343684673, + 0.009438600391149521, + 0.036487989127635956, + 0.024678945541381836, + 0.004040820989757776, + 0.0041430676355957985, + 0.004478313494473696, + 0.005574225913733244, + 0.005810759495943785, + 0.004886227194219828, + 0.00835573859512806, + 0.0054504466243088245, + 0.006371564231812954, + 0.004446186125278473, + 0.007479711901396513, + 0.011618453077971935, + 0.004764631390571594, + 0.007192270364612341, + 0.016317758709192276, + 0.006541862152516842, + 0.0019060169579461217, + 0.001373463892377913, + 0.0026098338421434164, + 0.0021895195823162794, + 0.001194542390294373, + 0.0014323381474241614, + 0.0017989743500947952, + 0.0037576889153569937, + 0.009556357748806477, + 0.005982999689877033, + 0.018907582387328148, + 0.00507294200360775, + 0.004181287717074156, + 0.0038140462711453438, + 0.004100488033145666, + 0.005672569386661053 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 5 + ], + "selected_reward": -0.004089802503585815 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000050/group_3/index_tensor([569, 569])_group_3_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000050/group_3/index_tensor([569, 569])_group_3_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..a3f5e3aec881feb5ede814ee998737ae136d11e1 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000050/group_3/index_tensor([569, 569])_group_3_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "static medium close-up, rendered in stark black-and-white ultra-high-speed photography, captures a majestic Snowy Owl in a moment of heightened alert. Perched on a snow-covered branch against a dark, out-of-focus background, the owl is a study in momentum. In an instant, its head dashes with incredible speed toward an unseen sound just off-camera. The entire sequence unfolds in extreme slow motion, transforming the violent action into an elegant balletic pirouette. The feathers on its neck and head ripple outward in mesmerizingly fluid waves, each barb and feather rendered in crystalline clarity. Its large, black, focused eyes gleam with a sharp, crystalline light as they lock onto a new target. As its head settles, a subtle, involuntary twitch energizes its powerful black talons, tightening their grip on the branch. The scene's high-contrast lighting creates a dramatic black-and-white contrast, casting deep shadows and accentuating the impeccable texture of the feathers, ultimately creating a powerful atmosphere of raw, focused instinct and sublime natural beauty.", + "seed": 29347, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000050/group_3/index_569_seed_29347_dp_6_rank_12_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 3, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.20347028374671935, + "vlm_logprob": -0.20347028374671935, + "vlm_yes_no_margin": 2.5, + "vlm_token_credit": -0.20347028374671935, + "vlm_yes_ratio": 0.8, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a Snowy Owl perched on a snow-covered branch?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the video rendered in stark black-and-white with high-contrast lighting?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the owl's head dash rapidly toward an unseen sound off-camera?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the feathers on the owl's neck and head shown rippling outward in fluid waves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot a static medium close-up?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.021520551294088364 + ], + "opposite_logprob_tokens": [ + -3.8965206146240234 + ], + "logprob_reward": -0.021520551294088364, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.021520551294088364, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09069783240556717, + 0.04618452116847038, + 0.11121898889541626, + 0.10276748239994049, + 0.041046205908060074, + 0.1130310595035553, + 0.17372651398181915, + 0.17660236358642578, + 0.04304927587509155, + 0.10167579352855682 + ], + "credit_map": [ + 0.009424257092177868, + 0.008727497421205044, + 0.012743600644171238, + 0.00878682266920805, + 0.01782796159386635, + 0.011449567973613739, + 0.01065769325941801, + 0.011080440133810043, + 0.005483395420014858, + 0.0058149900287389755, + 0.0068238540552556515, + 0.005029172170907259, + 0.006446001119911671, + 0.004208697937428951, + 0.00474977632984519, + 0.007628637365996838, + 0.007995521649718285, + 0.024262895807623863, + 0.00632951594889164, + 0.010725514031946659, + 0.012246417813003063, + 0.008699825964868069, + 0.027169590815901756, + 0.013789711520075798, + 0.009416797198355198, + 0.00679390411823988, + 0.006194813642650843, + 0.008538625203073025, + 0.012646909803152084, + 0.012580612674355507, + 0.015004130080342293, + 0.031591687351465225, + 0.010340415872633457, + 0.0035884964745491743, + 0.0027809112798422575, + 0.003968688193708658, + 0.0026831424329429865, + 0.008939067833125591, + 0.003822238417342305, + 0.0049232481978833675, + 0.010332220233976841, + 0.012672025710344315, + 0.019575556740164757, + 0.021521955728530884, + 0.013440161943435669, + 0.011612546630203724, + 0.008893826976418495, + 0.014982766471803188, + 0.024227039888501167, + 0.014589522033929825, + 0.016115864738821983, + 0.028971832245588303, + 0.02147911675274372, + 0.03404060751199722, + 0.017601314932107925, + 0.01670120842754841, + 0.03949064388871193, + 0.009711291640996933, + 0.04891519248485565, + 0.014515509828925133, + 0.014180387370288372, + 0.02547420747578144, + 0.010578569024801254, + 0.013736548833549023, + 0.0076057566329836845, + 0.007283039391040802, + 0.008337819017469883, + 0.004128138069063425, + 0.003928502090275288, + 0.004723392892628908, + 0.0039912923239171505, + 0.0030513389501720667, + 0.010939502157270908, + 0.012089638970792294, + 0.009030921384692192, + 0.009636913426220417, + 0.01315669808536768, + 0.0152875492349267, + 0.01996487006545067, + 0.011569707654416561 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 6, + 5 + ], + "selected_reward": -0.021520551294088364 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a Snowy Owl perched on a snow-covered branch?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the video rendered in stark black-and-white with high-contrast lighting?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the owl's head dash rapidly toward an unseen sound off-camera?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the feathers on the owl's neck and head shown rippling outward in fluid waves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot a static medium close-up?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.029833264648914337 + ], + "opposite_logprob_tokens": [ + -3.5298333168029785 + ], + "logprob_reward": -0.029833264648914337, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.029833264648914337, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.17067159712314606, + 0.14057974517345428, + 0.17268674075603485, + 0.08746892213821411, + 0.04496176913380623, + 0.11748148500919342, + 0.07701338827610016, + 0.06423832476139069, + 0.0651688426733017, + 0.05972922593355179 + ], + "credit_map": [ + 0.018287137150764465, + 0.01816796325147152, + 0.020770343020558357, + 0.010585714131593704, + 0.041689254343509674, + 0.024427128955721855, + 0.019024988636374474, + 0.01771906204521656, + 0.01961713843047619, + 0.015218019485473633, + 0.01646696776151657, + 0.015345322899520397, + 0.014510288834571838, + 0.022054200991988182, + 0.018589945510029793, + 0.018777847290039062, + 0.010578645393252373, + 0.03034212999045849, + 0.009530552662909031, + 0.01601146161556244, + 0.01699477806687355, + 0.017245033755898476, + 0.045154228806495667, + 0.026829877868294716, + 0.00966248195618391, + 0.008094449527561665, + 0.005407076794654131, + 0.00952698104083538, + 0.011573940515518188, + 0.011375716887414455, + 0.014909127727150917, + 0.016919145360589027, + 0.007975403219461441, + 0.005343357101082802, + 0.004452622961252928, + 0.0042008147574961185, + 0.004177391063421965, + 0.007027108687907457, + 0.0039053240325301886, + 0.007879742421209812, + 0.014536602422595024, + 0.013859105296432972, + 0.016860829666256905, + 0.018896888941526413, + 0.012429452501237392, + 0.013407479971647263, + 0.01023358665406704, + 0.017257537692785263, + 0.012312909588217735, + 0.007827159017324448, + 0.007737050298601389, + 0.00912103708833456, + 0.0063661979511380196, + 0.015673572197556496, + 0.005152884405106306, + 0.012822575867176056, + 0.008492955937981606, + 0.0053423629142344, + 0.007407305762171745, + 0.013663732446730137, + 0.01155377272516489, + 0.006835150998085737, + 0.006390361115336418, + 0.004552677273750305, + 0.014989030547440052, + 0.0075621223077178, + 0.017629653215408325, + 0.0045226155780255795, + 0.005912689492106438, + 0.005223780404776335, + 0.005257871933281422, + 0.004071079194545746, + 0.0076187169179320335, + 0.0058578019961714745, + 0.007532027084380388, + 0.0049453419633209705, + 0.00871515553444624, + 0.006346767768263817, + 0.014556651934981346, + 0.0041567618027329445 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 0, + 1 + ], + "selected_reward": -0.029833264648914337 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a Snowy Owl perched on a snow-covered branch?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the video rendered in stark black-and-white with high-contrast lighting?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the owl's head dash rapidly toward an unseen sound off-camera?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the feathers on the owl's neck and head shown rippling outward in fluid waves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot a static medium close-up?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -0.6933183670043945 + ], + "opposite_logprob_tokens": [ + -0.6933183670043945 + ], + "logprob_reward": -0.6933183670043945, + "yes_no_margin_reward": 0.0, + "token_credit_reward": -0.6933183670043945, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.07049548625946045, + 0.10312119871377945, + 0.1327899843454361, + 0.1031562089920044, + 0.19190546870231628, + 0.07803989946842194, + 0.11461172997951508, + 0.07364850491285324, + 0.0699293240904808, + 0.062302183359861374 + ], + "credit_map": [ + 0.007069894578307867, + 0.007713342551141977, + 0.010114272125065327, + 0.0072677647694945335, + 0.014516601338982582, + 0.008507254533469677, + 0.008880743756890297, + 0.0064256093464791775, + 0.012849901802837849, + 0.013597910292446613, + 0.013535619713366032, + 0.009547002613544464, + 0.012653217650949955, + 0.01739715225994587, + 0.007845289073884487, + 0.015695100650191307, + 0.011187777854502201, + 0.02059182897210121, + 0.01219804398715496, + 0.01701473817229271, + 0.017077676951885223, + 0.018305577337741852, + 0.021377865225076675, + 0.015036487951874733, + 0.019146213307976723, + 0.009135868400335312, + 0.0060095940716564655, + 0.007261202204972506, + 0.012876562774181366, + 0.011188621632754803, + 0.01285269483923912, + 0.024685455486178398, + 0.059361834079027176, + 0.007848808541893959, + 0.008908280171453953, + 0.0055517409928143024, + 0.010894187726080418, + 0.07384512573480606, + 0.00813926849514246, + 0.017356226220726967, + 0.00871119275689125, + 0.00785189401358366, + 0.009510799311101437, + 0.010475967079401016, + 0.010446316562592983, + 0.012955711223185062, + 0.005410047713667154, + 0.012677967548370361, + 0.016880564391613007, + 0.00963165145367384, + 0.010601974092423916, + 0.02001701295375824, + 0.01142650842666626, + 0.02415589429438114, + 0.011512834578752518, + 0.010385294444859028, + 0.01157955639064312, + 0.004680636804550886, + 0.013121635653078556, + 0.013472498394548893, + 0.010592878796160221, + 0.007079555187374353, + 0.007602682802826166, + 0.005519057158380747, + 0.009493274614214897, + 0.012384198606014252, + 0.013077479787170887, + 0.007111081853508949, + 0.0070900325663387775, + 0.0075786300003528595, + 0.006115879863500595, + 0.007078749593347311, + 0.008194840513169765, + 0.006947182584553957, + 0.008791315369307995, + 0.006453228648751974, + 0.006443867459893227, + 0.010156354866921902, + 0.009117737412452698, + 0.006197660230100155 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 2, + 6 + ], + "selected_reward": -0.6933183670043945 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a Snowy Owl perched on a snow-covered branch?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the video rendered in stark black-and-white with high-contrast lighting?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the owl's head dash rapidly toward an unseen sound off-camera?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the feathers on the owl's neck and head shown rippling outward in fluid waves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot a static medium close-up?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.020587678998708725 + ], + "opposite_logprob_tokens": [ + -3.895587682723999 + ], + "logprob_reward": -0.020587678998708725, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.020587678998708725, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0835767388343811, + 0.17808480560779572, + 0.17368829250335693, + 0.05852344632148743, + 0.11086586117744446, + 0.10797517001628876, + 0.09920484572649002, + 0.06400103867053986, + 0.07752469927072525, + 0.04655502736568451 + ], + "credit_map": [ + 0.012089462019503117, + 0.009579538367688656, + 0.00901826936751604, + 0.007993949577212334, + 0.011246728710830212, + 0.011549235321581364, + 0.012437929399311543, + 0.009661627933382988, + 0.03190235421061516, + 0.017880337312817574, + 0.02429468370974064, + 0.017219165340065956, + 0.022220443934202194, + 0.018319465219974518, + 0.020068569108843803, + 0.02617979422211647, + 0.011798303574323654, + 0.025639640167355537, + 0.015117225237190723, + 0.018593763932585716, + 0.01683441735804081, + 0.01797833852469921, + 0.04314787685871124, + 0.024578731507062912, + 0.006710194982588291, + 0.005462287459522486, + 0.004923379048705101, + 0.00625656871125102, + 0.007436740677803755, + 0.007976476103067398, + 0.006367330905050039, + 0.013390470296144485, + 0.03216696158051491, + 0.005658529233187437, + 0.004713043570518494, + 0.00590165052562952, + 0.006656410172581673, + 0.038845110684633255, + 0.006617680657655001, + 0.01030647847801447, + 0.012593477964401245, + 0.010830212384462357, + 0.010691860690712929, + 0.008848244324326515, + 0.0179886594414711, + 0.01982215978205204, + 0.007108809892088175, + 0.020091747865080833, + 0.01580391451716423, + 0.009914420545101166, + 0.005644319113343954, + 0.014260511845350266, + 0.0105151841416955, + 0.023303529247641563, + 0.009800534695386887, + 0.009962433949112892, + 0.009277161210775375, + 0.004468290600925684, + 0.010974018834531307, + 0.009037486277520657, + 0.008887697011232376, + 0.008472139947116375, + 0.0078632403165102, + 0.005021004006266594, + 0.008053911849856377, + 0.015761543065309525, + 0.008612271398305893, + 0.009839082136750221, + 0.009628436528146267, + 0.008943340741097927, + 0.008225452154874802, + 0.008460664190351963, + 0.009622369892895222, + 0.004039438907057047, + 0.004861125722527504, + 0.002425526501610875, + 0.003593183122575283, + 0.011703568510711193, + 0.004994965624064207, + 0.005314851179718971 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 2, + 4 + ], + "selected_reward": -0.020587678998708725 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a Snowy Owl perched on a snow-covered branch?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the video rendered in stark black-and-white with high-contrast lighting?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the owl's head dash rapidly toward an unseen sound off-camera?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the feathers on the owl's neck and head shown rippling outward in fluid waves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot a static medium close-up?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.25209155678749084 + ], + "opposite_logprob_tokens": [ + -1.5020915269851685 + ], + "logprob_reward": -0.25209155678749084, + "yes_no_margin_reward": 1.25, + "token_credit_reward": -0.25209155678749084, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07653617113828659, + 0.07340922206640244, + 0.17016027867794037, + 0.09340467303991318, + 0.17982706427574158, + 0.1672038584947586, + 0.08444760739803314, + 0.038978461176157, + 0.07980912923812866, + 0.03622356057167053 + ], + "credit_map": [ + 0.01027328334748745, + 0.00785531010478735, + 0.009141583926975727, + 0.0073613435961306095, + 0.011627670377492905, + 0.008717532269656658, + 0.013795140199363232, + 0.007764310576021671, + 0.009049588814377785, + 0.00576220266520977, + 0.010508129373192787, + 0.007196346763521433, + 0.007552687078714371, + 0.009562669321894646, + 0.007518692873418331, + 0.016258902847766876, + 0.011005975306034088, + 0.03352449834346771, + 0.007436677813529968, + 0.01731446385383606, + 0.014327166602015495, + 0.012826488353312016, + 0.04976911097764969, + 0.02395588532090187, + 0.017789656296372414, + 0.010021605528891087, + 0.006518605165183544, + 0.008156558498740196, + 0.011767124757170677, + 0.012270522303879261, + 0.015517300926148891, + 0.011363293044269085, + 0.05759834498167038, + 0.008312161080539227, + 0.006169713567942381, + 0.0076780784875154495, + 0.009657918475568295, + 0.06498096883296967, + 0.00760530075058341, + 0.017824571579694748, + 0.019440965726971626, + 0.017524858936667442, + 0.019017480313777924, + 0.014063617214560509, + 0.029736999422311783, + 0.027876131236553192, + 0.013533460907638073, + 0.026010338217020035, + 0.01617741584777832, + 0.007665992248803377, + 0.010609005577862263, + 0.013843613676726818, + 0.007831709459424019, + 0.01249734777957201, + 0.00694240303710103, + 0.008880111388862133, + 0.005491034127771854, + 0.0032311640679836273, + 0.008687249384820461, + 0.0053697554394602776, + 0.005344371311366558, + 0.0039492035284638405, + 0.0033608770463615656, + 0.0035448053386062384, + 0.014392613433301449, + 0.014822186902165413, + 0.014072694815695286, + 0.008159460499882698, + 0.007436799816787243, + 0.007850222289562225, + 0.007437734864652157, + 0.005637417081743479, + 0.00486096041277051, + 0.0046302382834255695, + 0.003298513125628233, + 0.003893621964380145, + 0.0043496908619999886, + 0.004679253790527582, + 0.006666073575615883, + 0.003845210885629058 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 2, + 5 + ], + "selected_reward": -0.25209155678749084 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "No", + "Yes", + "Yes" + ] + }, + { + "prompt": "static medium close-up, rendered in stark black-and-white ultra-high-speed photography, captures a majestic Snowy Owl in a moment of heightened alert. Perched on a snow-covered branch against a dark, out-of-focus background, the owl is a study in momentum. In an instant, its head dashes with incredible speed toward an unseen sound just off-camera. The entire sequence unfolds in extreme slow motion, transforming the violent action into an elegant balletic pirouette. The feathers on its neck and head ripple outward in mesmerizingly fluid waves, each barb and feather rendered in crystalline clarity. Its large, black, focused eyes gleam with a sharp, crystalline light as they lock onto a new target. As its head settles, a subtle, involuntary twitch energizes its powerful black talons, tightening their grip on the branch. The scene's high-contrast lighting creates a dramatic black-and-white contrast, casting deep shadows and accentuating the impeccable texture of the feathers, ultimately creating a powerful atmosphere of raw, focused instinct and sublime natural beauty.", + "seed": 29348, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000050/group_3/index_569_seed_29348_dp_6_rank_12_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 3, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.30405974239110944, + "vlm_logprob": -0.30405974239110944, + "vlm_yes_no_margin": 2.275000011920929, + "vlm_token_credit": -0.30405974239110944, + "vlm_yes_ratio": 0.8, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a Snowy Owl perched on a snow-covered branch?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the video rendered in stark black-and-white with high-contrast lighting?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the owl's head dash rapidly toward an unseen sound off-camera?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the feathers on the owl's neck and head shown rippling outward in fluid waves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot a static medium close-up?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.024701427668333054 + ], + "opposite_logprob_tokens": [ + -3.8997013568878174 + ], + "logprob_reward": -0.024701427668333054, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.024701427668333054, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09263955056667328, + 0.13508537411689758, + 0.027004798874258995, + 0.12699347734451294, + 0.21146856248378754, + 0.13389870524406433, + 0.07380608469247818, + 0.10002141445875168, + 0.07299762964248657, + 0.026084432378411293 + ], + "credit_map": [ + 0.007998982444405556, + 0.019421527162194252, + 0.01061361376196146, + 0.011855879798531532, + 0.010127981193363667, + 0.017053013667464256, + 0.0070678358897566795, + 0.008500713855028152, + 0.013023450039327145, + 0.023437904193997383, + 0.009599528275430202, + 0.018464503809809685, + 0.01958131045103073, + 0.01532598864287138, + 0.016122886911034584, + 0.019529804587364197, + 0.0032197325490415096, + 0.0033093367237597704, + 0.0024808773305267096, + 0.0034431719686836004, + 0.0036100964061915874, + 0.0034499492030590773, + 0.004481185227632523, + 0.003010446671396494, + 0.009890520013868809, + 0.04804025590419769, + 0.003758726641535759, + 0.012506342492997646, + 0.008535771630704403, + 0.0069894855841994286, + 0.005394797306507826, + 0.03187757357954979, + 0.015828033909201622, + 0.030583880841732025, + 0.020873259752988815, + 0.026255453005433083, + 0.021518217399716377, + 0.027966992929577827, + 0.0427665114402771, + 0.02567620761692524, + 0.020031461492180824, + 0.014383381232619286, + 0.017647933214902878, + 0.02245294488966465, + 0.020598504692316055, + 0.013151646591722965, + 0.01698610931634903, + 0.008646720089018345, + 0.00930496584624052, + 0.013552173972129822, + 0.0045691546984016895, + 0.00777639402076602, + 0.013915302231907845, + 0.010378616861999035, + 0.006958324927836657, + 0.0073511479422450066, + 0.019356392323970795, + 0.006312600336968899, + 0.007070691324770451, + 0.01406763680279255, + 0.006415576674044132, + 0.019285982474684715, + 0.012655982747673988, + 0.014856538735330105, + 0.01759236305952072, + 0.005089593585580587, + 0.004620764870196581, + 0.0068485853262245655, + 0.009930845350027084, + 0.007198219187557697, + 0.00943526066839695, + 0.01228199154138565, + 0.005690399091690779, + 0.0019559310749173164, + 0.0031623775139451027, + 0.002786155790090561, + 0.0020618492271751165, + 0.0020683277398347855, + 0.0034790418576449156, + 0.004880349151790142 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 1, + 5 + ], + "selected_reward": -0.024701427668333054 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a Snowy Owl perched on a snow-covered branch?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the video rendered in stark black-and-white with high-contrast lighting?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the owl's head dash rapidly toward an unseen sound off-camera?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the feathers on the owl's neck and head shown rippling outward in fluid waves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot a static medium close-up?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.055017709732055664 + ], + "opposite_logprob_tokens": [ + -2.9300177097320557 + ], + "logprob_reward": -0.055017709732055664, + "yes_no_margin_reward": 2.875, + "token_credit_reward": -0.055017709732055664, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09399276971817017, + 0.11779895424842834, + 0.032120127230882645, + 0.20922458171844482, + 0.056937698274850845, + 0.036304064095020294, + 0.09016355127096176, + 0.07163722068071365, + 0.14950792491436005, + 0.1423131823539734 + ], + "credit_map": [ + 0.009100276976823807, + 0.012795986607670784, + 0.014856782741844654, + 0.011835619807243347, + 0.008149081841111183, + 0.0209631510078907, + 0.007639953400939703, + 0.008651914075016975, + 0.011890405789017677, + 0.02115953527390957, + 0.009532970376312733, + 0.01859671249985695, + 0.013861468061804771, + 0.011707492172718048, + 0.011838086880743504, + 0.01921229250729084, + 0.002994601847603917, + 0.005630444269627333, + 0.002472242806106806, + 0.004089579451829195, + 0.004156704992055893, + 0.004819107707589865, + 0.004036833066493273, + 0.003920617047697306, + 0.016830023378133774, + 0.07964060455560684, + 0.006577790714800358, + 0.020560551434755325, + 0.013490266166627407, + 0.01099618710577488, + 0.00882486067712307, + 0.05230429023504257, + 0.003897417103871703, + 0.005415163468569517, + 0.00499768927693367, + 0.006817337591201067, + 0.006659793667495251, + 0.008450699970126152, + 0.014702814631164074, + 0.005996786057949066, + 0.004625501576811075, + 0.004540283232927322, + 0.006659358274191618, + 0.005045007914304733, + 0.004794489126652479, + 0.0030329369474202394, + 0.004671827424317598, + 0.002934661228209734, + 0.009556802920997143, + 0.012909868732094765, + 0.007373050786554813, + 0.010671333409845829, + 0.017877355217933655, + 0.014583446085453033, + 0.007833489216864109, + 0.009358208626508713, + 0.014713936485350132, + 0.004421011079102755, + 0.006892944220453501, + 0.009329404681921005, + 0.005809467285871506, + 0.011393814347684383, + 0.00733873900026083, + 0.011737904511392117, + 0.04219001159071922, + 0.008185908198356628, + 0.007330336142331362, + 0.012264363467693329, + 0.013584846630692482, + 0.013264061883091927, + 0.011542093008756638, + 0.04114629700779915, + 0.03704492002725601, + 0.0101028336212039, + 0.017443513497710228, + 0.012912674807012081, + 0.00979599915444851, + 0.01262286864221096, + 0.012211582623422146, + 0.030178779736161232 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 8, + 9 + ], + "selected_reward": -0.055017709732055664 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a Snowy Owl perched on a snow-covered branch?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the video rendered in stark black-and-white with high-contrast lighting?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the owl's head dash rapidly toward an unseen sound off-camera?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the feathers on the owl's neck and head shown rippling outward in fluid waves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot a static medium close-up?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -1.3134007453918457 + ], + "opposite_logprob_tokens": [ + -0.3134008049964905 + ], + "logprob_reward": -1.3134007453918457, + "yes_no_margin_reward": -0.9999999403953552, + "token_credit_reward": -1.3134007453918457, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.09102491289377213, + 0.05858614668250084, + 0.024750560522079468, + 0.3944430947303772, + 0.04594884440302849, + 0.034922316670417786, + 0.16443483531475067, + 0.08855108916759491, + 0.06414317339658737, + 0.03319501504302025 + ], + "credit_map": [ + 0.010855949483811855, + 0.011864453554153442, + 0.008407436311244965, + 0.011878578923642635, + 0.011088788509368896, + 0.01892571710050106, + 0.008906327188014984, + 0.009097667410969734, + 0.007229754235595465, + 0.011108843609690666, + 0.004237690940499306, + 0.009655894711613655, + 0.006728032603859901, + 0.005054536275565624, + 0.005590156186372042, + 0.008981237187981606, + 0.0023109219036996365, + 0.004181519616395235, + 0.0017479336820542812, + 0.0033766222186386585, + 0.003158244304358959, + 0.004831469152122736, + 0.0030613921117037535, + 0.002082459395751357, + 0.03246304765343666, + 0.14896650612354279, + 0.01325278077274561, + 0.038762714713811874, + 0.028234902769327164, + 0.023242920637130737, + 0.016414372250437737, + 0.09310586750507355, + 0.0041530984453856945, + 0.005559314973652363, + 0.0066690873354673386, + 0.007462713867425919, + 0.00506630027666688, + 0.006179848685860634, + 0.006378702819347382, + 0.004479778930544853, + 0.0042424979619681835, + 0.003978174179792404, + 0.005707809701561928, + 0.004716190043836832, + 0.004144683480262756, + 0.003675821702927351, + 0.0059531619772315025, + 0.0025039752945303917, + 0.022698312997817993, + 0.021632518619298935, + 0.017678141593933105, + 0.022827772423624992, + 0.018670041114091873, + 0.022501910105347633, + 0.01885092817246914, + 0.019575225189328194, + 0.021289698779582977, + 0.005876404233276844, + 0.005518815014511347, + 0.008985254913568497, + 0.00669056223705411, + 0.01732034794986248, + 0.009520646184682846, + 0.013349360786378384, + 0.009520720690488815, + 0.00640722643584013, + 0.005377913359552622, + 0.006806818302720785, + 0.012286948040127754, + 0.006921609863638878, + 0.008049794472754002, + 0.008772142231464386, + 0.005071921739727259, + 0.003985830582678318, + 0.003033016575500369, + 0.0031052231788635254, + 0.004013645462691784, + 0.0029978652019053698, + 0.0041825552470982075, + 0.006804958451539278 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 6, + 0 + ], + "selected_reward": -1.3134007453918457 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a Snowy Owl perched on a snow-covered branch?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the video rendered in stark black-and-white with high-contrast lighting?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the owl's head dash rapidly toward an unseen sound off-camera?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the feathers on the owl's neck and head shown rippling outward in fluid waves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot a static medium close-up?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03812311962246895 + ], + "opposite_logprob_tokens": [ + -3.28812313079834 + ], + "logprob_reward": -0.03812311962246895, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.03812311962246895, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09526050090789795, + 0.07130153477191925, + 0.048151277005672455, + 0.41661638021469116, + 0.04223693907260895, + 0.0714576318860054, + 0.10317033529281616, + 0.08064541220664978, + 0.04394679516553879, + 0.027213120833039284 + ], + "credit_map": [ + 0.013538935221731663, + 0.012527840211987495, + 0.008776087313890457, + 0.010943134315311909, + 0.013005275279283524, + 0.017595943063497543, + 0.009698893874883652, + 0.009174403734505177, + 0.007439911365509033, + 0.01345878280699253, + 0.004417553078383207, + 0.013927008956670761, + 0.008286669850349426, + 0.00727434828877449, + 0.00586656155064702, + 0.010630709119141102, + 0.00420552771538496, + 0.007841653190553188, + 0.002708324696868658, + 0.007281211670488119, + 0.00723104877397418, + 0.005621695425361395, + 0.008657410740852356, + 0.00460441131144762, + 0.033998094499111176, + 0.15649136900901794, + 0.013551915995776653, + 0.0406508594751358, + 0.02803412452340126, + 0.02811576798558235, + 0.01757670007646084, + 0.09819760173559189, + 0.003052645828574896, + 0.004805317148566246, + 0.0031692723277956247, + 0.006110192276537418, + 0.00831516832113266, + 0.004497454036027193, + 0.007058556191623211, + 0.005228335037827492, + 0.009956366382539272, + 0.007032248191535473, + 0.009478883817791939, + 0.013637850061058998, + 0.010380675084888935, + 0.0072409347631037235, + 0.008293899707496166, + 0.005436782259494066, + 0.009289045818150043, + 0.010908003896474838, + 0.00950212124735117, + 0.012576497159898281, + 0.013812352903187275, + 0.022630887106060982, + 0.010061304084956646, + 0.014390131458640099, + 0.014740931801497936, + 0.005553239490836859, + 0.005343459080904722, + 0.01209314540028572, + 0.0066038682125508785, + 0.01365253608673811, + 0.011497297324240208, + 0.011160941794514656, + 0.010026431642472744, + 0.0028326448518782854, + 0.0025679252576082945, + 0.004119376186281443, + 0.005505680106580257, + 0.00437142513692379, + 0.005433367099612951, + 0.00908995233476162, + 0.004879554733633995, + 0.002250364748761058, + 0.003311196807771921, + 0.0029120692051947117, + 0.0034106106031686068, + 0.0025774138048291206, + 0.0029572395142167807, + 0.0049146744422614574 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 6, + 0 + ], + "selected_reward": -0.03812311962246895 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a Snowy Owl perched on a snow-covered branch?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the video rendered in stark black-and-white with high-contrast lighting?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the owl's head dash rapidly toward an unseen sound off-camera?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the feathers on the owl's neck and head shown rippling outward in fluid waves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot a static medium close-up?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.08905570954084396 + ], + "opposite_logprob_tokens": [ + -2.4640557765960693 + ], + "logprob_reward": -0.08905570954084396, + "yes_no_margin_reward": 2.375, + "token_credit_reward": -0.08905570954084396, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0918436124920845, + 0.06106342375278473, + 0.0880395695567131, + 0.1329609751701355, + 0.12339659780263901, + 0.1297144740819931, + 0.21225705742835999, + 0.08400191366672516, + 0.05992717668414116, + 0.016795245930552483 + ], + "credit_map": [ + 0.013482706621289253, + 0.018888412043452263, + 0.009439007379114628, + 0.010846545919775963, + 0.009842288680374622, + 0.014344277791678905, + 0.007593853399157524, + 0.007406510412693024, + 0.00552713917568326, + 0.010931042023003101, + 0.004171127453446388, + 0.008822799660265446, + 0.006699549499899149, + 0.009389353916049004, + 0.006078697275370359, + 0.009443707764148712, + 0.011117557995021343, + 0.015301900915801525, + 0.006545105017721653, + 0.010167904198169708, + 0.01403459906578064, + 0.008699500933289528, + 0.011127851903438568, + 0.01104514580219984, + 0.010801511816680431, + 0.05126753821969032, + 0.0038488253485411406, + 0.01321305800229311, + 0.009294740855693817, + 0.006550601217895746, + 0.005070563405752182, + 0.03291412815451622, + 0.010024617426097393, + 0.013492682948708534, + 0.014771572314202785, + 0.02373075671494007, + 0.021706998348236084, + 0.01336422935128212, + 0.014244262129068375, + 0.012061472982168198, + 0.015690509229898453, + 0.010677560232579708, + 0.013303976505994797, + 0.027392327785491943, + 0.015460572205483913, + 0.012968909926712513, + 0.026012446731328964, + 0.008208162151277065, + 0.021223949268460274, + 0.02982289344072342, + 0.019209740683436394, + 0.025843339040875435, + 0.04430932551622391, + 0.033194009214639664, + 0.017944155260920525, + 0.020709620788693428, + 0.018779538571834564, + 0.006135633215308189, + 0.005636251997202635, + 0.007734269835054874, + 0.006693414878100157, + 0.015335530042648315, + 0.011381475254893303, + 0.01230578776448965, + 0.0156254880130291, + 0.00274799857288599, + 0.003337486181408167, + 0.005825157277286053, + 0.00746574392542243, + 0.004959434736520052, + 0.006693030707538128, + 0.013272834941744804, + 0.002812577411532402, + 0.0016134865581989288, + 0.0021403527352958918, + 0.0022266258019953966, + 0.0015248940326273441, + 0.0010547793935984373, + 0.0017152552027255297, + 0.003707274328917265 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 3, + 5 + ], + "selected_reward": -0.08905570954084396 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "No", + "Yes", + "Yes" + ] + }, + { + "prompt": "static medium close-up, rendered in stark black-and-white ultra-high-speed photography, captures a majestic Snowy Owl in a moment of heightened alert. Perched on a snow-covered branch against a dark, out-of-focus background, the owl is a study in momentum. In an instant, its head dashes with incredible speed toward an unseen sound just off-camera. The entire sequence unfolds in extreme slow motion, transforming the violent action into an elegant balletic pirouette. The feathers on its neck and head ripple outward in mesmerizingly fluid waves, each barb and feather rendered in crystalline clarity. Its large, black, focused eyes gleam with a sharp, crystalline light as they lock onto a new target. As its head settles, a subtle, involuntary twitch energizes its powerful black talons, tightening their grip on the branch. The scene's high-contrast lighting creates a dramatic black-and-white contrast, casting deep shadows and accentuating the impeccable texture of the feathers, ultimately creating a powerful atmosphere of raw, focused instinct and sublime natural beauty.", + "seed": 29349, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000050/group_3/index_569_seed_29349_dp_7_rank_14_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 3, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.2441688824445009, + "vlm_logprob": -0.2441688824445009, + "vlm_yes_no_margin": 2.2999999582767487, + "vlm_token_credit": -0.2441688824445009, + "vlm_yes_ratio": 0.8, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a Snowy Owl perched on a snow-covered branch?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the video rendered in stark black-and-white with high-contrast lighting?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the owl's head dash rapidly toward an unseen sound off-camera?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the feathers on the owl's neck and head shown rippling outward in fluid waves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot a static medium close-up?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.023474866524338722 + ], + "opposite_logprob_tokens": [ + -4.02347469329834 + ], + "logprob_reward": -0.023474866524338722, + "yes_no_margin_reward": 3.999999761581421, + "token_credit_reward": -0.023474866524338722, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13290347158908844, + 0.06653627753257751, + 0.02122403308749199, + 0.03772447258234024, + 0.49836716055870056, + 0.031549710780382156, + 0.036447156220674515, + 0.029265737161040306, + 0.02912842109799385, + 0.11685358732938766 + ], + "credit_map": [ + 0.009980914182960987, + 0.03650746867060661, + 0.006075062323361635, + 0.05330723524093628, + 0.0058627864345908165, + 0.007954593747854233, + 0.009706968441605568, + 0.003508431138470769, + 0.005170171614736319, + 0.007852931506931782, + 0.009032574482262135, + 0.00799353513866663, + 0.011291394010186195, + 0.008801981806755066, + 0.007945633493363857, + 0.008448049426078796, + 0.0019074127776548266, + 0.0019309744238853455, + 0.0015631109708920121, + 0.0024357892107218504, + 0.003561283927410841, + 0.002644215477630496, + 0.004223975352942944, + 0.002957272110506892, + 0.007090555038303137, + 0.004706219304352999, + 0.004625032190233469, + 0.004784673452377319, + 0.003828581189736724, + 0.0038352778647094965, + 0.005353804212063551, + 0.0035003311932086945, + 0.09500916302204132, + 0.2085842490196228, + 0.06322980672121048, + 0.02417517639696598, + 0.04208474978804588, + 0.020979896187782288, + 0.020636463537812233, + 0.023667633533477783, + 0.010417402721941471, + 0.0014978391118347645, + 0.0011564581654965878, + 0.0016775926342234015, + 0.0017091952031478286, + 0.01185790915042162, + 0.001436182064935565, + 0.0017971275374293327, + 0.0027178244199603796, + 0.0030270982533693314, + 0.003294349880889058, + 0.009221196174621582, + 0.0032569426111876965, + 0.007708793971687555, + 0.0037995760794728994, + 0.003421373199671507, + 0.0028880867175757885, + 0.0017005440313369036, + 0.003469157265499234, + 0.005110390018671751, + 0.00552511727437377, + 0.003852426540106535, + 0.003766090841963887, + 0.0029539226088672876, + 0.003469692077487707, + 0.002326538786292076, + 0.003474220633506775, + 0.003049924736842513, + 0.002983147045597434, + 0.003293008077889681, + 0.004202038981020451, + 0.006329849362373352, + 0.0261144507676363, + 0.007298727985471487, + 0.007413253653794527, + 0.00970897264778614, + 0.011076690629124641, + 0.029216343536973, + 0.009838360361754894, + 0.016186773777008057 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 0, + 9 + ], + "selected_reward": -0.023474866524338722 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a Snowy Owl perched on a snow-covered branch?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the video rendered in stark black-and-white with high-contrast lighting?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the owl's head dash rapidly toward an unseen sound off-camera?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the feathers on the owl's neck and head shown rippling outward in fluid waves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot a static medium close-up?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.08908875286579132 + ], + "opposite_logprob_tokens": [ + -2.4640886783599854 + ], + "logprob_reward": -0.08908875286579132, + "yes_no_margin_reward": 2.375, + "token_credit_reward": -0.08908875286579132, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.28945010900497437, + 0.06587772816419601, + 0.06266111880540848, + 0.026827311143279076, + 0.08046823740005493, + 0.06279056519269943, + 0.07254473865032196, + 0.062383607029914856, + 0.0529898926615715, + 0.22400672733783722 + ], + "credit_map": [ + 0.029659023508429527, + 0.06919766217470169, + 0.01689462922513485, + 0.09769759327173233, + 0.018813496455550194, + 0.01827852614223957, + 0.022562338039278984, + 0.01634684018790722, + 0.004984667524695396, + 0.008874453604221344, + 0.0067449891939759254, + 0.00630794744938612, + 0.010865681804716587, + 0.007548550609499216, + 0.010836465284228325, + 0.009714972227811813, + 0.006092668045312166, + 0.005550717934966087, + 0.005487961694598198, + 0.007586054969578981, + 0.010616415180265903, + 0.008029736578464508, + 0.01171258557587862, + 0.007584970910102129, + 0.0030051893554627895, + 0.003275536000728607, + 0.002781286835670471, + 0.003942640498280525, + 0.0036393438931554556, + 0.0028933135326951742, + 0.005198597442358732, + 0.002091403817757964, + 0.0076966979540884495, + 0.016593193635344505, + 0.01019156165421009, + 0.008738008327782154, + 0.008911419659852982, + 0.007278543431311846, + 0.012524684891104698, + 0.008534124121069908, + 0.021632887423038483, + 0.0033165172208100557, + 0.002250657184049487, + 0.0021149413660168648, + 0.0031228933949023485, + 0.025178907439112663, + 0.002618066966533661, + 0.0025556902401149273, + 0.006119374185800552, + 0.006786220706999302, + 0.006229287013411522, + 0.01714656502008438, + 0.013845948502421379, + 0.010993072763085365, + 0.007266782224178314, + 0.004157484974712133, + 0.006508318707346916, + 0.005535823758691549, + 0.006619599647819996, + 0.014621790498495102, + 0.0109592005610466, + 0.008139208890497684, + 0.0059199342504143715, + 0.004079732112586498, + 0.006727003958076239, + 0.005360995419323444, + 0.004842541646212339, + 0.00587780587375164, + 0.005968401674181223, + 0.006255250424146652, + 0.008275664411485195, + 0.009682230651378632, + 0.04709865525364876, + 0.013935871422290802, + 0.016151541844010353, + 0.023514429107308388, + 0.021098289638757706, + 0.04930262267589569, + 0.019616207107901573, + 0.0332891047000885 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 4 + ], + "selected_reward": -0.08908875286579132 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a Snowy Owl perched on a snow-covered branch?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the video rendered in stark black-and-white with high-contrast lighting?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the owl's head dash rapidly toward an unseen sound off-camera?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the feathers on the owl's neck and head shown rippling outward in fluid waves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot a static medium close-up?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -0.9742437601089478 + ], + "opposite_logprob_tokens": [ + -0.47424378991127014 + ], + "logprob_reward": -0.9742437601089478, + "yes_no_margin_reward": -0.4999999701976776, + "token_credit_reward": -0.9742437601089478, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.0736357569694519, + 0.04177884757518768, + 0.09226222336292267, + 0.0352013036608696, + 0.391547828912735, + 0.0894770622253418, + 0.061353374272584915, + 0.06110304966568947, + 0.10941001772880554, + 0.04423054680228233 + ], + "credit_map": [ + 0.01479994598776102, + 0.012404746375977993, + 0.00706266425549984, + 0.007311352528631687, + 0.008193044923245907, + 0.007109340745955706, + 0.010862806811928749, + 0.005891859065741301, + 0.0033338158391416073, + 0.004923775792121887, + 0.003872982691973448, + 0.006169851869344711, + 0.005592154338955879, + 0.00547929760068655, + 0.007549373898655176, + 0.004857594612985849, + 0.008054378442466259, + 0.00682146567851305, + 0.007131158374249935, + 0.01284466590732336, + 0.017607277259230614, + 0.015831656754016876, + 0.013902337290346622, + 0.010069279931485653, + 0.005319233518093824, + 0.0038270228542387486, + 0.0036009333562105894, + 0.004652905743569136, + 0.0038035197649151087, + 0.004624586086720228, + 0.0056304046884179115, + 0.003742696950212121, + 0.0756552666425705, + 0.1624729037284851, + 0.04663720726966858, + 0.020483387634158134, + 0.03296433761715889, + 0.017102904617786407, + 0.016911286860704422, + 0.019320523366332054, + 0.028569532558321953, + 0.006008361000567675, + 0.00360507913865149, + 0.004132356494665146, + 0.005294430535286665, + 0.03446920961141586, + 0.003264179453253746, + 0.004133915528655052, + 0.0049484726041555405, + 0.005781734827905893, + 0.00612157117575407, + 0.007495040539652109, + 0.011164125055074692, + 0.010783829726278782, + 0.009088488295674324, + 0.005970115307718515, + 0.006190595682710409, + 0.00481893215328455, + 0.005033926572650671, + 0.00948327500373125, + 0.0154439527541399, + 0.008111737668514252, + 0.006366484332829714, + 0.0056541478261351585, + 0.012168648652732372, + 0.01111126970499754, + 0.007958072237670422, + 0.009371945634484291, + 0.0334816500544548, + 0.013020019046962261, + 0.0134879844263196, + 0.00881043542176485, + 0.0059648617170751095, + 0.0031722700223326683, + 0.0035095291677862406, + 0.005211181007325649, + 0.007669082842767239, + 0.007185264490544796, + 0.005641151685267687, + 0.005877207033336163 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 8, + 2 + ], + "selected_reward": -0.9742437601089478 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a Snowy Owl perched on a snow-covered branch?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the video rendered in stark black-and-white with high-contrast lighting?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the owl's head dash rapidly toward an unseen sound off-camera?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the feathers on the owl's neck and head shown rippling outward in fluid waves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot a static medium close-up?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0337153896689415 + ], + "opposite_logprob_tokens": [ + -3.4087154865264893 + ], + "logprob_reward": -0.0337153896689415, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.0337153896689415, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.16325396299362183, + 0.03311768174171448, + 0.03969825804233551, + 0.020627234131097794, + 0.37230509519577026, + 0.22139695286750793, + 0.03846534341573715, + 0.03779076039791107, + 0.04360344633460045, + 0.029741207137703896 + ], + "credit_map": [ + 0.011718560010194778, + 0.04796792194247246, + 0.008114413358271122, + 0.06376861035823822, + 0.007755635306239128, + 0.0064687225967645645, + 0.012536638416349888, + 0.00492346566170454, + 0.004249109886586666, + 0.005197273567318916, + 0.0036556760314852, + 0.004231063649058342, + 0.004012063145637512, + 0.003892091568559408, + 0.004369140602648258, + 0.0035112665500491858, + 0.0030992981046438217, + 0.002914576791226864, + 0.0024808940943330526, + 0.004212353844195604, + 0.010109084658324718, + 0.006434119306504726, + 0.006042037159204483, + 0.004405903164297342, + 0.0030678233597427607, + 0.0021622066851705313, + 0.0020273879636079073, + 0.002740756142884493, + 0.00263571715913713, + 0.0027785981073975563, + 0.0030584551859647036, + 0.0021562916226685047, + 0.07360953092575073, + 0.15171390771865845, + 0.04490013048052788, + 0.02007184736430645, + 0.02952946349978447, + 0.016572380438447, + 0.017031637951731682, + 0.018876250833272934, + 0.07631268352270126, + 0.009920486249029636, + 0.012898958288133144, + 0.008416746743023396, + 0.009726985357701778, + 0.08630763739347458, + 0.00868981797248125, + 0.009123662486672401, + 0.004312853328883648, + 0.003015384543687105, + 0.003937366418540478, + 0.004459122661501169, + 0.00690309377387166, + 0.006966668646782637, + 0.004637766629457474, + 0.004233093466609716, + 0.004517885856330395, + 0.003105876734480262, + 0.00405771704390645, + 0.0069654579274356365, + 0.005803372710943222, + 0.005506525281816721, + 0.005479987245053053, + 0.002353937830775976, + 0.005297508090734482, + 0.004201879259198904, + 0.003994858358055353, + 0.004332599695771933, + 0.010977017693221569, + 0.005007825326174498, + 0.005034210626035929, + 0.004757551942020655, + 0.006021152250468731, + 0.0020652429666370153, + 0.0022286472376435995, + 0.0031089794356375933, + 0.0030658834148198366, + 0.006465440150350332, + 0.0037730501499027014, + 0.0030128133948892355 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 5, + 0 + ], + "selected_reward": -0.0337153896689415 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a Snowy Owl perched on a snow-covered branch?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the video rendered in stark black-and-white with high-contrast lighting?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the owl's head dash rapidly toward an unseen sound off-camera?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the feathers on the owl's neck and head shown rippling outward in fluid waves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot a static medium close-up?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.10032164305448532 + ], + "opposite_logprob_tokens": [ + -2.3503215312957764 + ], + "logprob_reward": -0.10032164305448532, + "yes_no_margin_reward": 2.25, + "token_credit_reward": -0.10032164305448532, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.16869740188121796, + 0.08968614041805267, + 0.05210151895880699, + 0.043619539588689804, + 0.36987048387527466, + 0.022212328389286995, + 0.04299774765968323, + 0.10579552501440048, + 0.058581892400979996, + 0.046437427401542664 + ], + "credit_map": [ + 0.013912639580667019, + 0.04270024225115776, + 0.010990072041749954, + 0.056195542216300964, + 0.011902018450200558, + 0.007444275543093681, + 0.016308888792991638, + 0.009243727661669254, + 0.010612419806420803, + 0.011319574899971485, + 0.011509178206324577, + 0.01110417302697897, + 0.01296462956815958, + 0.010663156397640705, + 0.009506972506642342, + 0.012006038799881935, + 0.006858091335743666, + 0.005981034599244595, + 0.004116542171686888, + 0.005096679553389549, + 0.00854196585714817, + 0.0065701683051884174, + 0.008377901278436184, + 0.006559138186275959, + 0.00812288373708725, + 0.0047194682992994785, + 0.005702702794224024, + 0.004089136142283678, + 0.005240997765213251, + 0.005419492721557617, + 0.006334914360195398, + 0.003989941440522671, + 0.0694156140089035, + 0.15407204627990723, + 0.04860825091600418, + 0.018267914652824402, + 0.03051643632352352, + 0.015453644096851349, + 0.01657356508076191, + 0.016963016241788864, + 0.004587011877447367, + 0.001772449817508459, + 0.0012473746901378036, + 0.002319372957572341, + 0.003594158450141549, + 0.005241977050900459, + 0.002173297107219696, + 0.0012766866711899638, + 0.0034926573280245066, + 0.004506194498389959, + 0.003285387298092246, + 0.005706552416086197, + 0.0100827906280756, + 0.00701958779245615, + 0.005290503613650799, + 0.003614074317738414, + 0.010841495357453823, + 0.008111958391964436, + 0.011572069488465786, + 0.01371860783547163, + 0.029595928266644478, + 0.01610500179231167, + 0.010281745344400406, + 0.005568717606365681, + 0.007936657406389713, + 0.005233601201325655, + 0.005730436183512211, + 0.0062725436873734, + 0.012078652158379555, + 0.006451525259763002, + 0.006332787219434977, + 0.008545683696866035, + 0.009847898036241531, + 0.002575306687504053, + 0.0025951270945370197, + 0.0047281947918236256, + 0.004687609151005745, + 0.011462959460914135, + 0.003921919967979193, + 0.006618409417569637 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 0, + 7 + ], + "selected_reward": -0.10032164305448532 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "No", + "Yes", + "Yes" + ] + }, + { + "prompt": "static medium close-up, rendered in stark black-and-white ultra-high-speed photography, captures a majestic Snowy Owl in a moment of heightened alert. Perched on a snow-covered branch against a dark, out-of-focus background, the owl is a study in momentum. In an instant, its head dashes with incredible speed toward an unseen sound just off-camera. The entire sequence unfolds in extreme slow motion, transforming the violent action into an elegant balletic pirouette. The feathers on its neck and head ripple outward in mesmerizingly fluid waves, each barb and feather rendered in crystalline clarity. Its large, black, focused eyes gleam with a sharp, crystalline light as they lock onto a new target. As its head settles, a subtle, involuntary twitch energizes its powerful black talons, tightening their grip on the branch. The scene's high-contrast lighting creates a dramatic black-and-white contrast, casting deep shadows and accentuating the impeccable texture of the feathers, ultimately creating a powerful atmosphere of raw, focused instinct and sublime natural beauty.", + "seed": 29350, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000050/group_3/index_569_seed_29350_dp_7_rank_14_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 3, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.24666484650224446, + "vlm_logprob": -0.24666484650224446, + "vlm_yes_no_margin": 2.8, + "vlm_token_credit": -0.24666484650224446, + "vlm_yes_ratio": 0.8, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a Snowy Owl perched on a snow-covered branch?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the video rendered in stark black-and-white with high-contrast lighting?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the owl's head dash rapidly toward an unseen sound off-camera?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the feathers on the owl's neck and head shown rippling outward in fluid waves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot a static medium close-up?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.008467604406177998 + ], + "opposite_logprob_tokens": [ + -5.008467674255371 + ], + "logprob_reward": -0.008467604406177998, + "yes_no_margin_reward": 5.0, + "token_credit_reward": -0.008467604406177998, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0638994500041008, + 0.05270018428564072, + 0.13375285267829895, + 0.050395116209983826, + 0.05024195834994316, + 0.20066680014133453, + 0.19255471229553223, + 0.0849965363740921, + 0.07426836341619492, + 0.09652407467365265 + ], + "credit_map": [ + 0.006137198768556118, + 0.004779890179634094, + 0.007072444539517164, + 0.00900694727897644, + 0.00558033213019371, + 0.007964488118886948, + 0.012316873297095299, + 0.011041276156902313, + 0.010196106508374214, + 0.00563997495919466, + 0.0032372719142585993, + 0.012917268089950085, + 0.005628330633044243, + 0.004805151838809252, + 0.004436805844306946, + 0.005839274264872074, + 0.0077077895402908325, + 0.004027160350233316, + 0.007757861632853746, + 0.006014229729771614, + 0.006266760639846325, + 0.041877731680870056, + 0.007130577228963375, + 0.05297073349356651, + 0.005084661766886711, + 0.006155129987746477, + 0.008334330283105373, + 0.006376340985298157, + 0.007025320082902908, + 0.005415665451437235, + 0.005671267863363028, + 0.006332398857921362, + 0.004180567339062691, + 0.00803663581609726, + 0.00492935162037611, + 0.005274081137031317, + 0.00867458526045084, + 0.00802443828433752, + 0.004526881966739893, + 0.006595417391508818, + 0.012527436017990112, + 0.014616823755204678, + 0.01813291758298874, + 0.011521030217409134, + 0.01667271926999092, + 0.06074107065796852, + 0.016430243849754333, + 0.05002455785870552, + 0.014499643817543983, + 0.022834572941064835, + 0.012402581982314587, + 0.015296664088964462, + 0.03541414067149162, + 0.02254767343401909, + 0.022772764787077904, + 0.04678667336702347, + 0.007349596358835697, + 0.012333108112215996, + 0.009777485392987728, + 0.010966704227030277, + 0.009958266280591488, + 0.017981434240937233, + 0.007720447611063719, + 0.008909494616091251, + 0.005148052703589201, + 0.005108457058668137, + 0.008569412864744663, + 0.006075498182326555, + 0.00610741414129734, + 0.014224140904843807, + 0.01114642433822155, + 0.017888963222503662, + 0.005487058777362108, + 0.004533746745437384, + 0.004020217340439558, + 0.0039878711104393005, + 0.0051741343922913074, + 0.03038661926984787, + 0.008412289433181286, + 0.03452214226126671 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 6, + 2 + ], + "selected_reward": -0.008467604406177998 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a Snowy Owl perched on a snow-covered branch?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the video rendered in stark black-and-white with high-contrast lighting?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the owl's head dash rapidly toward an unseen sound off-camera?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the feathers on the owl's neck and head shown rippling outward in fluid waves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot a static medium close-up?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.020622361451387405 + ], + "opposite_logprob_tokens": [ + -3.8956222534179688 + ], + "logprob_reward": -0.020622361451387405, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.020622361451387405, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14016593992710114, + 0.10375645756721497, + 0.2048109620809555, + 0.04279819875955582, + 0.0376807302236557, + 0.2609042823314667, + 0.0758909210562706, + 0.06017455458641052, + 0.01624230667948723, + 0.05757565051317215 + ], + "credit_map": [ + 0.01284544263035059, + 0.007794525474309921, + 0.018990125507116318, + 0.02336949296295643, + 0.013365586288273335, + 0.014570667408406734, + 0.021662117913365364, + 0.02756798453629017, + 0.023073533549904823, + 0.009931392036378384, + 0.005549553316086531, + 0.023945091292262077, + 0.009780687279999256, + 0.008398146368563175, + 0.00900912843644619, + 0.014068927615880966, + 0.013784448616206646, + 0.007980765774846077, + 0.013856634497642517, + 0.010270221158862114, + 0.012016888707876205, + 0.055204298347234726, + 0.01980936899781227, + 0.07188832759857178, + 0.004696267191320658, + 0.005237523000687361, + 0.010410710237920284, + 0.005094974767416716, + 0.004865819122642279, + 0.0035485494881868362, + 0.004574410151690245, + 0.0043699429370462894, + 0.002641626400873065, + 0.006691365037113428, + 0.0033929243218153715, + 0.0037650815211236477, + 0.00801777932792902, + 0.004641992039978504, + 0.004175191279500723, + 0.00435476703569293, + 0.01336014922708273, + 0.010492073372006416, + 0.00669882632791996, + 0.01108214445412159, + 0.018415633589029312, + 0.09468497335910797, + 0.015349062159657478, + 0.09082142263650894, + 0.00872198585420847, + 0.01968407817184925, + 0.005121259018778801, + 0.006767055485397577, + 0.00891194585710764, + 0.00589783163741231, + 0.008222751319408417, + 0.012564017437398434, + 0.009100374765694141, + 0.0060678510926663876, + 0.00704022403806448, + 0.0067336102947592735, + 0.005917307920753956, + 0.008095795288681984, + 0.008656933903694153, + 0.008562454022467136, + 0.0014116870006546378, + 0.002137900562956929, + 0.0018173514399677515, + 0.002322954125702381, + 0.002649173839017749, + 0.0014545266749337316, + 0.0021392733324319124, + 0.0023094392381608486, + 0.0035168223548680544, + 0.003030547872185707, + 0.0031382590532302856, + 0.0031024611089378595, + 0.004144939128309488, + 0.017008796334266663, + 0.0037459167651832104, + 0.019887907430529594 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 2, + 0 + ], + "selected_reward": -0.020622361451387405 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a Snowy Owl perched on a snow-covered branch?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the video rendered in stark black-and-white with high-contrast lighting?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the owl's head dash rapidly toward an unseen sound off-camera?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the feathers on the owl's neck and head shown rippling outward in fluid waves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot a static medium close-up?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -1.053873062133789 + ], + "opposite_logprob_tokens": [ + -0.42887309193611145 + ], + "logprob_reward": -1.053873062133789, + "yes_no_margin_reward": -0.625, + "token_credit_reward": -1.053873062133789, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.06918887048959732, + 0.09515900909900665, + 0.28176745772361755, + 0.0337187834084034, + 0.07719871401786804, + 0.09474223852157593, + 0.07240447402000427, + 0.05092961713671684, + 0.09813114255666733, + 0.1267596334218979 + ], + "credit_map": [ + 0.0069681997410953045, + 0.007742585614323616, + 0.009092269465327263, + 0.010949565097689629, + 0.0074204737320542336, + 0.008101379498839378, + 0.009469350799918175, + 0.009445054456591606, + 0.017424730584025383, + 0.008900555782020092, + 0.005975666455924511, + 0.024506237357854843, + 0.010140636935830116, + 0.006888359319418669, + 0.009816162288188934, + 0.011506669223308563, + 0.013715855777263641, + 0.005185824818909168, + 0.010124563239514828, + 0.009754206985235214, + 0.008004656061530113, + 0.09811440855264664, + 0.015172389335930347, + 0.12169556319713593, + 0.003425367409363389, + 0.003519178833812475, + 0.005729721859097481, + 0.004266820382326841, + 0.003853504778817296, + 0.003568734973669052, + 0.0054368856362998486, + 0.003918573260307312, + 0.007006781175732613, + 0.008039199747145176, + 0.01042244303971529, + 0.013964981772005558, + 0.009226427413523197, + 0.009403296746313572, + 0.01163231860846281, + 0.007503275293856859, + 0.0052134329453110695, + 0.0037240113597363234, + 0.004291662015020847, + 0.0044217356480658054, + 0.006748156622052193, + 0.031301531940698624, + 0.005025828257203102, + 0.03401589021086693, + 0.007840107195079327, + 0.014760123565793037, + 0.00692760432139039, + 0.006736410781741142, + 0.01345168799161911, + 0.00509665859863162, + 0.006986591499298811, + 0.010605291463434696, + 0.0075564999133348465, + 0.009182318113744259, + 0.005299180746078491, + 0.0050464714877307415, + 0.006033169571310282, + 0.00702269421890378, + 0.0037934051360934973, + 0.0069958860985934734, + 0.0040508294478058815, + 0.00781911239027977, + 0.005639321636408567, + 0.006712626200169325, + 0.012863344512879848, + 0.024109529331326485, + 0.007732484024018049, + 0.029203910380601883, + 0.007264142856001854, + 0.005351528991013765, + 0.00607449933886528, + 0.006950092036277056, + 0.005893781315535307, + 0.038461532443761826, + 0.012930918484926224, + 0.043833158910274506 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 9, + 8 + ], + "selected_reward": -1.053873062133789 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a Snowy Owl perched on a snow-covered branch?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the video rendered in stark black-and-white with high-contrast lighting?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the owl's head dash rapidly toward an unseen sound off-camera?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the feathers on the owl's neck and head shown rippling outward in fluid waves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot a static medium close-up?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02330170013010502 + ], + "opposite_logprob_tokens": [ + -3.773301601409912 + ], + "logprob_reward": -0.02330170013010502, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.02330170013010502, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09030566364526749, + 0.06866993010044098, + 0.2677570581436157, + 0.032130468636751175, + 0.03761637210845947, + 0.0323103666305542, + 0.03719956427812576, + 0.04908939450979233, + 0.18253207206726074, + 0.20238912105560303 + ], + "credit_map": [ + 0.008149741217494011, + 0.011405160650610924, + 0.010387144982814789, + 0.012393858283758163, + 0.008926154114305973, + 0.009387925267219543, + 0.012720249593257904, + 0.016935432329773903, + 0.01147878635674715, + 0.006697437725961208, + 0.004564421251416206, + 0.018723534420132637, + 0.00675599230453372, + 0.004951280076056719, + 0.005774262361228466, + 0.009724224917590618, + 0.017800062894821167, + 0.008831881918013096, + 0.013645694591104984, + 0.014375657774508, + 0.016764676198363304, + 0.08137315511703491, + 0.015884865075349808, + 0.09908107668161392, + 0.0032785541843622923, + 0.003587100189179182, + 0.0063468595035374165, + 0.0045302859507501125, + 0.003236205317080021, + 0.0029001992661505938, + 0.0050363908521831036, + 0.003214877797290683, + 0.003296750131994486, + 0.004424696788191795, + 0.005429233890026808, + 0.004851984325796366, + 0.00483569223433733, + 0.00639770133420825, + 0.004167511593550444, + 0.004212807398289442, + 0.002699574688449502, + 0.0038900519721210003, + 0.005574674345552921, + 0.00332845957018435, + 0.0033046305179595947, + 0.004118948243558407, + 0.0035523097030818462, + 0.005841720849275589, + 0.004110356792807579, + 0.006042001768946648, + 0.00301565695554018, + 0.004029016476124525, + 0.00593008054420352, + 0.0037045415956526995, + 0.004344257991760969, + 0.006023652385920286, + 0.004654552321881056, + 0.006340618245303631, + 0.007243890315294266, + 0.005881025921553373, + 0.005375238135457039, + 0.006484264973551035, + 0.006714059505611658, + 0.006395751144737005, + 0.009827151894569397, + 0.0128738759085536, + 0.014639402739703655, + 0.015282873064279556, + 0.015057257376611233, + 0.03943425044417381, + 0.02418823167681694, + 0.05122904106974602, + 0.009895889088511467, + 0.006028386764228344, + 0.005325286649167538, + 0.006878397427499294, + 0.009407821111381054, + 0.07007797807455063, + 0.012054654769599438, + 0.08272071927785873 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 9, + 8 + ], + "selected_reward": -0.02330170013010502 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a Snowy Owl perched on a snow-covered branch?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the video rendered in stark black-and-white with high-contrast lighting?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the owl's head dash rapidly toward an unseen sound off-camera?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the feathers on the owl's neck and head shown rippling outward in fluid waves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot a static medium close-up?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.12705950438976288 + ], + "opposite_logprob_tokens": [ + -2.1270594596862793 + ], + "logprob_reward": -0.12705950438976288, + "yes_no_margin_reward": 2.0, + "token_credit_reward": -0.12705950438976288, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06430498510599136, + 0.05188530310988426, + 0.33214321732521057, + 0.05267795920372009, + 0.05249509587883949, + 0.1501757651567459, + 0.040922317653894424, + 0.05375649407505989, + 0.13500699400901794, + 0.06663186848163605 + ], + "credit_map": [ + 0.005343195050954819, + 0.005747828632593155, + 0.0071043455973267555, + 0.008234190754592419, + 0.006819004192948341, + 0.008624322712421417, + 0.009368878789246082, + 0.013063217513263226, + 0.007000341545790434, + 0.006306719966232777, + 0.0034214972984045744, + 0.008818371221423149, + 0.008430843241512775, + 0.004629060626029968, + 0.006236635614186525, + 0.007041832432150841, + 0.018986277282238007, + 0.0069556813687086105, + 0.012031598016619682, + 0.01234141644090414, + 0.012422274798154831, + 0.1125243529677391, + 0.021105436608195305, + 0.13577619194984436, + 0.0061703030951321125, + 0.006022499408572912, + 0.009217637591063976, + 0.005607891362160444, + 0.007331155240535736, + 0.005684296600520611, + 0.0063508423045277596, + 0.006293334066867828, + 0.00482749380171299, + 0.00581257464364171, + 0.010082610882818699, + 0.008316456340253353, + 0.005583717487752438, + 0.0050485399551689625, + 0.007488435599952936, + 0.005335269495844841, + 0.008029710501432419, + 0.0069142673164606094, + 0.006053177639842033, + 0.005995930638164282, + 0.011593245901167393, + 0.05292055383324623, + 0.01047058030962944, + 0.048198290169239044, + 0.003837884869426489, + 0.007743647787719965, + 0.0031674387864768505, + 0.004359897691756487, + 0.006227240897715092, + 0.0041382405906915665, + 0.005982432048767805, + 0.005465537775307894, + 0.005990875419229269, + 0.007665070705115795, + 0.004695168696343899, + 0.006690292619168758, + 0.008637266233563423, + 0.007883942686021328, + 0.005774546414613724, + 0.006419328972697258, + 0.007838564924895763, + 0.00557291554287076, + 0.007867074571549892, + 0.007101330440491438, + 0.009272599592804909, + 0.04024156183004379, + 0.008560549467802048, + 0.04855239763855934, + 0.0038985663559287786, + 0.004208990838378668, + 0.004467417020350695, + 0.0033854262437671423, + 0.0044978465884923935, + 0.01975512132048607, + 0.0034605395048856735, + 0.02295796386897564 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 5, + 8 + ], + "selected_reward": -0.12705950438976288 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "No", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000050/group_4/index_tensor([7937, 7937])_group_4_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000050/group_4/index_tensor([7937, 7937])_group_4_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..8ab5e880255cc0cbfa16632b3860ec09afc927f0 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000050/group_4/index_tensor([7937, 7937])_group_4_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "In a slightly handheld, medium shot, a towering succubus stands quietly in the center of an ethereal, dark hall, the entire image imbued with the exquisite texture of a meticulously detailed digital illustration. Her cascading silver-white hair contrasts sharply with the massive black wings that emanate from behind her. Each feather gleams with a deep, obsidian sheen under the nebulous light filtering through the Gothic arched windows. She clad in a flowing, pure white silk gown that gently cascades to the floor. The cool light accentuates her fair skin and the dress, casting soft lavender shadows. With a subtle tremor of the camera, she slowly raises her right arm, gently twisting her wrist. Her long, white fingers gracefully extend toward the camera in an irresistible invitation. Her stunningly beautiful, cold face, pierced by deep, piercing eyes, creates a paradoxical atmosphere of sacredness and enchantment, imbued with an unreal, artistic beauty.", + "seed": 3446, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000050/group_4/index_7937_seed_3446_dp_8_rank_16_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 4, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.049138620216399434, + "vlm_logprob": -0.049138620216399434, + "vlm_yes_no_margin": 4.225, + "vlm_token_credit": -0.049138620216399434, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a towering succubus standing in the center of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the succubus have cascading silver-white hair and massive black wings with an obsidian sheen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the succubus wearing a flowing pure white silk gown that reaches the floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the succubus slowly raise her right arm and twist her wrist to extend her fingers toward the camera?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot slightly handheld with a subtle tremor?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.021772179752588272 + ], + "opposite_logprob_tokens": [ + -3.8967721462249756 + ], + "logprob_reward": -0.021772179752588272, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.021772179752588272, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15258516371250153, + 0.10178565233945847, + 0.05742930248379707, + 0.06905203312635422, + 0.06220095977187157, + 0.10948833078145981, + 0.3426034152507782, + 0.03329036012291908, + 0.032141152769327164, + 0.03942359983921051 + ], + "credit_map": [ + 0.021973954513669014, + 0.017995450645685196, + 0.021178921684622765, + 0.025307362899184227, + 0.021080007776618004, + 0.01508164219558239, + 0.01552412286400795, + 0.014443699270486832, + 0.010786079801619053, + 0.00954888854175806, + 0.0133175328373909, + 0.010818634182214737, + 0.019970593973994255, + 0.009919887408614159, + 0.01012964267283678, + 0.017294390127062798, + 0.0048164380714297295, + 0.008477236144244671, + 0.00756397470831871, + 0.007211142685264349, + 0.010437925346195698, + 0.003665834665298462, + 0.007524954155087471, + 0.007731799501925707, + 0.006557282991707325, + 0.007520087994635105, + 0.00888374075293541, + 0.011641839519143105, + 0.014515009708702564, + 0.00363369844853878, + 0.010020122863352299, + 0.006280248053371906, + 0.005943589378148317, + 0.005650160368531942, + 0.00816617626696825, + 0.011582780629396439, + 0.011257740668952465, + 0.003883295925334096, + 0.009745837189257145, + 0.005971381440758705, + 0.011347412131726742, + 0.01372806541621685, + 0.013935488648712635, + 0.015486017800867558, + 0.012071841396391392, + 0.01570567674934864, + 0.012664414942264557, + 0.014549404382705688, + 0.025917401537299156, + 0.016180161386728287, + 0.08339336514472961, + 0.03474236652255058, + 0.03276269510388374, + 0.015432051382958889, + 0.08955517411231995, + 0.044620200991630554, + 0.003420647233724594, + 0.002719595329836011, + 0.005885082297027111, + 0.0033867398742586374, + 0.005703429225832224, + 0.003792077535763383, + 0.004327925853431225, + 0.004054863005876541, + 0.0036022907588630915, + 0.0035354492720216513, + 0.004872775170952082, + 0.003603316145017743, + 0.005772379692643881, + 0.003870324231684208, + 0.003987145144492388, + 0.002897468628361821, + 0.003922976553440094, + 0.003288279753178358, + 0.007799691986292601, + 0.003682483686134219, + 0.00572696840390563, + 0.006643346510827541, + 0.003826626343652606, + 0.0045332289300858974 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 0, + 5 + ], + "selected_reward": -0.021772179752588272 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a towering succubus standing in the center of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the succubus have cascading silver-white hair and massive black wings with an obsidian sheen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the succubus wearing a flowing pure white silk gown that reaches the floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the succubus slowly raise her right arm and twist her wrist to extend her fingers toward the camera?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot slightly handheld with a subtle tremor?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0022096047177910805 + ], + "opposite_logprob_tokens": [ + -6.127209663391113 + ], + "logprob_reward": -0.0022096047177910805, + "yes_no_margin_reward": 6.125, + "token_credit_reward": -0.0022096047177910805, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1872205287218094, + 0.10837556421756744, + 0.06904279440641403, + 0.05915357545018196, + 0.026289617642760277, + 0.08439204096794128, + 0.18057221174240112, + 0.06759926676750183, + 0.06389622390270233, + 0.15345817804336548 + ], + "credit_map": [ + 0.024120735004544258, + 0.021231018006801605, + 0.025965247303247452, + 0.024265436455607414, + 0.0227586030960083, + 0.019018283113837242, + 0.019196320325136185, + 0.030664896592497826, + 0.015115047805011272, + 0.01238070148974657, + 0.01552268210798502, + 0.009984695352613926, + 0.009873658418655396, + 0.014737069606781006, + 0.008812195621430874, + 0.021949514746665955, + 0.0061800251714885235, + 0.01151623111218214, + 0.00824806559830904, + 0.009478077292442322, + 0.009616192430257797, + 0.005480614490807056, + 0.008425169624388218, + 0.010098417289555073, + 0.006127330008894205, + 0.005452994257211685, + 0.008377300575375557, + 0.014268509112298489, + 0.00967219565063715, + 0.003779926337301731, + 0.005986401345580816, + 0.005488920491188765, + 0.004451600834727287, + 0.002471624407917261, + 0.004950101021677256, + 0.002482365118339658, + 0.0029587573371827602, + 0.0024932946544140577, + 0.0027756381314247847, + 0.003706236369907856, + 0.009358057752251625, + 0.010360609740018845, + 0.010751495137810707, + 0.01233410369604826, + 0.007659717462956905, + 0.012170427478849888, + 0.009918652474880219, + 0.011838980950415134, + 0.017656909301877022, + 0.011781331151723862, + 0.03553508594632149, + 0.016928978264331818, + 0.019792474806308746, + 0.014242425560951233, + 0.04381069168448448, + 0.020824315026402473, + 0.007570264860987663, + 0.006308375392109156, + 0.015096699818968773, + 0.0077049462124705315, + 0.00685224961489439, + 0.005257480777800083, + 0.006585873197764158, + 0.012223377823829651, + 0.006236181128770113, + 0.0053369877859950066, + 0.01155907940119505, + 0.007222279440611601, + 0.007875112816691399, + 0.004059630446135998, + 0.013934752903878689, + 0.0076722013764083385, + 0.014606911689043045, + 0.010942361317574978, + 0.025292832404375076, + 0.01592710055410862, + 0.0258855652064085, + 0.017713015899062157, + 0.026037460193037987, + 0.01705292984843254 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 6, + 9 + ], + "selected_reward": -0.0022096047177910805 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a towering succubus standing in the center of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the succubus have cascading silver-white hair and massive black wings with an obsidian sheen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the succubus wearing a flowing pure white silk gown that reaches the floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the succubus slowly raise her right arm and twist her wrist to extend her fingers toward the camera?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot slightly handheld with a subtle tremor?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004095738288015127 + ], + "opposite_logprob_tokens": [ + -5.504095554351807 + ], + "logprob_reward": -0.004095738288015127, + "yes_no_margin_reward": 5.5, + "token_credit_reward": -0.004095738288015127, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12927252054214478, + 0.10202538222074509, + 0.08311895281076431, + 0.06946925818920135, + 0.03630894050002098, + 0.04697272554039955, + 0.32957303524017334, + 0.055092472583055496, + 0.06626059859991074, + 0.08190608769655228 + ], + "credit_map": [ + 0.013060163706541061, + 0.012712151743471622, + 0.016596928238868713, + 0.015778444707393646, + 0.02870524674654007, + 0.00867344718426466, + 0.014055829495191574, + 0.019690319895744324, + 0.010736626572906971, + 0.010556284338235855, + 0.013435517437756062, + 0.011526521295309067, + 0.02250969596207142, + 0.007619916461408138, + 0.014414804987609386, + 0.011226032860577106, + 0.007262146566063166, + 0.013087938539683819, + 0.010157151147723198, + 0.00879697222262621, + 0.017164045944809914, + 0.005670808255672455, + 0.010430758818984032, + 0.010549135506153107, + 0.006984173785895109, + 0.00795500073581934, + 0.00835814792662859, + 0.0153014175593853, + 0.010713331401348114, + 0.004561086185276508, + 0.007698190864175558, + 0.007897916249930859, + 0.005254742689430714, + 0.0030505447648465633, + 0.004826746415346861, + 0.004723106510937214, + 0.00791984610259533, + 0.0021877980325371027, + 0.005241699516773224, + 0.0031044590286910534, + 0.0044124918058514595, + 0.00400827219709754, + 0.005698117893189192, + 0.0060001276433467865, + 0.005741075612604618, + 0.004455998539924622, + 0.009633496403694153, + 0.007023145444691181, + 0.025555621832609177, + 0.01901593804359436, + 0.07527481019496918, + 0.03360000625252724, + 0.03401890769600868, + 0.014918156899511814, + 0.08718767762184143, + 0.04000192508101463, + 0.006501578260213137, + 0.005174577236175537, + 0.01236097700893879, + 0.006716551724821329, + 0.007551487069576979, + 0.004044344648718834, + 0.005418243817985058, + 0.0073247188702225685, + 0.005691303871572018, + 0.006888296455144882, + 0.007914851419627666, + 0.008341268636286259, + 0.008890416473150253, + 0.004605745896697044, + 0.015520146116614342, + 0.008408569730818272, + 0.008467656560242176, + 0.006213477812707424, + 0.015446294099092484, + 0.009079272858798504, + 0.011668423190712929, + 0.008104832842946053, + 0.01267791073769331, + 0.010248222388327122 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 0, + 1 + ], + "selected_reward": -0.004095738288015127 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a towering succubus standing in the center of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the succubus have cascading silver-white hair and massive black wings with an obsidian sheen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the succubus wearing a flowing pure white silk gown that reaches the floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the succubus slowly raise her right arm and twist her wrist to extend her fingers toward the camera?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot slightly handheld with a subtle tremor?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.016065318137407303 + ], + "opposite_logprob_tokens": [ + -4.1410651206970215 + ], + "logprob_reward": -0.016065318137407303, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.016065318137407303, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08168667554855347, + 0.08693250268697739, + 0.030044836923480034, + 0.17732492089271545, + 0.044532228261232376, + 0.04294509440660477, + 0.43223994970321655, + 0.038971737027168274, + 0.02747870609164238, + 0.037843335419893265 + ], + "credit_map": [ + 0.008256151340901852, + 0.006998763419687748, + 0.010439513251185417, + 0.013682584278285503, + 0.016724171116948128, + 0.0069846804253757, + 0.0097008366137743, + 0.008899977430701256, + 0.007642924785614014, + 0.007657283917069435, + 0.011645711027085781, + 0.015348432585597038, + 0.01911141909658909, + 0.005376845132559538, + 0.01014812383800745, + 0.010001754388213158, + 0.002419109456241131, + 0.003394084982573986, + 0.0038643083535134792, + 0.0034102832432836294, + 0.006649617571383715, + 0.0018325337441638112, + 0.005480646155774593, + 0.0029942539986222982, + 0.014622950926423073, + 0.015163464471697807, + 0.02452893927693367, + 0.05345503240823746, + 0.02121809870004654, + 0.008753987960517406, + 0.023586157709360123, + 0.015996288508176804, + 0.004019199870526791, + 0.003854301990941167, + 0.005687496159225702, + 0.009369080886244774, + 0.007552250754088163, + 0.0028366437181830406, + 0.007016684394329786, + 0.004196574911475182, + 0.004044025205075741, + 0.0036045559681952, + 0.005379389505833387, + 0.0049697356298565865, + 0.006070327013731003, + 0.004988687578588724, + 0.007426277734339237, + 0.006462092977017164, + 0.030320078134536743, + 0.015438240021467209, + 0.11250627040863037, + 0.04412047192454338, + 0.04026169329881668, + 0.014886847697198391, + 0.11264302581548691, + 0.06206332892179489, + 0.003304285928606987, + 0.002921894658356905, + 0.006680623162537813, + 0.004637531470507383, + 0.00576317822560668, + 0.0026735877618193626, + 0.004926434718072414, + 0.008064200170338154, + 0.002476176479831338, + 0.00224155536852777, + 0.004585770424455404, + 0.0030766527634114027, + 0.004398729652166367, + 0.0020792288705706596, + 0.00496989581733942, + 0.00365069811232388, + 0.003092915052548051, + 0.004802488721907139, + 0.005711214151233435, + 0.005056056659668684, + 0.006938687525689602, + 0.0029853915330022573, + 0.004578349646180868, + 0.004678231664001942 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 3, + 1 + ], + "selected_reward": -0.016065318137407303 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a towering succubus standing in the center of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the succubus have cascading silver-white hair and massive black wings with an obsidian sheen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the succubus wearing a flowing pure white silk gown that reaches the floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the succubus slowly raise her right arm and twist her wrist to extend her fingers toward the camera?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot slightly handheld with a subtle tremor?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.20155026018619537 + ], + "opposite_logprob_tokens": [ + -1.7015502452850342 + ], + "logprob_reward": -0.20155026018619537, + "yes_no_margin_reward": 1.5, + "token_credit_reward": -0.20155026018619537, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2101830542087555, + 0.12414415180683136, + 0.0339556522667408, + 0.07585643976926804, + 0.05218619853258133, + 0.12965017557144165, + 0.1562018245458603, + 0.0530734546482563, + 0.0912618413567543, + 0.07348722964525223 + ], + "credit_map": [ + 0.02121366746723652, + 0.02085995115339756, + 0.025759050622582436, + 0.030613020062446594, + 0.04284856468439102, + 0.016089454293251038, + 0.02753084897994995, + 0.02526850439608097, + 0.010294784791767597, + 0.01424215268343687, + 0.015267619863152504, + 0.018130432814359665, + 0.02700410783290863, + 0.008422589860856533, + 0.013858317397534847, + 0.016924140974879265, + 0.0036906199529767036, + 0.003942591603845358, + 0.004435610957443714, + 0.003418231848627329, + 0.005072451662272215, + 0.003382802475243807, + 0.004179032053798437, + 0.00583431264385581, + 0.008277293294668198, + 0.006708577275276184, + 0.012357317842543125, + 0.01279460545629263, + 0.01495509222149849, + 0.0047056348994374275, + 0.009136446751654148, + 0.006921466905623674, + 0.005011703819036484, + 0.003964032046496868, + 0.007071226369589567, + 0.007397856097668409, + 0.009518254548311234, + 0.004205755423754454, + 0.010238135233521461, + 0.004779234528541565, + 0.013099140487611294, + 0.012299048714339733, + 0.01782933622598648, + 0.024751251563429832, + 0.017002103850245476, + 0.010815314948558807, + 0.019992904737591743, + 0.013861076906323433, + 0.012693719007074833, + 0.010219758376479149, + 0.028115130960941315, + 0.015276256017386913, + 0.01674172654747963, + 0.010889367200434208, + 0.04409574717283249, + 0.01817011833190918, + 0.004544246941804886, + 0.005053984001278877, + 0.009329766035079956, + 0.0040107243694365025, + 0.0099601149559021, + 0.006678138859570026, + 0.005812393501400948, + 0.007684087380766869, + 0.010883928276598454, + 0.0051770261488854885, + 0.01892378367483616, + 0.008253243751823902, + 0.009748771786689758, + 0.005689298268407583, + 0.012566538527607918, + 0.020019249990582466, + 0.006242101546376944, + 0.005273252259939909, + 0.008705896325409412, + 0.005705923307687044, + 0.01082816906273365, + 0.009892832487821579, + 0.01787729188799858, + 0.008961766958236694 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 6, + 5 + ], + "selected_reward": -0.20155026018619537 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "In a slightly handheld, medium shot, a towering succubus stands quietly in the center of an ethereal, dark hall, the entire image imbued with the exquisite texture of a meticulously detailed digital illustration. Her cascading silver-white hair contrasts sharply with the massive black wings that emanate from behind her. Each feather gleams with a deep, obsidian sheen under the nebulous light filtering through the Gothic arched windows. She clad in a flowing, pure white silk gown that gently cascades to the floor. The cool light accentuates her fair skin and the dress, casting soft lavender shadows. With a subtle tremor of the camera, she slowly raises her right arm, gently twisting her wrist. Her long, white fingers gracefully extend toward the camera in an irresistible invitation. Her stunningly beautiful, cold face, pierced by deep, piercing eyes, creates a paradoxical atmosphere of sacredness and enchantment, imbued with an unreal, artistic beauty.", + "seed": 3447, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000050/group_4/index_7937_seed_3447_dp_8_rank_16_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 4, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.05433865683153272, + "vlm_logprob": -0.05433865683153272, + "vlm_yes_no_margin": 3.85, + "vlm_token_credit": -0.05433865683153272, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a towering succubus standing in the center of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the succubus have cascading silver-white hair and massive black wings with an obsidian sheen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the succubus wearing a flowing pure white silk gown that reaches the floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the succubus slowly raise her right arm and twist her wrist to extend her fingers toward the camera?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot slightly handheld with a subtle tremor?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.00845152884721756 + ], + "opposite_logprob_tokens": [ + -5.133451461791992 + ], + "logprob_reward": -0.00845152884721756, + "yes_no_margin_reward": 5.125, + "token_credit_reward": -0.00845152884721756, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12546314299106598, + 0.09143700450658798, + 0.08571627736091614, + 0.07195945084095001, + 0.19175668060779572, + 0.22947371006011963, + 0.05247173085808754, + 0.049628883600234985, + 0.03408040851354599, + 0.06801266968250275 + ], + "credit_map": [ + 0.02211729995906353, + 0.021315038204193115, + 0.014367516152560711, + 0.010007679462432861, + 0.020064212381839752, + 0.011394654400646687, + 0.012184090912342072, + 0.014012657105922699, + 0.009229783900082111, + 0.008203206583857536, + 0.013795399107038975, + 0.005676864180713892, + 0.011760947294533253, + 0.01889072172343731, + 0.008552540093660355, + 0.015327544882893562, + 0.011148111894726753, + 0.011401359923183918, + 0.01223637256771326, + 0.011793331243097782, + 0.01521254051476717, + 0.009759599342942238, + 0.006346491165459156, + 0.007818475365638733, + 0.00607313122600317, + 0.00950695015490055, + 0.007470118347555399, + 0.011598732322454453, + 0.017501050606369972, + 0.007073832210153341, + 0.005684751085937023, + 0.007050889544188976, + 0.024146409705281258, + 0.021144011989235878, + 0.03620379418134689, + 0.01577380858361721, + 0.025871414691209793, + 0.03126171976327896, + 0.013562248088419437, + 0.023793276399374008, + 0.03800150007009506, + 0.029156483709812164, + 0.027099106460809708, + 0.03257044777274132, + 0.02988310344517231, + 0.024749798700213432, + 0.018721070140600204, + 0.029292210936546326, + 0.005518186371773481, + 0.006113132461905479, + 0.005607239436358213, + 0.008167272433638573, + 0.010003908537328243, + 0.005896827206015587, + 0.0052443258464336395, + 0.005920844618231058, + 0.005648065824061632, + 0.005415673833340406, + 0.006446421146392822, + 0.006790748331695795, + 0.0064615411683917046, + 0.006848771125078201, + 0.004556296858936548, + 0.007461369503289461, + 0.0027897870168089867, + 0.003311648964881897, + 0.0034683148842304945, + 0.004568996839225292, + 0.004977771546691656, + 0.0071000405587255955, + 0.003036144655197859, + 0.004827703814953566, + 0.007427757605910301, + 0.007561194710433483, + 0.010790347121655941, + 0.008146612904965878, + 0.01006272342056036, + 0.006945193745195866, + 0.009354781359434128, + 0.00772405881434679 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 4, + 0 + ], + "selected_reward": -0.00845152884721756 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a towering succubus standing in the center of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the succubus have cascading silver-white hair and massive black wings with an obsidian sheen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the succubus wearing a flowing pure white silk gown that reaches the floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the succubus slowly raise her right arm and twist her wrist to extend her fingers toward the camera?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot slightly handheld with a subtle tremor?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.00221055606380105 + ], + "opposite_logprob_tokens": [ + -6.12721061706543 + ], + "logprob_reward": -0.00221055606380105, + "yes_no_margin_reward": 6.125, + "token_credit_reward": -0.00221055606380105, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10706200450658798, + 0.13075432181358337, + 0.059145428240299225, + 0.04680315777659416, + 0.28995779156684875, + 0.17555470764636993, + 0.04366503283381462, + 0.05762236937880516, + 0.04574943706393242, + 0.0436856783926487 + ], + "credit_map": [ + 0.010207487270236015, + 0.012469522655010223, + 0.013160087168216705, + 0.012643986381590366, + 0.017402874305844307, + 0.010971594601869583, + 0.008563844487071037, + 0.021642597392201424, + 0.013345311395823956, + 0.013072962872684002, + 0.020724356174468994, + 0.009185000322759151, + 0.018315903842449188, + 0.027472734451293945, + 0.011312690563499928, + 0.017325369641184807, + 0.008953829295933247, + 0.008197556249797344, + 0.009446561336517334, + 0.007560236845165491, + 0.00886622630059719, + 0.007262025494128466, + 0.003764358814805746, + 0.005094633903354406, + 0.005208578892052174, + 0.005946321412920952, + 0.006281254813075066, + 0.007091379724442959, + 0.009020921774208546, + 0.004824574571102858, + 0.003657958935946226, + 0.004772166721522808, + 0.03933161124587059, + 0.029749378561973572, + 0.061449166387319565, + 0.022133760154247284, + 0.0355973057448864, + 0.04453139007091522, + 0.022052448242902756, + 0.03511274978518486, + 0.02886766754090786, + 0.022400086745619774, + 0.020049920305609703, + 0.02258709818124771, + 0.022907136008143425, + 0.019464969635009766, + 0.014800012111663818, + 0.024477815255522728, + 0.0048230369575321674, + 0.00515463761985302, + 0.006184917408972979, + 0.00783515814691782, + 0.006656800862401724, + 0.004497083369642496, + 0.004461256787180901, + 0.004052143543958664, + 0.006176218390464783, + 0.005100926384329796, + 0.009416768327355385, + 0.007899493910372257, + 0.00794678833335638, + 0.008782446384429932, + 0.004892941098660231, + 0.007406790275126696, + 0.005696725565940142, + 0.004792539868503809, + 0.006376071833074093, + 0.004654623568058014, + 0.004824802279472351, + 0.006253620143979788, + 0.004196236375719309, + 0.008954817429184914, + 0.005548372864723206, + 0.0046560089103877544, + 0.007071062922477722, + 0.005021326709538698, + 0.006243360694497824, + 0.003907559905201197, + 0.006484032608568668, + 0.004753952380269766 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 5, + 1 + ], + "selected_reward": -0.00221055606380105 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a towering succubus standing in the center of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the succubus have cascading silver-white hair and massive black wings with an obsidian sheen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the succubus wearing a flowing pure white silk gown that reaches the floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the succubus slowly raise her right arm and twist her wrist to extend her fingers toward the camera?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot slightly handheld with a subtle tremor?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.038101088255643845 + ], + "opposite_logprob_tokens": [ + -3.2881011962890625 + ], + "logprob_reward": -0.038101088255643845, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.038101088255643845, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1697966754436493, + 0.08474947512149811, + 0.08051078766584396, + 0.13546650111675262, + 0.10863610357046127, + 0.14406602084636688, + 0.05973740667104721, + 0.061945509165525436, + 0.10971508175134659, + 0.045376457273960114 + ], + "credit_map": [ + 0.01646399311721325, + 0.019717706367373466, + 0.020202603191137314, + 0.02417914569377899, + 0.02646872214972973, + 0.020095331594347954, + 0.01203505415469408, + 0.030634116381406784, + 0.00930336955934763, + 0.009785395115613937, + 0.010683747939765453, + 0.0070763640105724335, + 0.012003994546830654, + 0.018153026700019836, + 0.006264216732233763, + 0.01147935725748539, + 0.01418527401983738, + 0.010633892379701138, + 0.013176579028367996, + 0.009097479283809662, + 0.011675307527184486, + 0.009127501398324966, + 0.005006040446460247, + 0.007608710788190365, + 0.015874125063419342, + 0.017594749107956886, + 0.016121717169880867, + 0.015610742382705212, + 0.021953172981739044, + 0.017679870128631592, + 0.013473189435899258, + 0.01715894043445587, + 0.011240457184612751, + 0.011558757163584232, + 0.015307055786252022, + 0.013704921118915081, + 0.01889236643910408, + 0.013537301681935787, + 0.008972023613750935, + 0.015423222444951534, + 0.022083520889282227, + 0.01903696544468403, + 0.016216043382883072, + 0.018348781391978264, + 0.017791185528039932, + 0.017947765067219734, + 0.009869411587715149, + 0.02277233824133873, + 0.007276459131389856, + 0.006482163444161415, + 0.00613883463665843, + 0.010042600333690643, + 0.009794065728783607, + 0.006131495349109173, + 0.006609652657061815, + 0.007262137718498707, + 0.00684445071965456, + 0.006817949470132589, + 0.008331893011927605, + 0.009049189276993275, + 0.007020832505077124, + 0.009770658798515797, + 0.005088725592941046, + 0.00902180653065443, + 0.012226910330355167, + 0.017096271738409996, + 0.012849696911871433, + 0.01080850325524807, + 0.013214943930506706, + 0.02040126919746399, + 0.009367546066641808, + 0.013749941252171993, + 0.005777483340352774, + 0.00521061522886157, + 0.0066929473541677, + 0.004610083997249603, + 0.008152177557349205, + 0.006316892337054014, + 0.004330197814851999, + 0.004286058247089386 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 5, + 3 + ], + "selected_reward": -0.038101088255643845 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a towering succubus standing in the center of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the succubus have cascading silver-white hair and massive black wings with an obsidian sheen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the succubus wearing a flowing pure white silk gown that reaches the floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the succubus slowly raise her right arm and twist her wrist to extend her fingers toward the camera?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot slightly handheld with a subtle tremor?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04305609315633774 + ], + "opposite_logprob_tokens": [ + -3.168056011199951 + ], + "logprob_reward": -0.04305609315633774, + "yes_no_margin_reward": 3.125, + "token_credit_reward": -0.04305609315633774, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08177868276834488, + 0.10568949580192566, + 0.07479840517044067, + 0.09046915173530579, + 0.17453604936599731, + 0.06045717000961304, + 0.08097150176763535, + 0.14754174649715424, + 0.14003364741802216, + 0.04372416436672211 + ], + "credit_map": [ + 0.010404991917312145, + 0.010401787236332893, + 0.010959679260849953, + 0.008317926898598671, + 0.013572920113801956, + 0.00798751413822174, + 0.007359285373240709, + 0.012774579226970673, + 0.010305252857506275, + 0.009606822393834591, + 0.016337374225258827, + 0.009344032034277916, + 0.012881060130894184, + 0.021405428647994995, + 0.009115684777498245, + 0.01669384352862835, + 0.007421011105179787, + 0.010644499212503433, + 0.009313383139669895, + 0.010205041617155075, + 0.014452020637691021, + 0.00862918235361576, + 0.005728425458073616, + 0.00840484257787466, + 0.012312299571931362, + 0.012495380826294422, + 0.010038462467491627, + 0.01576162874698639, + 0.016768725588917732, + 0.0075820996426045895, + 0.007571530528366566, + 0.007939024828374386, + 0.021580569446086884, + 0.025688018649816513, + 0.027757273986935616, + 0.020031312480568886, + 0.02016703598201275, + 0.026089850813150406, + 0.010537956841289997, + 0.02268403023481369, + 0.008260739967226982, + 0.00783286988735199, + 0.00911897886544466, + 0.004727378021925688, + 0.007698810193687677, + 0.007719853892922401, + 0.004023815505206585, + 0.011074724607169628, + 0.007865270599722862, + 0.0077679455280303955, + 0.015599162317812443, + 0.011436664499342442, + 0.011212768033146858, + 0.01053662970662117, + 0.007149028591811657, + 0.009404030628502369, + 0.01510486751794815, + 0.01579355262219906, + 0.015054789371788502, + 0.02595585584640503, + 0.02642478235065937, + 0.017527952790260315, + 0.01710384152829647, + 0.014576097950339317, + 0.015858592465519905, + 0.015646882355213165, + 0.014552624896168709, + 0.014089453965425491, + 0.019538193941116333, + 0.023258449509739876, + 0.011201109737157822, + 0.02588833123445511, + 0.004193748813122511, + 0.0031226230785250664, + 0.005011704284697771, + 0.004879418760538101, + 0.0076415762305259705, + 0.007931615225970745, + 0.004304850473999977, + 0.00663862656801939 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 7, + 8 + ], + "selected_reward": -0.04305609315633774 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a towering succubus standing in the center of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the succubus have cascading silver-white hair and massive black wings with an obsidian sheen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the succubus wearing a flowing pure white silk gown that reaches the floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the succubus slowly raise her right arm and twist her wrist to extend her fingers toward the camera?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot slightly handheld with a subtle tremor?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.1798740178346634 + ], + "opposite_logprob_tokens": [ + -1.804874062538147 + ], + "logprob_reward": -0.1798740178346634, + "yes_no_margin_reward": 1.625, + "token_credit_reward": -0.1798740178346634, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09438882768154144, + 0.06015266478061676, + 0.12694863975048065, + 0.1555359810590744, + 0.12948215007781982, + 0.13007615506649017, + 0.03516727685928345, + 0.09989555925130844, + 0.12652817368507385, + 0.04182460904121399 + ], + "credit_map": [ + 0.01382615976035595, + 0.011873459443449974, + 0.011199047788977623, + 0.013974852859973907, + 0.01369443815201521, + 0.008163096383213997, + 0.009302585385739803, + 0.012355183251202106, + 0.00531708775088191, + 0.005750949494540691, + 0.012021740898489952, + 0.007402041461318731, + 0.009202687069773674, + 0.007160344161093235, + 0.004510083701461554, + 0.008787726052105427, + 0.01694195717573166, + 0.01303460355848074, + 0.0204464141279459, + 0.017310984432697296, + 0.020130882039666176, + 0.015704918652772903, + 0.008324923925101757, + 0.015053937211632729, + 0.018343225121498108, + 0.01578531041741371, + 0.021217303350567818, + 0.026713434606790543, + 0.030081303790211678, + 0.014743439853191376, + 0.013333015143871307, + 0.015318933874368668, + 0.017450768500566483, + 0.015406384132802486, + 0.022315945476293564, + 0.011064884252846241, + 0.020229056477546692, + 0.017593516036868095, + 0.008655651472508907, + 0.016765927895903587, + 0.018383478745818138, + 0.014434652402997017, + 0.015659017488360405, + 0.015732862055301666, + 0.018229708075523376, + 0.013776160776615143, + 0.010473867878317833, + 0.02338639460504055, + 0.004072231240570545, + 0.003963094670325518, + 0.004316249862313271, + 0.005142492707818747, + 0.005866363178938627, + 0.0037812869995832443, + 0.0035461822990328074, + 0.004479374270886183, + 0.012967842631042004, + 0.01381934154778719, + 0.015187198296189308, + 0.013204232789576054, + 0.01525069773197174, + 0.010389451868832111, + 0.009349060244858265, + 0.009727725759148598, + 0.01242799125611782, + 0.016475381329655647, + 0.014836813323199749, + 0.01065649464726448, + 0.015201244503259659, + 0.02604695037007332, + 0.010704286396503448, + 0.020178990438580513, + 0.005686626303941011, + 0.0037039362359791994, + 0.006491629406809807, + 0.004398839548230171, + 0.008789927698671818, + 0.0030256493482738733, + 0.006050506606698036, + 0.0036774910986423492 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 5, + 4 + ], + "selected_reward": -0.1798740178346634 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "In a slightly handheld, medium shot, a towering succubus stands quietly in the center of an ethereal, dark hall, the entire image imbued with the exquisite texture of a meticulously detailed digital illustration. Her cascading silver-white hair contrasts sharply with the massive black wings that emanate from behind her. Each feather gleams with a deep, obsidian sheen under the nebulous light filtering through the Gothic arched windows. She clad in a flowing, pure white silk gown that gently cascades to the floor. The cool light accentuates her fair skin and the dress, casting soft lavender shadows. With a subtle tremor of the camera, she slowly raises her right arm, gently twisting her wrist. Her long, white fingers gracefully extend toward the camera in an irresistible invitation. Her stunningly beautiful, cold face, pierced by deep, piercing eyes, creates a paradoxical atmosphere of sacredness and enchantment, imbued with an unreal, artistic beauty.", + "seed": 3448, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000050/group_4/index_7937_seed_3448_dp_9_rank_18_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 4, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.054274854285176845, + "vlm_logprob": -0.054274854285176845, + "vlm_yes_no_margin": 4.374999952316284, + "vlm_token_credit": -0.054274854285176845, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a towering succubus standing in the center of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the succubus have cascading silver-white hair and massive black wings with an obsidian sheen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the succubus wearing a flowing pure white silk gown that reaches the floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the succubus slowly raise her right arm and twist her wrist to extend her fingers toward the camera?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot slightly handheld with a subtle tremor?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.020692074671387672 + ], + "opposite_logprob_tokens": [ + -4.020691871643066 + ], + "logprob_reward": -0.020692074671387672, + "yes_no_margin_reward": 3.999999761581421, + "token_credit_reward": -0.020692074671387672, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09606631845235825, + 0.20336374640464783, + 0.13607318699359894, + 0.09661165624856949, + 0.11286424845457077, + 0.1386694759130478, + 0.03782404214143753, + 0.05975319445133209, + 0.06298047304153442, + 0.05579368397593498 + ], + "credit_map": [ + 0.010552564635872841, + 0.012920753099024296, + 0.011339600197970867, + 0.014204297214746475, + 0.017780685797333717, + 0.008496159687638283, + 0.009722484275698662, + 0.011049770750105381, + 0.02102324366569519, + 0.024669572710990906, + 0.022205771878361702, + 0.020906727761030197, + 0.027861610054969788, + 0.012976394034922123, + 0.045853860676288605, + 0.02786657214164734, + 0.01641181856393814, + 0.019118638709187508, + 0.0231452789157629, + 0.014479988254606724, + 0.015099890530109406, + 0.015541229397058487, + 0.01576279290020466, + 0.016513554379343987, + 0.011705995537340641, + 0.011324632912874222, + 0.008922306820750237, + 0.008682142943143845, + 0.007754524238407612, + 0.007508796639740467, + 0.024036798626184464, + 0.016676465049386024, + 0.009897387586534023, + 0.008533048443496227, + 0.008752577006816864, + 0.008007089607417583, + 0.008200587704777718, + 0.005115245934575796, + 0.036292966455221176, + 0.02806534431874752, + 0.020997121930122375, + 0.015242106281220913, + 0.023481620475649834, + 0.019484372809529305, + 0.014082130044698715, + 0.01051647774875164, + 0.017640158534049988, + 0.017225494608283043, + 0.006368392147123814, + 0.0036196468863636255, + 0.005796748213469982, + 0.004743107128888369, + 0.004467471037060022, + 0.003118805354461074, + 0.004714903421700001, + 0.0049949646927416325, + 0.007445443421602249, + 0.006490013096481562, + 0.007491306401789188, + 0.007242188788950443, + 0.010160970501601696, + 0.003811456495895982, + 0.008008879609405994, + 0.009102934040129185, + 0.013082167133688927, + 0.005911890882998705, + 0.00829234253615141, + 0.005323235411196947, + 0.010919773951172829, + 0.004294147714972496, + 0.007510281167924404, + 0.007646630518138409, + 0.007646110374480486, + 0.0035635591484606266, + 0.009598596952855587, + 0.006391624920070171, + 0.0040059201419353485, + 0.004715828225016594, + 0.010304925963282585, + 0.009567119181156158 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 5, + 2 + ], + "selected_reward": -0.020692074671387672 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a towering succubus standing in the center of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the succubus have cascading silver-white hair and massive black wings with an obsidian sheen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the succubus wearing a flowing pure white silk gown that reaches the floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the succubus slowly raise her right arm and twist her wrist to extend her fingers toward the camera?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot slightly handheld with a subtle tremor?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0009328543092124164 + ], + "opposite_logprob_tokens": [ + -7.000932693481445 + ], + "logprob_reward": -0.0009328543092124164, + "yes_no_margin_reward": 7.0, + "token_credit_reward": -0.0009328543092124164, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.18586193025112152, + 0.12293455004692078, + 0.18465273082256317, + 0.04624980688095093, + 0.09991655498743057, + 0.13666163384914398, + 0.04236772656440735, + 0.03470240533351898, + 0.034550607204437256, + 0.11210200935602188 + ], + "credit_map": [ + 0.022304706275463104, + 0.02054363116621971, + 0.028348345309495926, + 0.022244079038500786, + 0.019984181970357895, + 0.021617023274302483, + 0.03162258863449097, + 0.0191973727196455, + 0.013728080317378044, + 0.013799209147691727, + 0.017286181449890137, + 0.010540571063756943, + 0.012256567366421223, + 0.010917143896222115, + 0.02881643921136856, + 0.01559036411345005, + 0.013145234435796738, + 0.008063658140599728, + 0.015342498198151588, + 0.012929245829582214, + 0.010565478354692459, + 0.00811049435287714, + 0.06649366021156311, + 0.050002455711364746, + 0.007294748909771442, + 0.0052276430651545525, + 0.0059048887342214584, + 0.004177414812147617, + 0.004145421087741852, + 0.007780067157000303, + 0.004821153357625008, + 0.00689847394824028, + 0.010374988429248333, + 0.00975042860955, + 0.009790905751287937, + 0.008633390069007874, + 0.008386289700865746, + 0.008126365020871162, + 0.02277836762368679, + 0.022075820714235306, + 0.020974868908524513, + 0.014427566900849342, + 0.02591969072818756, + 0.017965875566005707, + 0.01177616324275732, + 0.012055791914463043, + 0.015533927828073502, + 0.01800774596631527, + 0.006584807299077511, + 0.004535319283604622, + 0.005521851126104593, + 0.0043320306576788425, + 0.007826192304491997, + 0.003754209028556943, + 0.00451330142095685, + 0.005300013814121485, + 0.007530711125582457, + 0.003256687428802252, + 0.006154725328087807, + 0.0025796983391046524, + 0.0030315376352518797, + 0.0028685706201940775, + 0.004433067981153727, + 0.004847407341003418, + 0.004212589003145695, + 0.003373532323166728, + 0.004116058815270662, + 0.004156846087425947, + 0.005264843348413706, + 0.003163483925163746, + 0.005801526363939047, + 0.004461726639419794, + 0.014664212241768837, + 0.006100273225456476, + 0.02213648520410061, + 0.011059196665883064, + 0.009292354807257652, + 0.008046023547649384, + 0.021649757400155067, + 0.01915370486676693 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 5 + ], + "selected_reward": -0.0009328543092124164 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a towering succubus standing in the center of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the succubus have cascading silver-white hair and massive black wings with an obsidian sheen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the succubus wearing a flowing pure white silk gown that reaches the floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the succubus slowly raise her right arm and twist her wrist to extend her fingers toward the camera?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot slightly handheld with a subtle tremor?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0036168177612125874 + ], + "opposite_logprob_tokens": [ + -5.628616809844971 + ], + "logprob_reward": -0.0036168177612125874, + "yes_no_margin_reward": 5.625, + "token_credit_reward": -0.0036168177612125874, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06602905690670013, + 0.1811416745185852, + 0.2838646471500397, + 0.028059450909495354, + 0.05277391895651817, + 0.07367222011089325, + 0.04743606597185135, + 0.055019501596689224, + 0.0596202053129673, + 0.15238331258296967 + ], + "credit_map": [ + 0.006945817731320858, + 0.007887866348028183, + 0.0077253603376448154, + 0.008353409357368946, + 0.008538540452718735, + 0.007761453278362751, + 0.012269857339560986, + 0.006546756252646446, + 0.020242614671587944, + 0.020014280453324318, + 0.020492345094680786, + 0.01693359948694706, + 0.01935693994164467, + 0.014788221567869186, + 0.045610152184963226, + 0.023703519254922867, + 0.020760370418429375, + 0.016330059617757797, + 0.02552867867052555, + 0.018955998122692108, + 0.017367761582136154, + 0.012666505761444569, + 0.09541811794042587, + 0.07683715969324112, + 0.003345096716657281, + 0.003217987949028611, + 0.0032693026587367058, + 0.0028617207426577806, + 0.003568510292097926, + 0.004978409968316555, + 0.0033644246868789196, + 0.0034539964981377125, + 0.004855502862483263, + 0.004703342914581299, + 0.006152858026325703, + 0.004547927062958479, + 0.003571153851225972, + 0.003935175947844982, + 0.01110977865755558, + 0.013898181729018688, + 0.006845293566584587, + 0.008982661180198193, + 0.011828603222966194, + 0.011688265018165112, + 0.007334208115935326, + 0.005147348158061504, + 0.01116084959357977, + 0.01068499032407999, + 0.004245735239237547, + 0.004363636486232281, + 0.007249509450048208, + 0.006150187458842993, + 0.006699869874864817, + 0.003882588353008032, + 0.00902857817709446, + 0.005815957207232714, + 0.004290506709367037, + 0.004724712111055851, + 0.007356604561209679, + 0.006007920950651169, + 0.0050715068355202675, + 0.002473835600540042, + 0.012206312268972397, + 0.012888101860880852, + 0.0071705603040754795, + 0.004177025984972715, + 0.007346463389694691, + 0.0062131681479513645, + 0.006768181454390287, + 0.003454966237768531, + 0.013684595003724098, + 0.010805241763591766, + 0.016319172456860542, + 0.005613146349787712, + 0.020024629309773445, + 0.011121751740574837, + 0.009500526823103428, + 0.009675958193838596, + 0.045652035623788834, + 0.034476105123758316 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 1, + 9 + ], + "selected_reward": -0.0036168177612125874 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a towering succubus standing in the center of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the succubus have cascading silver-white hair and massive black wings with an obsidian sheen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the succubus wearing a flowing pure white silk gown that reaches the floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the succubus slowly raise her right arm and twist her wrist to extend her fingers toward the camera?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot slightly handheld with a subtle tremor?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.020576702430844307 + ], + "opposite_logprob_tokens": [ + -3.8955767154693604 + ], + "logprob_reward": -0.020576702430844307, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.020576702430844307, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.040016598999500275, + 0.028714070096611977, + 0.31263723969459534, + 0.05344439670443535, + 0.14909809827804565, + 0.04950275272130966, + 0.022357890382409096, + 0.058245956897735596, + 0.05633176118135452, + 0.22965119779109955 + ], + "credit_map": [ + 0.005266891326755285, + 0.0051657394506037235, + 0.0048103113658726215, + 0.006315009668469429, + 0.006748957559466362, + 0.0027792174369096756, + 0.004434643778949976, + 0.004495828878134489, + 0.0024830566253513098, + 0.0030809049494564533, + 0.002896001562476158, + 0.004817342385649681, + 0.004392336588352919, + 0.002789036137983203, + 0.0047026462852954865, + 0.0035527460277080536, + 0.026547415181994438, + 0.012551640160381794, + 0.023390790447592735, + 0.02308301068842411, + 0.01753385365009308, + 0.01346031203866005, + 0.11211656779050827, + 0.08395365625619888, + 0.00520683079957962, + 0.00567630585283041, + 0.005997444968670607, + 0.0049553606659173965, + 0.005120898596942425, + 0.006196274422109127, + 0.010183393023908138, + 0.010107888840138912, + 0.012129640206694603, + 0.01130027323961258, + 0.013680395670235157, + 0.010024704039096832, + 0.009274275973439217, + 0.008576973341405392, + 0.043678317219018936, + 0.04043352231383324, + 0.007445062976330519, + 0.004630683455616236, + 0.00942003633826971, + 0.0048975576646625996, + 0.00595908472314477, + 0.005532096140086651, + 0.004570263437926769, + 0.007047964725643396, + 0.0023289676755666733, + 0.0025706863962113857, + 0.0031449790112674236, + 0.002942084800451994, + 0.00369840394705534, + 0.001996254315599799, + 0.0027305521070957184, + 0.0029459611978381872, + 0.01449708640575409, + 0.005031305365264416, + 0.011151064187288284, + 0.00459654675796628, + 0.005370267201215029, + 0.0049112350679934025, + 0.004255199804902077, + 0.008433251641690731, + 0.005715178791433573, + 0.005875215400010347, + 0.005767793860286474, + 0.009238116443157196, + 0.00892528984695673, + 0.004310524556785822, + 0.010005513206124306, + 0.0064941281452775, + 0.027137337252497673, + 0.01208514254540205, + 0.0499817319214344, + 0.02334054931998253, + 0.017711402848362923, + 0.015024183318018913, + 0.04464820399880409, + 0.03972265124320984 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 9, + 4 + ], + "selected_reward": -0.020576702430844307 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a towering succubus standing in the center of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the succubus have cascading silver-white hair and massive black wings with an obsidian sheen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the succubus wearing a flowing pure white silk gown that reaches the floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the succubus slowly raise her right arm and twist her wrist to extend her fingers toward the camera?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot slightly handheld with a subtle tremor?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.22555582225322723 + ], + "opposite_logprob_tokens": [ + -1.6005557775497437 + ], + "logprob_reward": -0.22555582225322723, + "yes_no_margin_reward": 1.375, + "token_credit_reward": -0.22555582225322723, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10145637392997742, + 0.089007169008255, + 0.251476526260376, + 0.14916659891605377, + 0.10052419453859329, + 0.03772500157356262, + 0.027688100934028625, + 0.06800103187561035, + 0.0324704647064209, + 0.14248457551002502 + ], + "credit_map": [ + 0.011418141424655914, + 0.01020228210836649, + 0.013629625551402569, + 0.010147063061594963, + 0.009278794750571251, + 0.009777198545634747, + 0.02477959170937538, + 0.012223687954246998, + 0.009392770938575268, + 0.011994995176792145, + 0.008884277194738388, + 0.009745841845870018, + 0.013731527142226696, + 0.00543187977746129, + 0.018344540148973465, + 0.01148133259266615, + 0.021079707890748978, + 0.009451685473322868, + 0.018706748262047768, + 0.018704399466514587, + 0.014828567393124104, + 0.010267667472362518, + 0.09141800552606583, + 0.0670197382569313, + 0.016416192054748535, + 0.014529766514897346, + 0.016576984897255898, + 0.014886295422911644, + 0.011682147160172462, + 0.013679378665983677, + 0.0319497212767601, + 0.029446110129356384, + 0.015295448713004589, + 0.01291726902127266, + 0.012445795349776745, + 0.01354522630572319, + 0.01173762883991003, + 0.010570826940238476, + 0.01151424739509821, + 0.01249775942414999, + 0.00534815015271306, + 0.00452846847474575, + 0.006691322661936283, + 0.005268326494842768, + 0.003352686297148466, + 0.0031247418373823166, + 0.003655698848888278, + 0.0057556102983653545, + 0.005965957418084145, + 0.0027592123951762915, + 0.004119564779102802, + 0.0023684753105044365, + 0.0029198317788541317, + 0.0029993692878633738, + 0.0032109031453728676, + 0.003344786586239934, + 0.016972141340374947, + 0.004636913072317839, + 0.012757952325046062, + 0.0051766918040812016, + 0.005063280928879976, + 0.005380005110055208, + 0.0071702152490615845, + 0.010843826457858086, + 0.002747248625382781, + 0.0019135691691190004, + 0.004582114052027464, + 0.0030954908579587936, + 0.0036395627539604902, + 0.001723810681141913, + 0.007998162880539894, + 0.006770506501197815, + 0.015360192395746708, + 0.004176530055701733, + 0.011182865127921104, + 0.007087349891662598, + 0.008507230319082737, + 0.009379442781209946, + 0.04974605143070221, + 0.037044912576675415 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 3, + 9 + ], + "selected_reward": -0.22555582225322723 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "In a slightly handheld, medium shot, a towering succubus stands quietly in the center of an ethereal, dark hall, the entire image imbued with the exquisite texture of a meticulously detailed digital illustration. Her cascading silver-white hair contrasts sharply with the massive black wings that emanate from behind her. Each feather gleams with a deep, obsidian sheen under the nebulous light filtering through the Gothic arched windows. She clad in a flowing, pure white silk gown that gently cascades to the floor. The cool light accentuates her fair skin and the dress, casting soft lavender shadows. With a subtle tremor of the camera, she slowly raises her right arm, gently twisting her wrist. Her long, white fingers gracefully extend toward the camera in an irresistible invitation. Her stunningly beautiful, cold face, pierced by deep, piercing eyes, creates a paradoxical atmosphere of sacredness and enchantment, imbued with an unreal, artistic beauty.", + "seed": 3449, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000050/group_4/index_7937_seed_3449_dp_9_rank_18_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 4, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.0787962079513818, + "vlm_logprob": -0.0787962079513818, + "vlm_yes_no_margin": 4.174999988079071, + "vlm_token_credit": -0.0787962079513818, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a towering succubus standing in the center of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the succubus have cascading silver-white hair and massive black wings with an obsidian sheen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the succubus wearing a flowing pure white silk gown that reaches the floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the succubus slowly raise her right arm and twist her wrist to extend her fingers toward the camera?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot slightly handheld with a subtle tremor?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.010175495408475399 + ], + "opposite_logprob_tokens": [ + -4.760175704956055 + ], + "logprob_reward": -0.010175495408475399, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.010175495408475399, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3288937211036682, + 0.11204657703638077, + 0.043731119483709335, + 0.0731072872877121, + 0.048531752079725266, + 0.0695195198059082, + 0.06813783198595047, + 0.0393080972135067, + 0.07410789281129837, + 0.1426161229610443 + ], + "credit_map": [ + 0.030749790370464325, + 0.029522046446800232, + 0.04543856903910637, + 0.04334418848156929, + 0.058654192835092545, + 0.01751115918159485, + 0.05777846276760101, + 0.04589531570672989, + 0.011703042313456535, + 0.012930041179060936, + 0.011217831633985043, + 0.011588700115680695, + 0.013423952274024487, + 0.00987065676599741, + 0.026624636724591255, + 0.014687713235616684, + 0.005979438312351704, + 0.0046305484138429165, + 0.00552742462605238, + 0.003159027313813567, + 0.008106575347483158, + 0.004552888218313456, + 0.006249001249670982, + 0.0055262185633182526, + 0.005279687233269215, + 0.007273822091519833, + 0.005395144689828157, + 0.006926806643605232, + 0.026380563154816628, + 0.0033136100973933935, + 0.013078300282359123, + 0.005459355190396309, + 0.0038001067005097866, + 0.0037562171928584576, + 0.004094209522008896, + 0.009690199978649616, + 0.009996870532631874, + 0.002507331082597375, + 0.01084061712026596, + 0.003846200881525874, + 0.007477566134184599, + 0.007148088421672583, + 0.010955517180263996, + 0.00583425909280777, + 0.012122942134737968, + 0.005945706740021706, + 0.0098579041659832, + 0.010177535936236382, + 0.009293478913605213, + 0.006842035800218582, + 0.009640087373554707, + 0.003497902536764741, + 0.012708690017461777, + 0.005713653285056353, + 0.008595254272222519, + 0.011846725828945637, + 0.006049783900380135, + 0.003725939430296421, + 0.006047604139894247, + 0.003733833786100149, + 0.00543647026643157, + 0.0048072319477796555, + 0.0035385964438319206, + 0.005968633107841015, + 0.010051394812762737, + 0.004667592700570822, + 0.01114697940647602, + 0.006727100815623999, + 0.0151807377114892, + 0.005391099490225315, + 0.011721125803887844, + 0.009221857413649559, + 0.014872530475258827, + 0.01020000223070383, + 0.020056037232279778, + 0.012772072106599808, + 0.02466585487127304, + 0.023956380784511566, + 0.01755709759891033, + 0.018536146730184555 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 1 + ], + "selected_reward": -0.010175495408475399 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a towering succubus standing in the center of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the succubus have cascading silver-white hair and massive black wings with an obsidian sheen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the succubus wearing a flowing pure white silk gown that reaches the floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the succubus slowly raise her right arm and twist her wrist to extend her fingers toward the camera?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot slightly handheld with a subtle tremor?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0011940023396164179 + ], + "opposite_logprob_tokens": [ + -6.751194000244141 + ], + "logprob_reward": -0.0011940023396164179, + "yes_no_margin_reward": 6.75, + "token_credit_reward": -0.0011940023396164179, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10738702863454819, + 0.09590144455432892, + 0.029590701684355736, + 0.03612380847334862, + 0.33729901909828186, + 0.03874032571911812, + 0.12798471748828888, + 0.03164193406701088, + 0.07082747668027878, + 0.12450359761714935 + ], + "credit_map": [ + 0.012319988571107388, + 0.01242522057145834, + 0.012989144772291183, + 0.016224822029471397, + 0.017432641237974167, + 0.0061883325688540936, + 0.018475469201803207, + 0.011331412009894848, + 0.008343500085175037, + 0.010853374376893044, + 0.011718690395355225, + 0.008215160109102726, + 0.012256588786840439, + 0.0135843800380826, + 0.019227411597967148, + 0.011702347546815872, + 0.0027618331369012594, + 0.002943508792668581, + 0.004017442464828491, + 0.0035831904970109463, + 0.007067200727760792, + 0.0021480803843587637, + 0.0041592856869101524, + 0.0029101583641022444, + 0.0037157880142331123, + 0.0041783154010772705, + 0.004466549959033728, + 0.003304947167634964, + 0.008343525230884552, + 0.0024070448707789183, + 0.005471075884997845, + 0.0042365617118775845, + 0.02647067792713642, + 0.02273079939186573, + 0.03948662057518959, + 0.07304161787033081, + 0.031101325526833534, + 0.02098567970097065, + 0.0348876491189003, + 0.08859463781118393, + 0.0031580037903040648, + 0.004497384652495384, + 0.004330416675657034, + 0.0029672689270228148, + 0.010374420322477818, + 0.003248599823564291, + 0.0052591231651604176, + 0.004905107896775007, + 0.024443702772259712, + 0.009932482615113258, + 0.018517224118113518, + 0.005940732080489397, + 0.012534135952591896, + 0.017877615988254547, + 0.014090939424932003, + 0.02464788220822811, + 0.004854441620409489, + 0.002890383591875434, + 0.0053886245004832745, + 0.003413895843550563, + 0.004254072904586792, + 0.0034889907110482454, + 0.003107016207650304, + 0.00424451008439064, + 0.013455680571496487, + 0.004437664058059454, + 0.015670815482735634, + 0.00582881597802043, + 0.0071700322441756725, + 0.007622269913554192, + 0.005503992084413767, + 0.01113820355385542, + 0.017720326781272888, + 0.007095014676451683, + 0.022857295349240303, + 0.006158845964819193, + 0.016455862671136856, + 0.019462136551737785, + 0.011213724501430988, + 0.02354038879275322 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 6, + 9 + ], + "selected_reward": -0.0011940023396164179 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a towering succubus standing in the center of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the succubus have cascading silver-white hair and massive black wings with an obsidian sheen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the succubus wearing a flowing pure white silk gown that reaches the floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the succubus slowly raise her right arm and twist her wrist to extend her fingers toward the camera?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot slightly handheld with a subtle tremor?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.007632498629391193 + ], + "opposite_logprob_tokens": [ + -4.882632732391357 + ], + "logprob_reward": -0.007632498629391193, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.007632498629391193, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.05213477462530136, + 0.10143721848726273, + 0.05946800857782364, + 0.0915551632642746, + 0.38925600051879883, + 0.07214078307151794, + 0.07208935916423798, + 0.028768224641680717, + 0.0422200970351696, + 0.09093035012483597 + ], + "credit_map": [ + 0.004972083494067192, + 0.004301221575587988, + 0.005508561618626118, + 0.007545835338532925, + 0.010278765112161636, + 0.003925322089344263, + 0.009621956385672092, + 0.005981028079986572, + 0.011324659921228886, + 0.008419951424002647, + 0.010989774018526077, + 0.010664587840437889, + 0.014000504277646542, + 0.007990838028490543, + 0.026498543098568916, + 0.011548359878361225, + 0.0065215271897614, + 0.006486073601990938, + 0.00691471341997385, + 0.006177723873406649, + 0.011522826738655567, + 0.004375741817057133, + 0.011121062561869621, + 0.006348342634737492, + 0.007234282325953245, + 0.009181356057524681, + 0.007184790913015604, + 0.007503662724047899, + 0.035287126898765564, + 0.004494577180594206, + 0.0136277349665761, + 0.007041626609861851, + 0.028865760192275047, + 0.023457517847418785, + 0.030231311917304993, + 0.10206104069948196, + 0.046741824597120285, + 0.018827905878424644, + 0.06812836974859238, + 0.07094226032495499, + 0.006758457515388727, + 0.006611894816160202, + 0.00580467889085412, + 0.007724974304437637, + 0.022642415016889572, + 0.004655005410313606, + 0.010961493477225304, + 0.006981865502893925, + 0.01579335331916809, + 0.006203193683177233, + 0.008892446756362915, + 0.004674541298300028, + 0.00617370055988431, + 0.009181244298815727, + 0.007962455041706562, + 0.013208427466452122, + 0.004297112114727497, + 0.002381827449426055, + 0.006905083078891039, + 0.002844984410330653, + 0.0020260149613022804, + 0.003657165914773941, + 0.002283224370330572, + 0.004372812807559967, + 0.007299979217350483, + 0.002508468460291624, + 0.009651309810578823, + 0.003996473737061024, + 0.0038154905196279287, + 0.004331650212407112, + 0.0035932057071477175, + 0.007023522164672613, + 0.015582801774144173, + 0.005173942539840937, + 0.017413489520549774, + 0.0065652793273329735, + 0.01015079952776432, + 0.010924361646175385, + 0.007746127899736166, + 0.017373550683259964 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 1, + 3 + ], + "selected_reward": -0.007632498629391193 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a towering succubus standing in the center of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the succubus have cascading silver-white hair and massive black wings with an obsidian sheen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the succubus wearing a flowing pure white silk gown that reaches the floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the succubus slowly raise her right arm and twist her wrist to extend her fingers toward the camera?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot slightly handheld with a subtle tremor?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.026347124949097633 + ], + "opposite_logprob_tokens": [ + -3.6513471603393555 + ], + "logprob_reward": -0.026347124949097633, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.026347124949097633, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.37059977650642395, + 0.0758538767695427, + 0.032010313123464584, + 0.07885650545358658, + 0.03763765096664429, + 0.01765047200024128, + 0.026048747822642326, + 0.02756297029554844, + 0.06562948226928711, + 0.2681501507759094 + ], + "credit_map": [ + 0.031852830201387405, + 0.03398927301168442, + 0.0511859767138958, + 0.047737691551446915, + 0.06550222635269165, + 0.017103346064686775, + 0.07342078536748886, + 0.04980766400694847, + 0.007844437845051289, + 0.007288634777069092, + 0.008774803020060062, + 0.007725684903562069, + 0.010749880224466324, + 0.0050517488270998, + 0.020856523886322975, + 0.007562171667814255, + 0.002469112630933523, + 0.0036814333871006966, + 0.0033015848603099585, + 0.0034151608124375343, + 0.008862004615366459, + 0.0015022354200482368, + 0.006024339236319065, + 0.0027544477488845587, + 0.006915202829986811, + 0.008134119212627411, + 0.007769667077809572, + 0.006427846383303404, + 0.025091836228966713, + 0.004672837443649769, + 0.012515660375356674, + 0.007329344749450684, + 0.0027822491247206926, + 0.002336172154173255, + 0.0029861771035939455, + 0.01015288382768631, + 0.0036934928502887487, + 0.0018850585911422968, + 0.0076601943001151085, + 0.006141426041722298, + 0.001779636717401445, + 0.0021964304614812136, + 0.0017343664076179266, + 0.0019911075942218304, + 0.0031348264310508966, + 0.0012778573436662555, + 0.0035550238098949194, + 0.0019812260288745165, + 0.005139267537742853, + 0.0020488547161221504, + 0.0036943780723959208, + 0.001647229422815144, + 0.0018386136507615447, + 0.003755799261853099, + 0.0026249762158840895, + 0.0052996305748820305, + 0.005001472774893045, + 0.0024744959082454443, + 0.005122571252286434, + 0.001787657500244677, + 0.0027405121363699436, + 0.003588712541386485, + 0.0021384975407272577, + 0.004709054250270128, + 0.009136184118688107, + 0.004811425693333149, + 0.01057952456176281, + 0.011532985605299473, + 0.0064397817477583885, + 0.00595518946647644, + 0.00654588220641017, + 0.01062850933521986, + 0.03590574860572815, + 0.01524391770362854, + 0.047312743961811066, + 0.017392311245203018, + 0.03613074868917465, + 0.042442113161087036, + 0.031247783452272415, + 0.04247481748461723 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 3 + ], + "selected_reward": -0.026347124949097633 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a towering succubus standing in the center of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the succubus have cascading silver-white hair and massive black wings with an obsidian sheen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the succubus wearing a flowing pure white silk gown that reaches the floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the succubus slowly raise her right arm and twist her wrist to extend her fingers toward the camera?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot slightly handheld with a subtle tremor?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.34863191843032837 + ], + "opposite_logprob_tokens": [ + -1.2236318588256836 + ], + "logprob_reward": -0.34863191843032837, + "yes_no_margin_reward": 0.8749999403953552, + "token_credit_reward": -0.34863191843032837, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2267865687608719, + 0.11263886094093323, + 0.038939494639635086, + 0.19249096512794495, + 0.09242334216833115, + 0.029960468411445618, + 0.11403186619281769, + 0.044315945357084274, + 0.06084703654050827, + 0.08756546676158905 + ], + "credit_map": [ + 0.020274929702281952, + 0.015524319373071194, + 0.02914789505302906, + 0.025958634912967682, + 0.04321487247943878, + 0.012061195448040962, + 0.04495706409215927, + 0.0356476716697216, + 0.014730226248502731, + 0.0085672065615654, + 0.018489284440875053, + 0.013763467781245708, + 0.01563417911529541, + 0.008811542764306068, + 0.015681343153119087, + 0.016961604356765747, + 0.004329739138484001, + 0.004390456713736057, + 0.003967782948166132, + 0.004310924559831619, + 0.006857119966298342, + 0.0031488866079598665, + 0.0071831801906228065, + 0.004751404281705618, + 0.0158089529722929, + 0.018927309662103653, + 0.019718218594789505, + 0.014638346619904041, + 0.06626018136739731, + 0.009216470643877983, + 0.02975490503013134, + 0.01816658303141594, + 0.008405706845223904, + 0.0072095030918717384, + 0.00701659033074975, + 0.01949370466172695, + 0.013988631777465343, + 0.006704223807901144, + 0.02146412990987301, + 0.00814085640013218, + 0.0033950756769627333, + 0.004161739256232977, + 0.0033359837252646685, + 0.002590324031189084, + 0.005002726335078478, + 0.002909139497205615, + 0.0038174393121153116, + 0.0047480384819209576, + 0.018697338178753853, + 0.008434225805103779, + 0.01981702819466591, + 0.006323897745460272, + 0.010904805734753609, + 0.018434377387166023, + 0.010779174044728279, + 0.0206410214304924, + 0.008138401433825493, + 0.004209703765809536, + 0.007193346042186022, + 0.0029209814965724945, + 0.005575701128691435, + 0.005440461914986372, + 0.003983020782470703, + 0.006854328792542219, + 0.006021422334015369, + 0.003980865702033043, + 0.012740050442516804, + 0.0066071366891264915, + 0.006740259937942028, + 0.0038090720772743225, + 0.008235876448452473, + 0.01271235290914774, + 0.011625764891505241, + 0.006746504455804825, + 0.015986327081918716, + 0.006446820683777332, + 0.01268588937819004, + 0.008717745542526245, + 0.013375899754464626, + 0.01198051031678915 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 6 + ], + "selected_reward": -0.34863191843032837 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000050/group_5/index_tensor([8182, 8182])_group_5_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000050/group_5/index_tensor([8182, 8182])_group_5_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..7f2ee26c9e1e0fb57e5350611b6c02f486882854 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000050/group_5/index_tensor([8182, 8182])_group_5_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "In the distinctive style of a moving oil painting, a low-angle, wide-angle shot begins a slow, deliberate orbit around a lone olive tree. The entire scene is rendered with visible, impasto brushstrokes, giving every surface a tangible, sculpted texture. The ancient tree, the central subject, stands defiantly against a sun-drenched hill, its gnarled, twisted trunk an intricate tapestry of deep amber and cool gray applied with heavy, directional brushstrokes. As the camera gracefully curves, golden afternoon light saturates the frame, capturing the raised edges of the paint and casting long, dramatic, purplish shadows that stretch across the rolling ochre landscape. A soft, unseen breeze stirs, causing the tree's clusters of silvery-green stalks—each bearing a unique wisp of paint—to tremble subtly. Then, in a slow, poetic cascade, each leaf begins to separate, fluttering and spiraling downward, their silvery undersides shimmering in the warm light before coming to rest on the richly textured painted ground, creating a profound atmosphere of tranquility, timeless beauty and gentle decay.", + "seed": 25084, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000050/group_5/index_8182_seed_25084_dp_10_rank_20_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 5, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.022771171107888222, + "vlm_logprob": -0.022771171107888222, + "vlm_yes_no_margin": 4.75, + "vlm_token_credit": -0.022771171107888222, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video rendered in a moving oil painting style with visible impasto brushstrokes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the central subject a lone ancient olive tree with a gnarled, twisted trunk?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera perform a slow, deliberate orbit around the tree from a low-angle, wide-angle perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the tree's leaves separate and flutter downward in a slow cascade, creating silvery undersides?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene set on a sun-drenched hill with long, purplish shadows stretching across an ochre landscape?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.007305928971618414 + ], + "opposite_logprob_tokens": [ + -5.507306098937988 + ], + "logprob_reward": -0.007305928971618414, + "yes_no_margin_reward": 5.5, + "token_credit_reward": -0.007305928971618414, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.19812549650669098, + 0.15189668536186218, + 0.06168753653764725, + 0.037736501544713974, + 0.11873535811901093, + 0.07007052004337311, + 0.12431826442480087, + 0.05497824028134346, + 0.11255823820829391, + 0.06989315152168274 + ], + "credit_map": [ + 0.03275885432958603, + 0.027812574058771133, + 0.020406801253557205, + 0.022727686911821365, + 0.02751835435628891, + 0.0213791336864233, + 0.02344873547554016, + 0.022073352709412575, + 0.018252769485116005, + 0.024480074644088745, + 0.012475943192839622, + 0.021436838433146477, + 0.028922587633132935, + 0.014588721096515656, + 0.01737350970506668, + 0.01436623279005289, + 0.006835128180682659, + 0.011751260608434677, + 0.006211385130882263, + 0.00653200363740325, + 0.011737249791622162, + 0.005762946791946888, + 0.006486313417553902, + 0.006371247116476297, + 0.0049924650229513645, + 0.006887644063681364, + 0.003265442093834281, + 0.004594476893544197, + 0.005328167695552111, + 0.004016450606286526, + 0.0046310219913721085, + 0.004020834341645241, + 0.015189473517239094, + 0.018187567591667175, + 0.008047706447541714, + 0.02222413755953312, + 0.019043998792767525, + 0.011662278324365616, + 0.015795908868312836, + 0.008584295399487019, + 0.008074399083852768, + 0.010046008042991161, + 0.007105235941708088, + 0.009599796496331692, + 0.011506015434861183, + 0.009656810201704502, + 0.007822126150131226, + 0.006260131485760212, + 0.015813972800970078, + 0.0143505884334445, + 0.011186889372766018, + 0.015161179937422276, + 0.018451619893312454, + 0.016313226893544197, + 0.012651778757572174, + 0.020389003679156303, + 0.0047358013689517975, + 0.004424304701387882, + 0.004695408046245575, + 0.003996382467448711, + 0.014755979180335999, + 0.00461996253579855, + 0.004958965349942446, + 0.012791438959538937, + 0.006223524920642376, + 0.015424543991684914, + 0.009148583747446537, + 0.015160584822297096, + 0.02251473255455494, + 0.013929706066846848, + 0.01423611305654049, + 0.015920450910925865, + 0.007334550376981497, + 0.013017823919653893, + 0.006765814032405615, + 0.012771486304700375, + 0.009403751231729984, + 0.0077918157912790775, + 0.006914724130183458, + 0.005893188528716564 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 6 + ], + "selected_reward": -0.007305928971618414 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video rendered in a moving oil painting style with visible impasto brushstrokes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the central subject a lone ancient olive tree with a gnarled, twisted trunk?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera perform a slow, deliberate orbit around the tree from a low-angle, wide-angle perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the tree's leaves separate and flutter downward in a slow cascade, creating silvery undersides?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene set on a sun-drenched hill with long, purplish shadows stretching across an ochre landscape?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0097875464707613 + ], + "opposite_logprob_tokens": [ + -4.634787559509277 + ], + "logprob_reward": -0.0097875464707613, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.0097875464707613, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3926142454147339, + 0.15783417224884033, + 0.1494113653898239, + 0.04646586626768112, + 0.036807768046855927, + 0.04214337095618248, + 0.052601926028728485, + 0.01401260681450367, + 0.06296743452548981, + 0.045141227543354034 + ], + "credit_map": [ + 0.033925190567970276, + 0.04761337488889694, + 0.02655494026839733, + 0.03992521017789841, + 0.04949047788977623, + 0.0432259663939476, + 0.09194746613502502, + 0.05993163585662842, + 0.015239899046719074, + 0.02515561319887638, + 0.011350573040544987, + 0.023964548483490944, + 0.024160319939255714, + 0.01478572841733694, + 0.029638731852173805, + 0.013538764789700508, + 0.021824344992637634, + 0.030839912593364716, + 0.010532760061323643, + 0.018068503588438034, + 0.021476930007338524, + 0.014192184433341026, + 0.020326489582657814, + 0.012150229886174202, + 0.0062875873409211636, + 0.00946200080215931, + 0.00270663108676672, + 0.0064344885759055614, + 0.007079730276018381, + 0.004373050294816494, + 0.006592430640012026, + 0.0035299425944685936, + 0.0045643639750778675, + 0.006022712681442499, + 0.0025868266820907593, + 0.006075190380215645, + 0.005234440788626671, + 0.0033282467629760504, + 0.005641155410557985, + 0.003354830900207162, + 0.0035498396027833223, + 0.007881011813879013, + 0.00293283024802804, + 0.004598391707986593, + 0.005198115482926369, + 0.0046903397887945175, + 0.00399202574044466, + 0.009300812147557735, + 0.0071071963757276535, + 0.010207803919911385, + 0.0026014598552137613, + 0.010042746551334858, + 0.005704097915440798, + 0.006228439975529909, + 0.006464822683483362, + 0.004245360381901264, + 0.0015495026018470526, + 0.0015651355497539043, + 0.0013481369242072105, + 0.0018642644863575697, + 0.0018730622250586748, + 0.001712139928713441, + 0.001655971514992416, + 0.002444393467158079, + 0.005486834794282913, + 0.008735448122024536, + 0.007590930908918381, + 0.007406786549836397, + 0.012371922843158245, + 0.007139774039387703, + 0.006705010775476694, + 0.007530723232775927, + 0.005453044548630714, + 0.0055738650262355804, + 0.006491341162472963, + 0.004792992956936359, + 0.006371418014168739, + 0.005658959969878197, + 0.0064261313527822495, + 0.004373474512249231 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 2 + ], + "selected_reward": -0.0097875464707613 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video rendered in a moving oil painting style with visible impasto brushstrokes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the central subject a lone ancient olive tree with a gnarled, twisted trunk?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera perform a slow, deliberate orbit around the tree from a low-angle, wide-angle perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the tree's leaves separate and flutter downward in a slow cascade, creating silvery undersides?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene set on a sun-drenched hill with long, purplish shadows stretching across an ochre landscape?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005253320559859276 + ], + "opposite_logprob_tokens": [ + -5.255253314971924 + ], + "logprob_reward": -0.005253320559859276, + "yes_no_margin_reward": 5.25, + "token_credit_reward": -0.005253320559859276, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09058477729558945, + 0.1728033721446991, + 0.020658819004893303, + 0.01649513654410839, + 0.10171680897474289, + 0.11869074404239655, + 0.0914875939488411, + 0.05723707377910614, + 0.2516004145145416, + 0.07872532308101654 + ], + "credit_map": [ + 0.007843122817575932, + 0.01200538408011198, + 0.005999533925205469, + 0.01409293431788683, + 0.0199703648686409, + 0.011052877642214298, + 0.013113301247358322, + 0.006507255602627993, + 0.01956452801823616, + 0.02929418534040451, + 0.011394704692065716, + 0.02764919027686119, + 0.02685919590294361, + 0.015688538551330566, + 0.029532084241509438, + 0.012820917181670666, + 0.001924977288581431, + 0.003301488235592842, + 0.001468948321416974, + 0.003151783486828208, + 0.0040580397471785545, + 0.0020168498158454895, + 0.00297395302914083, + 0.0017627761699259281, + 0.002224623691290617, + 0.00322152697481215, + 0.0011773802107200027, + 0.00199192319996655, + 0.0025226250290870667, + 0.001492770272307098, + 0.002250357996672392, + 0.0016139267245307565, + 0.011316312476992607, + 0.015249422751367092, + 0.005962040740996599, + 0.019896360114216805, + 0.0184054933488369, + 0.010561712086200714, + 0.012519203126430511, + 0.0078062633983790874, + 0.01222054660320282, + 0.020958218723535538, + 0.006811619736254215, + 0.017377298325300217, + 0.016733478754758835, + 0.012372152879834175, + 0.01915782317519188, + 0.013059597462415695, + 0.01069850753992796, + 0.018550628796219826, + 0.007307104300707579, + 0.014962620101869106, + 0.008449316956102848, + 0.008179017342627048, + 0.012787430547177792, + 0.010552958585321903, + 0.004072186537086964, + 0.005512730218470097, + 0.0027642431668937206, + 0.004366847220808268, + 0.015304427593946457, + 0.005035689566284418, + 0.005482539068907499, + 0.01469840481877327, + 0.020307261496782303, + 0.041017282754182816, + 0.024525871500372887, + 0.03927721455693245, + 0.048497166484594345, + 0.018358033150434494, + 0.03201930969953537, + 0.02759825624525547, + 0.008274620398879051, + 0.012164433486759663, + 0.006771937943994999, + 0.01557072065770626, + 0.01076467428356409, + 0.010781656950712204, + 0.007479476276785135, + 0.006917798891663551 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 1, + 5 + ], + "selected_reward": -0.005253320559859276 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video rendered in a moving oil painting style with visible impasto brushstrokes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the central subject a lone ancient olive tree with a gnarled, twisted trunk?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera perform a slow, deliberate orbit around the tree from a low-angle, wide-angle perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the tree's leaves separate and flutter downward in a slow cascade, creating silvery undersides?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene set on a sun-drenched hill with long, purplish shadows stretching across an ochre landscape?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.08902353793382645 + ], + "opposite_logprob_tokens": [ + -2.4640235900878906 + ], + "logprob_reward": -0.08902353793382645, + "yes_no_margin_reward": 2.375, + "token_credit_reward": -0.08902353793382645, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15411919355392456, + 0.1251801699399948, + 0.03610004484653473, + 0.069563128054142, + 0.06459455192089081, + 0.08717595040798187, + 0.18534450232982635, + 0.11546650528907776, + 0.14075011014938354, + 0.02170587331056595 + ], + "credit_map": [ + 0.01388539094477892, + 0.018636805936694145, + 0.01150553859770298, + 0.01439813245087862, + 0.026167290285229683, + 0.015941999852657318, + 0.029485182836651802, + 0.0240988340228796, + 0.01492246799170971, + 0.02304280549287796, + 0.006980292499065399, + 0.018695063889026642, + 0.021577101200819016, + 0.011393354274332523, + 0.018263062462210655, + 0.010306020267307758, + 0.004260213579982519, + 0.005655516404658556, + 0.002327540423721075, + 0.005345073528587818, + 0.007607973646372557, + 0.004092380404472351, + 0.00432061031460762, + 0.00249073700979352, + 0.008472993038594723, + 0.014551362954080105, + 0.004001838620752096, + 0.009725317358970642, + 0.011774793267250061, + 0.007159480359405279, + 0.008756060153245926, + 0.005121274385601282, + 0.008338636718690395, + 0.012399036437273026, + 0.006189585663378239, + 0.012501212768256664, + 0.007620574440807104, + 0.005487461574375629, + 0.00823238305747509, + 0.0038256507832556963, + 0.011755052022635937, + 0.01144520752131939, + 0.0051549384370446205, + 0.009143669158220291, + 0.02013375796377659, + 0.013849596492946148, + 0.006684788968414068, + 0.009008935652673244, + 0.024090424180030823, + 0.026462139561772346, + 0.01293952390551567, + 0.05207178741693497, + 0.015097890049219131, + 0.015260545536875725, + 0.0264248326420784, + 0.012997336685657501, + 0.013943892903625965, + 0.01646159216761589, + 0.0072000110521912575, + 0.016103515401482582, + 0.017634039744734764, + 0.008812951855361462, + 0.019881727173924446, + 0.015428771264851093, + 0.011744227260351181, + 0.024104351177811623, + 0.0187576524913311, + 0.02493845857679844, + 0.0204755999147892, + 0.009409487247467041, + 0.01862005889415741, + 0.012700274586677551, + 0.002777448855340481, + 0.002547169802710414, + 0.0042690313421189785, + 0.0020295416470617056, + 0.002655635355040431, + 0.002884174231439829, + 0.002616270910948515, + 0.0019265994196757674 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 0, + 8 + ], + "selected_reward": -0.08902353793382645 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video rendered in a moving oil painting style with visible impasto brushstrokes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the central subject a lone ancient olive tree with a gnarled, twisted trunk?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera perform a slow, deliberate orbit around the tree from a low-angle, wide-angle perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the tree's leaves separate and flutter downward in a slow cascade, creating silvery undersides?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene set on a sun-drenched hill with long, purplish shadows stretching across an ochre landscape?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0024855216033756733 + ], + "opposite_logprob_tokens": [ + -6.002485752105713 + ], + "logprob_reward": -0.0024855216033756733, + "yes_no_margin_reward": 6.0, + "token_credit_reward": -0.0024855216033756733, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.23601043224334717, + 0.24800816178321838, + 0.04539286345243454, + 0.028117109090089798, + 0.050762541592121124, + 0.06985043734312057, + 0.0768849328160286, + 0.08484436571598053, + 0.08918453752994537, + 0.07094462960958481 + ], + "credit_map": [ + 0.023366378620266914, + 0.03116784058511257, + 0.010683379136025906, + 0.0175681971013546, + 0.06941995769739151, + 0.040847450494766235, + 0.022472701966762543, + 0.020484531298279762, + 0.02888523042201996, + 0.04197993129491806, + 0.017491118982434273, + 0.03602396324276924, + 0.04579657316207886, + 0.020846839994192123, + 0.03857317194342613, + 0.018411338329315186, + 0.00685928575694561, + 0.006888640578836203, + 0.00400161650031805, + 0.0047454992309212685, + 0.00900137796998024, + 0.0046986062079668045, + 0.005396004766225815, + 0.003801831742748618, + 0.003780996659770608, + 0.005323787219822407, + 0.0021259328350424767, + 0.004949169233441353, + 0.002896189922466874, + 0.0026887294370681047, + 0.003563560778275132, + 0.0027887437026947737, + 0.006889781914651394, + 0.005997007247060537, + 0.004277377855032682, + 0.004147674888372421, + 0.013955598697066307, + 0.00565419252961874, + 0.005645455792546272, + 0.004195451736450195, + 0.006783852819353342, + 0.007703679148107767, + 0.005106493830680847, + 0.006371417548507452, + 0.009816274046897888, + 0.013742947019636631, + 0.007762215565890074, + 0.012563562951982021, + 0.010369248688220978, + 0.00813551526516676, + 0.007935313507914543, + 0.007963947020471096, + 0.013043541461229324, + 0.010203700512647629, + 0.007877886295318604, + 0.011355778202414513, + 0.004523597192019224, + 0.006029467098414898, + 0.0026750401593744755, + 0.005563217680901289, + 0.029961466789245605, + 0.008092381991446018, + 0.003942643757909536, + 0.024056555703282356, + 0.006519358139485121, + 0.010343927890062332, + 0.00827100407332182, + 0.009795253165066242, + 0.017465420067310333, + 0.01145525649189949, + 0.014709089882671833, + 0.010625231079757214, + 0.009167869575321674, + 0.012119363993406296, + 0.007244161795824766, + 0.006802380550652742, + 0.011653056368231773, + 0.00897290650755167, + 0.00788650382310152, + 0.007098380010575056 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 8 + ], + "selected_reward": -0.0024855216033756733 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "In the distinctive style of a moving oil painting, a low-angle, wide-angle shot begins a slow, deliberate orbit around a lone olive tree. The entire scene is rendered with visible, impasto brushstrokes, giving every surface a tangible, sculpted texture. The ancient tree, the central subject, stands defiantly against a sun-drenched hill, its gnarled, twisted trunk an intricate tapestry of deep amber and cool gray applied with heavy, directional brushstrokes. As the camera gracefully curves, golden afternoon light saturates the frame, capturing the raised edges of the paint and casting long, dramatic, purplish shadows that stretch across the rolling ochre landscape. A soft, unseen breeze stirs, causing the tree's clusters of silvery-green stalks—each bearing a unique wisp of paint—to tremble subtly. Then, in a slow, poetic cascade, each leaf begins to separate, fluttering and spiraling downward, their silvery undersides shimmering in the warm light before coming to rest on the richly textured painted ground, creating a profound atmosphere of tranquility, timeless beauty and gentle decay.", + "seed": 25085, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000050/group_5/index_8182_seed_25085_dp_10_rank_20_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 5, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.017090207291767003, + "vlm_logprob": -0.017090207291767003, + "vlm_yes_no_margin": 5.0, + "vlm_token_credit": -0.017090207291767003, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video rendered in a moving oil painting style with visible impasto brushstrokes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the central subject a lone ancient olive tree with a gnarled, twisted trunk?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera perform a slow, deliberate orbit around the tree from a low-angle, wide-angle perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the tree's leaves separate and flutter downward in a slow cascade, creating silvery undersides?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene set on a sun-drenched hill with long, purplish shadows stretching across an ochre landscape?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.003916214685887098 + ], + "opposite_logprob_tokens": [ + -6.128916263580322 + ], + "logprob_reward": -0.003916214685887098, + "yes_no_margin_reward": 6.125, + "token_credit_reward": -0.003916214685887098, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.29937729239463806, + 0.03115849755704403, + 0.036577966064214706, + 0.04420989379286766, + 0.018423784524202347, + 0.013883274048566818, + 0.48964011669158936, + 0.012573523446917534, + 0.03378698602318764, + 0.020368685945868492 + ], + "credit_map": [ + 0.020647717639803886, + 0.025862229987978935, + 0.09311896562576294, + 0.018757645040750504, + 0.015412876382470131, + 0.10255924612283707, + 0.009212344884872437, + 0.013806264847517014, + 0.003778037615120411, + 0.006424633786082268, + 0.00399331608787179, + 0.004188702441751957, + 0.0027237185277044773, + 0.003192383097484708, + 0.003307388164103031, + 0.0035503176040947437, + 0.0032057755161076784, + 0.005537772551178932, + 0.005349200684577227, + 0.004040124360471964, + 0.004533948376774788, + 0.0031130551360547543, + 0.005630710627883673, + 0.005167381837964058, + 0.004064485896378756, + 0.009013094007968903, + 0.005346416495740414, + 0.006574660539627075, + 0.004584049340337515, + 0.0037590533029288054, + 0.004890195559710264, + 0.005977938417345285, + 0.002047904534265399, + 0.0019073145231232047, + 0.0025390249211341143, + 0.002802931470796466, + 0.001505256979726255, + 0.002965977182611823, + 0.002009630436077714, + 0.002645745174959302, + 0.0012138615129515529, + 0.0016583895776420832, + 0.002076428383588791, + 0.0019059767946600914, + 0.002116191666573286, + 0.0014947676099836826, + 0.0019092067377641797, + 0.0015084516489878297, + 0.060414526611566544, + 0.06286890059709549, + 0.04772031679749489, + 0.07029309868812561, + 0.03934069350361824, + 0.09859327226877213, + 0.050384726375341415, + 0.060024604201316833, + 0.0011106322053819895, + 0.0012147368397563696, + 0.0033109248615801334, + 0.0010596534702926874, + 0.0008842122042551637, + 0.0015499837463721633, + 0.0009031404624693096, + 0.0025402395986020565, + 0.00852140597999096, + 0.0023130816407501698, + 0.003051714738830924, + 0.0023114390205591917, + 0.0018166278023272753, + 0.010870972648262978, + 0.0019502976210787892, + 0.002951446920633316, + 0.0023086988367140293, + 0.0018021509749814868, + 0.002525890013203025, + 0.0016670719487592578, + 0.0017648310167714953, + 0.005049853585660458, + 0.0013946382096037269, + 0.0038555506616830826 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 0, + 3 + ], + "selected_reward": -0.003916214685887098 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video rendered in a moving oil painting style with visible impasto brushstrokes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the central subject a lone ancient olive tree with a gnarled, twisted trunk?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera perform a slow, deliberate orbit around the tree from a low-angle, wide-angle perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the tree's leaves separate and flutter downward in a slow cascade, creating silvery undersides?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene set on a sun-drenched hill with long, purplish shadows stretching across an ochre landscape?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0028229409363120794 + ], + "opposite_logprob_tokens": [ + -5.8778228759765625 + ], + "logprob_reward": -0.0028229409363120794, + "yes_no_margin_reward": 5.875, + "token_credit_reward": -0.0028229409363120794, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2918090522289276, + 0.1505943387746811, + 0.11233633756637573, + 0.0385267436504364, + 0.031483929604291916, + 0.08433376252651215, + 0.14812271296977997, + 0.031188352033495903, + 0.07929271459579468, + 0.0323120579123497 + ], + "credit_map": [ + 0.020820586010813713, + 0.03317556530237198, + 0.08082298934459686, + 0.016709640622138977, + 0.021092845126986504, + 0.08975299447774887, + 0.012337825261056423, + 0.017096633091568947, + 0.016177931800484657, + 0.03299395740032196, + 0.02076142467558384, + 0.014610390178859234, + 0.01852099783718586, + 0.0162592064589262, + 0.013788926415145397, + 0.01748151145875454, + 0.012323460541665554, + 0.02218531258404255, + 0.015809588134288788, + 0.00941804051399231, + 0.015431794337928295, + 0.011767786927521229, + 0.010409983806312084, + 0.014990374445915222, + 0.003955317195504904, + 0.007119889371097088, + 0.005583362653851509, + 0.004482506774365902, + 0.005327919963747263, + 0.0033413544297218323, + 0.003877934766933322, + 0.0048384596593678, + 0.0024765736889094114, + 0.006119986064732075, + 0.0039861868135631084, + 0.004122457001358271, + 0.004465634934604168, + 0.0024923549499362707, + 0.003145346650853753, + 0.004675390664488077, + 0.008823202922940254, + 0.014418835751712322, + 0.008384335786104202, + 0.006725731771439314, + 0.015597175806760788, + 0.00866896566003561, + 0.012100991792976856, + 0.009614533744752407, + 0.013610749505460262, + 0.020615749061107635, + 0.020250139757990837, + 0.02515653893351555, + 0.01201168168336153, + 0.01837725006043911, + 0.019489407539367676, + 0.018611202016472816, + 0.0042880624532699585, + 0.0032269929070025682, + 0.00535791227594018, + 0.002283546607941389, + 0.002592981792986393, + 0.004872472491115332, + 0.0023622307926416397, + 0.006204154342412949, + 0.02658732980489731, + 0.0053695873357355595, + 0.005991800222545862, + 0.002702348632737994, + 0.0031689468305557966, + 0.020527979359030724, + 0.003669766476377845, + 0.011274958029389381, + 0.004163787700235844, + 0.0029294437263160944, + 0.00661174301058054, + 0.004521268419921398, + 0.002650693990290165, + 0.00446162186563015, + 0.00267776008695364, + 0.00429574167355895 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 6 + ], + "selected_reward": -0.0028229409363120794 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video rendered in a moving oil painting style with visible impasto brushstrokes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the central subject a lone ancient olive tree with a gnarled, twisted trunk?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera perform a slow, deliberate orbit around the tree from a low-angle, wide-angle perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the tree's leaves separate and flutter downward in a slow cascade, creating silvery undersides?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene set on a sun-drenched hill with long, purplish shadows stretching across an ochre landscape?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.014193536713719368 + ], + "opposite_logprob_tokens": [ + -4.264193534851074 + ], + "logprob_reward": -0.014193536713719368, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.014193536713719368, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10280724614858627, + 0.04882345721125603, + 0.058011095970869064, + 0.04040437936782837, + 0.035404521971940994, + 0.06285113096237183, + 0.26140424609184265, + 0.14873363077640533, + 0.18432539701461792, + 0.05723490193486214 + ], + "credit_map": [ + 0.009779440239071846, + 0.00913927610963583, + 0.019541870802640915, + 0.01116282120347023, + 0.010575591586530209, + 0.01860916242003441, + 0.008604693226516247, + 0.015394396148622036, + 0.004945028107613325, + 0.014512153342366219, + 0.006769850384443998, + 0.00405684020370245, + 0.005150964483618736, + 0.004474293906241655, + 0.004236744716763496, + 0.004677581135183573, + 0.005378469824790955, + 0.011344486847519875, + 0.008395105600357056, + 0.00569703197106719, + 0.0066278777085244656, + 0.004532204940915108, + 0.007278568111360073, + 0.008757361210882664, + 0.004288253840059042, + 0.007982424460351467, + 0.00527736684307456, + 0.004408358596265316, + 0.005531701259315014, + 0.0029115574434399605, + 0.005786258261650801, + 0.004218460526317358, + 0.003136836225166917, + 0.006008116062730551, + 0.004840908106416464, + 0.005056786350905895, + 0.004020149353891611, + 0.0033445795997977257, + 0.00411924347281456, + 0.004877906758338213, + 0.00712073128670454, + 0.00946734007447958, + 0.007638760842382908, + 0.006585527211427689, + 0.009347092360258102, + 0.005964370910078287, + 0.008951060473918915, + 0.007776247803121805, + 0.04519426450133324, + 0.027118198573589325, + 0.02233518473803997, + 0.02741197496652603, + 0.019632011651992798, + 0.06834664195775986, + 0.01921813376247883, + 0.0321478396654129, + 0.03256813436746597, + 0.0065450649708509445, + 0.027584994211792946, + 0.004943685606122017, + 0.004955660086125135, + 0.01403274480253458, + 0.004574834369122982, + 0.05352853238582611, + 0.06395486742258072, + 0.009305869229137897, + 0.00934681948274374, + 0.004728988744318485, + 0.006411428097635508, + 0.0681658387184143, + 0.005559688434004784, + 0.016851913183927536, + 0.003651244333013892, + 0.00435419287532568, + 0.009523481130599976, + 0.006278615910559893, + 0.003892505308613181, + 0.011008047498762608, + 0.004500680603086948, + 0.014026139862835407 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 8, + 7 + ], + "selected_reward": -0.014193536713719368 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video rendered in a moving oil painting style with visible impasto brushstrokes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the central subject a lone ancient olive tree with a gnarled, twisted trunk?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera perform a slow, deliberate orbit around the tree from a low-angle, wide-angle perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the tree's leaves separate and flutter downward in a slow cascade, creating silvery undersides?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene set on a sun-drenched hill with long, purplish shadows stretching across an ochre landscape?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.06203282251954079 + ], + "opposite_logprob_tokens": [ + -2.81203293800354 + ], + "logprob_reward": -0.06203282251954079, + "yes_no_margin_reward": 2.75, + "token_credit_reward": -0.06203282251954079, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08432883024215698, + 0.1445487141609192, + 0.07828119397163391, + 0.2635120749473572, + 0.012152209877967834, + 0.11769143491983414, + 0.1441463977098465, + 0.10203157365322113, + 0.0330016054213047, + 0.020305991172790527 + ], + "credit_map": [ + 0.01126574445515871, + 0.015692638233304024, + 0.010276246815919876, + 0.008175958879292011, + 0.012607213109731674, + 0.00929173268377781, + 0.006424645893275738, + 0.010594647377729416, + 0.014198604971170425, + 0.04487941041588783, + 0.019712483510375023, + 0.012489675544202328, + 0.014168866910040379, + 0.01228839810937643, + 0.012814340181648731, + 0.013996935449540615, + 0.008462462574243546, + 0.017911532893776894, + 0.01111386064440012, + 0.007877287454903126, + 0.00779536459594965, + 0.00641716318204999, + 0.010386011563241482, + 0.008317516185343266, + 0.02442716807126999, + 0.06406086683273315, + 0.033357370644807816, + 0.03670200705528259, + 0.02628755196928978, + 0.019831698387861252, + 0.027785463258624077, + 0.031059956178069115, + 0.0010801509488373995, + 0.0018857966642826796, + 0.0016016424633562565, + 0.0015226986724883318, + 0.0018538612639531493, + 0.0011976255336776376, + 0.0013864005450159311, + 0.0016240329714491963, + 0.01265651360154152, + 0.018871113657951355, + 0.012470356188714504, + 0.012863007374107838, + 0.0196243729442358, + 0.011593823321163654, + 0.017644226551055908, + 0.011968025006353855, + 0.037034716457128525, + 0.013585152104496956, + 0.014602495357394218, + 0.009280702099204063, + 0.0076753213070333, + 0.03813682496547699, + 0.008955715224146843, + 0.01487548928707838, + 0.027889415621757507, + 0.009900925680994987, + 0.00977343413978815, + 0.005821865051984787, + 0.009295187890529633, + 0.008275344967842102, + 0.006384197622537613, + 0.024691196158528328, + 0.002396949799731374, + 0.0028358171693980694, + 0.0035160351544618607, + 0.003537273034453392, + 0.0027949344366788864, + 0.007703615818172693, + 0.003387670498341322, + 0.006829308345913887, + 0.0018801613477990031, + 0.0022268942557275295, + 0.0030911611393094063, + 0.0023399374913424253, + 0.0025473576970398426, + 0.0033013951033353806, + 0.0015486417105421424, + 0.003370442660525441 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 1, + 6 + ], + "selected_reward": -0.06203282251954079 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video rendered in a moving oil painting style with visible impasto brushstrokes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the central subject a lone ancient olive tree with a gnarled, twisted trunk?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera perform a slow, deliberate orbit around the tree from a low-angle, wide-angle perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the tree's leaves separate and flutter downward in a slow cascade, creating silvery undersides?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene set on a sun-drenched hill with long, purplish shadows stretching across an ochre landscape?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0024855216033756733 + ], + "opposite_logprob_tokens": [ + -6.002485752105713 + ], + "logprob_reward": -0.0024855216033756733, + "yes_no_margin_reward": 6.0, + "token_credit_reward": -0.0024855216033756733, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3595142960548401, + 0.08376839011907578, + 0.0567031167447567, + 0.0303661972284317, + 0.027218062430620193, + 0.09641436487436295, + 0.19223161041736603, + 0.07345202565193176, + 0.04609638825058937, + 0.034235499799251556 + ], + "credit_map": [ + 0.028620334342122078, + 0.04057341441512108, + 0.08863726258277893, + 0.03400582820177078, + 0.027149343863129616, + 0.0969809889793396, + 0.019134385511279106, + 0.02441275864839554, + 0.010265526361763477, + 0.01992039568722248, + 0.011507553048431873, + 0.00997367687523365, + 0.007340509444475174, + 0.00872060377150774, + 0.006765803322196007, + 0.00927432719618082, + 0.007691065315157175, + 0.00971351470798254, + 0.008792086504399776, + 0.00724892457947135, + 0.004858361091464758, + 0.0057752178981900215, + 0.005033295601606369, + 0.00759065430611372, + 0.0032295447308570147, + 0.004621175117790699, + 0.004201261792331934, + 0.002825746778398752, + 0.004047430586069822, + 0.003050662809982896, + 0.004128982312977314, + 0.004261394962668419, + 0.002152072498574853, + 0.0027631029952317476, + 0.004511742386966944, + 0.00390968844294548, + 0.0037391758523881435, + 0.0026956857182085514, + 0.0028075946029275656, + 0.00463900389149785, + 0.012935459613800049, + 0.016480397433042526, + 0.010745035484433174, + 0.007240836508572102, + 0.014156212098896503, + 0.01083285640925169, + 0.012496434152126312, + 0.011527142487466335, + 0.03751673176884651, + 0.028946679085493088, + 0.021752094849944115, + 0.027679748833179474, + 0.015088765881955624, + 0.019489560276269913, + 0.019823716953396797, + 0.021934328600764275, + 0.01877831108868122, + 0.005520680453628302, + 0.0078104231506586075, + 0.003941362723708153, + 0.0037383383605629206, + 0.006629596464335918, + 0.0037312167696654797, + 0.02330210618674755, + 0.010738302953541279, + 0.0039026206359267235, + 0.005742260254919529, + 0.0038162702694535255, + 0.0030188814271241426, + 0.011107736267149448, + 0.003300296375527978, + 0.004470025654882193, + 0.004463664256036282, + 0.003459525527432561, + 0.006874409969896078, + 0.003551329020410776, + 0.0029464594554156065, + 0.004676247015595436, + 0.003259852761402726, + 0.005004015751183033 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 6, + 5 + ], + "selected_reward": -0.0024855216033756733 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "In the distinctive style of a moving oil painting, a low-angle, wide-angle shot begins a slow, deliberate orbit around a lone olive tree. The entire scene is rendered with visible, impasto brushstrokes, giving every surface a tangible, sculpted texture. The ancient tree, the central subject, stands defiantly against a sun-drenched hill, its gnarled, twisted trunk an intricate tapestry of deep amber and cool gray applied with heavy, directional brushstrokes. As the camera gracefully curves, golden afternoon light saturates the frame, capturing the raised edges of the paint and casting long, dramatic, purplish shadows that stretch across the rolling ochre landscape. A soft, unseen breeze stirs, causing the tree's clusters of silvery-green stalks—each bearing a unique wisp of paint—to tremble subtly. Then, in a slow, poetic cascade, each leaf begins to separate, fluttering and spiraling downward, their silvery undersides shimmering in the warm light before coming to rest on the richly textured painted ground, creating a profound atmosphere of tranquility, timeless beauty and gentle decay.", + "seed": 25086, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000050/group_5/index_8182_seed_25086_dp_11_rank_22_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 5, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.029248665762133898, + "vlm_logprob": -0.029248665762133898, + "vlm_yes_no_margin": 4.925, + "vlm_token_credit": -0.029248665762133898, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video rendered in a moving oil painting style with visible impasto brushstrokes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the central subject a lone ancient olive tree with a gnarled, twisted trunk?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera perform a slow, deliberate orbit around the tree from a low-angle, wide-angle perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the tree's leaves separate and flutter downward in a slow cascade, creating silvery undersides?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene set on a sun-drenched hill with long, purplish shadows stretching across an ochre landscape?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.00845743902027607 + ], + "opposite_logprob_tokens": [ + -5.133457660675049 + ], + "logprob_reward": -0.00845743902027607, + "yes_no_margin_reward": 5.125, + "token_credit_reward": -0.00845743902027607, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.33659353852272034, + 0.035297542810440063, + 0.03700152784585953, + 0.033562880009412766, + 0.03749021515250206, + 0.18669815361499786, + 0.032479427754879, + 0.24619054794311523, + 0.03567939251661301, + 0.0190067570656538 + ], + "credit_map": [ + 0.024971451610326767, + 0.018751688301563263, + 0.10841461271047592, + 0.02837785705924034, + 0.022053668275475502, + 0.11131171137094498, + 0.00889306515455246, + 0.0138194989413023, + 0.0034677167423069477, + 0.004224476404488087, + 0.005286231637001038, + 0.0061490233056247234, + 0.00400240533053875, + 0.006585143972188234, + 0.0027037542313337326, + 0.0028787932824343443, + 0.0055002435110509396, + 0.0065537914633750916, + 0.005479907151311636, + 0.005263577215373516, + 0.0034341809805482626, + 0.004517503082752228, + 0.0024136980064213276, + 0.003838625270873308, + 0.004392610397189856, + 0.007203445304185152, + 0.003993655554950237, + 0.0037849401123821735, + 0.002837380627170205, + 0.00484296353533864, + 0.0024533637333661318, + 0.004054524470120668, + 0.004143626894801855, + 0.004477467853575945, + 0.006973291747272015, + 0.002776914043352008, + 0.0037575755268335342, + 0.005210540257394314, + 0.00443970737978816, + 0.005711092613637447, + 0.027493733912706375, + 0.011783597059547901, + 0.04299048334360123, + 0.016391564160585403, + 0.021101024001836777, + 0.027993833646178246, + 0.019642990082502365, + 0.01930094324052334, + 0.003495003329589963, + 0.00241765845566988, + 0.004653527867048979, + 0.0028289654292166233, + 0.003862903919070959, + 0.004574847407639027, + 0.004456403199583292, + 0.006190119776874781, + 0.01634879782795906, + 0.01829761452972889, + 0.06528206169605255, + 0.02284289337694645, + 0.023142220452427864, + 0.06523285806179047, + 0.019131788983941078, + 0.01591232791543007, + 0.0032974546775221825, + 0.0023421908263117075, + 0.0057864123955369, + 0.00281756860204041, + 0.0037669476587325335, + 0.008408726193010807, + 0.005154850427061319, + 0.004105244763195515, + 0.002080377424135804, + 0.0020505255088210106, + 0.0026726892683655024, + 0.002718826523050666, + 0.0025763672310858965, + 0.00242151552811265, + 0.002869007410481572, + 0.0016174488700926304 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 7, + 5 + ], + "selected_reward": -0.00845743902027607 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video rendered in a moving oil painting style with visible impasto brushstrokes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the central subject a lone ancient olive tree with a gnarled, twisted trunk?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera perform a slow, deliberate orbit around the tree from a low-angle, wide-angle perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the tree's leaves separate and flutter downward in a slow cascade, creating silvery undersides?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene set on a sun-drenched hill with long, purplish shadows stretching across an ochre landscape?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.002495748223736882 + ], + "opposite_logprob_tokens": [ + -6.002495765686035 + ], + "logprob_reward": -0.002495748223736882, + "yes_no_margin_reward": 6.0, + "token_credit_reward": -0.002495748223736882, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.31998491287231445, + 0.1925608366727829, + 0.10568039864301682, + 0.03130071610212326, + 0.030467113479971886, + 0.0865887701511383, + 0.024502065032720566, + 0.17354229092597961, + 0.026619749143719673, + 0.008753226138651371 + ], + "credit_map": [ + 0.018766319379210472, + 0.014043749310076237, + 0.11545061320066452, + 0.022793704643845558, + 0.01638462394475937, + 0.11742140352725983, + 0.00616303039714694, + 0.008961456827819347, + 0.016909733414649963, + 0.02914385311305523, + 0.029009491205215454, + 0.045381683856248856, + 0.025345781818032265, + 0.023982055485248566, + 0.010401695966720581, + 0.012386512011289597, + 0.010427729226648808, + 0.023045053705573082, + 0.013422856107354164, + 0.014559650793671608, + 0.014357408508658409, + 0.012285806238651276, + 0.007009231019765139, + 0.010572643019258976, + 0.003580022370442748, + 0.005037799943238497, + 0.003915450535714626, + 0.0081173125654459, + 0.00216846214607358, + 0.003253003116697073, + 0.002277708612382412, + 0.002950957277789712, + 0.0049554696306586266, + 0.002530396683141589, + 0.006612046621739864, + 0.0028950434643775225, + 0.0034234635531902313, + 0.00494141411036253, + 0.0021744989790022373, + 0.0029347767122089863, + 0.012880455702543259, + 0.005583086051046848, + 0.019810503348708153, + 0.007801563013345003, + 0.010298035107553005, + 0.012541926465928555, + 0.008912024088203907, + 0.008761175908148289, + 0.0023464227560907602, + 0.001920853042975068, + 0.004435988608747721, + 0.00206940365023911, + 0.003710411489009857, + 0.0029495793860405684, + 0.003036791691556573, + 0.004032612778246403, + 0.012407735921442509, + 0.014695331454277039, + 0.043386176228523254, + 0.016696440055966377, + 0.017082154750823975, + 0.04160261154174805, + 0.015326261520385742, + 0.012345563620328903, + 0.0022674081847071648, + 0.0028056285809725523, + 0.004610867705196142, + 0.0034723549615591764, + 0.004412794951349497, + 0.00409052474424243, + 0.003180817933753133, + 0.0017793510342016816, + 0.001120801898650825, + 0.0007416768348775804, + 0.0015294330660253763, + 0.0011944721918553114, + 0.0009183836518786848, + 0.0013135401532053947, + 0.0010934659512713552, + 0.0008414514595642686 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 7 + ], + "selected_reward": -0.002495748223736882 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video rendered in a moving oil painting style with visible impasto brushstrokes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the central subject a lone ancient olive tree with a gnarled, twisted trunk?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera perform a slow, deliberate orbit around the tree from a low-angle, wide-angle perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the tree's leaves separate and flutter downward in a slow cascade, creating silvery undersides?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene set on a sun-drenched hill with long, purplish shadows stretching across an ochre landscape?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.006737728603184223 + ], + "opposite_logprob_tokens": [ + -5.00673770904541 + ], + "logprob_reward": -0.006737728603184223, + "yes_no_margin_reward": 5.0, + "token_credit_reward": -0.006737728603184223, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2771545350551605, + 0.12237633019685745, + 0.053880803287029266, + 0.06714097410440445, + 0.05362137779593468, + 0.14709977805614471, + 0.055616267025470734, + 0.17742109298706055, + 0.020161474123597145, + 0.02552737481892109 + ], + "credit_map": [ + 0.018241634592413902, + 0.01786596141755581, + 0.08374077081680298, + 0.03114224597811699, + 0.02033948339521885, + 0.08426520228385925, + 0.008993403986096382, + 0.012565814889967442, + 0.022373400628566742, + 0.012533971108496189, + 0.01746082305908203, + 0.023785429075360298, + 0.00953473336994648, + 0.018319692462682724, + 0.008055823855102062, + 0.010312451981008053, + 0.006926195230334997, + 0.008825269527733326, + 0.008194519206881523, + 0.007164412643760443, + 0.004254243336617947, + 0.007992975413799286, + 0.003864564932882786, + 0.00665862113237381, + 0.009516957215964794, + 0.01105971448123455, + 0.008229835890233517, + 0.01025693491101265, + 0.006075307261198759, + 0.009027128107845783, + 0.00574920792132616, + 0.007225890178233385, + 0.006803763564676046, + 0.008473978377878666, + 0.008595605380833149, + 0.005756455473601818, + 0.004846867639571428, + 0.007964392192661762, + 0.004569281358271837, + 0.006611029617488384, + 0.020755700767040253, + 0.010928279720246792, + 0.030161604285240173, + 0.014506901614367962, + 0.01703464984893799, + 0.01940474472939968, + 0.018014995381236076, + 0.01629290170967579, + 0.00600068736821413, + 0.005551601760089397, + 0.009555202908813953, + 0.005313684698194265, + 0.007421884220093489, + 0.007797019090503454, + 0.005241000559180975, + 0.008735185489058495, + 0.010786019265651703, + 0.012837222777307034, + 0.047128450125455856, + 0.016782771795988083, + 0.016637852415442467, + 0.049451474100351334, + 0.013418300077319145, + 0.010379004292190075, + 0.0013754034880548716, + 0.0016014236025512218, + 0.003405382391065359, + 0.002181218471378088, + 0.0019561832305043936, + 0.002323530614376068, + 0.0036639668978750706, + 0.0036543665919452906, + 0.0023415416944772005, + 0.0018731310265138745, + 0.0053779613226652145, + 0.0026470867451280355, + 0.002765658311545849, + 0.0059946514666080475, + 0.0024971100501716137, + 0.002030235016718507 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 7, + 5 + ], + "selected_reward": -0.006737728603184223 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video rendered in a moving oil painting style with visible impasto brushstrokes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the central subject a lone ancient olive tree with a gnarled, twisted trunk?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera perform a slow, deliberate orbit around the tree from a low-angle, wide-angle perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the tree's leaves separate and flutter downward in a slow cascade, creating silvery undersides?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene set on a sun-drenched hill with long, purplish shadows stretching across an ochre landscape?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.12704186141490936 + ], + "opposite_logprob_tokens": [ + -2.127041816711426 + ], + "logprob_reward": -0.12704186141490936, + "yes_no_margin_reward": 2.0, + "token_credit_reward": -0.12704186141490936, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3865644931793213, + 0.053811214864254, + 0.032250914722681046, + 0.043277841061353683, + 0.04774220287799835, + 0.04742535576224327, + 0.04519379884004593, + 0.2985054850578308, + 0.03130603954195976, + 0.013922672718763351 + ], + "credit_map": [ + 0.024887332692742348, + 0.018376825377345085, + 0.140583798289299, + 0.018712490797042847, + 0.020592963322997093, + 0.14166481792926788, + 0.009576873853802681, + 0.012169424444437027, + 0.0049190097488462925, + 0.006547382567077875, + 0.008768431842327118, + 0.011147542856633663, + 0.006127743516117334, + 0.008620849810540676, + 0.0037837394047528505, + 0.0038965146522969007, + 0.005279442761093378, + 0.004125202540308237, + 0.004000842571258545, + 0.0038884219247847795, + 0.0031555762980133295, + 0.004640704952180386, + 0.002255005994811654, + 0.004905716981738806, + 0.005075538065284491, + 0.009387589059770107, + 0.004737709648907185, + 0.007499124389141798, + 0.004088240675628185, + 0.005019966047257185, + 0.002763643627986312, + 0.004706029314547777, + 0.0066040619276463985, + 0.00691035483032465, + 0.008107539266347885, + 0.0070496187545359135, + 0.004400419071316719, + 0.006472654175013304, + 0.0035160828847438097, + 0.00468146987259388, + 0.006796356290578842, + 0.0025558324996382, + 0.009498059749603271, + 0.004316217266023159, + 0.005311798304319382, + 0.006175922695547342, + 0.006172853987663984, + 0.006598315667361021, + 0.004916731268167496, + 0.0033232884015887976, + 0.006945163477212191, + 0.003994991537183523, + 0.005718827713280916, + 0.0057054683566093445, + 0.005986710544675589, + 0.008602617308497429, + 0.02055797167122364, + 0.024186039343476295, + 0.07574035227298737, + 0.02888038009405136, + 0.02960911951959133, + 0.0743396207690239, + 0.024724746122956276, + 0.0204672422260046, + 0.003391390899196267, + 0.002942661987617612, + 0.004855781327933073, + 0.003230849513784051, + 0.003399898298084736, + 0.004702869337052107, + 0.005196825601160526, + 0.0035857611801475286, + 0.0013202500995248556, + 0.0013374541886150837, + 0.002241195645183325, + 0.0017509125173091888, + 0.0018017352558672428, + 0.001928576035425067, + 0.0022472278214991093, + 0.0012953203404322267 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 7, + 1 + ], + "selected_reward": -0.12704186141490936 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video rendered in a moving oil painting style with visible impasto brushstrokes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the central subject a lone ancient olive tree with a gnarled, twisted trunk?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera perform a slow, deliberate orbit around the tree from a low-angle, wide-angle perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the tree's leaves separate and flutter downward in a slow cascade, creating silvery undersides?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene set on a sun-drenched hill with long, purplish shadows stretching across an ochre landscape?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.001510551548562944 + ], + "opposite_logprob_tokens": [ + -6.5015106201171875 + ], + "logprob_reward": -0.001510551548562944, + "yes_no_margin_reward": 6.5, + "token_credit_reward": -0.001510551548562944, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.537841796875, + 0.0642770528793335, + 0.05180349200963974, + 0.045120228081941605, + 0.03240291774272919, + 0.050067655742168427, + 0.05350440740585327, + 0.09903370589017868, + 0.038394179195165634, + 0.027554558590054512 + ], + "credit_map": [ + 0.02623988501727581, + 0.018141627311706543, + 0.20940011739730835, + 0.030174734070897102, + 0.018220417201519012, + 0.2121264934539795, + 0.009443702176213264, + 0.014094827696681023, + 0.01558079943060875, + 0.007615245878696442, + 0.009188026189804077, + 0.010106978006660938, + 0.004802055656909943, + 0.00875762663781643, + 0.0031661835964769125, + 0.005060137715190649, + 0.007441613357514143, + 0.008126900531351566, + 0.005949476733803749, + 0.006560441106557846, + 0.007564331404864788, + 0.006665237247943878, + 0.0034069649409502745, + 0.006088528316468, + 0.005411834921687841, + 0.0037174392491579056, + 0.006965680979192257, + 0.00309005961753428, + 0.004518676549196243, + 0.008610527031123638, + 0.003895563306286931, + 0.008910447359085083, + 0.003732935758307576, + 0.004466489423066378, + 0.004445901606231928, + 0.005245716776698828, + 0.0026908277068287134, + 0.0048517100512981415, + 0.002389324363321066, + 0.004580010659992695, + 0.007161546498537064, + 0.0029323145281523466, + 0.009530484676361084, + 0.004689330700784922, + 0.005863951984792948, + 0.005049378145486116, + 0.007007406558841467, + 0.007833242416381836, + 0.004499494098126888, + 0.004013039171695709, + 0.007484667003154755, + 0.003968643955886364, + 0.006095458287745714, + 0.006477754097431898, + 0.006322143133729696, + 0.014643210917711258, + 0.0063238684087991714, + 0.007722674403339624, + 0.025020821020007133, + 0.010625927709043026, + 0.009449221193790436, + 0.024826085194945335, + 0.008297745138406754, + 0.006767368875443935, + 0.0038072378374636173, + 0.002878666389733553, + 0.00530947744846344, + 0.0028818787541240454, + 0.0042594303376972675, + 0.006892390549182892, + 0.006452110596001148, + 0.005912985187023878, + 0.0021524240728467703, + 0.0012901433510705829, + 0.00812164880335331, + 0.0017455116612836719, + 0.001259781769476831, + 0.007826479151844978, + 0.002838826971128583, + 0.002319742226973176 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 7, + 1 + ], + "selected_reward": -0.001510551548562944 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "In the distinctive style of a moving oil painting, a low-angle, wide-angle shot begins a slow, deliberate orbit around a lone olive tree. The entire scene is rendered with visible, impasto brushstrokes, giving every surface a tangible, sculpted texture. The ancient tree, the central subject, stands defiantly against a sun-drenched hill, its gnarled, twisted trunk an intricate tapestry of deep amber and cool gray applied with heavy, directional brushstrokes. As the camera gracefully curves, golden afternoon light saturates the frame, capturing the raised edges of the paint and casting long, dramatic, purplish shadows that stretch across the rolling ochre landscape. A soft, unseen breeze stirs, causing the tree's clusters of silvery-green stalks—each bearing a unique wisp of paint—to tremble subtly. Then, in a slow, poetic cascade, each leaf begins to separate, fluttering and spiraling downward, their silvery undersides shimmering in the warm light before coming to rest on the richly textured painted ground, creating a profound atmosphere of tranquility, timeless beauty and gentle decay.", + "seed": 25087, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000050/group_5/index_8182_seed_25087_dp_11_rank_22_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 5, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.012719351099804044, + "vlm_logprob": -0.012719351099804044, + "vlm_yes_no_margin": 4.875, + "vlm_token_credit": -0.012719351099804044, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video rendered in a moving oil painting style with visible impasto brushstrokes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the central subject a lone ancient olive tree with a gnarled, twisted trunk?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera perform a slow, deliberate orbit around the tree from a low-angle, wide-angle perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the tree's leaves separate and flutter downward in a slow cascade, creating silvery undersides?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene set on a sun-drenched hill with long, purplish shadows stretching across an ochre landscape?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.003890090622007847 + ], + "opposite_logprob_tokens": [ + -6.253890037536621 + ], + "logprob_reward": -0.003890090622007847, + "yes_no_margin_reward": 6.25, + "token_credit_reward": -0.003890090622007847, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08187327533960342, + 0.10174824297428131, + 0.11808647960424423, + 0.2593269646167755, + 0.05201944336295128, + 0.07303069531917572, + 0.04335048049688339, + 0.030336324125528336, + 0.06997304409742355, + 0.17025505006313324 + ], + "credit_map": [ + 0.00749506801366806, + 0.012471412308514118, + 0.009848297573626041, + 0.01589183695614338, + 0.008502922020852566, + 0.008666002191603184, + 0.009024063125252724, + 0.009973671287298203, + 0.012136981822550297, + 0.011991286650300026, + 0.013698861002922058, + 0.014189671725034714, + 0.010164587758481503, + 0.014170236885547638, + 0.00696385744959116, + 0.01843275874853134, + 0.013120902702212334, + 0.01645560935139656, + 0.013682249002158642, + 0.011152260936796665, + 0.011733573861420155, + 0.02079160511493683, + 0.015479381196200848, + 0.0156709011644125, + 0.034339550882577896, + 0.028715336695313454, + 0.035288285464048386, + 0.02514691837131977, + 0.0327247679233551, + 0.05203902721405029, + 0.03001677803695202, + 0.0210562851279974, + 0.0043896581046283245, + 0.00797458365559578, + 0.006443073507398367, + 0.006092095747590065, + 0.008845796808600426, + 0.006258225068449974, + 0.006384829990565777, + 0.005631181411445141, + 0.008499905467033386, + 0.014447174966335297, + 0.009329322725534439, + 0.006177028641104698, + 0.006793218199163675, + 0.010827330872416496, + 0.007978284731507301, + 0.008978433907032013, + 0.004651482217013836, + 0.003856381168588996, + 0.00724384468048811, + 0.00474691204726696, + 0.00369561486877501, + 0.005671943537890911, + 0.004476342815905809, + 0.009007960557937622, + 0.002470614854246378, + 0.004048587754368782, + 0.0063802748918533325, + 0.0029506643768399954, + 0.002918762620538473, + 0.0035963275004178286, + 0.0035045603290200233, + 0.00446653226390481, + 0.008382038213312626, + 0.006775448564440012, + 0.014315306209027767, + 0.005477978382259607, + 0.008038068190217018, + 0.010885893367230892, + 0.006883676629513502, + 0.009214635007083416, + 0.015493853017687798, + 0.014074623584747314, + 0.034513574093580246, + 0.011486226692795753, + 0.012314860709011555, + 0.03497815504670143, + 0.016592931002378464, + 0.03080083802342415 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 9, + 2 + ], + "selected_reward": -0.003890090622007847 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video rendered in a moving oil painting style with visible impasto brushstrokes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the central subject a lone ancient olive tree with a gnarled, twisted trunk?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera perform a slow, deliberate orbit around the tree from a low-angle, wide-angle perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the tree's leaves separate and flutter downward in a slow cascade, creating silvery undersides?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene set on a sun-drenched hill with long, purplish shadows stretching across an ochre landscape?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0031945884693413973 + ], + "opposite_logprob_tokens": [ + -5.753194808959961 + ], + "logprob_reward": -0.0031945884693413973, + "yes_no_margin_reward": 5.75, + "token_credit_reward": -0.0031945884693413973, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10991103947162628, + 0.07434824109077454, + 0.11851035803556442, + 0.09946131706237793, + 0.052817776799201965, + 0.134066641330719, + 0.07269038259983063, + 0.028090260922908783, + 0.21443845331668854, + 0.09566549956798553 + ], + "credit_map": [ + 0.010927421972155571, + 0.014458337798714638, + 0.01442019920796156, + 0.020508257672190666, + 0.00864436011761427, + 0.010913524776697159, + 0.010450361296534538, + 0.019588571041822433, + 0.007887653075158596, + 0.009760526940226555, + 0.010326013900339603, + 0.011886120773851871, + 0.006397744175046682, + 0.007759240455925465, + 0.005105046555399895, + 0.015225895680487156, + 0.015461740083992481, + 0.017157964408397675, + 0.015908660367131233, + 0.009333403781056404, + 0.01660667173564434, + 0.019948484376072884, + 0.00850280188024044, + 0.015590634196996689, + 0.011531255207955837, + 0.011484233662486076, + 0.011848792433738708, + 0.00989212654531002, + 0.01422108057886362, + 0.01637105830013752, + 0.012073035351932049, + 0.012039735913276672, + 0.004846061579883099, + 0.007355270441621542, + 0.004867279436439276, + 0.006432634778320789, + 0.009838207624852657, + 0.006937708705663681, + 0.005800229497253895, + 0.006740388460457325, + 0.019105074927210808, + 0.016032300889492035, + 0.021621132269501686, + 0.012074967846274376, + 0.013692855834960938, + 0.01899660937488079, + 0.016088610514998436, + 0.016455087810754776, + 0.006559486500918865, + 0.008576100692152977, + 0.014440259896218777, + 0.005474223755300045, + 0.007492775097489357, + 0.0090003227815032, + 0.008571739308536053, + 0.012575468979775906, + 0.002904134104028344, + 0.003929333761334419, + 0.004926653113216162, + 0.0028889263048768044, + 0.002128370339050889, + 0.003967285621911287, + 0.002803064649924636, + 0.0045424955897033215, + 0.037150297313928604, + 0.03027396835386753, + 0.03399614617228508, + 0.026013288646936417, + 0.021606994792819023, + 0.02119223028421402, + 0.019504373893141747, + 0.024701179936528206, + 0.009973104111850262, + 0.00925753265619278, + 0.01722438633441925, + 0.010327392257750034, + 0.00909843947738409, + 0.015233112499117851, + 0.01036627683788538, + 0.014185260981321335 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 5, + 2 + ], + "selected_reward": -0.0031945884693413973 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video rendered in a moving oil painting style with visible impasto brushstrokes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the central subject a lone ancient olive tree with a gnarled, twisted trunk?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera perform a slow, deliberate orbit around the tree from a low-angle, wide-angle perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the tree's leaves separate and flutter downward in a slow cascade, creating silvery undersides?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene set on a sun-drenched hill with long, purplish shadows stretching across an ochre landscape?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.014192244037985802 + ], + "opposite_logprob_tokens": [ + -4.2641921043396 + ], + "logprob_reward": -0.014192244037985802, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.014192244037985802, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.140360027551651, + 0.08730141818523407, + 0.123989038169384, + 0.1546756476163864, + 0.07222020626068115, + 0.1804470270872116, + 0.05877114459872246, + 0.035699885338544846, + 0.03679942339658737, + 0.10973618924617767 + ], + "credit_map": [ + 0.016387702897191048, + 0.013913621194660664, + 0.020187683403491974, + 0.018985139206051826, + 0.011084487661719322, + 0.027579445391893387, + 0.012415874749422073, + 0.01980607584118843, + 0.010506276972591877, + 0.01198034081608057, + 0.01233847439289093, + 0.018035251647233963, + 0.007507510017603636, + 0.009171205572783947, + 0.00537923164665699, + 0.012383123859763145, + 0.01609690487384796, + 0.018636738881468773, + 0.014868608675897121, + 0.01289623137563467, + 0.012168414890766144, + 0.018370168283581734, + 0.01476957742124796, + 0.016182400286197662, + 0.016270022839307785, + 0.018364692106842995, + 0.018687913194298744, + 0.014410306699573994, + 0.03397771716117859, + 0.023636162281036377, + 0.014648075215518475, + 0.01468075904995203, + 0.00636320048943162, + 0.0123825678601861, + 0.007609998341649771, + 0.009809164330363274, + 0.011494861915707588, + 0.009558144956827164, + 0.007089156657457352, + 0.00791311077773571, + 0.021518649533391, + 0.026829220354557037, + 0.029335258528590202, + 0.0165566373616457, + 0.018271982669830322, + 0.02757450006902218, + 0.019357899203896523, + 0.021002883091568947, + 0.006792528089135885, + 0.004303559195250273, + 0.00978969968855381, + 0.0057339854538440704, + 0.004594914615154266, + 0.008048471063375473, + 0.006487768609076738, + 0.01302021648734808, + 0.0035670071374624968, + 0.005248869303613901, + 0.005663829855620861, + 0.004141532350331545, + 0.002953728660941124, + 0.004593288991600275, + 0.003951479680836201, + 0.00558015052229166, + 0.005351145286113024, + 0.005393119063228369, + 0.005595997907221317, + 0.0031532514840364456, + 0.00421147933229804, + 0.004875697661191225, + 0.004128902219235897, + 0.004089829511940479, + 0.010963916778564453, + 0.00965841393917799, + 0.019469307735562325, + 0.007363719400018454, + 0.012635678052902222, + 0.01873272843658924, + 0.012834616005420685, + 0.01807781308889389 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 3, + 0 + ], + "selected_reward": -0.014192244037985802 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video rendered in a moving oil painting style with visible impasto brushstrokes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the central subject a lone ancient olive tree with a gnarled, twisted trunk?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera perform a slow, deliberate orbit around the tree from a low-angle, wide-angle perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the tree's leaves separate and flutter downward in a slow cascade, creating silvery undersides?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene set on a sun-drenched hill with long, purplish shadows stretching across an ochre landscape?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03368876501917839 + ], + "opposite_logprob_tokens": [ + -3.40868878364563 + ], + "logprob_reward": -0.03368876501917839, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.03368876501917839, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.03925646096467972, + 0.13797609508037567, + 0.06456710398197174, + 0.3222401440143585, + 0.1067887544631958, + 0.19925284385681152, + 0.033875469118356705, + 0.01613042876124382, + 0.044112320989370346, + 0.03580041229724884 + ], + "credit_map": [ + 0.004966394975781441, + 0.0034918258897960186, + 0.006417745724320412, + 0.00433585187420249, + 0.0027947798371315002, + 0.00529866386204958, + 0.003510832553729415, + 0.008440366946160793, + 0.018428843468427658, + 0.017393527552485466, + 0.018475135788321495, + 0.020291712135076523, + 0.014389843679964542, + 0.02272094041109085, + 0.008703635074198246, + 0.01757245697081089, + 0.006854991894215345, + 0.010040909983217716, + 0.008818479254841805, + 0.006030178163200617, + 0.008167671039700508, + 0.01150722336024046, + 0.00510589312762022, + 0.008041754364967346, + 0.04477503523230553, + 0.034680042415857315, + 0.03992512822151184, + 0.03145904466509819, + 0.0480629988014698, + 0.056660331785678864, + 0.03739834949374199, + 0.0292791947722435, + 0.009370469488203526, + 0.018138136714696884, + 0.012143024243414402, + 0.008772785775363445, + 0.023857537657022476, + 0.014635354280471802, + 0.012448078952729702, + 0.007423367816954851, + 0.026092099025845528, + 0.026354392990469933, + 0.028486663475632668, + 0.019719861447811127, + 0.017068665474653244, + 0.027434896677732468, + 0.02565496414899826, + 0.028441308066248894, + 0.0027464609593153, + 0.0037241438403725624, + 0.005155672784894705, + 0.004030823707580566, + 0.0029530180618166924, + 0.004560598637908697, + 0.004098701756447554, + 0.006606045179069042, + 0.00130184565205127, + 0.0019371304661035538, + 0.002702664118260145, + 0.0017900991952046752, + 0.001686345785856247, + 0.0020892550237476826, + 0.0019877676386386156, + 0.002635322278365493, + 0.006774551700800657, + 0.005980595946311951, + 0.007917365990579128, + 0.0044854264706373215, + 0.003143958980217576, + 0.0064830235205590725, + 0.003732735989615321, + 0.005594663321971893, + 0.0034503862261772156, + 0.004739510826766491, + 0.00685653043910861, + 0.002586615039035678, + 0.00325372745282948, + 0.005991961341351271, + 0.0033766936976462603, + 0.005544987972825766 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 5, + 1 + ], + "selected_reward": -0.03368876501917839 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video rendered in a moving oil painting style with visible impasto brushstrokes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the central subject a lone ancient olive tree with a gnarled, twisted trunk?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera perform a slow, deliberate orbit around the tree from a low-angle, wide-angle perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the tree's leaves separate and flutter downward in a slow cascade, creating silvery undersides?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene set on a sun-drenched hill with long, purplish shadows stretching across an ochre landscape?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.008631067350506783 + ], + "opposite_logprob_tokens": [ + -4.758631229400635 + ], + "logprob_reward": -0.008631067350506783, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.008631067350506783, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09466488659381866, + 0.2145443558692932, + 0.05836601182818413, + 0.2747483551502228, + 0.047150079160928726, + 0.0708499327301979, + 0.04913916066288948, + 0.0199396014213562, + 0.057563360780477524, + 0.11303416639566422 + ], + "credit_map": [ + 0.010487331077456474, + 0.009337780997157097, + 0.011724689975380898, + 0.012388184666633606, + 0.0067024314776062965, + 0.017610279843211174, + 0.007413078565150499, + 0.01900111511349678, + 0.02767167240381241, + 0.02454891987144947, + 0.028412770479917526, + 0.028669800609350204, + 0.02227656915783882, + 0.034300148487091064, + 0.013051860965788364, + 0.03561263903975487, + 0.005926114972680807, + 0.011405172757804394, + 0.007718166336417198, + 0.00703052245080471, + 0.005123349837958813, + 0.007914385758340359, + 0.006805307697504759, + 0.00644299853593111, + 0.029126355424523354, + 0.03447841480374336, + 0.03685355558991432, + 0.0254579558968544, + 0.04013575240969658, + 0.0490667000412941, + 0.025854619219899178, + 0.03377503156661987, + 0.00541418744251132, + 0.0048446753062307835, + 0.00825697835534811, + 0.006064445246011019, + 0.004546811804175377, + 0.007953460328280926, + 0.004218574613332748, + 0.005850950255990028, + 0.009230333380401134, + 0.0093952352181077, + 0.012175588868558407, + 0.005310873035341501, + 0.006406415719538927, + 0.011070024222135544, + 0.007702977862209082, + 0.00955849327147007, + 0.006763147655874491, + 0.0046859318390488625, + 0.008443070575594902, + 0.004222025163471699, + 0.004686996340751648, + 0.006472853943705559, + 0.004929825197905302, + 0.008935314603149891, + 0.0023769945837557316, + 0.0023002037778496742, + 0.0027316859923303127, + 0.0022019280586391687, + 0.002051935764029622, + 0.0032240678556263447, + 0.0020800314377993345, + 0.0029727560468018055, + 0.004427838139235973, + 0.006260334979742765, + 0.010890952311456203, + 0.0058596148155629635, + 0.005708662327378988, + 0.009547441266477108, + 0.006117969751358032, + 0.008750558830797672, + 0.013329117558896542, + 0.013021916151046753, + 0.01710142008960247, + 0.009162908419966698, + 0.012047091498970985, + 0.01961277797818184, + 0.011865443550050259, + 0.01689349114894867 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 1, + 9 + ], + "selected_reward": -0.008631067350506783 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000050/group_6/index_tensor([7465, 7465])_group_6_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000050/group_6/index_tensor([7465, 7465])_group_6_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..8361540a06ce94b5f8eb8cffc111f11a03407318 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000050/group_6/index_tensor([7465, 7465])_group_6_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "slow, horizontal, medium shot draws the viewer's gaze to a classically appointed study bathed in warm candlelight. An aristocratic gentleman sits elegantly at a heavy mahogany desk. He wears a dark blue velvet Rococo-style jacket, delicately embroidered at the collar and cuffs with gold. He wears an off-white brocade waistcoat, a rich white lace scarf dangling from his collar. Beneath his meticulously styled, curly white wig lies a face defined by a serene expression. A single candlelight, emanating from the side of the frame, sculpts the contours of his face and the folds of his velvet jacket in a Caravaggio-esque chiaroscuro, illuminating a gleaming gold coin lying flat on the table. He calmly extends his right hand, carefully placing the tip of his slender index finger on the coin's edge. Then, with a steady, decisive, and rhythmic movement, he delicately flicks it upward. The gold coin instantly left the table, tracing a graceful golden arc through the air before swirling towards the other end of the desk, its surface reflecting a fluid light in the candlelight. Finally, with uncanny precision, the coin landed with a crisp, \"ding\" sound into a crystal goblet half-filled with crimson wine, stirring a tiny ripple in the liquid before slowly sinking to the bottom. The entire scene, with its composition, lighting, and color, rigorously adhered to the classical aesthetics of Renaissance painting, exuding an atmosphere of tranquility, focus, and supreme control, like a masterpiece in motion.", + "seed": 30693, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000050/group_6/index_7465_seed_30693_dp_12_rank_24_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 6, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.04622701806947589, + "vlm_logprob": -0.04622701806947589, + "vlm_yes_no_margin": 4.225, + "vlm_token_credit": -0.04622701806947589, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a classically appointed study with an aristocratic gentleman wearing a dark blue velvet Rococo-style jacket?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the lighting characterized by warm candlelight creating a Caravaggio-esque chiaroscuro effect?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the gentleman flick a gold coin from a mahogany desk into a crystal goblet half-filled with crimson wine?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the coin's motion depicted as a graceful golden arc through the air before landing in the goblet?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the overall composition and color palette adhering to the classical aesthetics of a Renaissance painting?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.12801870703697205 + ], + "opposite_logprob_tokens": [ + -2.253018617630005 + ], + "logprob_reward": -0.12801870703697205, + "yes_no_margin_reward": 2.125, + "token_credit_reward": -0.12801870703697205, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.44579440355300903, + 0.07353215664625168, + 0.04446232691407204, + 0.06133529916405678, + 0.0464278981089592, + 0.13913407921791077, + 0.031109467148780823, + 0.03343572840094566, + 0.0755792185664177, + 0.049189455807209015 + ], + "credit_map": [ + 0.04519680142402649, + 0.06446683406829834, + 0.04650229960680008, + 0.04827035218477249, + 0.08724088966846466, + 0.049720846116542816, + 0.06181718036532402, + 0.042579248547554016, + 0.008745160885155201, + 0.01113552413880825, + 0.007978896610438824, + 0.005280297715216875, + 0.01562055479735136, + 0.007624564692378044, + 0.009569927118718624, + 0.007577230222523212, + 0.004377462435513735, + 0.005937851965427399, + 0.0058077056892216206, + 0.004087789449840784, + 0.006054397206753492, + 0.0036833886988461018, + 0.00712289335206151, + 0.007390834856778383, + 0.005274471826851368, + 0.00839180313050747, + 0.004864690825343132, + 0.007618751376867294, + 0.009019480086863041, + 0.0058252387680113316, + 0.014575493521988392, + 0.005765371955931187, + 0.0030926510225981474, + 0.006184818223118782, + 0.00425094086676836, + 0.008000650443136692, + 0.008840946480631828, + 0.004818695597350597, + 0.0072728111408650875, + 0.003966387826949358, + 0.00823255255818367, + 0.0334978848695755, + 0.009664331562817097, + 0.033410586416721344, + 0.012770548462867737, + 0.010163434781134129, + 0.020882299169898033, + 0.01051243394613266, + 0.004128327593207359, + 0.003969815094023943, + 0.0030948698986321688, + 0.0033919333945959806, + 0.004158978816121817, + 0.0033732859883457422, + 0.005338134244084358, + 0.0036541216541081667, + 0.004679299425333738, + 0.004787751007825136, + 0.006276594940572977, + 0.0043441434390842915, + 0.004195083398371935, + 0.0032024839892983437, + 0.0029037296772003174, + 0.003046642290428281, + 0.011949765495955944, + 0.011136102490127087, + 0.015508543699979782, + 0.011877942830324173, + 0.007392589468508959, + 0.007183178327977657, + 0.005112739279866219, + 0.005418356973677874, + 0.007987127639353275, + 0.007935184985399246, + 0.009743296541273594, + 0.007728771772235632, + 0.00462706433609128, + 0.004358403384685516, + 0.0033535549882799387, + 0.0034560514613986015 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 5, + 8 + ], + "selected_reward": -0.12801870703697205 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a classically appointed study with an aristocratic gentleman wearing a dark blue velvet Rococo-style jacket?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the lighting characterized by warm candlelight creating a Caravaggio-esque chiaroscuro effect?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the gentleman flick a gold coin from a mahogany desk into a crystal goblet half-filled with crimson wine?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the coin's motion depicted as a graceful golden arc through the air before landing in the goblet?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the overall composition and color palette adhering to the classical aesthetics of a Renaissance painting?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0022232832852751017 + ], + "opposite_logprob_tokens": [ + -6.127223491668701 + ], + "logprob_reward": -0.0022232832852751017, + "yes_no_margin_reward": 6.125, + "token_credit_reward": -0.0022232832852751017, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.179575115442276, + 0.06678812950849533, + 0.08661431074142456, + 0.06738915294408798, + 0.09012465178966522, + 0.1315198689699173, + 0.14575456082820892, + 0.11423475295305252, + 0.06552881002426147, + 0.05247068405151367 + ], + "credit_map": [ + 0.019844580441713333, + 0.021810881793498993, + 0.019391724839806557, + 0.02384979836642742, + 0.02870926819741726, + 0.018769187852740288, + 0.02993801049888134, + 0.017261667177081108, + 0.006106274202466011, + 0.009116249158978462, + 0.007904436439275742, + 0.008064529858529568, + 0.013341717422008514, + 0.006938524544239044, + 0.009866111911833286, + 0.0054502831771969795, + 0.007113487459719181, + 0.012352172285318375, + 0.008794263936579227, + 0.012691219337284565, + 0.015289788134396076, + 0.010853986255824566, + 0.011769493110477924, + 0.007749902084469795, + 0.0076026697643101215, + 0.009273242205381393, + 0.005529145244508982, + 0.009761303663253784, + 0.010201659984886646, + 0.005324642173945904, + 0.012486633844673634, + 0.007209853734821081, + 0.007998306304216385, + 0.011694504879415035, + 0.010657384991645813, + 0.014011676423251629, + 0.01452790480107069, + 0.007281037513166666, + 0.015307017602026463, + 0.008646810427308083, + 0.008434784598648548, + 0.04761447384953499, + 0.005401380360126495, + 0.042456772178411484, + 0.0075407144613564014, + 0.004684916231781244, + 0.008510642684996128, + 0.006876180414110422, + 0.012059968896210194, + 0.019356651231646538, + 0.014484629034996033, + 0.014671768061816692, + 0.014961411245167255, + 0.014412175863981247, + 0.03971204534173012, + 0.016095899045467377, + 0.008130094036459923, + 0.014465157873928547, + 0.011197283864021301, + 0.012991819530725479, + 0.01294776052236557, + 0.011950785294175148, + 0.02655482292175293, + 0.01599702797830105, + 0.005405941978096962, + 0.009362025186419487, + 0.005724368616938591, + 0.006615803577005863, + 0.01259678415954113, + 0.010435279458761215, + 0.006708559114485979, + 0.008680048398673534, + 0.003311168635264039, + 0.0073228273540735245, + 0.004076128359884024, + 0.005256236996501684, + 0.011020692065358162, + 0.008338998071849346, + 0.008455832488834858, + 0.004688797984272242 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 6, + 5 + ], + "selected_reward": -0.0022232832852751017 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a classically appointed study with an aristocratic gentleman wearing a dark blue velvet Rococo-style jacket?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the lighting characterized by warm candlelight creating a Caravaggio-esque chiaroscuro effect?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the gentleman flick a gold coin from a mahogany desk into a crystal goblet half-filled with crimson wine?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the coin's motion depicted as a graceful golden arc through the air before landing in the goblet?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the overall composition and color palette adhering to the classical aesthetics of a Renaissance painting?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.08906083554029465 + ], + "opposite_logprob_tokens": [ + -2.4640607833862305 + ], + "logprob_reward": -0.08906083554029465, + "yes_no_margin_reward": 2.375, + "token_credit_reward": -0.08906083554029465, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.20817185938358307, + 0.07023530453443527, + 0.07964015007019043, + 0.07841179519891739, + 0.06734056770801544, + 0.1500028818845749, + 0.1561654806137085, + 0.11607489734888077, + 0.04157786816358566, + 0.03237922862172127 + ], + "credit_map": [ + 0.021329715847969055, + 0.02881183661520481, + 0.017268098890781403, + 0.022046364843845367, + 0.04287520423531532, + 0.022583991289138794, + 0.03577166423201561, + 0.017484990879893303, + 0.007416885811835527, + 0.011197037063539028, + 0.004955522250384092, + 0.006708238273859024, + 0.01625026948750019, + 0.007075166329741478, + 0.011739819310605526, + 0.004892371129244566, + 0.006298222113400698, + 0.011281274259090424, + 0.008101941086351871, + 0.010765619575977325, + 0.013148817233741283, + 0.008946452289819717, + 0.012331515550613403, + 0.008766308426856995, + 0.008186126127839088, + 0.011119949631392956, + 0.006435001268982887, + 0.013304109685122967, + 0.012071190401911736, + 0.006432028487324715, + 0.01420839224010706, + 0.006654998753219843, + 0.004655933473259211, + 0.009146425873041153, + 0.007588755805045366, + 0.01018450316041708, + 0.011832970194518566, + 0.006878822576254606, + 0.010476194322109222, + 0.006576966494321823, + 0.008729672990739346, + 0.0559639111161232, + 0.004671719390898943, + 0.049250949174165726, + 0.009264955297112465, + 0.004900221712887287, + 0.00959780439734459, + 0.007623630575835705, + 0.014296451583504677, + 0.020933352410793304, + 0.012010359205305576, + 0.012799770571291447, + 0.017686596140265465, + 0.01697884313762188, + 0.0438668355345726, + 0.017593255266547203, + 0.009079418145120144, + 0.014073329046368599, + 0.010329846292734146, + 0.011880654841661453, + 0.013857057318091393, + 0.014060459099709988, + 0.029283110052347183, + 0.013511015102267265, + 0.004865642171353102, + 0.0054029193706810474, + 0.004623827058821917, + 0.004630897659808397, + 0.006391928531229496, + 0.005327400751411915, + 0.004995069932192564, + 0.00534018175676465, + 0.0030048624612390995, + 0.004240960814058781, + 0.004149877466261387, + 0.0030665353406220675, + 0.00591925298795104, + 0.00492820842191577, + 0.003461747895926237, + 0.0036077857948839664 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 6, + 5 + ], + "selected_reward": -0.08906083554029465 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a classically appointed study with an aristocratic gentleman wearing a dark blue velvet Rococo-style jacket?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the lighting characterized by warm candlelight creating a Caravaggio-esque chiaroscuro effect?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the gentleman flick a gold coin from a mahogany desk into a crystal goblet half-filled with crimson wine?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the coin's motion depicted as a graceful golden arc through the air before landing in the goblet?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the overall composition and color palette adhering to the classical aesthetics of a Renaissance painting?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.008639458566904068 + ], + "opposite_logprob_tokens": [ + -4.758639335632324 + ], + "logprob_reward": -0.008639458566904068, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.008639458566904068, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13060982525348663, + 0.06548871099948883, + 0.06629686057567596, + 0.05818427726626396, + 0.07683230191469193, + 0.15032289922237396, + 0.16110067069530487, + 0.1722971349954605, + 0.056195106357336044, + 0.06267222762107849 + ], + "credit_map": [ + 0.014888959936797619, + 0.02000580169260502, + 0.009915478527545929, + 0.01677457056939602, + 0.02633560448884964, + 0.012503156438469887, + 0.02173481322824955, + 0.00845143012702465, + 0.005521914456039667, + 0.010227145627140999, + 0.006036823615431786, + 0.00775893172249198, + 0.01320831198245287, + 0.006829377263784409, + 0.010992195457220078, + 0.004914012271910906, + 0.0056855930015444756, + 0.009604010730981827, + 0.00722041679546237, + 0.008846898563206196, + 0.011048204265534878, + 0.00802976731210947, + 0.009015935473144054, + 0.006846027914434671, + 0.006479560397565365, + 0.007884793914854527, + 0.004741097334772348, + 0.008723296225070953, + 0.009958311915397644, + 0.00536407670006156, + 0.008112911134958267, + 0.006920226383954287, + 0.00593535928055644, + 0.010361742228269577, + 0.010080265812575817, + 0.011527719907462597, + 0.012332971207797527, + 0.007334110327064991, + 0.011468417942523956, + 0.007791716139763594, + 0.008599829860031605, + 0.05106474086642265, + 0.0076892501674592495, + 0.044320594519376755, + 0.010408326052129269, + 0.006150256842374802, + 0.01511893980205059, + 0.006970962043851614, + 0.013633409515023232, + 0.02189403958618641, + 0.011869294568896294, + 0.014232510700821877, + 0.02098868601024151, + 0.017893856391310692, + 0.03832820802927017, + 0.02226066403090954, + 0.013781002722680569, + 0.02084525302052498, + 0.015251641161739826, + 0.017052629962563515, + 0.023361900821328163, + 0.023222949355840683, + 0.03844033554196358, + 0.020341437309980392, + 0.004769886843860149, + 0.00827538687735796, + 0.004906040616333485, + 0.005225855857133865, + 0.011989318765699863, + 0.009008869528770447, + 0.006776740774512291, + 0.0052430033683776855, + 0.004266910720616579, + 0.009126362390816212, + 0.00570002943277359, + 0.006767779588699341, + 0.013812943361699581, + 0.010541247203946114, + 0.007347653619945049, + 0.0051092966459691525 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 6, + 5 + ], + "selected_reward": -0.008639458566904068 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a classically appointed study with an aristocratic gentleman wearing a dark blue velvet Rococo-style jacket?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the lighting characterized by warm candlelight creating a Caravaggio-esque chiaroscuro effect?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the gentleman flick a gold coin from a mahogany desk into a crystal goblet half-filled with crimson wine?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the coin's motion depicted as a graceful golden arc through the air before landing in the goblet?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the overall composition and color palette adhering to the classical aesthetics of a Renaissance painting?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0031928059179335833 + ], + "opposite_logprob_tokens": [ + -5.753192901611328 + ], + "logprob_reward": -0.0031928059179335833, + "yes_no_margin_reward": 5.75, + "token_credit_reward": -0.0031928059179335833, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2492421418428421, + 0.05942384526133537, + 0.06187340244650841, + 0.05947405472397804, + 0.08279915899038315, + 0.11758836358785629, + 0.1226893812417984, + 0.047927241772413254, + 0.12857598066329956, + 0.07040640711784363 + ], + "credit_map": [ + 0.036187514662742615, + 0.037803344428539276, + 0.03316139802336693, + 0.023277977481484413, + 0.03817072883248329, + 0.025156483054161072, + 0.03000652603805065, + 0.02547818049788475, + 0.008715051226317883, + 0.009569715708494186, + 0.008670150302350521, + 0.004577310290187597, + 0.009150560013949871, + 0.007059634197503328, + 0.0060775140300393105, + 0.005603909492492676, + 0.006168234162032604, + 0.009719218127429485, + 0.007132590748369694, + 0.006270926911383867, + 0.013921154662966728, + 0.0063195135444402695, + 0.007058910094201565, + 0.005282848607748747, + 0.0062177544459700584, + 0.008796163834631443, + 0.006347901187837124, + 0.007812870666384697, + 0.008186466060578823, + 0.00642607593908906, + 0.010651671327650547, + 0.00503515126183629, + 0.007910158485174179, + 0.011357363313436508, + 0.008227640762925148, + 0.011968141421675682, + 0.013174095191061497, + 0.009441843256354332, + 0.010570568963885307, + 0.010149345733225346, + 0.006520683411508799, + 0.02877645753324032, + 0.006518721114844084, + 0.023860380053520203, + 0.01351282000541687, + 0.00995567999780178, + 0.019453616812825203, + 0.008990008383989334, + 0.015914004296064377, + 0.02046212926506996, + 0.02061278186738491, + 0.019929468631744385, + 0.012549856677651405, + 0.009699546732008457, + 0.012098615989089012, + 0.011422972194850445, + 0.0055150603875517845, + 0.006443388294428587, + 0.0061860354617238045, + 0.006643162574619055, + 0.0064867762848734856, + 0.004986748564988375, + 0.006641053594648838, + 0.005025015212595463, + 0.013068932108581066, + 0.018563630059361458, + 0.0144984545186162, + 0.013979476876556873, + 0.024239813908934593, + 0.020704861730337143, + 0.013734111562371254, + 0.00978669710457325, + 0.00774461729452014, + 0.010114024393260479, + 0.010178649798035622, + 0.009360674768686295, + 0.010640445165336132, + 0.008258193731307983, + 0.008366766385734081, + 0.005743036977946758 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 6 + ], + "selected_reward": -0.0031928059179335833 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "slow, horizontal, medium shot draws the viewer's gaze to a classically appointed study bathed in warm candlelight. An aristocratic gentleman sits elegantly at a heavy mahogany desk. He wears a dark blue velvet Rococo-style jacket, delicately embroidered at the collar and cuffs with gold. He wears an off-white brocade waistcoat, a rich white lace scarf dangling from his collar. Beneath his meticulously styled, curly white wig lies a face defined by a serene expression. A single candlelight, emanating from the side of the frame, sculpts the contours of his face and the folds of his velvet jacket in a Caravaggio-esque chiaroscuro, illuminating a gleaming gold coin lying flat on the table. He calmly extends his right hand, carefully placing the tip of his slender index finger on the coin's edge. Then, with a steady, decisive, and rhythmic movement, he delicately flicks it upward. The gold coin instantly left the table, tracing a graceful golden arc through the air before swirling towards the other end of the desk, its surface reflecting a fluid light in the candlelight. Finally, with uncanny precision, the coin landed with a crisp, \"ding\" sound into a crystal goblet half-filled with crimson wine, stirring a tiny ripple in the liquid before slowly sinking to the bottom. The entire scene, with its composition, lighting, and color, rigorously adhered to the classical aesthetics of Renaissance painting, exuding an atmosphere of tranquility, focus, and supreme control, like a masterpiece in motion.", + "seed": 30694, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000050/group_6/index_7465_seed_30694_dp_12_rank_24_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 6, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.013427173858508468, + "vlm_logprob": -0.013427173858508468, + "vlm_yes_no_margin": 5.0, + "vlm_token_credit": -0.013427173858508468, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a classically appointed study with an aristocratic gentleman wearing a dark blue velvet Rococo-style jacket?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the lighting characterized by warm candlelight creating a Caravaggio-esque chiaroscuro effect?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the gentleman flick a gold coin from a mahogany desk into a crystal goblet half-filled with crimson wine?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the coin's motion depicted as a graceful golden arc through the air before landing in the goblet?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the overall composition and color palette adhering to the classical aesthetics of a Renaissance painting?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.020997844636440277 + ], + "opposite_logprob_tokens": [ + -5.895998001098633 + ], + "logprob_reward": -0.020997844636440277, + "yes_no_margin_reward": 5.875, + "token_credit_reward": -0.020997844636440277, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.21448677778244019, + 0.15873484313488007, + 0.05358130857348442, + 0.1395145058631897, + 0.08036239445209503, + 0.11693104356527328, + 0.07601027190685272, + 0.05394717678427696, + 0.05103509873151779, + 0.05539659038186073 + ], + "credit_map": [ + 0.02264750376343727, + 0.02454531565308571, + 0.028019322082400322, + 0.04558517038822174, + 0.021323051303625107, + 0.025502847507596016, + 0.021585294976830482, + 0.02527826651930809, + 0.01183807011693716, + 0.0054110935889184475, + 0.04638269916176796, + 0.01131072361022234, + 0.009434915147721767, + 0.05621969327330589, + 0.006407740525901318, + 0.011729909107089043, + 0.0043632872402668, + 0.006698959972709417, + 0.005742423236370087, + 0.01148635521531105, + 0.006078430917114019, + 0.005937272682785988, + 0.00815960019826889, + 0.005114978179335594, + 0.014406275004148483, + 0.02035418525338173, + 0.014033094979822636, + 0.02146890200674534, + 0.015187475830316544, + 0.0161467082798481, + 0.018101099878549576, + 0.019816769286990166, + 0.0059403893537819386, + 0.0076620914041996, + 0.010504105128347874, + 0.017396871000528336, + 0.008546472527086735, + 0.009634827263653278, + 0.008513075299561024, + 0.012164555490016937, + 0.013376627117395401, + 0.013883329927921295, + 0.014094818383455276, + 0.01903606578707695, + 0.013001027517020702, + 0.018455715849995613, + 0.011938714422285557, + 0.013144750148057938, + 0.00835476815700531, + 0.009075102396309376, + 0.014623431488871574, + 0.009240853600203991, + 0.007518970873206854, + 0.010151375085115433, + 0.007406883407384157, + 0.009638886898756027, + 0.006865428760647774, + 0.007826867513358593, + 0.008370911702513695, + 0.009293661452829838, + 0.006272795610129833, + 0.006559877190738916, + 0.003598481183871627, + 0.005159157328307629, + 0.005280323792248964, + 0.005845759529620409, + 0.007001105695962906, + 0.008237331174314022, + 0.005520564038306475, + 0.0066243563778698444, + 0.004632832482457161, + 0.00789282564073801, + 0.006299879867583513, + 0.007328244857490063, + 0.006956837605684996, + 0.006173168309032917, + 0.006535761523991823, + 0.007877848111093044, + 0.005204974673688412, + 0.009019878692924976 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 3 + ], + "selected_reward": -0.020997844636440277 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a classically appointed study with an aristocratic gentleman wearing a dark blue velvet Rococo-style jacket?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the lighting characterized by warm candlelight creating a Caravaggio-esque chiaroscuro effect?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the gentleman flick a gold coin from a mahogany desk into a crystal goblet half-filled with crimson wine?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the coin's motion depicted as a graceful golden arc through the air before landing in the goblet?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the overall composition and color palette adhering to the classical aesthetics of a Renaissance painting?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.00865836814045906 + ], + "opposite_logprob_tokens": [ + -4.758658409118652 + ], + "logprob_reward": -0.00865836814045906, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.00865836814045906, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1892262101173401, + 0.10108844190835953, + 0.07714994251728058, + 0.06880944222211838, + 0.19887714087963104, + 0.05857319012284279, + 0.06068901717662811, + 0.1365332305431366, + 0.07392938435077667, + 0.03512405604124069 + ], + "credit_map": [ + 0.026236101984977722, + 0.021733077242970467, + 0.02776806242763996, + 0.02595808170735836, + 0.023751679807901382, + 0.025631282478570938, + 0.018481722101569176, + 0.01966618373990059, + 0.009250171482563019, + 0.006840585730969906, + 0.02020006626844406, + 0.013239742256700993, + 0.009489268064498901, + 0.027670325711369514, + 0.004795041400939226, + 0.009603234007954597, + 0.011448300443589687, + 0.006719404831528664, + 0.008896672166883945, + 0.009214433841407299, + 0.01096743531525135, + 0.01339248288422823, + 0.008769202046096325, + 0.007742005400359631, + 0.00929258018732071, + 0.011562452651560307, + 0.008412063121795654, + 0.007454597391188145, + 0.01014481671154499, + 0.010867556557059288, + 0.004981460515409708, + 0.006093907169997692, + 0.019742198288440704, + 0.02014213055372238, + 0.03070375695824623, + 0.0220944881439209, + 0.018213974311947823, + 0.0363960862159729, + 0.017507078126072884, + 0.03407740592956543, + 0.005730829667299986, + 0.005579857621341944, + 0.0077682798728346825, + 0.008458847180008888, + 0.008028380572795868, + 0.012940002605319023, + 0.004242457449436188, + 0.0058245304971933365, + 0.006237381137907505, + 0.005545291118323803, + 0.0081415306776762, + 0.009878006763756275, + 0.008786524645984173, + 0.006698968820273876, + 0.00827012024819851, + 0.0071311877109110355, + 0.011939252726733685, + 0.012473434209823608, + 0.018467262387275696, + 0.021131115034222603, + 0.015231977216899395, + 0.020340770483016968, + 0.014252523891627789, + 0.022696882486343384, + 0.006975438911467791, + 0.007003006059676409, + 0.009015033021569252, + 0.010610448196530342, + 0.006811655592173338, + 0.012045321054756641, + 0.00648160558193922, + 0.01498686894774437, + 0.003441176377236843, + 0.004800933878868818, + 0.004098331555724144, + 0.004160634707659483, + 0.004079923499375582, + 0.004425934981554747, + 0.0034646284766495228, + 0.00665249302983284 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 0, + 7 + ], + "selected_reward": -0.00865836814045906 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a classically appointed study with an aristocratic gentleman wearing a dark blue velvet Rococo-style jacket?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the lighting characterized by warm candlelight creating a Caravaggio-esque chiaroscuro effect?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the gentleman flick a gold coin from a mahogany desk into a crystal goblet half-filled with crimson wine?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the coin's motion depicted as a graceful golden arc through the air before landing in the goblet?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the overall composition and color palette adhering to the classical aesthetics of a Renaissance painting?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02635478787124157 + ], + "opposite_logprob_tokens": [ + -3.6513547897338867 + ], + "logprob_reward": -0.02635478787124157, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.02635478787124157, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.33195188641548157, + 0.11414720863103867, + 0.07869697362184525, + 0.0634986087679863, + 0.17515946924686432, + 0.0438721589744091, + 0.0795479491353035, + 0.037513550370931625, + 0.04759497568011284, + 0.02801714651286602 + ], + "credit_map": [ + 0.047517627477645874, + 0.03459252417087555, + 0.04935121908783913, + 0.04865920916199684, + 0.03316638246178627, + 0.05872553214430809, + 0.03502843156456947, + 0.02491098642349243, + 0.01132902130484581, + 0.009158301167190075, + 0.016977321356534958, + 0.017085162922739983, + 0.0131848668679595, + 0.02311643958091736, + 0.006995145697146654, + 0.016300957649946213, + 0.013319646008312702, + 0.00647437060251832, + 0.00970100425183773, + 0.0121759669855237, + 0.009646022692322731, + 0.012878622859716415, + 0.007096649147570133, + 0.007404697127640247, + 0.006827771198004484, + 0.011854734271764755, + 0.00876732636243105, + 0.01044487301260233, + 0.005406809039413929, + 0.006618828978389502, + 0.007655513007193804, + 0.0059227533638477325, + 0.01798730343580246, + 0.0192578062415123, + 0.030473707243800163, + 0.04173756018280983, + 0.01282365433871746, + 0.02140844240784645, + 0.016726214438676834, + 0.014744793996214867, + 0.0051212431862950325, + 0.0038323034532368183, + 0.006252943072468042, + 0.0070421150885522366, + 0.005689195357263088, + 0.006936539895832539, + 0.0041428920812904835, + 0.004854931961745024, + 0.00847882591187954, + 0.006419656798243523, + 0.012387152761220932, + 0.01076175644993782, + 0.010756216943264008, + 0.013489366509020329, + 0.009022936224937439, + 0.00823204405605793, + 0.003987141884863377, + 0.003476326819509268, + 0.005755170714110136, + 0.004680943209677935, + 0.006443385034799576, + 0.005579568445682526, + 0.0035046113189309835, + 0.004086405970156193, + 0.004308982286602259, + 0.005381995812058449, + 0.007695641368627548, + 0.010052098892629147, + 0.003963991068303585, + 0.0053181881085038185, + 0.005082888063043356, + 0.005791196133941412, + 0.0028341214638203382, + 0.002685167361050844, + 0.004133969079703093, + 0.003458883613348007, + 0.003558175405487418, + 0.004434751346707344, + 0.003042801283299923, + 0.003869280917569995 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 1 + ], + "selected_reward": -0.02635478787124157 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a classically appointed study with an aristocratic gentleman wearing a dark blue velvet Rococo-style jacket?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the lighting characterized by warm candlelight creating a Caravaggio-esque chiaroscuro effect?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the gentleman flick a gold coin from a mahogany desk into a crystal goblet half-filled with crimson wine?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the coin's motion depicted as a graceful golden arc through the air before landing in the goblet?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the overall composition and color palette adhering to the classical aesthetics of a Renaissance painting?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.008638276718556881 + ], + "opposite_logprob_tokens": [ + -4.758638381958008 + ], + "logprob_reward": -0.008638276718556881, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.008638276718556881, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2626939117908478, + 0.10939997434616089, + 0.08323387056589127, + 0.11015058308839798, + 0.20391684770584106, + 0.04304889217019081, + 0.05615096539258957, + 0.04459260776638985, + 0.052690714597702026, + 0.03412161022424698 + ], + "credit_map": [ + 0.03466784581542015, + 0.02711779996752739, + 0.040994856506586075, + 0.03771042451262474, + 0.026301566511392593, + 0.04561168700456619, + 0.026490343734622, + 0.023799384012818336, + 0.010658723302185535, + 0.00643854308873415, + 0.025546269491314888, + 0.007626088336110115, + 0.009012196213006973, + 0.03391996771097183, + 0.005398330744355917, + 0.010799855925142765, + 0.015048973262310028, + 0.007879486307501793, + 0.009867358021438122, + 0.012760511599481106, + 0.008869374170899391, + 0.012961194850504398, + 0.00731841754168272, + 0.008528556674718857, + 0.012289352715015411, + 0.021683748811483383, + 0.012727399356663227, + 0.01905202493071556, + 0.01053167600184679, + 0.011791941709816456, + 0.01132894866168499, + 0.010745489038527012, + 0.016343750059604645, + 0.023755693808197975, + 0.03379575163125992, + 0.05158589407801628, + 0.018060460686683655, + 0.021661553531885147, + 0.022379552945494652, + 0.016334179788827896, + 0.004976347088813782, + 0.0039003435522317886, + 0.005277978256344795, + 0.006288087461143732, + 0.005807391833513975, + 0.008074863813817501, + 0.003209828631952405, + 0.005514050368219614, + 0.006068835966289043, + 0.004965805448591709, + 0.009466594085097313, + 0.007273656316101551, + 0.007498060818761587, + 0.009946485981345177, + 0.005232257768511772, + 0.005699269939213991, + 0.00400402070954442, + 0.004231282975524664, + 0.007141632493585348, + 0.006010737735778093, + 0.008128381334245205, + 0.00602181488648057, + 0.004104784224182367, + 0.004949953407049179, + 0.004305036272853613, + 0.0059906900860369205, + 0.007679332047700882, + 0.01021569687873125, + 0.00449973251670599, + 0.0066538178361952305, + 0.005707211326807737, + 0.0076391990296542645, + 0.00419362960383296, + 0.003644751152023673, + 0.0044393837451934814, + 0.00433293916285038, + 0.00453391345217824, + 0.004271142650395632, + 0.0039030052721500397, + 0.004802844952791929 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 3 + ], + "selected_reward": -0.008638276718556881 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a classically appointed study with an aristocratic gentleman wearing a dark blue velvet Rococo-style jacket?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the lighting characterized by warm candlelight creating a Caravaggio-esque chiaroscuro effect?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the gentleman flick a gold coin from a mahogany desk into a crystal goblet half-filled with crimson wine?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the coin's motion depicted as a graceful golden arc through the air before landing in the goblet?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the overall composition and color palette adhering to the classical aesthetics of a Renaissance painting?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.00248659192584455 + ], + "opposite_logprob_tokens": [ + -6.002486705780029 + ], + "logprob_reward": -0.00248659192584455, + "yes_no_margin_reward": 6.0, + "token_credit_reward": -0.00248659192584455, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3495728373527527, + 0.054281096905469894, + 0.04975089430809021, + 0.20860221982002258, + 0.08587823808193207, + 0.028075627982616425, + 0.03358198329806328, + 0.08914720267057419, + 0.06683375686407089, + 0.03427618369460106 + ], + "credit_map": [ + 0.03215353935956955, + 0.045618996024131775, + 0.04697597399353981, + 0.05314745754003525, + 0.0365750789642334, + 0.041812606155872345, + 0.04693247750401497, + 0.046356700360774994, + 0.0046128700487315655, + 0.003558329539373517, + 0.010889945551753044, + 0.0076099238358438015, + 0.0050767166540026665, + 0.013854297809302807, + 0.002820441732183099, + 0.005858567543327808, + 0.005549544934183359, + 0.00600814912468195, + 0.004835391417145729, + 0.008630738593637943, + 0.006757281720638275, + 0.005808579735457897, + 0.008044381625950336, + 0.004116817377507687, + 0.02422434650361538, + 0.03405335173010826, + 0.022768903523683548, + 0.027410216629505157, + 0.02390352636575699, + 0.02800286002457142, + 0.025075120851397514, + 0.023163877427577972, + 0.006848922464996576, + 0.007856842130422592, + 0.011402079835534096, + 0.017809947952628136, + 0.010471085086464882, + 0.01122547872364521, + 0.00839852076023817, + 0.011865353211760521, + 0.004575252532958984, + 0.003071814775466919, + 0.004267448093742132, + 0.0032640297431498766, + 0.003564372891560197, + 0.0047378093004226685, + 0.0020697496365755796, + 0.002525147283449769, + 0.004046058282256126, + 0.003524653147906065, + 0.005836915224790573, + 0.0038113584741950035, + 0.005142807494848967, + 0.004526687785983086, + 0.0021179437171667814, + 0.0045755538158118725, + 0.008000650443136692, + 0.009258019737899303, + 0.011708620004355907, + 0.013589262962341309, + 0.009733590297400951, + 0.01401106733828783, + 0.00790444016456604, + 0.014941539615392685, + 0.006224518176168203, + 0.005687085911631584, + 0.008786683902144432, + 0.0078565189614892, + 0.006241200026124716, + 0.01133886817842722, + 0.006803556345403194, + 0.013895321637392044, + 0.0032571041956543922, + 0.003596861381083727, + 0.004831950180232525, + 0.004535540007054806, + 0.005032102111726999, + 0.004295362625271082, + 0.002946865977719426, + 0.005780393723398447 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 7 + ], + "selected_reward": -0.00248659192584455 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "slow, horizontal, medium shot draws the viewer's gaze to a classically appointed study bathed in warm candlelight. An aristocratic gentleman sits elegantly at a heavy mahogany desk. He wears a dark blue velvet Rococo-style jacket, delicately embroidered at the collar and cuffs with gold. He wears an off-white brocade waistcoat, a rich white lace scarf dangling from his collar. Beneath his meticulously styled, curly white wig lies a face defined by a serene expression. A single candlelight, emanating from the side of the frame, sculpts the contours of his face and the folds of his velvet jacket in a Caravaggio-esque chiaroscuro, illuminating a gleaming gold coin lying flat on the table. He calmly extends his right hand, carefully placing the tip of his slender index finger on the coin's edge. Then, with a steady, decisive, and rhythmic movement, he delicately flicks it upward. The gold coin instantly left the table, tracing a graceful golden arc through the air before swirling towards the other end of the desk, its surface reflecting a fluid light in the candlelight. Finally, with uncanny precision, the coin landed with a crisp, \"ding\" sound into a crystal goblet half-filled with crimson wine, stirring a tiny ripple in the liquid before slowly sinking to the bottom. The entire scene, with its composition, lighting, and color, rigorously adhered to the classical aesthetics of Renaissance painting, exuding an atmosphere of tranquility, focus, and supreme control, like a masterpiece in motion.", + "seed": 30695, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000050/group_6/index_7465_seed_30695_dp_13_rank_26_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 6, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.045423412974923846, + "vlm_logprob": -0.045423412974923846, + "vlm_yes_no_margin": 3.875, + "vlm_token_credit": -0.045423412974923846, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a classically appointed study with an aristocratic gentleman wearing a dark blue velvet Rococo-style jacket?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the lighting characterized by warm candlelight creating a Caravaggio-esque chiaroscuro effect?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the gentleman flick a gold coin from a mahogany desk into a crystal goblet half-filled with crimson wine?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the coin's motion depicted as a graceful golden arc through the air before landing in the goblet?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the overall composition and color palette adhering to the classical aesthetics of a Renaissance painting?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.029974184930324554 + ], + "opposite_logprob_tokens": [ + -3.7799742221832275 + ], + "logprob_reward": -0.029974184930324554, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.029974184930324554, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3885931670665741, + 0.1150321438908577, + 0.041486434638500214, + 0.03422551974654198, + 0.07177470624446869, + 0.07025919109582901, + 0.09156882762908936, + 0.10830675065517426, + 0.05140146613121033, + 0.02735179103910923 + ], + "credit_map": [ + 0.04087373986840248, + 0.03916604816913605, + 0.0645066499710083, + 0.043793559074401855, + 0.054164279252290726, + 0.0398506224155426, + 0.05605591461062431, + 0.05018237233161926, + 0.012304430827498436, + 0.010431665927171707, + 0.01775715872645378, + 0.01343325525522232, + 0.014647324569523335, + 0.012279069051146507, + 0.018380269408226013, + 0.015798965469002724, + 0.005505271255970001, + 0.004390002228319645, + 0.007238726131618023, + 0.004214248154312372, + 0.005801453720778227, + 0.004492095671594143, + 0.004417408723384142, + 0.005427229218184948, + 0.004808914847671986, + 0.003058330388739705, + 0.007486211135983467, + 0.003560349578037858, + 0.004254089202731848, + 0.003869134932756424, + 0.003577051218599081, + 0.0036114377435296774, + 0.008087772876024246, + 0.010113738477230072, + 0.010661356151103973, + 0.012748608365654945, + 0.007592708338052034, + 0.008741126395761967, + 0.007062894757837057, + 0.006766504142433405, + 0.006924948189407587, + 0.010398752987384796, + 0.00854832585901022, + 0.01416687574237585, + 0.00698847696185112, + 0.008743244223296642, + 0.007241853978484869, + 0.007246715947985649, + 0.011249595321714878, + 0.00924752652645111, + 0.011610927991569042, + 0.011679905466735363, + 0.01225232519209385, + 0.01042838767170906, + 0.01260281726717949, + 0.01249734591692686, + 0.013379079289734364, + 0.012724513188004494, + 0.020767152309417725, + 0.014189622364938259, + 0.01335483230650425, + 0.012139859609305859, + 0.010003465227782726, + 0.01174822635948658, + 0.005195061210542917, + 0.005108184181153774, + 0.006146905943751335, + 0.007445985451340675, + 0.006917097605764866, + 0.006667350884526968, + 0.004964146763086319, + 0.008956730365753174, + 0.0026255848351866007, + 0.0027723568491637707, + 0.004186204634606838, + 0.0029217081610113382, + 0.005378905218094587, + 0.0033196748699992895, + 0.0029899210203438997, + 0.003157437313348055 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 7 + ], + "selected_reward": -0.029974184930324554 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a classically appointed study with an aristocratic gentleman wearing a dark blue velvet Rococo-style jacket?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the lighting characterized by warm candlelight creating a Caravaggio-esque chiaroscuro effect?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the gentleman flick a gold coin from a mahogany desk into a crystal goblet half-filled with crimson wine?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the coin's motion depicted as a graceful golden arc through the air before landing in the goblet?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the overall composition and color palette adhering to the classical aesthetics of a Renaissance painting?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.020606597885489464 + ], + "opposite_logprob_tokens": [ + -3.895606517791748 + ], + "logprob_reward": -0.020606597885489464, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.020606597885489464, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.4163804352283478, + 0.1651092767715454, + 0.08249937742948532, + 0.04178233817219734, + 0.06413257867097855, + 0.05428153648972511, + 0.0630580261349678, + 0.03518907353281975, + 0.034744925796985626, + 0.04282241314649582 + ], + "credit_map": [ + 0.06311681866645813, + 0.04904624819755554, + 0.04642122983932495, + 0.05281573161482811, + 0.06774844229221344, + 0.04547477513551712, + 0.04730398207902908, + 0.044453199952840805, + 0.02376176044344902, + 0.018146328628063202, + 0.02993096597492695, + 0.02458476833999157, + 0.021319186314940453, + 0.01779673434793949, + 0.011977902613580227, + 0.01759163849055767, + 0.011424238793551922, + 0.009966571815311909, + 0.01995520107448101, + 0.011575045995414257, + 0.008772213011980057, + 0.00846538320183754, + 0.0050644660368561745, + 0.007276261691004038, + 0.005691244266927242, + 0.004609748721122742, + 0.009832781739532948, + 0.005267277825623751, + 0.005678557325154543, + 0.0042801834642887115, + 0.002527045551687479, + 0.0038955011405050755, + 0.008418640121817589, + 0.00966628361493349, + 0.013831691816449165, + 0.008103113621473312, + 0.006725443061441183, + 0.007082585245370865, + 0.004736771807074547, + 0.005568051245063543, + 0.007732086349278688, + 0.008118534460663795, + 0.008293520659208298, + 0.007653662469238043, + 0.006293340586125851, + 0.007049309555441141, + 0.0043783835135400295, + 0.004762695636600256, + 0.008871413767337799, + 0.007931701838970184, + 0.012856507673859596, + 0.007568188477307558, + 0.007632460445165634, + 0.008000604808330536, + 0.005079853348433971, + 0.0051172999665141106, + 0.005356773734092712, + 0.004696568939834833, + 0.00526253879070282, + 0.004155549220740795, + 0.0043070088140666485, + 0.00440604193136096, + 0.0035746123176068068, + 0.003429979784414172, + 0.0047873458825051785, + 0.003580068936571479, + 0.0049912831746041775, + 0.0049856286495924, + 0.005632929969578981, + 0.0036859510000795126, + 0.003824713407084346, + 0.003257007570937276, + 0.004489933140575886, + 0.004894694779068232, + 0.0073372065089643, + 0.005511730443686247, + 0.0072699240408837795, + 0.004586947150528431, + 0.005446423310786486, + 0.0032855523750185966 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 2 + ], + "selected_reward": -0.020606597885489464 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a classically appointed study with an aristocratic gentleman wearing a dark blue velvet Rococo-style jacket?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the lighting characterized by warm candlelight creating a Caravaggio-esque chiaroscuro effect?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the gentleman flick a gold coin from a mahogany desk into a crystal goblet half-filled with crimson wine?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the coin's motion depicted as a graceful golden arc through the air before landing in the goblet?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the overall composition and color palette adhering to the classical aesthetics of a Renaissance painting?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.1603797823190689 + ], + "opposite_logprob_tokens": [ + -1.9103797674179077 + ], + "logprob_reward": -0.1603797823190689, + "yes_no_margin_reward": 1.75, + "token_credit_reward": -0.1603797823190689, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.29755064845085144, + 0.1732521653175354, + 0.06591811031103134, + 0.05684309080243111, + 0.06460932642221451, + 0.05614730343222618, + 0.09095478802919388, + 0.04559708386659622, + 0.07539510726928711, + 0.07373232394456863 + ], + "credit_map": [ + 0.03749031201004982, + 0.029198305681347847, + 0.0452083982527256, + 0.028202183544635773, + 0.04671845585107803, + 0.03313548117876053, + 0.036123719066381454, + 0.041473813354969025, + 0.02449812740087509, + 0.018745247274637222, + 0.035589899867773056, + 0.02075362019240856, + 0.019027922302484512, + 0.01853964664041996, + 0.015472535975277424, + 0.02062518335878849, + 0.008880655281245708, + 0.006893849931657314, + 0.0122510502114892, + 0.007998796179890633, + 0.010672970674932003, + 0.007264959160238504, + 0.00532012153416872, + 0.006635710131376982, + 0.009494367986917496, + 0.005721382796764374, + 0.01120070181787014, + 0.006857200991362333, + 0.007145577110350132, + 0.00660273153334856, + 0.004154312424361706, + 0.005666821263730526, + 0.00684647262096405, + 0.009811103343963623, + 0.010961641557514668, + 0.008731807582080364, + 0.009216626174747944, + 0.00703281257301569, + 0.0054215332493186, + 0.00658733444288373, + 0.008683385327458382, + 0.007434491999447346, + 0.006831774488091469, + 0.008152349852025509, + 0.00809124018996954, + 0.007165266200900078, + 0.005532478913664818, + 0.0042563192546367645, + 0.011365268379449844, + 0.013838106766343117, + 0.011423098854720592, + 0.011408261954784393, + 0.012816342525184155, + 0.011220907792448997, + 0.009903756901621819, + 0.008979052305221558, + 0.00677637429907918, + 0.0047940476797521114, + 0.009866958484053612, + 0.0048380461521446705, + 0.006655320059508085, + 0.0055360207334160805, + 0.003777110716328025, + 0.0033532099332660437, + 0.007457729894667864, + 0.007770189084112644, + 0.008835799060761929, + 0.009952149353921413, + 0.010617317631840706, + 0.009727598167955875, + 0.00964884739369154, + 0.011385486461222172, + 0.007895086891949177, + 0.008036388084292412, + 0.013977675698697567, + 0.009800146333873272, + 0.007648169528692961, + 0.008684810250997543, + 0.009692844934761524, + 0.007997202686965466 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 6 + ], + "selected_reward": -0.1603797823190689 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a classically appointed study with an aristocratic gentleman wearing a dark blue velvet Rococo-style jacket?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the lighting characterized by warm candlelight creating a Caravaggio-esque chiaroscuro effect?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the gentleman flick a gold coin from a mahogany desk into a crystal goblet half-filled with crimson wine?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the coin's motion depicted as a graceful golden arc through the air before landing in the goblet?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the overall composition and color palette adhering to the classical aesthetics of a Renaissance painting?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.012542770244181156 + ], + "opposite_logprob_tokens": [ + -4.387542724609375 + ], + "logprob_reward": -0.012542770244181156, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.012542770244181156, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2651260495185852, + 0.17022988200187683, + 0.0687585175037384, + 0.06700742989778519, + 0.06524815410375595, + 0.05548141524195671, + 0.09050949662923813, + 0.04409098997712135, + 0.10877195000648499, + 0.06477614492177963 + ], + "credit_map": [ + 0.03427636995911598, + 0.032745666801929474, + 0.0519041083753109, + 0.027586916461586952, + 0.03050796315073967, + 0.03209328651428223, + 0.026847563683986664, + 0.029164159670472145, + 0.02039157785475254, + 0.020990539342164993, + 0.0357617624104023, + 0.02197607234120369, + 0.01637287810444832, + 0.02041674219071865, + 0.017070762813091278, + 0.01724954880774021, + 0.007972431369125843, + 0.006887765601277351, + 0.015727899968624115, + 0.010005618445575237, + 0.009307841770350933, + 0.008045737631618977, + 0.004135640803724527, + 0.006675587501376867, + 0.009856222197413445, + 0.006434926763176918, + 0.015859132632613182, + 0.008573099039494991, + 0.008231810294091702, + 0.007620470132678747, + 0.0037059856113046408, + 0.006725786253809929, + 0.007999389432370663, + 0.007916931994259357, + 0.013124978169798851, + 0.006799603812396526, + 0.009789260104298592, + 0.007355961948633194, + 0.005146200768649578, + 0.007115829270333052, + 0.00890625175088644, + 0.006374265067279339, + 0.008014863356947899, + 0.0071948315016925335, + 0.007983808405697346, + 0.006959735881537199, + 0.005132571794092655, + 0.0049150846898555756, + 0.011689080856740475, + 0.014260019175708294, + 0.01034915167838335, + 0.01021253690123558, + 0.013215165585279465, + 0.012022830545902252, + 0.010005122050642967, + 0.008755592629313469, + 0.005549308378249407, + 0.0045513189397752285, + 0.00861714780330658, + 0.005037274211645126, + 0.00648593669757247, + 0.0051167928613722324, + 0.004501251969486475, + 0.004231959115713835, + 0.010809934698045254, + 0.011711693368852139, + 0.014429386705160141, + 0.013531092554330826, + 0.014638643711805344, + 0.01342758722603321, + 0.015467927791178226, + 0.014755679294466972, + 0.007817678153514862, + 0.007471889723092318, + 0.01309920847415924, + 0.008559348993003368, + 0.006205890327692032, + 0.007981217466294765, + 0.006524479016661644, + 0.0071164327673614025 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 8 + ], + "selected_reward": -0.012542770244181156 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a classically appointed study with an aristocratic gentleman wearing a dark blue velvet Rococo-style jacket?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the lighting characterized by warm candlelight creating a Caravaggio-esque chiaroscuro effect?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the gentleman flick a gold coin from a mahogany desk into a crystal goblet half-filled with crimson wine?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the coin's motion depicted as a graceful golden arc through the air before landing in the goblet?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the overall composition and color palette adhering to the classical aesthetics of a Renaissance painting?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0036137294955551624 + ], + "opposite_logprob_tokens": [ + -5.6286139488220215 + ], + "logprob_reward": -0.0036137294955551624, + "yes_no_margin_reward": 5.625, + "token_credit_reward": -0.0036137294955551624, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.39861586689949036, + 0.13766832649707794, + 0.0941312313079834, + 0.06283999234437943, + 0.05276712030172348, + 0.0356389582157135, + 0.054581157863140106, + 0.04052974283695221, + 0.044901736080646515, + 0.07832586765289307 + ], + "credit_map": [ + 0.06046347692608833, + 0.04212432727217674, + 0.10250872373580933, + 0.045209046453237534, + 0.04325779154896736, + 0.04409591481089592, + 0.023717451840639114, + 0.03723912686109543, + 0.02106945775449276, + 0.014584778808057308, + 0.020820047706365585, + 0.018322810530662537, + 0.01527347881346941, + 0.01527866069227457, + 0.013124149292707443, + 0.01919495314359665, + 0.012521062977612019, + 0.012559236958622932, + 0.012563485652208328, + 0.013745561242103577, + 0.010361182503402233, + 0.01011255756020546, + 0.010134723968803883, + 0.012133415788412094, + 0.008685302920639515, + 0.008716112934052944, + 0.009912130422890186, + 0.007989305071532726, + 0.009327881969511509, + 0.006634254474192858, + 0.0051319957710802555, + 0.006443011574447155, + 0.007701722905039787, + 0.009179942309856415, + 0.007070067338645458, + 0.00873314868658781, + 0.005378903821110725, + 0.0065797739662230015, + 0.0037573869340121746, + 0.004366171080619097, + 0.00563019048422575, + 0.004306044429540634, + 0.005099743604660034, + 0.00520407734438777, + 0.004853018093854189, + 0.0041041080839931965, + 0.003505438333377242, + 0.0029363390058279037, + 0.007993272505700588, + 0.00569403450936079, + 0.014428921975195408, + 0.005745661444962025, + 0.005391035228967667, + 0.006447550840675831, + 0.003927675541490316, + 0.0049530090764164925, + 0.005600097589194775, + 0.005587306339293718, + 0.008485808968544006, + 0.0050497339107096195, + 0.004405107349157333, + 0.004510799888521433, + 0.002735163550823927, + 0.00415572477504611, + 0.0055269505828619, + 0.003754303092136979, + 0.010458426550030708, + 0.006505418103188276, + 0.004923410248011351, + 0.005970983300358057, + 0.0028464975766837597, + 0.004915744531899691, + 0.008353663608431816, + 0.010580590926110744, + 0.01675717905163765, + 0.009235073812305927, + 0.009794902056455612, + 0.008675477467477322, + 0.006400434300303459, + 0.008528543636202812 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 2 + ], + "selected_reward": -0.0036137294955551624 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "slow, horizontal, medium shot draws the viewer's gaze to a classically appointed study bathed in warm candlelight. An aristocratic gentleman sits elegantly at a heavy mahogany desk. He wears a dark blue velvet Rococo-style jacket, delicately embroidered at the collar and cuffs with gold. He wears an off-white brocade waistcoat, a rich white lace scarf dangling from his collar. Beneath his meticulously styled, curly white wig lies a face defined by a serene expression. A single candlelight, emanating from the side of the frame, sculpts the contours of his face and the folds of his velvet jacket in a Caravaggio-esque chiaroscuro, illuminating a gleaming gold coin lying flat on the table. He calmly extends his right hand, carefully placing the tip of his slender index finger on the coin's edge. Then, with a steady, decisive, and rhythmic movement, he delicately flicks it upward. The gold coin instantly left the table, tracing a graceful golden arc through the air before swirling towards the other end of the desk, its surface reflecting a fluid light in the candlelight. Finally, with uncanny precision, the coin landed with a crisp, \"ding\" sound into a crystal goblet half-filled with crimson wine, stirring a tiny ripple in the liquid before slowly sinking to the bottom. The entire scene, with its composition, lighting, and color, rigorously adhered to the classical aesthetics of Renaissance painting, exuding an atmosphere of tranquility, focus, and supreme control, like a masterpiece in motion.", + "seed": 30696, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000050/group_6/index_7465_seed_30696_dp_13_rank_26_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 6, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.017264262354001404, + "vlm_logprob": -0.017264262354001404, + "vlm_yes_no_margin": 4.825, + "vlm_token_credit": -0.017264262354001404, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a classically appointed study with an aristocratic gentleman wearing a dark blue velvet Rococo-style jacket?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the lighting characterized by warm candlelight creating a Caravaggio-esque chiaroscuro effect?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the gentleman flick a gold coin from a mahogany desk into a crystal goblet half-filled with crimson wine?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the coin's motion depicted as a graceful golden arc through the air before landing in the goblet?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the overall composition and color palette adhering to the classical aesthetics of a Renaissance painting?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.05665462836623192 + ], + "opposite_logprob_tokens": [ + -3.431654691696167 + ], + "logprob_reward": -0.05665462836623192, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.05665462836623192, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2063014954328537, + 0.07391125708818436, + 0.024417364969849586, + 0.025179997086524963, + 0.013758047483861446, + 0.03193412721157074, + 0.5172954797744751, + 0.022133687511086464, + 0.054080184549093246, + 0.030988363549113274 + ], + "credit_map": [ + 0.022271616384387016, + 0.03228729963302612, + 0.03706153482198715, + 0.02396170049905777, + 0.017164137214422226, + 0.0244999248534441, + 0.025032779201865196, + 0.024022499099373817, + 0.0087039889767766, + 0.007148767355829477, + 0.010716618970036507, + 0.006293870974332094, + 0.007967168465256691, + 0.01211698167026043, + 0.005800721701234579, + 0.015163133852183819, + 0.0033649897668510675, + 0.004505848046392202, + 0.004269877448678017, + 0.0023488840088248253, + 0.003020172705873847, + 0.0019673772621899843, + 0.002060969127342105, + 0.002879244042560458, + 0.0025524755474179983, + 0.004653508309274912, + 0.0036780552472919226, + 0.002172892214730382, + 0.0016625639982521534, + 0.0030253019649535418, + 0.0029451961163431406, + 0.004490004386752844, + 0.0017853480530902743, + 0.0011715588625520468, + 0.002307642949745059, + 0.0015575354918837547, + 0.0008210940286517143, + 0.0011254597920924425, + 0.0014432065654546022, + 0.0035462009254842997, + 0.003543337807059288, + 0.0033691274002194405, + 0.006107018329203129, + 0.003386939410120249, + 0.0018535663839429617, + 0.0025948588736355305, + 0.0029380435589700937, + 0.008141236379742622, + 0.0660419911146164, + 0.01912146620452404, + 0.15105721354484558, + 0.03130079805850983, + 0.037906464189291, + 0.03866200149059296, + 0.13701030611991882, + 0.036195192486047745, + 0.002376040443778038, + 0.0019944191444665194, + 0.003346394980326295, + 0.002341880463063717, + 0.0020788859110325575, + 0.004716063383966684, + 0.0023138944525271654, + 0.002966109663248062, + 0.005411044228821993, + 0.0033484718296676874, + 0.009148484095931053, + 0.006508174818009138, + 0.003327961079776287, + 0.009470654651522636, + 0.003701334586367011, + 0.013164054602384567, + 0.005190314259380102, + 0.0038465247489511967, + 0.004678257275372744, + 0.002973314840346575, + 0.0028999270871281624, + 0.003660640213638544, + 0.0038479589857161045, + 0.00389142450876534 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 0, + 1 + ], + "selected_reward": -0.05665462836623192 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a classically appointed study with an aristocratic gentleman wearing a dark blue velvet Rococo-style jacket?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the lighting characterized by warm candlelight creating a Caravaggio-esque chiaroscuro effect?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the gentleman flick a gold coin from a mahogany desk into a crystal goblet half-filled with crimson wine?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the coin's motion depicted as a graceful golden arc through the air before landing in the goblet?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the overall composition and color palette adhering to the classical aesthetics of a Renaissance painting?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.003642592579126358 + ], + "opposite_logprob_tokens": [ + -5.628642559051514 + ], + "logprob_reward": -0.003642592579126358, + "yes_no_margin_reward": 5.625, + "token_credit_reward": -0.003642592579126358, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2187146544456482, + 0.1242537871003151, + 0.05422956496477127, + 0.056299906224012375, + 0.032610729336738586, + 0.09962847083806992, + 0.08269987255334854, + 0.06838690489530563, + 0.04821942001581192, + 0.21495658159255981 + ], + "credit_map": [ + 0.021390695124864578, + 0.027005460113286972, + 0.029673336073756218, + 0.037310756742954254, + 0.022829094901680946, + 0.032175272703170776, + 0.021152911707758904, + 0.02717713639140129, + 0.014897030778229237, + 0.016203157603740692, + 0.01763261668384075, + 0.014785709790885448, + 0.015578646212816238, + 0.0138379642739892, + 0.011599203571677208, + 0.019719481468200684, + 0.007214116398245096, + 0.007098323665559292, + 0.008202187716960907, + 0.0057601891458034515, + 0.006726598832756281, + 0.00699185486882925, + 0.004438328556716442, + 0.007797971833497286, + 0.006574623752385378, + 0.007743556518107653, + 0.007941790856420994, + 0.010800209827721119, + 0.003995833918452263, + 0.006438164506107569, + 0.005198346450924873, + 0.007607387378811836, + 0.0038394094444811344, + 0.004085746593773365, + 0.004962338600307703, + 0.005627741105854511, + 0.0024966513738036156, + 0.003408742370083928, + 0.0037068624515086412, + 0.0044832429848611355, + 0.013922713696956635, + 0.013060317374765873, + 0.0168298427015543, + 0.011074970476329327, + 0.008490745909512043, + 0.008768760599195957, + 0.008856232278048992, + 0.01862490549683571, + 0.009310619905591011, + 0.006697291042655706, + 0.02116304636001587, + 0.0073171318508684635, + 0.006053602322936058, + 0.006934179458767176, + 0.018116867169737816, + 0.007107148878276348, + 0.007097175810486078, + 0.005562897771596909, + 0.010305033065378666, + 0.00948828924447298, + 0.0056806099601089954, + 0.005198969505727291, + 0.007756360340863466, + 0.017297571524977684, + 0.005599911790341139, + 0.003242125501856208, + 0.008150906302034855, + 0.0038318089209496975, + 0.0029112137854099274, + 0.006826367694884539, + 0.005185625981539488, + 0.01247146911919117, + 0.026204677298665047, + 0.01357362512499094, + 0.03837284445762634, + 0.017919406294822693, + 0.02221248485147953, + 0.018677109852433205, + 0.040166594088077545, + 0.037829868495464325 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 1 + ], + "selected_reward": -0.003642592579126358 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a classically appointed study with an aristocratic gentleman wearing a dark blue velvet Rococo-style jacket?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the lighting characterized by warm candlelight creating a Caravaggio-esque chiaroscuro effect?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the gentleman flick a gold coin from a mahogany desk into a crystal goblet half-filled with crimson wine?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the coin's motion depicted as a graceful golden arc through the air before landing in the goblet?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the overall composition and color palette adhering to the classical aesthetics of a Renaissance painting?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.018194913864135742 + ], + "opposite_logprob_tokens": [ + -4.018195152282715 + ], + "logprob_reward": -0.018194913864135742, + "yes_no_margin_reward": 4.0, + "token_credit_reward": -0.018194913864135742, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13928274810314178, + 0.134449303150177, + 0.07730314880609512, + 0.09335432946681976, + 0.026678144931793213, + 0.10819398611783981, + 0.12538747489452362, + 0.0651998445391655, + 0.20577800273895264, + 0.02437298744916916 + ], + "credit_map": [ + 0.014195818454027176, + 0.01924768276512623, + 0.022115759551525116, + 0.020450742915272713, + 0.02067994698882103, + 0.013487357646226883, + 0.012231365777552128, + 0.01687408983707428, + 0.020098084583878517, + 0.017093529924750328, + 0.019992055371403694, + 0.013518856838345528, + 0.020578278228640556, + 0.016525045037269592, + 0.012592174112796783, + 0.014051285572350025, + 0.01028657890856266, + 0.006954372860491276, + 0.011309574358165264, + 0.00795056950300932, + 0.00956658273935318, + 0.008941993117332458, + 0.009093346074223518, + 0.013200139626860619, + 0.01197332888841629, + 0.009153293445706367, + 0.017285628244280815, + 0.010414347983896732, + 0.005509215407073498, + 0.007240642327815294, + 0.01520752813667059, + 0.016570355743169785, + 0.0032634278759360313, + 0.0041059460490942, + 0.004046273417770863, + 0.0036202764604240656, + 0.0022280646953731775, + 0.002932986943051219, + 0.0022804830223321915, + 0.004200688097625971, + 0.011628832668066025, + 0.013520670123398304, + 0.015515875071287155, + 0.018260061740875244, + 0.010749720968306065, + 0.01315463799983263, + 0.016511136665940285, + 0.008853064849972725, + 0.015489613637328148, + 0.010481184348464012, + 0.03019920364022255, + 0.0109279565513134, + 0.012832408770918846, + 0.008880216628313065, + 0.028798751533031464, + 0.007778160739690065, + 0.007079279515892267, + 0.005789026618003845, + 0.008803616277873516, + 0.009151875041425228, + 0.008718542754650116, + 0.005014269147068262, + 0.007970936596393585, + 0.012672308832406998, + 0.01821298897266388, + 0.014113867655396461, + 0.03300302103161812, + 0.016535203903913498, + 0.01658632420003414, + 0.023838115856051445, + 0.020129341632127762, + 0.06335917115211487, + 0.003977290820330381, + 0.002975283656269312, + 0.0035588929895311594, + 0.002473366679623723, + 0.0034759275149554014, + 0.0028447844088077545, + 0.0025437739677727222, + 0.00252367090433836 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 0, + 1 + ], + "selected_reward": -0.018194913864135742 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a classically appointed study with an aristocratic gentleman wearing a dark blue velvet Rococo-style jacket?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the lighting characterized by warm candlelight creating a Caravaggio-esque chiaroscuro effect?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the gentleman flick a gold coin from a mahogany desk into a crystal goblet half-filled with crimson wine?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the coin's motion depicted as a graceful golden arc through the air before landing in the goblet?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the overall composition and color palette adhering to the classical aesthetics of a Renaissance painting?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004636489786207676 + ], + "opposite_logprob_tokens": [ + -5.379636287689209 + ], + "logprob_reward": -0.004636489786207676, + "yes_no_margin_reward": 5.375, + "token_credit_reward": -0.004636489786207676, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1720210611820221, + 0.05241008102893829, + 0.042977891862392426, + 0.07904679328203201, + 0.043982550501823425, + 0.11176837980747223, + 0.3406364321708679, + 0.06406114250421524, + 0.06269388645887375, + 0.030401768162846565 + ], + "credit_map": [ + 0.019565528258681297, + 0.02432839199900627, + 0.030151454731822014, + 0.02535797841846943, + 0.01825336180627346, + 0.022055713459849358, + 0.014144864864647388, + 0.018163781613111496, + 0.007448036223649979, + 0.0066014803014695644, + 0.007843010127544403, + 0.006724324077367783, + 0.007971336133778095, + 0.004939972888678312, + 0.005836718715727329, + 0.005045201629400253, + 0.006539713125675917, + 0.0030944105237722397, + 0.007849300280213356, + 0.005624089390039444, + 0.003285432467237115, + 0.004147019237279892, + 0.005067440681159496, + 0.007370483595877886, + 0.01141588669270277, + 0.010571632534265518, + 0.013772363774478436, + 0.00900961086153984, + 0.00514069152995944, + 0.005791004281491041, + 0.010808536782860756, + 0.012537066824734211, + 0.005028915591537952, + 0.006658065598458052, + 0.006736909970641136, + 0.007026270963251591, + 0.0031438893638551235, + 0.005383338779211044, + 0.004611683543771505, + 0.005393472965806723, + 0.011456659063696861, + 0.014026558957993984, + 0.016456374898552895, + 0.018598275259137154, + 0.010164367966353893, + 0.010805553756654263, + 0.015553920529782772, + 0.014706664718687534, + 0.045172177255153656, + 0.016974136233329773, + 0.09952095150947571, + 0.021649032831192017, + 0.028400909155607224, + 0.01457715779542923, + 0.09306919574737549, + 0.021272871643304825, + 0.007536029908806086, + 0.006050468422472477, + 0.008553675375878811, + 0.012131499126553535, + 0.0055551775731146336, + 0.008615842089056969, + 0.008362618274986744, + 0.007255835458636284, + 0.006054253783077002, + 0.006319242529571056, + 0.008778784424066544, + 0.008247253485023975, + 0.008452453650534153, + 0.007820267230272293, + 0.007007314823567867, + 0.010014318861067295, + 0.003699239809066057, + 0.003065509023144841, + 0.004106717184185982, + 0.003372946521267295, + 0.003656514920294285, + 0.0026631539221853018, + 0.005775316618382931, + 0.004062370862811804 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 0, + 5 + ], + "selected_reward": -0.004636489786207676 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a classically appointed study with an aristocratic gentleman wearing a dark blue velvet Rococo-style jacket?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the lighting characterized by warm candlelight creating a Caravaggio-esque chiaroscuro effect?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the gentleman flick a gold coin from a mahogany desk into a crystal goblet half-filled with crimson wine?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the coin's motion depicted as a graceful golden arc through the air before landing in the goblet?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the overall composition and color palette adhering to the classical aesthetics of a Renaissance painting?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.00319268717430532 + ], + "opposite_logprob_tokens": [ + -5.753192901611328 + ], + "logprob_reward": -0.00319268717430532, + "yes_no_margin_reward": 5.75, + "token_credit_reward": -0.00319268717430532, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2882091701030731, + 0.0946069210767746, + 0.03879301995038986, + 0.07482968270778656, + 0.03929826617240906, + 0.12338121235370636, + 0.14549511671066284, + 0.05764470994472504, + 0.0396864227950573, + 0.0980554074048996 + ], + "credit_map": [ + 0.03029012493789196, + 0.02933403104543686, + 0.04640519991517067, + 0.039767272770404816, + 0.047809869050979614, + 0.026297904551029205, + 0.04598524421453476, + 0.022319549694657326, + 0.010645449161529541, + 0.009763019159436226, + 0.013731657527387142, + 0.019753865897655487, + 0.011575776152312756, + 0.008220957592129707, + 0.008357786573469639, + 0.012558410875499249, + 0.004095120821148157, + 0.00464565260335803, + 0.005802375730127096, + 0.006776471622288227, + 0.003903930075466633, + 0.004634005483239889, + 0.005336679518222809, + 0.003598783863708377, + 0.010213639587163925, + 0.014265956357121468, + 0.010284636169672012, + 0.014699583873152733, + 0.004590019583702087, + 0.005579723045229912, + 0.006564561743289232, + 0.008631564676761627, + 0.005347765516489744, + 0.005049362313002348, + 0.006901592016220093, + 0.0052400920540094376, + 0.0027164602652192116, + 0.0043391017243266106, + 0.004408699460327625, + 0.005295194685459137, + 0.013713416643440723, + 0.009783073328435421, + 0.02158837951719761, + 0.014001407660543919, + 0.007372788153588772, + 0.018470399081707, + 0.01449353713542223, + 0.02395820990204811, + 0.01873904839158058, + 0.009569973684847355, + 0.040005870163440704, + 0.009941581636667252, + 0.013650435954332352, + 0.008489117957651615, + 0.03502773120999336, + 0.010071341879665852, + 0.006492031738162041, + 0.005750310607254505, + 0.008270623162388802, + 0.0066644069738686085, + 0.007713173516094685, + 0.004998072981834412, + 0.006470214109867811, + 0.011285880580544472, + 0.004803551826626062, + 0.004033366683870554, + 0.005870944354683161, + 0.0042473748326301575, + 0.003942321985960007, + 0.007552759721875191, + 0.004287435673177242, + 0.0049486663192510605, + 0.013120410963892937, + 0.012849315069615841, + 0.016060059890151024, + 0.00959768332540989, + 0.008705765008926392, + 0.007970528677105904, + 0.011068493127822876, + 0.01868315041065216 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 6, + 5 + ], + "selected_reward": -0.00319268717430532 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000050/group_7/index_tensor([5162, 5162])_group_7_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000050/group_7/index_tensor([5162, 5162])_group_7_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..353af83b099fa84c764f4cd60d42abfae127629f --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000050/group_7/index_tensor([5162, 5162])_group_7_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "shot that begins with a wide angle and slowly zooms out, immersing the viewer in a deep, nearly pitch-black expanse of space, a backdrop dotted with distant, faint starlight. In the center of the frame, thousands of translucent ice-blue crystal fragments of varying sizes and shapes converge at high speed from all directions, as if drawn by an invisible gravitational force. Each fragment's surface is etched with the intricate contours of continents or oceans. They trail brilliant blue light trails as they dart, swirl, and tumble through the darkness, forming a dynamic, chaotic web of light. Then, as if on unified command, the fragments swiftly and decisively slow down and find their predetermined positions. With each perfectly precise rendezvous, a brief, dazzling flash of energy erupts, ultimately, within seconds, uniting into a complete, precise crystal globe radiating a soft, inner glow. As the camera continues to zoom out, this crystal clear blue planet quietly floats and slowly rotates against the dark background, and its internal continental plate structure is clearly visible. The entire picture, set against the backdrop of cold colors, creates a grand, magical and futuristic technological atmosphere of creation.", + "seed": 10638, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000050/group_7/index_5162_seed_10638_dp_14_rank_28_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 7, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.3268598140915856, + "vlm_logprob": -0.3268598140915856, + "vlm_yes_no_margin": 4.075, + "vlm_token_credit": -0.3268598140915856, + "vlm_yes_ratio": 0.8, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the video begin with a wide angle shot that slowly zooms out?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are thousands of translucent ice-blue crystal fragments visible converging from all directions at high speed?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the crystal fragments trail brilliant blue light as they move?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the fragments slow down, flash with energy, and unite to form a complete crystal globe with visible internal continental structures?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the final scene show the crystal globe floating and rotating against a dark background with faint starlight?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -1.605546474456787 + ], + "opposite_logprob_tokens": [ + -0.23054653406143188 + ], + "logprob_reward": -1.605546474456787, + "yes_no_margin_reward": -1.375, + "token_credit_reward": -1.605546474456787, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.19910936057567596, + 0.07862687855958939, + 0.06970337778329849, + 0.031059978529810905, + 0.03831521421670914, + 0.05467880889773369, + 0.36864256858825684, + 0.03451982140541077, + 0.1014709398150444, + 0.0238731000572443 + ], + "credit_map": [ + 0.01640043593943119, + 0.015100802294909954, + 0.026489637792110443, + 0.013502572663128376, + 0.04881739616394043, + 0.016850996762514114, + 0.016459329053759575, + 0.04548817500472069, + 0.007211398798972368, + 0.007841079495847225, + 0.006767331622540951, + 0.007603575009852648, + 0.007536967284977436, + 0.017188237980008125, + 0.009499452076852322, + 0.01497883815318346, + 0.007425228599458933, + 0.008136841468513012, + 0.011937003582715988, + 0.011460944078862667, + 0.006379778031259775, + 0.007361959666013718, + 0.006991260685026646, + 0.010010363534092903, + 0.003948078490793705, + 0.0037454331759363413, + 0.004467966966331005, + 0.004451428540050983, + 0.003507080255076289, + 0.002317688660696149, + 0.0051262471824884415, + 0.003496055491268635, + 0.003861085744574666, + 0.004874971695244312, + 0.006653169170022011, + 0.006438981741666794, + 0.005105182994157076, + 0.00244081299751997, + 0.0048076617531478405, + 0.0041333516128361225, + 0.007464143913239241, + 0.004336090758442879, + 0.012145834043622017, + 0.004336764570325613, + 0.0036366693675518036, + 0.012351484037935734, + 0.0044565508142113686, + 0.00595126673579216, + 0.041460826992988586, + 0.011866589076817036, + 0.11803857237100601, + 0.018555384129285812, + 0.014195333234965801, + 0.042167484760284424, + 0.014375745318830013, + 0.10798260569572449, + 0.0052625685930252075, + 0.003350189421325922, + 0.0047539714723825455, + 0.0026204714085906744, + 0.0022917822934687138, + 0.005624713376164436, + 0.003060361370444298, + 0.007555765099823475, + 0.014563219621777534, + 0.007237633224576712, + 0.013271911069750786, + 0.005639747250825167, + 0.0082668736577034, + 0.02052764594554901, + 0.008351937867701054, + 0.023611973971128464, + 0.0024807341396808624, + 0.00099347869399935, + 0.002425090642645955, + 0.0008978411788120866, + 0.0011540576815605164, + 0.00853542611002922, + 0.001615938264876604, + 0.005770531948655844 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 0, + 8 + ], + "selected_reward": -1.605546474456787 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the video begin with a wide angle shot that slowly zooms out?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are thousands of translucent ice-blue crystal fragments visible converging from all directions at high speed?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the crystal fragments trail brilliant blue light as they move?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the fragments slow down, flash with energy, and unite to form a complete crystal globe with visible internal continental structures?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the final scene show the crystal globe floating and rotating against a dark background with faint starlight?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.012580794282257557 + ], + "opposite_logprob_tokens": [ + -4.387580871582031 + ], + "logprob_reward": -0.012580794282257557, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.012580794282257557, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09555234014987946, + 0.05550017207860947, + 0.07113702595233917, + 0.04662239924073219, + 0.08019600808620453, + 0.05719081312417984, + 0.5255608558654785, + 0.019299134612083435, + 0.03968288004398346, + 0.009258401580154896 + ], + "credit_map": [ + 0.009187744930386543, + 0.006231487728655338, + 0.005546503234654665, + 0.006172089837491512, + 0.007735332474112511, + 0.023735150694847107, + 0.008485324680805206, + 0.02845870330929756, + 0.00792919471859932, + 0.00540470564737916, + 0.007850551046431065, + 0.006570232100784779, + 0.0054636625573039055, + 0.008452733047306538, + 0.005305023863911629, + 0.008524070493876934, + 0.00785062275826931, + 0.007156810723245144, + 0.007695548236370087, + 0.01465725526213646, + 0.008407753892242908, + 0.007189921103417873, + 0.010068637318909168, + 0.00811048038303852, + 0.007493219804018736, + 0.0052514090202748775, + 0.007814141921699047, + 0.005398562643676996, + 0.006619056686758995, + 0.003790829796344042, + 0.0049484772607684135, + 0.005306700710207224, + 0.015555083751678467, + 0.0093622962012887, + 0.0135972760617733, + 0.007610871456563473, + 0.00822874903678894, + 0.005426432937383652, + 0.008266662247478962, + 0.012148634530603886, + 0.009531461633741856, + 0.006571544334292412, + 0.010578404180705547, + 0.0058615487068891525, + 0.004985518753528595, + 0.006028851959854364, + 0.00616579269990325, + 0.007467692252248526, + 0.05678928643465042, + 0.019925490021705627, + 0.1554066240787506, + 0.032733142375946045, + 0.024174313992261887, + 0.13956089317798615, + 0.02701321430504322, + 0.06995789706707001, + 0.001814488903619349, + 0.0017245429335162044, + 0.0015913103707134724, + 0.0016486257081851363, + 0.0014634464168921113, + 0.004934209864586592, + 0.0023427228443324566, + 0.0037797875702381134, + 0.0050277868285775185, + 0.0024191378615796566, + 0.004787846002727747, + 0.0020296897273510695, + 0.0029307841323316097, + 0.009909319691359997, + 0.0029975897632539272, + 0.009580724872648716, + 0.0012216484174132347, + 0.0006712977192364633, + 0.0011610900983214378, + 0.0008046987932175398, + 0.0007191990152932703, + 0.002118032891303301, + 0.0009913587709888816, + 0.0015710764564573765 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 0, + 4 + ], + "selected_reward": -0.012580794282257557 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the video begin with a wide angle shot that slowly zooms out?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are thousands of translucent ice-blue crystal fragments visible converging from all directions at high speed?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the crystal fragments trail brilliant blue light as they move?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the fragments slow down, flash with energy, and unite to form a complete crystal globe with visible internal continental structures?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the final scene show the crystal globe floating and rotating against a dark background with faint starlight?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.002822346519678831 + ], + "opposite_logprob_tokens": [ + -5.877822399139404 + ], + "logprob_reward": -0.002822346519678831, + "yes_no_margin_reward": 5.875, + "token_credit_reward": -0.002822346519678831, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1234217956662178, + 0.04218420386314392, + 0.11145107448101044, + 0.03575630858540535, + 0.06419263035058975, + 0.08545751869678497, + 0.37412935495376587, + 0.032627690583467484, + 0.1016811802983284, + 0.02909824438393116 + ], + "credit_map": [ + 0.009409968741238117, + 0.007248887326568365, + 0.02601577155292034, + 0.010333796963095665, + 0.03075881488621235, + 0.008473063819110394, + 0.01226375624537468, + 0.018917741253972054, + 0.005250029731541872, + 0.004357327241450548, + 0.006699108984321356, + 0.0055329278111457825, + 0.0045394585467875, + 0.005936257541179657, + 0.004559547174721956, + 0.005309547297656536, + 0.011707275174558163, + 0.009727958589792252, + 0.012777786701917648, + 0.019881471991539, + 0.013083050027489662, + 0.01184664387255907, + 0.01791205443441868, + 0.014514835551381111, + 0.005054313689470291, + 0.004472621716558933, + 0.007066992577165365, + 0.0040541463531553745, + 0.0038641735445708036, + 0.0033427802845835686, + 0.0033332270104438066, + 0.004568053409457207, + 0.008404439315199852, + 0.008676139637827873, + 0.009009769186377525, + 0.007750213146209717, + 0.006556675769388676, + 0.00580985751003027, + 0.00772833963856101, + 0.010257191956043243, + 0.013767625205218792, + 0.008471802808344364, + 0.014926621690392494, + 0.008160930126905441, + 0.00612550089135766, + 0.01382787711918354, + 0.00811531487852335, + 0.012061848305165768, + 0.03968659043312073, + 0.014807038009166718, + 0.11473430693149567, + 0.022383378818631172, + 0.017561228945851326, + 0.07736078649759293, + 0.01859615370631218, + 0.06899987161159515, + 0.003192923730239272, + 0.002384983468800783, + 0.003401400987058878, + 0.0028161320369690657, + 0.002488326281309128, + 0.008655746467411518, + 0.003677045926451683, + 0.0060111284255981445, + 0.016825193539261818, + 0.007150986697524786, + 0.014301245100796223, + 0.005968633573502302, + 0.008615203201770782, + 0.019102109596133232, + 0.009224439039826393, + 0.020493369549512863, + 0.001897746347822249, + 0.001331231789663434, + 0.001954994397237897, + 0.001333829713985324, + 0.0013225453440099955, + 0.009651733562350273, + 0.001942545990459621, + 0.009663617238402367 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 0, + 2 + ], + "selected_reward": -0.002822346519678831 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the video begin with a wide angle shot that slowly zooms out?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are thousands of translucent ice-blue crystal fragments visible converging from all directions at high speed?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the crystal fragments trail brilliant blue light as they move?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the fragments slow down, flash with energy, and unite to form a complete crystal globe with visible internal continental structures?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the final scene show the crystal globe floating and rotating against a dark background with faint starlight?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0008137490367516875 + ], + "opposite_logprob_tokens": [ + -7.125813961029053 + ], + "logprob_reward": -0.0008137490367516875, + "yes_no_margin_reward": 7.125, + "token_credit_reward": -0.0008137490367516875, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1267639696598053, + 0.07674209028482437, + 0.0837922915816307, + 0.03907899558544159, + 0.08873680233955383, + 0.08913356810808182, + 0.24797731637954712, + 0.06118486821651459, + 0.08388251811265945, + 0.10270755738019943 + ], + "credit_map": [ + 0.01093029323965311, + 0.009653839282691479, + 0.010633580386638641, + 0.013078398071229458, + 0.028778312727808952, + 0.014088957570493221, + 0.016462286934256554, + 0.023138297721743584, + 0.008197417482733727, + 0.004660486709326506, + 0.007197869475930929, + 0.006257007364183664, + 0.006506064906716347, + 0.018770722672343254, + 0.006432739552110434, + 0.018719783052802086, + 0.008700309321284294, + 0.00826902687549591, + 0.008025070652365685, + 0.011440461501479149, + 0.01007700152695179, + 0.010060450062155724, + 0.014716229401528835, + 0.01250374410301447, + 0.005852338392287493, + 0.004192587919533253, + 0.00469361012801528, + 0.004647144582122564, + 0.005267406813800335, + 0.005262348800897598, + 0.004530246835201979, + 0.004633311182260513, + 0.01297821570187807, + 0.01320145558565855, + 0.012489108368754387, + 0.011890538036823273, + 0.011246508918702602, + 0.005118589382618666, + 0.011134355328977108, + 0.010678032413125038, + 0.011685973964631557, + 0.009048253297805786, + 0.012300909496843815, + 0.008139288052916527, + 0.006529731675982475, + 0.0169683825224638, + 0.00900694914162159, + 0.015454071573913097, + 0.03019976243376732, + 0.009328390471637249, + 0.07491052895784378, + 0.0135946711525321, + 0.01077701710164547, + 0.03602725639939308, + 0.010670678690075874, + 0.06246902421116829, + 0.009782323613762856, + 0.00548141822218895, + 0.007922360673546791, + 0.004071811214089394, + 0.006255762185901403, + 0.010092176496982574, + 0.006803791504353285, + 0.01077522523701191, + 0.01156721729785204, + 0.005773133132606745, + 0.01137415412813425, + 0.005509324837476015, + 0.0069346376694738865, + 0.01794087328016758, + 0.008234464563429356, + 0.01654870994389057, + 0.015904784202575684, + 0.005542371422052383, + 0.012243419885635376, + 0.006591347046196461, + 0.005855261348187923, + 0.028489859774708748, + 0.011569015681743622, + 0.016511494293808937 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 0, + 9 + ], + "selected_reward": -0.0008137490367516875 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the video begin with a wide angle shot that slowly zooms out?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are thousands of translucent ice-blue crystal fragments visible converging from all directions at high speed?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the crystal fragments trail brilliant blue light as they move?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the fragments slow down, flash with energy, and unite to form a complete crystal globe with visible internal continental structures?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the final scene show the crystal globe floating and rotating against a dark background with faint starlight?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.012535706162452698 + ], + "opposite_logprob_tokens": [ + -4.387535572052002 + ], + "logprob_reward": -0.012535706162452698, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.012535706162452698, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10703141242265701, + 0.04987618327140808, + 0.05667278543114662, + 0.035522837191820145, + 0.05418403446674347, + 0.1744832545518875, + 0.23924629390239716, + 0.11480651050806046, + 0.13039404153823853, + 0.03778263181447983 + ], + "credit_map": [ + 0.008870274759829044, + 0.007802105974406004, + 0.03003191202878952, + 0.012058897875249386, + 0.00819085631519556, + 0.007704836782068014, + 0.009094472974538803, + 0.023278065025806427, + 0.004288832191377878, + 0.004102571401745081, + 0.0037465922068804502, + 0.0056944782845675945, + 0.004093356896191835, + 0.011998485773801804, + 0.004792206920683384, + 0.011159660294651985, + 0.00503897201269865, + 0.005748589523136616, + 0.00489979749545455, + 0.008075137622654438, + 0.00795870553702116, + 0.006639828439801931, + 0.010537317954003811, + 0.0077744340524077415, + 0.005270475056022406, + 0.003267841413617134, + 0.005724016111344099, + 0.003754307283088565, + 0.00447918800637126, + 0.004920978099107742, + 0.004284551367163658, + 0.003821482416242361, + 0.00850350596010685, + 0.006499553564935923, + 0.010592482052743435, + 0.005177494138479233, + 0.005089888349175453, + 0.0058340043760836124, + 0.005306567996740341, + 0.007180534768849611, + 0.02494867891073227, + 0.01538885198533535, + 0.03982851281762123, + 0.011614027433097363, + 0.010639449581503868, + 0.04023626446723938, + 0.01246259268373251, + 0.01936488412320614, + 0.046088509261608124, + 0.01306032482534647, + 0.03735584765672684, + 0.011457913555204868, + 0.012840792536735535, + 0.04084775224328041, + 0.011510361917316914, + 0.06608479470014572, + 0.021741950884461403, + 0.008237418718636036, + 0.015738489106297493, + 0.006693082861602306, + 0.009688280522823334, + 0.024320220574736595, + 0.009842886589467525, + 0.018544184044003487, + 0.023207737132906914, + 0.00903342291712761, + 0.01861991174519062, + 0.007657038047909737, + 0.011278302408754826, + 0.024878837168216705, + 0.01234784908592701, + 0.023370938375592232, + 0.006571635138243437, + 0.003011775901541114, + 0.006637677550315857, + 0.00247026770375669, + 0.0027104762848466635, + 0.004928889684379101, + 0.0036438708193600178, + 0.007808037102222443 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 5, + 8 + ], + "selected_reward": -0.012535706162452698 + } + ], + "vlm_decoded_outputs": [ + "No", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "shot that begins with a wide angle and slowly zooms out, immersing the viewer in a deep, nearly pitch-black expanse of space, a backdrop dotted with distant, faint starlight. In the center of the frame, thousands of translucent ice-blue crystal fragments of varying sizes and shapes converge at high speed from all directions, as if drawn by an invisible gravitational force. Each fragment's surface is etched with the intricate contours of continents or oceans. They trail brilliant blue light trails as they dart, swirl, and tumble through the darkness, forming a dynamic, chaotic web of light. Then, as if on unified command, the fragments swiftly and decisively slow down and find their predetermined positions. With each perfectly precise rendezvous, a brief, dazzling flash of energy erupts, ultimately, within seconds, uniting into a complete, precise crystal globe radiating a soft, inner glow. As the camera continues to zoom out, this crystal clear blue planet quietly floats and slowly rotates against the dark background, and its internal continental plate structure is clearly visible. The entire picture, set against the backdrop of cold colors, creates a grand, magical and futuristic technological atmosphere of creation.", + "seed": 10639, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000050/group_7/index_5162_seed_10639_dp_14_rank_28_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 7, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.21799173888284712, + "vlm_logprob": -0.21799173888284712, + "vlm_yes_no_margin": 4.125, + "vlm_token_credit": -0.21799173888284712, + "vlm_yes_ratio": 0.8, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the video begin with a wide angle shot that slowly zooms out?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are thousands of translucent ice-blue crystal fragments visible converging from all directions at high speed?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the crystal fragments trail brilliant blue light as they move?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the fragments slow down, flash with energy, and unite to form a complete crystal globe with visible internal continental structures?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the final scene show the crystal globe floating and rotating against a dark background with faint starlight?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -1.065081238746643 + ], + "opposite_logprob_tokens": [ + -0.4400812089443207 + ], + "logprob_reward": -1.065081238746643, + "yes_no_margin_reward": -0.625, + "token_credit_reward": -1.065081238746643, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.07207345217466354, + 0.03027067892253399, + 0.045861098915338516, + 0.04562382027506828, + 0.04528366029262543, + 0.4765803813934326, + 0.09956041723489761, + 0.06521596014499664, + 0.04183564707636833, + 0.07769492268562317 + ], + "credit_map": [ + 0.00495935371145606, + 0.006368402857333422, + 0.008973581716418266, + 0.0128920404240489, + 0.007255176547914743, + 0.011677669361233711, + 0.01035211980342865, + 0.009595104493200779, + 0.002813195576891303, + 0.0037795891985297203, + 0.0030130213126540184, + 0.005533651448786259, + 0.0031574545428156853, + 0.003474997356534004, + 0.0037362342700362206, + 0.0047625331208109856, + 0.005339343566447496, + 0.007165173999965191, + 0.004627869930118322, + 0.005273289512842894, + 0.006939483340829611, + 0.005470776930451393, + 0.007046462036669254, + 0.003998700994998217, + 0.0031012948602437973, + 0.004048540256917477, + 0.005022489000111818, + 0.009474455378949642, + 0.004554852843284607, + 0.0035152090713381767, + 0.012178461998701096, + 0.003728513140231371, + 0.009542783722281456, + 0.003816191805526614, + 0.004687846172600985, + 0.004326515831053257, + 0.0030955802649259567, + 0.0049917069263756275, + 0.004450013395398855, + 0.010373023338615894, + 0.04048508405685425, + 0.057849522680044174, + 0.043988946825265884, + 0.05779657140374184, + 0.09599612653255463, + 0.053215622901916504, + 0.09233089536428452, + 0.034917596727609634, + 0.009901841171085835, + 0.010464534163475037, + 0.015486938878893852, + 0.013195288367569447, + 0.011367660015821457, + 0.01585117168724537, + 0.01445564441382885, + 0.008837337605655193, + 0.008591520600020885, + 0.006636394653469324, + 0.012715485878288746, + 0.0069679077714681625, + 0.0065255314111709595, + 0.011123057454824448, + 0.00478315819054842, + 0.00787290371954441, + 0.005340482108294964, + 0.003615791443735361, + 0.007404245436191559, + 0.0031077577732503414, + 0.003928658086806536, + 0.011810575611889362, + 0.002714363392442465, + 0.003913773223757744, + 0.00633532740175724, + 0.00639411062002182, + 0.014766436070203781, + 0.007659319322556257, + 0.008494321256875992, + 0.02002241648733616, + 0.006347251124680042, + 0.0076757436618208885 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 6, + 9 + ], + "selected_reward": -1.065081238746643 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the video begin with a wide angle shot that slowly zooms out?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are thousands of translucent ice-blue crystal fragments visible converging from all directions at high speed?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the crystal fragments trail brilliant blue light as they move?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the fragments slow down, flash with energy, and unite to form a complete crystal globe with visible internal continental structures?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the final scene show the crystal globe floating and rotating against a dark background with faint starlight?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.008665576577186584 + ], + "opposite_logprob_tokens": [ + -4.758665561676025 + ], + "logprob_reward": -0.008665576577186584, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.008665576577186584, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06288576126098633, + 0.11159893870353699, + 0.08540575951337814, + 0.026249393820762634, + 0.32538700103759766, + 0.21782974898815155, + 0.04135087877511978, + 0.05354779213666916, + 0.02933432161808014, + 0.046410348266363144 + ], + "credit_map": [ + 0.005040682852268219, + 0.00499624852091074, + 0.009477258659899235, + 0.008126567117869854, + 0.005872884765267372, + 0.00959964282810688, + 0.0077779837884008884, + 0.011994488537311554, + 0.013365818187594414, + 0.013150654733181, + 0.026937581598758698, + 0.01769668608903885, + 0.010339256376028061, + 0.008022469468414783, + 0.012456320226192474, + 0.00963014829903841, + 0.008551245555281639, + 0.010348441079258919, + 0.011966423131525517, + 0.011115550994873047, + 0.009904032573103905, + 0.007499159313738346, + 0.012793178670108318, + 0.013227725401520729, + 0.0020047121215611696, + 0.003715359605848789, + 0.002697631251066923, + 0.0032598432153463364, + 0.0048442017287015915, + 0.0027795669157058, + 0.004378599114716053, + 0.0025694798678159714, + 0.030765041708946228, + 0.014912787824869156, + 0.012944824062287807, + 0.020195232704281807, + 0.017331959679722786, + 0.11886237561702728, + 0.01809648610651493, + 0.09227830171585083, + 0.018882185220718384, + 0.0250407587736845, + 0.021795213222503662, + 0.02458280697464943, + 0.042466484010219574, + 0.024950485676527023, + 0.042671140283346176, + 0.017440687865018845, + 0.00372322928160429, + 0.004945850465446711, + 0.005063313990831375, + 0.005703842267394066, + 0.005191436503082514, + 0.006259080022573471, + 0.006603456102311611, + 0.0038606729358434677, + 0.006040800362825394, + 0.004808668978512287, + 0.015535751357674599, + 0.005286514759063721, + 0.005543138831853867, + 0.007026040460914373, + 0.0032279016450047493, + 0.006078972481191158, + 0.004419873934239149, + 0.002459282288327813, + 0.004012252204120159, + 0.0023224628530442715, + 0.0027563294861465693, + 0.005411186721175909, + 0.0021288448479026556, + 0.005824089050292969, + 0.00381060061044991, + 0.00417593028396368, + 0.007047191262245178, + 0.006001543719321489, + 0.005716368090361357, + 0.00812100525945425, + 0.004626709967851639, + 0.006911003030836582 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 5, + 1 + ], + "selected_reward": -0.008665576577186584 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the video begin with a wide angle shot that slowly zooms out?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are thousands of translucent ice-blue crystal fragments visible converging from all directions at high speed?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the crystal fragments trail brilliant blue light as they move?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the fragments slow down, flash with energy, and unite to form a complete crystal globe with visible internal continental structures?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the final scene show the crystal globe floating and rotating against a dark background with faint starlight?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.003197202691808343 + ], + "opposite_logprob_tokens": [ + -5.753197193145752 + ], + "logprob_reward": -0.003197202691808343, + "yes_no_margin_reward": 5.75, + "token_credit_reward": -0.003197202691808343, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.05064142867922783, + 0.13723434507846832, + 0.05938443914055824, + 0.020070170983672142, + 0.13754425942897797, + 0.33831995725631714, + 0.06767203658819199, + 0.048274654895067215, + 0.03205280005931854, + 0.10880592465400696 + ], + "credit_map": [ + 0.004530505742877722, + 0.004369029309600592, + 0.00783518236130476, + 0.007040912285447121, + 0.004551304969936609, + 0.008469498716294765, + 0.006054156459867954, + 0.007790838833898306, + 0.014581212773919106, + 0.011455027386546135, + 0.039073508232831955, + 0.04600587487220764, + 0.00640166737139225, + 0.004538009874522686, + 0.00908423587679863, + 0.006094809155911207, + 0.00739968242123723, + 0.007425767369568348, + 0.007045896258205175, + 0.006750848609954119, + 0.006238851696252823, + 0.005317667033523321, + 0.010084661655128002, + 0.009121065959334373, + 0.0018947317730635405, + 0.002854828955605626, + 0.002700755838304758, + 0.003145659575238824, + 0.00304711633361876, + 0.0013888106914237142, + 0.003295938717201352, + 0.0017423294484615326, + 0.013323966413736343, + 0.004565484821796417, + 0.004360793624073267, + 0.005589643027633429, + 0.006849889177829027, + 0.0519549660384655, + 0.008277153596282005, + 0.042622365057468414, + 0.029350610449910164, + 0.039410803467035294, + 0.03074849583208561, + 0.03993379324674606, + 0.06811337172985077, + 0.039727818220853806, + 0.06543680280447006, + 0.025598278269171715, + 0.005889959633350372, + 0.007335126865655184, + 0.00977562740445137, + 0.008719989098608494, + 0.007328798528760672, + 0.014770866371691227, + 0.007823060266673565, + 0.006028605159372091, + 0.0061280555091798306, + 0.0035204438026994467, + 0.014947877265512943, + 0.003979949280619621, + 0.003387351520359516, + 0.007781692780554295, + 0.0035316648427397013, + 0.00499762361869216, + 0.004076011013239622, + 0.003504412015900016, + 0.004167363978922367, + 0.003673068480566144, + 0.004065424669533968, + 0.005197176709771156, + 0.0027272554580122232, + 0.00464208796620369, + 0.010070115327835083, + 0.006880528759211302, + 0.017917513847351074, + 0.011749827302992344, + 0.009681970812380314, + 0.02370757795870304, + 0.009257595986127853, + 0.01954079046845436 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 4, + 1 + ], + "selected_reward": -0.003197202691808343 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the video begin with a wide angle shot that slowly zooms out?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are thousands of translucent ice-blue crystal fragments visible converging from all directions at high speed?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the crystal fragments trail brilliant blue light as they move?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the fragments slow down, flash with energy, and unite to form a complete crystal globe with visible internal continental structures?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the final scene show the crystal globe floating and rotating against a dark background with faint starlight?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0019418209558352828 + ], + "opposite_logprob_tokens": [ + -6.251941680908203 + ], + "logprob_reward": -0.0019418209558352828, + "yes_no_margin_reward": 6.25, + "token_credit_reward": -0.0019418209558352828, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07290961593389511, + 0.11085119098424911, + 0.050280071794986725, + 0.043586309999227524, + 0.3368012607097626, + 0.03818535804748535, + 0.10194867104291916, + 0.11951138824224472, + 0.030446434393525124, + 0.09547973424196243 + ], + "credit_map": [ + 0.005070322193205357, + 0.008640707470476627, + 0.007780556567013264, + 0.008960478007793427, + 0.008967535570263863, + 0.014164653606712818, + 0.007847713306546211, + 0.011477638967335224, + 0.010248357430100441, + 0.013849189504981041, + 0.019313903525471687, + 0.018562309443950653, + 0.015340556390583515, + 0.010275191627442837, + 0.013099558651447296, + 0.010162118822336197, + 0.005291758570820093, + 0.006891104392707348, + 0.005628152284771204, + 0.006859219633042812, + 0.00659013818949461, + 0.0042409407906234264, + 0.0073792715556919575, + 0.0073994859121739864, + 0.0035844561643898487, + 0.006791024934500456, + 0.004397530108690262, + 0.009547148831188679, + 0.004727828316390514, + 0.003950993064790964, + 0.0069346316158771515, + 0.0036526955664157867, + 0.0335933081805706, + 0.01490025781095028, + 0.019937681034207344, + 0.017954809591174126, + 0.0186904463917017, + 0.11749003827571869, + 0.02166460081934929, + 0.09257011115550995, + 0.003604779252782464, + 0.00381705304607749, + 0.00562542537227273, + 0.006539956666529179, + 0.005640371236950159, + 0.004206554964184761, + 0.0047906204126775265, + 0.003960595466196537, + 0.007209686562418938, + 0.01158999465405941, + 0.011313316412270069, + 0.014650201424956322, + 0.014202520251274109, + 0.014276399277150631, + 0.016307847574353218, + 0.012398702092468739, + 0.012800765223801136, + 0.012503267265856266, + 0.03222879394888878, + 0.012807345017790794, + 0.01390568632632494, + 0.01416503544896841, + 0.007910069078207016, + 0.013190421275794506, + 0.00271641300059855, + 0.004098089877516031, + 0.004211480263620615, + 0.004098902456462383, + 0.004527333192527294, + 0.004624291323125362, + 0.0027500016149133444, + 0.003419922199100256, + 0.007507842965424061, + 0.01303139142692089, + 0.011843042448163033, + 0.013527048751711845, + 0.01738703064620495, + 0.010556694120168686, + 0.01303319726139307, + 0.00859348475933075 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 7, + 1 + ], + "selected_reward": -0.0019418209558352828 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the video begin with a wide angle shot that slowly zooms out?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are thousands of translucent ice-blue crystal fragments visible converging from all directions at high speed?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the crystal fragments trail brilliant blue light as they move?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the fragments slow down, flash with energy, and unite to form a complete crystal globe with visible internal continental structures?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the final scene show the crystal globe floating and rotating against a dark background with faint starlight?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.011072855442762375 + ], + "opposite_logprob_tokens": [ + -4.511072635650635 + ], + "logprob_reward": -0.011072855442762375, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.011072855442762375, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11363840848207474, + 0.03634224086999893, + 0.035764455795288086, + 0.026474084705114365, + 0.10932157933712006, + 0.23792974650859833, + 0.07313332706689835, + 0.11416420340538025, + 0.03276018053293228, + 0.22047176957130432 + ], + "credit_map": [ + 0.010095501318573952, + 0.009228224866092205, + 0.019902333617210388, + 0.011442379094660282, + 0.009937231428921223, + 0.02584143541753292, + 0.010527550242841244, + 0.016663750633597374, + 0.00407150574028492, + 0.00568561814725399, + 0.004688900429755449, + 0.005883836653083563, + 0.004554532468318939, + 0.0030763742979615927, + 0.004524421412497759, + 0.0038570519536733627, + 0.003566871164366603, + 0.004864099435508251, + 0.004527594894170761, + 0.0046461643651127815, + 0.004433169029653072, + 0.004195557441562414, + 0.00612171646207571, + 0.003409281373023987, + 0.0016552655724808574, + 0.0034857457503676414, + 0.0024807420559227467, + 0.0057202791795134544, + 0.0030603548511862755, + 0.001493651419878006, + 0.006701383739709854, + 0.0018766616703942418, + 0.008441520854830742, + 0.004000777378678322, + 0.00419700937345624, + 0.006419127807021141, + 0.006488486658781767, + 0.03905851021409035, + 0.006186359561979771, + 0.03452979773283005, + 0.020061952993273735, + 0.028558634221553802, + 0.021555330604314804, + 0.029560504481196404, + 0.047332555055618286, + 0.02711997553706169, + 0.046165741980075836, + 0.01757504791021347, + 0.006804266478866339, + 0.007133188657462597, + 0.012716335244476795, + 0.009404070675373077, + 0.007737721782177687, + 0.014216906391084194, + 0.0084617855027318, + 0.006659049075096846, + 0.012378763407468796, + 0.011208868585526943, + 0.03188585117459297, + 0.012084678746759892, + 0.011559683829545975, + 0.01416925247758627, + 0.0077057103626430035, + 0.013171392492949963, + 0.0034402695018798113, + 0.0036382011603564024, + 0.004321018699556589, + 0.004271054640412331, + 0.004492003470659256, + 0.004885468631982803, + 0.0033197165466845036, + 0.004392447881400585, + 0.02092912793159485, + 0.0156894251704216, + 0.03191437944769859, + 0.026531895622611046, + 0.021995259448885918, + 0.03936733305454254, + 0.02173132821917534, + 0.04231302812695503 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 9, + 7 + ], + "selected_reward": -0.011072855442762375 + } + ], + "vlm_decoded_outputs": [ + "No", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "shot that begins with a wide angle and slowly zooms out, immersing the viewer in a deep, nearly pitch-black expanse of space, a backdrop dotted with distant, faint starlight. In the center of the frame, thousands of translucent ice-blue crystal fragments of varying sizes and shapes converge at high speed from all directions, as if drawn by an invisible gravitational force. Each fragment's surface is etched with the intricate contours of continents or oceans. They trail brilliant blue light trails as they dart, swirl, and tumble through the darkness, forming a dynamic, chaotic web of light. Then, as if on unified command, the fragments swiftly and decisively slow down and find their predetermined positions. With each perfectly precise rendezvous, a brief, dazzling flash of energy erupts, ultimately, within seconds, uniting into a complete, precise crystal globe radiating a soft, inner glow. As the camera continues to zoom out, this crystal clear blue planet quietly floats and slowly rotates against the dark background, and its internal continental plate structure is clearly visible. The entire picture, set against the backdrop of cold colors, creates a grand, magical and futuristic technological atmosphere of creation.", + "seed": 10640, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000050/group_7/index_5162_seed_10640_dp_15_rank_30_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 7, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.17285488927736878, + "vlm_logprob": -0.17285488927736878, + "vlm_yes_no_margin": 4.325, + "vlm_token_credit": -0.17285488927736878, + "vlm_yes_ratio": 0.8, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the video begin with a wide angle shot that slowly zooms out?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are thousands of translucent ice-blue crystal fragments visible converging from all directions at high speed?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the crystal fragments trail brilliant blue light as they move?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the fragments slow down, flash with energy, and unite to form a complete crystal globe with visible internal continental structures?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the final scene show the crystal globe floating and rotating against a dark background with faint starlight?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -0.8386370539665222 + ], + "opposite_logprob_tokens": [ + -0.5886370539665222 + ], + "logprob_reward": -0.8386370539665222, + "yes_no_margin_reward": -0.25, + "token_credit_reward": -0.8386370539665222, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.4290878176689148, + 0.019322089850902557, + 0.05199752002954483, + 0.0069425166584551334, + 0.022772330790758133, + 0.021904801949858665, + 0.08101250976324081, + 0.3436070382595062, + 0.01392744854092598, + 0.00942600890994072 + ], + "credit_map": [ + 0.023845620453357697, + 0.03411509469151497, + 0.018348626792430878, + 0.034687940031290054, + 0.17425799369812012, + 0.027103370055556297, + 0.026256414130330086, + 0.09047269076108932, + 0.0037680736277252436, + 0.0023231320083141327, + 0.0020260317251086235, + 0.002664815867319703, + 0.002154584741219878, + 0.002824150724336505, + 0.0022147216368466616, + 0.001346577424556017, + 0.005755979102104902, + 0.0037653837352991104, + 0.016274815425276756, + 0.005163051187992096, + 0.0032678968273103237, + 0.007589058950543404, + 0.005641730036586523, + 0.004539597313851118, + 0.0007663735887035728, + 0.0005827963468618691, + 0.0012945079943165183, + 0.001016218331642449, + 0.0005110689671710134, + 0.0009106328361667693, + 0.001113792648538947, + 0.0007471253629773855, + 0.0045180125162005424, + 0.00236120424233377, + 0.001953193685039878, + 0.0023229459766298532, + 0.00242761243134737, + 0.002316199243068695, + 0.0035617148969322443, + 0.003311445936560631, + 0.003009037347510457, + 0.001483932719565928, + 0.006479829084128141, + 0.0022032796405255795, + 0.0013340977020561695, + 0.0021806175354868174, + 0.002043948508799076, + 0.0031700562685728073, + 0.02167273312807083, + 0.005015047732740641, + 0.00684442138299346, + 0.004834660794585943, + 0.004403173457831144, + 0.02463620714843273, + 0.004155310802161694, + 0.009450949728488922, + 0.03086760640144348, + 0.007466585841029882, + 0.11194856464862823, + 0.012021269649267197, + 0.013774539344012737, + 0.13999046385288239, + 0.010025549679994583, + 0.017512425780296326, + 0.0028848829679191113, + 0.0015796538209542632, + 0.0014933164929971099, + 0.0010674023069441319, + 0.00119418790563941, + 0.002799812937155366, + 0.0013193401973694563, + 0.0015888507477939129, + 0.0010834598215296865, + 0.0006744757993146777, + 0.0010431029368191957, + 0.0007436333689838648, + 0.0006347679300233722, + 0.0018819328397512436, + 0.0007434898870997131, + 0.0026211452204734087 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 7, + 6 + ], + "selected_reward": -0.8386370539665222 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the video begin with a wide angle shot that slowly zooms out?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are thousands of translucent ice-blue crystal fragments visible converging from all directions at high speed?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the crystal fragments trail brilliant blue light as they move?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the fragments slow down, flash with energy, and unite to form a complete crystal globe with visible internal continental structures?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the final scene show the crystal globe floating and rotating against a dark background with faint starlight?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.00598602369427681 + ], + "opposite_logprob_tokens": [ + -5.130986213684082 + ], + "logprob_reward": -0.00598602369427681, + "yes_no_margin_reward": 5.125, + "token_credit_reward": -0.00598602369427681, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09082678705453873, + 0.044904716312885284, + 0.04107978940010071, + 0.014339514076709747, + 0.13938607275485992, + 0.036728329956531525, + 0.1522560864686966, + 0.4540793299674988, + 0.014241804368793964, + 0.012157629244029522 + ], + "credit_map": [ + 0.0059485044330358505, + 0.007648359052836895, + 0.006062276661396027, + 0.006734379567205906, + 0.024321874603629112, + 0.008591294288635254, + 0.008286205120384693, + 0.023233897984027863, + 0.009514493867754936, + 0.006403609178960323, + 0.0047440859489142895, + 0.005195134785026312, + 0.0036726638209074736, + 0.007391007617115974, + 0.005147325806319714, + 0.002836393192410469, + 0.003921075724065304, + 0.0045436774380505085, + 0.0033618242014199495, + 0.006346452981233597, + 0.0045559904538095, + 0.00447903061285615, + 0.008856485597789288, + 0.005015253089368343, + 0.001531305257230997, + 0.0015182328643277287, + 0.002035042503848672, + 0.002438911469653249, + 0.001571843633428216, + 0.0016044870717450976, + 0.002103545470163226, + 0.001536146504804492, + 0.033614736050367355, + 0.013563793152570724, + 0.010684676468372345, + 0.015620625577867031, + 0.010911580175161362, + 0.014441086910665035, + 0.025647353380918503, + 0.014902225695550442, + 0.003990269266068935, + 0.003847341751679778, + 0.005873069632798433, + 0.005307470448315144, + 0.003288162872195244, + 0.0035313747357577085, + 0.005364300217479467, + 0.005526341963559389, + 0.05472506955265999, + 0.007109476253390312, + 0.009041811339557171, + 0.005122743546962738, + 0.0061661056242883205, + 0.05668851360678673, + 0.005004135891795158, + 0.008398234844207764, + 0.033689722418785095, + 0.009718193672597408, + 0.15978197753429413, + 0.01700150966644287, + 0.017826814204454422, + 0.1802135854959488, + 0.013556641526520252, + 0.02229085937142372, + 0.0030442443676292896, + 0.0011624129256233573, + 0.0014608893543481827, + 0.001170382834970951, + 0.0009841760620474815, + 0.0032726519275456667, + 0.0012708880240097642, + 0.0018761585233733058, + 0.003345056204125285, + 0.0011667122598737478, + 0.0013046085368841887, + 0.0007873542490415275, + 0.0009048677165992558, + 0.002121982630342245, + 0.0009207630646415055, + 0.0016062839422374964 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 6, + 4 + ], + "selected_reward": -0.00598602369427681 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the video begin with a wide angle shot that slowly zooms out?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are thousands of translucent ice-blue crystal fragments visible converging from all directions at high speed?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the crystal fragments trail brilliant blue light as they move?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the fragments slow down, flash with energy, and unite to form a complete crystal globe with visible internal continental structures?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the final scene show the crystal globe floating and rotating against a dark background with faint starlight?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0046440837904810905 + ], + "opposite_logprob_tokens": [ + -5.37964391708374 + ], + "logprob_reward": -0.0046440837904810905, + "yes_no_margin_reward": 5.375, + "token_credit_reward": -0.0046440837904810905, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14215916395187378, + 0.019948089495301247, + 0.03440040349960327, + 0.020738862454891205, + 0.046554312109947205, + 0.038374729454517365, + 0.14994323253631592, + 0.5289626717567444, + 0.009842750616371632, + 0.009075733833014965 + ], + "credit_map": [ + 0.015709102153778076, + 0.012317406013607979, + 0.006602788809686899, + 0.011571145616471767, + 0.05624104663729668, + 0.011046931147575378, + 0.011040305718779564, + 0.01763044483959675, + 0.0031892128754407167, + 0.0023825219832360744, + 0.0019268587930127978, + 0.0022253706119954586, + 0.002069924958050251, + 0.0036620341707021, + 0.003042347962036729, + 0.0014498201198875904, + 0.003482807893306017, + 0.002916021505370736, + 0.007541747763752937, + 0.003674874547868967, + 0.0025392835959792137, + 0.004162526223808527, + 0.006227577105164528, + 0.003855568589642644, + 0.002191390609368682, + 0.0014631249941885471, + 0.0020664578769356012, + 0.0029522720724344254, + 0.0021099767182022333, + 0.004083266481757164, + 0.0031836330890655518, + 0.002688742708414793, + 0.006079958286136389, + 0.005579203832894564, + 0.003411837387830019, + 0.0048847077414393425, + 0.006971381139010191, + 0.0038521920796483755, + 0.009529631584882736, + 0.006245406344532967, + 0.004502241499722004, + 0.004226978402584791, + 0.006632993929088116, + 0.005627931095659733, + 0.003773841541260481, + 0.004480275325477123, + 0.00490545853972435, + 0.004225016571581364, + 0.03484058007597923, + 0.011911795474588871, + 0.01479270402342081, + 0.012834839522838593, + 0.009778600186109543, + 0.030860641971230507, + 0.011826975271105766, + 0.023097112774848938, + 0.11042191088199615, + 0.014913389459252357, + 0.09915172308683395, + 0.012610022909939289, + 0.0217965729534626, + 0.2219104915857315, + 0.014867215417325497, + 0.03329133987426758, + 0.0011940483236685395, + 0.0013844167115166783, + 0.0009169481927528977, + 0.0012842272408306599, + 0.0008862284594215453, + 0.0014901801478117704, + 0.000845454924274236, + 0.001841247663833201, + 0.00222618505358696, + 0.0008354869787581265, + 0.0011981602292507887, + 0.0006471864762715995, + 0.0005287750973366201, + 0.0007747514173388481, + 0.0008793636807240546, + 0.001985826762393117 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 6, + 0 + ], + "selected_reward": -0.0046440837904810905 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the video begin with a wide angle shot that slowly zooms out?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are thousands of translucent ice-blue crystal fragments visible converging from all directions at high speed?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the crystal fragments trail brilliant blue light as they move?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the fragments slow down, flash with energy, and unite to form a complete crystal globe with visible internal continental structures?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the final scene show the crystal globe floating and rotating against a dark background with faint starlight?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0008136299438774586 + ], + "opposite_logprob_tokens": [ + -7.1258134841918945 + ], + "logprob_reward": -0.0008136299438774586, + "yes_no_margin_reward": 7.125, + "token_credit_reward": -0.0008136299438774586, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.25564876198768616, + 0.014804716221988201, + 0.027484137564897537, + 0.012003451585769653, + 0.043122246861457825, + 0.2209019660949707, + 0.11939309537410736, + 0.276747465133667, + 0.015149340964853764, + 0.014744888991117477 + ], + "credit_map": [ + 0.012041079811751842, + 0.020137183368206024, + 0.010836878791451454, + 0.020718799903988838, + 0.10427682846784592, + 0.01680171675980091, + 0.013695606961846352, + 0.057140644639730453, + 0.002738221548497677, + 0.0013616293435916305, + 0.0016420678002759814, + 0.0019056674791499972, + 0.0013501623179763556, + 0.0019525596871972084, + 0.0022191156167536974, + 0.0016352906823158264, + 0.0022325259633362293, + 0.0026919099036604166, + 0.004770772065967321, + 0.003694531973451376, + 0.002284563845023513, + 0.003918493166565895, + 0.004522374831140041, + 0.0033689639531075954, + 0.0009567200322635472, + 0.0012407719623297453, + 0.001899220165796578, + 0.002041738713160157, + 0.0007810652023181319, + 0.0012144872453063726, + 0.002162463264539838, + 0.0017069841269403696, + 0.009875047020614147, + 0.004491840954869986, + 0.0034058985766023397, + 0.004912279546260834, + 0.003463927423581481, + 0.004561585374176502, + 0.0066746738739311695, + 0.0057369861751794815, + 0.02597726881504059, + 0.02156667597591877, + 0.04120238125324249, + 0.03230432793498039, + 0.021115414798259735, + 0.026373807340860367, + 0.03011227212846279, + 0.022249793633818626, + 0.020287219434976578, + 0.011275128461420536, + 0.016003647819161415, + 0.012872143648564816, + 0.008811075240373611, + 0.014809532091021538, + 0.010776307433843613, + 0.02455803006887436, + 0.09013067930936813, + 0.010409194976091385, + 0.018758058547973633, + 0.005240349099040031, + 0.011249596253037453, + 0.11104100197553635, + 0.00834254827350378, + 0.021575987339019775, + 0.0031432341784238815, + 0.0018726183334365487, + 0.0017050589667633176, + 0.0014837380731478333, + 0.0012859090929850936, + 0.00165915722027421, + 0.0013193089980632067, + 0.0026803140062838793, + 0.0027288903947919607, + 0.0016779890283942223, + 0.002094271592795849, + 0.0015057175187394023, + 0.0013335347175598145, + 0.0016460207989439368, + 0.0012388128088787198, + 0.0025196517817676067 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 0, + 5 + ], + "selected_reward": -0.0008136299438774586 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the video begin with a wide angle shot that slowly zooms out?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are thousands of translucent ice-blue crystal fragments visible converging from all directions at high speed?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the crystal fragments trail brilliant blue light as they move?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the fragments slow down, flash with energy, and unite to form a complete crystal globe with visible internal continental structures?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the final scene show the crystal globe floating and rotating against a dark background with faint starlight?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.014193654991686344 + ], + "opposite_logprob_tokens": [ + -4.264193534851074 + ], + "logprob_reward": -0.014193654991686344, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.014193654991686344, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.03109641931951046, + 0.014383838512003422, + 0.02298613265156746, + 0.011565104126930237, + 0.04755234345793724, + 0.03171898424625397, + 0.37640026211738586, + 0.4342919588088989, + 0.014776975847780704, + 0.01522795669734478 + ], + "credit_map": [ + 0.003349197329953313, + 0.0030977113638073206, + 0.002003683941438794, + 0.003076803172007203, + 0.009650010615587234, + 0.003415413899347186, + 0.002911758841946721, + 0.003591845277696848, + 0.0024306385312229395, + 0.001519039156846702, + 0.0017976885428652167, + 0.0016164148692041636, + 0.0014771424466744065, + 0.0021161686163395643, + 0.0021780396346002817, + 0.0012487079948186874, + 0.001849949941970408, + 0.0025051545817404985, + 0.0017281862674281001, + 0.0033174597192555666, + 0.0024948876816779375, + 0.0029768499080091715, + 0.0026392508298158646, + 0.005474394652992487, + 0.0014220352750271559, + 0.0009122947230935097, + 0.001800878206267953, + 0.0012530826497823, + 0.0009512454853393137, + 0.0024055400863289833, + 0.0012477664276957512, + 0.0015722626121714711, + 0.010917057283222675, + 0.0044265855103731155, + 0.003409840865060687, + 0.005370399449020624, + 0.004318538121879101, + 0.004656684119254351, + 0.009735304862260818, + 0.004717936273664236, + 0.004758105147629976, + 0.0028559155762195587, + 0.006105773150920868, + 0.0033295240718871355, + 0.0021559486631304026, + 0.003108362900093198, + 0.003612771863117814, + 0.005792587995529175, + 0.047069571912288666, + 0.03786929324269295, + 0.05251890420913696, + 0.04347265511751175, + 0.03246711939573288, + 0.03913778066635132, + 0.03937016427516937, + 0.08449482172727585, + 0.03800124302506447, + 0.009514213539659977, + 0.14251935482025146, + 0.01523546501994133, + 0.017505403608083725, + 0.17677147686481476, + 0.012886757031083107, + 0.02185807377099991, + 0.0030205948278307915, + 0.0017319420585408807, + 0.0014129177434369922, + 0.0013987981947138906, + 0.001248844782821834, + 0.00307063409127295, + 0.0011788207339122891, + 0.0017144243465736508, + 0.004300452768802643, + 0.0014928117161616683, + 0.0023432893212884665, + 0.0012645823881030083, + 0.0010414166608825326, + 0.001846996950916946, + 0.0010308480123057961, + 0.001907559228129685 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 6, + 4 + ], + "selected_reward": -0.014193654991686344 + } + ], + "vlm_decoded_outputs": [ + "No", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "shot that begins with a wide angle and slowly zooms out, immersing the viewer in a deep, nearly pitch-black expanse of space, a backdrop dotted with distant, faint starlight. In the center of the frame, thousands of translucent ice-blue crystal fragments of varying sizes and shapes converge at high speed from all directions, as if drawn by an invisible gravitational force. Each fragment's surface is etched with the intricate contours of continents or oceans. They trail brilliant blue light trails as they dart, swirl, and tumble through the darkness, forming a dynamic, chaotic web of light. Then, as if on unified command, the fragments swiftly and decisively slow down and find their predetermined positions. With each perfectly precise rendezvous, a brief, dazzling flash of energy erupts, ultimately, within seconds, uniting into a complete, precise crystal globe radiating a soft, inner glow. As the camera continues to zoom out, this crystal clear blue planet quietly floats and slowly rotates against the dark background, and its internal continental plate structure is clearly visible. The entire picture, set against the backdrop of cold colors, creates a grand, magical and futuristic technological atmosphere of creation.", + "seed": 10641, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000050/group_7/index_5162_seed_10641_dp_15_rank_30_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 7, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.13162488715024664, + "vlm_logprob": -0.13162488715024664, + "vlm_yes_no_margin": 4.75, + "vlm_token_credit": -0.13162488715024664, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the video begin with a wide angle shot that slowly zooms out?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are thousands of translucent ice-blue crystal fragments visible converging from all directions at high speed?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the crystal fragments trail brilliant blue light as they move?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the fragments slow down, flash with energy, and unite to form a complete crystal globe with visible internal continental structures?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the final scene show the crystal globe floating and rotating against a dark background with faint starlight?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.6429790258407593 + ], + "opposite_logprob_tokens": [ + -0.7679790258407593 + ], + "logprob_reward": -0.6429790258407593, + "yes_no_margin_reward": 0.125, + "token_credit_reward": -0.6429790258407593, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.17072080075740814, + 0.04259062558412552, + 0.20208629965782166, + 0.0379735492169857, + 0.011426773853600025, + 0.0815470963716507, + 0.08111913502216339, + 0.07613816112279892, + 0.2125418484210968, + 0.08385563641786575 + ], + "credit_map": [ + 0.01099641714245081, + 0.026388194411993027, + 0.011876783333718777, + 0.0332917682826519, + 0.02915780618786812, + 0.013751355931162834, + 0.02454913593828678, + 0.020709363743662834, + 0.005307675339281559, + 0.006624242290854454, + 0.004725418984889984, + 0.007013482041656971, + 0.00670737074688077, + 0.002314903074875474, + 0.006447630934417248, + 0.003449908923357725, + 0.03894113749265671, + 0.007092602085322142, + 0.018157051876187325, + 0.012578588910400867, + 0.01599987968802452, + 0.00867809820920229, + 0.06998317688703537, + 0.03065577521920204, + 0.0040629212744534016, + 0.0022457365412265062, + 0.004095796961337328, + 0.0026963113341480494, + 0.003076285822317004, + 0.00433893408626318, + 0.007346897851675749, + 0.010110671631991863, + 0.0013624051352962852, + 0.000948223692830652, + 0.0020146945025771856, + 0.0016348677454516292, + 0.0014128184411674738, + 0.0012590381084010005, + 0.0014324057847261429, + 0.001362321199849248, + 0.017791930586099625, + 0.0034500835463404655, + 0.026057420298457146, + 0.0057114893570542336, + 0.004162374883890152, + 0.009206674061715603, + 0.00505305640399456, + 0.010114079341292381, + 0.02019767463207245, + 0.004305832087993622, + 0.0184909850358963, + 0.005337981041520834, + 0.003613583277910948, + 0.015341040678322315, + 0.004401498939841986, + 0.00943054724484682, + 0.01218072697520256, + 0.004454412963241339, + 0.008094350807368755, + 0.0060657113790512085, + 0.005235452204942703, + 0.014154416508972645, + 0.0069976383820176125, + 0.018955454230308533, + 0.04297735542058945, + 0.007338942494243383, + 0.01723688468337059, + 0.007309900131076574, + 0.008901872672140598, + 0.08654262125492096, + 0.007419174537062645, + 0.034815121442079544, + 0.015057758428156376, + 0.004992373753339052, + 0.008754552341997623, + 0.006214678753167391, + 0.006947830785065889, + 0.025480935350060463, + 0.006213663145899773, + 0.01019385363906622 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 2, + 0 + ], + "selected_reward": -0.6429790258407593 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the video begin with a wide angle shot that slowly zooms out?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are thousands of translucent ice-blue crystal fragments visible converging from all directions at high speed?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the crystal fragments trail brilliant blue light as they move?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the fragments slow down, flash with energy, and unite to form a complete crystal globe with visible internal continental structures?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the final scene show the crystal globe floating and rotating against a dark background with faint starlight?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.008662977255880833 + ], + "opposite_logprob_tokens": [ + -4.758663177490234 + ], + "logprob_reward": -0.008662977255880833, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.008662977255880833, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.32780134677886963, + 0.07403655350208282, + 0.1521281599998474, + 0.04496907442808151, + 0.05639120936393738, + 0.09083530306816101, + 0.07154639065265656, + 0.03170856088399887, + 0.08892378211021423, + 0.06165960431098938 + ], + "credit_map": [ + 0.01773133873939514, + 0.032129205763339996, + 0.02652856893837452, + 0.09953609853982925, + 0.02422168292105198, + 0.02919095940887928, + 0.06383492797613144, + 0.034628573805093765, + 0.007971331477165222, + 0.00711654219776392, + 0.008130822330713272, + 0.01351611502468586, + 0.007208602502942085, + 0.0040986970998346806, + 0.018544211983680725, + 0.007450232747942209, + 0.017323801293969154, + 0.00987745076417923, + 0.02423401176929474, + 0.010641447268426418, + 0.009823113679885864, + 0.011685588397085667, + 0.028997119516134262, + 0.039545632898807526, + 0.006721767131239176, + 0.004774080589413643, + 0.006486278492957354, + 0.005004785023629665, + 0.004956225864589214, + 0.00475517800077796, + 0.005419397726655006, + 0.0068513620644807816, + 0.011777758598327637, + 0.0026028261054307222, + 0.00655576353892684, + 0.005479506682604551, + 0.006376101169735193, + 0.00318742822855711, + 0.015708450227975845, + 0.0047033741138875484, + 0.024108923971652985, + 0.0035037584602832794, + 0.03151616454124451, + 0.005189495626837015, + 0.004248992074280977, + 0.006982096936553717, + 0.005164689384400845, + 0.010121174156665802, + 0.017031053081154823, + 0.0040404885075986385, + 0.01576138101518154, + 0.004969945643097162, + 0.0030885692685842514, + 0.014120257459580898, + 0.003973233513534069, + 0.008561459369957447, + 0.0035607938189059496, + 0.002485572826117277, + 0.0030374687630683184, + 0.0027319358196109533, + 0.003273446811363101, + 0.008009965531527996, + 0.003388195065781474, + 0.00522118154913187, + 0.01599620096385479, + 0.003541481913998723, + 0.006768472027033567, + 0.0036873382050544024, + 0.00435743760317564, + 0.035229261964559555, + 0.0031948236282914877, + 0.016148768365383148, + 0.011917226016521454, + 0.003922952804714441, + 0.006871338002383709, + 0.004880663473159075, + 0.005583504214882851, + 0.016127649694681168, + 0.005246677435934544, + 0.007109588012099266 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 5 + ], + "selected_reward": -0.008662977255880833 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the video begin with a wide angle shot that slowly zooms out?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are thousands of translucent ice-blue crystal fragments visible converging from all directions at high speed?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the crystal fragments trail brilliant blue light as they move?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the fragments slow down, flash with energy, and unite to form a complete crystal globe with visible internal continental structures?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the final scene show the crystal globe floating and rotating against a dark background with faint starlight?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0036207374650985003 + ], + "opposite_logprob_tokens": [ + -5.628620624542236 + ], + "logprob_reward": -0.0036207374650985003, + "yes_no_margin_reward": 5.625, + "token_credit_reward": -0.0036207374650985003, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.23534683883190155, + 0.0648970976471901, + 0.08551400899887085, + 0.11866731196641922, + 0.054250143468379974, + 0.10582466423511505, + 0.08351251482963562, + 0.03268837183713913, + 0.17732715606689453, + 0.041971828788518906 + ], + "credit_map": [ + 0.012059123255312443, + 0.01999165304005146, + 0.014433315955102444, + 0.07550570368766785, + 0.02092560939490795, + 0.018580680713057518, + 0.044512342661619186, + 0.029338413849473, + 0.008412123657763004, + 0.006324842572212219, + 0.005753840319812298, + 0.010542180389165878, + 0.007618146017193794, + 0.004139530472457409, + 0.014669806696474552, + 0.00743663078173995, + 0.014976972714066505, + 0.005602212622761726, + 0.008573203347623348, + 0.007106224074959755, + 0.007618315517902374, + 0.005385285709053278, + 0.025196276605129242, + 0.011055517010390759, + 0.007473852951079607, + 0.003651014529168606, + 0.005815418437123299, + 0.0034863341134041548, + 0.0058338758535683155, + 0.009422910399734974, + 0.039856091141700745, + 0.043127819895744324, + 0.009151357226073742, + 0.003076220164075494, + 0.005800108425319195, + 0.005413347389549017, + 0.006828294601291418, + 0.0039996616542339325, + 0.01600637659430504, + 0.00397478137165308, + 0.027160484343767166, + 0.00449122441932559, + 0.020489903166890144, + 0.00677963811904192, + 0.007773983757942915, + 0.012299797497689724, + 0.011492164805531502, + 0.01533746812492609, + 0.02035818062722683, + 0.004627523943781853, + 0.018627293407917023, + 0.00561560271307826, + 0.0037120035849511623, + 0.016145871952176094, + 0.004665125161409378, + 0.009760911576449871, + 0.00791137758642435, + 0.0017250607488676906, + 0.003296236041933298, + 0.00202380889095366, + 0.001505660591647029, + 0.005291282199323177, + 0.002362519735470414, + 0.008572426624596119, + 0.0318748876452446, + 0.00945882685482502, + 0.012750680558383465, + 0.012296944856643677, + 0.01391480304300785, + 0.06050606444478035, + 0.012399215251207352, + 0.024125730618834496, + 0.00898036453872919, + 0.0018627027748152614, + 0.004657984245568514, + 0.0023418280761688948, + 0.0024760831147432327, + 0.013848377391695976, + 0.0025132468435913324, + 0.005291241221129894 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 3 + ], + "selected_reward": -0.0036207374650985003 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the video begin with a wide angle shot that slowly zooms out?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are thousands of translucent ice-blue crystal fragments visible converging from all directions at high speed?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the crystal fragments trail brilliant blue light as they move?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the fragments slow down, flash with energy, and unite to form a complete crystal globe with visible internal continental structures?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the final scene show the crystal globe floating and rotating against a dark background with faint starlight?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0009216589969582856 + ], + "opposite_logprob_tokens": [ + -7.000921726226807 + ], + "logprob_reward": -0.0009216589969582856, + "yes_no_margin_reward": 7.0, + "token_credit_reward": -0.0009216589969582856, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10009884089231491, + 0.042068351060152054, + 0.052390072494745255, + 0.3937891125679016, + 0.052792102098464966, + 0.07725545018911362, + 0.11146315932273865, + 0.05929418280720711, + 0.08613929897546768, + 0.024709494784474373 + ], + "credit_map": [ + 0.007630216423422098, + 0.0155557319521904, + 0.009356883354485035, + 0.017262017354369164, + 0.014992968179285526, + 0.009021632373332977, + 0.015422957018017769, + 0.010856431908905506, + 0.004260280169546604, + 0.007095423061400652, + 0.0046899584122002125, + 0.007213561795651913, + 0.005088177975267172, + 0.0027443859726190567, + 0.006685168016701937, + 0.0042913975194096565, + 0.006218448281288147, + 0.002038788516074419, + 0.005401249043643475, + 0.0038813056889921427, + 0.0040202452801167965, + 0.010121368803083897, + 0.01375189982354641, + 0.00695676589384675, + 0.013114486820995808, + 0.006306500174105167, + 0.011441806331276894, + 0.009252429939806461, + 0.02042747288942337, + 0.02274995855987072, + 0.1499626338481903, + 0.16053377091884613, + 0.006421612109988928, + 0.006873495411127806, + 0.007492851931601763, + 0.008703581057488918, + 0.006636872887611389, + 0.005107779987156391, + 0.006693416275084019, + 0.004862489178776741, + 0.02162778005003929, + 0.0030446115415543318, + 0.02793809026479721, + 0.004492687061429024, + 0.0030949367210268974, + 0.005214530974626541, + 0.0037638230714946985, + 0.008078989572823048, + 0.028806542977690697, + 0.005930798593908548, + 0.02686413750052452, + 0.0074937064200639725, + 0.0046425010077655315, + 0.019206946715712547, + 0.006298049818724394, + 0.012220474891364574, + 0.007092047482728958, + 0.0038739291485399008, + 0.005772080272436142, + 0.005293393973261118, + 0.005256934557110071, + 0.012701827101409435, + 0.006543655879795551, + 0.012760319747030735, + 0.018852975219488144, + 0.004427094012498856, + 0.007465815637260675, + 0.004291684366762638, + 0.004501772578805685, + 0.030788706615567207, + 0.0041816337034106255, + 0.011629621498286724, + 0.004059675615280867, + 0.002229983452707529, + 0.00297528482042253, + 0.0031085077207535505, + 0.002673099050298333, + 0.0036297859624028206, + 0.0024209001567214727, + 0.0036122589372098446 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 6, + 0 + ], + "selected_reward": -0.0009216589969582856 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the video begin with a wide angle shot that slowly zooms out?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are thousands of translucent ice-blue crystal fragments visible converging from all directions at high speed?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the crystal fragments trail brilliant blue light as they move?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the fragments slow down, flash with energy, and unite to form a complete crystal globe with visible internal continental structures?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the final scene show the crystal globe floating and rotating against a dark background with faint starlight?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.001940036192536354 + ], + "opposite_logprob_tokens": [ + -6.2519402503967285 + ], + "logprob_reward": -0.001940036192536354, + "yes_no_margin_reward": 6.25, + "token_credit_reward": -0.001940036192536354, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2309131920337677, + 0.021367017179727554, + 0.07571057230234146, + 0.2554439902305603, + 0.03960287570953369, + 0.025584625080227852, + 0.06781002134084702, + 0.05315406993031502, + 0.20939739048480988, + 0.021016232669353485 + ], + "credit_map": [ + 0.01297480147331953, + 0.029124705120921135, + 0.01668330654501915, + 0.06443609297275543, + 0.022514725103974342, + 0.017282195389270782, + 0.046243101358413696, + 0.021654285490512848, + 0.0023118830285966396, + 0.0029430072754621506, + 0.0023427689447999, + 0.003161500906571746, + 0.002865593647584319, + 0.0012526108184829354, + 0.00405497569590807, + 0.002434676745906472, + 0.018311893567442894, + 0.0036703143268823624, + 0.004249503370374441, + 0.004597826395183802, + 0.005130519159138203, + 0.006661593914031982, + 0.02514522150158882, + 0.007943701930344105, + 0.008572466671466827, + 0.0037941441405564547, + 0.0073553347028791904, + 0.006355323828756809, + 0.013202411122620106, + 0.014671072363853455, + 0.09771254658699036, + 0.10378067940473557, + 0.008184798993170261, + 0.002223389223217964, + 0.004111609887331724, + 0.004131030756980181, + 0.005244302563369274, + 0.0025365501642227173, + 0.010710356757044792, + 0.002460838295519352, + 0.003376163775101304, + 0.0017239291919395328, + 0.005570666398853064, + 0.002317165955901146, + 0.0016366435447707772, + 0.0036769534926861525, + 0.0027808528393507004, + 0.0045022498816251755, + 0.013385309837758541, + 0.003760308027267456, + 0.013563485816121101, + 0.005067586898803711, + 0.0032564369030296803, + 0.014963743276894093, + 0.004438159987330437, + 0.009374995715916157, + 0.013299335725605488, + 0.0026904502883553505, + 0.005112185142934322, + 0.0027946161571890116, + 0.0025593778118491173, + 0.009174266830086708, + 0.003265686333179474, + 0.014258148148655891, + 0.044004037976264954, + 0.008860375732183456, + 0.018066411837935448, + 0.008536562323570251, + 0.009820705279707909, + 0.08085305988788605, + 0.008695927448570728, + 0.030560314655303955, + 0.0020394979510456324, + 0.001722341519780457, + 0.002703465986996889, + 0.001774158445186913, + 0.0020633155945688486, + 0.005264868959784508, + 0.0019731817301362753, + 0.0034754010848701 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 0, + 8 + ], + "selected_reward": -0.001940036192536354 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000050/group_8/index_tensor([4133, 4133])_group_8_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000050/group_8/index_tensor([4133, 4133])_group_8_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..5c1b1fae784c7e9801599cb2354622e72709f95f --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000050/group_8/index_tensor([4133, 4133])_group_8_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "An epic aerial drone shot smoothly tracks a majestic golden eagle, its massive wingspan fully extended, effortlessly riding rising thermals as it gracefully circles above a vast, rugged valley. The raptor's plumage is a deep chocolate brown, and the golden late afternoon sun, striking it from a low angle, illuminates the signature golden feathers on the nape of its neck and crown, making it shimmer like a crown in the sky. Below, sheer granite cliffs and patches of deep green coniferous forest create a majestic landscape, while a silvery river winds through the valley floor, reflecting the light of the sky. The eagle glides in a wide, deliberate arc, its head turning with sharp, precise movements, its piercing amber eyes fixed intently downward, methodically scanning every inch of the landscape like a monarch surveying his kingdom. It occasionally adjusts its flying posture by slightly tilting its wings or slightly flapping its tail feathers. Its fleeting shadow sweeps across the rocks and treetops below like a ghost. The whole picture is full of primitive, untamed wild beauty, conveying a lonely and majestic kingly aura.", + "seed": 18131, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000050/group_8/index_4133_seed_18131_dp_16_rank_32_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 8, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.02123673791065812, + "vlm_logprob": -0.02123673791065812, + "vlm_yes_no_margin": 4.45, + "vlm_token_credit": -0.02123673791065812, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a golden eagle with a massive wingspan fully extended?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the eagle's plumage deep chocolate brown with golden feathers on the nape and crown?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the eagle gliding in a wide arc while riding rising thermals above a rugged valley?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the scene include sheer granite cliffs, deep green coniferous forest, and a silvery river below?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot an aerial drone view with the camera tracking the eagle smoothly?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.038955170661211014 + ], + "opposite_logprob_tokens": [ + -3.4139552116394043 + ], + "logprob_reward": -0.038955170661211014, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.038955170661211014, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.16885434091091156, + 0.06954243034124374, + 0.181236132979393, + 0.07914461940526962, + 0.03889651596546173, + 0.0427575446665287, + 0.09772057831287384, + 0.0588802769780159, + 0.15733864903450012, + 0.10562893748283386 + ], + "credit_map": [ + 0.010654116980731487, + 0.01551766972988844, + 0.031925883144140244, + 0.016549112275242805, + 0.012513474561274052, + 0.013636535033583641, + 0.01910250261425972, + 0.0489550419151783, + 0.0048533845692873, + 0.005725519265979528, + 0.012351876124739647, + 0.012241920456290245, + 0.00491121644154191, + 0.00824276264756918, + 0.011046773754060268, + 0.010168964974582195, + 0.012106992304325104, + 0.014218540862202644, + 0.016287410631775856, + 0.0357632115483284, + 0.016915839165449142, + 0.0297007467597723, + 0.02844519540667534, + 0.027798177674412727, + 0.007001511286944151, + 0.01177735161036253, + 0.01573512889444828, + 0.008502391166985035, + 0.008129577152431011, + 0.009152475744485855, + 0.00859637651592493, + 0.01024980004876852, + 0.005591124761849642, + 0.006327138748019934, + 0.0039859856478869915, + 0.004798132460564375, + 0.0064588990062475204, + 0.004928905051201582, + 0.004186292644590139, + 0.002620029728859663, + 0.003100333735346794, + 0.0070566958747804165, + 0.00514125544577837, + 0.007393490988761187, + 0.005460215732455254, + 0.003933621104806662, + 0.004070510622113943, + 0.006601415574550629, + 0.009056651964783669, + 0.009894399903714657, + 0.013819671235978603, + 0.011539619415998459, + 0.01420773658901453, + 0.01208739448338747, + 0.016785012558102608, + 0.010330085642635822, + 0.006057514809072018, + 0.006789360661059618, + 0.008868404664099216, + 0.009373124688863754, + 0.006208154838532209, + 0.007388327270746231, + 0.005619633011519909, + 0.008575751446187496, + 0.014320789836347103, + 0.0250724945217371, + 0.018677834421396255, + 0.02465442754328251, + 0.024343052878975868, + 0.012584245763719082, + 0.0191941000521183, + 0.01849168725311756, + 0.010781810618937016, + 0.008579662069678307, + 0.015075971372425556, + 0.015808841213583946, + 0.012708337977528572, + 0.010921517387032509, + 0.009612726047635078, + 0.02214006334543228 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 0, + 8 + ], + "selected_reward": -0.038955170661211014 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a golden eagle with a massive wingspan fully extended?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the eagle's plumage deep chocolate brown with golden feathers on the nape and crown?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the eagle gliding in a wide arc while riding rising thermals above a rugged valley?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the scene include sheer granite cliffs, deep green coniferous forest, and a silvery river below?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot an aerial drone view with the camera tracking the eagle smoothly?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04309023171663284 + ], + "opposite_logprob_tokens": [ + -3.168090343475342 + ], + "logprob_reward": -0.04309023171663284, + "yes_no_margin_reward": 3.125, + "token_credit_reward": -0.04309023171663284, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.24927225708961487, + 0.06648976355791092, + 0.07350707799196243, + 0.07430519908666611, + 0.05641227215528488, + 0.06416961550712585, + 0.09210819751024246, + 0.1494685858488083, + 0.05600214749574661, + 0.11826486140489578 + ], + "credit_map": [ + 0.024262811988592148, + 0.033241305500268936, + 0.038228318095207214, + 0.039516642689704895, + 0.018726058304309845, + 0.017471827566623688, + 0.0328546017408371, + 0.04497068375349045, + 0.006612938363105059, + 0.007537792902439833, + 0.007928925566375256, + 0.006757573690265417, + 0.006965640466660261, + 0.007792123593389988, + 0.014043528586626053, + 0.008851245976984501, + 0.004849240183830261, + 0.006163925398141146, + 0.006444400176405907, + 0.014952624216675758, + 0.008271678350865841, + 0.009705988690257072, + 0.013060498051345348, + 0.01005872804671526, + 0.006535038817673922, + 0.008441447280347347, + 0.01939443126320839, + 0.007304885890334845, + 0.005763450171798468, + 0.0071391453966498375, + 0.006751528941094875, + 0.012975269928574562, + 0.0065215881913900375, + 0.009064431302249432, + 0.005680320784449577, + 0.0053254603408277035, + 0.008864151313900948, + 0.009699275717139244, + 0.006745647639036179, + 0.004511397797614336, + 0.00427335686981678, + 0.015214395709335804, + 0.007392015308141708, + 0.00907022226601839, + 0.009387616999447346, + 0.006524167489260435, + 0.005216601770371199, + 0.007091240957379341, + 0.0061674565076828, + 0.007541022729128599, + 0.013642129488289356, + 0.011636057868599892, + 0.013981425203382969, + 0.011029298417270184, + 0.017481446266174316, + 0.010629368014633656, + 0.011610524728894234, + 0.011937150731682777, + 0.019042668864130974, + 0.021136578172445297, + 0.01749340631067753, + 0.03233347460627556, + 0.020669952034950256, + 0.01524482574313879, + 0.004615011624991894, + 0.006312637589871883, + 0.006044041831046343, + 0.008192910812795162, + 0.008903318084776402, + 0.006984878797084093, + 0.005967266391962767, + 0.008982083760201931, + 0.014363820664584637, + 0.009280604310333729, + 0.017728550359606743, + 0.019350575283169746, + 0.008663241751492023, + 0.01060702558606863, + 0.013325444422662258, + 0.024945596233010292 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 7, + 9 + ], + "selected_reward": -0.04309023171663284 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a golden eagle with a massive wingspan fully extended?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the eagle's plumage deep chocolate brown with golden feathers on the nape and crown?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the eagle gliding in a wide arc while riding rising thermals above a rugged valley?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the scene include sheer granite cliffs, deep green coniferous forest, and a silvery river below?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot an aerial drone view with the camera tracking the eagle smoothly?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0013383012264966965 + ], + "opposite_logprob_tokens": [ + -6.626338481903076 + ], + "logprob_reward": -0.0013383012264966965, + "yes_no_margin_reward": 6.625, + "token_credit_reward": -0.0013383012264966965, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.16645774245262146, + 0.07662507891654968, + 0.09447743743658066, + 0.09701084345579147, + 0.0618460550904274, + 0.10374318808317184, + 0.1029369905591011, + 0.09060437977313995, + 0.09016486257314682, + 0.1161334440112114 + ], + "credit_map": [ + 0.019131170585751534, + 0.01708928309381008, + 0.02625095471739769, + 0.019401824101805687, + 0.014652266167104244, + 0.01527672354131937, + 0.02662893757224083, + 0.028026599436998367, + 0.007230246439576149, + 0.008401704020798206, + 0.008109962567687035, + 0.006742033641785383, + 0.009693745523691177, + 0.008014369755983353, + 0.016969356685876846, + 0.011463656090199947, + 0.008066333830356598, + 0.006779466290026903, + 0.014233331196010113, + 0.017278537154197693, + 0.009142011404037476, + 0.010707784444093704, + 0.01511042844504118, + 0.013159545138478279, + 0.009368992410600185, + 0.015207184478640556, + 0.01459046732634306, + 0.00850790087133646, + 0.011153710074722767, + 0.013202697038650513, + 0.011332150548696518, + 0.01364773791283369, + 0.007077742833644152, + 0.011534962803125381, + 0.00697796605527401, + 0.009874406270682812, + 0.00794066023081541, + 0.007199038751423359, + 0.006462982390075922, + 0.004778297618031502, + 0.006903134752064943, + 0.030146224424242973, + 0.011290452443063259, + 0.012042572721838951, + 0.017639832571148872, + 0.010683887638151646, + 0.0060499003157019615, + 0.008987188339233398, + 0.009973490610718727, + 0.00905593205243349, + 0.014540934935212135, + 0.010635526850819588, + 0.013087336905300617, + 0.016362417489290237, + 0.014292373321950436, + 0.014988979324698448, + 0.008978570811450481, + 0.012565596960484982, + 0.015529538504779339, + 0.018562456592917442, + 0.007289659697562456, + 0.009237472899258137, + 0.008204832673072815, + 0.010236242786049843, + 0.008869356475770473, + 0.018794776871800423, + 0.00873886700719595, + 0.011512531898915768, + 0.011829622089862823, + 0.007476174738258123, + 0.008761407807469368, + 0.014182125218212605, + 0.02262621372938156, + 0.02445933222770691, + 0.012150970287621021, + 0.016425486654043198, + 0.00871118251234293, + 0.008953316137194633, + 0.007867613807320595, + 0.014939326792955399 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 5 + ], + "selected_reward": -0.0013383012264966965 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a golden eagle with a massive wingspan fully extended?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the eagle's plumage deep chocolate brown with golden feathers on the nape and crown?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the eagle gliding in a wide arc while riding rising thermals above a rugged valley?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the scene include sheer granite cliffs, deep green coniferous forest, and a silvery river below?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot an aerial drone view with the camera tracking the eagle smoothly?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.006736662704497576 + ], + "opposite_logprob_tokens": [ + -5.006736755371094 + ], + "logprob_reward": -0.006736662704497576, + "yes_no_margin_reward": 5.0, + "token_credit_reward": -0.006736662704497576, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09009332209825516, + 0.15771730244159698, + 0.09446068108081818, + 0.1417866349220276, + 0.06463825702667236, + 0.06490042060613632, + 0.041736900806427, + 0.14219246804714203, + 0.026598354801535606, + 0.1758756935596466 + ], + "credit_map": [ + 0.00655935937538743, + 0.011121079325675964, + 0.012868289835751057, + 0.014681996777653694, + 0.009490269236266613, + 0.008926973678171635, + 0.012732435949146748, + 0.013712919317185879, + 0.017356526106595993, + 0.015755439177155495, + 0.02557818777859211, + 0.022617274895310402, + 0.012013322673738003, + 0.019872725009918213, + 0.02238033339381218, + 0.022143498063087463, + 0.009221485815942287, + 0.008251527324318886, + 0.010843666270375252, + 0.021590188145637512, + 0.013739563524723053, + 0.008826559409499168, + 0.009882541373372078, + 0.012105151079595089, + 0.01256747916340828, + 0.020465785637497902, + 0.023391524329781532, + 0.013313107192516327, + 0.0179227776825428, + 0.019542420282959938, + 0.015698730945587158, + 0.018884817138314247, + 0.0067304885014891624, + 0.012404593639075756, + 0.006723958533257246, + 0.010988264344632626, + 0.009237808175384998, + 0.007197756785899401, + 0.007006410975009203, + 0.0043489751406013966, + 0.00451084366068244, + 0.020232511684298515, + 0.007097212132066488, + 0.007352051790803671, + 0.010073776356875896, + 0.0065240212716162205, + 0.0033183966297656298, + 0.00579160638153553, + 0.0034989474806934595, + 0.0049742963165044785, + 0.005621368531137705, + 0.003429556032642722, + 0.0057541080750525, + 0.00712115503847599, + 0.004896164406090975, + 0.00644130352884531, + 0.011193438433110714, + 0.011900331825017929, + 0.02200736664235592, + 0.023609953001141548, + 0.012507977895438671, + 0.02507198043167591, + 0.015709636732935905, + 0.020191781222820282, + 0.003052864456549287, + 0.0025598255451768637, + 0.0028758582193404436, + 0.004622949752956629, + 0.003124829614534974, + 0.00274388724938035, + 0.0036882977001369, + 0.0039298441261053085, + 0.025761032477021217, + 0.029443640261888504, + 0.021191366016864777, + 0.025871777907013893, + 0.011814373545348644, + 0.013346855528652668, + 0.018453145399689674, + 0.02999349683523178 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 1, + 7 + ], + "selected_reward": -0.006736662704497576 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a golden eagle with a massive wingspan fully extended?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the eagle's plumage deep chocolate brown with golden feathers on the nape and crown?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the eagle gliding in a wide arc while riding rising thermals above a rugged valley?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the scene include sheer granite cliffs, deep green coniferous forest, and a silvery river below?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot an aerial drone view with the camera tracking the eagle smoothly?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.016063323244452477 + ], + "opposite_logprob_tokens": [ + -4.141063213348389 + ], + "logprob_reward": -0.016063323244452477, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.016063323244452477, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1849924474954605, + 0.06535379588603973, + 0.08917352557182312, + 0.10277070850133896, + 0.16334520280361176, + 0.086357980966568, + 0.07137341797351837, + 0.07302452623844147, + 0.0930626317858696, + 0.07054581493139267 + ], + "credit_map": [ + 0.022322090342640877, + 0.025115644559264183, + 0.027634020894765854, + 0.025907544419169426, + 0.018347706645727158, + 0.017391188070178032, + 0.025202473625540733, + 0.023071780800819397, + 0.0058745527639985085, + 0.00573725625872612, + 0.011510134674608707, + 0.008900871500372887, + 0.005239973310381174, + 0.00904665607959032, + 0.009594024159014225, + 0.009450330398976803, + 0.007350248750299215, + 0.005775009281933308, + 0.012313020415604115, + 0.01798093691468239, + 0.009417224675416946, + 0.008664390072226524, + 0.01688477210700512, + 0.01078792568296194, + 0.008043186739087105, + 0.010798641480505466, + 0.021096277981996536, + 0.01024544145911932, + 0.014750055968761444, + 0.014057198539376259, + 0.013204410672187805, + 0.010575486347079277, + 0.02192523144185543, + 0.02261928841471672, + 0.017531801015138626, + 0.017825113609433174, + 0.023671967908740044, + 0.024513021111488342, + 0.025205792859196663, + 0.01005297526717186, + 0.005934190005064011, + 0.017354067414999008, + 0.009743019007146358, + 0.010233700275421143, + 0.0168780367821455, + 0.009287436492741108, + 0.0066552492789924145, + 0.010272282175719738, + 0.006359821651130915, + 0.009916245937347412, + 0.010625154711306095, + 0.006206630263477564, + 0.007493189536035061, + 0.01137485820800066, + 0.006939519662410021, + 0.01245799008756876, + 0.006185633596032858, + 0.006283447612076998, + 0.016729069873690605, + 0.016117911785840988, + 0.0050961291417479515, + 0.0070694126188755035, + 0.006322614848613739, + 0.009220310486853123, + 0.00938832014799118, + 0.010465282015502453, + 0.009012199006974697, + 0.013513349927961826, + 0.016048621386289597, + 0.011689603328704834, + 0.01208650041371584, + 0.010858754627406597, + 0.013306000269949436, + 0.0168228168040514, + 0.007963251322507858, + 0.008698781952261925, + 0.004640453495085239, + 0.003825897118076682, + 0.005233734846115112, + 0.010054877027869225 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 3 + ], + "selected_reward": -0.016063323244452477 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "An epic aerial drone shot smoothly tracks a majestic golden eagle, its massive wingspan fully extended, effortlessly riding rising thermals as it gracefully circles above a vast, rugged valley. The raptor's plumage is a deep chocolate brown, and the golden late afternoon sun, striking it from a low angle, illuminates the signature golden feathers on the nape of its neck and crown, making it shimmer like a crown in the sky. Below, sheer granite cliffs and patches of deep green coniferous forest create a majestic landscape, while a silvery river winds through the valley floor, reflecting the light of the sky. The eagle glides in a wide, deliberate arc, its head turning with sharp, precise movements, its piercing amber eyes fixed intently downward, methodically scanning every inch of the landscape like a monarch surveying his kingdom. It occasionally adjusts its flying posture by slightly tilting its wings or slightly flapping its tail feathers. Its fleeting shadow sweeps across the rocks and treetops below like a ghost. The whole picture is full of primitive, untamed wild beauty, conveying a lonely and majestic kingly aura.", + "seed": 18132, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000050/group_8/index_4133_seed_18132_dp_16_rank_32_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 8, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.02563773167785257, + "vlm_logprob": -0.02563773167785257, + "vlm_yes_no_margin": 4.175, + "vlm_token_credit": -0.02563773167785257, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a golden eagle with a massive wingspan fully extended?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the eagle's plumage deep chocolate brown with golden feathers on the nape and crown?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the eagle gliding in a wide arc while riding rising thermals above a rugged valley?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the scene include sheer granite cliffs, deep green coniferous forest, and a silvery river below?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot an aerial drone view with the camera tracking the eagle smoothly?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03511581942439079 + ], + "opposite_logprob_tokens": [ + -3.5351157188415527 + ], + "logprob_reward": -0.03511581942439079, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.03511581942439079, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07872981578111649, + 0.042518191039562225, + 0.28171348571777344, + 0.0682435929775238, + 0.06164167821407318, + 0.04899736866354942, + 0.14283321797847748, + 0.03108390048146248, + 0.1191818118095398, + 0.1250568926334381 + ], + "credit_map": [ + 0.007216466125100851, + 0.0076539223082363605, + 0.01332184299826622, + 0.008105040527880192, + 0.0076449098996818066, + 0.007424346636980772, + 0.01194505114108324, + 0.015418240800499916, + 0.003533388953655958, + 0.004778183996677399, + 0.006607536692172289, + 0.006137605290859938, + 0.0030078294221311808, + 0.004691690672188997, + 0.00747777009382844, + 0.00628418754786253, + 0.024957621470093727, + 0.03597099706530571, + 0.047647345811128616, + 0.026864098384976387, + 0.020590130239725113, + 0.032335065305233, + 0.042583297938108444, + 0.05076492950320244, + 0.006745498161762953, + 0.010288521647453308, + 0.009123940952122211, + 0.0070704505778849125, + 0.004914245568215847, + 0.012112281285226345, + 0.009635275229811668, + 0.008353378623723984, + 0.005871305242180824, + 0.011147292330861092, + 0.010203836485743523, + 0.005632384680211544, + 0.005703469272702932, + 0.0063379374332726, + 0.009925177320837975, + 0.006820277310907841, + 0.0056879883632063866, + 0.010409961454570293, + 0.007768283598124981, + 0.005823784973472357, + 0.0038397146854549646, + 0.0038076213095337152, + 0.0057077305391430855, + 0.0059522828087210655, + 0.020642587915062904, + 0.023942379280924797, + 0.02043088898062706, + 0.014116530306637287, + 0.016105519607663155, + 0.009403349831700325, + 0.018723584711551666, + 0.019468389451503754, + 0.004029216710478067, + 0.005174523219466209, + 0.0032861835788935423, + 0.004435109905898571, + 0.005063236691057682, + 0.0028201707173138857, + 0.003005324862897396, + 0.0032701343297958374, + 0.01293199509382248, + 0.03096689283847809, + 0.013451115228235722, + 0.01801386848092079, + 0.010475961491465569, + 0.00883535761386156, + 0.012745492160320282, + 0.01176113449037075, + 0.017500730231404305, + 0.012334389612078667, + 0.017072733491659164, + 0.01320337038487196, + 0.009556864388287067, + 0.01289695966988802, + 0.02098267525434494, + 0.021509164944291115 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 6, + 9 + ], + "selected_reward": -0.03511581942439079 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a golden eagle with a massive wingspan fully extended?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the eagle's plumage deep chocolate brown with golden feathers on the nape and crown?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the eagle gliding in a wide arc while riding rising thermals above a rugged valley?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the scene include sheer granite cliffs, deep green coniferous forest, and a silvery river below?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot an aerial drone view with the camera tracking the eagle smoothly?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.05499311164021492 + ], + "opposite_logprob_tokens": [ + -2.929993152618408 + ], + "logprob_reward": -0.05499311164021492, + "yes_no_margin_reward": 2.875, + "token_credit_reward": -0.05499311164021492, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.19874650239944458, + 0.0634656473994255, + 0.14315757155418396, + 0.10353375226259232, + 0.07562059164047241, + 0.08349434286355972, + 0.09744211286306381, + 0.05098907649517059, + 0.0921018198132515, + 0.09144855290651321 + ], + "credit_map": [ + 0.015492059290409088, + 0.020710114389657974, + 0.03183013200759888, + 0.01784536801278591, + 0.021544665098190308, + 0.021896934136748314, + 0.03900061920285225, + 0.0304266344755888, + 0.007146214600652456, + 0.009846807457506657, + 0.009776587598025799, + 0.0051425122655928135, + 0.005415778141468763, + 0.008994708769023418, + 0.008616725914180279, + 0.008526323363184929, + 0.011586214415729046, + 0.023510640487074852, + 0.020780127495527267, + 0.015220697969198227, + 0.00972591433674097, + 0.018878858536481857, + 0.02036488987505436, + 0.023090247064828873, + 0.00901830568909645, + 0.01688271388411522, + 0.01620507426559925, + 0.02284175530076027, + 0.009281948208808899, + 0.010359618812799454, + 0.012413956224918365, + 0.006530384533107281, + 0.006637132726609707, + 0.007755850907415152, + 0.01184253953397274, + 0.00778704509139061, + 0.009225196205079556, + 0.01058423612266779, + 0.014156916178762913, + 0.007631685119122267, + 0.008208466693758965, + 0.015851592645049095, + 0.011759936809539795, + 0.01225741021335125, + 0.007054072804749012, + 0.00695020193234086, + 0.00982549600303173, + 0.011587172746658325, + 0.009435291402041912, + 0.014425446279346943, + 0.008942749351263046, + 0.009816739708185196, + 0.010929575189948082, + 0.006812790408730507, + 0.022215742617845535, + 0.014863784424960613, + 0.005481797270476818, + 0.007807735353708267, + 0.005721434950828552, + 0.004918468650430441, + 0.010423460975289345, + 0.005600564181804657, + 0.0053275455720722675, + 0.005708073265850544, + 0.010134721174836159, + 0.024207882583141327, + 0.01087518036365509, + 0.013916215859353542, + 0.008050085976719856, + 0.007270150817930698, + 0.009355854243040085, + 0.008291739039123058, + 0.013262958265841007, + 0.0055190990678966045, + 0.02153969183564186, + 0.007966474629938602, + 0.007260837126523256, + 0.0143639100715518, + 0.008727644570171833, + 0.012807951308786869 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 3 + ], + "selected_reward": -0.05499311164021492 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a golden eagle with a massive wingspan fully extended?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the eagle's plumage deep chocolate brown with golden feathers on the nape and crown?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the eagle gliding in a wide arc while riding rising thermals above a rugged valley?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the scene include sheer granite cliffs, deep green coniferous forest, and a silvery river below?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot an aerial drone view with the camera tracking the eagle smoothly?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0019468179671093822 + ], + "opposite_logprob_tokens": [ + -6.251946926116943 + ], + "logprob_reward": -0.0019468179671093822, + "yes_no_margin_reward": 6.25, + "token_credit_reward": -0.0019468179671093822, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10262878239154816, + 0.21089190244674683, + 0.18310868740081787, + 0.104856476187706, + 0.08167414367198944, + 0.04949617013335228, + 0.07895001769065857, + 0.05297945439815521, + 0.06755673140287399, + 0.06785760819911957 + ], + "credit_map": [ + 0.010209046304225922, + 0.011247421614825726, + 0.01735730841755867, + 0.010353858582675457, + 0.007233513984829187, + 0.009318727999925613, + 0.017187688499689102, + 0.019721219316124916, + 0.018103880807757378, + 0.028669510036706924, + 0.03162044286727905, + 0.02532636560499668, + 0.01645505800843239, + 0.02329149655997753, + 0.03799060359597206, + 0.029434552416205406, + 0.013800018467009068, + 0.01760319620370865, + 0.02800934948027134, + 0.026729414239525795, + 0.015150681138038635, + 0.013056530617177486, + 0.04782586917281151, + 0.02093362808227539, + 0.008429454639554024, + 0.016348980367183685, + 0.016327163204550743, + 0.023679321631789207, + 0.010067799128592014, + 0.011301367543637753, + 0.012638615444302559, + 0.00606378260999918, + 0.006718371994793415, + 0.012018958106637001, + 0.012905968353152275, + 0.01015413273125887, + 0.006938411854207516, + 0.007135416846722364, + 0.016726311296224594, + 0.009076577611267567, + 0.006718640215694904, + 0.007276584394276142, + 0.009739522822201252, + 0.005640613380819559, + 0.003759517800062895, + 0.00492949690669775, + 0.005887593608349562, + 0.005544199142605066, + 0.009244071319699287, + 0.010156429372727871, + 0.015918392688035965, + 0.009795665740966797, + 0.009353344328701496, + 0.005402524024248123, + 0.009371956810355186, + 0.009707632474601269, + 0.006001242436468601, + 0.006860367488116026, + 0.005452598445117474, + 0.007071840111166239, + 0.009322063066065311, + 0.005622606258839369, + 0.005693390034139156, + 0.006955343298614025, + 0.007745417300611734, + 0.01504493784159422, + 0.007630015257745981, + 0.006971253082156181, + 0.009755125269293785, + 0.005854867864400148, + 0.00829069409519434, + 0.0062644160352647305, + 0.010223865509033203, + 0.00779133103787899, + 0.010096753016114235, + 0.006143809296190739, + 0.00758836604654789, + 0.005706359166651964, + 0.010626982897520065, + 0.00968013796955347 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 2, + 3 + ], + "selected_reward": -0.0019468179671093822 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a golden eagle with a massive wingspan fully extended?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the eagle's plumage deep chocolate brown with golden feathers on the nape and crown?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the eagle gliding in a wide arc while riding rising thermals above a rugged valley?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the scene include sheer granite cliffs, deep green coniferous forest, and a silvery river below?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot an aerial drone view with the camera tracking the eagle smoothly?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.009786365553736687 + ], + "opposite_logprob_tokens": [ + -4.634786128997803 + ], + "logprob_reward": -0.009786365553736687, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.009786365553736687, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.17551125586032867, + 0.165760338306427, + 0.04336894676089287, + 0.16814954578876495, + 0.13371621072292328, + 0.03364710882306099, + 0.0619765929877758, + 0.05196741968393326, + 0.07995647937059402, + 0.08594603091478348 + ], + "credit_map": [ + 0.01814236119389534, + 0.013944542966783047, + 0.03421102091670036, + 0.023992162197828293, + 0.012366300448775291, + 0.019555605947971344, + 0.024577008560299873, + 0.028722241520881653, + 0.018692871555685997, + 0.022624488919973373, + 0.028367625549435616, + 0.0187827181071043, + 0.011520004831254482, + 0.01781652309000492, + 0.025911297649145126, + 0.02204480953514576, + 0.004723466467112303, + 0.004853317979723215, + 0.006354200188070536, + 0.006359182298183441, + 0.003471291856840253, + 0.004858364816755056, + 0.006722392048686743, + 0.006026732735335827, + 0.015059741213917732, + 0.022820619866251945, + 0.023043088614940643, + 0.025251002982258797, + 0.011464166454970837, + 0.029726846143603325, + 0.023808537051081657, + 0.016975553706288338, + 0.012341173365712166, + 0.012932082638144493, + 0.031951431185007095, + 0.013159156776964664, + 0.007986893877387047, + 0.027345234528183937, + 0.01679598167538643, + 0.011204272508621216, + 0.0036062197759747505, + 0.004804131109267473, + 0.005501941777765751, + 0.004731422755867243, + 0.002932014176622033, + 0.003751601791009307, + 0.003735436126589775, + 0.004584342706948519, + 0.008642557077109814, + 0.007856186479330063, + 0.009592358954250813, + 0.006258547306060791, + 0.004967104643583298, + 0.003986728377640247, + 0.008586623705923557, + 0.01208648830652237, + 0.006364168133586645, + 0.005507236812263727, + 0.010910095646977425, + 0.006465148646384478, + 0.007969086989760399, + 0.0049909078516066074, + 0.004093108233064413, + 0.005667671095579863, + 0.007878558710217476, + 0.010739076882600784, + 0.01007638406008482, + 0.014870934188365936, + 0.008343063294887543, + 0.0068566943518817425, + 0.00946294516324997, + 0.011728827841579914, + 0.01132437027990818, + 0.01010450255125761, + 0.011470766738057137, + 0.015039948746562004, + 0.008327419869601727, + 0.005713859107345343, + 0.011528914794325829, + 0.012436251156032085 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 1 + ], + "selected_reward": -0.009786365553736687 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a golden eagle with a massive wingspan fully extended?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the eagle's plumage deep chocolate brown with golden feathers on the nape and crown?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the eagle gliding in a wide arc while riding rising thermals above a rugged valley?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the scene include sheer granite cliffs, deep green coniferous forest, and a silvery river below?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot an aerial drone view with the camera tracking the eagle smoothly?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.026346543803811073 + ], + "opposite_logprob_tokens": [ + -3.651346445083618 + ], + "logprob_reward": -0.026346543803811073, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.026346543803811073, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1050848588347435, + 0.05890671908855438, + 0.22755169868469238, + 0.0915115550160408, + 0.1661505252122879, + 0.07991859316825867, + 0.05732257664203644, + 0.043903954327106476, + 0.04978629946708679, + 0.11986316740512848 + ], + "credit_map": [ + 0.014866963028907776, + 0.008775431662797928, + 0.0198141410946846, + 0.008532275445759296, + 0.0079080481082201, + 0.015393834561109543, + 0.013363536447286606, + 0.016430633142590523, + 0.005214357748627663, + 0.008854863233864307, + 0.0090522151440382, + 0.007698448374867439, + 0.004165856633335352, + 0.006206871941685677, + 0.010274135507643223, + 0.007439977489411831, + 0.01656712405383587, + 0.021724404767155647, + 0.03477683663368225, + 0.03405959531664848, + 0.017902910709381104, + 0.014401703141629696, + 0.06317246705293655, + 0.024946682155132294, + 0.007271748036146164, + 0.014419485814869404, + 0.014600131660699844, + 0.021200377494096756, + 0.008621620945632458, + 0.009239910170435905, + 0.011033105663955212, + 0.005125175230205059, + 0.0169141236692667, + 0.01579386740922928, + 0.0391504280269146, + 0.01492458675056696, + 0.01068829745054245, + 0.03827809542417526, + 0.015515213832259178, + 0.01488592941313982, + 0.009386168792843819, + 0.015121575444936752, + 0.011246313340961933, + 0.010105394758284092, + 0.007040642201900482, + 0.0074743302538990974, + 0.008668893948197365, + 0.010875279083848, + 0.008802101947367191, + 0.007631176616996527, + 0.011958690360188484, + 0.005917152389883995, + 0.006458647083491087, + 0.0041133565828204155, + 0.005864240694791079, + 0.006577217485755682, + 0.005888922140002251, + 0.006035459693521261, + 0.008678821846842766, + 0.004987788386642933, + 0.004049757495522499, + 0.003624445991590619, + 0.004742251709103584, + 0.005896511022001505, + 0.006522217765450478, + 0.009311893954873085, + 0.005459447856992483, + 0.005047616548836231, + 0.004972788970917463, + 0.004031431395560503, + 0.006045411340892315, + 0.008395496755838394, + 0.019520999863743782, + 0.014205633662641048, + 0.013703910633921623, + 0.012320619076490402, + 0.011889352463185787, + 0.010451247915625572, + 0.020170802250504494, + 0.017600610852241516 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 4, + 9 + ], + "selected_reward": -0.026346543803811073 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "An epic aerial drone shot smoothly tracks a majestic golden eagle, its massive wingspan fully extended, effortlessly riding rising thermals as it gracefully circles above a vast, rugged valley. The raptor's plumage is a deep chocolate brown, and the golden late afternoon sun, striking it from a low angle, illuminates the signature golden feathers on the nape of its neck and crown, making it shimmer like a crown in the sky. Below, sheer granite cliffs and patches of deep green coniferous forest create a majestic landscape, while a silvery river winds through the valley floor, reflecting the light of the sky. The eagle glides in a wide, deliberate arc, its head turning with sharp, precise movements, its piercing amber eyes fixed intently downward, methodically scanning every inch of the landscape like a monarch surveying his kingdom. It occasionally adjusts its flying posture by slightly tilting its wings or slightly flapping its tail feathers. Its fleeting shadow sweeps across the rocks and treetops below like a ghost. The whole picture is full of primitive, untamed wild beauty, conveying a lonely and majestic kingly aura.", + "seed": 18133, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000050/group_8/index_4133_seed_18133_dp_17_rank_34_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 8, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.020762251084670426, + "vlm_logprob": -0.020762251084670426, + "vlm_yes_no_margin": 4.45, + "vlm_token_credit": -0.020762251084670426, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a golden eagle with a massive wingspan fully extended?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the eagle's plumage deep chocolate brown with golden feathers on the nape and crown?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the eagle gliding in a wide arc while riding rising thermals above a rugged valley?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the scene include sheer granite cliffs, deep green coniferous forest, and a silvery river below?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot an aerial drone view with the camera tracking the eagle smoothly?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01572917029261589 + ], + "opposite_logprob_tokens": [ + -4.515728950500488 + ], + "logprob_reward": -0.01572917029261589, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.01572917029261589, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08009602129459381, + 0.03984542936086655, + 0.043596301227808, + 0.0831465944647789, + 0.032407552003860474, + 0.06033581867814064, + 0.08955958485603333, + 0.15300102531909943, + 0.3196965456008911, + 0.09831517189741135 + ], + "credit_map": [ + 0.009384896606206894, + 0.008271838538348675, + 0.010169842280447483, + 0.011935328133404255, + 0.010484153404831886, + 0.009522484615445137, + 0.009592811577022076, + 0.010734658688306808, + 0.004521405790001154, + 0.0033071485813707113, + 0.006416653282940388, + 0.0048711663112044334, + 0.004247632343322039, + 0.006507935002446175, + 0.006101937498897314, + 0.0038715454284101725, + 0.004675636999309063, + 0.007092021405696869, + 0.005443560425192118, + 0.005823651794344187, + 0.003736976534128189, + 0.00547134131193161, + 0.005969494581222534, + 0.005383621901273727, + 0.007935585454106331, + 0.006647055968642235, + 0.009439386427402496, + 0.005578741431236267, + 0.009860844351351261, + 0.029814058914780617, + 0.004804839845746756, + 0.009066082537174225, + 0.0033342561218887568, + 0.0025910325348377228, + 0.0066909934394061565, + 0.005133589264005423, + 0.0031500759068876505, + 0.00392889603972435, + 0.004262817092239857, + 0.003315890673547983, + 0.0065302434377372265, + 0.009065928868949413, + 0.010517671704292297, + 0.00789201632142067, + 0.007809196598827839, + 0.005909132305532694, + 0.005264038220047951, + 0.007347587030380964, + 0.009358025155961514, + 0.011378365568816662, + 0.012651409022510052, + 0.011651256121695042, + 0.00915496051311493, + 0.010016926564276218, + 0.007090104278177023, + 0.018258538097143173, + 0.010381752625107765, + 0.010717366822063923, + 0.016772467643022537, + 0.01171757373958826, + 0.021064743399620056, + 0.04389861971139908, + 0.01699327491223812, + 0.021455245092511177, + 0.023943772539496422, + 0.03926239535212517, + 0.02864854782819748, + 0.0263944324105978, + 0.031147053465247154, + 0.03320251777768135, + 0.0548669695854187, + 0.08223085105419159, + 0.00991187896579504, + 0.019658494740724564, + 0.009046711027622223, + 0.008610071614384651, + 0.011724857613444328, + 0.016419388353824615, + 0.00974191166460514, + 0.013201849535107613 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 7, + 9 + ], + "selected_reward": -0.01572917029261589 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a golden eagle with a massive wingspan fully extended?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the eagle's plumage deep chocolate brown with golden feathers on the nape and crown?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the eagle gliding in a wide arc while riding rising thermals above a rugged valley?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the scene include sheer granite cliffs, deep green coniferous forest, and a silvery river below?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot an aerial drone view with the camera tracking the eagle smoothly?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.038117267191410065 + ], + "opposite_logprob_tokens": [ + -3.2881171703338623 + ], + "logprob_reward": -0.038117267191410065, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.038117267191410065, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09719929099082947, + 0.033514365553855896, + 0.0373321957886219, + 0.04815557599067688, + 0.048904430121183395, + 0.026517735794186592, + 0.09322980046272278, + 0.12481316179037094, + 0.272432804107666, + 0.21790063381195068 + ], + "credit_map": [ + 0.006136830896139145, + 0.010065170004963875, + 0.015417252667248249, + 0.022697720676660538, + 0.007855129428207874, + 0.009431007318198681, + 0.0133051797747612, + 0.012290999293327332, + 0.002608825685456395, + 0.0035990318283438683, + 0.0054947915486991405, + 0.004249255172908306, + 0.0028343903832137585, + 0.003234683070331812, + 0.006507512181997299, + 0.004985873121768236, + 0.0030619094613939524, + 0.0042195189744234085, + 0.005594723392277956, + 0.005309205036610365, + 0.004421155899763107, + 0.005016922019422054, + 0.005008291453123093, + 0.004700471647083759, + 0.0037059772294014692, + 0.006734804250299931, + 0.006991842295974493, + 0.005337139591574669, + 0.005152096040546894, + 0.00903861690312624, + 0.006224350072443485, + 0.004970752168446779, + 0.004822449758648872, + 0.004157039802521467, + 0.009741160087287426, + 0.007301281206309795, + 0.0041753738187253475, + 0.006380351725965738, + 0.00630183657631278, + 0.006024940870702267, + 0.003567156847566366, + 0.00399219011887908, + 0.00394379161298275, + 0.0032886893022805452, + 0.0037074631545692682, + 0.0023482723627239466, + 0.0026606987230479717, + 0.0030094750691205263, + 0.012397180311381817, + 0.011197743937373161, + 0.01275404542684555, + 0.014495154842734337, + 0.010136020369827747, + 0.010413680225610733, + 0.007519599515944719, + 0.014316371642053127, + 0.01250328030437231, + 0.008225339464843273, + 0.020777462050318718, + 0.014281067065894604, + 0.016295045614242554, + 0.021410711109638214, + 0.01712026074528694, + 0.014200004749000072, + 0.023794634267687798, + 0.03677134960889816, + 0.02941586822271347, + 0.02336500957608223, + 0.0263662226498127, + 0.032423652708530426, + 0.03900372236967087, + 0.06129235029220581, + 0.024723637849092484, + 0.05548392981290817, + 0.02093946933746338, + 0.012326424941420555, + 0.03141336515545845, + 0.0235544852912426, + 0.015760524198412895, + 0.03369880095124245 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 9, + 7 + ], + "selected_reward": -0.038117267191410065 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a golden eagle with a massive wingspan fully extended?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the eagle's plumage deep chocolate brown with golden feathers on the nape and crown?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the eagle gliding in a wide arc while riding rising thermals above a rugged valley?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the scene include sheer granite cliffs, deep green coniferous forest, and a silvery river below?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot an aerial drone view with the camera tracking the eagle smoothly?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0028196123894304037 + ], + "opposite_logprob_tokens": [ + -5.877819538116455 + ], + "logprob_reward": -0.0028196123894304037, + "yes_no_margin_reward": 5.875, + "token_credit_reward": -0.0028196123894304037, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.18615759909152985, + 0.08511295169591904, + 0.07960359007120132, + 0.1735602766275406, + 0.04409150779247284, + 0.03657861053943634, + 0.1269771158695221, + 0.10850797593593597, + 0.0887862890958786, + 0.07062406837940216 + ], + "credit_map": [ + 0.015656720846891403, + 0.028704725205898285, + 0.025288134813308716, + 0.034957706928253174, + 0.01667630486190319, + 0.014535891823470592, + 0.02372475527226925, + 0.026613350957632065, + 0.008338605985045433, + 0.004364862106740475, + 0.01862991973757744, + 0.007635669317096472, + 0.0080052949488163, + 0.009105910547077656, + 0.016700783744454384, + 0.012331902049481869, + 0.009536389261484146, + 0.00900109764188528, + 0.01664353348314762, + 0.009573211893439293, + 0.005074072629213333, + 0.008932254277169704, + 0.010185685940086842, + 0.010657341219484806, + 0.016451990231871605, + 0.01326140109449625, + 0.022620897740125656, + 0.01356846746057272, + 0.01981968618929386, + 0.05851272866129875, + 0.01138024590909481, + 0.017944859340786934, + 0.005272765178233385, + 0.0031668362207710743, + 0.008547050878405571, + 0.005992793943732977, + 0.0032105676364153624, + 0.006026905961334705, + 0.005670825019478798, + 0.006203762721270323, + 0.004390713293105364, + 0.005664291325956583, + 0.0052568926475942135, + 0.004790657199919224, + 0.004718683660030365, + 0.004020162392407656, + 0.003553330199792981, + 0.004183879587799311, + 0.01667064242064953, + 0.013189570046961308, + 0.015983449295163155, + 0.02241351455450058, + 0.015231523662805557, + 0.014348547905683517, + 0.0124649778008461, + 0.016674894839525223, + 0.007026435807347298, + 0.013199910521507263, + 0.012179521843791008, + 0.010315134190022945, + 0.02141701430082321, + 0.017878860235214233, + 0.013098379597067833, + 0.013392715714871883, + 0.00898281391710043, + 0.009565433487296104, + 0.013103902339935303, + 0.0095082838088274, + 0.008253300562500954, + 0.00612980080768466, + 0.014896932989358902, + 0.018345816060900688, + 0.008121803402900696, + 0.013183741830289364, + 0.011358750984072685, + 0.007798629347234964, + 0.007758011110126972, + 0.006084265653043985, + 0.0072079976089298725, + 0.009110868908464909 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 6 + ], + "selected_reward": -0.0028196123894304037 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a golden eagle with a massive wingspan fully extended?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the eagle's plumage deep chocolate brown with golden feathers on the nape and crown?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the eagle gliding in a wide arc while riding rising thermals above a rugged valley?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the scene include sheer granite cliffs, deep green coniferous forest, and a silvery river below?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot an aerial drone view with the camera tracking the eagle smoothly?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004092651419341564 + ], + "opposite_logprob_tokens": [ + -5.504092693328857 + ], + "logprob_reward": -0.004092651419341564, + "yes_no_margin_reward": 5.5, + "token_credit_reward": -0.004092651419341564, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.16650807857513428, + 0.11346138268709183, + 0.0709555372595787, + 0.06311419606208801, + 0.11931394040584564, + 0.032640039920806885, + 0.13996291160583496, + 0.10616154968738556, + 0.13285911083221436, + 0.05502316728234291 + ], + "credit_map": [ + 0.014033169485628605, + 0.012304779142141342, + 0.028865773230791092, + 0.040171656757593155, + 0.012711089104413986, + 0.014645249582827091, + 0.019462943077087402, + 0.0243134256452322, + 0.01258816383779049, + 0.00835404172539711, + 0.03357403352856636, + 0.013185516931116581, + 0.00822911225259304, + 0.007439560256898403, + 0.019685978069901466, + 0.010404987260699272, + 0.005349333398044109, + 0.007501029875129461, + 0.010033752769231796, + 0.009649470448493958, + 0.0077309864573180676, + 0.010458129458129406, + 0.009806598536670208, + 0.010426240041851997, + 0.00537287350744009, + 0.006624958943575621, + 0.009565283544361591, + 0.007451669778674841, + 0.006643175147473812, + 0.01578638143837452, + 0.005259914789348841, + 0.006409944500774145, + 0.012079410254955292, + 0.009556258097290993, + 0.020188990980386734, + 0.0176505446434021, + 0.011316011659801006, + 0.0140226474031806, + 0.021647879853844643, + 0.012852215208113194, + 0.002939085243269801, + 0.002676404081285, + 0.005069836042821407, + 0.004910461604595184, + 0.0027845478616654873, + 0.004705850034952164, + 0.0033096284605562687, + 0.006244228221476078, + 0.016242166981101036, + 0.011311397887766361, + 0.019644595682621002, + 0.030179183930158615, + 0.010720481164753437, + 0.013670243322849274, + 0.01449776254594326, + 0.023697083815932274, + 0.011389416642487049, + 0.009940559044480324, + 0.018512319773435593, + 0.01366589404642582, + 0.01416693814098835, + 0.012864340096712112, + 0.012137136422097683, + 0.013484948314726353, + 0.008712640963494778, + 0.014690009877085686, + 0.012543623335659504, + 0.017262334004044533, + 0.013298152945935726, + 0.01652272418141365, + 0.026277268305420876, + 0.023552369326353073, + 0.007346940692514181, + 0.005485795438289642, + 0.005248964764177799, + 0.0048071458004415035, + 0.006514618173241615, + 0.010683964937925339, + 0.007432339712977409, + 0.00750339962542057 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 6, + 8 + ], + "selected_reward": -0.004092651419341564 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a golden eagle with a massive wingspan fully extended?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the eagle's plumage deep chocolate brown with golden feathers on the nape and crown?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the eagle gliding in a wide arc while riding rising thermals above a rugged valley?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the scene include sheer granite cliffs, deep green coniferous forest, and a silvery river below?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot an aerial drone view with the camera tracking the eagle smoothly?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0430525541305542 + ], + "opposite_logprob_tokens": [ + -3.1680526733398438 + ], + "logprob_reward": -0.0430525541305542, + "yes_no_margin_reward": 3.125, + "token_credit_reward": -0.0430525541305542, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.17581844329833984, + 0.056793589144945145, + 0.04232863709330559, + 0.07052625715732574, + 0.10020481795072556, + 0.05538131669163704, + 0.050784531980752945, + 0.18684647977352142, + 0.177724689245224, + 0.08359118551015854 + ], + "credit_map": [ + 0.017006253823637962, + 0.02868862822651863, + 0.023572910577058792, + 0.03302106633782387, + 0.017269721254706383, + 0.011678076349198818, + 0.023621855303645134, + 0.02095993235707283, + 0.006244284566491842, + 0.0035371272824704647, + 0.013672860339283943, + 0.006329679861664772, + 0.004381610546261072, + 0.0075636678375303745, + 0.008722979575395584, + 0.0063413772732019424, + 0.0032239265274256468, + 0.0026461370289325714, + 0.010152937844395638, + 0.004667480010539293, + 0.003664734773337841, + 0.010220123454928398, + 0.003647500416263938, + 0.004105797503143549, + 0.006000084336847067, + 0.007364601828157902, + 0.018668873235583305, + 0.010974407196044922, + 0.004215274006128311, + 0.007287958636879921, + 0.007508005481213331, + 0.00850705150514841, + 0.013009225949645042, + 0.009220949374139309, + 0.0223431047052145, + 0.01344746258109808, + 0.007769853807985783, + 0.011299443431198597, + 0.011573958210647106, + 0.011540822684764862, + 0.005400438327342272, + 0.008844357915222645, + 0.0092744380235672, + 0.00711592473089695, + 0.005979804787784815, + 0.006707419641315937, + 0.005442957393825054, + 0.006615976803004742, + 0.005250796675682068, + 0.005338390823453665, + 0.0070059034042060375, + 0.008314908482134342, + 0.004265644121915102, + 0.005663455463945866, + 0.005445621907711029, + 0.0094998087733984, + 0.016198216006159782, + 0.019683845341205597, + 0.032685305923223495, + 0.021008271723985672, + 0.031424183398485184, + 0.020582426339387894, + 0.022147640585899353, + 0.023116592317819595, + 0.014380059204995632, + 0.027568982914090157, + 0.018261313438415527, + 0.017185362055897713, + 0.01853863336145878, + 0.021275125443935394, + 0.030888661742210388, + 0.029626553878188133, + 0.010655802674591541, + 0.02082504890859127, + 0.010957085527479649, + 0.0070901601575315, + 0.009497888386249542, + 0.00715899420902133, + 0.005770627874881029, + 0.011635580100119114 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 8, + 0 + ], + "selected_reward": -0.0430525541305542 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "An epic aerial drone shot smoothly tracks a majestic golden eagle, its massive wingspan fully extended, effortlessly riding rising thermals as it gracefully circles above a vast, rugged valley. The raptor's plumage is a deep chocolate brown, and the golden late afternoon sun, striking it from a low angle, illuminates the signature golden feathers on the nape of its neck and crown, making it shimmer like a crown in the sky. Below, sheer granite cliffs and patches of deep green coniferous forest create a majestic landscape, while a silvery river winds through the valley floor, reflecting the light of the sky. The eagle glides in a wide, deliberate arc, its head turning with sharp, precise movements, its piercing amber eyes fixed intently downward, methodically scanning every inch of the landscape like a monarch surveying his kingdom. It occasionally adjusts its flying posture by slightly tilting its wings or slightly flapping its tail feathers. Its fleeting shadow sweeps across the rocks and treetops below like a ghost. The whole picture is full of primitive, untamed wild beauty, conveying a lonely and majestic kingly aura.", + "seed": 18134, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000050/group_8/index_4133_seed_18134_dp_17_rank_34_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 8, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.03051055339165032, + "vlm_logprob": -0.03051055339165032, + "vlm_yes_no_margin": 4.125, + "vlm_token_credit": -0.03051055339165032, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a golden eagle with a massive wingspan fully extended?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the eagle's plumage deep chocolate brown with golden feathers on the nape and crown?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the eagle gliding in a wide arc while riding rising thermals above a rugged valley?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the scene include sheer granite cliffs, deep green coniferous forest, and a silvery river below?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot an aerial drone view with the camera tracking the eagle smoothly?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.08861330151557922 + ], + "opposite_logprob_tokens": [ + -2.588613271713257 + ], + "logprob_reward": -0.08861330151557922, + "yes_no_margin_reward": 2.5, + "token_credit_reward": -0.08861330151557922, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.19326698780059814, + 0.1276896893978119, + 0.11295133829116821, + 0.11387965083122253, + 0.09690846502780914, + 0.07710368931293488, + 0.07032999396324158, + 0.09488596767187119, + 0.032786086201667786, + 0.08019818365573883 + ], + "credit_map": [ + 0.020512422546744347, + 0.02773698791861534, + 0.020297575742006302, + 0.018027517944574356, + 0.029341639950871468, + 0.01993948593735695, + 0.029544517397880554, + 0.027866844087839127, + 0.013402435928583145, + 0.021676143631339073, + 0.009733246639370918, + 0.010349523276090622, + 0.01732463762164116, + 0.030679946765303612, + 0.010391267947852612, + 0.014132481999695301, + 0.010325885377824306, + 0.015233625657856464, + 0.008734578266739845, + 0.014068935066461563, + 0.011368214152753353, + 0.014853962697088718, + 0.021386245265603065, + 0.016979897394776344, + 0.015439758077263832, + 0.013668928295373917, + 0.019231967628002167, + 0.010366749949753284, + 0.011369860731065273, + 0.01088983565568924, + 0.014331740327179432, + 0.01858081854879856, + 0.014440184459090233, + 0.013638792559504509, + 0.01879134774208069, + 0.00558308232575655, + 0.004673758056014776, + 0.011921858415007591, + 0.008960922248661518, + 0.018898511305451393, + 0.013397837057709694, + 0.010474169626832008, + 0.012649375014007092, + 0.0062369718216359615, + 0.006396952085196972, + 0.010828020051121712, + 0.007460400462150574, + 0.00965995341539383, + 0.007265517953783274, + 0.005149161908775568, + 0.01802782714366913, + 0.006714135408401489, + 0.005826829932630062, + 0.013169553130865097, + 0.006302474532276392, + 0.007874494418501854, + 0.0064697712659835815, + 0.011866752058267593, + 0.014698858372867107, + 0.008832138031721115, + 0.009872321970760822, + 0.018411051481962204, + 0.009340289980173111, + 0.015394787304103374, + 0.0018991346005350351, + 0.0030086515471339226, + 0.0046056159771978855, + 0.004989439621567726, + 0.0038615018129348755, + 0.0049811131320893764, + 0.004895227961242199, + 0.00454539991915226, + 0.004850360564887524, + 0.014333009719848633, + 0.018469883129000664, + 0.0073105283081531525, + 0.0053818682208657265, + 0.013315500691533089, + 0.00810879748314619, + 0.008428245782852173 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 3 + ], + "selected_reward": -0.08861330151557922 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a golden eagle with a massive wingspan fully extended?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the eagle's plumage deep chocolate brown with golden feathers on the nape and crown?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the eagle gliding in a wide arc while riding rising thermals above a rugged valley?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the scene include sheer granite cliffs, deep green coniferous forest, and a silvery river below?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot an aerial drone view with the camera tracking the eagle smoothly?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.029821693897247314 + ], + "opposite_logprob_tokens": [ + -3.5298216342926025 + ], + "logprob_reward": -0.029821693897247314, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.029821693897247314, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15909536182880402, + 0.07156360894441605, + 0.04871230944991112, + 0.16670991480350494, + 0.18579590320587158, + 0.08903076499700546, + 0.07918456941843033, + 0.08738990873098373, + 0.03289852663874626, + 0.07961910218000412 + ], + "credit_map": [ + 0.014040116220712662, + 0.02068988047540188, + 0.016552696004509926, + 0.0162199717015028, + 0.019132694229483604, + 0.022935178130865097, + 0.02716818079352379, + 0.02235664799809456, + 0.0037110429257154465, + 0.007212442345917225, + 0.006195229943841696, + 0.008268555626273155, + 0.00680607371032238, + 0.019821185618638992, + 0.006403440609574318, + 0.01314564049243927, + 0.007521932944655418, + 0.005170864053070545, + 0.00431318674236536, + 0.008478670381009579, + 0.0049773394130170345, + 0.005710489582270384, + 0.006205214187502861, + 0.006334611214697361, + 0.016770359128713608, + 0.017735496163368225, + 0.023034479469060898, + 0.015505854971706867, + 0.01295049674808979, + 0.016003208234906197, + 0.024561826139688492, + 0.04014819487929344, + 0.02027982473373413, + 0.017921119928359985, + 0.03938944637775421, + 0.009100674651563168, + 0.010645677335560322, + 0.034867063164711, + 0.01830238476395607, + 0.0352897085249424, + 0.012972859665751457, + 0.013809765689074993, + 0.01450232695788145, + 0.006952528841793537, + 0.006387496367096901, + 0.012680940330028534, + 0.01015305332839489, + 0.011571794748306274, + 0.008800500072538853, + 0.005613875109702349, + 0.017392344772815704, + 0.008502908051013947, + 0.007779215928167105, + 0.013486227951943874, + 0.007045013830065727, + 0.01056448370218277, + 0.007242577150464058, + 0.012777716852724552, + 0.01320398785173893, + 0.008816610090434551, + 0.011383664794266224, + 0.014587546698749065, + 0.008126361295580864, + 0.011251447722315788, + 0.0022897629532963037, + 0.003412826219573617, + 0.005270959343761206, + 0.00512880040332675, + 0.00273972493596375, + 0.004590170923620462, + 0.004096817225217819, + 0.005369465798139572, + 0.00519800279289484, + 0.014591989107429981, + 0.01985519379377365, + 0.009201353415846825, + 0.005260570906102657, + 0.011522822082042694, + 0.006240047514438629, + 0.00774912815541029 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 3, + 0 + ], + "selected_reward": -0.029821693897247314 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a golden eagle with a massive wingspan fully extended?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the eagle's plumage deep chocolate brown with golden feathers on the nape and crown?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the eagle gliding in a wide arc while riding rising thermals above a rugged valley?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the scene include sheer granite cliffs, deep green coniferous forest, and a silvery river below?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot an aerial drone view with the camera tracking the eagle smoothly?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.002198899397626519 + ], + "opposite_logprob_tokens": [ + -6.127198696136475 + ], + "logprob_reward": -0.002198899397626519, + "yes_no_margin_reward": 6.125, + "token_credit_reward": -0.002198899397626519, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1342860460281372, + 0.1040196567773819, + 0.2632651925086975, + 0.08409769088029861, + 0.07665964961051941, + 0.09488698840141296, + 0.06192287802696228, + 0.0690155029296875, + 0.03665706515312195, + 0.07518934458494186 + ], + "credit_map": [ + 0.015570477582514286, + 0.011562827043235302, + 0.014543231576681137, + 0.011940217576920986, + 0.017035117372870445, + 0.013625201769173145, + 0.021344492211937904, + 0.028664471581578255, + 0.01271736528724432, + 0.017228983342647552, + 0.011387568898499012, + 0.011764755472540855, + 0.013468585908412933, + 0.014349817298352718, + 0.01039747055619955, + 0.012705106288194656, + 0.0252858754247427, + 0.0329875573515892, + 0.01977144181728363, + 0.036125171929597855, + 0.023680955171585083, + 0.037434518337249756, + 0.05157509818673134, + 0.036404553800821304, + 0.013567943125963211, + 0.012038649059832096, + 0.013084887526929379, + 0.007413005456328392, + 0.00739688565954566, + 0.009153109975159168, + 0.010206999257206917, + 0.01123620755970478, + 0.012294354848563671, + 0.007291593123227358, + 0.012419627048075199, + 0.004365645349025726, + 0.004496470559388399, + 0.013598388992249966, + 0.007249454036355019, + 0.01494411751627922, + 0.015061674639582634, + 0.013125374913215637, + 0.015215333551168442, + 0.005295844748616219, + 0.006280260626226664, + 0.014975341968238354, + 0.009988759644329548, + 0.014944391325116158, + 0.006970935966819525, + 0.004121636971831322, + 0.014870302751660347, + 0.006554778199642897, + 0.006206690799444914, + 0.010565339587628841, + 0.005484838504344225, + 0.007148353848606348, + 0.00622998271137476, + 0.010419109836220741, + 0.011060669086873531, + 0.007450834847986698, + 0.006631223484873772, + 0.010437823832035065, + 0.006291140802204609, + 0.010494714602828026, + 0.002692859387025237, + 0.0038252384401857853, + 0.006952198687940836, + 0.004290630109608173, + 0.0034081770572811365, + 0.004585691727697849, + 0.004120157565921545, + 0.006782112177461386, + 0.005052448716014624, + 0.010278342291712761, + 0.015327501110732555, + 0.008029306307435036, + 0.006388816982507706, + 0.013859418220818043, + 0.006519958842545748, + 0.009733553975820541 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 0, + 1 + ], + "selected_reward": -0.002198899397626519 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a golden eagle with a massive wingspan fully extended?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the eagle's plumage deep chocolate brown with golden feathers on the nape and crown?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the eagle gliding in a wide arc while riding rising thermals above a rugged valley?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the scene include sheer granite cliffs, deep green coniferous forest, and a silvery river below?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot an aerial drone view with the camera tracking the eagle smoothly?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.008636504411697388 + ], + "opposite_logprob_tokens": [ + -4.758636474609375 + ], + "logprob_reward": -0.008636504411697388, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.008636504411697388, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11781278252601624, + 0.1316639482975006, + 0.04381154477596283, + 0.056332822889089584, + 0.12020643055438995, + 0.15995275974273682, + 0.06077507510781288, + 0.07056364417076111, + 0.04678930342197418, + 0.192091703414917 + ], + "credit_map": [ + 0.01263151504099369, + 0.006791388615965843, + 0.013143545016646385, + 0.008466950617730618, + 0.010403696447610855, + 0.015633881092071533, + 0.01922043040394783, + 0.03152136877179146, + 0.011059715412557125, + 0.01733996905386448, + 0.006539024878293276, + 0.008507038466632366, + 0.010963738895952702, + 0.041371993720531464, + 0.010392015799880028, + 0.025490449741482735, + 0.005083993077278137, + 0.0047691320069134235, + 0.004351567476987839, + 0.006969088222831488, + 0.004578765947371721, + 0.0054753669537603855, + 0.006611633580178022, + 0.005971997976303101, + 0.0066708847880363464, + 0.0054477122612297535, + 0.005924084223806858, + 0.005977938883006573, + 0.00621109688654542, + 0.005257208365947008, + 0.007892374880611897, + 0.012951517477631569, + 0.01085931621491909, + 0.012170356698334217, + 0.02407875284552574, + 0.006857738364487886, + 0.008887365460395813, + 0.0191974900662899, + 0.015430904924869537, + 0.02272450178861618, + 0.019205160439014435, + 0.024012528359889984, + 0.03097071684896946, + 0.014456812292337418, + 0.011321834288537502, + 0.018213413655757904, + 0.01724906824529171, + 0.024523219093680382, + 0.006347817834466696, + 0.004593092482537031, + 0.012639124877750874, + 0.00901837833225727, + 0.006715511437505484, + 0.008687477558851242, + 0.006368271075189114, + 0.006405400112271309, + 0.00787426345050335, + 0.013235549442470074, + 0.011233753524720669, + 0.007042478770017624, + 0.009518514387309551, + 0.009023674763739109, + 0.0058077191933989525, + 0.00682769063860178, + 0.003362619783729315, + 0.004311438649892807, + 0.0067998263984918594, + 0.006215119734406471, + 0.0056915138848125935, + 0.008241770789027214, + 0.0057669333182275295, + 0.006400078069418669, + 0.01370836142450571, + 0.03114956058561802, + 0.04163433611392975, + 0.020430561155080795, + 0.0189390629529953, + 0.025472812354564667, + 0.0177068579941988, + 0.023050136864185333 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 5, + 1 + ], + "selected_reward": -0.008636504411697388 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a golden eagle with a massive wingspan fully extended?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the eagle's plumage deep chocolate brown with golden feathers on the nape and crown?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the eagle gliding in a wide arc while riding rising thermals above a rugged valley?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the scene include sheer granite cliffs, deep green coniferous forest, and a silvery river below?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot an aerial drone view with the camera tracking the eagle smoothly?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02328236773610115 + ], + "opposite_logprob_tokens": [ + -3.773282289505005 + ], + "logprob_reward": -0.02328236773610115, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.02328236773610115, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15922695398330688, + 0.1293305903673172, + 0.06390749663114548, + 0.09151958674192429, + 0.05432233214378357, + 0.1795492321252823, + 0.09716183692216873, + 0.054053895175457, + 0.09725463390350342, + 0.07367341965436935 + ], + "credit_map": [ + 0.014978793449699879, + 0.015121095813810825, + 0.017457708716392517, + 0.017643360421061516, + 0.017564229667186737, + 0.014893111772835255, + 0.025479288771748543, + 0.03608936071395874, + 0.010795987211167812, + 0.018713893368840218, + 0.01455008052289486, + 0.015396904200315475, + 0.019828787073493004, + 0.01865081489086151, + 0.016211766749620438, + 0.015182369388639927, + 0.006412905640900135, + 0.008022494614124298, + 0.005149209871888161, + 0.012633618898689747, + 0.005029745865613222, + 0.004947667941451073, + 0.01362660713493824, + 0.008085246197879314, + 0.0124315544962883, + 0.011904544197022915, + 0.018201516941189766, + 0.008348875679075718, + 0.007850565016269684, + 0.006494800094515085, + 0.011954857036471367, + 0.014332865364849567, + 0.009983707219362259, + 0.006239437498152256, + 0.006362411193549633, + 0.00415972713381052, + 0.005741075612604618, + 0.00652220007032156, + 0.006262322422116995, + 0.009051448665559292, + 0.020996930077672005, + 0.02624870464205742, + 0.030591674149036407, + 0.01694953627884388, + 0.01810530200600624, + 0.022808276116847992, + 0.022976098582148552, + 0.02087271772325039, + 0.009304806590080261, + 0.005626182071864605, + 0.022421784698963165, + 0.010602867230772972, + 0.009616324678063393, + 0.018254978582262993, + 0.009020274505019188, + 0.012314623221755028, + 0.004782059229910374, + 0.0059914360754191875, + 0.00909082405269146, + 0.005394711159169674, + 0.006319150328636169, + 0.007830839604139328, + 0.005621707066893578, + 0.009023170918226242, + 0.006148350890725851, + 0.007774342317134142, + 0.015083357691764832, + 0.01295074075460434, + 0.011255183257162571, + 0.016257761046290398, + 0.012197197414934635, + 0.01558770053088665, + 0.0037242204416543245, + 0.008865605108439922, + 0.012446969747543335, + 0.008362697437405586, + 0.008333186618983746, + 0.008751504123210907, + 0.014449828304350376, + 0.008739412762224674 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 0, + 1 + ], + "selected_reward": -0.02328236773610115 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000050/group_9/index_tensor([3898, 3898])_group_9_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000050/group_9/index_tensor([3898, 3898])_group_9_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..ff544a962badcbceb3a185242982d8f67f42aaf4 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000050/group_9/index_tensor([3898, 3898])_group_9_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "horizontal medium shot, panning smoothly to the right, tracks a massive kangaroo at the apex of a powerful leap, gracefully crossing a vast expanse of water. Nestled within a deep, dark cavern, the kangaroo's muscular body stretches to its full extent, from the tip of its long, thick tail, acting as a stabilizer, to its protruding nose. Its dense, rich, reddish-brown fur is softly outlined in the dim light. The scene's sole source of light comes from a dense, silent, floating swarm of bioluminescent fireflies, emitting a steady, ethereal blue glow that imbues the darkness with a mystical hue. This magical glow shimmers upon the surrounding damp, craggy rock faces and, more strikingly, is perfectly reflected upon the still, mirror-like black water below, creating a breathtaking symmetrical illusion that perfectly replicates the kangaroo's arched silhouette and its thousands of twinkling blue points of light, evoking a tranquil, dreamlike atmosphere imbued with the wonder of life.", + "seed": 11231, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000050/group_9/index_3898_seed_11231_dp_18_rank_36_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 9, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.06591942412778735, + "vlm_logprob": -0.06591942412778735, + "vlm_yes_no_margin": 3.85, + "vlm_token_credit": -0.06591942412778735, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot framed as a horizontal medium shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera pan smoothly to the right while tracking a massive kangaroo?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the kangaroo shown leaping across a vast expanse of water within a deep, dark cavern?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene illuminated by a swarm of bioluminescent fireflies emitting a steady blue glow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the blue glow from the fireflies reflected on the still, mirror-like black water below?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.12880127131938934 + ], + "opposite_logprob_tokens": [ + -2.3788013458251953 + ], + "logprob_reward": -0.12880127131938934, + "yes_no_margin_reward": 2.25, + "token_credit_reward": -0.12880127131938934, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15748795866966248, + 0.09881095588207245, + 0.0643148198723793, + 0.1441056877374649, + 0.05751274898648262, + 0.1086924746632576, + 0.19821354746818542, + 0.11302855610847473, + 0.034486498683691025, + 0.023346707224845886 + ], + "credit_map": [ + 0.028734469786286354, + 0.02027007006108761, + 0.018550721928477287, + 0.012618844397366047, + 0.019398704171180725, + 0.02101764641702175, + 0.017978928983211517, + 0.018918568268418312, + 0.018322715535759926, + 0.010339422151446342, + 0.009885852225124836, + 0.009684555232524872, + 0.012758786790072918, + 0.013638835400342941, + 0.012717780657112598, + 0.01146300695836544, + 0.007544454652816057, + 0.005911125801503658, + 0.0074971760623157024, + 0.007283744402229786, + 0.008053479716181755, + 0.007757178507745266, + 0.009977499954402447, + 0.01029016450047493, + 0.01815292239189148, + 0.01877748779952526, + 0.015073508024215698, + 0.01344182901084423, + 0.025798020884394646, + 0.01961248554289341, + 0.020971475169062614, + 0.012277972884476185, + 0.00632776552811265, + 0.007563376799225807, + 0.005063420627266169, + 0.007720274850726128, + 0.0077627431601285934, + 0.006590606179088354, + 0.012120121158659458, + 0.004364443942904472, + 0.011716144159436226, + 0.013479418121278286, + 0.019254442304372787, + 0.011390299536287785, + 0.015008394606411457, + 0.015049116685986519, + 0.011501873843371868, + 0.011292786337435246, + 0.015323374420404434, + 0.0489530935883522, + 0.017169009894132614, + 0.013053247705101967, + 0.0533914677798748, + 0.022395359352231026, + 0.013088167645037174, + 0.01483982428908348, + 0.009056014008820057, + 0.016759201884269714, + 0.011182266287505627, + 0.018608681857585907, + 0.013992726802825928, + 0.01900647208094597, + 0.010059447027742863, + 0.014363745227456093, + 0.0035519208759069443, + 0.00469087902456522, + 0.004402408841997385, + 0.0050195688381791115, + 0.0037567687686532736, + 0.005719280336052179, + 0.00336838373914361, + 0.0039772880263626575, + 0.0026528651360422373, + 0.0032141711562871933, + 0.0024242675863206387, + 0.002590601099655032, + 0.0027249278500676155, + 0.0042994385585188866, + 0.002014940371736884, + 0.003425496630370617 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 0, + 3 + ], + "selected_reward": -0.12880127131938934 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot framed as a horizontal medium shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera pan smoothly to the right while tracking a massive kangaroo?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the kangaroo shown leaping across a vast expanse of water within a deep, dark cavern?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene illuminated by a swarm of bioluminescent fireflies emitting a steady blue glow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the blue glow from the fireflies reflected on the still, mirror-like black water below?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.17989324033260345 + ], + "opposite_logprob_tokens": [ + -1.8048932552337646 + ], + "logprob_reward": -0.17989324033260345, + "yes_no_margin_reward": 1.625, + "token_credit_reward": -0.17989324033260345, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0653710812330246, + 0.05620197951793671, + 0.0818907842040062, + 0.2145666629076004, + 0.04336338862776756, + 0.10065975785255432, + 0.1536143273115158, + 0.16277086734771729, + 0.07921221852302551, + 0.0423489585518837 + ], + "credit_map": [ + 0.007862957194447517, + 0.005853793118149042, + 0.006639732513576746, + 0.00552194332703948, + 0.008224623277783394, + 0.009175964631140232, + 0.011097791604697704, + 0.010994281619787216, + 0.007986987940967083, + 0.008836818858981133, + 0.005140053573995829, + 0.006346561014652252, + 0.010040477849543095, + 0.007835247553884983, + 0.00488220015540719, + 0.005133630242198706, + 0.011044175364077091, + 0.012743036262691021, + 0.006337542086839676, + 0.007997842505574226, + 0.014981568790972233, + 0.010715789161622524, + 0.012266866862773895, + 0.005803963169455528, + 0.022372402250766754, + 0.03259120136499405, + 0.020031509920954704, + 0.01599358394742012, + 0.042936909943819046, + 0.029303114861249924, + 0.036404114216566086, + 0.014933813363313675, + 0.004151014145463705, + 0.005801481660455465, + 0.00424175476655364, + 0.006134995259344578, + 0.006255663000047207, + 0.0056594014167785645, + 0.0068693398497998714, + 0.004249734804034233, + 0.008844859898090363, + 0.012464086525142193, + 0.016638120636343956, + 0.011297415941953659, + 0.013475778512656689, + 0.015391988679766655, + 0.011088577099144459, + 0.011458927765488625, + 0.010621706955134869, + 0.04046030342578888, + 0.009344699792563915, + 0.013085850514471531, + 0.04395885393023491, + 0.016227461397647858, + 0.009699384681880474, + 0.010216065682470798, + 0.012454739771783352, + 0.02317539043724537, + 0.017711935564875603, + 0.032320693135261536, + 0.01981155201792717, + 0.020512595772743225, + 0.022204067558050156, + 0.014579897746443748, + 0.009418366476893425, + 0.010498094372451305, + 0.009361292235553265, + 0.014288355596363544, + 0.009194178506731987, + 0.011423525400459766, + 0.008077256381511688, + 0.006951153744012117, + 0.0034612291492521763, + 0.006102053448557854, + 0.0037232208997011185, + 0.0034790660720318556, + 0.005677727982401848, + 0.008654464036226273, + 0.003612055443227291, + 0.0076391384936869144 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 7, + 6 + ], + "selected_reward": -0.17989324033260345 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot framed as a horizontal medium shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera pan smoothly to the right while tracking a massive kangaroo?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the kangaroo shown leaping across a vast expanse of water within a deep, dark cavern?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene illuminated by a swarm of bioluminescent fireflies emitting a steady blue glow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the blue glow from the fireflies reflected on the still, mirror-like black water below?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.008639931678771973 + ], + "opposite_logprob_tokens": [ + -4.758639812469482 + ], + "logprob_reward": -0.008639931678771973, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.008639931678771973, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08665557950735092, + 0.18275362253189087, + 0.09519091248512268, + 0.12359428405761719, + 0.05397402122616768, + 0.06405434757471085, + 0.12880252301692963, + 0.1880461871623993, + 0.04705055058002472, + 0.029878001660108566 + ], + "credit_map": [ + 0.010500642471015453, + 0.00922095775604248, + 0.009278135374188423, + 0.00818129163235426, + 0.01061533484607935, + 0.01211731880903244, + 0.012069668620824814, + 0.01467222161591053, + 0.025885041803121567, + 0.017701976001262665, + 0.016386447474360466, + 0.014091823250055313, + 0.04228480905294418, + 0.025543931871652603, + 0.018567539751529694, + 0.022292034700512886, + 0.012705650180578232, + 0.014049266465008259, + 0.01024921890348196, + 0.011992843821644783, + 0.015095371752977371, + 0.012583917006850243, + 0.01041038241237402, + 0.008104249835014343, + 0.015613621100783348, + 0.01537590567022562, + 0.01519839745014906, + 0.009751922450959682, + 0.019363412633538246, + 0.016180388629436493, + 0.021832989528775215, + 0.010277635417878628, + 0.005928460042923689, + 0.005237935576587915, + 0.007089064922183752, + 0.007821379229426384, + 0.0064356159418821335, + 0.006293268874287605, + 0.009322745725512505, + 0.005845542531460524, + 0.005661812145262957, + 0.008640644140541553, + 0.007477991282939911, + 0.00916193425655365, + 0.008042656816542149, + 0.012435246258974075, + 0.006200677715241909, + 0.006433384492993355, + 0.00829008687287569, + 0.02886809967458248, + 0.00876597035676241, + 0.014167058281600475, + 0.03469289094209671, + 0.013894299045205116, + 0.008388174697756767, + 0.011735938489437103, + 0.013825740665197372, + 0.029596328735351562, + 0.017571715638041496, + 0.038397617638111115, + 0.023865781724452972, + 0.02269464172422886, + 0.019374392926692963, + 0.022719943895936012, + 0.0036465071607381105, + 0.00853120256215334, + 0.0048153214156627655, + 0.004214156419038773, + 0.007337084971368313, + 0.011224756017327309, + 0.0033559612929821014, + 0.003925558179616928, + 0.0030682068318128586, + 0.0044034854508936405, + 0.0032559558749198914, + 0.003196505829691887, + 0.004849584773182869, + 0.005339821334928274, + 0.002763428259640932, + 0.0030010128393769264 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 1, + 6 + ], + "selected_reward": -0.008639931678771973 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot framed as a horizontal medium shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera pan smoothly to the right while tracking a massive kangaroo?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the kangaroo shown leaping across a vast expanse of water within a deep, dark cavern?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene illuminated by a swarm of bioluminescent fireflies emitting a steady blue glow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the blue glow from the fireflies reflected on the still, mirror-like black water below?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.009777512401342392 + ], + "opposite_logprob_tokens": [ + -4.634777545928955 + ], + "logprob_reward": -0.009777512401342392, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.009777512401342392, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13351678848266602, + 0.043018653988838196, + 0.07524798810482025, + 0.05524319037795067, + 0.06508300453424454, + 0.08968408405780792, + 0.1247013658285141, + 0.30945342779159546, + 0.08005442470312119, + 0.023997098207473755 + ], + "credit_map": [ + 0.015297933481633663, + 0.008508414030075073, + 0.009511426091194153, + 0.009639091789722443, + 0.012625939212739468, + 0.021328870207071304, + 0.02695094421505928, + 0.029654162004590034, + 0.007134957704693079, + 0.004031885880976915, + 0.003304839599877596, + 0.004359271842986345, + 0.005614686291664839, + 0.006539804860949516, + 0.005093321204185486, + 0.006939886603504419, + 0.008870396763086319, + 0.007748157251626253, + 0.007307081483304501, + 0.006950379814952612, + 0.0074865552596747875, + 0.01012468058615923, + 0.01591324433684349, + 0.01084749773144722, + 0.007222842890769243, + 0.007091111037880182, + 0.006310319062322378, + 0.005131218582391739, + 0.008532763458788395, + 0.007648345548659563, + 0.007582046557217836, + 0.005724540911614895, + 0.006905506365001202, + 0.008177965879440308, + 0.007404709234833717, + 0.011153703555464745, + 0.007175162900239229, + 0.006716650910675526, + 0.011607561260461807, + 0.005941745359450579, + 0.007846093736588955, + 0.01071943435817957, + 0.013099328614771366, + 0.013378014788031578, + 0.010922909714281559, + 0.012505214661359787, + 0.01154958177357912, + 0.00966351106762886, + 0.008488120511174202, + 0.02599572390317917, + 0.010720520280301571, + 0.011382672935724258, + 0.02719225361943245, + 0.020022476091980934, + 0.01001068390905857, + 0.010888919234275818, + 0.020368807017803192, + 0.0425574965775013, + 0.04130474850535393, + 0.06334188580513, + 0.03311419486999512, + 0.03690267354249954, + 0.04736423119902611, + 0.024499399587512016, + 0.006042878609150648, + 0.012165098451077938, + 0.007638182025402784, + 0.007575132884085178, + 0.011812764219939709, + 0.02058137021958828, + 0.004769982770085335, + 0.00946900900453329, + 0.0026082838885486126, + 0.003099720226600766, + 0.003219189587980509, + 0.0033535894472151995, + 0.003017042763531208, + 0.0032765858341008425, + 0.0029623019509017467, + 0.0024603852070868015 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 0, + 6 + ], + "selected_reward": -0.009777512401342392 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot framed as a horizontal medium shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera pan smoothly to the right while tracking a massive kangaroo?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the kangaroo shown leaping across a vast expanse of water within a deep, dark cavern?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene illuminated by a swarm of bioluminescent fireflies emitting a steady blue glow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the blue glow from the fireflies reflected on the still, mirror-like black water below?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0024851649068295956 + ], + "opposite_logprob_tokens": [ + -6.002485275268555 + ], + "logprob_reward": -0.0024851649068295956, + "yes_no_margin_reward": 6.0, + "token_credit_reward": -0.0024851649068295956, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.39485207200050354, + 0.04934055358171463, + 0.04385606572031975, + 0.07859615236520767, + 0.13560380041599274, + 0.05323223024606705, + 0.0656224936246872, + 0.12578774988651276, + 0.03143073245882988, + 0.021678106859326363 + ], + "credit_map": [ + 0.04130614548921585, + 0.023303629830479622, + 0.04757321998476982, + 0.03588588535785675, + 0.03890400379896164, + 0.06285424530506134, + 0.07220464199781418, + 0.07282033562660217, + 0.00814316887408495, + 0.005581866018474102, + 0.004578125663101673, + 0.005337445996701717, + 0.005923437420278788, + 0.006247831042855978, + 0.007272511720657349, + 0.006256169639527798, + 0.00644419901072979, + 0.006456609349697828, + 0.0035295409616082907, + 0.00437126774340868, + 0.0063819908536970615, + 0.00628573726862669, + 0.0048792255111038685, + 0.005507499445229769, + 0.00858009047806263, + 0.009970863349735737, + 0.008010744117200375, + 0.006950823124498129, + 0.013724113814532757, + 0.010852377861738205, + 0.011952513828873634, + 0.008554628118872643, + 0.013174336403608322, + 0.01878979615867138, + 0.014063674956560135, + 0.025422774255275726, + 0.014351666904985905, + 0.020317552611231804, + 0.01842472143471241, + 0.011059285141527653, + 0.004723974969238043, + 0.007596331182867289, + 0.008869468234479427, + 0.006456525065004826, + 0.006667579989880323, + 0.008349694311618805, + 0.005073199979960918, + 0.0054954602383077145, + 0.0047409250400960445, + 0.01504724845290184, + 0.005002217832952738, + 0.00432969257235527, + 0.016117267310619354, + 0.010992967523634434, + 0.0037178383208811283, + 0.005674336571246386, + 0.008902360685169697, + 0.018995629623532295, + 0.015366350300610065, + 0.024133002385497093, + 0.015298205427825451, + 0.012883131392300129, + 0.018640384078025818, + 0.011568697169423103, + 0.0029781328048557043, + 0.004100702702999115, + 0.004650749731808901, + 0.005124558694660664, + 0.0030843608547002077, + 0.005398603156208992, + 0.0029456496704369783, + 0.003147978102788329, + 0.00200590118765831, + 0.002768169157207012, + 0.0025142929516732693, + 0.0026226553600281477, + 0.0034211757592856884, + 0.0033173495903611183, + 0.002178802154958248, + 0.0028497616294771433 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 7 + ], + "selected_reward": -0.0024851649068295956 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "horizontal medium shot, panning smoothly to the right, tracks a massive kangaroo at the apex of a powerful leap, gracefully crossing a vast expanse of water. Nestled within a deep, dark cavern, the kangaroo's muscular body stretches to its full extent, from the tip of its long, thick tail, acting as a stabilizer, to its protruding nose. Its dense, rich, reddish-brown fur is softly outlined in the dim light. The scene's sole source of light comes from a dense, silent, floating swarm of bioluminescent fireflies, emitting a steady, ethereal blue glow that imbues the darkness with a mystical hue. This magical glow shimmers upon the surrounding damp, craggy rock faces and, more strikingly, is perfectly reflected upon the still, mirror-like black water below, creating a breathtaking symmetrical illusion that perfectly replicates the kangaroo's arched silhouette and its thousands of twinkling blue points of light, evoking a tranquil, dreamlike atmosphere imbued with the wonder of life.", + "seed": 11232, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000050/group_9/index_3898_seed_11232_dp_18_rank_36_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 9, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.12000730270519852, + "vlm_logprob": -0.12000730270519852, + "vlm_yes_no_margin": 3.575, + "vlm_token_credit": -0.12000730270519852, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot framed as a horizontal medium shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera pan smoothly to the right while tracking a massive kangaroo?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the kangaroo shown leaping across a vast expanse of water within a deep, dark cavern?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene illuminated by a swarm of bioluminescent fireflies emitting a steady blue glow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the blue glow from the fireflies reflected on the still, mirror-like black water below?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.09912293404340744 + ], + "opposite_logprob_tokens": [ + -2.849123001098633 + ], + "logprob_reward": -0.09912293404340744, + "yes_no_margin_reward": 2.75, + "token_credit_reward": -0.09912293404340744, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.04233710467815399, + 0.02068387158215046, + 0.05270646885037422, + 0.11970416456460953, + 0.033079568296670914, + 0.027215227484703064, + 0.027204109355807304, + 0.05349528044462204, + 0.6143786311149597, + 0.009195561520755291 + ], + "credit_map": [ + 0.005391482729464769, + 0.005568195134401321, + 0.006169384345412254, + 0.007883131504058838, + 0.0031738420948386192, + 0.005095076747238636, + 0.00448619993403554, + 0.00456979451701045, + 0.0016454909928143024, + 0.0033386065624654293, + 0.0029368360992521048, + 0.0020829106215387583, + 0.0030081099830567837, + 0.0031688432209193707, + 0.0020950569305568933, + 0.002408017637208104, + 0.006158605217933655, + 0.005107735749334097, + 0.008024947717785835, + 0.00540435453876853, + 0.008285794407129288, + 0.007931791245937347, + 0.007204347290098667, + 0.004588892683386803, + 0.013574269600212574, + 0.012988853268325329, + 0.013306977227330208, + 0.019702520221471786, + 0.01640128344297409, + 0.01729370653629303, + 0.01572984643280506, + 0.010706709697842598, + 0.0033744527027010918, + 0.002971766982227564, + 0.005079582333564758, + 0.004138689022511244, + 0.004682013299316168, + 0.004902830813080072, + 0.004364351276308298, + 0.0035658825654536486, + 0.004003634676337242, + 0.0018757255747914314, + 0.004013759549707174, + 0.0038573360070586205, + 0.002773019252344966, + 0.003686128417029977, + 0.003951893653720617, + 0.0030537324491888285, + 0.0030883634462952614, + 0.0016957668121904135, + 0.002934908727183938, + 0.0028535029850900173, + 0.0025586781557649374, + 0.00423644483089447, + 0.0025814869441092014, + 0.007254958152770996, + 0.008864125236868858, + 0.006886304821819067, + 0.005852425470948219, + 0.006622341927140951, + 0.0066560604609549046, + 0.006981220096349716, + 0.0038983861450105906, + 0.0077344193123281, + 0.24077501893043518, + 0.018945330753922462, + 0.24134401977062225, + 0.03511521965265274, + 0.024876387789845467, + 0.0197147224098444, + 0.011627045460045338, + 0.02198086492717266, + 0.0016619133530184627, + 0.0008480979595333338, + 0.0012278195936232805, + 0.001056316657923162, + 0.0012262113159522414, + 0.0014701748732477427, + 0.0008878557709977031, + 0.0008171714725904167 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 3, + 7 + ], + "selected_reward": -0.09912293404340744 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot framed as a horizontal medium shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera pan smoothly to the right while tracking a massive kangaroo?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the kangaroo shown leaping across a vast expanse of water within a deep, dark cavern?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene illuminated by a swarm of bioluminescent fireflies emitting a steady blue glow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the blue glow from the fireflies reflected on the still, mirror-like black water below?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.4742688536643982 + ], + "opposite_logprob_tokens": [ + -0.9742688536643982 + ], + "logprob_reward": -0.4742688536643982, + "yes_no_margin_reward": 0.5, + "token_credit_reward": -0.4742688536643982, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07101938873529434, + 0.038676146417856216, + 0.05516045168042183, + 0.031201979145407677, + 0.03670788183808327, + 0.02328682132065296, + 0.02861919440329075, + 0.14920218288898468, + 0.4939977526664734, + 0.07212817668914795 + ], + "credit_map": [ + 0.00880164373666048, + 0.011663292534649372, + 0.00665011303499341, + 0.009505300782620907, + 0.00929565541446209, + 0.009000945836305618, + 0.008661137893795967, + 0.007441311608999968, + 0.0037021636962890625, + 0.00828090775758028, + 0.0036340279038995504, + 0.003312865737825632, + 0.006162191275507212, + 0.005984885152429342, + 0.004363551270216703, + 0.003235558047890663, + 0.005850503221154213, + 0.005267667584121227, + 0.007621989585459232, + 0.00636454951018095, + 0.009393314830958843, + 0.009459179826080799, + 0.0064531792886555195, + 0.004750068765133619, + 0.0038449650164693594, + 0.003011990338563919, + 0.004064928740262985, + 0.0041410718113183975, + 0.004995462950319052, + 0.004372649826109409, + 0.003329017199575901, + 0.003441896755248308, + 0.00402760598808527, + 0.00213656690903008, + 0.006666315719485283, + 0.004710296168923378, + 0.003692427184432745, + 0.006079084239900112, + 0.004524928517639637, + 0.0048706610687077045, + 0.002838369458913803, + 0.0023497999645769596, + 0.0031253458000719547, + 0.0028880960308015347, + 0.004824367817491293, + 0.0028351694345474243, + 0.002317027887329459, + 0.002108647022396326, + 0.003222087863832712, + 0.0023023225367069244, + 0.004321371670812368, + 0.002505096374079585, + 0.003521702019497752, + 0.005197965539991856, + 0.0036773968022316694, + 0.003871253691613674, + 0.0190598014742136, + 0.018238594755530357, + 0.020364126190543175, + 0.014791835099458694, + 0.024515459313988686, + 0.017202232033014297, + 0.009188016876578331, + 0.025842124596238136, + 0.19293667376041412, + 0.015567636117339134, + 0.19489102065563202, + 0.027861803770065308, + 0.01993900164961815, + 0.01603698544204235, + 0.010003963485360146, + 0.016760725528001785, + 0.010425403714179993, + 0.005891140084713697, + 0.008544380776584148, + 0.011394647881388664, + 0.009900663048028946, + 0.013430575840175152, + 0.005689672194421291, + 0.006851695943623781 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 7, + 9 + ], + "selected_reward": -0.4742688536643982 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot framed as a horizontal medium shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera pan smoothly to the right while tracking a massive kangaroo?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the kangaroo shown leaping across a vast expanse of water within a deep, dark cavern?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene illuminated by a swarm of bioluminescent fireflies emitting a steady blue glow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the blue glow from the fireflies reflected on the still, mirror-like black water below?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005949288606643677 + ], + "opposite_logprob_tokens": [ + -5.1309494972229 + ], + "logprob_reward": -0.005949288606643677, + "yes_no_margin_reward": 5.125, + "token_credit_reward": -0.005949288606643677, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06789982318878174, + 0.02932986430823803, + 0.17762798070907593, + 0.2827554941177368, + 0.06251297146081924, + 0.09579642117023468, + 0.03093738481402397, + 0.15824365615844727, + 0.026134328916668892, + 0.06876206398010254 + ], + "credit_map": [ + 0.006921129301190376, + 0.009700607508420944, + 0.008741114288568497, + 0.007135178428143263, + 0.010024701245129108, + 0.008264003321528435, + 0.008286852389574051, + 0.00882624089717865, + 0.003111473051831126, + 0.005798478610813618, + 0.0038217143155634403, + 0.002183650154620409, + 0.004638511687517166, + 0.004630551673471928, + 0.0026400128845125437, + 0.0025054721627384424, + 0.02363387495279312, + 0.014040903188288212, + 0.02030361071228981, + 0.019449740648269653, + 0.02142489142715931, + 0.02483094111084938, + 0.03417207673192024, + 0.019771944731473923, + 0.029622171074151993, + 0.030620615929365158, + 0.03157134726643562, + 0.0495203398168087, + 0.03861154988408089, + 0.04036565497517586, + 0.035695526748895645, + 0.02674826979637146, + 0.006683500483632088, + 0.003764664987102151, + 0.0071230498142540455, + 0.00919428188353777, + 0.006199368275702, + 0.010008987970650196, + 0.011267619207501411, + 0.008271493948996067, + 0.01998780108988285, + 0.007398279383778572, + 0.015568705275654793, + 0.010159460827708244, + 0.009320719167590141, + 0.013382842764258385, + 0.007620781660079956, + 0.012357821688055992, + 0.00444160308688879, + 0.0025254464708268642, + 0.0042843325063586235, + 0.002739296294748783, + 0.002683002734556794, + 0.006396929267793894, + 0.0026570248883217573, + 0.005209746304899454, + 0.019755637273192406, + 0.018396729603409767, + 0.021237097680568695, + 0.018041150644421577, + 0.024897601455450058, + 0.018493305891752243, + 0.01007625088095665, + 0.027345888316631317, + 0.005188385024666786, + 0.0024169590324163437, + 0.004179679788649082, + 0.003056805580854416, + 0.003201737767085433, + 0.0029485849663615227, + 0.0016027058009058237, + 0.0035394709557294846, + 0.010096930898725986, + 0.0060471296310424805, + 0.007425840012729168, + 0.009473511017858982, + 0.011213238351047039, + 0.012211153283715248, + 0.006268895696848631, + 0.006025375798344612 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 2, + 7 + ], + "selected_reward": -0.005949288606643677 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot framed as a horizontal medium shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera pan smoothly to the right while tracking a massive kangaroo?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the kangaroo shown leaping across a vast expanse of water within a deep, dark cavern?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene illuminated by a swarm of bioluminescent fireflies emitting a steady blue glow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the blue glow from the fireflies reflected on the still, mirror-like black water below?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01606203243136406 + ], + "opposite_logprob_tokens": [ + -4.141062259674072 + ], + "logprob_reward": -0.01606203243136406, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.01606203243136406, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.052826978266239166, + 0.03301670402288437, + 0.04052548110485077, + 0.07282070070505142, + 0.03067188523709774, + 0.03278139606118202, + 0.03685523569583893, + 0.16215506196022034, + 0.4900188148021698, + 0.048327796161174774 + ], + "credit_map": [ + 0.006106647662818432, + 0.005924745928496122, + 0.004875458776950836, + 0.005816448945552111, + 0.004388467408716679, + 0.006562695372849703, + 0.009125065058469772, + 0.010027450509369373, + 0.00340312160551548, + 0.005562864243984222, + 0.003269362263381481, + 0.003824439598247409, + 0.004934667609632015, + 0.005090044811367989, + 0.0033566723577678204, + 0.0035755278076976538, + 0.004348337650299072, + 0.0036827148869633675, + 0.0042832051403820515, + 0.004604030400514603, + 0.007705033756792545, + 0.006198100745677948, + 0.00534777669236064, + 0.004356278572231531, + 0.008762850426137447, + 0.00820022914558649, + 0.0073181684128940105, + 0.011675208806991577, + 0.010962313041090965, + 0.01005171611905098, + 0.00943051464855671, + 0.006419695448130369, + 0.0037261571269482374, + 0.0020563786383718252, + 0.005571424961090088, + 0.003325466066598892, + 0.002680977573618293, + 0.004831608384847641, + 0.003908106125891209, + 0.0045717628672719, + 0.005769941955804825, + 0.0021872229408472776, + 0.005990561563521624, + 0.003869730979204178, + 0.0033407798036932945, + 0.004529099445790052, + 0.004129410721361637, + 0.0029646442271769047, + 0.005046186037361622, + 0.0033278237096965313, + 0.004432810004800558, + 0.003393867053091526, + 0.004707180894911289, + 0.0043300543911755085, + 0.0031076394952833652, + 0.00850966852158308, + 0.02599022351205349, + 0.02265406772494316, + 0.0182491447776556, + 0.01344015821814537, + 0.02971874177455902, + 0.01577371172606945, + 0.01051186490803957, + 0.0258171409368515, + 0.19163201749324799, + 0.015684109181165695, + 0.19360016286373138, + 0.026002639904618263, + 0.02104666270315647, + 0.0160475242882967, + 0.009827904403209686, + 0.016177751123905182, + 0.0072122663259506226, + 0.004731713328510523, + 0.006287478841841221, + 0.006233175750821829, + 0.005442989524453878, + 0.009010262787342072, + 0.005155304912477732, + 0.00425460422411561 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 7, + 3 + ], + "selected_reward": -0.01606203243136406 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot framed as a horizontal medium shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera pan smoothly to the right while tracking a massive kangaroo?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the kangaroo shown leaping across a vast expanse of water within a deep, dark cavern?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene illuminated by a swarm of bioluminescent fireflies emitting a steady blue glow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the blue glow from the fireflies reflected on the still, mirror-like black water below?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004633404780179262 + ], + "opposite_logprob_tokens": [ + -5.37963342666626 + ], + "logprob_reward": -0.004633404780179262, + "yes_no_margin_reward": 5.375, + "token_credit_reward": -0.004633404780179262, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.022484924644231796, + 0.027080880478024483, + 0.06531889736652374, + 0.018764369189739227, + 0.018917864188551903, + 0.16444510221481323, + 0.012439168989658356, + 0.03416234627366066, + 0.6222833395004272, + 0.014103034511208534 + ], + "credit_map": [ + 0.002518116496503353, + 0.0037563208024948835, + 0.0016418315935879946, + 0.001968535827472806, + 0.00356995384208858, + 0.0029110321775078773, + 0.0031680925749242306, + 0.0029510436579585075, + 0.002040901454165578, + 0.00277003669179976, + 0.0043794200755655766, + 0.002430916065350175, + 0.0033555738627910614, + 0.0039509860798716545, + 0.003956656437367201, + 0.00419639190658927, + 0.008634114637970924, + 0.0059323725290596485, + 0.007352323271334171, + 0.007794877979904413, + 0.011202673427760601, + 0.010829482227563858, + 0.007632260210812092, + 0.005940796807408333, + 0.0025490110274404287, + 0.0016844056081026793, + 0.002737639704719186, + 0.0016546495025977492, + 0.00254642590880394, + 0.002965701976791024, + 0.0028659687377512455, + 0.001760568586178124, + 0.0023596349637955427, + 0.001344447722658515, + 0.002848377451300621, + 0.002433021320030093, + 0.0019778318237513304, + 0.0032124603167176247, + 0.002147312043234706, + 0.002594779245555401, + 0.034750234335660934, + 0.010052359662950039, + 0.032417070120573044, + 0.015938973054289818, + 0.013521726243197918, + 0.022827841341495514, + 0.013336257077753544, + 0.02160065434873104, + 0.001869145897217095, + 0.0008207883220165968, + 0.0026787002570927143, + 0.001126571442000568, + 0.0009852104121819139, + 0.0016276967944577336, + 0.0018729843432083726, + 0.0014580722199752927, + 0.006294941063970327, + 0.004885194823145866, + 0.003937850706279278, + 0.003961979411542416, + 0.004697585944086313, + 0.0038011441938579082, + 0.0022467095404863358, + 0.004336945246905088, + 0.2447168081998825, + 0.0194107536226511, + 0.24561084806919098, + 0.034105557948350906, + 0.0256057046353817, + 0.020069357007741928, + 0.011868894100189209, + 0.0208954606205225, + 0.0022178939543664455, + 0.0015551659744232893, + 0.00161654909607023, + 0.0017671718960627913, + 0.0021416640374809504, + 0.002098227385431528, + 0.001539206481538713, + 0.0011671571992337704 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 5, + 2 + ], + "selected_reward": -0.004633404780179262 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "horizontal medium shot, panning smoothly to the right, tracks a massive kangaroo at the apex of a powerful leap, gracefully crossing a vast expanse of water. Nestled within a deep, dark cavern, the kangaroo's muscular body stretches to its full extent, from the tip of its long, thick tail, acting as a stabilizer, to its protruding nose. Its dense, rich, reddish-brown fur is softly outlined in the dim light. The scene's sole source of light comes from a dense, silent, floating swarm of bioluminescent fireflies, emitting a steady, ethereal blue glow that imbues the darkness with a mystical hue. This magical glow shimmers upon the surrounding damp, craggy rock faces and, more strikingly, is perfectly reflected upon the still, mirror-like black water below, creating a breathtaking symmetrical illusion that perfectly replicates the kangaroo's arched silhouette and its thousands of twinkling blue points of light, evoking a tranquil, dreamlike atmosphere imbued with the wonder of life.", + "seed": 11233, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000050/group_9/index_3898_seed_11233_dp_19_rank_38_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 9, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.10656551322899759, + "vlm_logprob": -0.10656551322899759, + "vlm_yes_no_margin": 3.85, + "vlm_token_credit": -0.10656551322899759, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot framed as a horizontal medium shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera pan smoothly to the right while tracking a massive kangaroo?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the kangaroo shown leaping across a vast expanse of water within a deep, dark cavern?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene illuminated by a swarm of bioluminescent fireflies emitting a steady blue glow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the blue glow from the fireflies reflected on the still, mirror-like black water below?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.13000157475471497 + ], + "opposite_logprob_tokens": [ + -2.5050015449523926 + ], + "logprob_reward": -0.13000157475471497, + "yes_no_margin_reward": 2.375, + "token_credit_reward": -0.13000157475471497, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.17776767909526825, + 0.04817154258489609, + 0.05612177774310112, + 0.07210449129343033, + 0.08806298673152924, + 0.2545197606086731, + 0.0369190089404583, + 0.11204246431589127, + 0.09635107219219208, + 0.0579393245279789 + ], + "credit_map": [ + 0.019037801772356033, + 0.025715965777635574, + 0.01718507893383503, + 0.019999435171484947, + 0.028016507625579834, + 0.02043462172150612, + 0.021679965779185295, + 0.025698287412524223, + 0.00547331478446722, + 0.004798378329724073, + 0.007143209222704172, + 0.007695440202951431, + 0.0050437236204743385, + 0.0065205893479287624, + 0.005077616311609745, + 0.006419263314455748, + 0.005007671657949686, + 0.005105706863105297, + 0.009381044656038284, + 0.00773657625541091, + 0.004334065597504377, + 0.007626340724527836, + 0.007464561611413956, + 0.009465807117521763, + 0.006417184602469206, + 0.006362609099596739, + 0.011894416995346546, + 0.009672467596828938, + 0.006924968212842941, + 0.00874448660761118, + 0.010069288313388824, + 0.012019063346087933, + 0.00756771769374609, + 0.006738822441548109, + 0.01038826908916235, + 0.013363495469093323, + 0.014325967989861965, + 0.009203125722706318, + 0.010831247083842754, + 0.01564433053135872, + 0.037310708314180374, + 0.018380843102931976, + 0.03024369291961193, + 0.024007460102438927, + 0.02539719268679619, + 0.08786497265100479, + 0.012067132629454136, + 0.01924772933125496, + 0.004157538991421461, + 0.003269577631726861, + 0.0057856133207678795, + 0.005583884660154581, + 0.004992083180695772, + 0.0041869995184242725, + 0.004206289071589708, + 0.004737021401524544, + 0.010645966976881027, + 0.005341392010450363, + 0.036992397159338, + 0.020949624478816986, + 0.004685393534600735, + 0.016958141699433327, + 0.0071381921879947186, + 0.009331345558166504, + 0.008402300998568535, + 0.016629250720143318, + 0.01203130092471838, + 0.00906362198293209, + 0.01812797412276268, + 0.012772775255143642, + 0.011168941855430603, + 0.00815490074455738, + 0.005882161669433117, + 0.008018458262085915, + 0.007424991112202406, + 0.005305481143295765, + 0.009933426976203918, + 0.008418967947363853, + 0.004218440502882004, + 0.00873738992959261 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 0, + 7 + ], + "selected_reward": -0.13000157475471497 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot framed as a horizontal medium shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera pan smoothly to the right while tracking a massive kangaroo?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the kangaroo shown leaping across a vast expanse of water within a deep, dark cavern?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene illuminated by a swarm of bioluminescent fireflies emitting a steady blue glow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the blue glow from the fireflies reflected on the still, mirror-like black water below?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.38706734776496887 + ], + "opposite_logprob_tokens": [ + -1.1370673179626465 + ], + "logprob_reward": -0.38706734776496887, + "yes_no_margin_reward": 0.75, + "token_credit_reward": -0.38706734776496887, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10659117251634598, + 0.06112278997898102, + 0.0965057909488678, + 0.22122670710086823, + 0.051717471331357956, + 0.04890928789973259, + 0.14594008028507233, + 0.053581152111291885, + 0.14171408116817474, + 0.0726914256811142 + ], + "credit_map": [ + 0.011830466799438, + 0.0134012121707201, + 0.014278429560363293, + 0.01348007470369339, + 0.01567698083817959, + 0.013357640244066715, + 0.012573372572660446, + 0.011992992833256721, + 0.0064131892286241055, + 0.0072751534171402454, + 0.008113627322018147, + 0.008341704495251179, + 0.009123028255999088, + 0.00829703826457262, + 0.008041299879550934, + 0.0055177477188408375, + 0.009800625033676624, + 0.010807912796735764, + 0.011865010485053062, + 0.012475990690290928, + 0.012967227026820183, + 0.012834529392421246, + 0.0148209473118186, + 0.010933547280728817, + 0.02010028064250946, + 0.03640739992260933, + 0.01974295824766159, + 0.03060155361890793, + 0.034922316670417786, + 0.02394450455904007, + 0.03265159949660301, + 0.022856105118989944, + 0.005743833724409342, + 0.004741187207400799, + 0.00533412117511034, + 0.010852411389350891, + 0.007828408852219582, + 0.005983857903629541, + 0.005014914553612471, + 0.006218737922608852, + 0.006504192482680082, + 0.006780838128179312, + 0.005811677314341068, + 0.007402935065329075, + 0.00611648615449667, + 0.00708062993362546, + 0.004746687598526478, + 0.0044658370316028595, + 0.024874133989214897, + 0.023154929280281067, + 0.015034684911370277, + 0.018438635393977165, + 0.021998606622219086, + 0.01717550866305828, + 0.014422793872654438, + 0.010840789414942265, + 0.004225483629852533, + 0.004667897708714008, + 0.012555120512843132, + 0.008345398120582104, + 0.005677357316017151, + 0.008230216801166534, + 0.005097583867609501, + 0.004782091826200485, + 0.010423406027257442, + 0.020442064851522446, + 0.023532606661319733, + 0.017246050760149956, + 0.026281725615262985, + 0.016373464837670326, + 0.014537702314555645, + 0.012877064757049084, + 0.007405843585729599, + 0.010563992895185947, + 0.009755449369549751, + 0.006606653332710266, + 0.012269223108887672, + 0.009613224305212498, + 0.009980957955121994, + 0.006496083922684193 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 6, + 8 + ], + "selected_reward": -0.38706734776496887 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot framed as a horizontal medium shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera pan smoothly to the right while tracking a massive kangaroo?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the kangaroo shown leaping across a vast expanse of water within a deep, dark cavern?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene illuminated by a swarm of bioluminescent fireflies emitting a steady blue glow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the blue glow from the fireflies reflected on the still, mirror-like black water below?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004639219027012587 + ], + "opposite_logprob_tokens": [ + -5.379639148712158 + ], + "logprob_reward": -0.004639219027012587, + "yes_no_margin_reward": 5.375, + "token_credit_reward": -0.004639219027012587, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.21774083375930786, + 0.05801228806376457, + 0.10198049247264862, + 0.07975658029317856, + 0.06465791910886765, + 0.2332884818315506, + 0.07642394304275513, + 0.0395829901099205, + 0.040333762764930725, + 0.08822272717952728 + ], + "credit_map": [ + 0.020974522456526756, + 0.017340712249279022, + 0.02906915731728077, + 0.027121467515826225, + 0.022318357601761818, + 0.024861421436071396, + 0.03949118033051491, + 0.03656398504972458, + 0.0059127709828317165, + 0.005611106287688017, + 0.008573965169489384, + 0.008131160400807858, + 0.00697555486112833, + 0.0073545859195292, + 0.009374595247209072, + 0.006078544072806835, + 0.008527960628271103, + 0.008754056878387928, + 0.013323293067514896, + 0.013810982927680016, + 0.008737192489206791, + 0.012532321736216545, + 0.024177219718694687, + 0.012117454782128334, + 0.005783767439424992, + 0.006480616517364979, + 0.011803277768194675, + 0.01636965572834015, + 0.007220005616545677, + 0.009936378337442875, + 0.01228928379714489, + 0.009873585775494576, + 0.006179658696055412, + 0.003314863657578826, + 0.012842511758208275, + 0.010614296421408653, + 0.006411454640328884, + 0.00791558064520359, + 0.010113269090652466, + 0.007266276981681585, + 0.03970329836010933, + 0.009707896038889885, + 0.02643321454524994, + 0.016745079308748245, + 0.022909581661224365, + 0.09225477278232574, + 0.008307280950248241, + 0.017227336764335632, + 0.011552211828529835, + 0.010573144070804119, + 0.006804874166846275, + 0.01031836960464716, + 0.012791553512215614, + 0.01298158522695303, + 0.005979891866445541, + 0.005422305781394243, + 0.003940621856600046, + 0.004663494881242514, + 0.006257584318518639, + 0.0058467574417591095, + 0.004803539719432592, + 0.006585738621652126, + 0.0033585731871426105, + 0.004126677755266428, + 0.004312752280384302, + 0.004776023793965578, + 0.006251125130802393, + 0.005139546934515238, + 0.00729177612811327, + 0.004850208293646574, + 0.0029574923682957888, + 0.0047548385336995125, + 0.006557032465934753, + 0.009255749173462391, + 0.015995731577277184, + 0.008445538580417633, + 0.017340071499347687, + 0.01305522583425045, + 0.008477166295051575, + 0.009096206165850163 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 0, + 2 + ], + "selected_reward": -0.004639219027012587 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot framed as a horizontal medium shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera pan smoothly to the right while tracking a massive kangaroo?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the kangaroo shown leaping across a vast expanse of water within a deep, dark cavern?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene illuminated by a swarm of bioluminescent fireflies emitting a steady blue glow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the blue glow from the fireflies reflected on the still, mirror-like black water below?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.008634140715003014 + ], + "opposite_logprob_tokens": [ + -4.758634090423584 + ], + "logprob_reward": -0.008634140715003014, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.008634140715003014, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11166641861200333, + 0.0437227301299572, + 0.10896138101816177, + 0.07648298144340515, + 0.2130056470632553, + 0.19010542333126068, + 0.06134367361664772, + 0.09142570197582245, + 0.04030848667025566, + 0.06297750771045685 + ], + "credit_map": [ + 0.013291372917592525, + 0.009157371707260609, + 0.011848454363644123, + 0.015081916004419327, + 0.009587772190570831, + 0.013253024779260159, + 0.01919412426650524, + 0.020252380520105362, + 0.004340683110058308, + 0.00276330579072237, + 0.008492536842823029, + 0.007503253873437643, + 0.0032457879278808832, + 0.005594457034021616, + 0.006441464647650719, + 0.00534124206751585, + 0.009602312929928303, + 0.011813272722065449, + 0.013153093867003918, + 0.013957281596958637, + 0.013317866250872612, + 0.014371948316693306, + 0.02210240066051483, + 0.010643203742802143, + 0.0071694557555019855, + 0.010156144388020039, + 0.012860357761383057, + 0.009877854026854038, + 0.007943516597151756, + 0.00885713193565607, + 0.010078399442136288, + 0.009540123865008354, + 0.01812969706952572, + 0.01036680769175291, + 0.0309851486235857, + 0.056190233677625656, + 0.014199520461261272, + 0.03542398288846016, + 0.023112962022423744, + 0.024597292765975, + 0.0299796462059021, + 0.020707041025161743, + 0.023712895810604095, + 0.025604823604226112, + 0.017304839566349983, + 0.04744917154312134, + 0.009590592235326767, + 0.015756405889987946, + 0.009400238282978535, + 0.007469726260751486, + 0.009820307604968548, + 0.007251991890370846, + 0.005701449234038591, + 0.006629105657339096, + 0.0066664558835327625, + 0.00840439647436142, + 0.007707853335887194, + 0.004475811030715704, + 0.0292736254632473, + 0.018220635131001472, + 0.00463617080822587, + 0.0131620978936553, + 0.006159594748169184, + 0.007789914496243, + 0.003920709248632193, + 0.004602884873747826, + 0.008232604712247849, + 0.0058643207885324955, + 0.004187671933323145, + 0.005287215579301119, + 0.002845057751983404, + 0.005368021782487631, + 0.004760689567774534, + 0.004776580259203911, + 0.009715874679386616, + 0.0076814028434455395, + 0.012561745010316372, + 0.009079673327505589, + 0.00483695650473237, + 0.00956458505243063 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 5, + 0 + ], + "selected_reward": -0.008634140715003014 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot framed as a horizontal medium shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera pan smoothly to the right while tracking a massive kangaroo?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the kangaroo shown leaping across a vast expanse of water within a deep, dark cavern?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene illuminated by a swarm of bioluminescent fireflies emitting a steady blue glow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the blue glow from the fireflies reflected on the still, mirror-like black water below?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0024852838832885027 + ], + "opposite_logprob_tokens": [ + -6.002485275268555 + ], + "logprob_reward": -0.0024852838832885027, + "yes_no_margin_reward": 6.0, + "token_credit_reward": -0.0024852838832885027, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.25136691331863403, + 0.03041708841919899, + 0.08514007180929184, + 0.05223413556814194, + 0.05299101024866104, + 0.265328586101532, + 0.036475032567977905, + 0.1168719008564949, + 0.05267041549086571, + 0.05650489032268524 + ], + "credit_map": [ + 0.03308707848191261, + 0.020067719742655754, + 0.04138579219579697, + 0.03400491923093796, + 0.01760525442659855, + 0.029151374474167824, + 0.03357204422354698, + 0.04249272122979164, + 0.002764536999166012, + 0.0025303964503109455, + 0.005794628988951445, + 0.004517924506217241, + 0.0027356070932000875, + 0.004151959903538227, + 0.0030061479192227125, + 0.004915882367640734, + 0.007475982420146465, + 0.005646779201924801, + 0.016007620841264725, + 0.013841521926224232, + 0.006211171392351389, + 0.011727829463779926, + 0.015031352639198303, + 0.009197811596095562, + 0.005057807080447674, + 0.003708570497110486, + 0.009005718864500523, + 0.00713132880628109, + 0.005391326267272234, + 0.0062765032052993774, + 0.0074550919234752655, + 0.008207782171666622, + 0.0052155726589262486, + 0.002714063273742795, + 0.008315692655742168, + 0.009476288221776485, + 0.004210758488625288, + 0.006018850486725569, + 0.007072382140904665, + 0.009967396967113018, + 0.04471998289227486, + 0.013044928200542927, + 0.029054736718535423, + 0.021128956228494644, + 0.027934081852436066, + 0.09694644063711166, + 0.012932544574141502, + 0.019566882401704788, + 0.004775526002049446, + 0.0047766477800905704, + 0.005068614147603512, + 0.005157896783202887, + 0.003943095449358225, + 0.0036125527694821358, + 0.004528225399553776, + 0.0046124691143631935, + 0.009659113362431526, + 0.006067083217203617, + 0.03728410229086876, + 0.02179323509335518, + 0.008501390926539898, + 0.015653617680072784, + 0.007282288745045662, + 0.010631052777171135, + 0.004896750673651695, + 0.005889311898499727, + 0.00981587078422308, + 0.007046655751764774, + 0.008797143585979939, + 0.006590499542653561, + 0.004270830191671848, + 0.005363349337130785, + 0.004097474738955498, + 0.003925169352442026, + 0.010627429001033306, + 0.0058304076083004475, + 0.004691633395850658, + 0.00901609193533659, + 0.0042688725516200066, + 0.014047807082533836 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 0, + 7 + ], + "selected_reward": -0.0024852838832885027 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "horizontal medium shot, panning smoothly to the right, tracks a massive kangaroo at the apex of a powerful leap, gracefully crossing a vast expanse of water. Nestled within a deep, dark cavern, the kangaroo's muscular body stretches to its full extent, from the tip of its long, thick tail, acting as a stabilizer, to its protruding nose. Its dense, rich, reddish-brown fur is softly outlined in the dim light. The scene's sole source of light comes from a dense, silent, floating swarm of bioluminescent fireflies, emitting a steady, ethereal blue glow that imbues the darkness with a mystical hue. This magical glow shimmers upon the surrounding damp, craggy rock faces and, more strikingly, is perfectly reflected upon the still, mirror-like black water below, creating a breathtaking symmetrical illusion that perfectly replicates the kangaroo's arched silhouette and its thousands of twinkling blue points of light, evoking a tranquil, dreamlike atmosphere imbued with the wonder of life.", + "seed": 11234, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000050/group_9/index_3898_seed_11234_dp_19_rank_38_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 9, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.05000725355930626, + "vlm_logprob": -0.05000725355930626, + "vlm_yes_no_margin": 3.950000023841858, + "vlm_token_credit": -0.05000725355930626, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot framed as a horizontal medium shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera pan smoothly to the right while tracking a massive kangaroo?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the kangaroo shown leaping across a vast expanse of water within a deep, dark cavern?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene illuminated by a swarm of bioluminescent fireflies emitting a steady blue glow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the blue glow from the fireflies reflected on the still, mirror-like black water below?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.08470369875431061 + ], + "opposite_logprob_tokens": [ + -2.8347036838531494 + ], + "logprob_reward": -0.08470369875431061, + "yes_no_margin_reward": 2.75, + "token_credit_reward": -0.08470369875431061, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08799528330564499, + 0.024250853806734085, + 0.01600797101855278, + 0.05647728964686394, + 0.03608961030840874, + 0.0199667327105999, + 0.011264190077781677, + 0.03215403109788895, + 0.683054506778717, + 0.03273956850171089 + ], + "credit_map": [ + 0.006888372357934713, + 0.017635047435760498, + 0.0048154196701943874, + 0.006050032563507557, + 0.024568963795900345, + 0.015316519886255264, + 0.00735654029995203, + 0.005364381708204746, + 0.002219222020357847, + 0.0031680818647146225, + 0.002768418751657009, + 0.005634852219372988, + 0.002650127513334155, + 0.002606512513011694, + 0.002911381656304002, + 0.002292253077030182, + 0.002237872686237097, + 0.002212861552834511, + 0.002560387598350644, + 0.0009819346014410257, + 0.001996005652472377, + 0.002692891750484705, + 0.001703427406027913, + 0.0016225868603214622, + 0.009766777977347374, + 0.005310642067342997, + 0.011223641224205494, + 0.003574602771550417, + 0.006470089312642813, + 0.011166566051542759, + 0.004673043731600046, + 0.004291919991374016, + 0.0031693684868514538, + 0.004007777664810419, + 0.00348894065245986, + 0.0044075315818190575, + 0.007005245424807072, + 0.0056060487404465675, + 0.003540817415341735, + 0.0048638745211064816, + 0.002722769742831588, + 0.0019902854692190886, + 0.0019634785130620003, + 0.0018178860191255808, + 0.0034230120945721865, + 0.0025897487066686153, + 0.0024427915923297405, + 0.0030167591758072376, + 0.0021378009114414454, + 0.0009115926222875714, + 0.0014181106816977262, + 0.0011269745882600546, + 0.0016685063019394875, + 0.0018128686351701617, + 0.000916513439733535, + 0.0012718227226287127, + 0.003054836532101035, + 0.0055509344674646854, + 0.002424814971163869, + 0.0030227249953895807, + 0.00657243374735117, + 0.004160677548497915, + 0.003471356350928545, + 0.0038962510880082846, + 0.10087830573320389, + 0.08657922595739365, + 0.04212544485926628, + 0.03940126299858093, + 0.16154831647872925, + 0.13478600978851318, + 0.03165328875184059, + 0.0860825851559639, + 0.005427584983408451, + 0.003137403167784214, + 0.0030428932514041662, + 0.002141624456271529, + 0.008716817945241928, + 0.002985157538205385, + 0.002770204795524478, + 0.00451787980273366 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 0, + 3 + ], + "selected_reward": -0.08470369875431061 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot framed as a horizontal medium shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera pan smoothly to the right while tracking a massive kangaroo?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the kangaroo shown leaping across a vast expanse of water within a deep, dark cavern?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene illuminated by a swarm of bioluminescent fireflies emitting a steady blue glow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the blue glow from the fireflies reflected on the still, mirror-like black water below?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.14281047880649567 + ], + "opposite_logprob_tokens": [ + -2.017810583114624 + ], + "logprob_reward": -0.14281047880649567, + "yes_no_margin_reward": 1.8750001192092896, + "token_credit_reward": -0.14281047880649567, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06694157421588898, + 0.13034315407276154, + 0.05035722628235817, + 0.07863105088472366, + 0.04458249732851982, + 0.04518628120422363, + 0.027762223035097122, + 0.04265367239713669, + 0.43678727746009827, + 0.07675499469041824 + ], + "credit_map": [ + 0.006420921999961138, + 0.011309638619422913, + 0.005565972533077002, + 0.006151485722512007, + 0.010237167589366436, + 0.012220875360071659, + 0.007947694510221481, + 0.00708782346919179, + 0.010264556854963303, + 0.02155115269124508, + 0.011747376061975956, + 0.033883653581142426, + 0.021868767216801643, + 0.011287514120340347, + 0.01126309297978878, + 0.008477050811052322, + 0.008910950273275375, + 0.0031222065445035696, + 0.004450119566172361, + 0.004181098658591509, + 0.008318385109305382, + 0.011384828016161919, + 0.0062319496646523476, + 0.0037576896138489246, + 0.011947491206228733, + 0.00843691173940897, + 0.012938466854393482, + 0.0075107766315341, + 0.010081618092954159, + 0.014564495533704758, + 0.0073563517071306705, + 0.005794942378997803, + 0.006005711853504181, + 0.004294307436794043, + 0.006654982455074787, + 0.0048600840382277966, + 0.004356407094746828, + 0.008882623165845871, + 0.004446335136890411, + 0.005082045681774616, + 0.0062529961578547955, + 0.005555319134145975, + 0.0037774047814309597, + 0.0036533577367663383, + 0.009147708304226398, + 0.005827116779983044, + 0.004843090195208788, + 0.00612928532063961, + 0.005250775720924139, + 0.0024263274390250444, + 0.003148970426991582, + 0.0023299874737858772, + 0.0042486186139285564, + 0.004815421532839537, + 0.0016434428980574012, + 0.0038986781146377325, + 0.0049944110214710236, + 0.006920944433659315, + 0.0035542305558919907, + 0.004676613956689835, + 0.007020099554210901, + 0.005649706814438105, + 0.0052140941843390465, + 0.004623569082468748, + 0.05771683529019356, + 0.0749296024441719, + 0.02799033559858799, + 0.0267278291285038, + 0.11279810220003128, + 0.07121717929840088, + 0.024164404720067978, + 0.04124302789568901, + 0.022487524896860123, + 0.005195897072553635, + 0.006543864496052265, + 0.004144327715039253, + 0.024702640250325203, + 0.004954347386956215, + 0.004366319626569748, + 0.0043600741773843765 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 1, + 3 + ], + "selected_reward": -0.14281047880649567 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot framed as a horizontal medium shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera pan smoothly to the right while tracking a massive kangaroo?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the kangaroo shown leaping across a vast expanse of water within a deep, dark cavern?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene illuminated by a swarm of bioluminescent fireflies emitting a steady blue glow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the blue glow from the fireflies reflected on the still, mirror-like black water below?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.011072501540184021 + ], + "opposite_logprob_tokens": [ + -4.511072635650635 + ], + "logprob_reward": -0.011072501540184021, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.011072501540184021, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11290876567363739, + 0.2156757265329361, + 0.03258773311972618, + 0.06888432800769806, + 0.08189356327056885, + 0.04658734053373337, + 0.02666456624865532, + 0.022180484607815742, + 0.3503754734992981, + 0.04224199056625366 + ], + "credit_map": [ + 0.013809562660753727, + 0.013666299171745777, + 0.012568820267915726, + 0.011369719170033932, + 0.016404472291469574, + 0.01707005500793457, + 0.012493906542658806, + 0.015525934286415577, + 0.017665911465883255, + 0.03746378794312477, + 0.01746002957224846, + 0.04906097799539566, + 0.03701261803507805, + 0.022076217457652092, + 0.02128681167960167, + 0.01364938449114561, + 0.004239177331328392, + 0.0031710630282759666, + 0.005050983279943466, + 0.0029198434203863144, + 0.004071944858878851, + 0.005270914174616337, + 0.0032138247042894363, + 0.00464998185634613, + 0.015120931901037693, + 0.005299276672303677, + 0.016599271446466446, + 0.0037939660251140594, + 0.009059639647603035, + 0.009559777565300465, + 0.005399071611464024, + 0.004052388947457075, + 0.00991769041866064, + 0.0095142163336277, + 0.008414807729423046, + 0.008171811699867249, + 0.010695824399590492, + 0.0162220299243927, + 0.007622349075973034, + 0.011334828101098537, + 0.00605821143835783, + 0.0022838679142296314, + 0.004132006783038378, + 0.004323601722717285, + 0.003222407540306449, + 0.007212124299257994, + 0.004411389585584402, + 0.014943732880055904, + 0.003250992391258478, + 0.0021840091794729233, + 0.0031711554620414972, + 0.0023659116122871637, + 0.0024289374705404043, + 0.004539524670690298, + 0.0022161821834743023, + 0.006507852114737034, + 0.003111759200692177, + 0.002352063776925206, + 0.0026111239567399025, + 0.0028042178601026535, + 0.0029669825453311205, + 0.003272139700129628, + 0.002527443692088127, + 0.0025347552727907896, + 0.0466332882642746, + 0.05079816281795502, + 0.02093719132244587, + 0.02015126496553421, + 0.08814389258623123, + 0.0671125203371048, + 0.017762841656804085, + 0.038836315274238586, + 0.006006653420627117, + 0.00381780625320971, + 0.0074754985980689526, + 0.004849959630519152, + 0.005632178857922554, + 0.004198897164314985, + 0.003211155766621232, + 0.007049846928566694 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 1, + 0 + ], + "selected_reward": -0.011072501540184021 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot framed as a horizontal medium shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera pan smoothly to the right while tracking a massive kangaroo?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the kangaroo shown leaping across a vast expanse of water within a deep, dark cavern?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene illuminated by a swarm of bioluminescent fireflies emitting a steady blue glow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the blue glow from the fireflies reflected on the still, mirror-like black water below?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.00863473117351532 + ], + "opposite_logprob_tokens": [ + -4.758634567260742 + ], + "logprob_reward": -0.00863473117351532, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.00863473117351532, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.05449355021119118, + 0.44265997409820557, + 0.05969450995326042, + 0.046161457896232605, + 0.031053220853209496, + 0.06409143656492233, + 0.012437034398317337, + 0.12719480693340302, + 0.13879486918449402, + 0.023419106379151344 + ], + "credit_map": [ + 0.0071014841087162495, + 0.006108063738793135, + 0.004998897202312946, + 0.00580068351700902, + 0.00858383346349001, + 0.008106008172035217, + 0.007387794088572264, + 0.006406781729310751, + 0.03737640008330345, + 0.07611300051212311, + 0.036988284438848495, + 0.10450701415538788, + 0.07755620032548904, + 0.043867021799087524, + 0.04024813324213028, + 0.026003938168287277, + 0.011229416355490685, + 0.004691838286817074, + 0.005875967442989349, + 0.004960729740560055, + 0.008655888959765434, + 0.013269216753542423, + 0.006457850337028503, + 0.004553604405373335, + 0.0052514029666781425, + 0.004642785526812077, + 0.0070986757054924965, + 0.004981693811714649, + 0.005668845493346453, + 0.008654654026031494, + 0.004720059223473072, + 0.005143340677022934, + 0.005212012212723494, + 0.002758860820904374, + 0.004353623371571302, + 0.0028522175271064043, + 0.0032878825441002846, + 0.006263907067477703, + 0.0025819740258157253, + 0.003742741420865059, + 0.007054036948829889, + 0.0029108827002346516, + 0.00855210144072771, + 0.006884798873215914, + 0.004622187465429306, + 0.00972233060747385, + 0.006894468795508146, + 0.01745063066482544, + 0.0017447546124458313, + 0.001785681233741343, + 0.0012431801296770573, + 0.0012657020706683397, + 0.0017988497857004404, + 0.0017551557393744588, + 0.000851254619192332, + 0.0019924563821405172, + 0.013305097818374634, + 0.023004524409770966, + 0.008986327797174454, + 0.011361916549503803, + 0.025894399732351303, + 0.01608566753566265, + 0.013172554783523083, + 0.015384329482913017, + 0.020805396139621735, + 0.01755080558359623, + 0.008098239079117775, + 0.0077294353395700455, + 0.03480289876461029, + 0.02760348469018936, + 0.006279028952121735, + 0.015925586223602295, + 0.005665608681738377, + 0.0023882435634732246, + 0.002702932571992278, + 0.0010624434798955917, + 0.00693229865282774, + 0.0018266354454681277, + 0.0016800486482679844, + 0.001160894869826734 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 8, + 7 + ], + "selected_reward": -0.00863473117351532 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot framed as a horizontal medium shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera pan smoothly to the right while tracking a massive kangaroo?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the kangaroo shown leaping across a vast expanse of water within a deep, dark cavern?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene illuminated by a swarm of bioluminescent fireflies emitting a steady blue glow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the blue glow from the fireflies reflected on the still, mirror-like black water below?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0028148575220257044 + ], + "opposite_logprob_tokens": [ + -5.877814769744873 + ], + "logprob_reward": -0.0028148575220257044, + "yes_no_margin_reward": 5.875, + "token_credit_reward": -0.0028148575220257044, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14065611362457275, + 0.182890847325325, + 0.05528860166668892, + 0.1010693833231926, + 0.04784470424056053, + 0.07244472205638885, + 0.02540714479982853, + 0.0884697288274765, + 0.1984274685382843, + 0.08750129491090775 + ], + "credit_map": [ + 0.013850375078618526, + 0.029923660680651665, + 0.010403227061033249, + 0.01093768235296011, + 0.011487062089145184, + 0.036182478070259094, + 0.012386093847453594, + 0.01548553816974163, + 0.01566488668322563, + 0.03085171803832054, + 0.016038337722420692, + 0.044348087161779404, + 0.031737662851810455, + 0.01773039437830448, + 0.015766631811857224, + 0.01075312402099371, + 0.00827064085751772, + 0.006492414977401495, + 0.007650499232113361, + 0.00506138289347291, + 0.006721791811287403, + 0.009078415110707283, + 0.006354764103889465, + 0.0056586903519928455, + 0.013588561676442623, + 0.01030641421675682, + 0.018711376935243607, + 0.008215050213038921, + 0.012164667248725891, + 0.017508836463093758, + 0.008709781803190708, + 0.011864684522151947, + 0.009219817817211151, + 0.0034930126275867224, + 0.006596188060939312, + 0.004839818924665451, + 0.004566484596580267, + 0.009760870598256588, + 0.004252987448126078, + 0.00511552207171917, + 0.007548531051725149, + 0.004334666300565004, + 0.009617733769118786, + 0.007701570633798838, + 0.005549693014472723, + 0.01074425783008337, + 0.010393129661679268, + 0.01655513606965542, + 0.002443868201225996, + 0.0021411199122667313, + 0.0038341740146279335, + 0.0022376440465450287, + 0.0024721792433410883, + 0.00434248149394989, + 0.0021202010102570057, + 0.005815474782139063, + 0.007894244976341724, + 0.016408659517765045, + 0.0064359321258962154, + 0.007937672547996044, + 0.016589459031820297, + 0.011975826695561409, + 0.010578185319900513, + 0.010649742558598518, + 0.017357612028717995, + 0.027490276843309402, + 0.013006512075662613, + 0.012049118988215923, + 0.04556405171751976, + 0.048469651490449905, + 0.009740188717842102, + 0.024750055745244026, + 0.021204184740781784, + 0.008725839667022228, + 0.012849605642259121, + 0.0045182062312960625, + 0.023249858990311623, + 0.0059388079680502415, + 0.006379980128258467, + 0.004634794779121876 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 1, + 0 + ], + "selected_reward": -0.0028148575220257044 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000071/group_0/index_tensor([5902, 5902])_group_0_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000071/group_0/index_tensor([5902, 5902])_group_0_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..ba498717451aa5981f79c1781929f373c40ea0db --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000071/group_0/index_tensor([5902, 5902])_group_0_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "The camera opens with a wide, long shot, then slowly and steadily pans forward, capturing a tranquil, pastoral scene from a horizontal angle. As golden morning light slants across the horizon, an elderly villager stands alone in the center of a newly green rice paddy. Her graying hair is pulled back in a simple bun, and her face is etched with deep wrinkles from time, yet her expression is one of unwavering tranquility and contentment. She wears a faded blue homespun blouse and loose black trousers, her cloth shoes stained with morning mud. As the camera closes in, her movements begin: she slowly raises her arms toward the sky, as if embracing the first rays of morning sunlight, her fingertips trembling slightly with the stretch. Then, her body begins a gentle twirl, a simple movement, more like an impromptu expression of unity with nature, the hem of her cloth dress fluttering lightly as she turns. The soft sidelight casts a warm golden halo around her silhouette, illuminating the dewdrops in the field with a shimmering clarity. In the distance are rolling green mountains with a light morning mist lingering between the mountains. The whole picture is dominated by the tender green of the rice fields and the light blue of the sky, presenting an atmosphere that is warm, peaceful and full of hope for life.", + "seed": 4454, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000071/group_0/index_5902_seed_4454_dp_0_rank_0_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 0, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.11805116897448897, + "vlm_logprob": -0.11805116897448897, + "vlm_yes_no_margin": 2.625, + "vlm_token_credit": -0.11805116897448897, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an elderly villager with graying hair in a bun and deep wrinkles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the setting a newly green rice paddy with rolling green mountains in the distance?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera start with a wide long shot and slowly pan forward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the villager raise her arms toward the sky and then perform a gentle twirl?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the lighting golden morning light casting a warm halo and illuminating dewdrops?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.14410734176635742 + ], + "opposite_logprob_tokens": [ + -2.0191073417663574 + ], + "logprob_reward": -0.14410734176635742, + "yes_no_margin_reward": 1.875, + "token_credit_reward": -0.14410734176635742, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2976359724998474, + 0.03164994716644287, + 0.045782145112752914, + 0.024326950311660767, + 0.08106951415538788, + 0.10389197617769241, + 0.06779119372367859, + 0.05627701058983803, + 0.18297119438648224, + 0.1086040735244751 + ], + "credit_map": [ + 0.03628529608249664, + 0.033567383885383606, + 0.05232172831892967, + 0.025183845311403275, + 0.055921830236911774, + 0.03668494522571564, + 0.031015390530228615, + 0.02665555663406849, + 0.003951952792704105, + 0.00398225849494338, + 0.0029055513441562653, + 0.003317265771329403, + 0.0066925110295414925, + 0.004003433044999838, + 0.0036025417502969503, + 0.0031944303773343563, + 0.006102882791310549, + 0.004894379060715437, + 0.006621201056987047, + 0.005147934891283512, + 0.008661925792694092, + 0.00463356077671051, + 0.006559181958436966, + 0.0031610792502760887, + 0.0024247057735919952, + 0.0025633673649281263, + 0.005083794239908457, + 0.00369048397988081, + 0.0029186608735471964, + 0.002027772134169936, + 0.003144843503832817, + 0.0024733226746320724, + 0.015731969848275185, + 0.00839389767497778, + 0.005710234399884939, + 0.005671622697263956, + 0.00969952903687954, + 0.017399288713932037, + 0.012728121131658554, + 0.005734849721193314, + 0.013847052119672298, + 0.011102192103862762, + 0.01022670604288578, + 0.013296102173626423, + 0.01327574159950018, + 0.01582217402756214, + 0.016319170594215393, + 0.010002843104302883, + 0.009368564002215862, + 0.006300282198935747, + 0.01012286078184843, + 0.007991963066160679, + 0.006979453377425671, + 0.010174563154578209, + 0.00968178827315569, + 0.007171717006713152, + 0.006803401745855808, + 0.007160406094044447, + 0.003856792114675045, + 0.007268095389008522, + 0.008799892850220203, + 0.008444089442491531, + 0.00873864907771349, + 0.005205683875828981, + 0.024919340386986732, + 0.02141069807112217, + 0.01603003591299057, + 0.017801174893975258, + 0.039806440472602844, + 0.017584912478923798, + 0.028082292526960373, + 0.017336303368210793, + 0.015548128634691238, + 0.01286627259105444, + 0.014569859951734543, + 0.009345036000013351, + 0.019932089373469353, + 0.010945736430585384, + 0.014965319074690342, + 0.010431632399559021 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 9 + ], + "selected_reward": -0.14410734176635742 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an elderly villager with graying hair in a bun and deep wrinkles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the setting a newly green rice paddy with rolling green mountains in the distance?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera start with a wide long shot and slowly pan forward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the villager raise her arms toward the sky and then perform a gentle twirl?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the lighting golden morning light casting a warm halo and illuminating dewdrops?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004652034025639296 + ], + "opposite_logprob_tokens": [ + -5.37965202331543 + ], + "logprob_reward": -0.004652034025639296, + "yes_no_margin_reward": 5.375, + "token_credit_reward": -0.004652034025639296, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.29195815324783325, + 0.0701017826795578, + 0.051258474588394165, + 0.08872292190790176, + 0.05267738550901413, + 0.02649291604757309, + 0.16508348286151886, + 0.054170239716768265, + 0.1227557510137558, + 0.0767788216471672 + ], + "credit_map": [ + 0.03930666670203209, + 0.043210480362176895, + 0.037300702184438705, + 0.031657297164201736, + 0.050181444734334946, + 0.03727748990058899, + 0.029803529381752014, + 0.02322058379650116, + 0.009412682615220547, + 0.009228445589542389, + 0.008084854111075401, + 0.006929012481123209, + 0.011477547697722912, + 0.009237664751708508, + 0.008195312693715096, + 0.007536271121352911, + 0.008258549496531487, + 0.0068413252010941505, + 0.007624906487762928, + 0.005556226242333651, + 0.00642731599509716, + 0.004986317362636328, + 0.005852526053786278, + 0.0057113091461360455, + 0.012467443011701107, + 0.011795774102210999, + 0.01724163629114628, + 0.014117410406470299, + 0.006189829669892788, + 0.011340499855577946, + 0.006418157368898392, + 0.009152185171842575, + 0.006266723852604628, + 0.0072193327359855175, + 0.008991370908915997, + 0.008919182233512402, + 0.004937719088047743, + 0.005148549564182758, + 0.00655559916049242, + 0.0046389102935791016, + 0.004076335579156876, + 0.0043688672594726086, + 0.0031618948560208082, + 0.0029603117145597935, + 0.003304848913103342, + 0.0030729705467820168, + 0.0030100170988589525, + 0.002537672407925129, + 0.03738286346197128, + 0.01630125194787979, + 0.028623567894101143, + 0.014681101776659489, + 0.007672678679227829, + 0.036132920533418655, + 0.012749665416777134, + 0.011539439670741558, + 0.0076044234447181225, + 0.008984008803963661, + 0.0036030770279467106, + 0.00462000910192728, + 0.010882832109928131, + 0.00732563016936183, + 0.006632304284721613, + 0.004517958965152502, + 0.01726626418530941, + 0.015845853835344315, + 0.01443827711045742, + 0.014420673251152039, + 0.019901346415281296, + 0.009765781462192535, + 0.019162282347679138, + 0.011955288238823414, + 0.009776586666703224, + 0.010824394412338734, + 0.02088182233273983, + 0.011852500960230827, + 0.0066050575114786625, + 0.004565324168652296, + 0.005895272828638554, + 0.0063778674229979515 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 6, + 8 + ], + "selected_reward": -0.004652034025639296 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an elderly villager with graying hair in a bun and deep wrinkles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the setting a newly green rice paddy with rolling green mountains in the distance?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera start with a wide long shot and slowly pan forward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the villager raise her arms toward the sky and then perform a gentle twirl?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the lighting golden morning light casting a warm halo and illuminating dewdrops?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.2520795166492462 + ], + "opposite_logprob_tokens": [ + -1.5020794868469238 + ], + "logprob_reward": -0.2520795166492462, + "yes_no_margin_reward": 1.25, + "token_credit_reward": -0.2520795166492462, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.21620261669158936, + 0.06957565993070602, + 0.04683733731508255, + 0.055914197117090225, + 0.07110979408025742, + 0.06693453341722488, + 0.09606318175792694, + 0.11228891462087631, + 0.12936557829380035, + 0.13570816814899445 + ], + "credit_map": [ + 0.025058824568986893, + 0.024947691708803177, + 0.018388403579592705, + 0.02100032940506935, + 0.052498478442430496, + 0.025858594104647636, + 0.031579166650772095, + 0.01687113381922245, + 0.009279085323214531, + 0.006819991394877434, + 0.005870204418897629, + 0.009262490086257458, + 0.01615365967154503, + 0.009834027849137783, + 0.007458081003278494, + 0.004898127168416977, + 0.004818974062800407, + 0.005505912005901337, + 0.008807111531496048, + 0.006312159355729818, + 0.006865142844617367, + 0.005299621727317572, + 0.006018600892275572, + 0.003209814429283142, + 0.006887608207762241, + 0.005746038630604744, + 0.007296466734260321, + 0.0053311786614358425, + 0.00958921480923891, + 0.007455876097083092, + 0.008536658249795437, + 0.005071157123893499, + 0.011837663128972054, + 0.007821891456842422, + 0.005585756152868271, + 0.0055696251802146435, + 0.01001979224383831, + 0.013597025536000729, + 0.011347802355885506, + 0.005330237094312906, + 0.0072250221855938435, + 0.007593703456223011, + 0.007693154737353325, + 0.009602177888154984, + 0.007257115561515093, + 0.010016307234764099, + 0.011395268142223358, + 0.006151782814413309, + 0.01743382029235363, + 0.009109879843890667, + 0.007810899522155523, + 0.012547106482088566, + 0.009196465834975243, + 0.018469545990228653, + 0.012981558218598366, + 0.008513911627233028, + 0.01693371869623661, + 0.012859721668064594, + 0.008927303366363049, + 0.010929406620562077, + 0.02620389498770237, + 0.010960125364363194, + 0.016681095585227013, + 0.008793647401034832, + 0.017137940973043442, + 0.013774500228464603, + 0.01298910565674305, + 0.015314972959458828, + 0.030250653624534607, + 0.011923695914447308, + 0.01854442059993744, + 0.009430287405848503, + 0.018705075606703758, + 0.01674618013203144, + 0.019117219373583794, + 0.012777536176145077, + 0.02322508953511715, + 0.011459671892225742, + 0.018861835822463036, + 0.01481555588543415 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 8 + ], + "selected_reward": -0.2520795166492462 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an elderly villager with graying hair in a bun and deep wrinkles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the setting a newly green rice paddy with rolling green mountains in the distance?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera start with a wide long shot and slowly pan forward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the villager raise her arms toward the sky and then perform a gentle twirl?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the lighting golden morning light casting a warm halo and illuminating dewdrops?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.10033965110778809 + ], + "opposite_logprob_tokens": [ + -2.350339651107788 + ], + "logprob_reward": -0.10033965110778809, + "yes_no_margin_reward": 2.25, + "token_credit_reward": -0.10033965110778809, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1163189485669136, + 0.053728923201560974, + 0.045860655605793, + 0.044807177037000656, + 0.06100751459598541, + 0.04954737424850464, + 0.044959139078855515, + 0.07837296277284622, + 0.18812315165996552, + 0.31727415323257446 + ], + "credit_map": [ + 0.016870353370904922, + 0.015627026557922363, + 0.018441269174218178, + 0.013774141669273376, + 0.01542072556912899, + 0.012643926776945591, + 0.01527500431984663, + 0.008266499266028404, + 0.005488642491400242, + 0.0074326288886368275, + 0.007886365056037903, + 0.008180590346455574, + 0.005556796677410603, + 0.007452100981026888, + 0.006685737054795027, + 0.005046058911830187, + 0.005403711460530758, + 0.0054209791123867035, + 0.005425827577710152, + 0.005392313003540039, + 0.0058908239006996155, + 0.0056196232326328754, + 0.0076510049402713776, + 0.005056374240666628, + 0.00595318665727973, + 0.006258112378418446, + 0.004258384928107262, + 0.004669902380555868, + 0.007502813823521137, + 0.00591393793001771, + 0.006595015060156584, + 0.0036558222491294146, + 0.009608951397240162, + 0.007225858047604561, + 0.005890898406505585, + 0.004834387451410294, + 0.00982331857085228, + 0.009036082774400711, + 0.009878730401396751, + 0.004709288943558931, + 0.008219853043556213, + 0.006691352929919958, + 0.004804189316928387, + 0.004061195533722639, + 0.006634444929659367, + 0.007609191350638866, + 0.006295728962868452, + 0.00523142097517848, + 0.006522344890981913, + 0.0060977027751505375, + 0.005002384074032307, + 0.004577395971864462, + 0.006077553145587444, + 0.006961928680539131, + 0.00588664086535573, + 0.0038331870455294847, + 0.012037569656968117, + 0.009662562981247902, + 0.0059548658318817616, + 0.010325826704502106, + 0.010167551226913929, + 0.009120240807533264, + 0.013128678314387798, + 0.007975667715072632, + 0.02533288300037384, + 0.02262195572257042, + 0.01638314500451088, + 0.017391040921211243, + 0.03064935840666294, + 0.02273331955075264, + 0.03361182287335396, + 0.01939960941672325, + 0.047977034002542496, + 0.04150857776403427, + 0.025345012545585632, + 0.028088998049497604, + 0.056899406015872955, + 0.03285956755280495, + 0.04069947451353073, + 0.043896082788705826 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 8, + 0 + ], + "selected_reward": -0.10033965110778809 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an elderly villager with graying hair in a bun and deep wrinkles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the setting a newly green rice paddy with rolling green mountains in the distance?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera start with a wide long shot and slowly pan forward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the villager raise her arms toward the sky and then perform a gentle twirl?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the lighting golden morning light casting a warm halo and illuminating dewdrops?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.08907730132341385 + ], + "opposite_logprob_tokens": [ + -2.4640772342681885 + ], + "logprob_reward": -0.08907730132341385, + "yes_no_margin_reward": 2.375, + "token_credit_reward": -0.08907730132341385, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3183954060077667, + 0.09402001649141312, + 0.07380013912916183, + 0.05577417090535164, + 0.07204516977071762, + 0.06746123731136322, + 0.08503195643424988, + 0.03630473464727402, + 0.1426316797733307, + 0.05453544482588768 + ], + "credit_map": [ + 0.04455303028225899, + 0.03554093837738037, + 0.03444394841790199, + 0.03174782544374466, + 0.06935809552669525, + 0.03143075108528137, + 0.04427650198340416, + 0.02704431302845478, + 0.011324895545840263, + 0.00899716466665268, + 0.014379640109837055, + 0.016934078186750412, + 0.013904867693781853, + 0.011681832373142242, + 0.00967436097562313, + 0.007123174145817757, + 0.01081074494868517, + 0.008323532529175282, + 0.012111777439713478, + 0.008939463645219803, + 0.009468968026340008, + 0.009059186093509197, + 0.008374971337616444, + 0.006711492780596018, + 0.010132870636880398, + 0.00725587597116828, + 0.004065550398081541, + 0.003543747588992119, + 0.01322740875184536, + 0.00706890132278204, + 0.006867784075438976, + 0.0036120323929935694, + 0.011127878911793232, + 0.009306777268648148, + 0.00817668903619051, + 0.006883669178932905, + 0.008528589271008968, + 0.013682873919606209, + 0.009156695567071438, + 0.005181997548788786, + 0.008845598436892033, + 0.007503576576709747, + 0.006422320380806923, + 0.009791586548089981, + 0.009432647377252579, + 0.00951752532273531, + 0.010285054333508015, + 0.005662925541400909, + 0.01155251357704401, + 0.008685238659381866, + 0.01179822813719511, + 0.010737418197095394, + 0.008335424587130547, + 0.01074996404349804, + 0.01428928691893816, + 0.008883883245289326, + 0.003848177380859852, + 0.004832051694393158, + 0.0069183954037725925, + 0.004628994967788458, + 0.003290928201749921, + 0.004767392296344042, + 0.004384216386824846, + 0.003634579246863723, + 0.01613381877541542, + 0.021172039210796356, + 0.017570987343788147, + 0.017821436747908592, + 0.023938236758112907, + 0.012963713146746159, + 0.01805678755044937, + 0.014974655583500862, + 0.007332509383559227, + 0.0071867723017930984, + 0.005700353533029556, + 0.005433337762951851, + 0.01013971958309412, + 0.0046357824467122555, + 0.007445557974278927, + 0.006661410443484783 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 1 + ], + "selected_reward": -0.08907730132341385 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "The camera opens with a wide, long shot, then slowly and steadily pans forward, capturing a tranquil, pastoral scene from a horizontal angle. As golden morning light slants across the horizon, an elderly villager stands alone in the center of a newly green rice paddy. Her graying hair is pulled back in a simple bun, and her face is etched with deep wrinkles from time, yet her expression is one of unwavering tranquility and contentment. She wears a faded blue homespun blouse and loose black trousers, her cloth shoes stained with morning mud. As the camera closes in, her movements begin: she slowly raises her arms toward the sky, as if embracing the first rays of morning sunlight, her fingertips trembling slightly with the stretch. Then, her body begins a gentle twirl, a simple movement, more like an impromptu expression of unity with nature, the hem of her cloth dress fluttering lightly as she turns. The soft sidelight casts a warm golden halo around her silhouette, illuminating the dewdrops in the field with a shimmering clarity. In the distance are rolling green mountains with a light morning mist lingering between the mountains. The whole picture is dominated by the tender green of the rice fields and the light blue of the sky, presenting an atmosphere that is warm, peaceful and full of hope for life.", + "seed": 4455, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000071/group_0/index_5902_seed_4455_dp_0_rank_0_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 0, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.061686869245022535, + "vlm_logprob": -0.061686869245022535, + "vlm_yes_no_margin": 3.425, + "vlm_token_credit": -0.061686869245022535, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an elderly villager with graying hair in a bun and deep wrinkles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the setting a newly green rice paddy with rolling green mountains in the distance?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera start with a wide long shot and slowly pan forward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the villager raise her arms toward the sky and then perform a gentle twirl?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the lighting golden morning light casting a warm halo and illuminating dewdrops?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0444120392203331 + ], + "opposite_logprob_tokens": [ + -3.169412136077881 + ], + "logprob_reward": -0.0444120392203331, + "yes_no_margin_reward": 3.125, + "token_credit_reward": -0.0444120392203331, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15903571248054504, + 0.0554044209420681, + 0.04236431047320366, + 0.02567223459482193, + 0.03716077283024788, + 0.25792092084884644, + 0.07465355098247528, + 0.1981084644794464, + 0.06615810841321945, + 0.0835215151309967 + ], + "credit_map": [ + 0.02579914778470993, + 0.01766694337129593, + 0.02697685733437538, + 0.023985927924513817, + 0.013918213546276093, + 0.013254763558506966, + 0.016644680872559547, + 0.020789187401533127, + 0.006774973124265671, + 0.004461988806724548, + 0.006139055825769901, + 0.014533095993101597, + 0.005272896494716406, + 0.004191291518509388, + 0.006109985057264566, + 0.007921131327748299, + 0.0067061628215014935, + 0.0038812116254121065, + 0.0037149637937545776, + 0.005293062888085842, + 0.0048060608096420765, + 0.003215773031115532, + 0.0031928992830216885, + 0.011554178781807423, + 0.002113242633640766, + 0.0031517117749899626, + 0.0028371759690344334, + 0.002862189896404743, + 0.005488938186317682, + 0.003181425854563713, + 0.0027262528892606497, + 0.0033112948294728994, + 0.002603227039799094, + 0.0036558061838150024, + 0.00532075809314847, + 0.00408183503895998, + 0.00601730914786458, + 0.003970141522586346, + 0.006651816423982382, + 0.00485988287255168, + 0.016254782676696777, + 0.011660505086183548, + 0.09799937903881073, + 0.0908588394522667, + 0.01725205034017563, + 0.007587855216115713, + 0.007273295894265175, + 0.009034214541316032, + 0.0035749003291130066, + 0.004874980542808771, + 0.020111465826630592, + 0.016832657158374786, + 0.007207848597317934, + 0.010932283475995064, + 0.004981691017746925, + 0.00613772077485919, + 0.021794885396957397, + 0.01127825491130352, + 0.06992526352405548, + 0.029012465849518776, + 0.01780146174132824, + 0.012285959906876087, + 0.01416860893368721, + 0.02184157632291317, + 0.0050672488287091255, + 0.00821492075920105, + 0.008697492070496082, + 0.008642338216304779, + 0.008739527314901352, + 0.010908685624599457, + 0.009182868525385857, + 0.006705019623041153, + 0.006482456345111132, + 0.006764614023268223, + 0.013044130988419056, + 0.009270759299397469, + 0.00940293725579977, + 0.022867167368531227, + 0.00831174198538065, + 0.0073777069337666035 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 7, + 0 + ], + "selected_reward": -0.0444120392203331 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an elderly villager with graying hair in a bun and deep wrinkles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the setting a newly green rice paddy with rolling green mountains in the distance?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera start with a wide long shot and slowly pan forward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the villager raise her arms toward the sky and then perform a gentle twirl?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the lighting golden morning light casting a warm halo and illuminating dewdrops?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.002828290220350027 + ], + "opposite_logprob_tokens": [ + -5.877828121185303 + ], + "logprob_reward": -0.002828290220350027, + "yes_no_margin_reward": 5.875, + "token_credit_reward": -0.002828290220350027, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3081316649913788, + 0.039564404636621475, + 0.03763216361403465, + 0.05380712077021599, + 0.07365836948156357, + 0.20820938050746918, + 0.051396407186985016, + 0.06556746363639832, + 0.08547063916921616, + 0.07656237483024597 + ], + "credit_map": [ + 0.022574419155716896, + 0.029718199744820595, + 0.0844217836856842, + 0.06355087459087372, + 0.020534668117761612, + 0.03164436295628548, + 0.0220329686999321, + 0.03365437686443329, + 0.0028584289830178022, + 0.003494263393804431, + 0.004576995503157377, + 0.012544889003038406, + 0.003386883297935128, + 0.0029780010227113962, + 0.0037899543531239033, + 0.005934987682849169, + 0.007280933205038309, + 0.004139257594943047, + 0.004145406186580658, + 0.0031788614578545094, + 0.005534970201551914, + 0.004311271011829376, + 0.003491538343951106, + 0.00554992537945509, + 0.01030759047716856, + 0.007104318588972092, + 0.009128255769610405, + 0.007047263439744711, + 0.0035425222013145685, + 0.005313515197485685, + 0.004014365840703249, + 0.007349291350692511, + 0.005870846100151539, + 0.006699925288558006, + 0.015621460042893887, + 0.0098028015345335, + 0.0081028388813138, + 0.0059976959601044655, + 0.010594895109534264, + 0.010967908427119255, + 0.01093738991767168, + 0.010658959858119488, + 0.08075132966041565, + 0.07657930999994278, + 0.008462081663310528, + 0.0065600257366895676, + 0.005698811262845993, + 0.00856146588921547, + 0.0027242363430559635, + 0.0029745902866125107, + 0.016148611903190613, + 0.010539642535150051, + 0.004552003927528858, + 0.006087985821068287, + 0.002579968888312578, + 0.005789367947727442, + 0.006949238944798708, + 0.003125340212136507, + 0.02471119351685047, + 0.011014539748430252, + 0.004085582215338945, + 0.0034957798197865486, + 0.004063418600708246, + 0.008122369647026062, + 0.008265668526291847, + 0.00939995888620615, + 0.011125251650810242, + 0.012487848289310932, + 0.010076276957988739, + 0.01969246193766594, + 0.006119637284427881, + 0.00830353144556284, + 0.005338291171938181, + 0.007407826837152243, + 0.0070936656557023525, + 0.0059601739048957825, + 0.010412237606942654, + 0.02561294659972191, + 0.006475530564785004, + 0.00826170202344656 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 5, + 8 + ], + "selected_reward": -0.002828290220350027 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an elderly villager with graying hair in a bun and deep wrinkles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the setting a newly green rice paddy with rolling green mountains in the distance?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera start with a wide long shot and slowly pan forward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the villager raise her arms toward the sky and then perform a gentle twirl?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the lighting golden morning light casting a warm halo and illuminating dewdrops?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.179864764213562 + ], + "opposite_logprob_tokens": [ + -1.804864764213562 + ], + "logprob_reward": -0.179864764213562, + "yes_no_margin_reward": 1.625, + "token_credit_reward": -0.179864764213562, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.01593148522078991, + 0.005845947191119194, + 0.007589554414153099, + 0.005471945274621248, + 0.018768416717648506, + 0.8786517381668091, + 0.030289702117443085, + 0.020230116322636604, + 0.005412091501057148, + 0.01180899515748024 + ], + "credit_map": [ + 0.0018015486421063542, + 0.001974891871213913, + 0.0023560503032058477, + 0.0019938016775995493, + 0.002772248350083828, + 0.0017146988539025187, + 0.001455176156014204, + 0.0018630694830790162, + 0.0005964762531220913, + 0.0009054358233697712, + 0.0006299536325968802, + 0.0005700500914826989, + 0.0011244233464822173, + 0.0007155491621233523, + 0.0007440857589244843, + 0.0005599730066023767, + 0.0010885947849601507, + 0.0008477188530378044, + 0.0009155907318927348, + 0.0010662198765203357, + 0.0013602522667497396, + 0.0006850264617241919, + 0.0007043990190140903, + 0.0009217529441229999, + 0.0005917465314269066, + 0.000676293799187988, + 0.0005063185817562044, + 0.00045606092317029834, + 0.001460631494410336, + 0.0005846913554705679, + 0.0005769972340203822, + 0.0006192054133862257, + 0.0017338759498670697, + 0.0018452514195814729, + 0.002914209384471178, + 0.0022999956272542477, + 0.003514561802148819, + 0.0019665975123643875, + 0.0019011268159374595, + 0.002592795994132757, + 0.038806941360235214, + 0.02375096082687378, + 0.374329149723053, + 0.3693094849586487, + 0.009413027204573154, + 0.013327062129974365, + 0.0177838746458292, + 0.031931206583976746, + 0.0019970261491835117, + 0.0016957790357992053, + 0.009621583856642246, + 0.007536550983786583, + 0.0024194642901420593, + 0.003940778784453869, + 0.0013662255369126797, + 0.001712293829768896, + 0.0024975575506687164, + 0.0008376213954761624, + 0.008602505549788475, + 0.0034847331698983908, + 0.0006439924472942948, + 0.0006292131729424, + 0.0009114936692640185, + 0.0026229992508888245, + 0.00042624230263754725, + 0.0005136714898981154, + 0.0005482450360432267, + 0.0006527210935018957, + 0.0011314384173601866, + 0.0005500486004166305, + 0.000922619248740375, + 0.0006671053124591708, + 0.0009514070698060095, + 0.0014710606774315238, + 0.0012854012893512845, + 0.0016397166764363647, + 0.001743731671012938, + 0.001632891595363617, + 0.0016206287546083331, + 0.001464157598093152 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 6, + 7 + ], + "selected_reward": -0.179864764213562 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an elderly villager with graying hair in a bun and deep wrinkles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the setting a newly green rice paddy with rolling green mountains in the distance?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera start with a wide long shot and slowly pan forward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the villager raise her arms toward the sky and then perform a gentle twirl?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the lighting golden morning light casting a warm halo and illuminating dewdrops?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02634979411959648 + ], + "opposite_logprob_tokens": [ + -3.6513497829437256 + ], + "logprob_reward": -0.02634979411959648, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.02634979411959648, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.030557138845324516, + 0.009191079996526241, + 0.031254444271326065, + 0.010041596367955208, + 0.015342782251536846, + 0.6159589886665344, + 0.15902461111545563, + 0.0623183436691761, + 0.007376200519502163, + 0.058934785425662994 + ], + "credit_map": [ + 0.006087465211749077, + 0.004837908782064915, + 0.0052358293905854225, + 0.004458969924598932, + 0.002185788005590439, + 0.00198458693921566, + 0.0022571810986846685, + 0.003509411122649908, + 0.0008824639371596277, + 0.000887686328496784, + 0.0007411706610582769, + 0.0021846042945981026, + 0.001469047274440527, + 0.0009366649901494384, + 0.0009584627114236355, + 0.0011309797409921885, + 0.0037284246645867825, + 0.0027224591467529535, + 0.002043855609372258, + 0.004709477536380291, + 0.00483668502420187, + 0.002117706695571542, + 0.0026692550163716078, + 0.00842658057808876, + 0.001205333974212408, + 0.00114484503865242, + 0.0012298764195293188, + 0.001126763061620295, + 0.0020209834910929203, + 0.0009473589016124606, + 0.0009362102719023824, + 0.0014302252093330026, + 0.0019510244019329548, + 0.0011481147957965732, + 0.002542967675253749, + 0.0028622562531381845, + 0.002156869973987341, + 0.0012248357525095344, + 0.0012328341836109757, + 0.0022238795645534992, + 0.027330251410603523, + 0.017036762088537216, + 0.2615518867969513, + 0.2570812404155731, + 0.008172674104571342, + 0.009348294697701931, + 0.012989336624741554, + 0.022448524832725525, + 0.01177920587360859, + 0.009892363101243973, + 0.048513755202293396, + 0.028368638828396797, + 0.01413488108664751, + 0.023683197796344757, + 0.008060310035943985, + 0.014592260122299194, + 0.007374254520982504, + 0.0025485658552497625, + 0.027265911921858788, + 0.01042084489017725, + 0.0017276633298024535, + 0.0018905180040746927, + 0.0019942985381931067, + 0.009096288122236729, + 0.0006507353973574936, + 0.0007869815453886986, + 0.0007222366984933615, + 0.0010884226066991687, + 0.0010209328029304743, + 0.0013754748506471515, + 0.0008850852027535439, + 0.0008463311824016273, + 0.0039570811204612255, + 0.005524932872503996, + 0.006713217590004206, + 0.005892367102205753, + 0.0066432105377316475, + 0.019422346726059914, + 0.005005776882171631, + 0.005775852128863335 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 6, + 7 + ], + "selected_reward": -0.02634979411959648 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an elderly villager with graying hair in a bun and deep wrinkles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the setting a newly green rice paddy with rolling green mountains in the distance?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera start with a wide long shot and slowly pan forward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the villager raise her arms toward the sky and then perform a gentle twirl?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the lighting golden morning light casting a warm halo and illuminating dewdrops?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.05497945845127106 + ], + "opposite_logprob_tokens": [ + -2.9299795627593994 + ], + "logprob_reward": -0.05497945845127106, + "yes_no_margin_reward": 2.875, + "token_credit_reward": -0.05497945845127106, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1101280227303505, + 0.031105583533644676, + 0.01984480209648609, + 0.017003891989588737, + 0.02179175801575184, + 0.6206085681915283, + 0.03506017103791237, + 0.10153576731681824, + 0.016680611297488213, + 0.026240818202495575 + ], + "credit_map": [ + 0.025150565430521965, + 0.01757422462105751, + 0.01228154357522726, + 0.011849645525217056, + 0.009776446968317032, + 0.008570737205445766, + 0.007341375108808279, + 0.01758350245654583, + 0.007252666167914867, + 0.004416229669004679, + 0.003058508737012744, + 0.004142554476857185, + 0.0016284359153360128, + 0.0027798032388091087, + 0.0023760695476084948, + 0.005451318342238665, + 0.0019936831668019295, + 0.002066283719614148, + 0.004538060165941715, + 0.0031854782719165087, + 0.002440337324514985, + 0.0020452584140002728, + 0.0015611809212714434, + 0.0020145236048847437, + 0.0018642450449988246, + 0.0018324622651562095, + 0.0032848636619746685, + 0.0024276108015328646, + 0.0030354263726621866, + 0.0016707531176507473, + 0.0015551985707134008, + 0.0013333343667909503, + 0.0016416582511737943, + 0.0021967014763504267, + 0.004122281912714243, + 0.0038433971349149942, + 0.0020119035616517067, + 0.001685332739725709, + 0.0026739456225186586, + 0.0036165392957627773, + 0.0298242699354887, + 0.01785111241042614, + 0.26448342204093933, + 0.2579837739467621, + 0.005888932850211859, + 0.009538532234728336, + 0.012553920038044453, + 0.02248462848365307, + 0.003362244926393032, + 0.0021705187391489744, + 0.011515302583575249, + 0.0067850337363779545, + 0.0025524008087813854, + 0.004307972267270088, + 0.0015410414198413491, + 0.0028256606310606003, + 0.011341888457536697, + 0.003831214737147093, + 0.04404418170452118, + 0.016842244192957878, + 0.0029436680488288403, + 0.004019058775156736, + 0.003291147295385599, + 0.015222371555864811, + 0.001262044650502503, + 0.002105219755321741, + 0.002045400906354189, + 0.002167521743103862, + 0.0018327889265492558, + 0.003964128904044628, + 0.0016349718207493424, + 0.0016685365699231625, + 0.0018971115350723267, + 0.0023273194674402475, + 0.0033429565373808146, + 0.0028868606314063072, + 0.0033780999947339296, + 0.007274260278791189, + 0.0023007357958704233, + 0.0028334783855825663 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 0, + 7 + ], + "selected_reward": -0.05497945845127106 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "The camera opens with a wide, long shot, then slowly and steadily pans forward, capturing a tranquil, pastoral scene from a horizontal angle. As golden morning light slants across the horizon, an elderly villager stands alone in the center of a newly green rice paddy. Her graying hair is pulled back in a simple bun, and her face is etched with deep wrinkles from time, yet her expression is one of unwavering tranquility and contentment. She wears a faded blue homespun blouse and loose black trousers, her cloth shoes stained with morning mud. As the camera closes in, her movements begin: she slowly raises her arms toward the sky, as if embracing the first rays of morning sunlight, her fingertips trembling slightly with the stretch. Then, her body begins a gentle twirl, a simple movement, more like an impromptu expression of unity with nature, the hem of her cloth dress fluttering lightly as she turns. The soft sidelight casts a warm golden halo around her silhouette, illuminating the dewdrops in the field with a shimmering clarity. In the distance are rolling green mountains with a light morning mist lingering between the mountains. The whole picture is dominated by the tender green of the rice fields and the light blue of the sky, presenting an atmosphere that is warm, peaceful and full of hope for life.", + "seed": 4456, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000071/group_0/index_5902_seed_4456_dp_1_rank_2_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 0, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.07693388075567782, + "vlm_logprob": -0.07693388075567782, + "vlm_yes_no_margin": 3.55, + "vlm_token_credit": -0.07693388075567782, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an elderly villager with graying hair in a bun and deep wrinkles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the setting a newly green rice paddy with rolling green mountains in the distance?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera start with a wide long shot and slowly pan forward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the villager raise her arms toward the sky and then perform a gentle twirl?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the lighting golden morning light casting a warm halo and illuminating dewdrops?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.05018831789493561 + ], + "opposite_logprob_tokens": [ + -3.0501883029937744 + ], + "logprob_reward": -0.05018831789493561, + "yes_no_margin_reward": 3.0, + "token_credit_reward": -0.05018831789493561, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10443074256181717, + 0.10940226167440414, + 0.04541836678981781, + 0.14466559886932373, + 0.12111064046621323, + 0.11473048478364944, + 0.043488673865795135, + 0.0653865858912468, + 0.045618653297424316, + 0.20574796199798584 + ], + "credit_map": [ + 0.009418196976184845, + 0.011650106869637966, + 0.017525160685181618, + 0.013968659564852715, + 0.014086372219026089, + 0.009996681474149227, + 0.011019067838788033, + 0.016766507178544998, + 0.0074158902280032635, + 0.011004327796399593, + 0.026908989995718002, + 0.014962241984903812, + 0.00847351923584938, + 0.008819135837256908, + 0.007400685455650091, + 0.02441748045384884, + 0.003988438751548529, + 0.005584449041634798, + 0.006702587008476257, + 0.005513698328286409, + 0.00952207576483488, + 0.0030108920764178038, + 0.004791188985109329, + 0.0063050370663404465, + 0.007027810905128717, + 0.013778634369373322, + 0.02732347883284092, + 0.0307819452136755, + 0.01594826579093933, + 0.010967467911541462, + 0.012638568878173828, + 0.026199430227279663, + 0.010088426060974598, + 0.012804617173969746, + 0.014939983375370502, + 0.014605957083404064, + 0.019977973774075508, + 0.01679232344031334, + 0.014985723420977592, + 0.016915645450353622, + 0.010890555568039417, + 0.009719225578010082, + 0.02906191349029541, + 0.006425843574106693, + 0.014109943062067032, + 0.011475726962089539, + 0.009827363304793835, + 0.023219915106892586, + 0.003490197705104947, + 0.005282122641801834, + 0.006683142855763435, + 0.007865888066589832, + 0.0053688157349824905, + 0.003793540643528104, + 0.004835744388401508, + 0.006169228348881006, + 0.005956584587693214, + 0.011364784091711044, + 0.007812061812728643, + 0.006538998801261187, + 0.010883113369345665, + 0.009322992525994778, + 0.004825741518288851, + 0.008682317100465298, + 0.0034978073090314865, + 0.007739819120615721, + 0.005877586547285318, + 0.007281601428985596, + 0.005664182361215353, + 0.004598483443260193, + 0.0060036820359528065, + 0.004955492448061705, + 0.01644069328904152, + 0.03188036382198334, + 0.028159167617559433, + 0.023143921047449112, + 0.025864973664283752, + 0.024906259030103683, + 0.030917402356863022, + 0.024435201659798622 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 3, + 4 + ], + "selected_reward": -0.05018831789493561 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an elderly villager with graying hair in a bun and deep wrinkles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the setting a newly green rice paddy with rolling green mountains in the distance?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera start with a wide long shot and slowly pan forward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the villager raise her arms toward the sky and then perform a gentle twirl?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the lighting golden morning light casting a warm halo and illuminating dewdrops?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0024994343984872103 + ], + "opposite_logprob_tokens": [ + -6.002499580383301 + ], + "logprob_reward": -0.0024994343984872103, + "yes_no_margin_reward": 6.0, + "token_credit_reward": -0.0024994343984872103, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10863232612609863, + 0.23554296791553497, + 0.1793101280927658, + 0.10478205233812332, + 0.03437216952443123, + 0.10969337821006775, + 0.047144901007413864, + 0.06154930591583252, + 0.0794079601764679, + 0.039564814418554306 + ], + "credit_map": [ + 0.01248258538544178, + 0.007640928495675325, + 0.023546993732452393, + 0.019277309998869896, + 0.009754136204719543, + 0.007393844425678253, + 0.009876134805381298, + 0.018660390749573708, + 0.009394796565175056, + 0.009173737838864326, + 0.09194692224264145, + 0.010774858295917511, + 0.007098004221916199, + 0.01225180272012949, + 0.009540868923068047, + 0.08536196500062943, + 0.01459257397800684, + 0.0215193759649992, + 0.031508173793554306, + 0.01475326344370842, + 0.031422436237335205, + 0.01243039034307003, + 0.012930719181895256, + 0.04015319049358368, + 0.009042016230523586, + 0.010882917791604996, + 0.01486263982951641, + 0.01854824647307396, + 0.019793422892689705, + 0.008807414211332798, + 0.008533812128007412, + 0.01431156788021326, + 0.004997375886887312, + 0.004683252424001694, + 0.0055700624361634254, + 0.0035320764873176813, + 0.00443680165335536, + 0.002672507194802165, + 0.003349033882841468, + 0.005131056532263756, + 0.007619477342814207, + 0.003519648453220725, + 0.0393797904253006, + 0.004953351803123951, + 0.008333018980920315, + 0.00843549519777298, + 0.006187458988279104, + 0.031265128403902054, + 0.004608894698321819, + 0.004559446591883898, + 0.009386928752064705, + 0.006757044233381748, + 0.004384177271276712, + 0.006314015481621027, + 0.004881720989942551, + 0.006252665538340807, + 0.005722873844206333, + 0.007237305864691734, + 0.008952979929745197, + 0.0061825942248106, + 0.010286224074661732, + 0.01071128062903881, + 0.0052834185771644115, + 0.007172625977545977, + 0.008044999092817307, + 0.016588827595114708, + 0.009037424810230732, + 0.009259321726858616, + 0.010129544883966446, + 0.009401601739227772, + 0.009527305141091347, + 0.0074189333245158195, + 0.0032582185231149197, + 0.008280016481876373, + 0.003744447138160467, + 0.005835448391735554, + 0.004659052938222885, + 0.0035063992254436016, + 0.0054553491063416, + 0.004825878422707319 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 2, + 5 + ], + "selected_reward": -0.0024994343984872103 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an elderly villager with graying hair in a bun and deep wrinkles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the setting a newly green rice paddy with rolling green mountains in the distance?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera start with a wide long shot and slowly pan forward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the villager raise her arms toward the sky and then perform a gentle twirl?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the lighting golden morning light casting a warm halo and illuminating dewdrops?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.2813144326210022 + ], + "opposite_logprob_tokens": [ + -1.4063143730163574 + ], + "logprob_reward": -0.2813144326210022, + "yes_no_margin_reward": 1.125, + "token_credit_reward": -0.2813144326210022, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.04726820066571236, + 0.09467491507530212, + 0.09757053852081299, + 0.10254883021116257, + 0.047081079334020615, + 0.23479437828063965, + 0.05893528833985329, + 0.04330538585782051, + 0.24392829835414886, + 0.02989308536052704 + ], + "credit_map": [ + 0.003951354883611202, + 0.0053119538351893425, + 0.007787328213453293, + 0.0051215230487287045, + 0.006622920278459787, + 0.004631454590708017, + 0.007051115855574608, + 0.006790549028664827, + 0.007077885791659355, + 0.008251674473285675, + 0.02042766660451889, + 0.0151326609775424, + 0.009185819886624813, + 0.00791162345558405, + 0.005673789419233799, + 0.021013788878917694, + 0.008419555611908436, + 0.01392705924808979, + 0.01908799819648266, + 0.007456666324287653, + 0.020558780059218407, + 0.006302554160356522, + 0.007302605547010899, + 0.014515317976474762, + 0.0070853084325790405, + 0.009976856410503387, + 0.01302530337125063, + 0.0077980258502066135, + 0.028802961111068726, + 0.009844714775681496, + 0.01307836826890707, + 0.01293728593736887, + 0.0036266471724957228, + 0.0033163491170853376, + 0.010039729066193104, + 0.004475429188460112, + 0.00791070144623518, + 0.0031276235822588205, + 0.007251657545566559, + 0.007332942448556423, + 0.041280124336481094, + 0.022803442552685738, + 0.023059900850057602, + 0.0205074492841959, + 0.03574748337268829, + 0.043468061834573746, + 0.020965905860066414, + 0.02696201391518116, + 0.005196997430175543, + 0.009748686105012894, + 0.009182805195450783, + 0.00988298375159502, + 0.00576169416308403, + 0.00468627130612731, + 0.0076082865707576275, + 0.006867567077279091, + 0.0038193194195628166, + 0.0049390788190066814, + 0.005572912748903036, + 0.003505383152514696, + 0.007616968359798193, + 0.009198367595672607, + 0.0033914430532604456, + 0.005261910613626242, + 0.026181688532233238, + 0.04595166817307472, + 0.03145604580640793, + 0.02635718509554863, + 0.0346902497112751, + 0.029015524312853813, + 0.027231642976403236, + 0.023044293746352196, + 0.002004757057875395, + 0.0027504509780555964, + 0.0038156353402882814, + 0.004416675306856632, + 0.0025644232518970966, + 0.004095243755728006, + 0.005541118793189526, + 0.004704780876636505 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 5, + 3 + ], + "selected_reward": -0.2813144326210022 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an elderly villager with graying hair in a bun and deep wrinkles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the setting a newly green rice paddy with rolling green mountains in the distance?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera start with a wide long shot and slowly pan forward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the villager raise her arms toward the sky and then perform a gentle twirl?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the lighting golden morning light casting a warm halo and illuminating dewdrops?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.012545477598905563 + ], + "opposite_logprob_tokens": [ + -4.387545585632324 + ], + "logprob_reward": -0.012545477598905563, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.012545477598905563, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.23037000000476837, + 0.17486371099948883, + 0.14395423233509064, + 0.1944178342819214, + 0.0390830859541893, + 0.03940587863326073, + 0.03449216112494469, + 0.041871290653944016, + 0.06269004195928574, + 0.03885183483362198 + ], + "credit_map": [ + 0.012309717945754528, + 0.018011389300227165, + 0.06319107115268707, + 0.01995866373181343, + 0.025059878826141357, + 0.013543393462896347, + 0.027159351855516434, + 0.05113649368286133, + 0.006960618309676647, + 0.004587734118103981, + 0.06884031742811203, + 0.007208652328699827, + 0.008395264856517315, + 0.007381211034953594, + 0.007307896856218576, + 0.06418200582265854, + 0.012815812602639198, + 0.021369045600295067, + 0.03135548532009125, + 0.01700202375650406, + 0.018725350499153137, + 0.013380458578467369, + 0.010465947911143303, + 0.018840093165636063, + 0.01871608756482601, + 0.01579960063099861, + 0.021697308868169785, + 0.02617076225578785, + 0.040408797562122345, + 0.01921592652797699, + 0.016627956181764603, + 0.03578139841556549, + 0.0031551048159599304, + 0.003711940720677376, + 0.005052740219980478, + 0.004832498729228973, + 0.006991610862314701, + 0.003960541915148497, + 0.004968772642314434, + 0.006409874651581049, + 0.0033211379777640104, + 0.004446703940629959, + 0.003916833084076643, + 0.004719094838947058, + 0.008614370599389076, + 0.0038479981012642384, + 0.005135405343025923, + 0.0054043312557041645, + 0.003268321743234992, + 0.003907026723027229, + 0.007502231281250715, + 0.0047599319368600845, + 0.0037850914523005486, + 0.0030908628832548857, + 0.0036779192741960287, + 0.004500769078731537, + 0.004036527592688799, + 0.00522567005828023, + 0.005622855853289366, + 0.0028186540585011244, + 0.008830679580569267, + 0.005412308033555746, + 0.0038429005071520805, + 0.006081689149141312, + 0.006319673731923103, + 0.013744749128818512, + 0.008791210129857063, + 0.006541636772453785, + 0.005984764080494642, + 0.00655379518866539, + 0.007766075897961855, + 0.006988131906837225, + 0.003754032077267766, + 0.005319093819707632, + 0.005863710772246122, + 0.004414988681674004, + 0.004483269527554512, + 0.005345860496163368, + 0.005150540266185999, + 0.004520336166024208 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 1 + ], + "selected_reward": -0.012545477598905563 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an elderly villager with graying hair in a bun and deep wrinkles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the setting a newly green rice paddy with rolling green mountains in the distance?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera start with a wide long shot and slowly pan forward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the villager raise her arms toward the sky and then perform a gentle twirl?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the lighting golden morning light casting a warm halo and illuminating dewdrops?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03812174126505852 + ], + "opposite_logprob_tokens": [ + -3.2881217002868652 + ], + "logprob_reward": -0.03812174126505852, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.03812174126505852, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1682313084602356, + 0.07460744678974152, + 0.08068858087062836, + 0.10993124544620514, + 0.10096309334039688, + 0.09376443922519684, + 0.06900069117546082, + 0.09767011553049088, + 0.03706524893641472, + 0.16807781159877777 + ], + "credit_map": [ + 0.015944834798574448, + 0.009708273224532604, + 0.04529615119099617, + 0.01765366643667221, + 0.017175571992993355, + 0.010379155166447163, + 0.01205501426011324, + 0.04001864418387413, + 0.004903742577880621, + 0.005622057244181633, + 0.022387461736798286, + 0.004885992966592312, + 0.005929422099143267, + 0.006333937402814627, + 0.0048223864287137985, + 0.01972244307398796, + 0.004443507641553879, + 0.007605563849210739, + 0.022737106308341026, + 0.004352393560111523, + 0.0077249761670827866, + 0.004103072918951511, + 0.004101670812815428, + 0.025620294734835625, + 0.006890855263918638, + 0.006758518051356077, + 0.022513562813401222, + 0.02140676975250244, + 0.012811827473342419, + 0.00827791541814804, + 0.009233364835381508, + 0.02203843928873539, + 0.00896204449236393, + 0.009113232605159283, + 0.01837919093668461, + 0.00824963953346014, + 0.016057487577199936, + 0.007962052710354328, + 0.014748409390449524, + 0.017491033300757408, + 0.00505555234849453, + 0.005773676559329033, + 0.030313586816191673, + 0.004804092925041914, + 0.011053192429244518, + 0.006730012595653534, + 0.006475990638136864, + 0.02355833165347576, + 0.005787249654531479, + 0.008471923880279064, + 0.011983933858573437, + 0.008917160332202911, + 0.009527312591671944, + 0.009411420673131943, + 0.006219091359525919, + 0.008682599291205406, + 0.008719375357031822, + 0.012822438962757587, + 0.01665390282869339, + 0.01025441475212574, + 0.017274782061576843, + 0.013846896588802338, + 0.007600430399179459, + 0.010497869923710823, + 0.0031546526588499546, + 0.005404775962233543, + 0.005508077796548605, + 0.00564460689201951, + 0.004318391438573599, + 0.0036435630172491074, + 0.005251559894531965, + 0.004139621742069721, + 0.014862081967294216, + 0.028866322711110115, + 0.019559133797883987, + 0.019717583432793617, + 0.021902136504650116, + 0.01979275792837143, + 0.02432224340736866, + 0.019055558368563652 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 3 + ], + "selected_reward": -0.03812174126505852 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "The camera opens with a wide, long shot, then slowly and steadily pans forward, capturing a tranquil, pastoral scene from a horizontal angle. As golden morning light slants across the horizon, an elderly villager stands alone in the center of a newly green rice paddy. Her graying hair is pulled back in a simple bun, and her face is etched with deep wrinkles from time, yet her expression is one of unwavering tranquility and contentment. She wears a faded blue homespun blouse and loose black trousers, her cloth shoes stained with morning mud. As the camera closes in, her movements begin: she slowly raises her arms toward the sky, as if embracing the first rays of morning sunlight, her fingertips trembling slightly with the stretch. Then, her body begins a gentle twirl, a simple movement, more like an impromptu expression of unity with nature, the hem of her cloth dress fluttering lightly as she turns. The soft sidelight casts a warm golden halo around her silhouette, illuminating the dewdrops in the field with a shimmering clarity. In the distance are rolling green mountains with a light morning mist lingering between the mountains. The whole picture is dominated by the tender green of the rice fields and the light blue of the sky, presenting an atmosphere that is warm, peaceful and full of hope for life.", + "seed": 4457, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000071/group_0/index_5902_seed_4457_dp_1_rank_2_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 0, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.10506416577845812, + "vlm_logprob": -0.10506416577845812, + "vlm_yes_no_margin": 3.275, + "vlm_token_credit": -0.10506416577845812, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an elderly villager with graying hair in a bun and deep wrinkles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the setting a newly green rice paddy with rolling green mountains in the distance?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera start with a wide long shot and slowly pan forward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the villager raise her arms toward the sky and then perform a gentle twirl?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the lighting golden morning light casting a warm halo and illuminating dewdrops?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.08089606463909149 + ], + "opposite_logprob_tokens": [ + -2.5808961391448975 + ], + "logprob_reward": -0.08089606463909149, + "yes_no_margin_reward": 2.5, + "token_credit_reward": -0.08089606463909149, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.17971564829349518, + 0.08069348335266113, + 0.1407000571489334, + 0.1156235784292221, + 0.08501572161912918, + 0.039825186133384705, + 0.06938248127698898, + 0.028779424726963043, + 0.19421014189720154, + 0.06605426222085953 + ], + "credit_map": [ + 0.0212255846709013, + 0.024935375899076462, + 0.029566291719675064, + 0.0167184267193079, + 0.029258986935019493, + 0.01322655938565731, + 0.021058687940239906, + 0.023725735023617744, + 0.008957436308264732, + 0.009849279187619686, + 0.011501576751470566, + 0.01636376418173313, + 0.011926738545298576, + 0.005065277684479952, + 0.008803799748420715, + 0.008225610479712486, + 0.025852816179394722, + 0.03597670793533325, + 0.028142422437667847, + 0.010581562295556068, + 0.01299984846264124, + 0.005329134874045849, + 0.009087970480322838, + 0.012729601934552193, + 0.018913475796580315, + 0.017474764958024025, + 0.010254457592964172, + 0.008494184352457523, + 0.02619098499417305, + 0.007468039635568857, + 0.016321228817105293, + 0.010506441816687584, + 0.009381660260260105, + 0.008891345001757145, + 0.008834091946482658, + 0.009732149541378021, + 0.01631920225918293, + 0.006822356954216957, + 0.01320687960833311, + 0.011828036047518253, + 0.0035593193024396896, + 0.0038543378468602896, + 0.0041779628954827785, + 0.0070389872416853905, + 0.006486127618700266, + 0.0034467200748622417, + 0.004617445170879364, + 0.006644285283982754, + 0.0039119175635278225, + 0.0036492906510829926, + 0.02176726423203945, + 0.009257650934159756, + 0.00476115383207798, + 0.003852910827845335, + 0.0048791831359267235, + 0.017303114756941795, + 0.002531274687498808, + 0.0022864092607051134, + 0.004849462769925594, + 0.004160980693995953, + 0.004131345544010401, + 0.0020172749646008015, + 0.0029701553285121918, + 0.005832522176206112, + 0.025677816942334175, + 0.02942628040909767, + 0.0339311882853508, + 0.02513187564909458, + 0.029863035306334496, + 0.012951274402439594, + 0.02106773853302002, + 0.016160929575562477, + 0.008734619244933128, + 0.007527000270783901, + 0.012699941173195839, + 0.005857424810528755, + 0.013280033133924007, + 0.004629196133464575, + 0.0069576590321958065, + 0.006368386093527079 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 0, + 2 + ], + "selected_reward": -0.08089606463909149 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an elderly villager with graying hair in a bun and deep wrinkles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the setting a newly green rice paddy with rolling green mountains in the distance?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera start with a wide long shot and slowly pan forward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the villager raise her arms toward the sky and then perform a gentle twirl?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the lighting golden morning light casting a warm halo and illuminating dewdrops?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0032010050490498543 + ], + "opposite_logprob_tokens": [ + -5.753201007843018 + ], + "logprob_reward": -0.0032010050490498543, + "yes_no_margin_reward": 5.75, + "token_credit_reward": -0.0032010050490498543, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3232838809490204, + 0.10637380182743073, + 0.08847329020500183, + 0.0493217371404171, + 0.0586080327630043, + 0.03415204584598541, + 0.0717916414141655, + 0.1264582872390747, + 0.09067235887050629, + 0.050864964723587036 + ], + "credit_map": [ + 0.056973040103912354, + 0.054082486778497696, + 0.04058236628770828, + 0.04830222576856613, + 0.043182648718357086, + 0.02042291685938835, + 0.01923724263906479, + 0.04050096869468689, + 0.014041177928447723, + 0.011737082153558731, + 0.013886927627027035, + 0.019800003618001938, + 0.011931691318750381, + 0.01372580137103796, + 0.0076675317250192165, + 0.0135835912078619, + 0.028068965300917625, + 0.02705204114317894, + 0.00413266196846962, + 0.008328139781951904, + 0.008719510398805141, + 0.003513262141495943, + 0.003781548235565424, + 0.004877160768955946, + 0.006704798899590969, + 0.006608906667679548, + 0.005272237118333578, + 0.007169454824179411, + 0.006489027291536331, + 0.0048132846131920815, + 0.003944478929042816, + 0.00831955298781395, + 0.007078876718878746, + 0.008765244856476784, + 0.007712732069194317, + 0.009889231994748116, + 0.005728508811444044, + 0.006716609001159668, + 0.003937508445233107, + 0.008779320865869522, + 0.00239050155505538, + 0.0030328072607517242, + 0.004736592993140221, + 0.005737078841775656, + 0.004141691606491804, + 0.003892882727086544, + 0.0036042043939232826, + 0.006616286002099514, + 0.005420002155005932, + 0.00725129246711731, + 0.015747658908367157, + 0.01357616949826479, + 0.005049490835517645, + 0.005601033102720976, + 0.0049842712469398975, + 0.014161726459860802, + 0.0105298375710845, + 0.008313065394759178, + 0.029370497912168503, + 0.02418440394103527, + 0.013083218596875668, + 0.0053938766941428185, + 0.008626890368759632, + 0.026956496760249138, + 0.008001473732292652, + 0.009432822465896606, + 0.026053663343191147, + 0.009498491883277893, + 0.007470743730664253, + 0.004431379493325949, + 0.008695105090737343, + 0.01708868145942688, + 0.008659443818032742, + 0.010810825042426586, + 0.005971300881356001, + 0.004589763469994068, + 0.004228837322443724, + 0.007014282513409853, + 0.004011018201708794, + 0.005579497665166855 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 7, + 1 + ], + "selected_reward": -0.0032010050490498543 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an elderly villager with graying hair in a bun and deep wrinkles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the setting a newly green rice paddy with rolling green mountains in the distance?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera start with a wide long shot and slowly pan forward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the villager raise her arms toward the sky and then perform a gentle twirl?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the lighting golden morning light casting a warm halo and illuminating dewdrops?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.387034147977829 + ], + "opposite_logprob_tokens": [ + -1.1370341777801514 + ], + "logprob_reward": -0.387034147977829, + "yes_no_margin_reward": 0.75, + "token_credit_reward": -0.387034147977829, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.16272921860218048, + 0.09502623975276947, + 0.10204613208770752, + 0.039721839129924774, + 0.025041718035936356, + 0.06005796417593956, + 0.1690041422843933, + 0.09520100802183151, + 0.19154387712478638, + 0.05962783098220825 + ], + "credit_map": [ + 0.02415391430258751, + 0.019700922071933746, + 0.03777248039841652, + 0.017050165683031082, + 0.012431754730641842, + 0.011122669093310833, + 0.012466223910450935, + 0.028031084686517715, + 0.01325663086026907, + 0.00961109809577465, + 0.010072238743305206, + 0.01676252856850624, + 0.01771361567080021, + 0.007020236924290657, + 0.009054461494088173, + 0.011535427533090115, + 0.017886755988001823, + 0.014371348544955254, + 0.023766016587615013, + 0.008226358331739902, + 0.014007135294377804, + 0.005593865644186735, + 0.007095795124769211, + 0.01109885610640049, + 0.005421753972768784, + 0.006304266396909952, + 0.005139879882335663, + 0.0042173839174211025, + 0.005594454240053892, + 0.0035835406742990017, + 0.0041671935468912125, + 0.005293367896229029, + 0.0026109479367733, + 0.0032659154385328293, + 0.003085644915699959, + 0.004095786716789007, + 0.003991888370364904, + 0.001809799112379551, + 0.00233712256886065, + 0.0038446139078587294, + 0.0052636186592280865, + 0.005965705029666424, + 0.006719541270285845, + 0.007983241230249405, + 0.012173709459602833, + 0.004200557246804237, + 0.007016933057457209, + 0.010734658688306808, + 0.013791455887258053, + 0.011880542151629925, + 0.04153018444776535, + 0.027495160698890686, + 0.023324303328990936, + 0.010334408842027187, + 0.014238815754652023, + 0.026409272104501724, + 0.006084492430090904, + 0.005853292066603899, + 0.022530201822519302, + 0.017484135925769806, + 0.010754692368209362, + 0.004476245958358049, + 0.012949350290000439, + 0.015068593434989452, + 0.027115892618894577, + 0.03042389266192913, + 0.026634052395820618, + 0.02395440824329853, + 0.01589817926287651, + 0.012494214810431004, + 0.019767461344599724, + 0.03525577858090401, + 0.00810597650706768, + 0.008135974407196045, + 0.00664651021361351, + 0.004167637787759304, + 0.01436323206871748, + 0.005735511891543865, + 0.006909333635121584, + 0.0055636572651565075 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 6, + 0 + ], + "selected_reward": -0.387034147977829 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an elderly villager with graying hair in a bun and deep wrinkles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the setting a newly green rice paddy with rolling green mountains in the distance?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera start with a wide long shot and slowly pan forward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the villager raise her arms toward the sky and then perform a gentle twirl?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the lighting golden morning light casting a warm halo and illuminating dewdrops?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01607177034020424 + ], + "opposite_logprob_tokens": [ + -4.141071796417236 + ], + "logprob_reward": -0.01607177034020424, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.01607177034020424, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07414595782756805, + 0.0754522979259491, + 0.1140509843826294, + 0.019098324701189995, + 0.016019754111766815, + 0.02771904319524765, + 0.18113043904304504, + 0.16664724051952362, + 0.2777518928050995, + 0.04798407852649689 + ], + "credit_map": [ + 0.006630655378103256, + 0.006771918851882219, + 0.01821126416325569, + 0.01271531730890274, + 0.008828112855553627, + 0.004368740599602461, + 0.006244946271181107, + 0.010374999605119228, + 0.009668763726949692, + 0.010003997944295406, + 0.006200314965099096, + 0.007796557154506445, + 0.019714856520295143, + 0.005865932907909155, + 0.007905663922429085, + 0.008296209387481213, + 0.026058891788125038, + 0.03268943354487419, + 0.018642328679561615, + 0.0064034354873001575, + 0.011679662391543388, + 0.004091399721801281, + 0.006274333223700523, + 0.008211498148739338, + 0.00193876086268574, + 0.00261557474732399, + 0.0017995680682361126, + 0.0026418378110975027, + 0.00496214022859931, + 0.0018139370949938893, + 0.0016973308520391583, + 0.0016291745705530047, + 0.0016311485087499022, + 0.002152196131646633, + 0.002117978176102042, + 0.0019833215046674013, + 0.0029039171058684587, + 0.00118661776650697, + 0.0019768571946769953, + 0.002067715860903263, + 0.002334534889087081, + 0.004010038450360298, + 0.003039821982383728, + 0.0036827572621405125, + 0.005870526656508446, + 0.0024156318977475166, + 0.003920746501535177, + 0.0024449853226542473, + 0.008107985369861126, + 0.0073625510558485985, + 0.06275101751089096, + 0.0310087688267231, + 0.005416858941316605, + 0.007790567819029093, + 0.008290424942970276, + 0.050402265042066574, + 0.01742740534245968, + 0.016144167631864548, + 0.022824592888355255, + 0.015887951478362083, + 0.03023824468255043, + 0.010349671356379986, + 0.015205664560198784, + 0.03856954351067543, + 0.0396728441119194, + 0.04676096886396408, + 0.04990726709365845, + 0.04107176512479782, + 0.021853778511285782, + 0.01846116967499256, + 0.027635252103209496, + 0.03238886222243309, + 0.006529463455080986, + 0.005129901226609945, + 0.011133046820759773, + 0.004078458063304424, + 0.006530805956572294, + 0.002700509736314416, + 0.005508562549948692, + 0.0063733323477208614 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 6, + 7 + ], + "selected_reward": -0.01607177034020424 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an elderly villager with graying hair in a bun and deep wrinkles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the setting a newly green rice paddy with rolling green mountains in the distance?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera start with a wide long shot and slowly pan forward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the villager raise her arms toward the sky and then perform a gentle twirl?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the lighting golden morning light casting a warm halo and illuminating dewdrops?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03811784088611603 + ], + "opposite_logprob_tokens": [ + -3.2881178855895996 + ], + "logprob_reward": -0.03811784088611603, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.03811784088611603, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3819054365158081, + 0.07191933691501617, + 0.10894601792097092, + 0.03724981099367142, + 0.07755046337842941, + 0.033879250288009644, + 0.06584524363279343, + 0.05733725428581238, + 0.10624620318412781, + 0.05912092328071594 + ], + "credit_map": [ + 0.0598326213657856, + 0.056612931191921234, + 0.06662198156118393, + 0.04928644746541977, + 0.06446365267038345, + 0.018467310816049576, + 0.028169073164463043, + 0.038451433181762695, + 0.0099534522742033, + 0.007062782067805529, + 0.009773477911949158, + 0.01669272407889366, + 0.0066544064320623875, + 0.0057886564172804356, + 0.0058665345422923565, + 0.01012730784714222, + 0.026523564010858536, + 0.03227237984538078, + 0.020864229649305344, + 0.006860305089503527, + 0.006671003066003323, + 0.002106491941958666, + 0.0047394591383636, + 0.00890858843922615, + 0.0032241512089967728, + 0.003598113777115941, + 0.005314345937222242, + 0.006385795772075653, + 0.006431246642023325, + 0.003190114162862301, + 0.004324100911617279, + 0.00478194747120142, + 0.007541902828961611, + 0.008134691044688225, + 0.0113517539575696, + 0.0162507351487875, + 0.009352969005703926, + 0.007495465222746134, + 0.005179210100322962, + 0.012243740260601044, + 0.003056590212509036, + 0.004517791792750359, + 0.003034994238987565, + 0.00478333467617631, + 0.007655350491404533, + 0.002733507426455617, + 0.004134016577154398, + 0.003963666968047619, + 0.005611531902104616, + 0.004130931571125984, + 0.017275920137763023, + 0.007920214906334877, + 0.00469440221786499, + 0.004482812713831663, + 0.0035059372894465923, + 0.01822349987924099, + 0.009642224758863449, + 0.0063451314345002174, + 0.006599383894354105, + 0.007345681078732014, + 0.005651287268847227, + 0.0032237728592008352, + 0.004752682987600565, + 0.013777092099189758, + 0.008577406406402588, + 0.006919138133525848, + 0.03947455808520317, + 0.013685454614460468, + 0.00797386933118105, + 0.004249072168022394, + 0.006973145063966513, + 0.018393564969301224, + 0.009718003682792187, + 0.008097358979284763, + 0.01429902296513319, + 0.00591092836111784, + 0.008310464210808277, + 0.003451138036325574, + 0.004652919247746468, + 0.004681091755628586 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 8 + ], + "selected_reward": -0.03811784088611603 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000071/group_1/index_tensor([5722, 5722])_group_1_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000071/group_1/index_tensor([5722, 5722])_group_1_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..206fe19ba58e6411ca98b159d1660758b158db63 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000071/group_1/index_tensor([5722, 5722])_group_1_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "slow, smooth, and slightly bird's-eye view of the drone gracefully accompanies a bright yellow two-person kayak as it leisurely paddles across the crystal-clear turquoise waters of a high mountain lake. The lake's surface, as tranquil as a mirror, perfectly reflects the surrounding jagged, majestic snow-capped peaks and the lush, deep green evergreen forests lining its shores, making the line between reality and reflection almost indistinguishable. Inside, two friends paddle in a fluid, synchronized rhythm. The woman in the front, her blond ponytail peeking from beneath a dark gray beanie, wears a vibrant turquoise waterproof jacket and a calm, focused expression. Behind her, her male companion, clad in a burnt orange fleece pullover and clutching a black carbon fiber paddle, glances up from time to time with a smile of awe at the majestic mountain scenery. Their oars dipped almost silently into the water, the only things breaking the glassy surface were tiny concentric ripples that spread out from where the oars entered the water, and an almost imperceptible breeze that stirred a subtle shimmer across the surface of the lake around the boats. The soft golden afternoon sunlight bathed the scene, warming the textures of the snow-capped peaks and making the water shimmer, creating an atmosphere of grandeur, tranquility, and profound tranquility.", + "seed": 22108, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000071/group_1/index_5722_seed_22108_dp_2_rank_4_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 1, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.024258521432057024, + "vlm_logprob": -0.024258521432057024, + "vlm_yes_no_margin": 4.124999952316284, + "vlm_token_credit": -0.024258521432057024, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera angle a slightly elevated bird's-eye view?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a bright yellow two-person kayak visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are two people paddling in a synchronized rhythm inside the kayak?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the woman in the front wearing a dark gray beanie and a vibrant turquoise waterproof jacket?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the man behind her wearing a burnt orange fleece pullover and holding a black paddle?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04780564829707146 + ], + "opposite_logprob_tokens": [ + -3.2978055477142334 + ], + "logprob_reward": -0.04780564829707146, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.04780564829707146, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.45973291993141174, + 0.16755612194538116, + 0.02121371030807495, + 0.06899949908256531, + 0.052984680980443954, + 0.017635170370340347, + 0.07250630855560303, + 0.06636877357959747, + 0.03896920010447502, + 0.03403366729617119 + ], + "credit_map": [ + 0.01918799802660942, + 0.026774615049362183, + 0.024378228932619095, + 0.01820998825132847, + 0.028408348560333252, + 0.1340288519859314, + 0.03078114427626133, + 0.1779637634754181, + 0.010128166526556015, + 0.009779484011232853, + 0.05293186008930206, + 0.009711993858218193, + 0.006741353310644627, + 0.01584690436720848, + 0.007236026227474213, + 0.055180322378873825, + 0.00334074255079031, + 0.0024361812975257635, + 0.0031754588708281517, + 0.0024339971132576466, + 0.0019322247244417667, + 0.0031383540481328964, + 0.0023952587507665157, + 0.0023614931851625443, + 0.007702097296714783, + 0.00507658114656806, + 0.01018255203962326, + 0.006600289139896631, + 0.008822571486234665, + 0.007861258462071419, + 0.008481807075440884, + 0.014272334054112434, + 0.005648475140333176, + 0.0049575623124837875, + 0.0060504344291985035, + 0.006344558205455542, + 0.004730915650725365, + 0.008767942897975445, + 0.004496775567531586, + 0.011988012120127678, + 0.0018614060245454311, + 0.0011882069520652294, + 0.001610824023373425, + 0.001314331777393818, + 0.0021530066151171923, + 0.003594190813601017, + 0.0014581838622689247, + 0.004455021116882563, + 0.0078083244152367115, + 0.005805154796689749, + 0.0078734727576375, + 0.0067184907384216785, + 0.014668626710772514, + 0.009892617352306843, + 0.012943942099809647, + 0.0067956699058413506, + 0.006258232053369284, + 0.005215445999056101, + 0.007138861808925867, + 0.00902676023542881, + 0.016993781551718712, + 0.008997703902423382, + 0.005196575541049242, + 0.007541409693658352, + 0.0038380499463528395, + 0.002866063266992569, + 0.006681526079773903, + 0.004512176848948002, + 0.005747795104980469, + 0.004903365857899189, + 0.0061835600063204765, + 0.004236659966409206, + 0.004460844211280346, + 0.003019146155565977, + 0.0033663995563983917, + 0.004466304089874029, + 0.0046868640929460526, + 0.005110146943479776, + 0.004542033188045025, + 0.004381929524242878 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 6 + ], + "selected_reward": -0.04780564829707146 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera angle a slightly elevated bird's-eye view?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a bright yellow two-person kayak visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are two people paddling in a synchronized rhythm inside the kayak?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the woman in the front wearing a dark gray beanie and a vibrant turquoise waterproof jacket?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the man behind her wearing a burnt orange fleece pullover and holding a black paddle?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.002219476969912648 + ], + "opposite_logprob_tokens": [ + -6.1272196769714355 + ], + "logprob_reward": -0.002219476969912648, + "yes_no_margin_reward": 6.125, + "token_credit_reward": -0.002219476969912648, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.29965075850486755, + 0.4017069935798645, + 0.06645982712507248, + 0.05038408190011978, + 0.029839521273970604, + 0.025221263989806175, + 0.017642410472035408, + 0.0215778648853302, + 0.03153533861041069, + 0.05598193407058716 + ], + "credit_map": [ + 0.01774624176323414, + 0.015907257795333862, + 0.05808045342564583, + 0.022488651797175407, + 0.01828589290380478, + 0.04203736037015915, + 0.02736523374915123, + 0.09773971140384674, + 0.0205309446901083, + 0.021101927384734154, + 0.12134112417697906, + 0.021588217467069626, + 0.015740888193249702, + 0.039787326008081436, + 0.0180863868445158, + 0.1435302197933197, + 0.007327335420995951, + 0.007178038824349642, + 0.01508351694792509, + 0.00569411413744092, + 0.007770347408950329, + 0.007621277589350939, + 0.008581914938986301, + 0.007203289307653904, + 0.006450408138334751, + 0.004408380948007107, + 0.00884886085987091, + 0.004863034002482891, + 0.005377244669944048, + 0.006405479274690151, + 0.007020954508334398, + 0.0070097255520522594, + 0.0025799013674259186, + 0.0022687390446662903, + 0.0030191477853804827, + 0.003150711767375469, + 0.00326430331915617, + 0.005399858113378286, + 0.0036144459154456854, + 0.006542416289448738, + 0.0020718658342957497, + 0.0015438684495165944, + 0.0020544242579489946, + 0.0017493024934083223, + 0.0018877453403547406, + 0.00661154231056571, + 0.0023993290960788727, + 0.006903189700096846, + 0.0020989372860640287, + 0.0015112824039533734, + 0.002290199976414442, + 0.0016171348979696631, + 0.003558030119165778, + 0.0019236654043197632, + 0.002580021508038044, + 0.0020631395746022463, + 0.003127175383269787, + 0.0017995478119701147, + 0.002947642933577299, + 0.002427757019177079, + 0.0023218190763145685, + 0.0041495440527796745, + 0.002205523196607828, + 0.002598857507109642, + 0.003691211575642228, + 0.002571715973317623, + 0.003812455339357257, + 0.00327680422924459, + 0.00487609812989831, + 0.005863507743924856, + 0.004721042234450579, + 0.0027225036174058914, + 0.0056772977113723755, + 0.004190483596175909, + 0.004876221530139446, + 0.004924517124891281, + 0.005376595538109541, + 0.010074262507259846, + 0.005413881968706846, + 0.015448681078851223 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 2 + ], + "selected_reward": -0.002219476969912648 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera angle a slightly elevated bird's-eye view?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a bright yellow two-person kayak visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are two people paddling in a synchronized rhythm inside the kayak?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the woman in the front wearing a dark gray beanie and a vibrant turquoise waterproof jacket?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the man behind her wearing a burnt orange fleece pullover and holding a black paddle?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.029800519347190857 + ], + "opposite_logprob_tokens": [ + -3.5298004150390625 + ], + "logprob_reward": -0.029800519347190857, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.029800519347190857, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.44639864563941956, + 0.0972239151597023, + 0.079267717897892, + 0.09125670045614243, + 0.04600629583001137, + 0.014894473366439342, + 0.029888853430747986, + 0.05777604132890701, + 0.06036939099431038, + 0.07691790163516998 + ], + "credit_map": [ + 0.01978294365108013, + 0.0276956669986248, + 0.060882389545440674, + 0.02489314414560795, + 0.02446078509092331, + 0.125662699341774, + 0.033990368247032166, + 0.12903064489364624, + 0.005010505206882954, + 0.0047531467862427235, + 0.020421715453267097, + 0.0054224166087806225, + 0.005693543702363968, + 0.014086590148508549, + 0.007060085888952017, + 0.03477591276168823, + 0.011982272379100323, + 0.009027644991874695, + 0.020226944237947464, + 0.006256659049540758, + 0.006489147432148457, + 0.009962372481822968, + 0.007259981241077185, + 0.008062693290412426, + 0.005874096415936947, + 0.00715955626219511, + 0.012061120010912418, + 0.00700747175142169, + 0.006716374773532152, + 0.022880034521222115, + 0.0071179517544806, + 0.022440094500780106, + 0.004884869791567326, + 0.005044175311923027, + 0.004846910480409861, + 0.004717388190329075, + 0.00535705266520381, + 0.006427014246582985, + 0.00418197363615036, + 0.0105469124391675, + 0.0014675369020551443, + 0.0010466011008247733, + 0.001996189123019576, + 0.0012965102214366198, + 0.0016305249882861972, + 0.0018833341309800744, + 0.002255138708278537, + 0.0033186376094818115, + 0.0026252148672938347, + 0.003231630427762866, + 0.002711247419938445, + 0.003911226522177458, + 0.005439163651317358, + 0.00325203500688076, + 0.005587561056017876, + 0.003130774013698101, + 0.006715731229633093, + 0.004452724941074848, + 0.006560655776411295, + 0.005733765661716461, + 0.008820698596537113, + 0.010613956488668919, + 0.007503438740968704, + 0.0073750694282352924, + 0.006112983450293541, + 0.004965163767337799, + 0.00810862984508276, + 0.0067215571179986, + 0.00992046482861042, + 0.007989257574081421, + 0.010783685371279716, + 0.005767648573964834, + 0.007818859070539474, + 0.0063242679461836815, + 0.007372280582785606, + 0.008568346500396729, + 0.007045138627290726, + 0.01589915342628956, + 0.008400038816034794, + 0.015489816665649414 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 3 + ], + "selected_reward": -0.029800519347190857 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera angle a slightly elevated bird's-eye view?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a bright yellow two-person kayak visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are two people paddling in a synchronized rhythm inside the kayak?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the woman in the front wearing a dark gray beanie and a vibrant turquoise waterproof jacket?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the man behind her wearing a burnt orange fleece pullover and holding a black paddle?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0181814506649971 + ], + "opposite_logprob_tokens": [ + -4.018181324005127 + ], + "logprob_reward": -0.0181814506649971, + "yes_no_margin_reward": 3.999999761581421, + "token_credit_reward": -0.0181814506649971, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.24507981538772583, + 0.25684356689453125, + 0.06409439444541931, + 0.05360350385308266, + 0.03477149084210396, + 0.04356342554092407, + 0.14837278425693512, + 0.052122972905635834, + 0.05520830303430557, + 0.0463397353887558 + ], + "credit_map": [ + 0.012162245810031891, + 0.018467584624886513, + 0.05448222905397415, + 0.014759715646505356, + 0.014996089972555637, + 0.03718377277255058, + 0.015918586403131485, + 0.07710958272218704, + 0.013705240562558174, + 0.015130133368074894, + 0.07702445238828659, + 0.014233899302780628, + 0.010589350014925003, + 0.025257717818021774, + 0.010862285271286964, + 0.09004047513008118, + 0.008336358703672886, + 0.008120257407426834, + 0.006916198413819075, + 0.007956498302519321, + 0.010130640119314194, + 0.008038868196308613, + 0.007473514415323734, + 0.0071220602840185165, + 0.006176828872412443, + 0.004641375504434109, + 0.011498299427330494, + 0.004502743016928434, + 0.006683376617729664, + 0.00979922991245985, + 0.004420188255608082, + 0.005881461780518293, + 0.0025910602416843176, + 0.0026730364188551903, + 0.003183545544743538, + 0.0032263516914099455, + 0.0037769500631839037, + 0.007260928396135569, + 0.0029381958302110434, + 0.009121423587203026, + 0.0037801116704940796, + 0.002975720213726163, + 0.003855442861095071, + 0.0027418993413448334, + 0.0041473316960036755, + 0.010288436897099018, + 0.004398739896714687, + 0.011375744827091694, + 0.01586727984249592, + 0.012752275913953781, + 0.01671304740011692, + 0.013295182958245277, + 0.03413264453411102, + 0.019540684297680855, + 0.02361050248146057, + 0.012461178004741669, + 0.005803863517940044, + 0.003678536042571068, + 0.007159252651035786, + 0.007607119623571634, + 0.009821275249123573, + 0.005848706234246492, + 0.007039298769086599, + 0.005164919421076775, + 0.004926901776343584, + 0.003992064390331507, + 0.005997414700686932, + 0.004890087526291609, + 0.009306172840297222, + 0.01092058140784502, + 0.00934159941971302, + 0.005833480507135391, + 0.007916093803942204, + 0.0032286453060805798, + 0.003987989854067564, + 0.0036822755355387926, + 0.0045083919540047646, + 0.004751745145767927, + 0.006120976060628891, + 0.012143615633249283 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 6 + ], + "selected_reward": -0.0181814506649971 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera angle a slightly elevated bird's-eye view?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a bright yellow two-person kayak visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are two people paddling in a synchronized rhythm inside the kayak?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the woman in the front wearing a dark gray beanie and a vibrant turquoise waterproof jacket?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the man behind her wearing a burnt orange fleece pullover and holding a black paddle?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.023285511881113052 + ], + "opposite_logprob_tokens": [ + -3.7732856273651123 + ], + "logprob_reward": -0.023285511881113052, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.023285511881113052, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2320283055305481, + 0.2447628378868103, + 0.030544601380825043, + 0.0886983647942543, + 0.03366806358098984, + 0.023705370724201202, + 0.03021407686173916, + 0.0553557425737381, + 0.13497188687324524, + 0.12605074048042297 + ], + "credit_map": [ + 0.01177113875746727, + 0.014912840910255909, + 0.06439055502414703, + 0.0175168514251709, + 0.013733625411987305, + 0.023193148896098137, + 0.019007263705134392, + 0.06750288605690002, + 0.012377865612506866, + 0.011958315968513489, + 0.06942401081323624, + 0.012697377242147923, + 0.009700420312583447, + 0.027127616107463837, + 0.011992285959422588, + 0.08948495239019394, + 0.003146155970171094, + 0.0033549736253917217, + 0.005372840445488691, + 0.0031307144090533257, + 0.004919256083667278, + 0.003122232621535659, + 0.003610832616686821, + 0.003887598169967532, + 0.007183259818702936, + 0.0066238222643733025, + 0.012128209695219994, + 0.0059227352030575275, + 0.00780893350020051, + 0.020514309406280518, + 0.006737987510859966, + 0.021779118105769157, + 0.0028104125522077084, + 0.002885326510295272, + 0.0031896554864943027, + 0.0036710849963128567, + 0.003794731106609106, + 0.00602392852306366, + 0.003116344567388296, + 0.008176579140126705, + 0.0020194719545543194, + 0.0018287490820512176, + 0.0027165573555976152, + 0.0022639185190200806, + 0.004703914280980825, + 0.002894564298912883, + 0.0044289943762123585, + 0.0028492009732872248, + 0.0034378820564597845, + 0.0029936593491584063, + 0.00415787473320961, + 0.0045582023449242115, + 0.0033452201168984175, + 0.003355399239808321, + 0.004157835617661476, + 0.004208005033433437, + 0.006440526805818081, + 0.004618184641003609, + 0.006747512146830559, + 0.006484283600002527, + 0.012909711338579655, + 0.007031935267150402, + 0.0055631063878536224, + 0.005560486111789942, + 0.015622296370565891, + 0.008987333625555038, + 0.013419382274150848, + 0.010821565054357052, + 0.02061678282916546, + 0.031089268624782562, + 0.020742041990160942, + 0.01367320865392685, + 0.017227090895175934, + 0.00877054687589407, + 0.009455088526010513, + 0.008907608687877655, + 0.012087499722838402, + 0.03339120373129845, + 0.01189426053315401, + 0.024317458271980286 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 8 + ], + "selected_reward": -0.023285511881113052 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "slow, smooth, and slightly bird's-eye view of the drone gracefully accompanies a bright yellow two-person kayak as it leisurely paddles across the crystal-clear turquoise waters of a high mountain lake. The lake's surface, as tranquil as a mirror, perfectly reflects the surrounding jagged, majestic snow-capped peaks and the lush, deep green evergreen forests lining its shores, making the line between reality and reflection almost indistinguishable. Inside, two friends paddle in a fluid, synchronized rhythm. The woman in the front, her blond ponytail peeking from beneath a dark gray beanie, wears a vibrant turquoise waterproof jacket and a calm, focused expression. Behind her, her male companion, clad in a burnt orange fleece pullover and clutching a black carbon fiber paddle, glances up from time to time with a smile of awe at the majestic mountain scenery. Their oars dipped almost silently into the water, the only things breaking the glassy surface were tiny concentric ripples that spread out from where the oars entered the water, and an almost imperceptible breeze that stirred a subtle shimmer across the surface of the lake around the boats. The soft golden afternoon sunlight bathed the scene, warming the textures of the snow-capped peaks and making the water shimmer, creating an atmosphere of grandeur, tranquility, and profound tranquility.", + "seed": 22109, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000071/group_1/index_5722_seed_22109_dp_2_rank_4_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 1, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.03616128807188943, + "vlm_logprob": -0.03616128807188943, + "vlm_yes_no_margin": 4.175, + "vlm_token_credit": -0.03616128807188943, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera angle a slightly elevated bird's-eye view?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a bright yellow two-person kayak visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are two people paddling in a synchronized rhythm inside the kayak?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the woman in the front wearing a dark gray beanie and a vibrant turquoise waterproof jacket?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the man behind her wearing a burnt orange fleece pullover and holding a black paddle?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07791396975517273 + ], + "opposite_logprob_tokens": [ + -2.827913999557495 + ], + "logprob_reward": -0.07791396975517273, + "yes_no_margin_reward": 2.75, + "token_credit_reward": -0.07791396975517273, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.6182641983032227, + 0.026873603463172913, + 0.06408559530973434, + 0.012424430809915066, + 0.06708932667970657, + 0.03056059218943119, + 0.0495566688477993, + 0.05252384766936302, + 0.06238425150513649, + 0.016237501055002213 + ], + "credit_map": [ + 0.09722970426082611, + 0.013704030774533749, + 0.16657127439975739, + 0.021305708214640617, + 0.022632472217082977, + 0.20110340416431427, + 0.0173176359385252, + 0.07839998602867126, + 0.005770978983491659, + 0.0027601353358477354, + 0.002331333700567484, + 0.0015954850241541862, + 0.002124600810930133, + 0.008362418040633202, + 0.0021485518664121628, + 0.0017800999339669943, + 0.005364470183849335, + 0.005707313772290945, + 0.006371959112584591, + 0.009906606748700142, + 0.010913494043052197, + 0.006462065037339926, + 0.012586576864123344, + 0.00677311047911644, + 0.0017844198737293482, + 0.0015511625679209828, + 0.0013011593837291002, + 0.0017929227324202657, + 0.0016586908604949713, + 0.0014317735331133008, + 0.0016980089712888002, + 0.0012062931200489402, + 0.006036971230059862, + 0.006763630546629429, + 0.004504309967160225, + 0.009001336991786957, + 0.007493910379707813, + 0.010158509947359562, + 0.013486827723681927, + 0.009643837809562683, + 0.0036911501083523035, + 0.0037952186539769173, + 0.0023980040568858385, + 0.0032447490375488997, + 0.004730232059955597, + 0.0036157749127596617, + 0.0052727218717336655, + 0.0038127428852021694, + 0.006276268977671862, + 0.0069967578165233135, + 0.004768002778291702, + 0.004736361093819141, + 0.012020031921565533, + 0.0065302313305437565, + 0.005137735512107611, + 0.0030912794172763824, + 0.011620152741670609, + 0.006196241360157728, + 0.00596161512658, + 0.004884818568825722, + 0.008454613387584686, + 0.005426851566880941, + 0.005356999114155769, + 0.0046225558035075665, + 0.004846405703574419, + 0.0072008538991212845, + 0.0035874820314347744, + 0.005331007298082113, + 0.01509032305330038, + 0.012212528847157955, + 0.009855372831225395, + 0.0042602852918207645, + 0.0023317497689276934, + 0.001925514661706984, + 0.001521677477285266, + 0.0021496734116226435, + 0.0027611448895186186, + 0.00177689956035465, + 0.0016927467659115791, + 0.0020780949853360653 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 2 + ], + "selected_reward": -0.07791396975517273 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera angle a slightly elevated bird's-eye view?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a bright yellow two-person kayak visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are two people paddling in a synchronized rhythm inside the kayak?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the woman in the front wearing a dark gray beanie and a vibrant turquoise waterproof jacket?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the man behind her wearing a burnt orange fleece pullover and holding a black paddle?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0007485924870707095 + ], + "opposite_logprob_tokens": [ + -7.250748634338379 + ], + "logprob_reward": -0.0007485924870707095, + "yes_no_margin_reward": 7.25, + "token_credit_reward": -0.0007485924870707095, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13419923186302185, + 0.03368112072348595, + 0.10694028437137604, + 0.026108352467417717, + 0.09560348093509674, + 0.05514782667160034, + 0.03085549920797348, + 0.43573716282844543, + 0.050956785678863525, + 0.03077024407684803 + ], + "credit_map": [ + 0.026598382741212845, + 0.004723004065454006, + 0.021551379933953285, + 0.006712454836815596, + 0.006320144049823284, + 0.05063244700431824, + 0.007779441773891449, + 0.009881976991891861, + 0.005419427528977394, + 0.002781769959256053, + 0.0027872445061802864, + 0.003496038494631648, + 0.003361204406246543, + 0.0069474163465201855, + 0.006357760634273291, + 0.002530258148908615, + 0.008946090005338192, + 0.011323108337819576, + 0.012449458241462708, + 0.01919650472700596, + 0.011836505495011806, + 0.010024745017290115, + 0.016429362818598747, + 0.01673450879752636, + 0.005048736929893494, + 0.004787337966263294, + 0.0033759060315787792, + 0.002474077045917511, + 0.0026949585881084204, + 0.0033487144391983747, + 0.0020998562686145306, + 0.002278766129165888, + 0.011002602986991405, + 0.01238447055220604, + 0.007231836207211018, + 0.01603088155388832, + 0.01014803722500801, + 0.012379518710076809, + 0.01423546951264143, + 0.01219066884368658, + 0.005573130678385496, + 0.005428101867437363, + 0.005093730054795742, + 0.008875171653926373, + 0.009117268025875092, + 0.008521868847310543, + 0.006117958575487137, + 0.006420597899705172, + 0.0036268422845751047, + 0.0031969889532774687, + 0.002890173811465502, + 0.003010503714904189, + 0.0056826276704669, + 0.0035155368968844414, + 0.004104949999600649, + 0.0048278761096298695, + 0.09958507865667343, + 0.04879005253314972, + 0.0510881282389164, + 0.046489689499139786, + 0.07458112388849258, + 0.04322712868452072, + 0.03968258947134018, + 0.03229338303208351, + 0.004501203540712595, + 0.006510403007268906, + 0.0030217301100492477, + 0.003307259175926447, + 0.01370514277368784, + 0.011469319462776184, + 0.005858856253325939, + 0.0025828727521002293, + 0.0023042347747832537, + 0.0031926468946039677, + 0.0028445501811802387, + 0.005309544038027525, + 0.005599225405603647, + 0.004026030655950308, + 0.0042586661875247955, + 0.003235345706343651 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 0, + 2 + ], + "selected_reward": -0.0007485924870707095 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera angle a slightly elevated bird's-eye view?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a bright yellow two-person kayak visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are two people paddling in a synchronized rhythm inside the kayak?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the woman in the front wearing a dark gray beanie and a vibrant turquoise waterproof jacket?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the man behind her wearing a burnt orange fleece pullover and holding a black paddle?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07001570612192154 + ], + "opposite_logprob_tokens": [ + -2.6950156688690186 + ], + "logprob_reward": -0.07001570612192154, + "yes_no_margin_reward": 2.625, + "token_credit_reward": -0.07001570612192154, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.4214348793029785, + 0.05100996047258377, + 0.04791457578539848, + 0.051169488579034805, + 0.1187281534075737, + 0.047717079520225525, + 0.07151968032121658, + 0.10311755537986755, + 0.06256306171417236, + 0.02482556365430355 + ], + "credit_map": [ + 0.062123194336891174, + 0.007069087587296963, + 0.12244571000337601, + 0.016217291355133057, + 0.014683116227388382, + 0.13046666979789734, + 0.013130314648151398, + 0.05529946833848953, + 0.007382753770798445, + 0.006708846427500248, + 0.003736024722456932, + 0.004506411962211132, + 0.004604123532772064, + 0.012315826490521431, + 0.0073832194320857525, + 0.00437275180593133, + 0.005198013968765736, + 0.006761118303984404, + 0.0059610470198094845, + 0.005859899800270796, + 0.007261024322360754, + 0.005848861299455166, + 0.006473054178059101, + 0.004551555495709181, + 0.005109566263854504, + 0.005255111493170261, + 0.006686786655336618, + 0.005648937541991472, + 0.005899571347981691, + 0.007147555239498615, + 0.005842871963977814, + 0.009579083882272243, + 0.017577646300196648, + 0.014581813476979733, + 0.008617517538368702, + 0.012372342869639397, + 0.012771288864314556, + 0.01647319830954075, + 0.020448079332709312, + 0.01588626019656658, + 0.0042031328193843365, + 0.0062291887588799, + 0.0034717805683612823, + 0.005511108785867691, + 0.006479965057224035, + 0.0057591949589550495, + 0.009662344120442867, + 0.00640035979449749, + 0.009255201555788517, + 0.008735430426895618, + 0.005236085969954729, + 0.006532374769449234, + 0.021377814933657646, + 0.009645598009228706, + 0.005892907734960318, + 0.00484426598995924, + 0.021046658977866173, + 0.010977785103023052, + 0.0106132160872221, + 0.008971705101430416, + 0.019617829471826553, + 0.011952429078519344, + 0.013216113671660423, + 0.006721806712448597, + 0.006170849781483412, + 0.005091235972940922, + 0.003776237601414323, + 0.006106764543801546, + 0.011792514473199844, + 0.012089206837117672, + 0.011994685977697372, + 0.005541566293686628, + 0.002559884451329708, + 0.003179187187924981, + 0.0018613348947837949, + 0.002594016259536147, + 0.004829266108572483, + 0.003248263383284211, + 0.002815134124830365, + 0.003738475264981389 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 7 + ], + "selected_reward": -0.07001570612192154 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera angle a slightly elevated bird's-eye view?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a bright yellow two-person kayak visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are two people paddling in a synchronized rhythm inside the kayak?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the woman in the front wearing a dark gray beanie and a vibrant turquoise waterproof jacket?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the man behind her wearing a burnt orange fleece pullover and holding a black paddle?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.016062267124652863 + ], + "opposite_logprob_tokens": [ + -4.141062259674072 + ], + "logprob_reward": -0.016062267124652863, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.016062267124652863, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.18864113092422485, + 0.06901007145643234, + 0.07086123526096344, + 0.04581640288233757, + 0.09600004553794861, + 0.09168544411659241, + 0.11062518507242203, + 0.27090373635292053, + 0.025078481063246727, + 0.03137829527258873 + ], + "credit_map": [ + 0.01993374153971672, + 0.010853738524019718, + 0.05938402935862541, + 0.012447378598153591, + 0.009173233062028885, + 0.029387524351477623, + 0.011324835941195488, + 0.036136627197265625, + 0.013329165987670422, + 0.004859642591327429, + 0.004195149522274733, + 0.004719672258943319, + 0.005154536571353674, + 0.023326359689235687, + 0.009363039396703243, + 0.004062501713633537, + 0.009896024130284786, + 0.012637995183467865, + 0.008787938393652439, + 0.005973244551569223, + 0.01190438773483038, + 0.006747187115252018, + 0.00774230994284153, + 0.007172142621129751, + 0.0064907134510576725, + 0.004510440398007631, + 0.003525535576045513, + 0.004783584736287594, + 0.007561107166111469, + 0.005562418140470982, + 0.008311313576996326, + 0.005071286112070084, + 0.009311949834227562, + 0.007657586131244898, + 0.0061087035574018955, + 0.007816416211426258, + 0.013487469404935837, + 0.017094982787966728, + 0.02279750630259514, + 0.01172542292624712, + 0.00726543040946126, + 0.008568407036364079, + 0.0073357392102479935, + 0.010899674147367477, + 0.010178616270422935, + 0.011654756031930447, + 0.019048716872930527, + 0.016734100878238678, + 0.022173317149281502, + 0.017609452828764915, + 0.00989722739905119, + 0.009817318990826607, + 0.01875997707247734, + 0.013706854544579983, + 0.011742954142391682, + 0.006918078754097223, + 0.06386086344718933, + 0.030041087418794632, + 0.030406739562749863, + 0.024013668298721313, + 0.04979517310857773, + 0.0274359118193388, + 0.02727401629090309, + 0.018076246604323387, + 0.0036823891568928957, + 0.0025442230980843306, + 0.0019641711842268705, + 0.002425347687676549, + 0.0051596625708043575, + 0.0035846310202032328, + 0.0037229706067591906, + 0.001995082013309002, + 0.0024341342505067587, + 0.004295941907912493, + 0.002755493624135852, + 0.0037290919572114944, + 0.005592373665422201, + 0.0038760609459131956, + 0.004189355298876762, + 0.004505841992795467 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 0, + 6 + ], + "selected_reward": -0.016062267124652863 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera angle a slightly elevated bird's-eye view?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a bright yellow two-person kayak visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are two people paddling in a synchronized rhythm inside the kayak?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the woman in the front wearing a dark gray beanie and a vibrant turquoise waterproof jacket?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the man behind her wearing a burnt orange fleece pullover and holding a black paddle?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01606590487062931 + ], + "opposite_logprob_tokens": [ + -4.141066074371338 + ], + "logprob_reward": -0.01606590487062931, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.01606590487062931, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.161081001162529, + 0.028742149472236633, + 0.031039774417877197, + 0.026883065700531006, + 0.14904095232486725, + 0.13820254802703857, + 0.12498452514410019, + 0.26243409514427185, + 0.04776095971465111, + 0.029830897226929665 + ], + "credit_map": [ + 0.024533582851290703, + 0.0071321530267596245, + 0.03988130763173103, + 0.006550085265189409, + 0.0064598992466926575, + 0.04850589856505394, + 0.008060095831751823, + 0.01995798386633396, + 0.004349475260823965, + 0.0027557609137147665, + 0.0028546194080263376, + 0.0026844472158700228, + 0.0025359713472425938, + 0.0064714215695858, + 0.004891782999038696, + 0.0021986709907650948, + 0.004443412646651268, + 0.004084866493940353, + 0.0041604009456932545, + 0.0029439562931656837, + 0.004967216867953539, + 0.00326332775875926, + 0.0038553206250071526, + 0.0033212744165211916, + 0.0023983148857951164, + 0.0024804212152957916, + 0.0020106537267565727, + 0.005605210084468126, + 0.0032424391247332096, + 0.0031776465475559235, + 0.005135256331413984, + 0.002833122154697776, + 0.014179921709001064, + 0.01272293645888567, + 0.009917733259499073, + 0.01835852488875389, + 0.01830117031931877, + 0.022666435688734055, + 0.030468758195638657, + 0.022425465285778046, + 0.01635028049349785, + 0.014985050074756145, + 0.012163545936346054, + 0.012988242320716381, + 0.02091851457953453, + 0.0185709185898304, + 0.02281658537685871, + 0.019409408792853355, + 0.025365401059389114, + 0.018355242908000946, + 0.01137276366353035, + 0.014319141395390034, + 0.01672245003283024, + 0.01563093438744545, + 0.012755146250128746, + 0.010463438928127289, + 0.06247467175126076, + 0.029579641297459602, + 0.03115050122141838, + 0.02343985065817833, + 0.044205375015735626, + 0.025881022214889526, + 0.026578109711408615, + 0.019124921411275864, + 0.008038109168410301, + 0.007070633582770824, + 0.003625048091635108, + 0.002678539603948593, + 0.010896614752709866, + 0.006624233443289995, + 0.006111907307058573, + 0.002715872135013342, + 0.002736859256401658, + 0.004099218640476465, + 0.0025582199450582266, + 0.0035123389679938555, + 0.0049846963956952095, + 0.0037152061704546213, + 0.0040840087458491325, + 0.004140347708016634 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 0, + 4 + ], + "selected_reward": -0.01606590487062931 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "slow, smooth, and slightly bird's-eye view of the drone gracefully accompanies a bright yellow two-person kayak as it leisurely paddles across the crystal-clear turquoise waters of a high mountain lake. The lake's surface, as tranquil as a mirror, perfectly reflects the surrounding jagged, majestic snow-capped peaks and the lush, deep green evergreen forests lining its shores, making the line between reality and reflection almost indistinguishable. Inside, two friends paddle in a fluid, synchronized rhythm. The woman in the front, her blond ponytail peeking from beneath a dark gray beanie, wears a vibrant turquoise waterproof jacket and a calm, focused expression. Behind her, her male companion, clad in a burnt orange fleece pullover and clutching a black carbon fiber paddle, glances up from time to time with a smile of awe at the majestic mountain scenery. Their oars dipped almost silently into the water, the only things breaking the glassy surface were tiny concentric ripples that spread out from where the oars entered the water, and an almost imperceptible breeze that stirred a subtle shimmer across the surface of the lake around the boats. The soft golden afternoon sunlight bathed the scene, warming the textures of the snow-capped peaks and making the water shimmer, creating an atmosphere of grandeur, tranquility, and profound tranquility.", + "seed": 22110, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000071/group_1/index_5722_seed_22110_dp_3_rank_6_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 1, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.03113936653826386, + "vlm_logprob": -0.03113936653826386, + "vlm_yes_no_margin": 4.025, + "vlm_token_credit": -0.03113936653826386, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera angle a slightly elevated bird's-eye view?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a bright yellow two-person kayak visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are two people paddling in a synchronized rhythm inside the kayak?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the woman in the front wearing a dark gray beanie and a vibrant turquoise waterproof jacket?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the man behind her wearing a burnt orange fleece pullover and holding a black paddle?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.037363987416028976 + ], + "opposite_logprob_tokens": [ + -3.6623640060424805 + ], + "logprob_reward": -0.037363987416028976, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.037363987416028976, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.6930768489837646, + 0.010241221636533737, + 0.009789521805942059, + 0.03611135482788086, + 0.0197892002761364, + 0.03155820816755295, + 0.06272787600755692, + 0.015334879979491234, + 0.0310851838439703, + 0.09028573334217072 + ], + "credit_map": [ + 0.18249811232089996, + 0.04267770051956177, + 0.039633672684431076, + 0.04886307939887047, + 0.035680584609508514, + 0.050421833992004395, + 0.04206513613462448, + 0.2512366771697998, + 0.0013560419902205467, + 0.0009900813456624746, + 0.001041620154865086, + 0.0009078201255761087, + 0.0009427882032468915, + 0.0011962386779487133, + 0.000886151974555105, + 0.002920478116720915, + 0.0010116009507328272, + 0.00135610846336931, + 0.0008245924254879355, + 0.0012141307815909386, + 0.001535930554382503, + 0.001002884702757001, + 0.001029278850182891, + 0.0018149939132854342, + 0.004635341931134462, + 0.00467808498069644, + 0.002368433866649866, + 0.0057321530766785145, + 0.005786368623375893, + 0.004382219631224871, + 0.00411138404160738, + 0.004417366813868284, + 0.0027810060419142246, + 0.0028793306555598974, + 0.0012902689632028341, + 0.0019566374830901623, + 0.004693293944001198, + 0.0026488678995519876, + 0.0017090983456000686, + 0.0018306946149095893, + 0.004599843639880419, + 0.006668642163276672, + 0.0024056590627878904, + 0.0047876774333417416, + 0.004642146173864603, + 0.0031427829526364803, + 0.0027179676108062267, + 0.0025934865698218346, + 0.018586421385407448, + 0.016726434230804443, + 0.0036453171633183956, + 0.004721281584352255, + 0.007673337124288082, + 0.0045939558185637, + 0.0025715655647218227, + 0.004209557548165321, + 0.0026222995948046446, + 0.0038144460413604975, + 0.0009373071370646358, + 0.0008962289430201054, + 0.003070427803322673, + 0.0012130544055253267, + 0.0011581321014091372, + 0.0016229826724156737, + 0.0041413442231714725, + 0.004676838405430317, + 0.00244008912704885, + 0.0021861400455236435, + 0.003753556404262781, + 0.004128275904804468, + 0.002720494056120515, + 0.00703844428062439, + 0.00783112645149231, + 0.012399879284203053, + 0.0035000003408640623, + 0.005531998351216316, + 0.033854708075523376, + 0.006631411612033844, + 0.00624918332323432, + 0.014287416823208332 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 6 + ], + "selected_reward": -0.037363987416028976 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera angle a slightly elevated bird's-eye view?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a bright yellow two-person kayak visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are two people paddling in a synchronized rhythm inside the kayak?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the woman in the front wearing a dark gray beanie and a vibrant turquoise waterproof jacket?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the man behind her wearing a burnt orange fleece pullover and holding a black paddle?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0017391807632520795 + ], + "opposite_logprob_tokens": [ + -6.376739025115967 + ], + "logprob_reward": -0.0017391807632520795, + "yes_no_margin_reward": 6.375, + "token_credit_reward": -0.0017391807632520795, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.7253230214118958, + 0.006900853011757135, + 0.018419615924358368, + 0.03503191098570824, + 0.008666791021823883, + 0.041635118424892426, + 0.1011393740773201, + 0.0066367085091769695, + 0.009183354675769806, + 0.04706323891878128 + ], + "credit_map": [ + 0.1904040277004242, + 0.044382285326719284, + 0.04143138974905014, + 0.05149061977863312, + 0.03681066259741783, + 0.052590180188417435, + 0.044459033757448196, + 0.26375481486320496, + 0.0006050266674719751, + 0.0008697787416167557, + 0.001182547421194613, + 0.0010879701003432274, + 0.0006521293544210494, + 0.0005643075564876199, + 0.0009775562211871147, + 0.000961537123657763, + 0.001701165921986103, + 0.002396455267444253, + 0.0011626725317910314, + 0.002428988693282008, + 0.004507747944444418, + 0.0016145947156473994, + 0.0021338891237974167, + 0.0024741014931350946, + 0.002557199215516448, + 0.003526938846334815, + 0.001541019999422133, + 0.0021399029064923525, + 0.01269571203738451, + 0.0023897963110357523, + 0.002181294606998563, + 0.008000046946108341, + 0.0014786969404667616, + 0.0008733129943720996, + 0.0007199309184215963, + 0.000564180372748524, + 0.002355896169319749, + 0.0011089576873928308, + 0.0006014050450176001, + 0.0009644106612540781, + 0.0066736117005348206, + 0.007934702560305595, + 0.00315451854839921, + 0.006181743461638689, + 0.0062988754361867905, + 0.00381017429754138, + 0.003805689048022032, + 0.003775801043957472, + 0.03234395012259483, + 0.026842694729566574, + 0.006427126005291939, + 0.006554445717483759, + 0.011462323367595673, + 0.0066240094602108, + 0.0041426713578403, + 0.006742153316736221, + 0.0007024090737104416, + 0.0018742723623290658, + 0.0004975288757123053, + 0.000586515583563596, + 0.0015097936848178506, + 0.0005335385212674737, + 0.00042332318844273686, + 0.0005093273357488215, + 0.0012498853029683232, + 0.0012821592390537262, + 0.0006540186586789787, + 0.000659649318549782, + 0.0011543447617441416, + 0.0013302044244483113, + 0.0007082451484166086, + 0.0021448477637022734, + 0.010827730409801006, + 0.004998191725462675, + 0.0026367066893726587, + 0.003408763324841857, + 0.01430015079677105, + 0.002676744246855378, + 0.004111460875719786, + 0.004103491548448801 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 6, + 9 + ], + "selected_reward": -0.0017391807632520795 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera angle a slightly elevated bird's-eye view?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a bright yellow two-person kayak visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are two people paddling in a synchronized rhythm inside the kayak?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the woman in the front wearing a dark gray beanie and a vibrant turquoise waterproof jacket?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the man behind her wearing a burnt orange fleece pullover and holding a black paddle?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07002159208059311 + ], + "opposite_logprob_tokens": [ + -2.695021629333496 + ], + "logprob_reward": -0.07002159208059311, + "yes_no_margin_reward": 2.625, + "token_credit_reward": -0.07002159208059311, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.6477111577987671, + 0.015370824374258518, + 0.02501596510410309, + 0.03682997077703476, + 0.02824844792485237, + 0.062098048627376556, + 0.07910262048244476, + 0.017412235960364342, + 0.061293959617614746, + 0.0269168671220541 + ], + "credit_map": [ + 0.16851012408733368, + 0.0396844744682312, + 0.03709939494729042, + 0.04718894511461258, + 0.03362702205777168, + 0.04729277640581131, + 0.039723195135593414, + 0.23458518087863922, + 0.0017188723431900144, + 0.0022652919869869947, + 0.0013085956452414393, + 0.001914482214488089, + 0.0018358032684773207, + 0.0017255248967558146, + 0.0013530009891837835, + 0.0032492505852133036, + 0.0015525423223152757, + 0.002320069121196866, + 0.0013668378815054893, + 0.0008884024573490024, + 0.007619902491569519, + 0.0018728942377492785, + 0.0011926895240321755, + 0.008202624507248402, + 0.005434664897620678, + 0.004447300452739, + 0.0019153819885104895, + 0.003183982102200389, + 0.011817391961812973, + 0.0031694304198026657, + 0.002780779730528593, + 0.0040810368955135345, + 0.004209490027278662, + 0.0035874771419912577, + 0.0019099408527836204, + 0.002234142739325762, + 0.00854487344622612, + 0.003523023799061775, + 0.0022373339161276817, + 0.002002161694690585, + 0.010018481872975826, + 0.0121000399813056, + 0.004643222317099571, + 0.008915632031857967, + 0.010041294619441032, + 0.005666965153068304, + 0.005431991070508957, + 0.0052804118022322655, + 0.022081071510910988, + 0.021147768944501877, + 0.0052887024357914925, + 0.006382782012224197, + 0.010198757983744144, + 0.006152162794023752, + 0.002904399298131466, + 0.004946970380842686, + 0.0027013798244297504, + 0.003271797439083457, + 0.001423165202140808, + 0.0013420670293271542, + 0.0032239235006272793, + 0.0012659995118156075, + 0.0018967651994898915, + 0.0022871363908052444, + 0.004390657413750887, + 0.009009134955704212, + 0.00381508469581604, + 0.008308667689561844, + 0.011932254768908024, + 0.005102223251014948, + 0.006058139260858297, + 0.012677785009145737, + 0.004957413300871849, + 0.004333668388426304, + 0.001875624991953373, + 0.0030792129691690207, + 0.005011828150600195, + 0.0028151716105639935, + 0.0021536066196858883, + 0.002690334804356098 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 6, + 5 + ], + "selected_reward": -0.07002159208059311 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera angle a slightly elevated bird's-eye view?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a bright yellow two-person kayak visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are two people paddling in a synchronized rhythm inside the kayak?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the woman in the front wearing a dark gray beanie and a vibrant turquoise waterproof jacket?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the man behind her wearing a burnt orange fleece pullover and holding a black paddle?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.023284347727894783 + ], + "opposite_logprob_tokens": [ + -3.773284435272217 + ], + "logprob_reward": -0.023284347727894783, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.023284347727894783, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.5184702277183533, + 0.012527721002697945, + 0.028577180579304695, + 0.058162324130535126, + 0.018384523689746857, + 0.07146192342042923, + 0.1614275872707367, + 0.00921652838587761, + 0.04933007061481476, + 0.07244199514389038 + ], + "credit_map": [ + 0.13943588733673096, + 0.030074691399931908, + 0.02969992347061634, + 0.03132547810673714, + 0.025159301236271858, + 0.03887610882520676, + 0.028051147237420082, + 0.19584761559963226, + 0.0014937056694179773, + 0.0017489000456407666, + 0.0013216602383181453, + 0.001680839341133833, + 0.0015409025363624096, + 0.0013289605267345905, + 0.0012470318470150232, + 0.0021657187025994062, + 0.0018820951227098703, + 0.002258639084175229, + 0.0018458033446222544, + 0.0012332547921687365, + 0.008060252293944359, + 0.0021641585044562817, + 0.0018261477816849947, + 0.009306828491389751, + 0.0034852020908147097, + 0.0057436455972492695, + 0.0025331575889140368, + 0.003228405024856329, + 0.01971878856420517, + 0.004064843524247408, + 0.003379488829523325, + 0.016008790582418442, + 0.0027979607693850994, + 0.0029364728834480047, + 0.0014023135881870985, + 0.0018955389969050884, + 0.0024042336735874414, + 0.0022532183211296797, + 0.0014938723761588335, + 0.0032009114511311054, + 0.011876624077558517, + 0.014260136522352695, + 0.005599875934422016, + 0.010139087215065956, + 0.010866361670196056, + 0.0064556491561234, + 0.0057054501958191395, + 0.0065587363205850124, + 0.0491182915866375, + 0.04386785626411438, + 0.009485924616456032, + 0.01125787477940321, + 0.019345184788107872, + 0.01119073573499918, + 0.006087529007345438, + 0.011074176989495754, + 0.001533741015009582, + 0.0011151289800181985, + 0.0011352914152666926, + 0.0012647139374166727, + 0.0008118533296510577, + 0.0010266473982483149, + 0.0010670514311641455, + 0.0012621002970263362, + 0.006049322430044413, + 0.009790340438485146, + 0.0030351753812283278, + 0.0034539245534688234, + 0.008875660598278046, + 0.006014672107994556, + 0.00353897619061172, + 0.008571991696953773, + 0.010122961364686489, + 0.010374643839895725, + 0.0031909488607198, + 0.006500551477074623, + 0.024078980088233948, + 0.0035939400549978018, + 0.004785951226949692, + 0.009794019162654877 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 6, + 9 + ], + "selected_reward": -0.023284347727894783 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera angle a slightly elevated bird's-eye view?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a bright yellow two-person kayak visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are two people paddling in a synchronized rhythm inside the kayak?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the woman in the front wearing a dark gray beanie and a vibrant turquoise waterproof jacket?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the man behind her wearing a burnt orange fleece pullover and holding a black paddle?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02328772470355034 + ], + "opposite_logprob_tokens": [ + -3.773287773132324 + ], + "logprob_reward": -0.02328772470355034, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.02328772470355034, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1580689549446106, + 0.01814836636185646, + 0.02879389375448227, + 0.05349315330386162, + 0.08358200639486313, + 0.19870854914188385, + 0.18947741389274597, + 0.02255028486251831, + 0.18465308845043182, + 0.06252426654100418 + ], + "credit_map": [ + 0.025230644270777702, + 0.016572562977671623, + 0.012195715680718422, + 0.025633184239268303, + 0.01418283674865961, + 0.01139483880251646, + 0.015484195202589035, + 0.03737497702240944, + 0.0028627472929656506, + 0.0016385875642299652, + 0.0018389658071100712, + 0.001846924889832735, + 0.002039311919361353, + 0.0026580486446619034, + 0.0021443620789796114, + 0.003119417931884527, + 0.0032714607659727335, + 0.005121774040162563, + 0.0025099690537899733, + 0.00462856562808156, + 0.0033956882543861866, + 0.0027861655689775944, + 0.004099007695913315, + 0.002981261583045125, + 0.008947022259235382, + 0.0066382125951349735, + 0.003507762448862195, + 0.005575275048613548, + 0.01415981724858284, + 0.0053557464852929115, + 0.004511100705713034, + 0.00479821627959609, + 0.014157106168568134, + 0.011631423607468605, + 0.005778299644589424, + 0.007256621029227972, + 0.01967845857143402, + 0.010521968826651573, + 0.006414899602532387, + 0.008143228478729725, + 0.0331977903842926, + 0.03729945048689842, + 0.015482407063245773, + 0.028459321707487106, + 0.03008728101849556, + 0.018964260816574097, + 0.01751626469194889, + 0.017701776698231697, + 0.05724206194281578, + 0.051431626081466675, + 0.011060875840485096, + 0.013473776169121265, + 0.02304033748805523, + 0.012943428941071033, + 0.00702886376529932, + 0.013256444595754147, + 0.0027160239405930042, + 0.0036037741228938103, + 0.0021587249357253313, + 0.002841951558366418, + 0.0036408279556781054, + 0.0017509176395833492, + 0.0033322342205792665, + 0.0025058300234377384, + 0.015922702848911285, + 0.02684902399778366, + 0.010825351811945438, + 0.022473618388175964, + 0.03595424443483353, + 0.01598532311618328, + 0.01653948426246643, + 0.04010333493351936, + 0.007289391942322254, + 0.010090778581798077, + 0.003714236430823803, + 0.007484850939363241, + 0.016793524846434593, + 0.007929062470793724, + 0.003649568883702159, + 0.005572855938225985 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 6, + 8 + ], + "selected_reward": -0.02328772470355034 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "slow, smooth, and slightly bird's-eye view of the drone gracefully accompanies a bright yellow two-person kayak as it leisurely paddles across the crystal-clear turquoise waters of a high mountain lake. The lake's surface, as tranquil as a mirror, perfectly reflects the surrounding jagged, majestic snow-capped peaks and the lush, deep green evergreen forests lining its shores, making the line between reality and reflection almost indistinguishable. Inside, two friends paddle in a fluid, synchronized rhythm. The woman in the front, her blond ponytail peeking from beneath a dark gray beanie, wears a vibrant turquoise waterproof jacket and a calm, focused expression. Behind her, her male companion, clad in a burnt orange fleece pullover and clutching a black carbon fiber paddle, glances up from time to time with a smile of awe at the majestic mountain scenery. Their oars dipped almost silently into the water, the only things breaking the glassy surface were tiny concentric ripples that spread out from where the oars entered the water, and an almost imperceptible breeze that stirred a subtle shimmer across the surface of the lake around the boats. The soft golden afternoon sunlight bathed the scene, warming the textures of the snow-capped peaks and making the water shimmer, creating an atmosphere of grandeur, tranquility, and profound tranquility.", + "seed": 22111, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000071/group_1/index_5722_seed_22111_dp_3_rank_6_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 1, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.03864805363118649, + "vlm_logprob": -0.03864805363118649, + "vlm_yes_no_margin": 3.9, + "vlm_token_credit": -0.03864805363118649, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera angle a slightly elevated bird's-eye view?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a bright yellow two-person kayak visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are two people paddling in a synchronized rhythm inside the kayak?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the woman in the front wearing a dark gray beanie and a vibrant turquoise waterproof jacket?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the man behind her wearing a burnt orange fleece pullover and holding a black paddle?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.043821070343256 + ], + "opposite_logprob_tokens": [ + -3.793821096420288 + ], + "logprob_reward": -0.043821070343256, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.043821070343256, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.19854503870010376, + 0.08218981325626373, + 0.09129278361797333, + 0.3051086664199829, + 0.019093694165349007, + 0.06320425868034363, + 0.06205315142869949, + 0.02590501308441162, + 0.04924440383911133, + 0.10336313396692276 + ], + "credit_map": [ + 0.031669944524765015, + 0.023008238524198532, + 0.025786269456148148, + 0.029735665768384933, + 0.025859015062451363, + 0.02223741076886654, + 0.024566616863012314, + 0.015681887045502663, + 0.02045295387506485, + 0.007977249100804329, + 0.00853306706994772, + 0.009672186337411404, + 0.007909342646598816, + 0.00953742116689682, + 0.01147800125181675, + 0.006629583425819874, + 0.018145252019166946, + 0.008300046436488628, + 0.010341770946979523, + 0.012141481973230839, + 0.009977438487112522, + 0.014985673129558563, + 0.008842842653393745, + 0.00855827797204256, + 0.032586921006441116, + 0.025376364588737488, + 0.039577219635248184, + 0.06618622690439224, + 0.02350272238254547, + 0.056053951382637024, + 0.04208746924996376, + 0.019737770780920982, + 0.002546773524954915, + 0.002402971964329481, + 0.0018641200149431825, + 0.0027592983096837997, + 0.0027995097916573286, + 0.002523138653486967, + 0.0020008154679089785, + 0.0021970653906464577, + 0.008795562200248241, + 0.007043186109513044, + 0.007656961679458618, + 0.0075479354709386826, + 0.008185145445168018, + 0.008248290978372097, + 0.008887039497494698, + 0.006840134039521217, + 0.015541507862508297, + 0.007622816599905491, + 0.006472119130194187, + 0.004989716690033674, + 0.007634819019585848, + 0.009057410061359406, + 0.00512807909399271, + 0.005606680642813444, + 0.0028684830758720636, + 0.002238508081063628, + 0.004506861325353384, + 0.0024082139134407043, + 0.0023297499865293503, + 0.0067673334851861, + 0.003064457792788744, + 0.001721405191347003, + 0.008235323242843151, + 0.0060503617860376835, + 0.006868378724902868, + 0.002963871229439974, + 0.006954703014343977, + 0.010656759142875671, + 0.002774057909846306, + 0.00474094832316041, + 0.013870816677808762, + 0.011537316255271435, + 0.015096178278326988, + 0.01054491475224495, + 0.008794451132416725, + 0.01641891524195671, + 0.012759780511260033, + 0.014340764842927456 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 0, + 9 + ], + "selected_reward": -0.043821070343256 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera angle a slightly elevated bird's-eye view?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a bright yellow two-person kayak visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are two people paddling in a synchronized rhythm inside the kayak?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the woman in the front wearing a dark gray beanie and a vibrant turquoise waterproof jacket?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the man behind her wearing a burnt orange fleece pullover and holding a black paddle?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0025258325040340424 + ], + "opposite_logprob_tokens": [ + -6.002525806427002 + ], + "logprob_reward": -0.0025258325040340424, + "yes_no_margin_reward": 6.0, + "token_credit_reward": -0.0025258325040340424, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.21416118741035461, + 0.10700789839029312, + 0.1473112255334854, + 0.08172264695167542, + 0.03436724841594696, + 0.07048668712377548, + 0.1173991858959198, + 0.04655013978481293, + 0.07774645835161209, + 0.10324728488922119 + ], + "credit_map": [ + 0.037799008190631866, + 0.017159441486001015, + 0.0217108353972435, + 0.020279742777347565, + 0.02632860839366913, + 0.04092646762728691, + 0.03272683918476105, + 0.017230240628123283, + 0.015774425119161606, + 0.01076425425708294, + 0.013141106814146042, + 0.009485799819231033, + 0.01569315232336521, + 0.01788128726184368, + 0.01339721493422985, + 0.010870660655200481, + 0.02812577411532402, + 0.016858914867043495, + 0.016925696283578873, + 0.021580174565315247, + 0.014171720482409, + 0.019089996814727783, + 0.018040215596556664, + 0.01251874677836895, + 0.010117989964783192, + 0.009081210009753704, + 0.010939523577690125, + 0.010007751174271107, + 0.009692703373730183, + 0.017506975680589676, + 0.0074925473891198635, + 0.006883947178721428, + 0.0039418465457856655, + 0.004775145556777716, + 0.003685147501528263, + 0.004802078474313021, + 0.004782693926244974, + 0.005499981343746185, + 0.003983433824032545, + 0.0028969186823815107, + 0.011534045450389385, + 0.0074071334674954414, + 0.007580946199595928, + 0.011393208056688309, + 0.0071446942165493965, + 0.008528845384716988, + 0.00901174545288086, + 0.007886066101491451, + 0.017713872715830803, + 0.01940569095313549, + 0.012127557769417763, + 0.01067268941551447, + 0.01846710778772831, + 0.018326876685023308, + 0.009045729413628578, + 0.011639663018286228, + 0.00814273115247488, + 0.005691175814718008, + 0.005761141423135996, + 0.00537598691880703, + 0.006247531156986952, + 0.006207540165632963, + 0.0044532762840390205, + 0.004670755006372929, + 0.00867683719843626, + 0.008772911503911018, + 0.011723094619810581, + 0.008504902943968773, + 0.007083964999765158, + 0.01172620803117752, + 0.007563989143818617, + 0.013694551773369312, + 0.014747458510100842, + 0.008708354085683823, + 0.013179810717701912, + 0.009173366241157055, + 0.011402538977563381, + 0.013469872064888477, + 0.021375026553869247, + 0.01119085494428873 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 6 + ], + "selected_reward": -0.0025258325040340424 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera angle a slightly elevated bird's-eye view?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a bright yellow two-person kayak visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are two people paddling in a synchronized rhythm inside the kayak?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the woman in the front wearing a dark gray beanie and a vibrant turquoise waterproof jacket?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the man behind her wearing a burnt orange fleece pullover and holding a black paddle?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.1003231555223465 + ], + "opposite_logprob_tokens": [ + -2.35032320022583 + ], + "logprob_reward": -0.1003231555223465, + "yes_no_margin_reward": 2.25, + "token_credit_reward": -0.1003231555223465, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2655720114707947, + 0.07865802943706512, + 0.12738430500030518, + 0.047890741378068924, + 0.053907446563243866, + 0.07926969975233078, + 0.12090077251195908, + 0.03816738352179527, + 0.04523113742470741, + 0.143018439412117 + ], + "credit_map": [ + 0.052649181336164474, + 0.029636163264513016, + 0.02861884795129299, + 0.02960244007408619, + 0.03218758478760719, + 0.03458836302161217, + 0.03879252076148987, + 0.019496925175189972, + 0.013854130171239376, + 0.007531045936048031, + 0.00859752669930458, + 0.008988218382000923, + 0.008559132926166058, + 0.012500247918069363, + 0.012157022021710873, + 0.006470709573477507, + 0.019289875403046608, + 0.016872910782694817, + 0.017538294196128845, + 0.01347125880420208, + 0.016199877485632896, + 0.018542416393756866, + 0.013796349056065083, + 0.011673323810100555, + 0.004257232882082462, + 0.004481676500290632, + 0.006616898346692324, + 0.007541378028690815, + 0.005279495846480131, + 0.009102918207645416, + 0.005052859429270029, + 0.005558277480304241, + 0.006163144484162331, + 0.008261813782155514, + 0.00444809440523386, + 0.009010803885757923, + 0.0086141936480999, + 0.006820515729486942, + 0.0058813998475670815, + 0.004707477521151304, + 0.01510708685964346, + 0.007969565689563751, + 0.011162977665662766, + 0.010103103704750538, + 0.007154588121920824, + 0.012657520361244678, + 0.00749848410487175, + 0.0076163713820278645, + 0.01725192368030548, + 0.015776850283145905, + 0.01879280060529709, + 0.010021368972957134, + 0.016915375366806984, + 0.021776685491204262, + 0.008705130778253078, + 0.011660642921924591, + 0.005820963066071272, + 0.00671223783865571, + 0.004619147628545761, + 0.003772176569327712, + 0.004656446166336536, + 0.006213303189724684, + 0.0033259561751037836, + 0.0030471505597233772, + 0.006030161399394274, + 0.005294781178236008, + 0.005903289653360844, + 0.004092817194759846, + 0.004861391615122557, + 0.008608360774815083, + 0.004675069358199835, + 0.00576526764780283, + 0.014244109392166138, + 0.013631458394229412, + 0.020804643630981445, + 0.020575085654854774, + 0.016497652977705002, + 0.01919666863977909, + 0.01757282204926014, + 0.020496005192399025 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 2 + ], + "selected_reward": -0.1003231555223465 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera angle a slightly elevated bird's-eye view?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a bright yellow two-person kayak visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are two people paddling in a synchronized rhythm inside the kayak?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the woman in the front wearing a dark gray beanie and a vibrant turquoise waterproof jacket?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the man behind her wearing a burnt orange fleece pullover and holding a black paddle?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.023282134905457497 + ], + "opposite_logprob_tokens": [ + -3.773282051086426 + ], + "logprob_reward": -0.023282134905457497, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.023282134905457497, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.21042993664741516, + 0.0681442841887474, + 0.08343550562858582, + 0.07988516986370087, + 0.04186186194419861, + 0.10436157882213593, + 0.13347354531288147, + 0.07772365212440491, + 0.08298122137784958, + 0.11770322918891907 + ], + "credit_map": [ + 0.034073103219270706, + 0.028005558997392654, + 0.034144651144742966, + 0.022846272215247154, + 0.027447080239653587, + 0.027842119336128235, + 0.0198932234197855, + 0.016177935525774956, + 0.01565682701766491, + 0.006619806867092848, + 0.007981548085808754, + 0.0076296767219901085, + 0.007181853987276554, + 0.00858947541564703, + 0.008937959559261799, + 0.005547143518924713, + 0.013404492288827896, + 0.009625792503356934, + 0.010037533938884735, + 0.010570892132818699, + 0.011120663955807686, + 0.012466639280319214, + 0.01017844956368208, + 0.006031043827533722, + 0.008691497147083282, + 0.009295198135077953, + 0.010550024919211864, + 0.014941063709557056, + 0.006073182914406061, + 0.015997104346752167, + 0.007804828230291605, + 0.00653226999565959, + 0.0057276044972240925, + 0.006754606030881405, + 0.0037610260769724846, + 0.006103826686739922, + 0.0034416252747178078, + 0.004530155565589666, + 0.005587757099419832, + 0.0059552621096372604, + 0.01549693662673235, + 0.010540199466049671, + 0.013173884712159634, + 0.020205358043313026, + 0.012714456766843796, + 0.012337105348706245, + 0.009662991389632225, + 0.010230652056634426, + 0.01601783186197281, + 0.009858903475105762, + 0.03018243983387947, + 0.008845717646181583, + 0.013047917746007442, + 0.03088241070508957, + 0.009155613370239735, + 0.015482708811759949, + 0.00872163474559784, + 0.00949518010020256, + 0.011126355268061161, + 0.00980595126748085, + 0.00754201877862215, + 0.014798402786254883, + 0.011331033892929554, + 0.004903079476207495, + 0.007921428419649601, + 0.008631854318082333, + 0.011738568544387817, + 0.008692663162946701, + 0.008461078628897667, + 0.016111750155687332, + 0.012219516560435295, + 0.009204357862472534, + 0.014422540552914143, + 0.012190092355012894, + 0.011270065791904926, + 0.015578025951981544, + 0.011030948720872402, + 0.012228644452989101, + 0.02611890621483326, + 0.014864004217088223 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 6, + 9 + ], + "selected_reward": -0.023282134905457497 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera angle a slightly elevated bird's-eye view?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a bright yellow two-person kayak visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are two people paddling in a synchronized rhythm inside the kayak?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the woman in the front wearing a dark gray beanie and a vibrant turquoise waterproof jacket?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the man behind her wearing a burnt orange fleece pullover and holding a black paddle?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.023288074880838394 + ], + "opposite_logprob_tokens": [ + -3.7732880115509033 + ], + "logprob_reward": -0.023288074880838394, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.023288074880838394, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13526543974876404, + 0.06643729656934738, + 0.07314866036176682, + 0.03809928521513939, + 0.05662749707698822, + 0.1214749813079834, + 0.11574502289295197, + 0.08225623518228531, + 0.11446911841630936, + 0.1964765191078186 + ], + "credit_map": [ + 0.019049767404794693, + 0.017000317573547363, + 0.021134337410330772, + 0.0142449289560318, + 0.02072245068848133, + 0.01790248043835163, + 0.014672848396003246, + 0.010538291186094284, + 0.014953361824154854, + 0.007024710066616535, + 0.008320826105773449, + 0.007140027359127998, + 0.010004107840359211, + 0.007615806069225073, + 0.006568222306668758, + 0.0048102280125021935, + 0.010571807622909546, + 0.008130927570164204, + 0.010622493922710419, + 0.00792099628597498, + 0.008376939222216606, + 0.009542373940348625, + 0.012031870894134045, + 0.00595124252140522, + 0.005211292766034603, + 0.004379144869744778, + 0.005210057366639376, + 0.005716185085475445, + 0.004735749680548906, + 0.005088984500616789, + 0.003674097126349807, + 0.004083769395947456, + 0.008259058929979801, + 0.007749454118311405, + 0.00599269475787878, + 0.008106812834739685, + 0.005977243650704622, + 0.008240179158747196, + 0.006427657324820757, + 0.005874388851225376, + 0.017496084794402122, + 0.01213198434561491, + 0.016257740557193756, + 0.021559521555900574, + 0.012120532803237438, + 0.014891715720295906, + 0.013774723745882511, + 0.013242661952972412, + 0.012828539125621319, + 0.01024911180138588, + 0.03050481155514717, + 0.005983024835586548, + 0.011347983032464981, + 0.031270429491996765, + 0.00616849260404706, + 0.007392617873847485, + 0.008056516759097576, + 0.008058940060436726, + 0.014897088520228863, + 0.012044218368828297, + 0.007182247471064329, + 0.014260306023061275, + 0.010570698417723179, + 0.007186207454651594, + 0.011630777269601822, + 0.012409231625497341, + 0.016289589926600456, + 0.011952823027968407, + 0.011563703417778015, + 0.02204766683280468, + 0.015607372857630253, + 0.01296794693917036, + 0.024663373827934265, + 0.018000395968556404, + 0.01830870844423771, + 0.026572400704026222, + 0.014618399553000927, + 0.023686641827225685, + 0.040806595236063004, + 0.029819970950484276 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 0, + 5 + ], + "selected_reward": -0.023288074880838394 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000071/group_10/index_tensor([7020, 7020])_group_10_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000071/group_10/index_tensor([7020, 7020])_group_10_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..6f108a91d067f1d81da89c7f03ff41e914946b6f --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000071/group_10/index_tensor([7020, 7020])_group_10_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "low, static shot, shot from the ground, immerses the viewer in the depths of a dark and oppressive bamboo forest. Tall, emerald green bamboo stalks intertwine densely, forming a natural dome. Only faint, cold, moonlit rays penetrate the gaps, casting dappled shadows on the damp, decaying leaves. Suddenly, with lightning speed, a massive, muscular white tiger pounces from the deepest shadows of the bamboo forest. Its powerful hind legs stomp on the slippery fallen leaves, its entire explosive body stretched out in mid-air, like a flash of white and black lightning, shattering the silence of the forest. Its pure white fur, dotted with deep black stripes, stands out in the dim light. Its icy blue eyes flash with a deadly coldness. Its mouth opens sharply, revealing canine teeth as sharp as daggers. Its front paws thrust forward, claws fully extended, charging straight at the camera. The action is swift and decisive, full of the unrivaled aura of a predator. The high-contrast light and shadow and the cool overall tone create a heart-stopping sense of tension and primitive wild beauty.", + "seed": 24543, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000071/group_10/index_7020_seed_24543_dp_20_rank_40_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 10, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.021829846035689115, + "vlm_logprob": -0.021829846035689115, + "vlm_yes_no_margin": 4.425, + "vlm_token_credit": -0.021829846035689115, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned at a low angle, shot from the ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene set in a dark bamboo forest with tall, emerald green stalks forming a dense dome?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a massive, muscular white tiger with black stripes pounce from the shadows onto slippery fallen leaves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the tiger shown in mid-air with its front paws thrust forward and claws fully extended, charging straight at the camera?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the lighting characterized by high-contrast light and shadow with a cool overall tone?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07143516838550568 + ], + "opposite_logprob_tokens": [ + -3.4464352130889893 + ], + "logprob_reward": -0.07143516838550568, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.07143516838550568, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15155674517154694, + 0.1882299929857254, + 0.08994709700345993, + 0.04845709353685379, + 0.048345644026994705, + 0.04906368628144264, + 0.022166188806295395, + 0.07452359795570374, + 0.21818692982196808, + 0.10952300578355789 + ], + "credit_map": [ + 0.010756654664874077, + 0.014644362963736057, + 0.030044225975871086, + 0.023192647844552994, + 0.011195600964128971, + 0.028466828167438507, + 0.016408827155828476, + 0.016847606748342514, + 0.014379460364580154, + 0.019282113760709763, + 0.041681837290525436, + 0.019883954897522926, + 0.02870708331465721, + 0.022118577733635902, + 0.020908121019601822, + 0.021268848329782486, + 0.009007414802908897, + 0.007401473820209503, + 0.014515087008476257, + 0.015347445383667946, + 0.008387906476855278, + 0.013574878685176373, + 0.009919087402522564, + 0.01179380714893341, + 0.004728768020868301, + 0.005446385592222214, + 0.007896112278103828, + 0.005970778875052929, + 0.006325647700577974, + 0.006749655585736036, + 0.0063409567810595036, + 0.004998787306249142, + 0.004475933033972979, + 0.004360707011073828, + 0.00621439004316926, + 0.009453317150473595, + 0.003616977483034134, + 0.008500526659190655, + 0.0059504155069589615, + 0.0057733794674277306, + 0.006267879158258438, + 0.004136446863412857, + 0.007291275542229414, + 0.006923132110387087, + 0.004050148651003838, + 0.006840939167886972, + 0.006540803238749504, + 0.007013062946498394, + 0.003223143983632326, + 0.0022017296869307756, + 0.0027034070808440447, + 0.003638703841716051, + 0.0016870878171175718, + 0.003725510323420167, + 0.0021446289028972387, + 0.0028419778682291508, + 0.007730402052402496, + 0.007178676780313253, + 0.010799483396112919, + 0.00659240921959281, + 0.009277540259063244, + 0.017145859077572823, + 0.007434213533997536, + 0.008365019224584103, + 0.021154997870326042, + 0.020194323733448982, + 0.037579264491796494, + 0.02690145932137966, + 0.01269274391233921, + 0.03287145495414734, + 0.030440514907240868, + 0.036352161318063736, + 0.01137698907405138, + 0.011108183301985264, + 0.016959790140390396, + 0.022960638627409935, + 0.008932309225201607, + 0.018787330016493797, + 0.010188414715230465, + 0.009209350682795048 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 1, + 0 + ], + "selected_reward": -0.07143516838550568 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned at a low angle, shot from the ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene set in a dark bamboo forest with tall, emerald green stalks forming a dense dome?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a massive, muscular white tiger with black stripes pounce from the shadows onto slippery fallen leaves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the tiger shown in mid-air with its front paws thrust forward and claws fully extended, charging straight at the camera?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the lighting characterized by high-contrast light and shadow with a cool overall tone?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0067544239573180676 + ], + "opposite_logprob_tokens": [ + -5.006754398345947 + ], + "logprob_reward": -0.0067544239573180676, + "yes_no_margin_reward": 5.0, + "token_credit_reward": -0.0067544239573180676, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11758642643690109, + 0.051521360874176025, + 0.05958903953433037, + 0.03266730159521103, + 0.07988365739583969, + 0.0267027635127306, + 0.0493837334215641, + 0.0308791883289814, + 0.4719362258911133, + 0.07985032349824905 + ], + "credit_map": [ + 0.010167892090976238, + 0.0123780257999897, + 0.0164667759090662, + 0.01948985829949379, + 0.010665079578757286, + 0.01860344596207142, + 0.011757510714232922, + 0.01805783435702324, + 0.006446752231568098, + 0.005849729757755995, + 0.007792882155627012, + 0.008310175500810146, + 0.004195324145257473, + 0.00854136236011982, + 0.004683684092015028, + 0.005701454356312752, + 0.006033700425177813, + 0.005745782051235437, + 0.013152810744941235, + 0.009361696429550648, + 0.0070196352899074554, + 0.008213705383241177, + 0.004134036134928465, + 0.005927675869315863, + 0.0031367025803774595, + 0.002956208074465394, + 0.004421762656420469, + 0.004422701895236969, + 0.002555574057623744, + 0.008058738894760609, + 0.0034603900276124477, + 0.003655223175883293, + 0.006664730142802, + 0.006246781907975674, + 0.01485347468405962, + 0.012893321923911572, + 0.0062292758375406265, + 0.01587630622088909, + 0.007463926449418068, + 0.009655839763581753, + 0.0024810098111629486, + 0.0023822300136089325, + 0.00562268728390336, + 0.0030302631203085184, + 0.002259288914501667, + 0.005292481742799282, + 0.0026224502362310886, + 0.003012355649843812, + 0.006009128410369158, + 0.0035216575488448143, + 0.008785011246800423, + 0.006003396585583687, + 0.0035396309103816748, + 0.010424626059830189, + 0.0037479607854038477, + 0.007352324202656746, + 0.0027466793544590473, + 0.0024971780367195606, + 0.0058736237697303295, + 0.0031086488161236048, + 0.004126251675188541, + 0.005515335593372583, + 0.002890708390623331, + 0.004120763391256332, + 0.04022293910384178, + 0.04527325928211212, + 0.09211468696594238, + 0.041676606982946396, + 0.041212186217308044, + 0.09960358589887619, + 0.042643796652555466, + 0.0691891610622406, + 0.009614143520593643, + 0.007683746051043272, + 0.012962776236236095, + 0.014614192768931389, + 0.005416495725512505, + 0.014565176330506802, + 0.007964610122144222, + 0.0070291850715875626 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 0, + 4 + ], + "selected_reward": -0.0067544239573180676 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned at a low angle, shot from the ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene set in a dark bamboo forest with tall, emerald green stalks forming a dense dome?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a massive, muscular white tiger with black stripes pounce from the shadows onto slippery fallen leaves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the tiger shown in mid-air with its front paws thrust forward and claws fully extended, charging straight at the camera?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the lighting characterized by high-contrast light and shadow with a cool overall tone?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.012537943199276924 + ], + "opposite_logprob_tokens": [ + -4.387537956237793 + ], + "logprob_reward": -0.012537943199276924, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.012537943199276924, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0760338231921196, + 0.07464322447776794, + 0.04112210497260094, + 0.025531290099024773, + 0.0443035326898098, + 0.04601404443383217, + 0.048471514135599136, + 0.11166155338287354, + 0.39965757727622986, + 0.1325613558292389 + ], + "credit_map": [ + 0.006639424711465836, + 0.00767469173297286, + 0.008936602622270584, + 0.017894575372338295, + 0.004871522542089224, + 0.009629875421524048, + 0.01075773872435093, + 0.009629393927752972, + 0.007846017368137836, + 0.0054648020304739475, + 0.015049368143081665, + 0.010920953936874866, + 0.00717836944386363, + 0.00842446368187666, + 0.011679379269480705, + 0.008079876191914082, + 0.005455596372485161, + 0.0032032278832048178, + 0.006623651832342148, + 0.0052217827178537846, + 0.00353153795003891, + 0.004975041840225458, + 0.006989337503910065, + 0.005121930968016386, + 0.0021347219590097666, + 0.0030935199465602636, + 0.004059919621795416, + 0.004462774842977524, + 0.0028342504519969225, + 0.002830172423273325, + 0.0034107817336916924, + 0.002705147722736001, + 0.0048035369254648685, + 0.0044068447314202785, + 0.008836056105792522, + 0.0059212916530668736, + 0.0032870634458959103, + 0.006572792772203684, + 0.005105119198560715, + 0.0053708311170339584, + 0.004406563006341457, + 0.0038884838577359915, + 0.009679964743554592, + 0.00540057523176074, + 0.0034280046820640564, + 0.008219696581363678, + 0.005792686250060797, + 0.005198071710765362, + 0.005297032650560141, + 0.003430270357057452, + 0.007898060604929924, + 0.00981763657182455, + 0.002885624533519149, + 0.006993996445089579, + 0.005738845095038414, + 0.006410046946257353, + 0.012057536281645298, + 0.011354976333677769, + 0.014878617599606514, + 0.013675779104232788, + 0.011995838023722172, + 0.023753151297569275, + 0.011297843419015408, + 0.012647813186049461, + 0.056293290108442307, + 0.035912517458200455, + 0.07345710694789886, + 0.035608693957328796, + 0.03539714962244034, + 0.07138871401548386, + 0.03730478137731552, + 0.05429532006382942, + 0.017430119216442108, + 0.012731385417282581, + 0.022422976791858673, + 0.027219345793128014, + 0.011921227909624577, + 0.019529147073626518, + 0.011188195087015629, + 0.010118947364389896 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 9, + 7 + ], + "selected_reward": -0.012537943199276924 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned at a low angle, shot from the ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene set in a dark bamboo forest with tall, emerald green stalks forming a dense dome?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a massive, muscular white tiger with black stripes pounce from the shadows onto slippery fallen leaves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the tiger shown in mid-air with its front paws thrust forward and claws fully extended, charging straight at the camera?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the lighting characterized by high-contrast light and shadow with a cool overall tone?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.009780699387192726 + ], + "opposite_logprob_tokens": [ + -4.6347808837890625 + ], + "logprob_reward": -0.009780699387192726, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.009780699387192726, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08421508222818375, + 0.06848858296871185, + 0.13815738260746002, + 0.028731707483530045, + 0.05185393616557121, + 0.05733923614025116, + 0.094692163169384, + 0.12552642822265625, + 0.24418450891971588, + 0.10681101679801941 + ], + "credit_map": [ + 0.0069891526363790035, + 0.009323405101895332, + 0.01091914251446724, + 0.015348472632467747, + 0.006650071125477552, + 0.014432248659431934, + 0.010270259343087673, + 0.010282333008944988, + 0.004319849889725447, + 0.004976123105734587, + 0.010812126100063324, + 0.01152624748647213, + 0.009334606118500233, + 0.010036162100732327, + 0.011751197278499603, + 0.005732272285968065, + 0.011859825812280178, + 0.009199099615216255, + 0.03451882302761078, + 0.01622580923140049, + 0.007964849472045898, + 0.020845351740717888, + 0.021443510428071022, + 0.016100097447633743, + 0.0020665067713707685, + 0.0029855621978640556, + 0.003900849726051092, + 0.004331786185503006, + 0.0035934753250330687, + 0.0037373153027147055, + 0.00466435169801116, + 0.0034518595784902573, + 0.005256096366792917, + 0.004496311768889427, + 0.007719544693827629, + 0.009457334876060486, + 0.004833357874304056, + 0.006094388198107481, + 0.005962217226624489, + 0.00803468469530344, + 0.004906703718006611, + 0.0041600726544857025, + 0.010045822709798813, + 0.008351047523319721, + 0.0045077321119606495, + 0.007902356795966625, + 0.010490873828530312, + 0.006974625401198864, + 0.009916402399539948, + 0.007834806106984615, + 0.015584642067551613, + 0.01849965751171112, + 0.008485452271997929, + 0.015198118053376675, + 0.010732464492321014, + 0.00844061840325594, + 0.013544493354856968, + 0.0088798301294446, + 0.020485375076532364, + 0.014275692403316498, + 0.011501223780214787, + 0.027310660108923912, + 0.012903486378490925, + 0.0166256632655859, + 0.030710844323039055, + 0.020299173891544342, + 0.04509773477911949, + 0.02206694893538952, + 0.020135538652539253, + 0.04517439752817154, + 0.023124894127249718, + 0.03757496923208237, + 0.014903521165251732, + 0.010415054857730865, + 0.018793949857354164, + 0.017071492969989777, + 0.011836608871817589, + 0.015526232309639454, + 0.00999140553176403, + 0.008272751234471798 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 2, + 7 + ], + "selected_reward": -0.009780699387192726 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned at a low angle, shot from the ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene set in a dark bamboo forest with tall, emerald green stalks forming a dense dome?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a massive, muscular white tiger with black stripes pounce from the shadows onto slippery fallen leaves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the tiger shown in mid-air with its front paws thrust forward and claws fully extended, charging straight at the camera?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the lighting characterized by high-contrast light and shadow with a cool overall tone?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.008640995249152184 + ], + "opposite_logprob_tokens": [ + -4.758640766143799 + ], + "logprob_reward": -0.008640995249152184, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.008640995249152184, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3080778419971466, + 0.16277971863746643, + 0.07975984364748001, + 0.0496373251080513, + 0.02681472897529602, + 0.045011378824710846, + 0.05802442878484726, + 0.061404723674058914, + 0.08072280883789062, + 0.1277671605348587 + ], + "credit_map": [ + 0.04131098464131355, + 0.041782528162002563, + 0.04244321212172508, + 0.03145020827651024, + 0.03322560340166092, + 0.05106479674577713, + 0.02289365604519844, + 0.04390687122941017, + 0.025683371350169182, + 0.021899627521634102, + 0.022308895364403725, + 0.019210755825042725, + 0.019138673320412636, + 0.019061880186200142, + 0.0164356492459774, + 0.019040865823626518, + 0.009290183894336224, + 0.008308030664920807, + 0.011101321317255497, + 0.009787597693502903, + 0.012137176468968391, + 0.01321189571171999, + 0.006998635362833738, + 0.0089249974116683, + 0.004525491502135992, + 0.0055493856780231, + 0.008369557559490204, + 0.005685378331691027, + 0.006275802385061979, + 0.007663446478545666, + 0.005540019366890192, + 0.0060282424092292786, + 0.004202688112854958, + 0.0032712563406676054, + 0.0032657559495419264, + 0.00408905278891325, + 0.00230141612701118, + 0.004211554303765297, + 0.002364095998927951, + 0.00310890912078321, + 0.005023561418056488, + 0.005316352937370539, + 0.00621991790831089, + 0.006481834687292576, + 0.004398935474455357, + 0.006740807555615902, + 0.004830559249967337, + 0.005999408196657896, + 0.006019935477524996, + 0.0048492602072656155, + 0.010161221027374268, + 0.010603793896734715, + 0.004618051461875439, + 0.009294130839407444, + 0.0062657189555466175, + 0.006212317384779453, + 0.008960414677858353, + 0.005096076522022486, + 0.010758085176348686, + 0.006011719815433025, + 0.004498150199651718, + 0.011463968083262444, + 0.005712899379432201, + 0.008903410285711288, + 0.0072502633556723595, + 0.008652650751173496, + 0.0114067941904068, + 0.009450635872781277, + 0.005205464083701372, + 0.01573106460273266, + 0.012212813831865788, + 0.010813115164637566, + 0.01694587618112564, + 0.012708495371043682, + 0.023575516417622566, + 0.01914336159825325, + 0.010353107936680317, + 0.022291023284196854, + 0.012043238617479801, + 0.010706550441682339 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 9 + ], + "selected_reward": -0.008640995249152184 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "low, static shot, shot from the ground, immerses the viewer in the depths of a dark and oppressive bamboo forest. Tall, emerald green bamboo stalks intertwine densely, forming a natural dome. Only faint, cold, moonlit rays penetrate the gaps, casting dappled shadows on the damp, decaying leaves. Suddenly, with lightning speed, a massive, muscular white tiger pounces from the deepest shadows of the bamboo forest. Its powerful hind legs stomp on the slippery fallen leaves, its entire explosive body stretched out in mid-air, like a flash of white and black lightning, shattering the silence of the forest. Its pure white fur, dotted with deep black stripes, stands out in the dim light. Its icy blue eyes flash with a deadly coldness. Its mouth opens sharply, revealing canine teeth as sharp as daggers. Its front paws thrust forward, claws fully extended, charging straight at the camera. The action is swift and decisive, full of the unrivaled aura of a predator. The high-contrast light and shadow and the cool overall tone create a heart-stopping sense of tension and primitive wild beauty.", + "seed": 24544, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000071/group_10/index_7020_seed_24544_dp_20_rank_40_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 10, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.019574110582470893, + "vlm_logprob": -0.019574110582470893, + "vlm_yes_no_margin": 4.625, + "vlm_token_credit": -0.019574110582470893, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned at a low angle, shot from the ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene set in a dark bamboo forest with tall, emerald green stalks forming a dense dome?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a massive, muscular white tiger with black stripes pounce from the shadows onto slippery fallen leaves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the tiger shown in mid-air with its front paws thrust forward and claws fully extended, charging straight at the camera?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the lighting characterized by high-contrast light and shadow with a cool overall tone?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.06718553602695465 + ], + "opposite_logprob_tokens": [ + -3.442185640335083 + ], + "logprob_reward": -0.06718553602695465, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.06718553602695465, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07411614805459976, + 0.24896666407585144, + 0.057275112718343735, + 0.2477504312992096, + 0.02110053412616253, + 0.19344034790992737, + 0.01132502593100071, + 0.04027744010090828, + 0.04984520375728607, + 0.05590318143367767 + ], + "credit_map": [ + 0.007103305775672197, + 0.005781244952231646, + 0.012649910524487495, + 0.007018775679171085, + 0.010612930171191692, + 0.013506625778973103, + 0.010196950286626816, + 0.007246396504342556, + 0.012677299790084362, + 0.006675443612039089, + 0.03913768380880356, + 0.009163578040897846, + 0.012189364992082119, + 0.045970991253852844, + 0.10321762412786484, + 0.019934648647904396, + 0.0024071831721812487, + 0.001991418655961752, + 0.00785063486546278, + 0.004380374681204557, + 0.0031426921486854553, + 0.01190352626144886, + 0.02133007161319256, + 0.004269205499440432, + 0.0196949765086174, + 0.013138115406036377, + 0.03363976627588272, + 0.024737214669585228, + 0.020207080990076065, + 0.057462774217128754, + 0.03440055623650551, + 0.0444699265062809, + 0.0027198379393666983, + 0.0020158966071903706, + 0.003350313287228346, + 0.002373333089053631, + 0.0027013150975108147, + 0.002795059932395816, + 0.0026035644114017487, + 0.0025412128306925297, + 0.04019421711564064, + 0.01143925916403532, + 0.040219470858573914, + 0.007729016710072756, + 0.011636541225016117, + 0.032759491354227066, + 0.009211788885295391, + 0.04025054723024368, + 0.0019345030887052417, + 0.0011631258530542254, + 0.0015859303530305624, + 0.001538111362606287, + 0.001060779090039432, + 0.0013530125143006444, + 0.0011387132108211517, + 0.001550849643535912, + 0.005745531991124153, + 0.0023921076208353043, + 0.011138814501464367, + 0.002838841639459133, + 0.003348486963659525, + 0.005821644328534603, + 0.0028910785913467407, + 0.006100931204855442, + 0.007356484420597553, + 0.004187731072306633, + 0.008838899433612823, + 0.006880358327180147, + 0.004108142573386431, + 0.008504295721650124, + 0.003787609050050378, + 0.006181679666042328, + 0.006469910964369774, + 0.006671705283224583, + 0.014778822660446167, + 0.002849434968084097, + 0.0026316174771636724, + 0.007102178875356913, + 0.0038342878688126802, + 0.01156521961092949 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 3, + 5 + ], + "selected_reward": -0.06718553602695465 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned at a low angle, shot from the ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene set in a dark bamboo forest with tall, emerald green stalks forming a dense dome?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a massive, muscular white tiger with black stripes pounce from the shadows onto slippery fallen leaves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the tiger shown in mid-air with its front paws thrust forward and claws fully extended, charging straight at the camera?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the lighting characterized by high-contrast light and shadow with a cool overall tone?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.007643027231097221 + ], + "opposite_logprob_tokens": [ + -4.882643222808838 + ], + "logprob_reward": -0.007643027231097221, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.007643027231097221, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09673888236284256, + 0.5068871974945068, + 0.04481707140803337, + 0.11247031390666962, + 0.03940592333674431, + 0.05724338814616203, + 0.0281915795058012, + 0.03586840629577637, + 0.02978822961449623, + 0.04858902841806412 + ], + "credit_map": [ + 0.007584025617688894, + 0.006992161273956299, + 0.014885099604725838, + 0.010313819162547588, + 0.009841401129961014, + 0.015176115557551384, + 0.020070094615221024, + 0.011876163072884083, + 0.0210218895226717, + 0.01087079755961895, + 0.07498753070831299, + 0.017014753073453903, + 0.023786934092640877, + 0.11686470359563828, + 0.20716820657253265, + 0.03517237305641174, + 0.0022016940638422966, + 0.0017019199440255761, + 0.006490339059382677, + 0.004270206205546856, + 0.0025062316562980413, + 0.008745317347347736, + 0.015911562368273735, + 0.0029898020438849926, + 0.010006231255829334, + 0.0058211395516991615, + 0.015318203717470169, + 0.01182984746992588, + 0.009154481813311577, + 0.028933009132742882, + 0.012298781424760818, + 0.01910862885415554, + 0.004665958229452372, + 0.002438881201669574, + 0.006817437708377838, + 0.0033519312273710966, + 0.0030103318858891726, + 0.008775165304541588, + 0.0043356409296393394, + 0.0060105761513113976, + 0.011846700683236122, + 0.004073506221175194, + 0.011758945882320404, + 0.0034709139727056026, + 0.0037255259230732918, + 0.008290371857583523, + 0.003961984533816576, + 0.010115441866219044, + 0.0038104199338704348, + 0.0035489711444824934, + 0.0034288610331714153, + 0.0028431431856006384, + 0.0036891945637762547, + 0.0040684649720788, + 0.0031615321058779955, + 0.0036409900058060884, + 0.004903488792479038, + 0.003303087316453457, + 0.008046939969062805, + 0.00238309009000659, + 0.0032664192840456963, + 0.006099123042076826, + 0.0027876514941453934, + 0.005078605376183987, + 0.004381377249956131, + 0.002646493259817362, + 0.006331700831651688, + 0.0038123917765915394, + 0.003377832705155015, + 0.0030882416758686304, + 0.0022791940718889236, + 0.003870996180921793, + 0.00597805343568325, + 0.0031402460299432278, + 0.015096445567905903, + 0.0015865571331232786, + 0.0022285240702331066, + 0.00521115493029356, + 0.002902732929214835, + 0.012445315718650818 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 3, + 0 + ], + "selected_reward": -0.007643027231097221 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned at a low angle, shot from the ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene set in a dark bamboo forest with tall, emerald green stalks forming a dense dome?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a massive, muscular white tiger with black stripes pounce from the shadows onto slippery fallen leaves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the tiger shown in mid-air with its front paws thrust forward and claws fully extended, charging straight at the camera?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the lighting characterized by high-contrast light and shadow with a cool overall tone?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.012535941787064075 + ], + "opposite_logprob_tokens": [ + -4.38753604888916 + ], + "logprob_reward": -0.012535941787064075, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.012535941787064075, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.05850477144122124, + 0.08319631963968277, + 0.05111207067966461, + 0.04738027974963188, + 0.06215886399149895, + 0.3324960172176361, + 0.037498265504837036, + 0.1668180525302887, + 0.05011073499917984, + 0.11072467267513275 + ], + "credit_map": [ + 0.006104399915784597, + 0.004298286512494087, + 0.008290729485452175, + 0.00826192181557417, + 0.0061589195393025875, + 0.007046191021800041, + 0.008579008281230927, + 0.009765305556356907, + 0.0042597148567438126, + 0.0035975452046841383, + 0.006542865186929703, + 0.0040491316467523575, + 0.006173399742692709, + 0.02878219448029995, + 0.022608010098338127, + 0.007183443289250135, + 0.0025151181034743786, + 0.0023490770254284143, + 0.006460272707045078, + 0.0076605393551290035, + 0.0028623053804039955, + 0.008071848191320896, + 0.015355129726231098, + 0.005837779026478529, + 0.0033754853066056967, + 0.003040790092200041, + 0.0059483712539076805, + 0.004313056357204914, + 0.0027564724441617727, + 0.011663391254842281, + 0.0043030669912695885, + 0.01197963859885931, + 0.006170183885842562, + 0.005917642265558243, + 0.00815355870872736, + 0.006168834399431944, + 0.008973364718258381, + 0.009193737991154194, + 0.007948551326990128, + 0.009632986970245838, + 0.06581305712461472, + 0.01951669156551361, + 0.06917670369148254, + 0.012984081171452999, + 0.019971154630184174, + 0.057360779494047165, + 0.016406884416937828, + 0.07126665115356445, + 0.004265256226062775, + 0.0036003119312226772, + 0.004787173122167587, + 0.004433643072843552, + 0.004484141711145639, + 0.005038559436798096, + 0.004654718097299337, + 0.006234459113329649, + 0.03007890097796917, + 0.014091764576733112, + 0.03373430296778679, + 0.012908308766782284, + 0.013483753427863121, + 0.026793595403432846, + 0.01498460117727518, + 0.02074280008673668, + 0.005860569886863232, + 0.005683986935764551, + 0.005815648008137941, + 0.007502800319343805, + 0.007860320620238781, + 0.005541130434721708, + 0.005815653596073389, + 0.006030619144439697, + 0.014348973520100117, + 0.008694845251739025, + 0.028412897139787674, + 0.00864966120570898, + 0.008145402185618877, + 0.009347619488835335, + 0.008089409209787846, + 0.02503584325313568 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 7, + 9 + ], + "selected_reward": -0.012535941787064075 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned at a low angle, shot from the ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene set in a dark bamboo forest with tall, emerald green stalks forming a dense dome?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a massive, muscular white tiger with black stripes pounce from the shadows onto slippery fallen leaves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the tiger shown in mid-air with its front paws thrust forward and claws fully extended, charging straight at the camera?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the lighting characterized by high-contrast light and shadow with a cool overall tone?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005251067224889994 + ], + "opposite_logprob_tokens": [ + -5.255250930786133 + ], + "logprob_reward": -0.005251067224889994, + "yes_no_margin_reward": 5.25, + "token_credit_reward": -0.005251067224889994, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08603210002183914, + 0.03914732113480568, + 0.03022637963294983, + 0.03983457386493683, + 0.06918440014123917, + 0.38336503505706787, + 0.02026350051164627, + 0.0785445049405098, + 0.09055965393781662, + 0.1628425419330597 + ], + "credit_map": [ + 0.007138772867619991, + 0.004908070433884859, + 0.01279742643237114, + 0.005746557377278805, + 0.010641884058713913, + 0.013196296989917755, + 0.023325830698013306, + 0.008277254179120064, + 0.002908910857513547, + 0.00288438075222075, + 0.004717659670859575, + 0.003343658521771431, + 0.0043439604341983795, + 0.011480896733701229, + 0.0042783282697200775, + 0.005189522635191679, + 0.00202299770899117, + 0.002063408959656954, + 0.00393776036798954, + 0.0065318867564201355, + 0.0026529631577432156, + 0.0036844140850007534, + 0.006471623200923204, + 0.0028613244649022818, + 0.005522876046597958, + 0.0023032845929265022, + 0.007643519435077906, + 0.004556854721158743, + 0.0029775763396173716, + 0.006258894689381123, + 0.005047904793173075, + 0.005523662082850933, + 0.00739951990544796, + 0.00485297990962863, + 0.008404005318880081, + 0.00886447262018919, + 0.007140440866351128, + 0.010023972019553185, + 0.012503188103437424, + 0.009995820000767708, + 0.0795619860291481, + 0.024441810324788094, + 0.07821564376354218, + 0.0211950670927763, + 0.02422219142317772, + 0.05955066159367561, + 0.024552829563617706, + 0.07162486016750336, + 0.002778943395242095, + 0.0018203547224402428, + 0.002628563903272152, + 0.002109150169417262, + 0.0028426209464669228, + 0.0026808453258126974, + 0.0024418989196419716, + 0.0029611221980303526, + 0.009086166508495808, + 0.0048484597355127335, + 0.02327008731663227, + 0.0053016371093690395, + 0.005957571789622307, + 0.011701608076691628, + 0.0073426468297839165, + 0.011036324314773083, + 0.014445521868765354, + 0.00774385454133153, + 0.012784660793840885, + 0.011288999579846859, + 0.009364807046949863, + 0.012445459142327309, + 0.009264777414500713, + 0.013221561908721924, + 0.0212097130715847, + 0.007492974866181612, + 0.0517272874712944, + 0.005717993713915348, + 0.0071909199468791485, + 0.016659392043948174, + 0.009265992790460587, + 0.043578263372182846 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 9, + 8 + ], + "selected_reward": -0.005251067224889994 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned at a low angle, shot from the ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene set in a dark bamboo forest with tall, emerald green stalks forming a dense dome?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a massive, muscular white tiger with black stripes pounce from the shadows onto slippery fallen leaves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the tiger shown in mid-air with its front paws thrust forward and claws fully extended, charging straight at the camera?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the lighting characterized by high-contrast light and shadow with a cool overall tone?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005254980642348528 + ], + "opposite_logprob_tokens": [ + -5.255254745483398 + ], + "logprob_reward": -0.005254980642348528, + "yes_no_margin_reward": 5.25, + "token_credit_reward": -0.005254980642348528, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.23245477676391602, + 0.1113988384604454, + 0.11831015348434448, + 0.06039775162935257, + 0.03204379230737686, + 0.0905320942401886, + 0.049207355827093124, + 0.06913048774003983, + 0.07139313220977783, + 0.16513165831565857 + ], + "credit_map": [ + 0.01769007369875908, + 0.01172711793333292, + 0.04012338072061539, + 0.01702781952917576, + 0.02141265757381916, + 0.02092790976166725, + 0.07883617281913757, + 0.02470964938402176, + 0.011182096786797047, + 0.008088238537311554, + 0.015504157170653343, + 0.00785116758197546, + 0.011808821931481361, + 0.027228783816099167, + 0.014766695909202099, + 0.014968873001635075, + 0.00687992200255394, + 0.004662666469812393, + 0.01766398176550865, + 0.00897279754281044, + 0.0063514066860079765, + 0.01560832280665636, + 0.04507813975214958, + 0.01309292297810316, + 0.006261568516492844, + 0.003882475197315216, + 0.011696279980242252, + 0.007102095987647772, + 0.004942277446389198, + 0.008133688941597939, + 0.009872385300695896, + 0.00850698072463274, + 0.0049486528150737286, + 0.003749711671844125, + 0.004270029719918966, + 0.0038750472012907267, + 0.0038174607325345278, + 0.003830699250102043, + 0.0038253357633948326, + 0.0037268565502017736, + 0.01703852415084839, + 0.006804121658205986, + 0.016718823462724686, + 0.007697409018874168, + 0.005953952670097351, + 0.011897275224328041, + 0.007724348455667496, + 0.01669764332473278, + 0.008385563269257545, + 0.0051271324045956135, + 0.008883061818778515, + 0.003149254247546196, + 0.005611212458461523, + 0.007254552561789751, + 0.004334667697548866, + 0.0064619132317602634, + 0.009518388658761978, + 0.008473528549075127, + 0.016037359833717346, + 0.0047239502891898155, + 0.006083453074097633, + 0.009353977628052235, + 0.005655789747834206, + 0.00928403902798891, + 0.010091176256537437, + 0.006673640571534634, + 0.011879034340381622, + 0.009108348749577999, + 0.006257808767259121, + 0.010996988043189049, + 0.005584611091762781, + 0.010801517404615879, + 0.017938602715730667, + 0.024050170555710793, + 0.04444101080298424, + 0.007583406288176775, + 0.009431014768779278, + 0.020321568474173546, + 0.008006111718714237, + 0.03335976228117943 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 2 + ], + "selected_reward": -0.005254980642348528 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "low, static shot, shot from the ground, immerses the viewer in the depths of a dark and oppressive bamboo forest. Tall, emerald green bamboo stalks intertwine densely, forming a natural dome. Only faint, cold, moonlit rays penetrate the gaps, casting dappled shadows on the damp, decaying leaves. Suddenly, with lightning speed, a massive, muscular white tiger pounces from the deepest shadows of the bamboo forest. Its powerful hind legs stomp on the slippery fallen leaves, its entire explosive body stretched out in mid-air, like a flash of white and black lightning, shattering the silence of the forest. Its pure white fur, dotted with deep black stripes, stands out in the dim light. Its icy blue eyes flash with a deadly coldness. Its mouth opens sharply, revealing canine teeth as sharp as daggers. Its front paws thrust forward, claws fully extended, charging straight at the camera. The action is swift and decisive, full of the unrivaled aura of a predator. The high-contrast light and shadow and the cool overall tone create a heart-stopping sense of tension and primitive wild beauty.", + "seed": 24545, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000071/group_10/index_7020_seed_24545_dp_21_rank_42_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 10, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.02155723173636943, + "vlm_logprob": -0.02155723173636943, + "vlm_yes_no_margin": 4.925, + "vlm_token_credit": -0.02155723173636943, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned at a low angle, shot from the ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene set in a dark bamboo forest with tall, emerald green stalks forming a dense dome?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a massive, muscular white tiger with black stripes pounce from the shadows onto slippery fallen leaves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the tiger shown in mid-air with its front paws thrust forward and claws fully extended, charging straight at the camera?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the lighting characterized by high-contrast light and shadow with a cool overall tone?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0854433923959732 + ], + "opposite_logprob_tokens": [ + -3.2104434967041016 + ], + "logprob_reward": -0.0854433923959732, + "yes_no_margin_reward": 3.125, + "token_credit_reward": -0.0854433923959732, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15513606369495392, + 0.04505551978945732, + 0.07624556124210358, + 0.06319330632686615, + 0.08905594050884247, + 0.15089024603366852, + 0.2121935933828354, + 0.09988652914762497, + 0.05457863211631775, + 0.053764600306749344 + ], + "credit_map": [ + 0.011275026947259903, + 0.01792784407734871, + 0.02085126005113125, + 0.01937965489923954, + 0.023477982729673386, + 0.010448098182678223, + 0.021827926859259605, + 0.029948275536298752, + 0.004109981469810009, + 0.006274422165006399, + 0.005843953229486942, + 0.008099301718175411, + 0.004334751516580582, + 0.0038285665214061737, + 0.007862616330385208, + 0.0047019245103001595, + 0.006444504018872976, + 0.006798173766583204, + 0.009490038268268108, + 0.007994390092790127, + 0.0160862784832716, + 0.006498547270894051, + 0.00917956791818142, + 0.01375405490398407, + 0.004725996404886246, + 0.006321858614683151, + 0.007943899370729923, + 0.007121525704860687, + 0.005339946132153273, + 0.006540945265442133, + 0.014365770854055882, + 0.010833362117409706, + 0.00471684243530035, + 0.007541272789239883, + 0.013837399892508984, + 0.005342279095202684, + 0.00553548289462924, + 0.006307120900601149, + 0.029799383133649826, + 0.015976155176758766, + 0.02369655855000019, + 0.013440296053886414, + 0.021482350304722786, + 0.013208495453000069, + 0.0128543172031641, + 0.0315123088657856, + 0.012087313458323479, + 0.022608596831560135, + 0.021443719044327736, + 0.025325307622551918, + 0.029447881504893303, + 0.023806210607290268, + 0.028102293610572815, + 0.027027912437915802, + 0.03269625082612038, + 0.024344034492969513, + 0.008327452465891838, + 0.004663468338549137, + 0.031445879489183426, + 0.004166787955909967, + 0.005077110603451729, + 0.03527239337563515, + 0.004173422232270241, + 0.006760008167475462, + 0.004404788836836815, + 0.0057090152986347675, + 0.00803367979824543, + 0.006690701469779015, + 0.008275015279650688, + 0.005703068803995848, + 0.008194465190172195, + 0.007567896042019129, + 0.0071909865364432335, + 0.005856646224856377, + 0.0062733665108680725, + 0.005774856545031071, + 0.007532070856541395, + 0.0063371118158102036, + 0.0076863872818648815, + 0.007113175000995398 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 0, + 5 + ], + "selected_reward": -0.0854433923959732 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned at a low angle, shot from the ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene set in a dark bamboo forest with tall, emerald green stalks forming a dense dome?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a massive, muscular white tiger with black stripes pounce from the shadows onto slippery fallen leaves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the tiger shown in mid-air with its front paws thrust forward and claws fully extended, charging straight at the camera?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the lighting characterized by high-contrast light and shadow with a cool overall tone?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.009795809164643288 + ], + "opposite_logprob_tokens": [ + -4.634795665740967 + ], + "logprob_reward": -0.009795809164643288, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.009795809164643288, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.28448665142059326, + 0.10035359114408493, + 0.06511291116476059, + 0.12581264972686768, + 0.13236500322818756, + 0.05932890623807907, + 0.0461634136736393, + 0.0631645992398262, + 0.08861104398965836, + 0.03460126742720604 + ], + "credit_map": [ + 0.017188379541039467, + 0.031740494072437286, + 0.03755057230591774, + 0.029795264825224876, + 0.05859595909714699, + 0.027912110090255737, + 0.04141399264335632, + 0.04028988257050514, + 0.009398168884217739, + 0.009622381068766117, + 0.013555451296269894, + 0.012625759467482567, + 0.012247846461832523, + 0.011923003010451794, + 0.014621919952332973, + 0.016359059140086174, + 0.003614748129621148, + 0.004594364669173956, + 0.006950022652745247, + 0.005016944371163845, + 0.0065028429962694645, + 0.007211815565824509, + 0.01503247581422329, + 0.0161896999925375, + 0.00793842226266861, + 0.006860359106212854, + 0.04290279746055603, + 0.009761277586221695, + 0.007255473639816046, + 0.007156602572649717, + 0.013938198797404766, + 0.029999518766999245, + 0.009151761420071125, + 0.00616074213758111, + 0.03351559117436409, + 0.00642450200393796, + 0.01177846360951662, + 0.010850858874619007, + 0.04307135194540024, + 0.011411737650632858, + 0.008469718508422375, + 0.004639442078769207, + 0.009080239571630955, + 0.0039030457846820354, + 0.005211811978369951, + 0.013339088298380375, + 0.004927570465952158, + 0.009757987223565578, + 0.005633190739899874, + 0.0053077274933457375, + 0.006880202330648899, + 0.0047377366572618484, + 0.005531370639801025, + 0.005232290364801884, + 0.006820141337811947, + 0.006020756904035807, + 0.008099956437945366, + 0.005236751865595579, + 0.012939654290676117, + 0.004952580202370882, + 0.006042621098458767, + 0.013872063718736172, + 0.005791978444904089, + 0.0062289913184940815, + 0.005198675207793713, + 0.00835714116692543, + 0.0157820712774992, + 0.012038796208798885, + 0.010443543083965778, + 0.008349940180778503, + 0.01651444286108017, + 0.011926431208848953, + 0.003481585066765547, + 0.00327119417488575, + 0.003598120529204607, + 0.003982126247137785, + 0.006179529242217541, + 0.0042808628641068935, + 0.006094896234571934, + 0.0037129539996385574 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 3 + ], + "selected_reward": -0.009795809164643288 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned at a low angle, shot from the ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene set in a dark bamboo forest with tall, emerald green stalks forming a dense dome?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a massive, muscular white tiger with black stripes pounce from the shadows onto slippery fallen leaves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the tiger shown in mid-air with its front paws thrust forward and claws fully extended, charging straight at the camera?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the lighting characterized by high-contrast light and shadow with a cool overall tone?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.00673500495031476 + ], + "opposite_logprob_tokens": [ + -5.006734848022461 + ], + "logprob_reward": -0.00673500495031476, + "yes_no_margin_reward": 5.0, + "token_credit_reward": -0.00673500495031476, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.18782523274421692, + 0.04522828012704849, + 0.09959474205970764, + 0.07707206904888153, + 0.1335667073726654, + 0.1098480224609375, + 0.12500634789466858, + 0.10508016496896744, + 0.0735294371843338, + 0.043249018490314484 + ], + "credit_map": [ + 0.01787526160478592, + 0.020569292828440666, + 0.026024391874670982, + 0.03262186795473099, + 0.019302062690258026, + 0.015315369702875614, + 0.028674663975834846, + 0.027442313730716705, + 0.0045248642563819885, + 0.004618540406227112, + 0.006500570569187403, + 0.00703062629327178, + 0.0063419099897146225, + 0.0034532975405454636, + 0.006611032877117395, + 0.006147436331957579, + 0.005492879077792168, + 0.006511659361422062, + 0.01024325005710125, + 0.0072848014533519745, + 0.00916432123631239, + 0.007688015699386597, + 0.02423138916492462, + 0.028978437185287476, + 0.0043926541693508625, + 0.006152759306132793, + 0.02311812713742256, + 0.006443530321121216, + 0.007054190617054701, + 0.005197864957153797, + 0.006553757470101118, + 0.018159184604883194, + 0.007712614256888628, + 0.006828268989920616, + 0.012880401685833931, + 0.00654191616922617, + 0.008200984448194504, + 0.007316227070987225, + 0.04687612131237984, + 0.037210166454315186, + 0.013502368703484535, + 0.011255884543061256, + 0.018325652927160263, + 0.009365043602883816, + 0.008360298350453377, + 0.02046765387058258, + 0.007640088442713022, + 0.02093103528022766, + 0.012883597053587437, + 0.017284216359257698, + 0.020311491563916206, + 0.012829506769776344, + 0.013239027000963688, + 0.012168951332569122, + 0.015019535087049007, + 0.021270016208291054, + 0.012737283483147621, + 0.009135250933468342, + 0.016278836876153946, + 0.013176641426980495, + 0.011742948554456234, + 0.01607467606663704, + 0.016382811591029167, + 0.009551714174449444, + 0.006812975276261568, + 0.004992367699742317, + 0.01232963427901268, + 0.0077795046381652355, + 0.00790180079638958, + 0.008534151129424572, + 0.014574087224900723, + 0.010604921728372574, + 0.00484246714040637, + 0.005133458413183689, + 0.0048415083438158035, + 0.0052422694861888885, + 0.00592429842799902, + 0.005036042537540197, + 0.0067786723375320435, + 0.005450305063277483 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 6 + ], + "selected_reward": -0.00673500495031476 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned at a low angle, shot from the ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene set in a dark bamboo forest with tall, emerald green stalks forming a dense dome?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a massive, muscular white tiger with black stripes pounce from the shadows onto slippery fallen leaves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the tiger shown in mid-air with its front paws thrust forward and claws fully extended, charging straight at the camera?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the lighting characterized by high-contrast light and shadow with a cool overall tone?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004095382057130337 + ], + "opposite_logprob_tokens": [ + -5.504095554351807 + ], + "logprob_reward": -0.004095382057130337, + "yes_no_margin_reward": 5.5, + "token_credit_reward": -0.004095382057130337, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06576644629240036, + 0.025544730946421623, + 0.03712902590632439, + 0.04618040844798088, + 0.10279999673366547, + 0.1608450561761856, + 0.3299919366836548, + 0.05019357427954674, + 0.13322314620018005, + 0.048325736075639725 + ], + "credit_map": [ + 0.0062868595123291016, + 0.006344578228890896, + 0.00993989035487175, + 0.009069690480828285, + 0.009567584842443466, + 0.00444386713206768, + 0.011018458753824234, + 0.009095514193177223, + 0.002108062617480755, + 0.0022333883680403233, + 0.0031457499135285616, + 0.003921991214156151, + 0.0037677197251468897, + 0.0025035711005330086, + 0.0036239689216017723, + 0.0042402781546115875, + 0.003542265621945262, + 0.0045862686820328236, + 0.004921967163681984, + 0.00564867677167058, + 0.0036228392273187637, + 0.004520421847701073, + 0.005460190586745739, + 0.0048263948410749435, + 0.003462989116087556, + 0.0049956743605434895, + 0.009729916229844093, + 0.004589430056512356, + 0.005782051011919975, + 0.004494648426771164, + 0.004549875855445862, + 0.008575821295380592, + 0.008122370578348637, + 0.006617498118430376, + 0.028266379609704018, + 0.006252993829548359, + 0.00776346679776907, + 0.007804710417985916, + 0.027506202459335327, + 0.010466375388205051, + 0.025189245119690895, + 0.015476890839636326, + 0.02650604024529457, + 0.016723722219467163, + 0.013238574378192425, + 0.0278385691344738, + 0.012227038852870464, + 0.023644983768463135, + 0.03485511243343353, + 0.03290416672825813, + 0.05860697105526924, + 0.02969272993505001, + 0.029974879696965218, + 0.03572435677051544, + 0.04355712607502937, + 0.06467659026384354, + 0.005654764361679554, + 0.005592959467321634, + 0.007228214293718338, + 0.0044780755415558815, + 0.008426058106124401, + 0.007004813756793737, + 0.00819394551217556, + 0.0036147458013147116, + 0.008498508483171463, + 0.013283319771289825, + 0.023361189290881157, + 0.01789994351565838, + 0.016142407432198524, + 0.011903275735676289, + 0.025068288668990135, + 0.017066210508346558, + 0.006059034261852503, + 0.004173509310930967, + 0.006650594063103199, + 0.005022445693612099, + 0.006977055687457323, + 0.006407636217772961, + 0.008643228560686111, + 0.0043922364711761475 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 5, + 8 + ], + "selected_reward": -0.004095382057130337 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned at a low angle, shot from the ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene set in a dark bamboo forest with tall, emerald green stalks forming a dense dome?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a massive, muscular white tiger with black stripes pounce from the shadows onto slippery fallen leaves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the tiger shown in mid-air with its front paws thrust forward and claws fully extended, charging straight at the camera?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the lighting characterized by high-contrast light and shadow with a cool overall tone?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.001716570113785565 + ], + "opposite_logprob_tokens": [ + -6.376716613769531 + ], + "logprob_reward": -0.001716570113785565, + "yes_no_margin_reward": 6.375, + "token_credit_reward": -0.001716570113785565, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13794243335723877, + 0.06583580374717712, + 0.14589671790599823, + 0.1606854796409607, + 0.07083509862422943, + 0.05504317954182625, + 0.0979592502117157, + 0.10133767127990723, + 0.10637632012367249, + 0.0580880381166935 + ], + "credit_map": [ + 0.010395224206149578, + 0.015887662768363953, + 0.024616748094558716, + 0.01998649537563324, + 0.021185118705034256, + 0.010053934529423714, + 0.017134258523583412, + 0.01868297904729843, + 0.005045892205089331, + 0.007226939778774977, + 0.010606321506202221, + 0.009365534409880638, + 0.008333822712302208, + 0.005058520473539829, + 0.010653004981577396, + 0.009545771405100822, + 0.010318750515580177, + 0.012496409937739372, + 0.013876642100512981, + 0.012661401182413101, + 0.027223289012908936, + 0.012239864096045494, + 0.034463416785001755, + 0.022616952657699585, + 0.011010888032615185, + 0.0077506923116743565, + 0.05442733317613602, + 0.011418270878493786, + 0.00895377341657877, + 0.010086105205118656, + 0.017103418707847595, + 0.039934996515512466, + 0.005642654374241829, + 0.004670613445341587, + 0.013166552409529686, + 0.003857507137581706, + 0.006940113380551338, + 0.008652245625853539, + 0.018476907163858414, + 0.009428506717085838, + 0.005985686089843512, + 0.006041559856384993, + 0.007066172547638416, + 0.006441107019782066, + 0.0049185254611074924, + 0.009787194430828094, + 0.007442381698638201, + 0.007360553368926048, + 0.008819346316158772, + 0.012767861597239971, + 0.012406669557094574, + 0.011263399384915829, + 0.012604025192558765, + 0.009499242529273033, + 0.016856810078024864, + 0.013741900213062763, + 0.008837063796818256, + 0.004281588830053806, + 0.031953636556863785, + 0.0045705377124249935, + 0.005758124869316816, + 0.03546709194779396, + 0.004398518241941929, + 0.006071098614484072, + 0.012510878965258598, + 0.01636219397187233, + 0.02064845897257328, + 0.00978444330394268, + 0.007718788925558329, + 0.012707769870758057, + 0.011315216310322285, + 0.015328570269048214, + 0.010681352578103542, + 0.007536701858043671, + 0.006300642620772123, + 0.00492869783192873, + 0.0074241007678210735, + 0.007551735732704401, + 0.009136452339589596, + 0.004528358578681946 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 2, + 0 + ], + "selected_reward": -0.001716570113785565 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "low, static shot, shot from the ground, immerses the viewer in the depths of a dark and oppressive bamboo forest. Tall, emerald green bamboo stalks intertwine densely, forming a natural dome. Only faint, cold, moonlit rays penetrate the gaps, casting dappled shadows on the damp, decaying leaves. Suddenly, with lightning speed, a massive, muscular white tiger pounces from the deepest shadows of the bamboo forest. Its powerful hind legs stomp on the slippery fallen leaves, its entire explosive body stretched out in mid-air, like a flash of white and black lightning, shattering the silence of the forest. Its pure white fur, dotted with deep black stripes, stands out in the dim light. Its icy blue eyes flash with a deadly coldness. Its mouth opens sharply, revealing canine teeth as sharp as daggers. Its front paws thrust forward, claws fully extended, charging straight at the camera. The action is swift and decisive, full of the unrivaled aura of a predator. The high-contrast light and shadow and the cool overall tone create a heart-stopping sense of tension and primitive wild beauty.", + "seed": 24546, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000071/group_10/index_7020_seed_24546_dp_21_rank_42_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 10, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.023051506746560334, + "vlm_logprob": -0.023051506746560334, + "vlm_yes_no_margin": 4.525, + "vlm_token_credit": -0.023051506746560334, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned at a low angle, shot from the ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene set in a dark bamboo forest with tall, emerald green stalks forming a dense dome?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a massive, muscular white tiger with black stripes pounce from the shadows onto slippery fallen leaves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the tiger shown in mid-air with its front paws thrust forward and claws fully extended, charging straight at the camera?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the lighting characterized by high-contrast light and shadow with a cool overall tone?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.077882319688797 + ], + "opposite_logprob_tokens": [ + -3.8278822898864746 + ], + "logprob_reward": -0.077882319688797, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.077882319688797, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.17872287333011627, + 0.05417785048484802, + 0.053980711847543716, + 0.026050925254821777, + 0.017200054600834846, + 0.01880306378006935, + 0.3489413857460022, + 0.028439976274967194, + 0.07216290384531021, + 0.20152029395103455 + ], + "credit_map": [ + 0.015740497037768364, + 0.011098112910985947, + 0.01708928681910038, + 0.01180009264498949, + 0.022327609360218048, + 0.03279540687799454, + 0.032092660665512085, + 0.0357792042195797, + 0.004943821579217911, + 0.006295854691416025, + 0.008538148365914822, + 0.009153461083769798, + 0.004886465147137642, + 0.008566364645957947, + 0.006402341183274984, + 0.005391390528529882, + 0.0039045766461640596, + 0.0023806949611753225, + 0.00286434986628592, + 0.004204270895570517, + 0.004492766223847866, + 0.017198478803038597, + 0.00454500550404191, + 0.014390566386282444, + 0.0037101898342370987, + 0.0025762387085705996, + 0.0027269062120467424, + 0.002192459534853697, + 0.00357816438190639, + 0.0051538897678256035, + 0.003281591460108757, + 0.0028314844239503145, + 0.0019961015786975622, + 0.001914937049150467, + 0.0021042427979409695, + 0.0030387931037694216, + 0.0014457469806075096, + 0.0026329588145017624, + 0.002204226329922676, + 0.0018630492268130183, + 0.0017640426522120833, + 0.002474689856171608, + 0.0021520517766475677, + 0.0033146897330880165, + 0.0017693392001092434, + 0.002307348418980837, + 0.002347608795389533, + 0.0026732920669019222, + 0.015739427879452705, + 0.011316992342472076, + 0.031707003712654114, + 0.010528329759836197, + 0.01114464458078146, + 0.12621353566646576, + 0.008136310614645481, + 0.13415510952472687, + 0.0027434767689555883, + 0.0025049701798707247, + 0.0023261161986738443, + 0.0034373044036328793, + 0.006670748814940453, + 0.0033386487048119307, + 0.004747836384922266, + 0.002670874120667577, + 0.005595771130174398, + 0.0038271727971732616, + 0.008766247890889645, + 0.002513451501727104, + 0.003162301378324628, + 0.02242301218211651, + 0.0027203222271054983, + 0.02315462753176689, + 0.018721502274274826, + 0.004113547969609499, + 0.07464577257633209, + 0.011120332404971123, + 0.008873816579580307, + 0.07198458164930344, + 0.003573774127289653, + 0.00848696194589138 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 9, + 0 + ], + "selected_reward": -0.077882319688797 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned at a low angle, shot from the ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene set in a dark bamboo forest with tall, emerald green stalks forming a dense dome?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a massive, muscular white tiger with black stripes pounce from the shadows onto slippery fallen leaves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the tiger shown in mid-air with its front paws thrust forward and claws fully extended, charging straight at the camera?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the lighting characterized by high-contrast light and shadow with a cool overall tone?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.009800058789551258 + ], + "opposite_logprob_tokens": [ + -4.634799957275391 + ], + "logprob_reward": -0.009800058789551258, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.009800058789551258, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09681713581085205, + 0.14116664230823517, + 0.11807425320148468, + 0.019686194136738777, + 0.020703841000795364, + 0.0321103110909462, + 0.17819912731647491, + 0.06235906854271889, + 0.07072605937719345, + 0.26015743613243103 + ], + "credit_map": [ + 0.008318980224430561, + 0.005497774574905634, + 0.008934355340898037, + 0.00925733894109726, + 0.006666639354079962, + 0.019583435729146004, + 0.016871390864253044, + 0.0216872189193964, + 0.013006012886762619, + 0.012838023714721203, + 0.02652369625866413, + 0.016005808487534523, + 0.012735405936837196, + 0.02758653461933136, + 0.018954670056700706, + 0.013516481034457684, + 0.007388828322291374, + 0.00473202345892787, + 0.006414854899048805, + 0.006635400932282209, + 0.008806626312434673, + 0.038846857845783234, + 0.00850353017449379, + 0.03674612566828728, + 0.0027369263116270304, + 0.0018235464813187718, + 0.00291411648504436, + 0.0019914170261472464, + 0.0017605910543352365, + 0.0026477135252207518, + 0.0032793486025184393, + 0.002532533835619688, + 0.0032805821392685175, + 0.002702409401535988, + 0.002848658012226224, + 0.0026342892087996006, + 0.002099934034049511, + 0.002303755609318614, + 0.0022922807838767767, + 0.0025419313460588455, + 0.0038889236748218536, + 0.002231532009318471, + 0.004989915527403355, + 0.0019393516704440117, + 0.0023709828965365887, + 0.006176351103931665, + 0.0018190355040133, + 0.008694217540323734, + 0.007201562635600567, + 0.00617721863090992, + 0.016117332503199577, + 0.006874213460832834, + 0.006645950488746166, + 0.06259073317050934, + 0.004762233234941959, + 0.06782987713813782, + 0.005233731120824814, + 0.006381509825587273, + 0.004683529958128929, + 0.006005282513797283, + 0.01529709529131651, + 0.009849483147263527, + 0.008277902379631996, + 0.006630532443523407, + 0.004441999364644289, + 0.002444810699671507, + 0.007837659679353237, + 0.0016112668672576547, + 0.002953536808490753, + 0.024103736504912376, + 0.0022915785666555166, + 0.025041470304131508, + 0.025024278089404106, + 0.007437485735863447, + 0.09485757350921631, + 0.014962584711611271, + 0.012304525822401047, + 0.0878969058394432, + 0.005805560387670994, + 0.01186851691454649 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 6, + 1 + ], + "selected_reward": -0.009800058789551258 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned at a low angle, shot from the ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene set in a dark bamboo forest with tall, emerald green stalks forming a dense dome?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a massive, muscular white tiger with black stripes pounce from the shadows onto slippery fallen leaves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the tiger shown in mid-air with its front paws thrust forward and claws fully extended, charging straight at the camera?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the lighting characterized by high-contrast light and shadow with a cool overall tone?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.009783060289919376 + ], + "opposite_logprob_tokens": [ + -4.6347832679748535 + ], + "logprob_reward": -0.009783060289919376, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.009783060289919376, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08837364614009857, + 0.060841672122478485, + 0.028384696692228317, + 0.0626787394285202, + 0.02022569254040718, + 0.032953836023807526, + 0.27241721749305725, + 0.07489617168903351, + 0.16595333814620972, + 0.19327491521835327 + ], + "credit_map": [ + 0.006175799295306206, + 0.008758174255490303, + 0.010637045837938786, + 0.018137257546186447, + 0.009798587299883366, + 0.012846295721828938, + 0.01268343161791563, + 0.009337058290839195, + 0.004205197095870972, + 0.005436343606561422, + 0.007607391104102135, + 0.008739246055483818, + 0.005351661704480648, + 0.01638428308069706, + 0.007743364199995995, + 0.005374186672270298, + 0.0028483220376074314, + 0.001706001115962863, + 0.002186110243201256, + 0.002108369953930378, + 0.0026933588087558746, + 0.007861792109906673, + 0.0032690854277461767, + 0.0057116602547466755, + 0.006619535852223635, + 0.007744381669908762, + 0.005431856494396925, + 0.00501226307824254, + 0.01445662509649992, + 0.008950188755989075, + 0.008756807073950768, + 0.0057070814073085785, + 0.002685907296836376, + 0.0025159211363643408, + 0.002302954439073801, + 0.0031884273048490286, + 0.0017533485079184175, + 0.0023491857573390007, + 0.0029894071631133556, + 0.0024405401200056076, + 0.004835909232497215, + 0.003963003400713205, + 0.0042458889074623585, + 0.005081462673842907, + 0.0026256651617586613, + 0.0038504409603774548, + 0.003019723342731595, + 0.0053317430429160595, + 0.014743818901479244, + 0.009966231882572174, + 0.023379001766443253, + 0.0064880275167524815, + 0.009386468678712845, + 0.10012536495923996, + 0.00643338356167078, + 0.10189495235681534, + 0.007773369550704956, + 0.006223500706255436, + 0.00796596985310316, + 0.007370969746261835, + 0.014664409682154655, + 0.01153483521193266, + 0.01257280632853508, + 0.0067903115414083, + 0.008837435394525528, + 0.006165712606161833, + 0.01753769814968109, + 0.0037095516454428434, + 0.0058772508054971695, + 0.05751173943281174, + 0.004626390989869833, + 0.06168755143880844, + 0.017669962719082832, + 0.004970189183950424, + 0.06946217268705368, + 0.010454144328832626, + 0.008604559116065502, + 0.07031599432229996, + 0.00410835538059473, + 0.007689557503908873 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 9, + 8 + ], + "selected_reward": -0.009783060289919376 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned at a low angle, shot from the ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene set in a dark bamboo forest with tall, emerald green stalks forming a dense dome?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a massive, muscular white tiger with black stripes pounce from the shadows onto slippery fallen leaves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the tiger shown in mid-air with its front paws thrust forward and claws fully extended, charging straight at the camera?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the lighting characterized by high-contrast light and shadow with a cool overall tone?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.012537825852632523 + ], + "opposite_logprob_tokens": [ + -4.387537956237793 + ], + "logprob_reward": -0.012537825852632523, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.012537825852632523, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15003244578838348, + 0.0335514210164547, + 0.028555523604154587, + 0.04149509221315384, + 0.026245901361107826, + 0.06665617972612381, + 0.0335976742208004, + 0.1482977271080017, + 0.40259242057800293, + 0.06897567957639694 + ], + "credit_map": [ + 0.018907833844423294, + 0.011554148979485035, + 0.013262294232845306, + 0.012879018671810627, + 0.013441485352814198, + 0.025735024362802505, + 0.02173478528857231, + 0.03251785412430763, + 0.003288298612460494, + 0.0022686105221509933, + 0.0035684281028807163, + 0.004694141913205385, + 0.002673584269359708, + 0.00768081471323967, + 0.006811466533690691, + 0.002566073788329959, + 0.0031288007739931345, + 0.0030626326333731413, + 0.002952044364064932, + 0.007352957036346197, + 0.002848429372534156, + 0.0029716615099459887, + 0.0034598999191075563, + 0.0027790949679911137, + 0.006398990284651518, + 0.0049776374362409115, + 0.0037755793891847134, + 0.003946211654692888, + 0.005647447891533375, + 0.007275878917425871, + 0.005721664056181908, + 0.003751680487766862, + 0.003924382384866476, + 0.0023176700342446566, + 0.00278476276434958, + 0.00480095436796546, + 0.002148118568584323, + 0.0030304621905088425, + 0.0038475780747830868, + 0.0033919699490070343, + 0.006119732744991779, + 0.006829398218542337, + 0.008648269809782505, + 0.010480007156729698, + 0.005721446126699448, + 0.010283497162163258, + 0.007523908279836178, + 0.011049912311136723, + 0.0019604554399847984, + 0.0019380549201741815, + 0.0033621673937886953, + 0.002626189962029457, + 0.0019854058045893908, + 0.009012621827423573, + 0.002211712533608079, + 0.010501064360141754, + 0.012537849135696888, + 0.014673758298158646, + 0.011980372481048107, + 0.012812047265470028, + 0.03811878710985184, + 0.02488100528717041, + 0.020172616466879845, + 0.013121269643306732, + 0.019665833562612534, + 0.014085660688579082, + 0.04237031564116478, + 0.005808122456073761, + 0.010960325598716736, + 0.14639945328235626, + 0.010199915617704391, + 0.15310275554656982, + 0.0067137000150978565, + 0.002273938851431012, + 0.023901360109448433, + 0.003875486785545945, + 0.0033321459777653217, + 0.02402968890964985, + 0.0020536112133413553, + 0.002795747946947813 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 0, + 7 + ], + "selected_reward": -0.012537825852632523 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned at a low angle, shot from the ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene set in a dark bamboo forest with tall, emerald green stalks forming a dense dome?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a massive, muscular white tiger with black stripes pounce from the shadows onto slippery fallen leaves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the tiger shown in mid-air with its front paws thrust forward and claws fully extended, charging straight at the camera?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the lighting characterized by high-contrast light and shadow with a cool overall tone?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005254269111901522 + ], + "opposite_logprob_tokens": [ + -5.25525426864624 + ], + "logprob_reward": -0.005254269111901522, + "yes_no_margin_reward": 5.25, + "token_credit_reward": -0.005254269111901522, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06506698578596115, + 0.10212244093418121, + 0.10869795083999634, + 0.023550672456622124, + 0.07277339696884155, + 0.022823195904493332, + 0.2703405022621155, + 0.0651632621884346, + 0.16516432166099548, + 0.10429723560810089 + ], + "credit_map": [ + 0.009008638560771942, + 0.00689611304551363, + 0.006230220664292574, + 0.011798353865742683, + 0.007961688563227654, + 0.008874554187059402, + 0.00879326555877924, + 0.005504157859832048, + 0.007935267873108387, + 0.007643247023224831, + 0.01786061003804207, + 0.007673776242882013, + 0.007857858203351498, + 0.03300156816840172, + 0.010171369649469852, + 0.009978750720620155, + 0.007889333181083202, + 0.005019180476665497, + 0.005861780606210232, + 0.0063356091268360615, + 0.008056827820837498, + 0.036432407796382904, + 0.007831418886780739, + 0.031271398067474365, + 0.002757363487035036, + 0.00228017196059227, + 0.002627612790092826, + 0.003101617330685258, + 0.003747682087123394, + 0.003044764744117856, + 0.0035560389515012503, + 0.002435422735288739, + 0.010635373182594776, + 0.00980236753821373, + 0.008711027912795544, + 0.007080865558236837, + 0.008236327208578587, + 0.010499974712729454, + 0.00941466074436903, + 0.008392803370952606, + 0.00516129843890667, + 0.002954286988824606, + 0.003131190314888954, + 0.0021196608431637287, + 0.0016321371076628566, + 0.0027604098431766033, + 0.0016250897897407413, + 0.0034391237422823906, + 0.011893074959516525, + 0.009002489037811756, + 0.024652760475873947, + 0.00820131879299879, + 0.009045077487826347, + 0.09637827426195145, + 0.006870707962661982, + 0.10429681092500687, + 0.00518026715144515, + 0.006761134136468172, + 0.005223456304520369, + 0.006052599288523197, + 0.016300437971949577, + 0.010801710188388824, + 0.008403448387980461, + 0.0064402129501104355, + 0.007827126421034336, + 0.004964470397680998, + 0.01703687570989132, + 0.002646715147420764, + 0.004411289934068918, + 0.06133708357810974, + 0.004265929106622934, + 0.06267483532428741, + 0.008410166949033737, + 0.004187081009149551, + 0.03496147692203522, + 0.007329599931836128, + 0.004929947666823864, + 0.03569332882761955, + 0.004142174497246742, + 0.0046434588730335236 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 8, + 2 + ], + "selected_reward": -0.005254269111901522 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000071/group_11/index_tensor([7867, 7867])_group_11_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000071/group_11/index_tensor([7867, 7867])_group_11_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..5c2eb782ac2120ad850a9e6d0cd8806c0480ce27 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000071/group_11/index_tensor([7867, 7867])_group_11_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "low-angle tracking shot moves smoothly alongside a massive crimson combine harvester as it churns steadily and relentlessly across a vast field. The camera's position emphasizes the immense scale of the monolithic steel machine, its body coated in a fine dust that dulls its surface in the hazy sunlight. Its massive black rubber tires press deep tracks into the parched soil. In front, a wide, intricate harvester rotates methodically, its metal reel gathering tall, golden wheat stalks onto the precisely moving cutting bar below. The scene is bathed in the warm, diffuse light of the afternoon sun, filtered through a thin atmospheric haze, softening the horizon where ripe wheat fields stretch like a golden ocean. As the machine advances, its wake leaves a neat trail of stubble, while a stream of chopped straw and chaff erupts from the chute behind, catching the amber glow and shimmering in the air. The overall color palette is a rich tapestry of gold, amber, and the crimson of the machinery, creating a powerful yet tranquil atmosphere of agricultural abundance. The entire scene is captured with a subtle, nostalgic film texture that adds a timeless and painterly quality to the powerful, rhythmic movements of the harvest.", + "seed": 9806, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000071/group_11/index_7867_seed_9806_dp_22_rank_44_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 11, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.04568333704955876, + "vlm_logprob": -0.04568333704955876, + "vlm_yes_no_margin": 3.95, + "vlm_token_credit": -0.04568333704955876, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a massive crimson combine harvester?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the harvester moving across a vast field of tall, golden wheat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the harvester's metal reel rotating to gather wheat stalks onto a cutting bar?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the scene feature a low-angle camera shot emphasizing the machine's scale?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the environment bathed in warm, diffuse afternoon sunlight with a thin atmospheric haze?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.023466017097234726 + ], + "opposite_logprob_tokens": [ + -4.023466110229492 + ], + "logprob_reward": -0.023466017097234726, + "yes_no_margin_reward": 4.0, + "token_credit_reward": -0.023466017097234726, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.49776944518089294, + 0.07550054043531418, + 0.06909511983394623, + 0.06508135795593262, + 0.07711677998304367, + 0.07871045917272568, + 0.058696381747722626, + 0.030233141034841537, + 0.033517587929964066, + 0.014279200695455074 + ], + "credit_map": [ + 0.050039492547512054, + 0.054321229457855225, + 0.028782829642295837, + 0.018105169758200645, + 0.15570712089538574, + 0.11534970998764038, + 0.04051404073834419, + 0.03494982793927193, + 0.009024139493703842, + 0.013228334486484528, + 0.007106008008122444, + 0.010475612245500088, + 0.008279559202492237, + 0.007592482957988977, + 0.00878797098994255, + 0.011006436310708523, + 0.007615883368998766, + 0.010217824019491673, + 0.005571866873651743, + 0.011032470501959324, + 0.01274370588362217, + 0.004962652921676636, + 0.009591471403837204, + 0.007359246723353863, + 0.006542761344462633, + 0.009676056914031506, + 0.00687075313180685, + 0.0141226751729846, + 0.00849299505352974, + 0.004550046753138304, + 0.00955794658511877, + 0.005268122535198927, + 0.008151344954967499, + 0.013105225749313831, + 0.007680220529437065, + 0.008785977028310299, + 0.010829250328242779, + 0.005998852197080851, + 0.013494814746081829, + 0.009071094915270805, + 0.011138572357594967, + 0.010000813752412796, + 0.00932853389531374, + 0.01619132235646248, + 0.007059743162244558, + 0.0048260134644806385, + 0.014556670561432838, + 0.005608791019767523, + 0.006302486173808575, + 0.00794992409646511, + 0.00599902356043458, + 0.009470172226428986, + 0.0071860128082334995, + 0.004399503115564585, + 0.010897176340222359, + 0.006492083892226219, + 0.0037216541822999716, + 0.004058683291077614, + 0.003122470108792186, + 0.003701040055602789, + 0.005126014817506075, + 0.0024465364404022694, + 0.004969663452357054, + 0.003087080316618085, + 0.00413254089653492, + 0.004265908617526293, + 0.004632009193301201, + 0.004601823631674051, + 0.0037060093600302935, + 0.0025568949058651924, + 0.006080428138375282, + 0.0035419755149632692, + 0.001568878535181284, + 0.001738701481372118, + 0.0013375317212194204, + 0.002180410549044609, + 0.0017313988646492362, + 0.0012496496783569455, + 0.0030718990601599216, + 0.0014007319696247578 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 5, + 4 + ], + "selected_reward": -0.023466017097234726 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a massive crimson combine harvester?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the harvester moving across a vast field of tall, golden wheat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the harvester's metal reel rotating to gather wheat stalks onto a cutting bar?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the scene feature a low-angle camera shot emphasizing the machine's scale?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the environment bathed in warm, diffuse afternoon sunlight with a thin atmospheric haze?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.16040973365306854 + ], + "opposite_logprob_tokens": [ + -1.910409688949585 + ], + "logprob_reward": -0.16040973365306854, + "yes_no_margin_reward": 1.75, + "token_credit_reward": -0.16040973365306854, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.40359118580818176, + 0.13522493839263916, + 0.04266902804374695, + 0.07665165513753891, + 0.1047518327832222, + 0.10356327146291733, + 0.06504865735769272, + 0.026537751778960228, + 0.02978430688381195, + 0.012177434749901295 + ], + "credit_map": [ + 0.05954906344413757, + 0.04729857295751572, + 0.03759417682886124, + 0.03217126801609993, + 0.04371348395943642, + 0.06451526284217834, + 0.032304711639881134, + 0.08644463121891022, + 0.012028008699417114, + 0.016829559579491615, + 0.013058316893875599, + 0.01573874056339264, + 0.016882741823792458, + 0.01605384796857834, + 0.014769094996154308, + 0.029864618554711342, + 0.004796603694558144, + 0.006694044452160597, + 0.003538897493854165, + 0.004898309241980314, + 0.005886897910386324, + 0.0035141240805387497, + 0.00854356586933136, + 0.004796583205461502, + 0.006993579678237438, + 0.01001492328941822, + 0.010096157900989056, + 0.019867124035954475, + 0.006385653279721737, + 0.005628087557852268, + 0.011812412180006504, + 0.0058537074364721775, + 0.009591016918420792, + 0.016779912635684013, + 0.010321672074496746, + 0.009923090226948261, + 0.011647756211459637, + 0.008020523935556412, + 0.016598766669631004, + 0.02186909131705761, + 0.010448433458805084, + 0.012830073945224285, + 0.008855421096086502, + 0.015035131014883518, + 0.011453989893198013, + 0.0063167414627969265, + 0.025312334299087524, + 0.013311131857335567, + 0.007355355191975832, + 0.0091655058786273, + 0.005025689490139484, + 0.01102079451084137, + 0.005984024610370398, + 0.005318570416420698, + 0.011594077572226524, + 0.009584635496139526, + 0.0032039128709584475, + 0.003344352124258876, + 0.002337245037779212, + 0.004038180224597454, + 0.0031689107418060303, + 0.002092709532007575, + 0.004418328404426575, + 0.003934112377464771, + 0.004206929821521044, + 0.004176344256848097, + 0.002857920480892062, + 0.0040977103635668755, + 0.0035795639269053936, + 0.0025690351612865925, + 0.005023975390940905, + 0.0032728249207139015, + 0.0018016623798757792, + 0.0016257099341601133, + 0.001179501530714333, + 0.0012469090288504958, + 0.001294839195907116, + 0.0011140276910737157, + 0.0023527960292994976, + 0.0015619886107742786 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 4 + ], + "selected_reward": -0.16040973365306854 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a massive crimson combine harvester?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the harvester moving across a vast field of tall, golden wheat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the harvester's metal reel rotating to gather wheat stalks onto a cutting bar?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the scene feature a low-angle camera shot emphasizing the machine's scale?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the environment bathed in warm, diffuse afternoon sunlight with a thin atmospheric haze?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.033723801374435425 + ], + "opposite_logprob_tokens": [ + -3.408723831176758 + ], + "logprob_reward": -0.033723801374435425, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.033723801374435425, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.5812935829162598, + 0.05815109610557556, + 0.02910603955388069, + 0.033639416098594666, + 0.055624715983867645, + 0.13383455574512482, + 0.06858183443546295, + 0.01471509225666523, + 0.013179834000766277, + 0.011873899027705193 + ], + "credit_map": [ + 0.035103633999824524, + 0.06256715953350067, + 0.02860071323812008, + 0.023936251178383827, + 0.23293009400367737, + 0.12006235867738724, + 0.04836457222700119, + 0.029728783294558525, + 0.006307270843535662, + 0.010165995918214321, + 0.005845637060701847, + 0.007968274876475334, + 0.008723863400518894, + 0.00637308182194829, + 0.007245098706334829, + 0.005521873943507671, + 0.0036834790371358395, + 0.005120144225656986, + 0.002343772677704692, + 0.0043597230687737465, + 0.0036886411253362894, + 0.002238478511571884, + 0.0039801518432796, + 0.003691649530082941, + 0.0034413961693644524, + 0.004476187285035849, + 0.004980606958270073, + 0.008172987960278988, + 0.0026857564225792885, + 0.002565306844189763, + 0.004781005438417196, + 0.0025361678563058376, + 0.0045672329142689705, + 0.009687412530183792, + 0.005819819401949644, + 0.005717945750802755, + 0.007537766359746456, + 0.004141306970268488, + 0.008031195029616356, + 0.010122032836079597, + 0.014143332839012146, + 0.016828330233693123, + 0.01161081250756979, + 0.02207871712744236, + 0.013648920692503452, + 0.008485173806548119, + 0.0318143367767334, + 0.015224935486912727, + 0.006471320055425167, + 0.008647620677947998, + 0.005021829158067703, + 0.01152542419731617, + 0.005259125493466854, + 0.004845520947128534, + 0.01873389631509781, + 0.00807709526270628, + 0.0014402575325220823, + 0.001638547983020544, + 0.0014329442055895925, + 0.002184579148888588, + 0.0018276836490258574, + 0.001193763455376029, + 0.002491055056452751, + 0.0025062605272978544, + 0.0014404072426259518, + 0.0016894923755899072, + 0.0014682885957881808, + 0.0015974949346855283, + 0.0018841982819139957, + 0.0011669526575133204, + 0.002236988628283143, + 0.0016960117500275373, + 0.001165179768577218, + 0.0014438243815675378, + 0.00136026821564883, + 0.0019979120697826147, + 0.0012956279097124934, + 0.0010238953400403261, + 0.002577368402853608, + 0.001009822590276599 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 5, + 6 + ], + "selected_reward": -0.033723801374435425 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a massive crimson combine harvester?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the harvester moving across a vast field of tall, golden wheat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the harvester's metal reel rotating to gather wheat stalks onto a cutting bar?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the scene feature a low-angle camera shot emphasizing the machine's scale?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the environment bathed in warm, diffuse afternoon sunlight with a thin atmospheric haze?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.007626228500157595 + ], + "opposite_logprob_tokens": [ + -4.882626056671143 + ], + "logprob_reward": -0.007626228500157595, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.007626228500157595, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.21877850592136383, + 0.2599138915538788, + 0.16599270701408386, + 0.06440793722867966, + 0.06581293046474457, + 0.07582630217075348, + 0.034506265074014664, + 0.05143006518483162, + 0.040038254112005234, + 0.02329312637448311 + ], + "credit_map": [ + 0.019458025693893433, + 0.027956437319517136, + 0.019093802198767662, + 0.024947483092546463, + 0.06435190886259079, + 0.021315883845090866, + 0.026490388438105583, + 0.015164574608206749, + 0.02811288647353649, + 0.039487581700086594, + 0.03275872394442558, + 0.028395969420671463, + 0.02745506912469864, + 0.032659437507390976, + 0.027679042890667915, + 0.04336518421769142, + 0.021003160625696182, + 0.02243799902498722, + 0.01910484954714775, + 0.023761972784996033, + 0.021572059020400047, + 0.012150559574365616, + 0.022285105660557747, + 0.02367699146270752, + 0.0044841468334198, + 0.008817912079393864, + 0.008032457903027534, + 0.01755256950855255, + 0.006367473863065243, + 0.004538693465292454, + 0.010842449963092804, + 0.0037722301203757524, + 0.005827291402965784, + 0.010027092881500721, + 0.005629370454698801, + 0.006019290536642075, + 0.00800326094031334, + 0.004591325763612986, + 0.01569136045873165, + 0.010023941285908222, + 0.009277170524001122, + 0.011679275892674923, + 0.00633241469040513, + 0.008946520276367664, + 0.00871242769062519, + 0.0050325337797403336, + 0.016861123964190483, + 0.008984841406345367, + 0.004034499172121286, + 0.006236208137124777, + 0.0037388585042208433, + 0.004428211133927107, + 0.0032448098063468933, + 0.0026174478698521852, + 0.004675562959164381, + 0.005530667025595903, + 0.0048082321882247925, + 0.005707967560738325, + 0.004773706663399935, + 0.0072714900597929955, + 0.006194698624312878, + 0.00456186942756176, + 0.010306006297469139, + 0.007806096691638231, + 0.005848064552992582, + 0.005662105046212673, + 0.005966417491436005, + 0.004724628292024136, + 0.005949134472757578, + 0.003405016614124179, + 0.005159941501915455, + 0.0033229440450668335, + 0.002745432313531637, + 0.0037555070593953133, + 0.0025188676081597805, + 0.003066102508455515, + 0.003209976013749838, + 0.0024132593534886837, + 0.0025271992199122906, + 0.0030567815992981195 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 2 + ], + "selected_reward": -0.007626228500157595 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a massive crimson combine harvester?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the harvester moving across a vast field of tall, golden wheat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the harvester's metal reel rotating to gather wheat stalks onto a cutting bar?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the scene feature a low-angle camera shot emphasizing the machine's scale?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the environment bathed in warm, diffuse afternoon sunlight with a thin atmospheric haze?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0031909046228975058 + ], + "opposite_logprob_tokens": [ + -5.753190994262695 + ], + "logprob_reward": -0.0031909046228975058, + "yes_no_margin_reward": 5.75, + "token_credit_reward": -0.0031909046228975058, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.4249657988548279, + 0.16190965473651886, + 0.09968746453523636, + 0.044642556458711624, + 0.06656903773546219, + 0.06487243622541428, + 0.028521064668893814, + 0.028566787019371986, + 0.05744638666510582, + 0.022818775847554207 + ], + "credit_map": [ + 0.0652705654501915, + 0.04838571324944496, + 0.032974567264318466, + 0.01843954809010029, + 0.10184964537620544, + 0.09141284972429276, + 0.02619284763932228, + 0.04044011980295181, + 0.012742232531309128, + 0.024551786482334137, + 0.015997106209397316, + 0.01705009862780571, + 0.02220236137509346, + 0.014110133983194828, + 0.01665043830871582, + 0.03860550373792648, + 0.010988601483404636, + 0.019920943304896355, + 0.008159312419593334, + 0.013549498282372952, + 0.01573285274207592, + 0.006686347536742687, + 0.012922869063913822, + 0.011727049946784973, + 0.004386563785374165, + 0.007734078913927078, + 0.003331627231091261, + 0.007007640320807695, + 0.006767009850591421, + 0.003185052191838622, + 0.006218656897544861, + 0.0060119302943348885, + 0.005687168333679438, + 0.008673247881233692, + 0.00549683254212141, + 0.005722083151340485, + 0.007741742767393589, + 0.004723499994724989, + 0.01653957925736904, + 0.011984892189502716, + 0.006653838790953159, + 0.00822316762059927, + 0.0066161672584712505, + 0.01243795920163393, + 0.007212316617369652, + 0.004134376533329487, + 0.01274328026920557, + 0.006851336918771267, + 0.0025148498825728893, + 0.003924845717847347, + 0.003005108330398798, + 0.0040886178612709045, + 0.003036334179341793, + 0.002080504084005952, + 0.006070151459425688, + 0.0038006561808288097, + 0.0032056616619229317, + 0.0034287467133253813, + 0.002985141472890973, + 0.004269776400178671, + 0.0035333875566720963, + 0.002491577295586467, + 0.004337409045547247, + 0.004315088968724012, + 0.008889056742191315, + 0.007460505701601505, + 0.006165644619613886, + 0.007934636436402798, + 0.005770297721028328, + 0.004967243876308203, + 0.009334777481853962, + 0.0069242254830896854, + 0.003718324238434434, + 0.00358239421620965, + 0.0024255539756268263, + 0.0026501815300434828, + 0.002808129182085395, + 0.0020976620726287365, + 0.003387185512110591, + 0.002149348147213459 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 2 + ], + "selected_reward": -0.0031909046228975058 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "low-angle tracking shot moves smoothly alongside a massive crimson combine harvester as it churns steadily and relentlessly across a vast field. The camera's position emphasizes the immense scale of the monolithic steel machine, its body coated in a fine dust that dulls its surface in the hazy sunlight. Its massive black rubber tires press deep tracks into the parched soil. In front, a wide, intricate harvester rotates methodically, its metal reel gathering tall, golden wheat stalks onto the precisely moving cutting bar below. The scene is bathed in the warm, diffuse light of the afternoon sun, filtered through a thin atmospheric haze, softening the horizon where ripe wheat fields stretch like a golden ocean. As the machine advances, its wake leaves a neat trail of stubble, while a stream of chopped straw and chaff erupts from the chute behind, catching the amber glow and shimmering in the air. The overall color palette is a rich tapestry of gold, amber, and the crimson of the machinery, creating a powerful yet tranquil atmosphere of agricultural abundance. The entire scene is captured with a subtle, nostalgic film texture that adds a timeless and painterly quality to the powerful, rhythmic movements of the harvest.", + "seed": 9807, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000071/group_11/index_7867_seed_9807_dp_22_rank_44_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 11, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.021674007875844836, + "vlm_logprob": -0.021674007875844836, + "vlm_yes_no_margin": 4.3, + "vlm_token_credit": -0.021674007875844836, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a massive crimson combine harvester?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the harvester moving across a vast field of tall, golden wheat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the harvester's metal reel rotating to gather wheat stalks onto a cutting bar?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the scene feature a low-angle camera shot emphasizing the machine's scale?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the environment bathed in warm, diffuse afternoon sunlight with a thin atmospheric haze?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0641951933503151 + ], + "opposite_logprob_tokens": [ + -2.814195156097412 + ], + "logprob_reward": -0.0641951933503151, + "yes_no_margin_reward": 2.75, + "token_credit_reward": -0.0641951933503151, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14234383404254913, + 0.06561353802680969, + 0.06473708152770996, + 0.09538615494966507, + 0.12570811808109283, + 0.0988372340798378, + 0.08397993445396423, + 0.17735421657562256, + 0.10883360356092453, + 0.037206266075372696 + ], + "credit_map": [ + 0.011675521731376648, + 0.02031658962368965, + 0.012121975421905518, + 0.022735681384801865, + 0.01930275931954384, + 0.02036747708916664, + 0.018671268597245216, + 0.0171525739133358, + 0.006555711850523949, + 0.007251559291034937, + 0.005424074828624725, + 0.011367924511432648, + 0.010832635685801506, + 0.007516599725931883, + 0.009101858362555504, + 0.007563176564872265, + 0.007655618246644735, + 0.006271969992667437, + 0.005701087880879641, + 0.006068442948162556, + 0.009710731916129589, + 0.005112334620207548, + 0.004885789938271046, + 0.019331105053424835, + 0.01351466216146946, + 0.009236600250005722, + 0.005149160046130419, + 0.012996269389986992, + 0.0334581583738327, + 0.006466202903538942, + 0.007366092409938574, + 0.00719900568947196, + 0.016779284924268723, + 0.021191926673054695, + 0.013371257111430168, + 0.013693388551473618, + 0.01837971620261669, + 0.018437761813402176, + 0.013077029958367348, + 0.010777740739285946, + 0.008009920828044415, + 0.015140433795750141, + 0.006943169049918652, + 0.006725518964231014, + 0.021205419674515724, + 0.009409786202013493, + 0.011884335428476334, + 0.019518649205565453, + 0.009345347993075848, + 0.010273195803165436, + 0.01074532326310873, + 0.005859580356627703, + 0.011389946565032005, + 0.008146233856678009, + 0.016451090574264526, + 0.011769218370318413, + 0.015854643657803535, + 0.023552773520350456, + 0.023826219141483307, + 0.021830111742019653, + 0.03235827386379242, + 0.019446318969130516, + 0.021247075870633125, + 0.019238805398344994, + 0.010290386155247688, + 0.013658380135893822, + 0.009547604247927666, + 0.007842382416129112, + 0.017341259866952896, + 0.014530197717249393, + 0.010912385769188404, + 0.02471100725233555, + 0.005670777056366205, + 0.006193073932081461, + 0.00328178983181715, + 0.0044133421033620834, + 0.004308945965021849, + 0.0028425483033061028, + 0.004530981183052063, + 0.005964805372059345 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 0, + 4 + ], + "selected_reward": -0.0641951933503151 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a massive crimson combine harvester?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the harvester moving across a vast field of tall, golden wheat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the harvester's metal reel rotating to gather wheat stalks onto a cutting bar?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the scene feature a low-angle camera shot emphasizing the machine's scale?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the environment bathed in warm, diffuse afternoon sunlight with a thin atmospheric haze?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.011092307977378368 + ], + "opposite_logprob_tokens": [ + -4.511092185974121 + ], + "logprob_reward": -0.011092307977378368, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.011092307977378368, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.04987187311053276, + 0.043687235563993454, + 0.027199234813451767, + 0.11774439364671707, + 0.03980487585067749, + 0.02432137355208397, + 0.07026909291744232, + 0.0319802351295948, + 0.5700162649154663, + 0.02510540373623371 + ], + "credit_map": [ + 0.007984294556081295, + 0.004188537131994963, + 0.003973905462771654, + 0.005824638064950705, + 0.01365412212908268, + 0.0046948958188295364, + 0.0058539146557450294, + 0.003697568317875266, + 0.004120613448321819, + 0.005225121974945068, + 0.004392917267978191, + 0.007261662278324366, + 0.005836485419422388, + 0.006802536081522703, + 0.0045515610836446285, + 0.005496337544173002, + 0.004051586147397757, + 0.003541958285495639, + 0.0024468458723276854, + 0.0025654116179794073, + 0.003952996805310249, + 0.0023728162050247192, + 0.0018970802193507552, + 0.006370541173964739, + 0.019502539187669754, + 0.009747917763888836, + 0.006593887694180012, + 0.01592447981238365, + 0.03931045159697533, + 0.008051225915551186, + 0.009411699138581753, + 0.009202207438647747, + 0.004059113562107086, + 0.005941718351095915, + 0.004901289939880371, + 0.004929033108055592, + 0.006019136868417263, + 0.004125008825212717, + 0.007052181288599968, + 0.0027774013578891754, + 0.004143169615417719, + 0.003610580926761031, + 0.0030284051317721605, + 0.0016071094432845712, + 0.004025829955935478, + 0.0036027547903358936, + 0.0019644659478217363, + 0.0023390573915094137, + 0.007239066995680332, + 0.007610658183693886, + 0.007915639318525791, + 0.005571142304688692, + 0.011498101986944675, + 0.008820120245218277, + 0.012702862732112408, + 0.00891150627285242, + 0.0021121001336723566, + 0.003000677330419421, + 0.003128848737105727, + 0.0026094818022102118, + 0.0038717102725058794, + 0.002166910795494914, + 0.007320166565477848, + 0.007770343217998743, + 0.04648426175117493, + 0.05873371288180351, + 0.04749877750873566, + 0.039326369762420654, + 0.06419222801923752, + 0.09013783931732178, + 0.06397324800491333, + 0.15966984629631042, + 0.0031731734052300453, + 0.0042899735271930695, + 0.0031969030387699604, + 0.003299225587397814, + 0.00247632316313684, + 0.0020662681199610233, + 0.0034170548897236586, + 0.0031864847987890244 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 3, + 6 + ], + "selected_reward": -0.011092307977378368 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a massive crimson combine harvester?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the harvester moving across a vast field of tall, golden wheat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the harvester's metal reel rotating to gather wheat stalks onto a cutting bar?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the scene feature a low-angle camera shot emphasizing the machine's scale?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the environment bathed in warm, diffuse afternoon sunlight with a thin atmospheric haze?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.014207405038177967 + ], + "opposite_logprob_tokens": [ + -4.264207363128662 + ], + "logprob_reward": -0.014207405038177967, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.014207405038177967, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.05001652240753174, + 0.061267465353012085, + 0.044240258634090424, + 0.06550894677639008, + 0.062467362731695175, + 0.04340377077460289, + 0.08456910401582718, + 0.032620664685964584, + 0.5243451595306396, + 0.03156079351902008 + ], + "credit_map": [ + 0.006567093078047037, + 0.005185967311263084, + 0.005213517230004072, + 0.0067824628204107285, + 0.009210660122334957, + 0.005089552141726017, + 0.005313619505614042, + 0.006653642747551203, + 0.005981281399726868, + 0.006243569310754538, + 0.0062933932058513165, + 0.011438047513365746, + 0.008574607782065868, + 0.006123075261712074, + 0.008799237199127674, + 0.007814249955117702, + 0.0069063263945281506, + 0.0041518728248775005, + 0.003701738314703107, + 0.005828360561281443, + 0.006549046840518713, + 0.0042977361008524895, + 0.003921082708984613, + 0.008884090930223465, + 0.00625110138207674, + 0.006344866938889027, + 0.004226149525493383, + 0.007772441953420639, + 0.02240968868136406, + 0.007759795058518648, + 0.004370365757495165, + 0.006374524440616369, + 0.008577907457947731, + 0.009414873085916042, + 0.00669960118830204, + 0.008355604484677315, + 0.009300131350755692, + 0.006924779620021582, + 0.008303822949528694, + 0.004890630021691322, + 0.0036110999062657356, + 0.005874919705092907, + 0.004678505938500166, + 0.0034634938929229975, + 0.00892582256346941, + 0.004768994636833668, + 0.0037813526578247547, + 0.008299576118588448, + 0.00793454609811306, + 0.010178581811487675, + 0.006314077880233526, + 0.00924411229789257, + 0.013361074961721897, + 0.00976806040853262, + 0.017437029629945755, + 0.01033160462975502, + 0.002597479149699211, + 0.0029632507357746363, + 0.0033747318666428328, + 0.004113759845495224, + 0.004663438536226749, + 0.0018530504312366247, + 0.00656533520668745, + 0.006489617750048637, + 0.05181222781538963, + 0.061281848698854446, + 0.04769953712821007, + 0.040912870317697525, + 0.06830441206693649, + 0.0648215264081955, + 0.05515587329864502, + 0.1343568116426468, + 0.004643305670469999, + 0.003874150337651372, + 0.0030342615209519863, + 0.0027672515716403723, + 0.003656154265627265, + 0.002358333906158805, + 0.004870573058724403, + 0.006356762256473303 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 6, + 3 + ], + "selected_reward": -0.014207405038177967 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a massive crimson combine harvester?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the harvester moving across a vast field of tall, golden wheat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the harvester's metal reel rotating to gather wheat stalks onto a cutting bar?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the scene feature a low-angle camera shot emphasizing the machine's scale?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the environment bathed in warm, diffuse afternoon sunlight with a thin atmospheric haze?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.016060156747698784 + ], + "opposite_logprob_tokens": [ + -4.1410603523254395 + ], + "logprob_reward": -0.016060156747698784, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.016060156747698784, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07586172223091125, + 0.056810904294252396, + 0.05028577148914337, + 0.2605118155479431, + 0.11849714815616608, + 0.15448173880577087, + 0.052344806492328644, + 0.035419438034296036, + 0.15918263792991638, + 0.036604028195142746 + ], + "credit_map": [ + 0.007724138908088207, + 0.00788975041359663, + 0.0077783288434147835, + 0.009850661270320415, + 0.011588895693421364, + 0.009966455399990082, + 0.009505070745944977, + 0.01155842561274767, + 0.005451801232993603, + 0.008337881416082382, + 0.00671015540137887, + 0.007308542728424072, + 0.008381016552448273, + 0.006644978653639555, + 0.0064500425942242146, + 0.007526485715061426, + 0.005679309833794832, + 0.0039473166689276695, + 0.006387609988451004, + 0.005131684709340334, + 0.005616546142846346, + 0.0046581970527768135, + 0.005242629442363977, + 0.013622477650642395, + 0.02799520082771778, + 0.0272693932056427, + 0.011812455952167511, + 0.03508121147751808, + 0.10860070586204529, + 0.01555962860584259, + 0.017499566078186035, + 0.016693634912371635, + 0.020533880218863487, + 0.014325413852930069, + 0.017186511307954788, + 0.01391716580837965, + 0.0167388953268528, + 0.012221242301166058, + 0.012178352102637291, + 0.011395681649446487, + 0.008668634109199047, + 0.0224111657589674, + 0.016269635409116745, + 0.013335713185369968, + 0.033330343663692474, + 0.01258520781993866, + 0.02157554030418396, + 0.02630549483001232, + 0.004907091148197651, + 0.005306035280227661, + 0.009945139288902283, + 0.0075753522105515, + 0.005681196693331003, + 0.00491364998742938, + 0.008475578390061855, + 0.0055407630279660225, + 0.004161369055509567, + 0.0033929168712347746, + 0.004943174310028553, + 0.004572548903524876, + 0.004845515824854374, + 0.003305374877527356, + 0.005515105556696653, + 0.004683431703597307, + 0.015934282913804054, + 0.01751548796892166, + 0.017753688618540764, + 0.012167559936642647, + 0.020910581573843956, + 0.02187422849237919, + 0.014466820284724236, + 0.03855997696518898, + 0.005242571234703064, + 0.0035029167775064707, + 0.004665391519665718, + 0.0038971391040831804, + 0.004119168967008591, + 0.0035129825118929148, + 0.005419781897217035, + 0.006244073621928692 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 8, + 5 + ], + "selected_reward": -0.016060156747698784 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a massive crimson combine harvester?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the harvester moving across a vast field of tall, golden wheat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the harvester's metal reel rotating to gather wheat stalks onto a cutting bar?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the scene feature a low-angle camera shot emphasizing the machine's scale?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the environment bathed in warm, diffuse afternoon sunlight with a thin atmospheric haze?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.002814976265653968 + ], + "opposite_logprob_tokens": [ + -5.877814769744873 + ], + "logprob_reward": -0.002814976265653968, + "yes_no_margin_reward": 5.875, + "token_credit_reward": -0.002814976265653968, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12043465673923492, + 0.1454041749238968, + 0.025751030072569847, + 0.09631533175706863, + 0.08238513022661209, + 0.10523612797260284, + 0.051605936139822006, + 0.09879014641046524, + 0.2298911064863205, + 0.04418637603521347 + ], + "credit_map": [ + 0.01349552907049656, + 0.01584024727344513, + 0.011075577698647976, + 0.016345767304301262, + 0.02319076471030712, + 0.016490716487169266, + 0.013354482129216194, + 0.010641585104167461, + 0.015304156579077244, + 0.017789483070373535, + 0.01319645345211029, + 0.027820419520139694, + 0.02039538323879242, + 0.014639824628829956, + 0.02293424680829048, + 0.01332420576363802, + 0.004035858903080225, + 0.004338452592492104, + 0.0021194240543991327, + 0.0026336279697716236, + 0.0035215013194829226, + 0.00291860057041049, + 0.002391090150922537, + 0.0037924747448414564, + 0.01630476489663124, + 0.008955574594438076, + 0.005418715067207813, + 0.012765803374350071, + 0.03090130165219307, + 0.00632515549659729, + 0.008221734315156937, + 0.007422278635203838, + 0.011511505581438541, + 0.012308705598115921, + 0.011017457582056522, + 0.01036926731467247, + 0.00927917193621397, + 0.007404808420687914, + 0.009387804195284843, + 0.01110641099512577, + 0.00923017505556345, + 0.018034664914011955, + 0.00948121678084135, + 0.008297913707792759, + 0.021976875141263008, + 0.009144613519310951, + 0.011803634464740753, + 0.01726704277098179, + 0.007154886610805988, + 0.0061497376300394535, + 0.009464995935559273, + 0.004177918191999197, + 0.006032124161720276, + 0.003588360734283924, + 0.008581562899053097, + 0.006456349976360798, + 0.008937345817685127, + 0.013774064369499683, + 0.013243322260677814, + 0.013490540906786919, + 0.013830971904098988, + 0.008515615016222, + 0.01335038710385561, + 0.013647900894284248, + 0.024131489917635918, + 0.02788405306637287, + 0.0212715957313776, + 0.018389929085969925, + 0.031819649040699005, + 0.02834172733128071, + 0.023233624175190926, + 0.05481903254985809, + 0.007498936261981726, + 0.006845154799520969, + 0.005216510035097599, + 0.0064305891282856464, + 0.004503004252910614, + 0.003141381312161684, + 0.004679476842284203, + 0.005871322937309742 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 1, + 0 + ], + "selected_reward": -0.002814976265653968 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "low-angle tracking shot moves smoothly alongside a massive crimson combine harvester as it churns steadily and relentlessly across a vast field. The camera's position emphasizes the immense scale of the monolithic steel machine, its body coated in a fine dust that dulls its surface in the hazy sunlight. Its massive black rubber tires press deep tracks into the parched soil. In front, a wide, intricate harvester rotates methodically, its metal reel gathering tall, golden wheat stalks onto the precisely moving cutting bar below. The scene is bathed in the warm, diffuse light of the afternoon sun, filtered through a thin atmospheric haze, softening the horizon where ripe wheat fields stretch like a golden ocean. As the machine advances, its wake leaves a neat trail of stubble, while a stream of chopped straw and chaff erupts from the chute behind, catching the amber glow and shimmering in the air. The overall color palette is a rich tapestry of gold, amber, and the crimson of the machinery, creating a powerful yet tranquil atmosphere of agricultural abundance. The entire scene is captured with a subtle, nostalgic film texture that adds a timeless and painterly quality to the powerful, rhythmic movements of the harvest.", + "seed": 9808, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000071/group_11/index_7867_seed_9808_dp_23_rank_46_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 11, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.03155030887573958, + "vlm_logprob": -0.03155030887573958, + "vlm_yes_no_margin": 4.025, + "vlm_token_credit": -0.03155030887573958, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a massive crimson combine harvester?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the harvester moving across a vast field of tall, golden wheat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the harvester's metal reel rotating to gather wheat stalks onto a cutting bar?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the scene feature a low-angle camera shot emphasizing the machine's scale?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the environment bathed in warm, diffuse afternoon sunlight with a thin atmospheric haze?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.10243277996778488 + ], + "opposite_logprob_tokens": [ + -2.3524327278137207 + ], + "logprob_reward": -0.10243277996778488, + "yes_no_margin_reward": 2.25, + "token_credit_reward": -0.10243277996778488, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12400826066732407, + 0.19745995104312897, + 0.11856429278850555, + 0.28963997960090637, + 0.03787514939904213, + 0.09607712179422379, + 0.03893083706498146, + 0.026810482144355774, + 0.01885688304901123, + 0.051777005195617676 + ], + "credit_map": [ + 0.011926036328077316, + 0.009700567461550236, + 0.018099742010235786, + 0.010149208828806877, + 0.020101288333535194, + 0.029689421877264977, + 0.01129924412816763, + 0.013042774051427841, + 0.013626079075038433, + 0.023977195844054222, + 0.023324396461248398, + 0.016481975093483925, + 0.029789678752422333, + 0.05253613740205765, + 0.018124263733625412, + 0.019600238651037216, + 0.015135071240365505, + 0.01079714298248291, + 0.01555543951690197, + 0.01157514750957489, + 0.01812969148159027, + 0.02809893898665905, + 0.010436574928462505, + 0.008836310356855392, + 0.07238265872001648, + 0.09721800684928894, + 0.0075142765417695045, + 0.010353239253163338, + 0.016826501116156578, + 0.02689657174050808, + 0.011708579026162624, + 0.04674019664525986, + 0.004216267727315426, + 0.0025550569407641888, + 0.0034669023007154465, + 0.0028965454548597336, + 0.005548601504415274, + 0.009946243837475777, + 0.003590361448004842, + 0.005655176471918821, + 0.00786937028169632, + 0.00758693553507328, + 0.01220861729234457, + 0.008471932262182236, + 0.009794837795197964, + 0.01618790812790394, + 0.010728335939347744, + 0.023229196667671204, + 0.004965631291270256, + 0.004880124237388372, + 0.005054405890405178, + 0.004350146744400263, + 0.0032246923074126244, + 0.007651975378394127, + 0.0043620578944683075, + 0.004441808443516493, + 0.005888459272682667, + 0.002268789568915963, + 0.0028069443069398403, + 0.0021049347706139088, + 0.001428186777047813, + 0.007674668915569782, + 0.0019162810640409589, + 0.0027222218923270702, + 0.0019663285929709673, + 0.0016311078798025846, + 0.0022417206782847643, + 0.002717524766921997, + 0.0018478917190805078, + 0.0030621185433119535, + 0.0031994087621569633, + 0.00219078385271132, + 0.005238473415374756, + 0.004580351058393717, + 0.006176651921123266, + 0.0069445399567484856, + 0.0037169004790484905, + 0.009090536274015903, + 0.008923838846385479, + 0.007105722092092037 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 1, + 0 + ], + "selected_reward": -0.10243277996778488 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a massive crimson combine harvester?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the harvester moving across a vast field of tall, golden wheat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the harvester's metal reel rotating to gather wheat stalks onto a cutting bar?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the scene feature a low-angle camera shot emphasizing the machine's scale?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the environment bathed in warm, diffuse afternoon sunlight with a thin atmospheric haze?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.023310085758566856 + ], + "opposite_logprob_tokens": [ + -3.7733101844787598 + ], + "logprob_reward": -0.023310085758566856, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.023310085758566856, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07897796481847763, + 0.06772240996360779, + 0.08168244361877441, + 0.404127299785614, + 0.05116691067814827, + 0.1255841702222824, + 0.06063733622431755, + 0.09564291685819626, + 0.012616968713700771, + 0.02184160239994526 + ], + "credit_map": [ + 0.007259846664965153, + 0.0061610909178853035, + 0.014090432785451412, + 0.010754101909697056, + 0.010565461590886116, + 0.010275428183376789, + 0.009344812482595444, + 0.010526794008910656, + 0.004810976795852184, + 0.006437638308852911, + 0.01382655743509531, + 0.010237671434879303, + 0.00870512891560793, + 0.01021884847432375, + 0.007627430371940136, + 0.005858154967427254, + 0.006266337353736162, + 0.006000495050102472, + 0.018564419820904732, + 0.010742847807705402, + 0.010960840620100498, + 0.01287635788321495, + 0.007456766441464424, + 0.008814381435513496, + 0.0981987863779068, + 0.13192784786224365, + 0.014896403066813946, + 0.016886111348867416, + 0.02486451342701912, + 0.036642421036958694, + 0.017558138817548752, + 0.06315304338932037, + 0.007496926933526993, + 0.004849415272474289, + 0.004951172973960638, + 0.004100285470485687, + 0.0032149136532098055, + 0.01613633707165718, + 0.003555967938154936, + 0.006861889734864235, + 0.009980633854866028, + 0.009829072281718254, + 0.021250875666737556, + 0.013840142637491226, + 0.009965403005480766, + 0.019598564133048058, + 0.014867930673062801, + 0.0262515377253294, + 0.007442456670105457, + 0.008758158423006535, + 0.008479527197778225, + 0.008082752116024494, + 0.004352692048996687, + 0.009887780994176865, + 0.006795787252485752, + 0.006838179659098387, + 0.022435294464230537, + 0.008884665556252003, + 0.006159968674182892, + 0.005028588697314262, + 0.0038078920915722847, + 0.030854105949401855, + 0.0073610288091003895, + 0.011111377738416195, + 0.0014368247939273715, + 0.0010275725508108735, + 0.0021733168978244066, + 0.002685906132683158, + 0.0010376771679148078, + 0.0016979334177449346, + 0.0010850016260519624, + 0.001472736126743257, + 0.0024948061909526587, + 0.002434652764350176, + 0.002866768976673484, + 0.003342597745358944, + 0.0024718407075852156, + 0.002866287948563695, + 0.003163878573104739, + 0.0022007692605257034 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 5, + 7 + ], + "selected_reward": -0.023310085758566856 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a massive crimson combine harvester?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the harvester moving across a vast field of tall, golden wheat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the harvester's metal reel rotating to gather wheat stalks onto a cutting bar?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the scene feature a low-angle camera shot emphasizing the machine's scale?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the environment bathed in warm, diffuse afternoon sunlight with a thin atmospheric haze?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.014204349368810654 + ], + "opposite_logprob_tokens": [ + -4.264204502105713 + ], + "logprob_reward": -0.014204349368810654, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.014204349368810654, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08308973163366318, + 0.14480571448802948, + 0.4395078122615814, + 0.1139008179306984, + 0.02492441236972809, + 0.07715633511543274, + 0.06848635524511337, + 0.011499524116516113, + 0.014683982357382774, + 0.021945364773273468 + ], + "credit_map": [ + 0.009447700344026089, + 0.007639714516699314, + 0.01658243127167225, + 0.013155405409634113, + 0.010554948821663857, + 0.010570266284048557, + 0.008587323129177094, + 0.006551939528435469, + 0.013436345383524895, + 0.01397975068539381, + 0.024636805057525635, + 0.014235582202672958, + 0.017585746943950653, + 0.02110026590526104, + 0.01336030662059784, + 0.02647089771926403, + 0.041963834315538406, + 0.038146790117025375, + 0.08137985318899155, + 0.05738084763288498, + 0.06347651779651642, + 0.07776544243097305, + 0.049903180450201035, + 0.0294913612306118, + 0.010773461312055588, + 0.011600826866924763, + 0.019948823377490044, + 0.01571841351687908, + 0.016072221100330353, + 0.017677204683423042, + 0.011673614382743835, + 0.010436257347464561, + 0.002211417304351926, + 0.0013273477088660002, + 0.004176636692136526, + 0.002827777760103345, + 0.0026631979271769524, + 0.00396312540397048, + 0.0022174480836838484, + 0.005537461955100298, + 0.006782689597457647, + 0.006183161400258541, + 0.011423245072364807, + 0.009519661776721478, + 0.006269753910601139, + 0.012571503408253193, + 0.010351678356528282, + 0.014054637402296066, + 0.009936668910086155, + 0.007666043005883694, + 0.007854497991502285, + 0.006351068615913391, + 0.005135539919137955, + 0.016630899161100388, + 0.006615267600864172, + 0.008296366780996323, + 0.001555939670652151, + 0.0009605237282812595, + 0.0019539918284863234, + 0.0014733248390257359, + 0.0009926880011335015, + 0.0017406478291377425, + 0.001221314538270235, + 0.0016010936815291643, + 0.0012991956900805235, + 0.0012106772046536207, + 0.001826726016588509, + 0.0022736084647476673, + 0.0022965066600590944, + 0.0026406976394355297, + 0.0013930581044405699, + 0.0017435118788853288, + 0.002531471662223339, + 0.0019203046103939414, + 0.0036791383754462004, + 0.0024687673430889845, + 0.002384874038398266, + 0.0037586481776088476, + 0.002697089919820428, + 0.0025050686672329903 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 1, + 3 + ], + "selected_reward": -0.014204349368810654 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a massive crimson combine harvester?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the harvester moving across a vast field of tall, golden wheat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the harvester's metal reel rotating to gather wheat stalks onto a cutting bar?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the scene feature a low-angle camera shot emphasizing the machine's scale?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the environment bathed in warm, diffuse afternoon sunlight with a thin atmospheric haze?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.014190481044352055 + ], + "opposite_logprob_tokens": [ + -4.264190673828125 + ], + "logprob_reward": -0.014190481044352055, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.014190481044352055, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09971155226230621, + 0.08912837505340576, + 0.22972550988197327, + 0.29612162709236145, + 0.12648148834705353, + 0.04469962790608406, + 0.034218430519104004, + 0.04242036119103432, + 0.01722905784845352, + 0.020263949409127235 + ], + "credit_map": [ + 0.008733287453651428, + 0.010237174108624458, + 0.013736733235418797, + 0.009547783061861992, + 0.012984665110707283, + 0.01726093515753746, + 0.011222432367503643, + 0.015988541767001152, + 0.007311588153243065, + 0.009570661000907421, + 0.009480210952460766, + 0.007994513027369976, + 0.012763396836817265, + 0.0229048989713192, + 0.008061019703745842, + 0.011042087338864803, + 0.024186676368117332, + 0.01904582418501377, + 0.03928109630942345, + 0.028834152966737747, + 0.030548930168151855, + 0.04614993929862976, + 0.02516036480665207, + 0.016518525779247284, + 0.0671989768743515, + 0.09030498564243317, + 0.011821301653981209, + 0.012661325745284557, + 0.01744132488965988, + 0.031005043536424637, + 0.012880617752671242, + 0.052808038890361786, + 0.01686404086649418, + 0.009042317047715187, + 0.010430239140987396, + 0.008527910336852074, + 0.011760391294956207, + 0.0281364805996418, + 0.010259756818413734, + 0.03146035224199295, + 0.004419832024723291, + 0.005376051180064678, + 0.005387816112488508, + 0.00653810752555728, + 0.007642917800694704, + 0.006299323868006468, + 0.0048304274678230286, + 0.00420515239238739, + 0.003616098314523697, + 0.004488558042794466, + 0.005922267679125071, + 0.0056653437204658985, + 0.0039329687133431435, + 0.0038557203952223063, + 0.003530168207362294, + 0.003207304747775197, + 0.0075950357131659985, + 0.004116903990507126, + 0.00529381912201643, + 0.0061221010982990265, + 0.003449982265010476, + 0.008430292829871178, + 0.00314020155929029, + 0.004272024147212505, + 0.001809613429941237, + 0.00146666937507689, + 0.0021703264210373163, + 0.002649247646331787, + 0.0017177780391648412, + 0.0037592893932014704, + 0.0016908813267946243, + 0.0019652522169053555, + 0.001990323653444648, + 0.0019897734746336937, + 0.0024902764707803726, + 0.003276451490819454, + 0.0013742886949330568, + 0.003090046579018235, + 0.0034665714483708143, + 0.0025862178299576044 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 2, + 4 + ], + "selected_reward": -0.014190481044352055 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a massive crimson combine harvester?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the harvester moving across a vast field of tall, golden wheat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the harvester's metal reel rotating to gather wheat stalks onto a cutting bar?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the scene feature a low-angle camera shot emphasizing the machine's scale?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the environment bathed in warm, diffuse afternoon sunlight with a thin atmospheric haze?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.003613848239183426 + ], + "opposite_logprob_tokens": [ + -5.6286139488220215 + ], + "logprob_reward": -0.003613848239183426, + "yes_no_margin_reward": 5.625, + "token_credit_reward": -0.003613848239183426, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08031298220157623, + 0.06622330844402313, + 0.10089584439992905, + 0.4881463050842285, + 0.07308586686849594, + 0.02207641862332821, + 0.03231057897210121, + 0.09524763375520706, + 0.020768295973539352, + 0.02093283087015152 + ], + "credit_map": [ + 0.006682788021862507, + 0.01037551462650299, + 0.008398872800171375, + 0.012372853234410286, + 0.011253148317337036, + 0.015706362202763557, + 0.008741212077438831, + 0.006782226264476776, + 0.0055991895496845245, + 0.004403312224894762, + 0.006955714430660009, + 0.003236276563256979, + 0.006675015203654766, + 0.016595877707004547, + 0.00382618373259902, + 0.018931742757558823, + 0.012678096070885658, + 0.008907717652618885, + 0.015172474086284637, + 0.010086636058986187, + 0.01567055471241474, + 0.02314009517431259, + 0.008709513582289219, + 0.00653074262663722, + 0.11764051020145416, + 0.15804779529571533, + 0.019847307354211807, + 0.022948671132326126, + 0.03192511945962906, + 0.04573836177587509, + 0.01852007955312729, + 0.07347840815782547, + 0.007601909339427948, + 0.004296630155295134, + 0.009024002589285374, + 0.0058221882209181786, + 0.006639424245804548, + 0.012847998179495335, + 0.005717715714126825, + 0.021135997027158737, + 0.002746673533692956, + 0.0031699170358479023, + 0.002693428425118327, + 0.00285880989395082, + 0.001892133615911007, + 0.003423268673941493, + 0.0025575526524335146, + 0.002734632231295109, + 0.004223675467073917, + 0.0037115258164703846, + 0.003583296900615096, + 0.0033752513118088245, + 0.0026987935416400433, + 0.00708951149135828, + 0.0038271809462457895, + 0.0038013430312275887, + 0.02171960286796093, + 0.00813203677535057, + 0.010722030885517597, + 0.008112753741443157, + 0.003773745149374008, + 0.026539646089076996, + 0.006548142991960049, + 0.009699671529233456, + 0.002247451338917017, + 0.0019736106041818857, + 0.0030990508385002613, + 0.003941703587770462, + 0.001798543962650001, + 0.003318286268040538, + 0.0019961001817137003, + 0.002393548609688878, + 0.0023442290257662535, + 0.002040871186181903, + 0.0031843779142946005, + 0.00337899848818779, + 0.0013897618046030402, + 0.002992141991853714, + 0.0029703981708735228, + 0.002632051007822156 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 2, + 7 + ], + "selected_reward": -0.003613848239183426 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "low-angle tracking shot moves smoothly alongside a massive crimson combine harvester as it churns steadily and relentlessly across a vast field. The camera's position emphasizes the immense scale of the monolithic steel machine, its body coated in a fine dust that dulls its surface in the hazy sunlight. Its massive black rubber tires press deep tracks into the parched soil. In front, a wide, intricate harvester rotates methodically, its metal reel gathering tall, golden wheat stalks onto the precisely moving cutting bar below. The scene is bathed in the warm, diffuse light of the afternoon sun, filtered through a thin atmospheric haze, softening the horizon where ripe wheat fields stretch like a golden ocean. As the machine advances, its wake leaves a neat trail of stubble, while a stream of chopped straw and chaff erupts from the chute behind, catching the amber glow and shimmering in the air. The overall color palette is a rich tapestry of gold, amber, and the crimson of the machinery, creating a powerful yet tranquil atmosphere of agricultural abundance. The entire scene is captured with a subtle, nostalgic film texture that adds a timeless and painterly quality to the powerful, rhythmic movements of the harvest.", + "seed": 9809, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000071/group_11/index_7867_seed_9809_dp_23_rank_46_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 11, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.0634718507528305, + "vlm_logprob": -0.0634718507528305, + "vlm_yes_no_margin": 3.424999976158142, + "vlm_token_credit": -0.0634718507528305, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a massive crimson combine harvester?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the harvester moving across a vast field of tall, golden wheat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the harvester's metal reel rotating to gather wheat stalks onto a cutting bar?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the scene feature a low-angle camera shot emphasizing the machine's scale?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the environment bathed in warm, diffuse afternoon sunlight with a thin atmospheric haze?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.09256277233362198 + ], + "opposite_logprob_tokens": [ + -2.467562675476074 + ], + "logprob_reward": -0.09256277233362198, + "yes_no_margin_reward": 2.375, + "token_credit_reward": -0.09256277233362198, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10950140655040741, + 0.05302473157644272, + 0.1419350802898407, + 0.10382465273141861, + 0.1206042543053627, + 0.14738522469997406, + 0.08427827060222626, + 0.044599514454603195, + 0.05710557848215103, + 0.13774128258228302 + ], + "credit_map": [ + 0.016607139259576797, + 0.01411343365907669, + 0.012984891422092915, + 0.00974220223724842, + 0.01227263268083334, + 0.00785936787724495, + 0.021536698564887047, + 0.014385041780769825, + 0.007855493575334549, + 0.010297765955328941, + 0.006430248264223337, + 0.0064595043659210205, + 0.005521436687558889, + 0.004377518780529499, + 0.005048658698797226, + 0.0070341117680072784, + 0.02075986936688423, + 0.022048966959118843, + 0.017691904678940773, + 0.030014963820576668, + 0.01331466343253851, + 0.010072645731270313, + 0.011783276684582233, + 0.016248803585767746, + 0.014227936044335365, + 0.019752785563468933, + 0.013157512061297894, + 0.02119576930999756, + 0.009201456792652607, + 0.005539206322282553, + 0.0094111617654562, + 0.011338835582137108, + 0.028034241870045662, + 0.024753006175160408, + 0.012607875280082226, + 0.014824540354311466, + 0.010544599033892155, + 0.010132353752851486, + 0.00877048633992672, + 0.010937158018350601, + 0.02404729090631008, + 0.022687865421175957, + 0.023909611627459526, + 0.01854771561920643, + 0.021512819454073906, + 0.014660452492535114, + 0.012876180000603199, + 0.009143301285803318, + 0.014091790653765202, + 0.009242664091289043, + 0.008473177440464497, + 0.01001057494431734, + 0.012378212995827198, + 0.007949494756758213, + 0.015842901542782784, + 0.0062894560396671295, + 0.00415382394567132, + 0.00970891211181879, + 0.004392007831484079, + 0.008525075390934944, + 0.005796543322503567, + 0.004179652314633131, + 0.004457238595932722, + 0.0033862642012536526, + 0.009111442603170872, + 0.009145112708210945, + 0.004024269990622997, + 0.0024439527187496424, + 0.0037331636995077133, + 0.004811024758964777, + 0.004479192662984133, + 0.01935742422938347, + 0.017292119562625885, + 0.01422474067658186, + 0.036301493644714355, + 0.013444043695926666, + 0.01278417557477951, + 0.011716371402144432, + 0.012171683833003044, + 0.01980665512382984 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 2, + 9 + ], + "selected_reward": -0.09256277233362198 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a massive crimson combine harvester?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the harvester moving across a vast field of tall, golden wheat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the harvester's metal reel rotating to gather wheat stalks onto a cutting bar?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the scene feature a low-angle camera shot emphasizing the machine's scale?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the environment bathed in warm, diffuse afternoon sunlight with a thin atmospheric haze?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07006527483463287 + ], + "opposite_logprob_tokens": [ + -2.6950652599334717 + ], + "logprob_reward": -0.07006527483463287, + "yes_no_margin_reward": 2.625, + "token_credit_reward": -0.07006527483463287, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.058842387050390244, + 0.07693061977624893, + 0.04315144941210747, + 0.11233145743608475, + 0.10671951621770859, + 0.04236464202404022, + 0.10249253362417221, + 0.03682589903473854, + 0.10747639834880829, + 0.31286507844924927 + ], + "credit_map": [ + 0.009060915559530258, + 0.006778731010854244, + 0.0081632724031806, + 0.0045675612054765224, + 0.006297925487160683, + 0.004891985096037388, + 0.011184384115040302, + 0.007897616364061832, + 0.012519900687038898, + 0.011524839326739311, + 0.012408015318214893, + 0.008954011835157871, + 0.007316693663597107, + 0.007797646336257458, + 0.006939239799976349, + 0.009470273740589619, + 0.007720521185547113, + 0.007387745194137096, + 0.004850305616855621, + 0.009555657394230366, + 0.0038468961138278246, + 0.0021135450806468725, + 0.003427319461479783, + 0.004249457735568285, + 0.017166582867503166, + 0.014555907808244228, + 0.022471420466899872, + 0.014654566533863544, + 0.009999163448810577, + 0.012574060820043087, + 0.009134690277278423, + 0.011775069870054722, + 0.02796236239373684, + 0.023677710443735123, + 0.011488104239106178, + 0.011157731525599957, + 0.008289665915071964, + 0.00784052163362503, + 0.007873143069446087, + 0.008430279791355133, + 0.008536594919860363, + 0.005701666232198477, + 0.003779594786465168, + 0.004454747773706913, + 0.005822758190333843, + 0.005204719956964254, + 0.004102780018001795, + 0.004761780146509409, + 0.014217802323400974, + 0.01374923624098301, + 0.009039309807121754, + 0.012724818661808968, + 0.012617457658052444, + 0.015808066353201866, + 0.01374710164964199, + 0.010588732548058033, + 0.0044395290315151215, + 0.005593759473413229, + 0.004097648896276951, + 0.00486439373344183, + 0.004168160725384951, + 0.0046141492202878, + 0.0048692477867007256, + 0.004179009702056646, + 0.02132873237133026, + 0.013223733752965927, + 0.0061596897430717945, + 0.006423778831958771, + 0.005407940596342087, + 0.009406701661646366, + 0.008327742107212543, + 0.037198081612586975, + 0.03779555857181549, + 0.02706601656973362, + 0.08473199605941772, + 0.039163846522569656, + 0.0243757963180542, + 0.021168258041143417, + 0.030785854905843735, + 0.04777776449918747 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 3, + 8 + ], + "selected_reward": -0.07006527483463287 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a massive crimson combine harvester?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the harvester moving across a vast field of tall, golden wheat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the harvester's metal reel rotating to gather wheat stalks onto a cutting bar?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the scene feature a low-angle camera shot emphasizing the machine's scale?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the environment bathed in warm, diffuse afternoon sunlight with a thin atmospheric haze?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.14283879101276398 + ], + "opposite_logprob_tokens": [ + -2.017838716506958 + ], + "logprob_reward": -0.14283879101276398, + "yes_no_margin_reward": 1.8749998807907104, + "token_credit_reward": -0.14283879101276398, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0696604996919632, + 0.07595545053482056, + 0.04662102088332176, + 0.08996330201625824, + 0.10595375299453735, + 0.06268520653247833, + 0.07400503754615784, + 0.02502346970140934, + 0.18389980494976044, + 0.2662324905395508 + ], + "credit_map": [ + 0.013181157410144806, + 0.014131041243672371, + 0.00963672436773777, + 0.007085559424012899, + 0.004018719308078289, + 0.0038820446934551, + 0.009914997965097427, + 0.007810261566191912, + 0.01176062598824501, + 0.008906486444175243, + 0.01983902044594288, + 0.010717492550611496, + 0.004654875956475735, + 0.006214515306055546, + 0.005241514649242163, + 0.008620917797088623, + 0.008310147561132908, + 0.007664232049137354, + 0.006503002252429724, + 0.009128730744123459, + 0.0041761258617043495, + 0.0026017306372523308, + 0.0041884188540279865, + 0.004048631060868502, + 0.012728295288980007, + 0.011857864446938038, + 0.016747841611504555, + 0.010569140315055847, + 0.008809389546513557, + 0.010961190797388554, + 0.007997026666998863, + 0.01029255148023367, + 0.01735483855009079, + 0.013060089200735092, + 0.01399314682930708, + 0.0170760340988636, + 0.010385303758084774, + 0.00971536710858345, + 0.01117203664034605, + 0.013196941465139389, + 0.008848648518323898, + 0.008308317512273788, + 0.007544160354882479, + 0.006575746927410364, + 0.013438810594379902, + 0.0071716345846652985, + 0.006136459764093161, + 0.004661427345126867, + 0.012650934979319572, + 0.010790129192173481, + 0.0069205607287585735, + 0.009886999614536762, + 0.010229632258415222, + 0.009873485192656517, + 0.00555735407397151, + 0.008095940575003624, + 0.002738700248301029, + 0.0040941690094769, + 0.0026009518187493086, + 0.003081468166783452, + 0.0037412522360682487, + 0.0028709887992590666, + 0.0026797305326908827, + 0.003216208890080452, + 0.024403836578130722, + 0.041315656155347824, + 0.007254495285451412, + 0.009283334016799927, + 0.016073422506451607, + 0.01651804707944393, + 0.010798901319503784, + 0.05825211480259895, + 0.031855881214141846, + 0.027263164520263672, + 0.07174692302942276, + 0.03261594846844673, + 0.02478102222084999, + 0.0178377702832222, + 0.024156702682375908, + 0.03597505763173103 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 8, + 4 + ], + "selected_reward": -0.14283879101276398 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a massive crimson combine harvester?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the harvester moving across a vast field of tall, golden wheat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the harvester's metal reel rotating to gather wheat stalks onto a cutting bar?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the scene feature a low-angle camera shot emphasizing the machine's scale?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the environment bathed in warm, diffuse afternoon sunlight with a thin atmospheric haze?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005947155877947807 + ], + "opposite_logprob_tokens": [ + -5.130947113037109 + ], + "logprob_reward": -0.005947155877947807, + "yes_no_margin_reward": 5.125, + "token_credit_reward": -0.005947155877947807, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15323172509670258, + 0.07956106960773468, + 0.04457136243581772, + 0.08160914480686188, + 0.12463977932929993, + 0.08993780612945557, + 0.10798163712024689, + 0.07990703731775284, + 0.08435820043087006, + 0.15420228242874146 + ], + "credit_map": [ + 0.026075849309563637, + 0.019635464996099472, + 0.01792251132428646, + 0.01625489816069603, + 0.01094600185751915, + 0.011494075879454613, + 0.0282365120947361, + 0.02266640029847622, + 0.011924820020794868, + 0.017344916239380836, + 0.011185050010681152, + 0.009060926735401154, + 0.007342757657170296, + 0.004829224199056625, + 0.0075778653845191, + 0.010295500047504902, + 0.007466692943125963, + 0.006824852433055639, + 0.005052928347140551, + 0.007093283347785473, + 0.00529817771166563, + 0.0031110376585274935, + 0.004376094322651625, + 0.005348291713744402, + 0.010685993358492851, + 0.010572386905550957, + 0.00832501333206892, + 0.017466740682721138, + 0.008106631226837635, + 0.0072311051189899445, + 0.00876875501126051, + 0.010452505201101303, + 0.03187870234251022, + 0.030924975872039795, + 0.01076213363558054, + 0.013803435489535332, + 0.010817713104188442, + 0.008191825821995735, + 0.008280654437839985, + 0.009980321861803532, + 0.014122288674116135, + 0.011986427009105682, + 0.013100924901664257, + 0.012188483029603958, + 0.01053531002253294, + 0.00644130352884531, + 0.01092230062931776, + 0.010640756227076054, + 0.013101620599627495, + 0.012061080895364285, + 0.009127598255872726, + 0.013137714006006718, + 0.01310227531939745, + 0.01703796349465847, + 0.018205653876066208, + 0.012207720428705215, + 0.009393462911248207, + 0.015336258336901665, + 0.007713606581091881, + 0.01502217072993517, + 0.009096336551010609, + 0.008575848303735256, + 0.008602186106145382, + 0.0061671617440879345, + 0.0065259565599262714, + 0.02606850303709507, + 0.0044336942955851555, + 0.003311978653073311, + 0.012272566556930542, + 0.007883329875767231, + 0.005324052181094885, + 0.018538109958171844, + 0.017928346991539, + 0.018328433856368065, + 0.03776386380195618, + 0.020329557359218597, + 0.014904121868312359, + 0.01322258822619915, + 0.014439393766224384, + 0.01728595606982708 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 0, + 4 + ], + "selected_reward": -0.005947155877947807 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a massive crimson combine harvester?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the harvester moving across a vast field of tall, golden wheat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the harvester's metal reel rotating to gather wheat stalks onto a cutting bar?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the scene feature a low-angle camera shot emphasizing the machine's scale?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the environment bathed in warm, diffuse afternoon sunlight with a thin atmospheric haze?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.00594525970518589 + ], + "opposite_logprob_tokens": [ + -5.130945205688477 + ], + "logprob_reward": -0.00594525970518589, + "yes_no_margin_reward": 5.125, + "token_credit_reward": -0.00594525970518589, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13432611525058746, + 0.05961829051375389, + 0.06434884667396545, + 0.07163360714912415, + 0.14598555862903595, + 0.09095386415719986, + 0.07359515130519867, + 0.034648116677999496, + 0.05128272995352745, + 0.2736077308654785 + ], + "credit_map": [ + 0.02507673390209675, + 0.01777852326631546, + 0.017280537635087967, + 0.01126149669289589, + 0.010337881743907928, + 0.007808702532202005, + 0.026651965454220772, + 0.018130268901586533, + 0.008150272071361542, + 0.0095951147377491, + 0.005731017328798771, + 0.006036817096173763, + 0.007464746478945017, + 0.008269446901977062, + 0.005411176476627588, + 0.008959694765508175, + 0.010664360597729683, + 0.010392709635198116, + 0.009298757649958134, + 0.014958508312702179, + 0.005054702516645193, + 0.0029729518573731184, + 0.004416355863213539, + 0.006590502802282572, + 0.010192389599978924, + 0.01017408724874258, + 0.009528156369924545, + 0.012809745967388153, + 0.00690366979688406, + 0.007013056427240372, + 0.0069063398987054825, + 0.00810615811496973, + 0.038802582770586014, + 0.0326441265642643, + 0.01487263198941946, + 0.01619766652584076, + 0.010833566077053547, + 0.010622539557516575, + 0.010750370100140572, + 0.011262085288763046, + 0.012745310552418232, + 0.01177151408046484, + 0.015054293908178806, + 0.012892693281173706, + 0.012276425957679749, + 0.00803537480533123, + 0.009274810552597046, + 0.008903436362743378, + 0.011533188633620739, + 0.011988784186542034, + 0.008224144577980042, + 0.009596369229257107, + 0.008810737170279026, + 0.008243759162724018, + 0.0071159834042191505, + 0.008082186803221703, + 0.0034007253125309944, + 0.004903238266706467, + 0.0035134360659867525, + 0.005354944616556168, + 0.003830138361081481, + 0.0026324945501983166, + 0.0050077117048203945, + 0.006005425937473774, + 0.005365682300180197, + 0.015217272564768791, + 0.004226206336170435, + 0.004011970479041338, + 0.008103002794086933, + 0.006455093156546354, + 0.002837453270331025, + 0.0050660474225878716, + 0.03518005833029747, + 0.021905653178691864, + 0.07567880302667618, + 0.029708050191402435, + 0.022003185003995895, + 0.018971074372529984, + 0.02978583611547947, + 0.04037506878376007 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 4, + 0 + ], + "selected_reward": -0.00594525970518589 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000071/group_12/index_tensor([6638, 6638])_group_12_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000071/group_12/index_tensor([6638, 6638])_group_12_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..ed94422b88539f66c165af08524a12c3d7924ab6 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000071/group_12/index_tensor([6638, 6638])_group_12_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "steady medium shot, shot from a level angle, shows a female reporter with a neat dark brown ponytail standing on a vast stretch of sand. Wearing a windproof royal blue jacket and a small clip-on microphone clipped to her collar, she reports from the scene with professional demeanor. Beside her, a magnificent and intricate sand castle rises, its soaring spires, delicate arches, and intricately carved walls. The bright afternoon sun clearly defines its textures and shadows, while the shimmering azure sea looms in the background. Suddenly, as she calmly reports, the highest spire of the sand castle begins to tremble slightly. Then, as if suddenly losing its internal support, the entire castle collapses inward with rapidity and violence. Countless grains of dry sand are flung into the air, forming a pervasive mist. Sunlight penetrates this mist, creating beams of light reminiscent of the Tyndall effect, filling the entire scene with a sense of chaotic movement. A look of pure astonishment flashed across the female reporter's face. She subconsciously took a half step back and raised her arms to try to cover herself. This sudden collapse instantly transformed the originally peaceful reporting scene into an unexpected moment full of dramatic tension.", + "seed": 14140, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000071/group_12/index_6638_seed_14140_dp_24_rank_48_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 12, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.03893596893176436, + "vlm_logprob": -0.03893596893176436, + "vlm_yes_no_margin": 4.5, + "vlm_token_credit": -0.03893596893176436, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a steady medium shot taken from a level angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a female reporter with a neat dark brown ponytail wearing a royal blue jacket and a clip-on microphone visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a magnificent sand castle with soaring spires and intricate walls stand beside the reporter before collapsing?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the sand castle collapse rapidly and violently, flinging sand into the air to create a mist with visible light beams?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the reporter react with astonishment by taking a half step back and raising her arms to cover herself?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.13530227541923523 + ], + "opposite_logprob_tokens": [ + -4.8853020668029785 + ], + "logprob_reward": -0.13530227541923523, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.13530227541923523, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2735377848148346, + 0.048465996980667114, + 0.029683420434594154, + 0.09928284585475922, + 0.04677119478583336, + 0.05970855429768562, + 0.062469009310007095, + 0.09690161049365997, + 0.14429505169391632, + 0.13888461887836456 + ], + "credit_map": [ + 0.04952719062566757, + 0.020718438550829887, + 0.016722746193408966, + 0.018128644675016403, + 0.02690035291016102, + 0.07725441455841064, + 0.02879638411104679, + 0.03548960015177727, + 0.0067613692954182625, + 0.0051732310093939304, + 0.004817428067326546, + 0.007351083215326071, + 0.006414621137082577, + 0.003645764198154211, + 0.005776155740022659, + 0.008526344783604145, + 0.0050992779433727264, + 0.002647613873705268, + 0.003309479681774974, + 0.003884334350004792, + 0.0034380534198135138, + 0.005379111040383577, + 0.0029276865534484386, + 0.002997864270582795, + 0.012704240158200264, + 0.011582131497561932, + 0.01527380757033825, + 0.025632210075855255, + 0.009183645248413086, + 0.008264423348009586, + 0.008141688071191311, + 0.00850069522857666, + 0.007318801712244749, + 0.006031745113432407, + 0.004130398388952017, + 0.00774720311164856, + 0.004106421023607254, + 0.008681932464241982, + 0.004638568498194218, + 0.0041161272674798965, + 0.006846914067864418, + 0.010732769034802914, + 0.008877694606781006, + 0.014634816907346249, + 0.005318616982549429, + 0.004385792650282383, + 0.004056259989738464, + 0.004855689592659473, + 0.00921387318521738, + 0.00925490539520979, + 0.012629839591681957, + 0.014553979970514774, + 0.005193141754716635, + 0.0034557862672954798, + 0.004229282960295677, + 0.003938197623938322, + 0.02117948606610298, + 0.02094304747879505, + 0.01598852500319481, + 0.016647759824991226, + 0.005485079251229763, + 0.00420268764719367, + 0.006334313191473484, + 0.00612070644274354, + 0.02547387033700943, + 0.031032653525471687, + 0.028690172359347343, + 0.02930920198559761, + 0.009413794614374638, + 0.006040193606168032, + 0.0075604673475027084, + 0.00677470164373517, + 0.025113215669989586, + 0.02012297697365284, + 0.030029911547899246, + 0.021185195073485374, + 0.011639852076768875, + 0.010462148115038872, + 0.010239914990961552, + 0.01009140070527792 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 9 + ], + "selected_reward": -0.13530227541923523 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a steady medium shot taken from a level angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a female reporter with a neat dark brown ponytail wearing a royal blue jacket and a clip-on microphone visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a magnificent sand castle with soaring spires and intricate walls stand beside the reporter before collapsing?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the sand castle collapse rapidly and violently, flinging sand into the air to create a mist with visible light beams?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the reporter react with astonishment by taking a half step back and raising her arms to cover herself?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.011102329008281231 + ], + "opposite_logprob_tokens": [ + -4.511102199554443 + ], + "logprob_reward": -0.011102329008281231, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.011102329008281231, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.4876249134540558, + 0.08635545521974564, + 0.039440516382455826, + 0.030624331906437874, + 0.02596723660826683, + 0.044266264885663986, + 0.06483735144138336, + 0.08986411988735199, + 0.04512413591146469, + 0.08589566498994827 + ], + "credit_map": [ + 0.08575652539730072, + 0.023899592459201813, + 0.02877846173942089, + 0.027792826294898987, + 0.04792345315217972, + 0.13689906895160675, + 0.04030667617917061, + 0.09626831859350204, + 0.014283182099461555, + 0.004519525915384293, + 0.004933785181492567, + 0.007963568903505802, + 0.006278208456933498, + 0.019531484693288803, + 0.008004321716725826, + 0.020841378718614578, + 0.005378291942179203, + 0.004041666630655527, + 0.003109337529167533, + 0.00680698873475194, + 0.003683059709146619, + 0.005015140399336815, + 0.004616159945726395, + 0.006789873354136944, + 0.0038133705966174603, + 0.002310599433258176, + 0.002180041978135705, + 0.0035576841328293085, + 0.002874988131225109, + 0.004460036288946867, + 0.004463916644454002, + 0.006963694002479315, + 0.0037313485518097878, + 0.001648889621719718, + 0.0029515638016164303, + 0.003407191252335906, + 0.002537880325689912, + 0.005631437990814447, + 0.0030606063082814217, + 0.0029983206186443567, + 0.00872891116887331, + 0.009207730181515217, + 0.006059092469513416, + 0.0072951652109622955, + 0.003554208669811487, + 0.0033507179468870163, + 0.0029328002128750086, + 0.003137638559564948, + 0.011298760771751404, + 0.010388204827904701, + 0.011363973841071129, + 0.012304294854402542, + 0.006114783231168985, + 0.004001778084784746, + 0.00469098798930645, + 0.004674570634961128, + 0.015999889001250267, + 0.014003721065819263, + 0.007908878847956657, + 0.008019529283046722, + 0.016430702060461044, + 0.007941323332488537, + 0.011302800849080086, + 0.008257279172539711, + 0.007270380388945341, + 0.008116642944514751, + 0.005517846904695034, + 0.006448197178542614, + 0.005573541857302189, + 0.0045789750292897224, + 0.00393599784001708, + 0.0036825549323111773, + 0.017103489488363266, + 0.012816096656024456, + 0.009063700214028358, + 0.00931817851960659, + 0.012378433719277382, + 0.00983239896595478, + 0.008337502367794514, + 0.007045871578156948 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 7, + 1 + ], + "selected_reward": -0.011102329008281231 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a steady medium shot taken from a level angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a female reporter with a neat dark brown ponytail wearing a royal blue jacket and a clip-on microphone visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a magnificent sand castle with soaring spires and intricate walls stand beside the reporter before collapsing?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the sand castle collapse rapidly and violently, flinging sand into the air to create a mist with visible light beams?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the reporter react with astonishment by taking a half step back and raising her arms to cover herself?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005953436251729727 + ], + "opposite_logprob_tokens": [ + -5.130953311920166 + ], + "logprob_reward": -0.005953436251729727, + "yes_no_margin_reward": 5.125, + "token_credit_reward": -0.005953436251729727, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.20861117541790009, + 0.07250599563121796, + 0.11637520045042038, + 0.06667162477970123, + 0.04680786654353142, + 0.14113613963127136, + 0.09448949992656708, + 0.09124904870986938, + 0.04053563252091408, + 0.1216178610920906 + ], + "credit_map": [ + 0.035492539405822754, + 0.008723605424165726, + 0.012827965430915356, + 0.014212632551789284, + 0.01805930770933628, + 0.050843652337789536, + 0.02083447389304638, + 0.04761699587106705, + 0.01075087022036314, + 0.00796440988779068, + 0.005931340157985687, + 0.006447321269661188, + 0.005244291387498379, + 0.014224284328520298, + 0.0065726470202207565, + 0.01537083089351654, + 0.02161419577896595, + 0.006751967594027519, + 0.008021571673452854, + 0.010572848841547966, + 0.011534713208675385, + 0.03218815103173256, + 0.01191113144159317, + 0.013780618086457253, + 0.008140278980135918, + 0.008910944685339928, + 0.006184547673910856, + 0.015182724222540855, + 0.005654493812471628, + 0.005917397327721119, + 0.006379496771842241, + 0.010301738046109676, + 0.006103979889303446, + 0.005684796720743179, + 0.005455089267343283, + 0.004929418209940195, + 0.0047468882985413074, + 0.004657967481762171, + 0.005577806383371353, + 0.009651916101574898, + 0.01802874729037285, + 0.026480782777071, + 0.02198578417301178, + 0.03420736640691757, + 0.012256735004484653, + 0.011558039113879204, + 0.0072249132208526134, + 0.009393766522407532, + 0.013897297903895378, + 0.016538485884666443, + 0.02412242256104946, + 0.023603152483701706, + 0.005016970913857222, + 0.003517685690894723, + 0.003363798838108778, + 0.004429688677191734, + 0.0168191846460104, + 0.01803944818675518, + 0.015234281308948994, + 0.01750655099749565, + 0.007345590274780989, + 0.0042505827732384205, + 0.006099451333284378, + 0.005953962914645672, + 0.006880856119096279, + 0.00650014728307724, + 0.009266084060072899, + 0.008818957023322582, + 0.0027073032688349485, + 0.0021994917187839746, + 0.001983171096071601, + 0.0021796238142997026, + 0.018209222704172134, + 0.014306551776826382, + 0.012253872118890285, + 0.017979858443140984, + 0.01286250352859497, + 0.017736444249749184, + 0.01403284166008234, + 0.014236564747989178 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 5, + 9 + ], + "selected_reward": -0.005953436251729727 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a steady medium shot taken from a level angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a female reporter with a neat dark brown ponytail wearing a royal blue jacket and a clip-on microphone visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a magnificent sand castle with soaring spires and intricate walls stand beside the reporter before collapsing?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the sand castle collapse rapidly and violently, flinging sand into the air to create a mist with visible light beams?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the reporter react with astonishment by taking a half step back and raising her arms to cover herself?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.008640876971185207 + ], + "opposite_logprob_tokens": [ + -4.758640766143799 + ], + "logprob_reward": -0.008640876971185207, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.008640876971185207, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.20941226184368134, + 0.04720088839530945, + 0.031146574765443802, + 0.10018855333328247, + 0.06703035533428192, + 0.18757514655590057, + 0.12098725140094757, + 0.03436746448278427, + 0.02843335084617138, + 0.17365813255310059 + ], + "credit_map": [ + 0.027709001675248146, + 0.01719847321510315, + 0.014202095568180084, + 0.021584825590252876, + 0.016675248742103577, + 0.03167421743273735, + 0.03272974118590355, + 0.0476386733353138, + 0.0057174330577254295, + 0.004389260895550251, + 0.005510581657290459, + 0.005861527752131224, + 0.0035790023393929005, + 0.005053291097283363, + 0.006514584179967642, + 0.010575207881629467, + 0.0041020517237484455, + 0.002694074297323823, + 0.004651282448321581, + 0.005188588984310627, + 0.0023864428512752056, + 0.002634888980537653, + 0.0034028340596705675, + 0.006086413748562336, + 0.011242561042308807, + 0.011669176630675793, + 0.009282740764319897, + 0.024209238588809967, + 0.007031123153865337, + 0.010324074886739254, + 0.009240865707397461, + 0.01718878000974655, + 0.009197166189551353, + 0.011290584690868855, + 0.008241881616413593, + 0.011867898516356945, + 0.00520146731287241, + 0.006360185798257589, + 0.006520877126604319, + 0.008350295014679432, + 0.023441163823008537, + 0.034643951803445816, + 0.023054488003253937, + 0.03934112936258316, + 0.019972234964370728, + 0.022312769666314125, + 0.012341916561126709, + 0.012467483058571815, + 0.02091696299612522, + 0.020523332059383392, + 0.022734196856617928, + 0.03186866641044617, + 0.008051029406487942, + 0.005051552318036556, + 0.005327476654201746, + 0.006514037027955055, + 0.006473070941865444, + 0.006517138797789812, + 0.0062239887192845345, + 0.007525639142841101, + 0.0020112961065024137, + 0.001543981023132801, + 0.0017209193902090192, + 0.0023514290805906057, + 0.004597817547619343, + 0.0050904774107038975, + 0.006592800375074148, + 0.005728383548557758, + 0.0018840515986084938, + 0.0015332995681092143, + 0.0014528263127431273, + 0.0015536947175860405, + 0.027488593012094498, + 0.02144778147339821, + 0.02510671131312847, + 0.022034229710698128, + 0.018450336530804634, + 0.02227032370865345, + 0.02057827077805996, + 0.01628189906477928 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 5, + 9 + ], + "selected_reward": -0.008640876971185207 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a steady medium shot taken from a level angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a female reporter with a neat dark brown ponytail wearing a royal blue jacket and a clip-on microphone visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a magnificent sand castle with soaring spires and intricate walls stand beside the reporter before collapsing?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the sand castle collapse rapidly and violently, flinging sand into the air to create a mist with visible light beams?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the reporter react with astonishment by taking a half step back and raising her arms to cover herself?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03368092700839043 + ], + "opposite_logprob_tokens": [ + -3.4086809158325195 + ], + "logprob_reward": -0.03368092700839043, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.03368092700839043, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.17584562301635742, + 0.04682702198624611, + 0.028465155512094498, + 0.04811198636889458, + 0.047813933342695236, + 0.04612485691905022, + 0.04582289233803749, + 0.048211194574832916, + 0.07676102966070175, + 0.4360162913799286 + ], + "credit_map": [ + 0.024337494745850563, + 0.010523047298192978, + 0.010349929332733154, + 0.01345139741897583, + 0.009379800409078598, + 0.02388572320342064, + 0.0245207529515028, + 0.05939747765660286, + 0.00785424280911684, + 0.002998292213305831, + 0.0026770078111439943, + 0.0038065481930971146, + 0.0036961224395781755, + 0.011281960643827915, + 0.003986649680882692, + 0.010526200756430626, + 0.004160769283771515, + 0.002699153497815132, + 0.0022389760706573725, + 0.003261848818510771, + 0.0024133631959557533, + 0.004261322319507599, + 0.0028877570293843746, + 0.00654196459800005, + 0.0053802612237632275, + 0.004588061012327671, + 0.004172513727098703, + 0.011355442926287651, + 0.003335529938340187, + 0.00354884029366076, + 0.005384956020861864, + 0.01034637913107872, + 0.0074734860099852085, + 0.0036475646775215864, + 0.004970672074705362, + 0.004405560903251171, + 0.004106673412024975, + 0.011999422684311867, + 0.005418409593403339, + 0.005792141426354647, + 0.007189515512436628, + 0.010145971551537514, + 0.006468781735748053, + 0.006399938371032476, + 0.0031791338697075844, + 0.0036932043731212616, + 0.003062387928366661, + 0.005985924042761326, + 0.008611819706857204, + 0.009046515449881554, + 0.006637968588620424, + 0.01006859727203846, + 0.0033327890560030937, + 0.003338189097121358, + 0.0022890605032444, + 0.002497951267287135, + 0.007142615504562855, + 0.007539018522948027, + 0.009305077604949474, + 0.007194333244115114, + 0.005575041752308607, + 0.003373422659933567, + 0.0037884775083512068, + 0.004293210804462433, + 0.010322241112589836, + 0.01077159121632576, + 0.01198045164346695, + 0.010054333135485649, + 0.011254140175879002, + 0.007427691016346216, + 0.007015532813966274, + 0.007935053668916225, + 0.06962285935878754, + 0.04964801296591759, + 0.041688017547130585, + 0.05353187397122383, + 0.042577680200338364, + 0.0656246617436409, + 0.05779944360256195, + 0.05552375689148903 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 0, + 8 + ], + "selected_reward": -0.03368092700839043 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "steady medium shot, shot from a level angle, shows a female reporter with a neat dark brown ponytail standing on a vast stretch of sand. Wearing a windproof royal blue jacket and a small clip-on microphone clipped to her collar, she reports from the scene with professional demeanor. Beside her, a magnificent and intricate sand castle rises, its soaring spires, delicate arches, and intricately carved walls. The bright afternoon sun clearly defines its textures and shadows, while the shimmering azure sea looms in the background. Suddenly, as she calmly reports, the highest spire of the sand castle begins to tremble slightly. Then, as if suddenly losing its internal support, the entire castle collapses inward with rapidity and violence. Countless grains of dry sand are flung into the air, forming a pervasive mist. Sunlight penetrates this mist, creating beams of light reminiscent of the Tyndall effect, filling the entire scene with a sense of chaotic movement. A look of pure astonishment flashed across the female reporter's face. She subconsciously took a half step back and raised her arms to try to cover herself. This sudden collapse instantly transformed the originally peaceful reporting scene into an unexpected moment full of dramatic tension.", + "seed": 14141, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000071/group_12/index_6638_seed_14141_dp_24_rank_48_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 12, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.05188741330057382, + "vlm_logprob": -0.05188741330057382, + "vlm_yes_no_margin": 3.775, + "vlm_token_credit": -0.05188741330057382, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a steady medium shot taken from a level angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a female reporter with a neat dark brown ponytail wearing a royal blue jacket and a clip-on microphone visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a magnificent sand castle with soaring spires and intricate walls stand beside the reporter before collapsing?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the sand castle collapse rapidly and violently, flinging sand into the air to create a mist with visible light beams?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the reporter react with astonishment by taking a half step back and raising her arms to cover herself?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.14930881559848785 + ], + "opposite_logprob_tokens": [ + -3.899308919906616 + ], + "logprob_reward": -0.14930881559848785, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.14930881559848785, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.250144362449646, + 0.036968305706977844, + 0.03503480181097984, + 0.03216289356350899, + 0.08162452280521393, + 0.03739457204937935, + 0.09514056146144867, + 0.2247464507818222, + 0.06657491624355316, + 0.14020858705043793 + ], + "credit_map": [ + 0.0261662770062685, + 0.03034375235438347, + 0.018118543550372124, + 0.015114787966012955, + 0.04359434172511101, + 0.02775450050830841, + 0.05251019448041916, + 0.036541979759931564, + 0.004951108247041702, + 0.0045249625109136105, + 0.0023528188467025757, + 0.002205661963671446, + 0.004411853384226561, + 0.006298319902271032, + 0.006368522997945547, + 0.005855055991560221, + 0.00467600068077445, + 0.0038526460994035006, + 0.0019711677450686693, + 0.0021329056471586227, + 0.007410668767988682, + 0.004978425335139036, + 0.0044577536173164845, + 0.005555231589823961, + 0.002969499910250306, + 0.002500733360648155, + 0.0019870009273290634, + 0.0021295789629220963, + 0.002956605516374111, + 0.006381363142281771, + 0.006468768231570721, + 0.006769341416656971, + 0.01250417809933424, + 0.00414237380027771, + 0.003462641965597868, + 0.0032965491991490126, + 0.02510013058781624, + 0.020997853949666023, + 0.005251961760222912, + 0.0068688285537064075, + 0.004674513358622789, + 0.00618572486564517, + 0.0028935750015079975, + 0.0033534718677401543, + 0.006247927434742451, + 0.0040866779163479805, + 0.005131354555487633, + 0.004821327514946461, + 0.008510233834385872, + 0.00554265221580863, + 0.006728044245392084, + 0.007223688531666994, + 0.013900856487452984, + 0.020481236279010773, + 0.014817094430327415, + 0.017936762422323227, + 0.029995812103152275, + 0.01203321386128664, + 0.05861325189471245, + 0.017202045768499374, + 0.03841027989983559, + 0.014742569997906685, + 0.029437489807605743, + 0.024311790242791176, + 0.006693453527987003, + 0.007865706458687782, + 0.006492867134511471, + 0.005480599123984575, + 0.008801647461950779, + 0.009890664368867874, + 0.013292397372424603, + 0.008057584054768085, + 0.013169213198125362, + 0.017800074070692062, + 0.018781516700983047, + 0.01571323722600937, + 0.021207313984632492, + 0.016631657257676125, + 0.021028665825724602, + 0.01587691716849804 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 7, + 9 + ], + "selected_reward": -0.14930881559848785 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a steady medium shot taken from a level angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a female reporter with a neat dark brown ponytail wearing a royal blue jacket and a clip-on microphone visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a magnificent sand castle with soaring spires and intricate walls stand beside the reporter before collapsing?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the sand castle collapse rapidly and violently, flinging sand into the air to create a mist with visible light beams?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the reporter react with astonishment by taking a half step back and raising her arms to cover herself?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.009809031151235104 + ], + "opposite_logprob_tokens": [ + -4.6348090171813965 + ], + "logprob_reward": -0.009809031151235104, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.009809031151235104, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.19198937714099884, + 0.07276985049247742, + 0.043871182948350906, + 0.01763070933520794, + 0.1633872389793396, + 0.07448522001504898, + 0.0973493903875351, + 0.12130545824766159, + 0.08208394050598145, + 0.13512766361236572 + ], + "credit_map": [ + 0.023351438343524933, + 0.024234525859355927, + 0.020264632999897003, + 0.02280030958354473, + 0.039514604955911636, + 0.01252036914229393, + 0.020096275955438614, + 0.02920721098780632, + 0.00998660083860159, + 0.009831193834543228, + 0.007064310368150473, + 0.007204536348581314, + 0.0073821330443024635, + 0.008347274735569954, + 0.01104571484029293, + 0.011908091604709625, + 0.0078116971999406815, + 0.004134071990847588, + 0.0032132097985595465, + 0.00304278451949358, + 0.004640914965420961, + 0.004662516061216593, + 0.007633603177964687, + 0.008732385002076626, + 0.0024228275287896395, + 0.0026128427125513554, + 0.0018527802312746644, + 0.0025845009367913008, + 0.0020328001119196415, + 0.0021535600535571575, + 0.0017595720710232854, + 0.002211824292317033, + 0.027667680755257607, + 0.00742174219340086, + 0.008156772702932358, + 0.006885791663080454, + 0.05212385207414627, + 0.03809388726949692, + 0.010126279667019844, + 0.012911232188344002, + 0.0072518158704042435, + 0.012262847274541855, + 0.0076748100109398365, + 0.007542500272393227, + 0.009934852831065655, + 0.008813712745904922, + 0.012960419058799744, + 0.00804426334798336, + 0.009832034818828106, + 0.011962548829615116, + 0.012989533133804798, + 0.010483378544449806, + 0.014306873083114624, + 0.011644601821899414, + 0.012589284218847752, + 0.013541138730943203, + 0.013262106105685234, + 0.011071554385125637, + 0.025491507723927498, + 0.015715187415480614, + 0.01831529662013054, + 0.010259948670864105, + 0.014298728667199612, + 0.012891127727925777, + 0.0062171947211027145, + 0.006729491986334324, + 0.009503258392214775, + 0.005568544380366802, + 0.010901921428740025, + 0.0071654412895441055, + 0.02150651067495346, + 0.014491571113467216, + 0.013596423901617527, + 0.014255688525736332, + 0.02636713907122612, + 0.019047703593969345, + 0.015935314819216728, + 0.009313318878412247, + 0.01650897227227688, + 0.020103104412555695 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 9 + ], + "selected_reward": -0.009809031151235104 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a steady medium shot taken from a level angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a female reporter with a neat dark brown ponytail wearing a royal blue jacket and a clip-on microphone visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a magnificent sand castle with soaring spires and intricate walls stand beside the reporter before collapsing?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the sand castle collapse rapidly and violently, flinging sand into the air to create a mist with visible light beams?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the reporter react with astonishment by taking a half step back and raising her arms to cover herself?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.014206347987055779 + ], + "opposite_logprob_tokens": [ + -4.264206409454346 + ], + "logprob_reward": -0.014206347987055779, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.014206347987055779, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.052618518471717834, + 0.08092861622571945, + 0.2513095736503601, + 0.03352559730410576, + 0.11769598722457886, + 0.03329787775874138, + 0.06559444218873978, + 0.09933503717184067, + 0.04200456663966179, + 0.22368980944156647 + ], + "credit_map": [ + 0.006195242516696453, + 0.007325355429202318, + 0.005335095804184675, + 0.005069573875516653, + 0.00573606276884675, + 0.007490728050470352, + 0.006722717545926571, + 0.008743743412196636, + 0.007831059396266937, + 0.007104994263499975, + 0.00472942553460598, + 0.004587991163134575, + 0.008215014822781086, + 0.01554900873452425, + 0.01824701949954033, + 0.014664104208350182, + 0.026834828779101372, + 0.019122570753097534, + 0.010541205294430256, + 0.011109771206974983, + 0.023470183834433556, + 0.0416565015912056, + 0.05627625435590744, + 0.0622982531785965, + 0.004058866761624813, + 0.0028079384937882423, + 0.0016496048774570227, + 0.0015771404141560197, + 0.0035855728201568127, + 0.00522894412279129, + 0.007316452916711569, + 0.007301074452698231, + 0.023453807458281517, + 0.006028389558196068, + 0.0033927757758647203, + 0.0027033265214413404, + 0.04074516519904137, + 0.027000978589057922, + 0.006105991546064615, + 0.008265554904937744, + 0.0029305892530828714, + 0.004188760183751583, + 0.001952844439074397, + 0.002223641611635685, + 0.005289604887366295, + 0.005234603304415941, + 0.0070493752136826515, + 0.004428455606102943, + 0.007443328853696585, + 0.003955277614295483, + 0.007579383440315723, + 0.0055085355415940285, + 0.009692680090665817, + 0.011238757520914078, + 0.008471472188830376, + 0.0117050064727664, + 0.012794614769518375, + 0.005784145090728998, + 0.015329097397625446, + 0.009342639707028866, + 0.015604466199874878, + 0.009396006353199482, + 0.016284899786114693, + 0.014799161814153194, + 0.0034841480664908886, + 0.0038062836974859238, + 0.00690164091065526, + 0.004340941086411476, + 0.005723841488361359, + 0.004904177039861679, + 0.00779036246240139, + 0.005053170956671238, + 0.020535051822662354, + 0.027817342430353165, + 0.04205002263188362, + 0.038966938853263855, + 0.028496308252215385, + 0.01681540720164776, + 0.023070180788636208, + 0.02593855746090412 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 9, + 4 + ], + "selected_reward": -0.014206347987055779 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a steady medium shot taken from a level angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a female reporter with a neat dark brown ponytail wearing a royal blue jacket and a clip-on microphone visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a magnificent sand castle with soaring spires and intricate walls stand beside the reporter before collapsing?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the sand castle collapse rapidly and violently, flinging sand into the air to create a mist with visible light beams?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the reporter react with astonishment by taking a half step back and raising her arms to cover herself?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04306522756814957 + ], + "opposite_logprob_tokens": [ + -3.168065309524536 + ], + "logprob_reward": -0.04306522756814957, + "yes_no_margin_reward": 3.125, + "token_credit_reward": -0.04306522756814957, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1698399931192398, + 0.11107514798641205, + 0.1358506828546524, + 0.013532496057450771, + 0.08316056430339813, + 0.03167200833559036, + 0.045501239597797394, + 0.09334392100572586, + 0.05519923195242882, + 0.26082465052604675 + ], + "credit_map": [ + 0.017179405316710472, + 0.020545925945043564, + 0.012306232936680317, + 0.014850247651338577, + 0.029453737661242485, + 0.017121696844697, + 0.028420673683285713, + 0.029962098225951195, + 0.009697885252535343, + 0.0090072276070714, + 0.006068797316402197, + 0.006097808480262756, + 0.011452410370111465, + 0.021462149918079376, + 0.02480989322066307, + 0.022478986531496048, + 0.01612599939107895, + 0.01110865082591772, + 0.006161594297736883, + 0.0063472711481153965, + 0.012853282503783703, + 0.020861579105257988, + 0.029223941266536713, + 0.03316837176680565, + 0.001464633154682815, + 0.0013853327836841345, + 0.0010229452745988965, + 0.0011297727469354868, + 0.0016764532774686813, + 0.00218356866389513, + 0.0022140806540846825, + 0.0024557122960686684, + 0.014598693698644638, + 0.004062062595039606, + 0.002314870711416006, + 0.0017701647011563182, + 0.027913810685276985, + 0.022420376539230347, + 0.003904949175193906, + 0.006175647955387831, + 0.0028352185618132353, + 0.0033387383446097374, + 0.0016475060256198049, + 0.0023342836648225784, + 0.0039006390143185854, + 0.005119870416820049, + 0.007935039699077606, + 0.004560717381536961, + 0.0050346218049526215, + 0.003452925244346261, + 0.005236402153968811, + 0.0036037792451679707, + 0.0061193290166556835, + 0.006169650703668594, + 0.0068718912079930305, + 0.009012645110487938, + 0.00985630787909031, + 0.004976787604391575, + 0.012195536866784096, + 0.008948025293648243, + 0.013567546382546425, + 0.008754625916481018, + 0.01738082990050316, + 0.01766427606344223, + 0.005078502930700779, + 0.0037352752406150103, + 0.0072417776100337505, + 0.00432828301563859, + 0.008532731793820858, + 0.007606094237416983, + 0.01177382841706276, + 0.006902744993567467, + 0.02454528585076332, + 0.03009217604994774, + 0.05076891928911209, + 0.043480388820171356, + 0.032096732407808304, + 0.018037747591733932, + 0.028319019824266434, + 0.03348439931869507 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 0, + 2 + ], + "selected_reward": -0.04306522756814957 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a steady medium shot taken from a level angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a female reporter with a neat dark brown ponytail wearing a royal blue jacket and a clip-on microphone visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a magnificent sand castle with soaring spires and intricate walls stand beside the reporter before collapsing?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the sand castle collapse rapidly and violently, flinging sand into the air to create a mist with visible light beams?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the reporter react with astonishment by taking a half step back and raising her arms to cover herself?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.043047644197940826 + ], + "opposite_logprob_tokens": [ + -3.1680476665496826 + ], + "logprob_reward": -0.043047644197940826, + "yes_no_margin_reward": 3.125, + "token_credit_reward": -0.043047644197940826, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06508447974920273, + 0.049859486520290375, + 0.016831113025546074, + 0.03184419870376587, + 0.07395760715007782, + 0.026820404455065727, + 0.06993123888969421, + 0.1173420175909996, + 0.1267833113670349, + 0.4215461313724518 + ], + "credit_map": [ + 0.00834202766418457, + 0.006765902042388916, + 0.006828620098531246, + 0.0068583120591938496, + 0.009947081096470356, + 0.006601156201213598, + 0.010171735659241676, + 0.009569640271365643, + 0.004747786559164524, + 0.004403586499392986, + 0.003150500124320388, + 0.0031794728711247444, + 0.0049230726435780525, + 0.008680211380124092, + 0.01036527007818222, + 0.010409586131572723, + 0.001890829880721867, + 0.0020944608841091394, + 0.0011478860396891832, + 0.0016388241201639175, + 0.0024582091718912125, + 0.0025595168117433786, + 0.0023663120809942484, + 0.0026750750839710236, + 0.002872797194868326, + 0.0029489558655768633, + 0.001981037901714444, + 0.0019116137409582734, + 0.0036205719225108624, + 0.00611448148265481, + 0.005835356656461954, + 0.006559381261467934, + 0.012169728986918926, + 0.003673288971185684, + 0.0023043763358145952, + 0.001984789501875639, + 0.023736117407679558, + 0.0193681251257658, + 0.004596475511789322, + 0.006124704610556364, + 0.002459554001688957, + 0.0035606659948825836, + 0.001832532580010593, + 0.0019085524836555123, + 0.002853955375030637, + 0.004506418015807867, + 0.006262366659939289, + 0.003436358878389001, + 0.0069672297686338425, + 0.004784123040735722, + 0.009114853106439114, + 0.008354765363037586, + 0.00924512930214405, + 0.009019996039569378, + 0.010808080434799194, + 0.011637061834335327, + 0.018118493258953094, + 0.005915764253586531, + 0.027420008555054665, + 0.010370488278567791, + 0.021198878064751625, + 0.009333770722150803, + 0.01286817155778408, + 0.012116446159780025, + 0.01061213482171297, + 0.010705439373850822, + 0.01560274139046669, + 0.008621440269052982, + 0.018863193690776825, + 0.007900199852883816, + 0.03384565934538841, + 0.02063250169157982, + 0.04408007860183716, + 0.05284227803349495, + 0.07696743309497833, + 0.07192067056894302, + 0.05171572044491768, + 0.03088260628283024, + 0.043774284422397614, + 0.04936304688453674 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 8, + 7 + ], + "selected_reward": -0.043047644197940826 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "steady medium shot, shot from a level angle, shows a female reporter with a neat dark brown ponytail standing on a vast stretch of sand. Wearing a windproof royal blue jacket and a small clip-on microphone clipped to her collar, she reports from the scene with professional demeanor. Beside her, a magnificent and intricate sand castle rises, its soaring spires, delicate arches, and intricately carved walls. The bright afternoon sun clearly defines its textures and shadows, while the shimmering azure sea looms in the background. Suddenly, as she calmly reports, the highest spire of the sand castle begins to tremble slightly. Then, as if suddenly losing its internal support, the entire castle collapses inward with rapidity and violence. Countless grains of dry sand are flung into the air, forming a pervasive mist. Sunlight penetrates this mist, creating beams of light reminiscent of the Tyndall effect, filling the entire scene with a sense of chaotic movement. A look of pure astonishment flashed across the female reporter's face. She subconsciously took a half step back and raised her arms to try to cover herself. This sudden collapse instantly transformed the originally peaceful reporting scene into an unexpected moment full of dramatic tension.", + "seed": 14142, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000071/group_12/index_6638_seed_14142_dp_25_rank_50_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 12, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.04009365017991513, + "vlm_logprob": -0.04009365017991513, + "vlm_yes_no_margin": 5.075, + "vlm_token_credit": -0.04009365017991513, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a steady medium shot taken from a level angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a female reporter with a neat dark brown ponytail wearing a royal blue jacket and a clip-on microphone visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a magnificent sand castle with soaring spires and intricate walls stand beside the reporter before collapsing?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the sand castle collapse rapidly and violently, flinging sand into the air to create a mist with visible light beams?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the reporter react with astonishment by taking a half step back and raising her arms to cover herself?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.16946282982826233 + ], + "opposite_logprob_tokens": [ + -4.79446268081665 + ], + "logprob_reward": -0.16946282982826233, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.16946282982826233, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09141413867473602, + 0.024805236607789993, + 0.015215097926557064, + 0.027388595044612885, + 0.02344249188899994, + 0.021532472223043442, + 0.03563836216926575, + 0.46971675753593445, + 0.11232254654169083, + 0.17852434515953064 + ], + "credit_map": [ + 0.013182003982365131, + 0.014220578595995903, + 0.011686716228723526, + 0.018153682351112366, + 0.009343880228698254, + 0.008278298191726208, + 0.008540628477931023, + 0.008008360862731934, + 0.003537268377840519, + 0.004569158423691988, + 0.0036170172970741987, + 0.0043524885550141335, + 0.0023808153346180916, + 0.002462448086589575, + 0.00202712113969028, + 0.0018589183455333114, + 0.0021972174290567636, + 0.0026014079339802265, + 0.0027595481369644403, + 0.00309428246691823, + 0.001022780081257224, + 0.0012762078549712896, + 0.0011236253194510937, + 0.0011400298681110144, + 0.004134437069296837, + 0.003938388079404831, + 0.005152896512299776, + 0.0051171728409826756, + 0.002167320577427745, + 0.002907075686380267, + 0.001786923734471202, + 0.0021843821741640568, + 0.0034907618537545204, + 0.0017317559104412794, + 0.004712223075330257, + 0.0021420959383249283, + 0.0026247482746839523, + 0.004333954304456711, + 0.0019319953862577677, + 0.0024749573785811663, + 0.0028003747574985027, + 0.0018356159562245011, + 0.00514353159815073, + 0.0027446772437542677, + 0.0021904276218265295, + 0.0028944346122443676, + 0.0018176307203248143, + 0.0021057799458503723, + 0.004530259408056736, + 0.004990207497030497, + 0.005647369660437107, + 0.00912032462656498, + 0.00277703651227057, + 0.003746080445125699, + 0.0026284425985068083, + 0.0021986437495797873, + 0.04375810548663139, + 0.0529780238866806, + 0.07465813308954239, + 0.14972320199012756, + 0.044579438865184784, + 0.03301297500729561, + 0.05132979899644852, + 0.019677089527249336, + 0.009748095646500587, + 0.014941686764359474, + 0.018813587725162506, + 0.027533702552318573, + 0.017594704404473305, + 0.007759337779134512, + 0.01084789540618658, + 0.0050835320726037025, + 0.019502896815538406, + 0.019251124933362007, + 0.04512941837310791, + 0.04282670095562935, + 0.016580618917942047, + 0.009845566935837269, + 0.014211504720151424, + 0.011176511645317078 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 9, + 8 + ], + "selected_reward": -0.16946282982826233 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a steady medium shot taken from a level angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a female reporter with a neat dark brown ponytail wearing a royal blue jacket and a clip-on microphone visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a magnificent sand castle with soaring spires and intricate walls stand beside the reporter before collapsing?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the sand castle collapse rapidly and violently, flinging sand into the air to create a mist with visible light beams?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the reporter react with astonishment by taking a half step back and raising her arms to cover herself?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.016091477125883102 + ], + "opposite_logprob_tokens": [ + -4.141091346740723 + ], + "logprob_reward": -0.016091477125883102, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.016091477125883102, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.19199752807617188, + 0.05522413179278374, + 0.02648242563009262, + 0.03266424313187599, + 0.0389096699655056, + 0.03703881800174713, + 0.06414700299501419, + 0.2327776700258255, + 0.13363215327262878, + 0.18712638318538666 + ], + "credit_map": [ + 0.0270948838442564, + 0.014857606031000614, + 0.03154444694519043, + 0.023699980229139328, + 0.024269862100481987, + 0.031190859153866768, + 0.020706983283162117, + 0.018632901832461357, + 0.006781542208045721, + 0.004507138859480619, + 0.009109171107411385, + 0.009543782100081444, + 0.005592788103967905, + 0.008215547539293766, + 0.005902844946831465, + 0.005571314599364996, + 0.003425509901717305, + 0.0031295963563024998, + 0.0036774147301912308, + 0.005360854789614677, + 0.0027272289153188467, + 0.0035074520856142044, + 0.0023051195312291384, + 0.0023492472246289253, + 0.004422695375978947, + 0.004449647851288319, + 0.005618204828351736, + 0.007182139903306961, + 0.0027086306363344193, + 0.003506614826619625, + 0.002295498037710786, + 0.002480813069269061, + 0.004835331346839666, + 0.003140924032777548, + 0.006356752477586269, + 0.004344344604760408, + 0.004977026954293251, + 0.006497783586382866, + 0.003774563316255808, + 0.004982943646609783, + 0.005271122790873051, + 0.0033318016212433577, + 0.006304750684648752, + 0.0042419834062457085, + 0.00572916679084301, + 0.004782489500939846, + 0.003835082519799471, + 0.003542419755831361, + 0.011183568276464939, + 0.007831800729036331, + 0.017144542187452316, + 0.0114213852211833, + 0.004845388699322939, + 0.004325155168771744, + 0.00424201088026166, + 0.003153156954795122, + 0.02109626494348049, + 0.027853108942508698, + 0.035542044788599014, + 0.0745389461517334, + 0.021236611530184746, + 0.015165381133556366, + 0.026497764512896538, + 0.0108475461602211, + 0.01747352071106434, + 0.01830504834651947, + 0.018798962235450745, + 0.0326002836227417, + 0.017572438344359398, + 0.008777081966400146, + 0.013942103832960129, + 0.006162704434245825, + 0.02248917520046234, + 0.018499769270420074, + 0.033102359622716904, + 0.02472357451915741, + 0.037921562790870667, + 0.01640349254012108, + 0.02025546319782734, + 0.013730984181165695 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 0, + 9 + ], + "selected_reward": -0.016091477125883102 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a steady medium shot taken from a level angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a female reporter with a neat dark brown ponytail wearing a royal blue jacket and a clip-on microphone visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a magnificent sand castle with soaring spires and intricate walls stand beside the reporter before collapsing?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the sand castle collapse rapidly and violently, flinging sand into the air to create a mist with visible light beams?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the reporter react with astonishment by taking a half step back and raising her arms to cover herself?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0031969649717211723 + ], + "opposite_logprob_tokens": [ + -5.753197193145752 + ], + "logprob_reward": -0.0031969649717211723, + "yes_no_margin_reward": 5.75, + "token_credit_reward": -0.0031969649717211723, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.16034775972366333, + 0.032981839030981064, + 0.03359053656458855, + 0.01904633454978466, + 0.01732509955763817, + 0.028245287016034126, + 0.19764471054077148, + 0.16336926817893982, + 0.2038891762495041, + 0.143559992313385 + ], + "credit_map": [ + 0.022600403055548668, + 0.015909112989902496, + 0.026727896183729172, + 0.028328074142336845, + 0.015994975343346596, + 0.019739016890525818, + 0.01679980382323265, + 0.014248485676944256, + 0.004830635152757168, + 0.00543473195284605, + 0.003994459751993418, + 0.006114518269896507, + 0.00334235280752182, + 0.0029780531767755747, + 0.0032013452146202326, + 0.003085743635892868, + 0.003692604135721922, + 0.00445950822904706, + 0.004698790609836578, + 0.009857517667114735, + 0.003084515919908881, + 0.0024559299927204847, + 0.0029764065984636545, + 0.0023652634117752314, + 0.0024476530961692333, + 0.0024534722324460745, + 0.0030413882341235876, + 0.003382254159078002, + 0.0016595162451267242, + 0.0023285336792469025, + 0.0018472453812137246, + 0.001886271988041699, + 0.0022466620430350304, + 0.0020878291688859463, + 0.00232551246881485, + 0.003293293993920088, + 0.0019238422391936183, + 0.002196374349296093, + 0.0016199600649997592, + 0.0016316256951540709, + 0.003686488838866353, + 0.0027602387126535177, + 0.007710792124271393, + 0.0041895220056176186, + 0.003462698310613632, + 0.00261182663962245, + 0.002047381130978465, + 0.001776339951902628, + 0.03043343871831894, + 0.025520188733935356, + 0.05093845725059509, + 0.04542945325374603, + 0.013496769592165947, + 0.012753397226333618, + 0.010858139023184776, + 0.008214876987040043, + 0.01948019117116928, + 0.02121887542307377, + 0.021248482167720795, + 0.04892772063612938, + 0.019091807305812836, + 0.010063007473945618, + 0.015119466930627823, + 0.008219709619879723, + 0.023189540952444077, + 0.02749096229672432, + 0.030939804390072823, + 0.05955788493156433, + 0.018485920503735542, + 0.016593249514698982, + 0.017199018970131874, + 0.01043279841542244, + 0.01730356179177761, + 0.015818441286683083, + 0.03358439728617668, + 0.025091463699936867, + 0.015544753521680832, + 0.01144255232065916, + 0.016332440078258514, + 0.008442371152341366 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 6, + 7 + ], + "selected_reward": -0.0031969649717211723 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a steady medium shot taken from a level angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a female reporter with a neat dark brown ponytail wearing a royal blue jacket and a clip-on microphone visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a magnificent sand castle with soaring spires and intricate walls stand beside the reporter before collapsing?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the sand castle collapse rapidly and violently, flinging sand into the air to create a mist with visible light beams?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the reporter react with astonishment by taking a half step back and raising her arms to cover herself?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0019427727675065398 + ], + "opposite_logprob_tokens": [ + -6.2519426345825195 + ], + "logprob_reward": -0.0019427727675065398, + "yes_no_margin_reward": 6.25, + "token_credit_reward": -0.0019427727675065398, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12324607372283936, + 0.03430595248937607, + 0.027631614357233047, + 0.014596297405660152, + 0.03438806161284447, + 0.03246362507343292, + 0.14487983286380768, + 0.20715919137001038, + 0.2271731197834015, + 0.15415623784065247 + ], + "credit_map": [ + 0.01615053415298462, + 0.011925716884434223, + 0.01662377640604973, + 0.014573469758033752, + 0.02063216082751751, + 0.01836421526968479, + 0.013346133753657341, + 0.011630067601799965, + 0.004331014584749937, + 0.004041677340865135, + 0.004356401041150093, + 0.005649849772453308, + 0.004492402542382479, + 0.004607030656188726, + 0.003788253292441368, + 0.0030393197666853666, + 0.0035706846974790096, + 0.0033441910054534674, + 0.003691291669383645, + 0.005975050386041403, + 0.0036269621923565865, + 0.002895752666518092, + 0.0022683909628540277, + 0.0022592900786548853, + 0.0019395698327571154, + 0.002477717585861683, + 0.0024353456683456898, + 0.0022152408491820097, + 0.0014886499848216772, + 0.001708770520053804, + 0.0011662583565339446, + 0.001164746005088091, + 0.0035318457521498203, + 0.004807991907000542, + 0.00486020976677537, + 0.007331282831728458, + 0.004019568674266338, + 0.003932267893105745, + 0.002968357875943184, + 0.002936534583568573, + 0.004326339345425367, + 0.003677139524370432, + 0.004317696671932936, + 0.004273812286555767, + 0.0066936807706952095, + 0.0029224895406514406, + 0.0035351829137653112, + 0.0027172837872058153, + 0.019638769328594208, + 0.02473796159029007, + 0.02591126598417759, + 0.03788929060101509, + 0.01041287463158369, + 0.009785888716578484, + 0.009548918344080448, + 0.006954866461455822, + 0.020151738077402115, + 0.026887398213148117, + 0.025896470993757248, + 0.07318431884050369, + 0.01583004742860794, + 0.015869902446866035, + 0.018265260383486748, + 0.011074049398303032, + 0.027808763086795807, + 0.02361770160496235, + 0.043536994606256485, + 0.051088523119688034, + 0.023155726492404938, + 0.02287233993411064, + 0.02014675736427307, + 0.01494632288813591, + 0.01868986152112484, + 0.016805406659841537, + 0.028662893921136856, + 0.03513282537460327, + 0.01654985547065735, + 0.01040206104516983, + 0.01884179189801216, + 0.009071552194654942 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 7, + 9 + ], + "selected_reward": -0.0019427727675065398 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a steady medium shot taken from a level angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a female reporter with a neat dark brown ponytail wearing a royal blue jacket and a clip-on microphone visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a magnificent sand castle with soaring spires and intricate walls stand beside the reporter before collapsing?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the sand castle collapse rapidly and violently, flinging sand into the air to create a mist with visible light beams?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the reporter react with astonishment by taking a half step back and raising her arms to cover herself?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.009774206206202507 + ], + "opposite_logprob_tokens": [ + -4.634774208068848 + ], + "logprob_reward": -0.009774206206202507, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.009774206206202507, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06363347172737122, + 0.021886395290493965, + 0.012927046976983547, + 0.020581286400556564, + 0.021997738629579544, + 0.03568697348237038, + 0.033081211149692535, + 0.16163668036460876, + 0.10417849570512772, + 0.524390697479248 + ], + "credit_map": [ + 0.00860241986811161, + 0.005359167233109474, + 0.013420161791145802, + 0.00931867491453886, + 0.00709163723513484, + 0.007731160614639521, + 0.005835180636495352, + 0.006275066174566746, + 0.0029959497042000294, + 0.0022717993706464767, + 0.0034722236450761557, + 0.0029044405091553926, + 0.00248918985016644, + 0.002959273522719741, + 0.0023738951422274113, + 0.0024196237791329622, + 0.0017038228688761592, + 0.001683464040979743, + 0.0018855666276067495, + 0.0019581986125558615, + 0.0011910394532606006, + 0.0014132361393421888, + 0.0012691393494606018, + 0.0018225798849016428, + 0.003118277760222554, + 0.001914827385917306, + 0.003615793539211154, + 0.0029312388505786657, + 0.0020116225350648165, + 0.002646790351718664, + 0.0019528536358848214, + 0.0023898815270513296, + 0.0022300882264971733, + 0.00255391001701355, + 0.003019115887582302, + 0.004256729502230883, + 0.0028415087144821882, + 0.0023962811101228, + 0.0023877734784036875, + 0.0023123309947550297, + 0.004533490631729364, + 0.0027754579205065966, + 0.006875795777887106, + 0.0038003376685082912, + 0.005564174149185419, + 0.003784734755754471, + 0.004084883723407984, + 0.004268096759915352, + 0.00413659168407321, + 0.00362310535274446, + 0.007758387364447117, + 0.0038720325101166964, + 0.0028209739830344915, + 0.005501218140125275, + 0.0026992429047822952, + 0.0026696575805544853, + 0.01501122210174799, + 0.018544157966971397, + 0.025689788162708282, + 0.05052660405635834, + 0.01583697274327278, + 0.012668564915657043, + 0.01579786278307438, + 0.007561510428786278, + 0.014476160518825054, + 0.014168497174978256, + 0.015105052851140499, + 0.015131880529224873, + 0.01562599651515484, + 0.011103223077952862, + 0.011152351275086403, + 0.0074153258465230465, + 0.052149511873722076, + 0.05636263266205788, + 0.07986133545637131, + 0.07965469360351562, + 0.09617011994123459, + 0.04935303330421448, + 0.05815788730978966, + 0.05268150940537453 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 7, + 8 + ], + "selected_reward": -0.009774206206202507 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "steady medium shot, shot from a level angle, shows a female reporter with a neat dark brown ponytail standing on a vast stretch of sand. Wearing a windproof royal blue jacket and a small clip-on microphone clipped to her collar, she reports from the scene with professional demeanor. Beside her, a magnificent and intricate sand castle rises, its soaring spires, delicate arches, and intricately carved walls. The bright afternoon sun clearly defines its textures and shadows, while the shimmering azure sea looms in the background. Suddenly, as she calmly reports, the highest spire of the sand castle begins to tremble slightly. Then, as if suddenly losing its internal support, the entire castle collapses inward with rapidity and violence. Countless grains of dry sand are flung into the air, forming a pervasive mist. Sunlight penetrates this mist, creating beams of light reminiscent of the Tyndall effect, filling the entire scene with a sense of chaotic movement. A look of pure astonishment flashed across the female reporter's face. She subconsciously took a half step back and raised her arms to try to cover herself. This sudden collapse instantly transformed the originally peaceful reporting scene into an unexpected moment full of dramatic tension.", + "seed": 14143, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000071/group_12/index_6638_seed_14143_dp_25_rank_50_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 12, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.07034972254186869, + "vlm_logprob": -0.07034972254186869, + "vlm_yes_no_margin": 3.875, + "vlm_token_credit": -0.07034972254186869, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a steady medium shot taken from a level angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a female reporter with a neat dark brown ponytail wearing a royal blue jacket and a clip-on microphone visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a magnificent sand castle with soaring spires and intricate walls stand beside the reporter before collapsing?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the sand castle collapse rapidly and violently, flinging sand into the air to create a mist with visible light beams?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the reporter react with astonishment by taking a half step back and raising her arms to cover herself?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.1531940996646881 + ], + "opposite_logprob_tokens": [ + -4.653193950653076 + ], + "logprob_reward": -0.1531940996646881, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.1531940996646881, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.23461517691612244, + 0.10317429900169373, + 0.03901732712984085, + 0.039465975016355515, + 0.03418797627091408, + 0.04520754516124725, + 0.09578938782215118, + 0.07382294535636902, + 0.16417714953422546, + 0.1705421805381775 + ], + "credit_map": [ + 0.03579956293106079, + 0.048072364181280136, + 0.03045235015451908, + 0.03899654000997543, + 0.022816989570856094, + 0.01707598753273487, + 0.0224081389605999, + 0.01899324730038643, + 0.01448371447622776, + 0.023304810747504234, + 0.011541693471372128, + 0.013349255546927452, + 0.007080056704580784, + 0.007099304348230362, + 0.011006969958543777, + 0.015308489091694355, + 0.005656978581100702, + 0.004532678984105587, + 0.011364901438355446, + 0.004203066695481539, + 0.0031486551743000746, + 0.004247189499437809, + 0.0023464460391551256, + 0.0035174109507352114, + 0.0060048517771065235, + 0.0054937126114964485, + 0.002894600620493293, + 0.0035486482083797455, + 0.006499140989035368, + 0.005720298737287521, + 0.003960773814469576, + 0.005343945696949959, + 0.005217278376221657, + 0.004578319378197193, + 0.004301820881664753, + 0.0042021083645522594, + 0.005515930708497763, + 0.004831073340028524, + 0.002408815547823906, + 0.003132629906758666, + 0.005764641333371401, + 0.00824221596121788, + 0.004983759950846434, + 0.00625535286962986, + 0.005864758975803852, + 0.006129273679107428, + 0.003994353581219912, + 0.003973190672695637, + 0.016872532665729523, + 0.02734225057065487, + 0.011262739077210426, + 0.01607252098619938, + 0.009348796680569649, + 0.00707387225702405, + 0.004125007428228855, + 0.003691666992381215, + 0.010133711621165276, + 0.015686385333538055, + 0.013367881067097187, + 0.016909660771489143, + 0.006155865732580423, + 0.004589633550494909, + 0.0037233401089906693, + 0.003256465308368206, + 0.021655602380633354, + 0.020018091425299644, + 0.030115215107798576, + 0.0471356175839901, + 0.013426237739622593, + 0.01231992058455944, + 0.013533074408769608, + 0.0059733884409070015, + 0.018040576949715614, + 0.024718016386032104, + 0.019991064444184303, + 0.04248800873756409, + 0.01814102753996849, + 0.014731965027749538, + 0.018417254090309143, + 0.01401426550000906 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 8 + ], + "selected_reward": -0.1531940996646881 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a steady medium shot taken from a level angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a female reporter with a neat dark brown ponytail wearing a royal blue jacket and a clip-on microphone visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a magnificent sand castle with soaring spires and intricate walls stand beside the reporter before collapsing?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the sand castle collapse rapidly and violently, flinging sand into the air to create a mist with visible light beams?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the reporter react with astonishment by taking a half step back and raising her arms to cover herself?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.014226914383471012 + ], + "opposite_logprob_tokens": [ + -4.264226913452148 + ], + "logprob_reward": -0.014226914383471012, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.014226914383471012, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15225626528263092, + 0.32754114270210266, + 0.06374472379684448, + 0.019716765731573105, + 0.042001139372587204, + 0.020863760262727737, + 0.024942319840192795, + 0.04019561782479286, + 0.17908447980880737, + 0.12965375185012817 + ], + "credit_map": [ + 0.02378351241350174, + 0.023624362424016, + 0.021869290620088577, + 0.019234076142311096, + 0.021300677210092545, + 0.019190041348338127, + 0.009975130669772625, + 0.013279175385832787, + 0.0375426821410656, + 0.03280556946992874, + 0.03717048466205597, + 0.03203153610229492, + 0.027449389919638634, + 0.026130810379981995, + 0.054549701511859894, + 0.07986093312501907, + 0.00838306825608015, + 0.007808259222656488, + 0.011408356949687004, + 0.009025192819535732, + 0.006933125201612711, + 0.005769952200353146, + 0.006877906154841185, + 0.0075388685800135136, + 0.0026433239690959454, + 0.0028643549885600805, + 0.004368211608380079, + 0.002250377554446459, + 0.0016636373475193977, + 0.001725645735859871, + 0.0016936402535066009, + 0.002507574623450637, + 0.0066196508705616, + 0.004466456361114979, + 0.004120035097002983, + 0.004018051084131002, + 0.007357500959187746, + 0.007284937892109156, + 0.0034188157878816128, + 0.004715693648904562, + 0.0028089291881769896, + 0.0029259410221129656, + 0.002663363702595234, + 0.0033204422798007727, + 0.0028637389186769724, + 0.0028568946290761232, + 0.0015134175773710012, + 0.0019110323628410697, + 0.004008108284324408, + 0.0048431456089019775, + 0.003673700150102377, + 0.0043423473834991455, + 0.002617094898596406, + 0.002593317301943898, + 0.0015053384704515338, + 0.0013592676259577274, + 0.00661394651979208, + 0.00847721379250288, + 0.00794672966003418, + 0.0060898433439433575, + 0.004379004705697298, + 0.002529695862904191, + 0.0024373820051550865, + 0.00172180135268718, + 0.028384560719132423, + 0.023430174216628075, + 0.028585514053702354, + 0.04767956957221031, + 0.017544753849506378, + 0.01496156770735979, + 0.011608240194618702, + 0.006890109274536371, + 0.01859128475189209, + 0.020889097824692726, + 0.017535142600536346, + 0.033720482140779495, + 0.01124696247279644, + 0.009861079044640064, + 0.00929376482963562, + 0.008515945635735989 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 8, + 0 + ], + "selected_reward": -0.014226914383471012 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a steady medium shot taken from a level angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a female reporter with a neat dark brown ponytail wearing a royal blue jacket and a clip-on microphone visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a magnificent sand castle with soaring spires and intricate walls stand beside the reporter before collapsing?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the sand castle collapse rapidly and violently, flinging sand into the air to create a mist with visible light beams?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the reporter react with astonishment by taking a half step back and raising her arms to cover herself?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.014201059006154537 + ], + "opposite_logprob_tokens": [ + -4.2642011642456055 + ], + "logprob_reward": -0.014201059006154537, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.014201059006154537, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1550493687391281, + 0.06439267098903656, + 0.022990668192505836, + 0.01363993901759386, + 0.030047260224819183, + 0.018697211518883705, + 0.02680193819105625, + 0.06642063707113266, + 0.2885291278362274, + 0.3134312331676483 + ], + "credit_map": [ + 0.021892618387937546, + 0.03151172772049904, + 0.03900577500462532, + 0.02779935859143734, + 0.010810837149620056, + 0.009443586692214012, + 0.007118185982108116, + 0.00746726430952549, + 0.009602888487279415, + 0.014623407274484634, + 0.014653917402029037, + 0.010965561494231224, + 0.004045609384775162, + 0.002784763928502798, + 0.0034068028908222914, + 0.004309718031436205, + 0.0036298229824751616, + 0.003842269303277135, + 0.007294724229723215, + 0.002960996935144067, + 0.001555510563775897, + 0.0012840639101341367, + 0.0011672073742374778, + 0.0012560728937387466, + 0.0019829741213470697, + 0.002071729628369212, + 0.0026680806186050177, + 0.0015360650140792131, + 0.0014935086946934462, + 0.0015130996471270919, + 0.0010093600722029805, + 0.0013651209883391857, + 0.005021655932068825, + 0.004155977629125118, + 0.0076688723638653755, + 0.004252489656209946, + 0.003555850824341178, + 0.0019402917241677642, + 0.0015827777097001672, + 0.0018693400779739022, + 0.0028002532199025154, + 0.0028428169898688793, + 0.004067253787070513, + 0.0028459334280341864, + 0.002043298212811351, + 0.001489765476435423, + 0.0012128842063248158, + 0.001395004685036838, + 0.0044618044048547745, + 0.0033637136220932007, + 0.005530274007469416, + 0.004537642467767, + 0.002893993165344, + 0.0019280922133475542, + 0.0022307911422103643, + 0.0018556237919256091, + 0.009435879997909069, + 0.009721899405121803, + 0.014168685302138329, + 0.01600559800863266, + 0.0064896512776613235, + 0.004731596447527409, + 0.0037843568716198206, + 0.0020829662680625916, + 0.04365696385502815, + 0.039011821150779724, + 0.05265704542398453, + 0.08060476928949356, + 0.02827686071395874, + 0.020936984568834305, + 0.01503772847354412, + 0.008346944116055965, + 0.039757467806339264, + 0.050178010016679764, + 0.05733305588364601, + 0.07972995936870575, + 0.03460273519158363, + 0.023953402414917946, + 0.017829973250627518, + 0.010046613402664661 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 8, + 0 + ], + "selected_reward": -0.014201059006154537 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a steady medium shot taken from a level angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a female reporter with a neat dark brown ponytail wearing a royal blue jacket and a clip-on microphone visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a magnificent sand castle with soaring spires and intricate walls stand beside the reporter before collapsing?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the sand castle collapse rapidly and violently, flinging sand into the air to create a mist with visible light beams?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the reporter react with astonishment by taking a half step back and raising her arms to cover herself?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.00978341419249773 + ], + "opposite_logprob_tokens": [ + -4.6347832679748535 + ], + "logprob_reward": -0.00978341419249773, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.00978341419249773, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07582604885101318, + 0.03672351315617561, + 0.013592471368610859, + 0.01671365648508072, + 0.01745016686618328, + 0.011026344262063503, + 0.023870205506682396, + 0.035948626697063446, + 0.25025808811187744, + 0.5185908675193787 + ], + "credit_map": [ + 0.009824380278587341, + 0.014268150553107262, + 0.019859207794070244, + 0.011923310346901417, + 0.006297399289906025, + 0.0058586779050529, + 0.0036066402681171894, + 0.004188281949609518, + 0.005385174881666899, + 0.007829591631889343, + 0.008350924588739872, + 0.005753139965236187, + 0.00270407204516232, + 0.0019977320916950703, + 0.0021230888087302446, + 0.00257978867739439, + 0.0019854302518069744, + 0.0028359920252114534, + 0.0026383984368294477, + 0.002157312585040927, + 0.0010492991423234344, + 0.0011730939149856567, + 0.0008582797017879784, + 0.0008946641464717686, + 0.002219654619693756, + 0.002933328505605459, + 0.0022238637320697308, + 0.002638329053297639, + 0.0019045094959437847, + 0.001759598613716662, + 0.0012585049262270331, + 0.001775868353433907, + 0.0027128043584525585, + 0.003009447129443288, + 0.0025460286997258663, + 0.0025732354260981083, + 0.0022691101767122746, + 0.001715476275421679, + 0.0011266934452578425, + 0.001497371238656342, + 0.0016419519670307636, + 0.001779905753210187, + 0.0014934880891814828, + 0.0016171231400221586, + 0.0018521938472986221, + 0.0010833482956513762, + 0.0007939311908558011, + 0.0007644016877748072, + 0.003590668784454465, + 0.005126298870891333, + 0.004187623038887978, + 0.004260540474206209, + 0.0021284990943968296, + 0.0015970998210832477, + 0.0014714900171384215, + 0.001507985987700522, + 0.005409535486251116, + 0.007976552471518517, + 0.006452458444982767, + 0.006714537739753723, + 0.003545589977875352, + 0.002651285845786333, + 0.002017493126913905, + 0.001181171159259975, + 0.03834545612335205, + 0.03408695012331009, + 0.04490937292575836, + 0.06855769455432892, + 0.02466006763279438, + 0.0185822993516922, + 0.013515095226466656, + 0.0076011489145457745, + 0.07496098428964615, + 0.08560407906770706, + 0.09045641124248505, + 0.12770617008209229, + 0.058203358203172684, + 0.03594353049993515, + 0.029214834794402122, + 0.01650148630142212 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 8, + 0 + ], + "selected_reward": -0.00978341419249773 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a steady medium shot taken from a level angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a female reporter with a neat dark brown ponytail wearing a royal blue jacket and a clip-on microphone visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a magnificent sand castle with soaring spires and intricate walls stand beside the reporter before collapsing?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the sand castle collapse rapidly and violently, flinging sand into the air to create a mist with visible light beams?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the reporter react with astonishment by taking a half step back and raising her arms to cover herself?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.16034312546253204 + ], + "opposite_logprob_tokens": [ + -1.9103431701660156 + ], + "logprob_reward": -0.16034312546253204, + "yes_no_margin_reward": 1.75, + "token_credit_reward": -0.16034312546253204, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11938856542110443, + 0.10950622707605362, + 0.054564181715250015, + 0.03486945107579231, + 0.046625662595033646, + 0.03468259051442146, + 0.06126708164811134, + 0.05675441026687622, + 0.21997873485088348, + 0.26236310601234436 + ], + "credit_map": [ + 0.016828684136271477, + 0.014305989257991314, + 0.027295954525470734, + 0.01331630814820528, + 0.012499876320362091, + 0.016367996111512184, + 0.009217808954417706, + 0.009555939584970474, + 0.015480058267712593, + 0.008686245419085026, + 0.020565001294016838, + 0.012592869810760021, + 0.010475633665919304, + 0.00683682132512331, + 0.01324590016156435, + 0.021623697131872177, + 0.006837553810328245, + 0.0069367000833153725, + 0.009565898217260838, + 0.004994864575564861, + 0.006743970327079296, + 0.004638975486159325, + 0.006511830259114504, + 0.008334385231137276, + 0.005141595844179392, + 0.005362586583942175, + 0.004540710709989071, + 0.004645140841603279, + 0.005759993102401495, + 0.00399890448898077, + 0.002232320373877883, + 0.003188192145898938, + 0.005965978838503361, + 0.006475705653429031, + 0.010472170077264309, + 0.006931763142347336, + 0.0039029112085700035, + 0.004337461199611425, + 0.003989800810813904, + 0.004549866076558828, + 0.005261299666017294, + 0.005804784595966339, + 0.005306617356836796, + 0.0050416430458426476, + 0.004401777870953083, + 0.004361739382147789, + 0.0021259624045342207, + 0.0023787650279700756, + 0.008923194371163845, + 0.013797920197248459, + 0.010849859565496445, + 0.011464559473097324, + 0.005205583292990923, + 0.0047090626321733, + 0.0036078719422221184, + 0.0027090238872915506, + 0.00883080530911684, + 0.012078114785254002, + 0.008721450343728065, + 0.012592894956469536, + 0.004658444784581661, + 0.003626983379945159, + 0.003686648327857256, + 0.00255906512029469, + 0.03332507982850075, + 0.029704494401812553, + 0.03197000175714493, + 0.04624483361840248, + 0.024208085611462593, + 0.028396349400281906, + 0.015564308501780033, + 0.01056557148694992, + 0.037868887186050415, + 0.04267780855298042, + 0.04786393418908119, + 0.06518016010522842, + 0.02332836575806141, + 0.021614152938127518, + 0.013254490680992603, + 0.010575302876532078 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 8, + 0 + ], + "selected_reward": -0.16034312546253204 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000071/group_13/index_tensor([3425, 3425])_group_13_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000071/group_13/index_tensor([3425, 3425])_group_13_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..3e5374441b0414e7def1f3ba3a06189091bbe749 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000071/group_13/index_tensor([3425, 3425])_group_13_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "dynamic low-angle tracking shot captures two magnificent dragon boats racing side by side, capturing them neck and neck in a fierce competition. The entire scene comes to life in a vibrant, hand-drawn style reminiscent of high-energy animation. The camera moves swiftly yet steadily, keeping pace with the boats as they jockey for position in the churning, stylized water. The left-hand boat's hull is a saturated crimson, while its rival's is a vibrant imperial yellow, both adorned with meticulously rendered gold and green scales. Their fierce, sculpted dragon-headed prows, bared teeth and glaring eyes, seem to propel them forward with every powerful stroke of the oarsmen. Within each boat, a team of athletes, depicted in bold, expressive ink lines, moves in perfect and powerful sync, their muscles tense and faces set in grim determination. Their oars plunge deep into the azure water, creating explosions of white foam and vibrant speed lines that emphasize the raw power of their movements. The water itself is a turbulent blend of deep blue and green, its surface broken by dynamic, hand-drawn waves and spray. In the background, the sun-drenched riverbank, packed with cheering crowds, presents a colorful, impressionistic blur of movement and excitement. Bright, direct sunlight makes the bold, saturated colors of the boats and water burst, creating a high-contrast, visually dynamic atmosphere filled with the intense energy of the race.", + "seed": 8333, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000071/group_13/index_3425_seed_8333_dp_26_rank_52_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 13, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.03060174435377121, + "vlm_logprob": -0.03060174435377121, + "vlm_yes_no_margin": 3.75, + "vlm_token_credit": -0.03060174435377121, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are two dragon boats racing side by side in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the left-hand boat's hull a saturated crimson color?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the rival boat's hull a vibrant imperial yellow color?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera moving swiftly yet steadily to track the boats?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the visual style a vibrant, hand-drawn animation with bold ink lines?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.015913154929876328 + ], + "opposite_logprob_tokens": [ + -4.265913009643555 + ], + "logprob_reward": -0.015913154929876328, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.015913154929876328, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2753286361694336, + 0.05389617383480072, + 0.14606386423110962, + 0.039054032415151596, + 0.054550666362047195, + 0.09874283522367477, + 0.11500275135040283, + 0.11389120668172836, + 0.06596853584051132, + 0.0375012569129467 + ], + "credit_map": [ + 0.02204729989171028, + 0.08977528661489487, + 0.015796761959791183, + 0.02185896970331669, + 0.02574816718697548, + 0.027141835540533066, + 0.022101126611232758, + 0.05085921287536621, + 0.00659802695736289, + 0.006290555465966463, + 0.004007128067314625, + 0.004537447355687618, + 0.010850563645362854, + 0.0072939167730510235, + 0.006827652454376221, + 0.0074908859096467495, + 0.018753763288259506, + 0.019170479848980904, + 0.021893296390771866, + 0.016593001782894135, + 0.014480876736342907, + 0.016886664554476738, + 0.01259908638894558, + 0.02568671479821205, + 0.0054438114166259766, + 0.01027686521410942, + 0.002436371287330985, + 0.0051876250654459, + 0.0036077669356018305, + 0.003448022762313485, + 0.003975693602114916, + 0.004677880555391312, + 0.007735010236501694, + 0.007926314137876034, + 0.009220484644174576, + 0.005819236859679222, + 0.007962798699736595, + 0.0064993612468242645, + 0.004327060654759407, + 0.005060406401753426, + 0.0065192123875021935, + 0.03818560391664505, + 0.003935776650905609, + 0.004799656104296446, + 0.008424091152846813, + 0.026646926999092102, + 0.004017512314021587, + 0.00621406314894557, + 0.011034821160137653, + 0.010755972005426884, + 0.013421833515167236, + 0.015175482258200645, + 0.011294646188616753, + 0.01588284596800804, + 0.021253997460007668, + 0.016183165833353996, + 0.008098499849438667, + 0.016051655635237694, + 0.005122816655784845, + 0.007690331432968378, + 0.008711379021406174, + 0.03442257642745972, + 0.015416573733091354, + 0.018377387896180153, + 0.01820223033428192, + 0.005545430351048708, + 0.003776607569307089, + 0.003857380012050271, + 0.0038962254766374826, + 0.008200247772037983, + 0.004354908596724272, + 0.018135512247681618, + 0.008040097542107105, + 0.01227191649377346, + 0.002324033295735717, + 0.002389113651588559, + 0.0032429706770926714, + 0.003302875440567732, + 0.002589714713394642, + 0.0033405404537916183 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 6 + ], + "selected_reward": -0.015913154929876328 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are two dragon boats racing side by side in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the left-hand boat's hull a saturated crimson color?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the rival boat's hull a vibrant imperial yellow color?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera moving swiftly yet steadily to track the boats?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the visual style a vibrant, hand-drawn animation with bold ink lines?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.05496727302670479 + ], + "opposite_logprob_tokens": [ + -2.929967164993286 + ], + "logprob_reward": -0.05496727302670479, + "yes_no_margin_reward": 2.875, + "token_credit_reward": -0.05496727302670479, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14795400202274323, + 0.11105433851480484, + 0.10712265223264694, + 0.07462134212255478, + 0.04996025934815407, + 0.0531306155025959, + 0.02199716493487358, + 0.24694722890853882, + 0.09932070225477219, + 0.08789171278476715 + ], + "credit_map": [ + 0.012977022677659988, + 0.02887747995555401, + 0.01634027063846588, + 0.019256656989455223, + 0.014762169681489468, + 0.016979176551103592, + 0.024137310683727264, + 0.014623921364545822, + 0.01216178759932518, + 0.011124981567263603, + 0.007995525375008583, + 0.00913996808230877, + 0.026384489610791206, + 0.020609602332115173, + 0.013376081362366676, + 0.010261907242238522, + 0.013726516626775265, + 0.013566335663199425, + 0.01568611152470112, + 0.011996316723525524, + 0.012239700183272362, + 0.015618166886270046, + 0.011761393398046494, + 0.012528110295534134, + 0.01173052005469799, + 0.012502356432378292, + 0.006281577982008457, + 0.005868807435035706, + 0.012833033688366413, + 0.010926675982773304, + 0.006542464718222618, + 0.007935904897749424, + 0.005960559938102961, + 0.0068734330125153065, + 0.006804257165640593, + 0.004607719369232655, + 0.010746844112873077, + 0.0058382670395076275, + 0.003867692779749632, + 0.005261484067887068, + 0.0028971072752028704, + 0.018910501152276993, + 0.002551046898588538, + 0.0018572253175079823, + 0.003899856237694621, + 0.017468491569161415, + 0.0020982222631573677, + 0.0034481666516512632, + 0.0027329917065799236, + 0.0026318798772990704, + 0.002130210632458329, + 0.002494388958439231, + 0.0023545913863927126, + 0.00486746896058321, + 0.0015155113069340587, + 0.003270122455433011, + 0.01563982479274273, + 0.03575483709573746, + 0.010859361849725246, + 0.013840919360518456, + 0.016573334112763405, + 0.0899767056107521, + 0.022385360673069954, + 0.04191688075661659, + 0.03229391202330589, + 0.022986583411693573, + 0.0055866967886686325, + 0.006171980407088995, + 0.005189140792936087, + 0.009541950188577175, + 0.0041069937869906425, + 0.013443444855511189, + 0.026593849062919617, + 0.021130813285708427, + 0.007957404479384422, + 0.0066564870066940784, + 0.004750818945467472, + 0.005973948165774345, + 0.005857378244400024, + 0.008971012197434902 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 0, + 1 + ], + "selected_reward": -0.05496727302670479 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are two dragon boats racing side by side in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the left-hand boat's hull a saturated crimson color?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the rival boat's hull a vibrant imperial yellow color?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera moving swiftly yet steadily to track the boats?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the visual style a vibrant, hand-drawn animation with bold ink lines?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03809810429811478 + ], + "opposite_logprob_tokens": [ + -3.288098096847534 + ], + "logprob_reward": -0.03809810429811478, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.03809810429811478, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.17393901944160461, + 0.142033651471138, + 0.32449930906295776, + 0.07421459257602692, + 0.04474997520446777, + 0.0477701872587204, + 0.023791693150997162, + 0.03921745717525482, + 0.0631680116057396, + 0.06661605834960938 + ], + "credit_map": [ + 0.014484331011772156, + 0.058155614882707596, + 0.01274531614035368, + 0.013191708363592625, + 0.015769504010677338, + 0.017576906830072403, + 0.01642356626689434, + 0.02559206821024418, + 0.015133512206375599, + 0.01483754999935627, + 0.013511956669390202, + 0.01179358921945095, + 0.034990083426237106, + 0.021881479769945145, + 0.01697356440126896, + 0.012911910191178322, + 0.040501441806554794, + 0.03714548796415329, + 0.03451431915163994, + 0.031152335926890373, + 0.03763888403773308, + 0.04189432039856911, + 0.03393075615167618, + 0.06772179156541824, + 0.010912716388702393, + 0.012343665584921837, + 0.006135535892099142, + 0.005935020744800568, + 0.010335405357182026, + 0.013425451703369617, + 0.0064922901801764965, + 0.008634510450065136, + 0.005091211758553982, + 0.005670078564435244, + 0.0027440472040325403, + 0.0029900202061980963, + 0.0151152228936553, + 0.005138508975505829, + 0.00424331845715642, + 0.0037575680762529373, + 0.0035767785739153624, + 0.013727187179028988, + 0.003344009630382061, + 0.002714816015213728, + 0.004957971163094044, + 0.012396169826388359, + 0.002903454238548875, + 0.004149796906858683, + 0.0027800886891782284, + 0.003086348529905081, + 0.002884314628317952, + 0.0024320913944393396, + 0.0020576247479766607, + 0.00457425182685256, + 0.002586304908618331, + 0.003390668425709009, + 0.003915765788406134, + 0.007738794665783644, + 0.0024407070595771074, + 0.004621167667210102, + 0.0036467609461396933, + 0.007322491146624088, + 0.004223460331559181, + 0.005308311898261309, + 0.020611055195331573, + 0.010079778730869293, + 0.002837572945281863, + 0.0028396209236234426, + 0.0033582996111363173, + 0.0067783924750983715, + 0.0033049241174012423, + 0.013358366675674915, + 0.013564515858888626, + 0.02340092882514, + 0.005482373293489218, + 0.005870240740478039, + 0.003935145679861307, + 0.005151822231709957, + 0.0047666532918810844, + 0.004444381222128868 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 0, + 1 + ], + "selected_reward": -0.03809810429811478 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are two dragon boats racing side by side in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the left-hand boat's hull a saturated crimson color?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the rival boat's hull a vibrant imperial yellow color?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera moving swiftly yet steadily to track the boats?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the visual style a vibrant, hand-drawn animation with bold ink lines?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03808777406811714 + ], + "opposite_logprob_tokens": [ + -3.288087844848633 + ], + "logprob_reward": -0.03808777406811714, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.03808777406811714, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.5604124665260315, + 0.0811714380979538, + 0.09705346822738647, + 0.03892308846116066, + 0.042217936366796494, + 0.04154644533991814, + 0.010940725915133953, + 0.02743581496179104, + 0.06547999382019043, + 0.03481866419315338 + ], + "credit_map": [ + 0.04798382893204689, + 0.22274714708328247, + 0.01637587510049343, + 0.028752872720360756, + 0.05038601905107498, + 0.0478673055768013, + 0.027587756514549255, + 0.11871160566806793, + 0.007797101512551308, + 0.008104216307401657, + 0.007664680015295744, + 0.006737468298524618, + 0.017793679609894753, + 0.015889378264546394, + 0.00995495356619358, + 0.007229955866932869, + 0.01023419201374054, + 0.011903746984899044, + 0.011411365121603012, + 0.015554135665297508, + 0.010908115655183792, + 0.010178747586905956, + 0.011455181054770947, + 0.015407973900437355, + 0.0048695942386984825, + 0.005897520110011101, + 0.0033717805054038763, + 0.0047955880872905254, + 0.00624821474775672, + 0.005070067010819912, + 0.003699746448546648, + 0.004970571491867304, + 0.003298546653240919, + 0.005200068466365337, + 0.002301333472132683, + 0.0025454869028180838, + 0.01494162529706955, + 0.006592877209186554, + 0.004128491971641779, + 0.003209502901881933, + 0.0032645417377352715, + 0.015492884442210197, + 0.002034880220890045, + 0.001753434189595282, + 0.0030083884485065937, + 0.01088270265609026, + 0.0023279415909200907, + 0.0027816700749099255, + 0.001166178728453815, + 0.0010918775806203485, + 0.0013100032228976488, + 0.0015518047148361802, + 0.0011343812802806497, + 0.0014658861327916384, + 0.0015992504777386785, + 0.0016213428461924195, + 0.0023270072415471077, + 0.004050051793456078, + 0.0014958862448111176, + 0.0019773845560848713, + 0.0016950966091826558, + 0.008564187213778496, + 0.002559133805334568, + 0.0047670649364590645, + 0.01434949692338705, + 0.00575061934068799, + 0.0030159575399011374, + 0.0026916582137346268, + 0.0035784451756626368, + 0.011197964660823345, + 0.00529694277793169, + 0.019598903134465218, + 0.008379204198718071, + 0.013122454285621643, + 0.0013529440620914102, + 0.0018198012840002775, + 0.002128192689269781, + 0.002863573841750622, + 0.0026081956457346678, + 0.0025442917831242085 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 1 + ], + "selected_reward": -0.03808777406811714 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are two dragon boats racing side by side in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the left-hand boat's hull a saturated crimson color?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the rival boat's hull a vibrant imperial yellow color?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera moving swiftly yet steadily to track the boats?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the visual style a vibrant, hand-drawn animation with bold ink lines?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0059424154460430145 + ], + "opposite_logprob_tokens": [ + -5.130942344665527 + ], + "logprob_reward": -0.0059424154460430145, + "yes_no_margin_reward": 5.125, + "token_credit_reward": -0.0059424154460430145, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09862630069255829, + 0.12604279816150665, + 0.38415542244911194, + 0.10542642325162888, + 0.05573679879307747, + 0.05007360130548477, + 0.03811503201723099, + 0.01664569601416588, + 0.10256246477365494, + 0.022615429013967514 + ], + "credit_map": [ + 0.007152625359594822, + 0.02655031345784664, + 0.007196103222668171, + 0.007335649337619543, + 0.008171898312866688, + 0.010445629246532917, + 0.00810952391475439, + 0.0236645620316267, + 0.016630034893751144, + 0.011842709966003895, + 0.008081352338194847, + 0.009297405369579792, + 0.02712368033826351, + 0.027305137366056442, + 0.013281157240271568, + 0.012481327168643475, + 0.04192040488123894, + 0.03830558806657791, + 0.04009281098842621, + 0.05459161847829819, + 0.05090220645070076, + 0.051973290741443634, + 0.04221361130475998, + 0.06415589898824692, + 0.0172125156968832, + 0.020089466124773026, + 0.007579233031719923, + 0.006556226871907711, + 0.021235806867480278, + 0.012794925831258297, + 0.009663213044404984, + 0.010295029729604721, + 0.004879385232925415, + 0.007101273629814386, + 0.004553542472422123, + 0.0038382462225854397, + 0.016334762796759605, + 0.009423873387277126, + 0.004559595137834549, + 0.005046122707426548, + 0.0024426819290965796, + 0.018685724586248398, + 0.0013579620281234384, + 0.0016429764218628407, + 0.005074809305369854, + 0.016814637929201126, + 0.002116258256137371, + 0.0019385480554774404, + 0.003941575065255165, + 0.004839523229748011, + 0.0045197345316410065, + 0.003920168150216341, + 0.0030907215550541878, + 0.007523525971919298, + 0.003576042130589485, + 0.006703742779791355, + 0.0016754635144025087, + 0.0026993481442332268, + 0.0017462620744481683, + 0.002674182178452611, + 0.0016930245328694582, + 0.001673222635872662, + 0.002319199964404106, + 0.002164992969483137, + 0.015317060984671116, + 0.014574750326573849, + 0.008040759712457657, + 0.011306030675768852, + 0.008437631651759148, + 0.016108792275190353, + 0.013283815234899521, + 0.015493624843657017, + 0.005620195530354977, + 0.00818728283047676, + 0.001243472215719521, + 0.0011688239173963666, + 0.0015251950826495886, + 0.0017244401387870312, + 0.0015741238603368402, + 0.0015718959039077163 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 1, + 3 + ], + "selected_reward": -0.0059424154460430145 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "dynamic low-angle tracking shot captures two magnificent dragon boats racing side by side, capturing them neck and neck in a fierce competition. The entire scene comes to life in a vibrant, hand-drawn style reminiscent of high-energy animation. The camera moves swiftly yet steadily, keeping pace with the boats as they jockey for position in the churning, stylized water. The left-hand boat's hull is a saturated crimson, while its rival's is a vibrant imperial yellow, both adorned with meticulously rendered gold and green scales. Their fierce, sculpted dragon-headed prows, bared teeth and glaring eyes, seem to propel them forward with every powerful stroke of the oarsmen. Within each boat, a team of athletes, depicted in bold, expressive ink lines, moves in perfect and powerful sync, their muscles tense and faces set in grim determination. Their oars plunge deep into the azure water, creating explosions of white foam and vibrant speed lines that emphasize the raw power of their movements. The water itself is a turbulent blend of deep blue and green, its surface broken by dynamic, hand-drawn waves and spray. In the background, the sun-drenched riverbank, packed with cheering crowds, presents a colorful, impressionistic blur of movement and excitement. Bright, direct sunlight makes the bold, saturated colors of the boats and water burst, creating a high-contrast, visually dynamic atmosphere filled with the intense energy of the race.", + "seed": 8334, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000071/group_13/index_3425_seed_8334_dp_26_rank_52_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 13, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.029113611113280057, + "vlm_logprob": -0.029113611113280057, + "vlm_yes_no_margin": 3.85, + "vlm_token_credit": -0.029113611113280057, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are two dragon boats racing side by side in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the left-hand boat's hull a saturated crimson color?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the rival boat's hull a vibrant imperial yellow color?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera moving swiftly yet steadily to track the boats?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the visual style a vibrant, hand-drawn animation with bold ink lines?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01012569759041071 + ], + "opposite_logprob_tokens": [ + -4.885125637054443 + ], + "logprob_reward": -0.01012569759041071, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.01012569759041071, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14989140629768372, + 0.26558563113212585, + 0.03442777693271637, + 0.028150487691164017, + 0.07683870196342468, + 0.026137635111808777, + 0.20264257490634918, + 0.06352763622999191, + 0.07458724081516266, + 0.07821088284254074 + ], + "credit_map": [ + 0.018242135643959045, + 0.02496412768959999, + 0.017703954130411148, + 0.015811126679182053, + 0.02237371727824211, + 0.018086083233356476, + 0.01720358617603779, + 0.015506686642765999, + 0.017905378714203835, + 0.018613873049616814, + 0.04474152252078056, + 0.07666491717100143, + 0.021523965522646904, + 0.05831397324800491, + 0.014715690165758133, + 0.013106315396726131, + 0.0030810364987701178, + 0.0040840026922523975, + 0.0048956358805298805, + 0.004731378983706236, + 0.003425753442570567, + 0.004074169788509607, + 0.0056367674842476845, + 0.0044990344904363155, + 0.003720876295119524, + 0.003257238771766424, + 0.002027877140790224, + 0.004624166991561651, + 0.0029588581528514624, + 0.003194296732544899, + 0.004343834239989519, + 0.004023341462016106, + 0.01149908546358347, + 0.008438872173428535, + 0.007392219267785549, + 0.005538577679544687, + 0.010957461781799793, + 0.011508801952004433, + 0.010830734856426716, + 0.010672954842448235, + 0.0019047469832003117, + 0.003446006216108799, + 0.00386648322455585, + 0.0027356818318367004, + 0.004268353804945946, + 0.0025174012407660484, + 0.0038017795886844397, + 0.0035971838515251875, + 0.017876535654067993, + 0.028402188792824745, + 0.02056463062763214, + 0.019437791779637337, + 0.03951748088002205, + 0.030134446918964386, + 0.026440631598234177, + 0.020268883556127548, + 0.008693577721714973, + 0.006113979499787092, + 0.007468755356967449, + 0.007321270182728767, + 0.0069436123594641685, + 0.008905375376343727, + 0.009960959665477276, + 0.008120107464492321, + 0.006380326580256224, + 0.005351313389837742, + 0.007142314221709967, + 0.011550486087799072, + 0.007674276363104582, + 0.008659940212965012, + 0.0151410773396492, + 0.012687508016824722, + 0.007553498260676861, + 0.011457582004368305, + 0.006085183471441269, + 0.006543065886944532, + 0.013803980313241482, + 0.011522021144628525, + 0.010835213586688042, + 0.01041034422814846 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 6, + 0 + ], + "selected_reward": -0.01012569759041071 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are two dragon boats racing side by side in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the left-hand boat's hull a saturated crimson color?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the rival boat's hull a vibrant imperial yellow color?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera moving swiftly yet steadily to track the boats?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the visual style a vibrant, hand-drawn animation with bold ink lines?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04307219386100769 + ], + "opposite_logprob_tokens": [ + -3.16807222366333 + ], + "logprob_reward": -0.04307219386100769, + "yes_no_margin_reward": 3.125, + "token_credit_reward": -0.04307219386100769, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09685900062322617, + 0.1957322657108307, + 0.03855486214160919, + 0.0416528582572937, + 0.07564821094274521, + 0.0335284061729908, + 0.22030112147331238, + 0.08262843638658524, + 0.1165923997759819, + 0.09850247204303741 + ], + "credit_map": [ + 0.013726233504712582, + 0.013056174851953983, + 0.012605143710970879, + 0.008065979927778244, + 0.01178586483001709, + 0.014277900569140911, + 0.010528228245675564, + 0.01281347032636404, + 0.012653329409658909, + 0.01389844436198473, + 0.0335528589785099, + 0.057642437517642975, + 0.015619054436683655, + 0.04536054655909538, + 0.00998096913099289, + 0.007024615071713924, + 0.004107278306037188, + 0.005215302109718323, + 0.004463609773665667, + 0.004243879579007626, + 0.006621938198804855, + 0.004418566357344389, + 0.005038048140704632, + 0.004446238279342651, + 0.006242158357053995, + 0.004891638178378344, + 0.005258871708065271, + 0.006290493533015251, + 0.003836223157122731, + 0.004703401122242212, + 0.0054869260638952255, + 0.004943144507706165, + 0.00849775318056345, + 0.010855386033654213, + 0.009408297948539257, + 0.008710905909538269, + 0.00982873234897852, + 0.008240466006100178, + 0.009816893376410007, + 0.010289782658219337, + 0.003576673101633787, + 0.0052233487367630005, + 0.00452894764021039, + 0.003012788947671652, + 0.005809555295854807, + 0.003948761150240898, + 0.0029153015930205584, + 0.004513029009103775, + 0.023694481700658798, + 0.021798893809318542, + 0.033441923558712006, + 0.035626381635665894, + 0.022109296172857285, + 0.026211537420749664, + 0.02784564718604088, + 0.02957296371459961, + 0.011916263028979301, + 0.011309484951198101, + 0.009173691272735596, + 0.008020853623747826, + 0.011314939707517624, + 0.010757872834801674, + 0.01103050448000431, + 0.009104822762310505, + 0.012126335874199867, + 0.009091528132557869, + 0.016536490991711617, + 0.026503536850214005, + 0.010104836896061897, + 0.012954295612871647, + 0.014411806128919125, + 0.014863571152091026, + 0.008913715369999409, + 0.012259624898433685, + 0.011628939770162106, + 0.012875113636255264, + 0.013254942372441292, + 0.013603280298411846, + 0.014448379166424274, + 0.011518476530909538 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 1, + 8 + ], + "selected_reward": -0.04307219386100769 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are two dragon boats racing side by side in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the left-hand boat's hull a saturated crimson color?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the rival boat's hull a vibrant imperial yellow color?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera moving swiftly yet steadily to track the boats?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the visual style a vibrant, hand-drawn animation with bold ink lines?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.029810817912220955 + ], + "opposite_logprob_tokens": [ + -3.529810905456543 + ], + "logprob_reward": -0.029810817912220955, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.029810817912220955, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08892007172107697, + 0.49370867013931274, + 0.029457757249474525, + 0.029039649292826653, + 0.10259996354579926, + 0.03884807601571083, + 0.08802910149097443, + 0.05257350951433182, + 0.033146653324365616, + 0.04367656633257866 + ], + "credit_map": [ + 0.009885203093290329, + 0.014215035364031792, + 0.011166075244545937, + 0.01144943293184042, + 0.011828039772808552, + 0.011777741834521294, + 0.010777332819998264, + 0.007821213454008102, + 0.0318957157433033, + 0.0347149483859539, + 0.08871646225452423, + 0.1505490094423294, + 0.03920635208487511, + 0.11037873476743698, + 0.0238152127712965, + 0.014432242140173912, + 0.002262700814753771, + 0.0028148945420980453, + 0.003202722640708089, + 0.005330786108970642, + 0.002429405227303505, + 0.002587414812296629, + 0.0052286419086158276, + 0.00560119329020381, + 0.004225991200655699, + 0.003247651504352689, + 0.002611310686916113, + 0.004368986468762159, + 0.0031633067410439253, + 0.003706080839037895, + 0.003837222931906581, + 0.0038791003171354532, + 0.00845287460833788, + 0.010649415664374828, + 0.012281612493097782, + 0.009679937735199928, + 0.01741795241832733, + 0.01438793446868658, + 0.014145612716674805, + 0.015584633685648441, + 0.004423327278345823, + 0.0059445383958518505, + 0.0026868386194109917, + 0.002471794141456485, + 0.00823749415576458, + 0.0061518000438809395, + 0.0035377240274101496, + 0.005394566338509321, + 0.009416256099939346, + 0.015054602175951004, + 0.00681461114436388, + 0.006964826490730047, + 0.01938239298760891, + 0.01333540678024292, + 0.008279445581138134, + 0.008781556971371174, + 0.0070581394247710705, + 0.006388356909155846, + 0.005957321263849735, + 0.0059433747082948685, + 0.0071006580255925655, + 0.006769982166588306, + 0.007518197875469923, + 0.005837481003254652, + 0.003356205066666007, + 0.0025627536233514547, + 0.004637658130377531, + 0.007162539288401604, + 0.0022951140999794006, + 0.004068783018738031, + 0.004596571438014507, + 0.004467029590159655, + 0.004254171624779701, + 0.005323657300323248, + 0.004896906204521656, + 0.0053476979956030846, + 0.005061089061200619, + 0.006207547150552273, + 0.007072338834404945, + 0.005513162352144718 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 4, + 0 + ], + "selected_reward": -0.029810817912220955 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are two dragon boats racing side by side in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the left-hand boat's hull a saturated crimson color?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the rival boat's hull a vibrant imperial yellow color?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera moving swiftly yet steadily to track the boats?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the visual style a vibrant, hand-drawn animation with bold ink lines?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.05493725836277008 + ], + "opposite_logprob_tokens": [ + -2.9299373626708984 + ], + "logprob_reward": -0.05493725836277008, + "yes_no_margin_reward": 2.875, + "token_credit_reward": -0.05493725836277008, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13367295265197754, + 0.15477430820465088, + 0.09498155117034912, + 0.04526272043585777, + 0.10619974881410599, + 0.0851719006896019, + 0.1191888302564621, + 0.17618398368358612, + 0.038625773042440414, + 0.04593818262219429 + ], + "credit_map": [ + 0.018738990649580956, + 0.01906649023294449, + 0.016226455569267273, + 0.011240450665354729, + 0.017263270914554596, + 0.017146000638604164, + 0.017092246562242508, + 0.01689906232059002, + 0.009590881876647472, + 0.011707414872944355, + 0.025998011231422424, + 0.04333677142858505, + 0.012613323517143726, + 0.035881903022527695, + 0.007927675731480122, + 0.00771833723410964, + 0.010948729701340199, + 0.01021510735154152, + 0.010499388910830021, + 0.013419841416180134, + 0.012127112597227097, + 0.014173789881169796, + 0.01381576806306839, + 0.00978181604295969, + 0.00878274254500866, + 0.0040537696331739426, + 0.004748825449496508, + 0.007347229868173599, + 0.003338204463943839, + 0.005573061294853687, + 0.005814236123114824, + 0.005604654084891081, + 0.008454362861812115, + 0.008579290471971035, + 0.023767981678247452, + 0.019316498190164566, + 0.00811485294252634, + 0.010449220426380634, + 0.011575642973184586, + 0.01594190113246441, + 0.01377034280449152, + 0.014422844164073467, + 0.006059154402464628, + 0.0065261851996183395, + 0.013949243351817131, + 0.0097079211845994, + 0.010166643187403679, + 0.010569578967988491, + 0.011073878034949303, + 0.02234143391251564, + 0.010212103836238384, + 0.013683110475540161, + 0.023473026230931282, + 0.014143574982881546, + 0.011936796829104424, + 0.012324913404881954, + 0.018418265506625175, + 0.019208256155252457, + 0.01832536607980728, + 0.01767093501985073, + 0.026879578828811646, + 0.021067146211862564, + 0.032650090754032135, + 0.021964356303215027, + 0.0032165953889489174, + 0.0029824767261743546, + 0.0038629984483122826, + 0.007705233991146088, + 0.0034145463723689318, + 0.0039936695247888565, + 0.00731287058442831, + 0.0061373841017484665, + 0.005253439303487539, + 0.0058687422424554825, + 0.005228268448263407, + 0.005437884945422411, + 0.007533705793321133, + 0.006599149666726589, + 0.004724453669041395, + 0.005292545072734356 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 1, + 0 + ], + "selected_reward": -0.05493725836277008 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are two dragon boats racing side by side in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the left-hand boat's hull a saturated crimson color?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the rival boat's hull a vibrant imperial yellow color?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera moving swiftly yet steadily to track the boats?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the visual style a vibrant, hand-drawn animation with bold ink lines?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.007622087839990854 + ], + "opposite_logprob_tokens": [ + -4.882622241973877 + ], + "logprob_reward": -0.007622087839990854, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.007622087839990854, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12213604152202606, + 0.25135523080825806, + 0.0923214703798294, + 0.021386252716183662, + 0.07083411514759064, + 0.02146424911916256, + 0.214945450425148, + 0.0879441648721695, + 0.0279936995357275, + 0.08961941301822662 + ], + "credit_map": [ + 0.020447788760066032, + 0.019801199436187744, + 0.010066328570246696, + 0.010789486579596996, + 0.01933836005628109, + 0.018012361600995064, + 0.011273667216300964, + 0.01240683812648058, + 0.01547479908913374, + 0.018055548891425133, + 0.044943176209926605, + 0.07630268484354019, + 0.019630925729870796, + 0.0574178472161293, + 0.011957305483520031, + 0.007572903297841549, + 0.008602669462561607, + 0.011443606577813625, + 0.011332445777952671, + 0.009665300138294697, + 0.011354281567037106, + 0.010893771424889565, + 0.015689799562096596, + 0.01333958376199007, + 0.002176266862079501, + 0.003532882547006011, + 0.0020620885770767927, + 0.002280439483001828, + 0.0020583656150847673, + 0.004214719403535128, + 0.0024959032889455557, + 0.002565583446994424, + 0.006459190510213375, + 0.00495548453181982, + 0.014791768975555897, + 0.015403599478304386, + 0.004112029913812876, + 0.006930150091648102, + 0.007970917969942093, + 0.01021096482872963, + 0.00170851347502321, + 0.0033917331602424383, + 0.001510310685262084, + 0.0015730608720332384, + 0.004597348161041737, + 0.003090009093284607, + 0.002890445524826646, + 0.00270282500423491, + 0.022910289466381073, + 0.033282727003097534, + 0.01843329332768917, + 0.022688454017043114, + 0.04393351823091507, + 0.026416469365358353, + 0.023875420913100243, + 0.023405255749821663, + 0.01204290147870779, + 0.00848256517201662, + 0.010059843771159649, + 0.01171113457530737, + 0.011461620219051838, + 0.010067441500723362, + 0.013215390965342522, + 0.01090325228869915, + 0.002629699418321252, + 0.0031532710418105125, + 0.0020134798251092434, + 0.0029858634807169437, + 0.006047774106264114, + 0.004546253476291895, + 0.0034600135404616594, + 0.003157340921461582, + 0.009130078367888927, + 0.011915977112948895, + 0.006661982275545597, + 0.006475646514445543, + 0.014996941201388836, + 0.011238962411880493, + 0.016104882583022118, + 0.013094933703541756 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 6, + 0 + ], + "selected_reward": -0.007622087839990854 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "dynamic low-angle tracking shot captures two magnificent dragon boats racing side by side, capturing them neck and neck in a fierce competition. The entire scene comes to life in a vibrant, hand-drawn style reminiscent of high-energy animation. The camera moves swiftly yet steadily, keeping pace with the boats as they jockey for position in the churning, stylized water. The left-hand boat's hull is a saturated crimson, while its rival's is a vibrant imperial yellow, both adorned with meticulously rendered gold and green scales. Their fierce, sculpted dragon-headed prows, bared teeth and glaring eyes, seem to propel them forward with every powerful stroke of the oarsmen. Within each boat, a team of athletes, depicted in bold, expressive ink lines, moves in perfect and powerful sync, their muscles tense and faces set in grim determination. Their oars plunge deep into the azure water, creating explosions of white foam and vibrant speed lines that emphasize the raw power of their movements. The water itself is a turbulent blend of deep blue and green, its surface broken by dynamic, hand-drawn waves and spray. In the background, the sun-drenched riverbank, packed with cheering crowds, presents a colorful, impressionistic blur of movement and excitement. Bright, direct sunlight makes the bold, saturated colors of the boats and water burst, creating a high-contrast, visually dynamic atmosphere filled with the intense energy of the race.", + "seed": 8335, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000071/group_13/index_3425_seed_8335_dp_27_rank_54_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 13, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.06282691797241569, + "vlm_logprob": -0.06282691797241569, + "vlm_yes_no_margin": 3.25, + "vlm_token_credit": -0.06282691797241569, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are two dragon boats racing side by side in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the left-hand boat's hull a saturated crimson color?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the rival boat's hull a vibrant imperial yellow color?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera moving swiftly yet steadily to track the boats?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the visual style a vibrant, hand-drawn animation with bold ink lines?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.023412451148033142 + ], + "opposite_logprob_tokens": [ + -3.8984124660491943 + ], + "logprob_reward": -0.023412451148033142, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.023412451148033142, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13278181850910187, + 0.08731653541326523, + 0.04397130757570267, + 0.09476231783628464, + 0.11696404218673706, + 0.15119822323322296, + 0.11520370095968246, + 0.09561027586460114, + 0.1243448331952095, + 0.037846941500902176 + ], + "credit_map": [ + 0.011241558007895947, + 0.016372626647353172, + 0.026294441893696785, + 0.009749841876327991, + 0.03293471038341522, + 0.014311393722891808, + 0.007377321831882, + 0.014499927870929241, + 0.016599273309111595, + 0.012322932481765747, + 0.011783760040998459, + 0.00725160026922822, + 0.017803024500608444, + 0.010046783834695816, + 0.006011339835822582, + 0.005497821606695652, + 0.004506000317633152, + 0.005742247682064772, + 0.0045946091413497925, + 0.004541684407740831, + 0.008246475830674171, + 0.007203736808151007, + 0.0036342027597129345, + 0.005502352025359869, + 0.02566220983862877, + 0.005323741119354963, + 0.03459782898426056, + 0.006802375894039869, + 0.006437285337597132, + 0.0058418274857103825, + 0.005397636443376541, + 0.004699413664638996, + 0.011934127658605576, + 0.009703338146209717, + 0.013470351696014404, + 0.020987261086702347, + 0.022489363327622414, + 0.013934672810137272, + 0.012721612118184566, + 0.011723308824002743, + 0.015048932284116745, + 0.006684363353997469, + 0.04940468445420265, + 0.010228106752038002, + 0.047254350036382675, + 0.01031390018761158, + 0.006837871856987476, + 0.0054260133765637875, + 0.013030854053795338, + 0.02919209934771061, + 0.027018405497074127, + 0.00866315234452486, + 0.010849833488464355, + 0.010370288975536823, + 0.006801367737352848, + 0.00927770882844925, + 0.008747505024075508, + 0.009265754371881485, + 0.013067868538200855, + 0.013873426243662834, + 0.020472751930356026, + 0.010143352672457695, + 0.012330388650298119, + 0.007709232624620199, + 0.012105650268495083, + 0.010522178374230862, + 0.013005596585571766, + 0.018975893035531044, + 0.01599534973502159, + 0.013576015830039978, + 0.013514152728021145, + 0.026650002226233482, + 0.003837710479274392, + 0.004667528904974461, + 0.006585198920220137, + 0.0034977842587977648, + 0.00659517478197813, + 0.004509831313043833, + 0.004275545012205839, + 0.0038781671319156885 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 0, + 8 + ], + "selected_reward": -0.023412451148033142 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are two dragon boats racing side by side in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the left-hand boat's hull a saturated crimson color?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the rival boat's hull a vibrant imperial yellow color?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera moving swiftly yet steadily to track the boats?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the visual style a vibrant, hand-drawn animation with bold ink lines?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.08903630077838898 + ], + "opposite_logprob_tokens": [ + -2.464036226272583 + ], + "logprob_reward": -0.08903630077838898, + "yes_no_margin_reward": 2.375, + "token_credit_reward": -0.08903630077838898, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.17900222539901733, + 0.21332819759845734, + 0.053021159023046494, + 0.09437938779592514, + 0.07155094295740128, + 0.0656704306602478, + 0.043393347412347794, + 0.09872523695230484, + 0.06727201491594315, + 0.11365711688995361 + ], + "credit_map": [ + 0.01352180726826191, + 0.014573995023965836, + 0.045820362865924835, + 0.019109351560473442, + 0.037071824073791504, + 0.01792195439338684, + 0.01735682599246502, + 0.01362609677016735, + 0.02303469553589821, + 0.03456706181168556, + 0.026724468916654587, + 0.0210232213139534, + 0.03397899493575096, + 0.03851809352636337, + 0.017729638144373894, + 0.01775202713906765, + 0.005101790186017752, + 0.009640857577323914, + 0.005483498331159353, + 0.006910368334501982, + 0.010042084380984306, + 0.006770401727408171, + 0.004842992871999741, + 0.004229165147989988, + 0.021783173084259033, + 0.007107312325388193, + 0.03329203650355339, + 0.006283563561737537, + 0.006928454153239727, + 0.0076508186757564545, + 0.005255961790680885, + 0.006078063510358334, + 0.009280787780880928, + 0.008751694113016129, + 0.007382520940154791, + 0.008282278664410114, + 0.009762396104633808, + 0.01268912572413683, + 0.008049017749726772, + 0.007353116758167744, + 0.00652511278167367, + 0.00362176401540637, + 0.01319237519055605, + 0.006013275124132633, + 0.020168330520391464, + 0.005568636581301689, + 0.004353195428848267, + 0.0062277428805828094, + 0.005108718294650316, + 0.007023219019174576, + 0.012178696691989899, + 0.0032671212684363127, + 0.005077289417386055, + 0.0035880303476005793, + 0.003132424782961607, + 0.004017846658825874, + 0.007094348315149546, + 0.006748109124600887, + 0.017816754058003426, + 0.010862487368285656, + 0.02368600107729435, + 0.013939120806753635, + 0.00891124363988638, + 0.009667174890637398, + 0.007108728867024183, + 0.0077286954037845135, + 0.006959551014006138, + 0.0068977512419223785, + 0.007796802558004856, + 0.010068628937005997, + 0.006941510830074549, + 0.01377034466713667, + 0.008386981673538685, + 0.009285294450819492, + 0.015317333862185478, + 0.013020051643252373, + 0.017201419919729233, + 0.016451675444841385, + 0.01809833012521267, + 0.01589602790772915 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 9 + ], + "selected_reward": -0.08903630077838898 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are two dragon boats racing side by side in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the left-hand boat's hull a saturated crimson color?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the rival boat's hull a vibrant imperial yellow color?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera moving swiftly yet steadily to track the boats?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the visual style a vibrant, hand-drawn animation with bold ink lines?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.12705960869789124 + ], + "opposite_logprob_tokens": [ + -2.1270596981048584 + ], + "logprob_reward": -0.12705960869789124, + "yes_no_margin_reward": 2.0, + "token_credit_reward": -0.12705960869789124, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.18750132620334625, + 0.20280422270298004, + 0.06519505381584167, + 0.05860692635178566, + 0.11608674377202988, + 0.0399048738181591, + 0.09357719868421555, + 0.152970090508461, + 0.042426805943250656, + 0.04092670604586601 + ], + "credit_map": [ + 0.014416128396987915, + 0.022808322682976723, + 0.05460743233561516, + 0.01405381504446268, + 0.04346276447176933, + 0.015506073832511902, + 0.010776285082101822, + 0.011870539747178555, + 0.035901088267564774, + 0.02189043164253235, + 0.02374039962887764, + 0.010202870704233646, + 0.042176052927970886, + 0.03855441138148308, + 0.012491993606090546, + 0.017846984788775444, + 0.01127413660287857, + 0.008090920746326447, + 0.010308848693966866, + 0.00489639351144433, + 0.01109579112380743, + 0.008748302236199379, + 0.00525486608967185, + 0.005525801330804825, + 0.017723675817251205, + 0.003934007603675127, + 0.01826823316514492, + 0.0036960355937480927, + 0.005313390400260687, + 0.0037010503001511097, + 0.003031390719115734, + 0.0029391469433903694, + 0.011988392099738121, + 0.010542926378548145, + 0.01665438339114189, + 0.01969732902944088, + 0.019827138632535934, + 0.013313603587448597, + 0.013027328066527843, + 0.01103565189987421, + 0.00460833590477705, + 0.004039203282445669, + 0.0036176524590700865, + 0.00708335405215621, + 0.006770200561732054, + 0.004210880026221275, + 0.004798359237611294, + 0.0047768945805728436, + 0.009491849690675735, + 0.018097223713994026, + 0.017235206440091133, + 0.008425723761320114, + 0.012505413964390755, + 0.009145519696176052, + 0.007287708576768637, + 0.01138855516910553, + 0.016633212566375732, + 0.011342734098434448, + 0.023366034030914307, + 0.02087690308690071, + 0.024171750992536545, + 0.017196571454405785, + 0.02432156167924404, + 0.01506134308874607, + 0.0038393198046833277, + 0.0054189846850931644, + 0.005115247331559658, + 0.00886375829577446, + 0.005809470545500517, + 0.0039968634955585, + 0.0030569504015147686, + 0.006326219532638788, + 0.00413915840908885, + 0.005321315489709377, + 0.006200840231031179, + 0.0046470025554299355, + 0.005844796542078257, + 0.007083639968186617, + 0.00416887691244483, + 0.0035210803616791964 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 7 + ], + "selected_reward": -0.12705960869789124 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are two dragon boats racing side by side in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the left-hand boat's hull a saturated crimson color?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the rival boat's hull a vibrant imperial yellow color?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera moving swiftly yet steadily to track the boats?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the visual style a vibrant, hand-drawn animation with bold ink lines?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0699942484498024 + ], + "opposite_logprob_tokens": [ + -2.6949942111968994 + ], + "logprob_reward": -0.0699942484498024, + "yes_no_margin_reward": 2.625, + "token_credit_reward": -0.0699942484498024, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.21201591193675995, + 0.20545922219753265, + 0.056174080818891525, + 0.07597257941961288, + 0.04644227400422096, + 0.22004303336143494, + 0.05258303880691528, + 0.06758763641119003, + 0.033842794597148895, + 0.029879547655582428 + ], + "credit_map": [ + 0.012842318043112755, + 0.015903213992714882, + 0.06110227108001709, + 0.015253729186952114, + 0.06597338616847992, + 0.017147326841950417, + 0.012338299304246902, + 0.01145536545664072, + 0.0193222276866436, + 0.017557037994265556, + 0.0615704320371151, + 0.012229482643306255, + 0.06020022928714752, + 0.018686335533857346, + 0.007308740168809891, + 0.008584734052419662, + 0.007589514367282391, + 0.007216323167085648, + 0.006548451259732246, + 0.007830305024981499, + 0.009118158370256424, + 0.0071130977012217045, + 0.005499332211911678, + 0.0052588991820812225, + 0.009132004342973232, + 0.008264501579105854, + 0.01606052741408348, + 0.008985561318695545, + 0.0077194939367473125, + 0.010533325374126434, + 0.0062240236438810825, + 0.009053145535290241, + 0.00409537460654974, + 0.004943748936057091, + 0.004567848984152079, + 0.00906024593859911, + 0.010780872777104378, + 0.003853290341794491, + 0.005321437027305365, + 0.0038194565568119287, + 0.02310776337981224, + 0.009278761222958565, + 0.06716448813676834, + 0.01243374403566122, + 0.0779499039053917, + 0.013503268361091614, + 0.009636735543608665, + 0.006968364119529724, + 0.007913880050182343, + 0.013077435083687305, + 0.007483449764549732, + 0.004190295469015837, + 0.005233969073742628, + 0.006023120600730181, + 0.0030104652978479862, + 0.005650424398481846, + 0.005739870946854353, + 0.003306723665446043, + 0.01486310176551342, + 0.007305680774152279, + 0.016711143776774406, + 0.008160064928233624, + 0.0073303766548633575, + 0.004170671571046114, + 0.0027484814636409283, + 0.0024820317048579454, + 0.0027179515454918146, + 0.0037350363563746214, + 0.003462755586951971, + 0.006009110249578953, + 0.0035379661712795496, + 0.0091494619846344, + 0.0025172815658152103, + 0.003079455578699708, + 0.00332559528760612, + 0.0033222155179828405, + 0.00413033040240407, + 0.004587464965879917, + 0.005110582802444696, + 0.0038066203705966473 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 0, + 1 + ], + "selected_reward": -0.0699942484498024 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are two dragon boats racing side by side in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the left-hand boat's hull a saturated crimson color?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the rival boat's hull a vibrant imperial yellow color?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera moving swiftly yet steadily to track the boats?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the visual style a vibrant, hand-drawn animation with bold ink lines?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004631980787962675 + ], + "opposite_logprob_tokens": [ + -5.379631996154785 + ], + "logprob_reward": -0.004631980787962675, + "yes_no_margin_reward": 5.375, + "token_credit_reward": -0.004631980787962675, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1853010505437851, + 0.1829717457294464, + 0.046447038650512695, + 0.089653380215168, + 0.050648197531700134, + 0.22870761156082153, + 0.039915479719638824, + 0.030456215143203735, + 0.03997882455587387, + 0.1059204488992691 + ], + "credit_map": [ + 0.013313408009707928, + 0.013996215537190437, + 0.041250161826610565, + 0.018668735399842262, + 0.049704186618328094, + 0.013454261235892773, + 0.021105101332068443, + 0.0138089619576931, + 0.016300706192851067, + 0.026632199063897133, + 0.036818213760852814, + 0.010959068313241005, + 0.05141780152916908, + 0.023233303800225258, + 0.008803172968327999, + 0.008807278238236904, + 0.007879139855504036, + 0.0058161052875220776, + 0.005408070515841246, + 0.004639435093849897, + 0.006775088142603636, + 0.007649086881428957, + 0.003385825315490365, + 0.004894285462796688, + 0.02573523297905922, + 0.006626022048294544, + 0.029314734041690826, + 0.006418140139430761, + 0.0062378933653235435, + 0.00498985406011343, + 0.0050400253385305405, + 0.005291477311402559, + 0.005325159057974815, + 0.004925826098769903, + 0.005485219415277243, + 0.008223233744502068, + 0.009871229529380798, + 0.0066317119635641575, + 0.005535881035029888, + 0.0046499352902174, + 0.023868940770626068, + 0.01113587524741888, + 0.06666288524866104, + 0.014420900493860245, + 0.08023997396230698, + 0.014336436055600643, + 0.010211708024144173, + 0.007830888032913208, + 0.004263534210622311, + 0.009205497801303864, + 0.00680523319169879, + 0.0037331439089030027, + 0.004652088508009911, + 0.0042482404969632626, + 0.0025026460643857718, + 0.004505098331719637, + 0.00519133685156703, + 0.0028799811843782663, + 0.0049058557488024235, + 0.003467799397185445, + 0.004080070182681084, + 0.004480316769331694, + 0.002555222949013114, + 0.002895632293075323, + 0.003663421608507633, + 0.004146421793848276, + 0.004083759151399136, + 0.007021467667073011, + 0.007852092385292053, + 0.00427296943962574, + 0.003008150029927492, + 0.00593054061755538, + 0.008739066310226917, + 0.010852755047380924, + 0.013777161948382854, + 0.010736866854131222, + 0.013049622066318989, + 0.013797545805573463, + 0.019186872988939285, + 0.01578056439757347 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 0, + 1 + ], + "selected_reward": -0.004631980787962675 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "dynamic low-angle tracking shot captures two magnificent dragon boats racing side by side, capturing them neck and neck in a fierce competition. The entire scene comes to life in a vibrant, hand-drawn style reminiscent of high-energy animation. The camera moves swiftly yet steadily, keeping pace with the boats as they jockey for position in the churning, stylized water. The left-hand boat's hull is a saturated crimson, while its rival's is a vibrant imperial yellow, both adorned with meticulously rendered gold and green scales. Their fierce, sculpted dragon-headed prows, bared teeth and glaring eyes, seem to propel them forward with every powerful stroke of the oarsmen. Within each boat, a team of athletes, depicted in bold, expressive ink lines, moves in perfect and powerful sync, their muscles tense and faces set in grim determination. Their oars plunge deep into the azure water, creating explosions of white foam and vibrant speed lines that emphasize the raw power of their movements. The water itself is a turbulent blend of deep blue and green, its surface broken by dynamic, hand-drawn waves and spray. In the background, the sun-drenched riverbank, packed with cheering crowds, presents a colorful, impressionistic blur of movement and excitement. Bright, direct sunlight makes the bold, saturated colors of the boats and water burst, creating a high-contrast, visually dynamic atmosphere filled with the intense energy of the race.", + "seed": 8336, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000071/group_13/index_3425_seed_8336_dp_27_rank_54_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 13, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.029480168875306845, + "vlm_logprob": -0.029480168875306845, + "vlm_yes_no_margin": 4.025, + "vlm_token_credit": -0.029480168875306845, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are two dragon boats racing side by side in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the left-hand boat's hull a saturated crimson color?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the rival boat's hull a vibrant imperial yellow color?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera moving swiftly yet steadily to track the boats?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the visual style a vibrant, hand-drawn animation with bold ink lines?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.012266317382454872 + ], + "opposite_logprob_tokens": [ + -5.637266159057617 + ], + "logprob_reward": -0.012266317382454872, + "yes_no_margin_reward": 5.625, + "token_credit_reward": -0.012266317382454872, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.8598144054412842, + 0.01083480566740036, + 0.019413689151406288, + 0.01958366483449936, + 0.015748227015137672, + 0.019232306629419327, + 0.023997699841856956, + 0.008673381991684437, + 0.018265381455421448, + 0.00443639513105154 + ], + "credit_map": [ + 0.05185945704579353, + 0.05176347494125366, + 0.026255877688527107, + 0.04413338378071785, + 0.04048632085323334, + 0.24136541783809662, + 0.3112582266330719, + 0.0926922932267189, + 0.0010811274405568838, + 0.0008767289691604674, + 0.0013418298913165927, + 0.0011867450084537268, + 0.001331994542852044, + 0.0019495764281600714, + 0.0016990670701488853, + 0.0013677356764674187, + 0.0021029398776590824, + 0.0021685801912099123, + 0.004212496802210808, + 0.0020890631712973118, + 0.0023585769813507795, + 0.002375199692323804, + 0.002341191517189145, + 0.0017656432464718819, + 0.004083795007318258, + 0.002899828599765897, + 0.0017858986975625157, + 0.0016858599847182631, + 0.001658766414038837, + 0.002079325495287776, + 0.0028740938287228346, + 0.002516097854822874, + 0.0030268363188952208, + 0.0015942230820655823, + 0.0014239095617085695, + 0.0021721224766224623, + 0.002879339735955, + 0.0015626308741047978, + 0.0016637115040794015, + 0.0014254546258598566, + 0.0029004726093262434, + 0.0018672285368666053, + 0.0019239505054429173, + 0.002493957057595253, + 0.0028639822266995907, + 0.0021358104422688484, + 0.002955675357952714, + 0.0020912312902510166, + 0.0033570455852895975, + 0.0030487533658742905, + 0.002054505283012986, + 0.002946313703432679, + 0.00268744258210063, + 0.005393983330577612, + 0.0014548092149198055, + 0.003054849337786436, + 0.0011697270674631, + 0.0010961617808789015, + 0.0009319956880062819, + 0.0011415586341172457, + 0.0009011440561152995, + 0.0016500391066074371, + 0.0008476240327581763, + 0.0009351318585686386, + 0.0016964408569037914, + 0.00163507170509547, + 0.0017293692799285054, + 0.004608990624547005, + 0.0012531650718301535, + 0.0013770291116088629, + 0.0013284824090078473, + 0.004636833444237709, + 0.0006442873273044825, + 0.0006348396418616176, + 0.0007020212942734361, + 0.0004641092964448035, + 0.0004446418897714466, + 0.0005208335933275521, + 0.00046645678230561316, + 0.0005592058296315372 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 6, + 3 + ], + "selected_reward": -0.012266317382454872 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are two dragon boats racing side by side in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the left-hand boat's hull a saturated crimson color?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the rival boat's hull a vibrant imperial yellow color?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera moving swiftly yet steadily to track the boats?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the visual style a vibrant, hand-drawn animation with bold ink lines?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04307493194937706 + ], + "opposite_logprob_tokens": [ + -3.1680748462677 + ], + "logprob_reward": -0.04307493194937706, + "yes_no_margin_reward": 3.125, + "token_credit_reward": -0.04307493194937706, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.6971490979194641, + 0.03036629967391491, + 0.044959764927625656, + 0.09364116936922073, + 0.008039554581046104, + 0.03000195510685444, + 0.02125350944697857, + 0.03994245082139969, + 0.010041603818535805, + 0.02460464835166931 + ], + "credit_map": [ + 0.04214150458574295, + 0.041896186769008636, + 0.020847130566835403, + 0.03654264286160469, + 0.03284738212823868, + 0.19770000874996185, + 0.25162336230278015, + 0.07355090975761414, + 0.0034495005384087563, + 0.0028901405166834593, + 0.004792098421603441, + 0.0030094569083303213, + 0.004941603168845177, + 0.006231572013348341, + 0.002963321516290307, + 0.0020886061247438192, + 0.006221082992851734, + 0.0036141476593911648, + 0.006214572116732597, + 0.0036849151365458965, + 0.005233625415712595, + 0.007474284619092941, + 0.008915575221180916, + 0.003601564560085535, + 0.014047633856534958, + 0.007926455698907375, + 0.006592741701751947, + 0.009196409024298191, + 0.006031435914337635, + 0.005504924338310957, + 0.02651265822350979, + 0.017828911542892456, + 0.0014453164767473936, + 0.0009230458526872098, + 0.0007300310535356402, + 0.0009188189869746566, + 0.0013616776559501886, + 0.000782718590926379, + 0.0009338448871858418, + 0.0009441003203392029, + 0.005050624255090952, + 0.0035668034106492996, + 0.002905648434534669, + 0.003378160996362567, + 0.0038130017928779125, + 0.0035342497285455465, + 0.0040163034573197365, + 0.0037371621001511812, + 0.003202148713171482, + 0.0025509772822260857, + 0.0022016875445842743, + 0.002192308660596609, + 0.0023722515907138586, + 0.003149727126583457, + 0.0022974645253270864, + 0.0032869449350982904, + 0.004867228679358959, + 0.007150415331125259, + 0.003213397925719619, + 0.003982337191700935, + 0.005856951232999563, + 0.006830353755503893, + 0.003234465140849352, + 0.004807302728295326, + 0.0009092268883250654, + 0.0009496964048594236, + 0.000716847600415349, + 0.002186262747272849, + 0.0009013741509988904, + 0.001322887372225523, + 0.0007506298716180027, + 0.002304678549990058, + 0.0030741561204195023, + 0.0026384671218693256, + 0.0036158461589366198, + 0.002540378598496318, + 0.0028167939744889736, + 0.00409048143774271, + 0.0024303016252815723, + 0.0033982230816036463 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 2 + ], + "selected_reward": -0.04307493194937706 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are two dragon boats racing side by side in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the left-hand boat's hull a saturated crimson color?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the rival boat's hull a vibrant imperial yellow color?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera moving swiftly yet steadily to track the boats?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the visual style a vibrant, hand-drawn animation with bold ink lines?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04307344928383827 + ], + "opposite_logprob_tokens": [ + -3.1680734157562256 + ], + "logprob_reward": -0.04307344928383827, + "yes_no_margin_reward": 3.125, + "token_credit_reward": -0.04307344928383827, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.49929389357566833, + 0.10192541778087616, + 0.05456287041306496, + 0.19172403216362, + 0.013294916599988937, + 0.04824581369757652, + 0.059996820986270905, + 0.012538000009953976, + 0.012527541257441044, + 0.005890641827136278 + ], + "credit_map": [ + 0.029474688693881035, + 0.03157675638794899, + 0.015925208106637, + 0.026194194331765175, + 0.024392493069171906, + 0.13858391344547272, + 0.1795487403869629, + 0.05359794944524765, + 0.012147445231676102, + 0.009406780824065208, + 0.012515942566096783, + 0.00993050541728735, + 0.019492652267217636, + 0.021170439198613167, + 0.011158157140016556, + 0.006103492341935635, + 0.005915337707847357, + 0.004827296361327171, + 0.00573639664798975, + 0.004486251622438431, + 0.009392664767801762, + 0.009636850096285343, + 0.010555332526564598, + 0.00401274161413312, + 0.0183757022023201, + 0.009297792799770832, + 0.011225732043385506, + 0.017083797603845596, + 0.013166029937565327, + 0.013891643844544888, + 0.06584227830171585, + 0.042841073125600815, + 0.00176422786898911, + 0.0013459582114592195, + 0.0013411447871476412, + 0.0017709521343931556, + 0.002337924437597394, + 0.001547517254948616, + 0.0015599624020978808, + 0.001627230318263173, + 0.006629405543208122, + 0.004274771083146334, + 0.003333679400384426, + 0.004913756158202887, + 0.00660617183893919, + 0.0082086231559515, + 0.00790433119982481, + 0.006375073920935392, + 0.008481438271701336, + 0.006897693034261465, + 0.005067720077931881, + 0.005110837519168854, + 0.006970641203224659, + 0.015318639576435089, + 0.004776885267347097, + 0.007372962776571512, + 0.0012227375991642475, + 0.0014936531661078334, + 0.0012810557382181287, + 0.001838603988289833, + 0.0015994524583220482, + 0.0016939281485974789, + 0.0013093419838696718, + 0.0020992266945540905, + 0.001542392885312438, + 0.0015699077630415559, + 0.0011069560423493385, + 0.0013475756859406829, + 0.001423359615728259, + 0.0031125247478485107, + 0.001096948515623808, + 0.0013278760015964508, + 0.0007026463863439858, + 0.0005689627141691744, + 0.0007419602479785681, + 0.0004824348143301904, + 0.0007338319555856287, + 0.0014700257452204823, + 0.0005540729034692049, + 0.0006367075257003307 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 1 + ], + "selected_reward": -0.04307344928383827 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are two dragon boats racing side by side in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the left-hand boat's hull a saturated crimson color?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the rival boat's hull a vibrant imperial yellow color?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera moving swiftly yet steadily to track the boats?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the visual style a vibrant, hand-drawn animation with bold ink lines?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0430445596575737 + ], + "opposite_logprob_tokens": [ + -3.1680445671081543 + ], + "logprob_reward": -0.0430445596575737, + "yes_no_margin_reward": 3.125, + "token_credit_reward": -0.0430445596575737, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.8290290236473083, + 0.014488950371742249, + 0.014153429307043552, + 0.05361180752515793, + 0.02525206096470356, + 0.018294651061296463, + 0.03403559327125549, + 0.0025275214575231075, + 0.006978721357882023, + 0.0016281979624181986 + ], + "credit_map": [ + 0.04961596429347992, + 0.049344878643751144, + 0.02534477785229683, + 0.04301775246858597, + 0.03901039436459541, + 0.23143169283866882, + 0.3013181984424591, + 0.0899454727768898, + 0.001788930967450142, + 0.0014598091365769506, + 0.0016183769330382347, + 0.0026092601474374533, + 0.0018526199273765087, + 0.0021668956615030766, + 0.0016993626486510038, + 0.0012936970451846719, + 0.0018415579106658697, + 0.0015272110467776656, + 0.0029416466131806374, + 0.0014123745495453477, + 0.0016728354385122657, + 0.0017220382578670979, + 0.0017332436982542276, + 0.0013025233056396246, + 0.005666294600814581, + 0.0025821239687502384, + 0.0027656962629407644, + 0.0047286623157560825, + 0.003542395308613777, + 0.0036805961281061172, + 0.018508736044168472, + 0.012137305922806263, + 0.004113682545721531, + 0.0026642943266779184, + 0.0018803675193339586, + 0.0031835767440497875, + 0.00464701047167182, + 0.0026637392584234476, + 0.0033190855756402016, + 0.0027803066186606884, + 0.0023449452128261328, + 0.0015306032728403807, + 0.0019183158874511719, + 0.0025538462214171886, + 0.0030760634690523148, + 0.0021961918100714684, + 0.002675804076716304, + 0.0019988848362118006, + 0.0041233692318201065, + 0.003731951816007495, + 0.0027672420255839825, + 0.0036595219280570745, + 0.005491912830621004, + 0.0065153539180755615, + 0.0032505940180271864, + 0.004495650064200163, + 0.0004002895439043641, + 0.00040471984539180994, + 0.0002650473907124251, + 0.00021205894881859422, + 0.00033853031345643103, + 0.00036128360079601407, + 0.00019789760699495673, + 0.0003476945566944778, + 0.0007751443772576749, + 0.0007154892082326114, + 0.0006020423606969416, + 0.0012338279047980905, + 0.0007573344628326595, + 0.0012012815568596125, + 0.00046696848585270345, + 0.0012266337871551514, + 0.00027839484391734004, + 0.0002390884910710156, + 0.00023296437575481832, + 0.00016176937788259238, + 0.00016510086425114423, + 0.0002089708432322368, + 0.000138339280965738, + 0.00020357005996629596 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 6 + ], + "selected_reward": -0.0430445596575737 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are two dragon boats racing side by side in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the left-hand boat's hull a saturated crimson color?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the rival boat's hull a vibrant imperial yellow color?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera moving swiftly yet steadily to track the boats?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the visual style a vibrant, hand-drawn animation with bold ink lines?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0059415861032903194 + ], + "opposite_logprob_tokens": [ + -5.130941390991211 + ], + "logprob_reward": -0.0059415861032903194, + "yes_no_margin_reward": 5.125, + "token_credit_reward": -0.0059415861032903194, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.750803530216217, + 0.04290333390235901, + 0.026767155155539513, + 0.06912453472614288, + 0.03110029734671116, + 0.015275119803845882, + 0.01389189250767231, + 0.01707649976015091, + 0.02599002607166767, + 0.007067553699016571 + ], + "credit_map": [ + 0.045961491763591766, + 0.04462708532810211, + 0.0231434628367424, + 0.03999495133757591, + 0.0343497060239315, + 0.21198567748069763, + 0.2711489796638489, + 0.07959217578172684, + 0.006795969326049089, + 0.0037277338560670614, + 0.004322421737015247, + 0.006746930070221424, + 0.005219050217419863, + 0.0074896542355418205, + 0.004948786459863186, + 0.0036527905613183975, + 0.004742826800793409, + 0.0024924506433308125, + 0.004642164334654808, + 0.00168040266726166, + 0.004693503957241774, + 0.004089075140655041, + 0.0024150898680090904, + 0.002011646283790469, + 0.01291834656149149, + 0.009451562538743019, + 0.005650372244417667, + 0.0065478007309138775, + 0.005569948349148035, + 0.005373658612370491, + 0.013113810680806637, + 0.010499042458832264, + 0.003290948225185275, + 0.002514368388801813, + 0.0018682352965697646, + 0.0035905218683183193, + 0.005875677336007357, + 0.004154921043664217, + 0.0051432144828140736, + 0.004662413615733385, + 0.0019998156931251287, + 0.0015115400310605764, + 0.0015265546971932054, + 0.0020931900944560766, + 0.0017037322977557778, + 0.001867686747573316, + 0.0025792550295591354, + 0.0019933469593524933, + 0.0019796211272478104, + 0.0016625216230750084, + 0.001296766335144639, + 0.0015733831096440554, + 0.0017764520598575473, + 0.0028225898277014494, + 0.0012615795712918043, + 0.0015189796686172485, + 0.001954158302396536, + 0.0018990867538377643, + 0.0014597463887184858, + 0.0014054152416065335, + 0.0015408394392579794, + 0.001475257333368063, + 0.003153587458655238, + 0.00418840954080224, + 0.0023905886337161064, + 0.0027433421928435564, + 0.0020406125113368034, + 0.005662720650434494, + 0.002099719364196062, + 0.003530807327479124, + 0.002047968562692404, + 0.005474269390106201, + 0.0010806444333866239, + 0.0007472934084944427, + 0.0009741985704749823, + 0.0008421146194450557, + 0.0007481012726202607, + 0.001078653847798705, + 0.0007692635990679264, + 0.0008272844715975225 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 1 + ], + "selected_reward": -0.0059415861032903194 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000071/group_14/index_tensor([1504, 1504])_group_14_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000071/group_14/index_tensor([1504, 1504])_group_14_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..0c6635557e408f2a0495ec78cfdf70222129c156 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000071/group_14/index_tensor([1504, 1504])_group_14_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "Unfolding in an ultra-detailed artistic style, this mesmerizing time-lapse, captured from static, ground-level macro footage, immerses the viewer in the scene. The camera closely follows a gnarled ginseng root, shaped like a sleeping figure with twisted limbs, nestled in dark, rich, fertile soil dotted with tiny pebbles. A soft beam of light penetrates the darkness from above, creating a dramatic black-and-white contrast effect that meticulously highlights the root's woody, fibrous skin, deep cracks, and tiny root hairs. Crystal-like water droplets begin to fall in a rhythmic, deliberate sequence, impacting the soil and roots; each drop glints like a gem in the light before being greedily absorbed, darkening the earth and tracing a shimmering path along the root's contours. As time passes faster, the root visibly plumps and swells with latent energy, its color deepening to a healthier, more saturated brown. The sequence climaxes when the surface of the soil just above the root “head” begins to crack and bulge, and then a tiny, vibrant green sprout miraculously forces its way through the earth, steadily pushing upward and unfolding two tiny, shimmering cotyledons that catch the light, creating an atmosphere of quiet yet profound wonder at the powerful magic of life renewal.", + "seed": 1071, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000071/group_14/index_1504_seed_1071_dp_28_rank_56_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 14, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.06215846110135317, + "vlm_logprob": -0.06215846110135317, + "vlm_yes_no_margin": 3.75, + "vlm_token_credit": -0.06215846110135317, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned at a static, ground-level macro perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a gnarled ginseng root shaped like a sleeping figure with twisted limbs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the scene feature a soft beam of light creating a black-and-white contrast effect?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do crystal-like water droplets fall rhythmically, glint like gems, and darken the soil upon impact?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video show the root swelling and a tiny green sprout pushing upward to unfold two cotyledons?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.021370267495512962 + ], + "opposite_logprob_tokens": [ + -5.7713704109191895 + ], + "logprob_reward": -0.021370267495512962, + "yes_no_margin_reward": 5.75, + "token_credit_reward": -0.021370267495512962, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.18475839495658875, + 0.059372954070568085, + 0.0797034353017807, + 0.043387942016124725, + 0.07738350331783295, + 0.032839540392160416, + 0.08304276317358017, + 0.17041683197021484, + 0.06954804062843323, + 0.19954663515090942 + ], + "credit_map": [ + 0.016720615327358246, + 0.023109665140509605, + 0.023569658398628235, + 0.039538394659757614, + 0.015897851437330246, + 0.015062381513416767, + 0.02045980654656887, + 0.030400017276406288, + 0.007158674765378237, + 0.007234199438244104, + 0.009435976855456829, + 0.011010410264134407, + 0.0055561065673828125, + 0.004657106939703226, + 0.006448693573474884, + 0.007871784269809723, + 0.007486221846193075, + 0.008988470770418644, + 0.008357915095984936, + 0.012686076574027538, + 0.011469959281384945, + 0.007699881214648485, + 0.011794806458055973, + 0.011220106855034828, + 0.004584394861012697, + 0.003894777735695243, + 0.00832308642566204, + 0.00630428921431303, + 0.00613373750820756, + 0.0030685733072459698, + 0.005573351867496967, + 0.005505730863660574, + 0.007766024209558964, + 0.006441363599151373, + 0.010339641012251377, + 0.01374615728855133, + 0.008849767968058586, + 0.006816324777901173, + 0.015262194909155369, + 0.008162031881511211, + 0.004273892380297184, + 0.004019707441329956, + 0.005642471835017204, + 0.0051820469088852406, + 0.0025165711995214224, + 0.0028752200305461884, + 0.005457697901874781, + 0.0028719326946884394, + 0.011331112124025822, + 0.009261513128876686, + 0.012664996087551117, + 0.012470093555748463, + 0.007453275844454765, + 0.009417058899998665, + 0.012980042025446892, + 0.007464674301445484, + 0.014506271108984947, + 0.04221459850668907, + 0.010814323090016842, + 0.019093023613095284, + 0.011496908962726593, + 0.010270552709698677, + 0.04684945195913315, + 0.015171707607805729, + 0.010432039387524128, + 0.00646649906411767, + 0.008482090197503567, + 0.010482709854841232, + 0.009108874946832657, + 0.007114298641681671, + 0.010649795643985271, + 0.006811737548559904, + 0.0330541618168354, + 0.018508197739720345, + 0.027440477162599564, + 0.022441565990447998, + 0.02115195244550705, + 0.018398776650428772, + 0.028373263776302338, + 0.030178237706422806 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 0, + 7 + ], + "selected_reward": -0.021370267495512962 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned at a static, ground-level macro perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a gnarled ginseng root shaped like a sleeping figure with twisted limbs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the scene feature a soft beam of light creating a black-and-white contrast effect?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do crystal-like water droplets fall rhythmically, glint like gems, and darken the soil upon impact?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video show the root swelling and a tiny green sprout pushing upward to unfold two cotyledons?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02636430785059929 + ], + "opposite_logprob_tokens": [ + -3.651364326477051 + ], + "logprob_reward": -0.02636430785059929, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.02636430785059929, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06943368166685104, + 0.09102891385555267, + 0.11154749989509583, + 0.030023781582713127, + 0.04103294759988785, + 0.04241413250565529, + 0.028947582468390465, + 0.40828418731689453, + 0.05770935118198395, + 0.11957793682813644 + ], + "credit_map": [ + 0.008608602918684483, + 0.007969400845468044, + 0.009720477275550365, + 0.012195851653814316, + 0.008848157711327076, + 0.005543196573853493, + 0.009006321430206299, + 0.0075416723266243935, + 0.013096480630338192, + 0.013132133521139622, + 0.008444448933005333, + 0.011809761635959148, + 0.01216261088848114, + 0.005323187448084354, + 0.015715915709733963, + 0.011344367638230324, + 0.01162912417203188, + 0.014127086848020554, + 0.01626020111143589, + 0.013573591597378254, + 0.018203092738986015, + 0.00987337063997984, + 0.013838252052664757, + 0.014042770490050316, + 0.004230082035064697, + 0.005010385066270828, + 0.003374797059223056, + 0.0030000812839716673, + 0.004208425525575876, + 0.0028947792015969753, + 0.0033713150769472122, + 0.003933915868401527, + 0.002743995049968362, + 0.004400257021188736, + 0.0056875720620155334, + 0.004413597285747528, + 0.008119146339595318, + 0.0043513840064406395, + 0.007569507695734501, + 0.0037474874407052994, + 0.004440737888216972, + 0.005763460416346788, + 0.00840055849403143, + 0.006343604996800423, + 0.0049026357010006905, + 0.0035166959278285503, + 0.0053300922736525536, + 0.003716345876455307, + 0.003360005794093013, + 0.0031580242794007063, + 0.0030423139687627554, + 0.00350740528665483, + 0.005779098253697157, + 0.0036750948056578636, + 0.0035540407989174128, + 0.002871600678190589, + 0.03938324376940727, + 0.10886068642139435, + 0.020921295508742332, + 0.04025142639875412, + 0.02861694060266018, + 0.019892243668437004, + 0.11374842375516891, + 0.036609947681427, + 0.008285535499453545, + 0.006070547737181187, + 0.007610643748193979, + 0.0064296601340174675, + 0.010141671635210514, + 0.004576321225613356, + 0.009114403277635574, + 0.005480571184307337, + 0.010373464785516262, + 0.011485892347991467, + 0.01363849826157093, + 0.01670767553150654, + 0.012336201034486294, + 0.015755483880639076, + 0.01808561198413372, + 0.021195117384195328 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 9, + 2 + ], + "selected_reward": -0.02636430785059929 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned at a static, ground-level macro perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a gnarled ginseng root shaped like a sleeping figure with twisted limbs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the scene feature a soft beam of light creating a black-and-white contrast effect?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do crystal-like water droplets fall rhythmically, glint like gems, and darken the soil upon impact?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video show the root swelling and a tiny green sprout pushing upward to unfold two cotyledons?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.22557227313518524 + ], + "opposite_logprob_tokens": [ + -1.6005722284317017 + ], + "logprob_reward": -0.22557227313518524, + "yes_no_margin_reward": 1.375, + "token_credit_reward": -0.22557227313518524, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11516924947500229, + 0.06539667397737503, + 0.06544608622789383, + 0.02813100256025791, + 0.060368746519088745, + 0.11795766651630402, + 0.07605312764644623, + 0.06327225267887115, + 0.2766037583351135, + 0.1316014677286148 + ], + "credit_map": [ + 0.01351233758032322, + 0.011864501051604748, + 0.018774691969156265, + 0.01495493296533823, + 0.011299201287329197, + 0.007677845191210508, + 0.023643679916858673, + 0.013442039489746094, + 0.009936672635376453, + 0.009128809906542301, + 0.005725144408643246, + 0.010111717507243156, + 0.006351557560265064, + 0.004277932923287153, + 0.010274396277964115, + 0.00959043763577938, + 0.0055038114078342915, + 0.005621312651783228, + 0.008664690889418125, + 0.012186802923679352, + 0.007256384938955307, + 0.005916275549679995, + 0.013305981643497944, + 0.006990816444158554, + 0.004142712336033583, + 0.005076443310827017, + 0.0036120328586548567, + 0.003950720187276602, + 0.0029428701382130384, + 0.002440104028210044, + 0.003398437751457095, + 0.002567676827311516, + 0.003839056473225355, + 0.005615228321403265, + 0.0063157761469483376, + 0.008483344689011574, + 0.010513458400964737, + 0.007144736126065254, + 0.012387031689286232, + 0.006070102099329233, + 0.007358393631875515, + 0.01169651746749878, + 0.0135797755792737, + 0.020937709137797356, + 0.02176349051296711, + 0.009771814569830894, + 0.022567924112081528, + 0.010282022878527641, + 0.014970737509429455, + 0.008597146719694138, + 0.007032007910311222, + 0.012552694417536259, + 0.006435461342334747, + 0.005120756104588509, + 0.014447885565459728, + 0.006896429695188999, + 0.007161475718021393, + 0.014814398251473904, + 0.0046706609427928925, + 0.007044997997581959, + 0.004883492831140757, + 0.004857389722019434, + 0.01502828299999237, + 0.004811552818864584, + 0.026539385318756104, + 0.03149297833442688, + 0.031812768429517746, + 0.03338620439171791, + 0.06606114655733109, + 0.026492709293961525, + 0.03793666139245033, + 0.02288188226521015, + 0.011192567646503448, + 0.017574984580278397, + 0.01394015271216631, + 0.018917372450232506, + 0.013718456961214542, + 0.01496639009565115, + 0.019308099523186684, + 0.021983420476317406 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 9, + 5 + ], + "selected_reward": -0.22557227313518524 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned at a static, ground-level macro perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a gnarled ginseng root shaped like a sleeping figure with twisted limbs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the scene feature a soft beam of light creating a black-and-white contrast effect?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do crystal-like water droplets fall rhythmically, glint like gems, and darken the soil upon impact?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video show the root swelling and a tiny green sprout pushing upward to unfold two cotyledons?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.023291684687137604 + ], + "opposite_logprob_tokens": [ + -3.77329158782959 + ], + "logprob_reward": -0.023291684687137604, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.023291684687137604, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13855569064617157, + 0.04566830396652222, + 0.09587427973747253, + 0.09605740010738373, + 0.11629116535186768, + 0.18371577560901642, + 0.08388575911521912, + 0.043259210884571075, + 0.13043898344039917, + 0.06625346094369888 + ], + "credit_map": [ + 0.01813015155494213, + 0.01748741790652275, + 0.027237804606556892, + 0.019840190187096596, + 0.013313311152160168, + 0.009825867600739002, + 0.01405361108481884, + 0.018667323514819145, + 0.003816324518993497, + 0.004300416447222233, + 0.005352420266717672, + 0.006323942914605141, + 0.007163741625845432, + 0.003049554768949747, + 0.01037667691707611, + 0.005285223480314016, + 0.007965948432683945, + 0.013545393012464046, + 0.0193150844424963, + 0.013582726009190083, + 0.008060979656875134, + 0.004975356161594391, + 0.016508542001247406, + 0.011920241639018059, + 0.017148006707429886, + 0.009463552385568619, + 0.011884640902280807, + 0.011688850820064545, + 0.008575736545026302, + 0.006726028863340616, + 0.012633354403078556, + 0.017937228083610535, + 0.014217868447303772, + 0.016468709334731102, + 0.025669075548648834, + 0.016193848103284836, + 0.01463482715189457, + 0.008618902415037155, + 0.012422320432960987, + 0.008065622299909592, + 0.027595458552241325, + 0.024414056912064552, + 0.036882709711790085, + 0.02147662825882435, + 0.017679894343018532, + 0.010540958493947983, + 0.029583537951111794, + 0.015542524866759777, + 0.009636118076741695, + 0.008862190879881382, + 0.010635968297719955, + 0.01535901427268982, + 0.009378800168633461, + 0.007621947675943375, + 0.015912532806396484, + 0.006479187402874231, + 0.003887810045853257, + 0.008950126357376575, + 0.0038776546716690063, + 0.004934108816087246, + 0.003276858478784561, + 0.0030517117120325565, + 0.011336883530020714, + 0.003944057505577803, + 0.012819220311939716, + 0.014685658738017082, + 0.014541903510689735, + 0.017474155873060226, + 0.03307999670505524, + 0.011263522319495678, + 0.015960633754730225, + 0.01061388198286295, + 0.007485994137823582, + 0.0069420672953128815, + 0.00996171310544014, + 0.00902058556675911, + 0.008617048151791096, + 0.006046330090612173, + 0.010313553735613823, + 0.007866167463362217 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 0, + 8 + ], + "selected_reward": -0.023291684687137604 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned at a static, ground-level macro perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a gnarled ginseng root shaped like a sleeping figure with twisted limbs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the scene feature a soft beam of light creating a black-and-white contrast effect?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do crystal-like water droplets fall rhythmically, glint like gems, and darken the soil upon impact?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video show the root swelling and a tiny green sprout pushing upward to unfold two cotyledons?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.014193772338330746 + ], + "opposite_logprob_tokens": [ + -4.264193534851074 + ], + "logprob_reward": -0.014193772338330746, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.014193772338330746, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06208143010735512, + 0.09170053154230118, + 0.06322543323040009, + 0.07207126915454865, + 0.08160831779241562, + 0.09304466843605042, + 0.02032802626490593, + 0.11850618571043015, + 0.05397937074303627, + 0.34345483779907227 + ], + "credit_map": [ + 0.007000068202614784, + 0.008345534093677998, + 0.010005115531384945, + 0.009084640070796013, + 0.005072108935564756, + 0.004311603959649801, + 0.009508023038506508, + 0.008754328824579716, + 0.00921107642352581, + 0.009862948209047318, + 0.012190061621367931, + 0.017912408336997032, + 0.014885948970913887, + 0.006852657534182072, + 0.00830760970711708, + 0.012477812357246876, + 0.005943316966295242, + 0.00791154708713293, + 0.00981473084539175, + 0.007669053506106138, + 0.010613911785185337, + 0.005440964829176664, + 0.007771574426442385, + 0.008060329593718052, + 0.008226322941482067, + 0.012341017834842205, + 0.009280549362301826, + 0.0135834701359272, + 0.006703425198793411, + 0.004960312973707914, + 0.009091151878237724, + 0.007885011844336987, + 0.006681269034743309, + 0.008892694488167763, + 0.013321836479008198, + 0.013614418916404247, + 0.012344657443463802, + 0.008076608180999756, + 0.012129775248467922, + 0.006547046359628439, + 0.01249684114009142, + 0.016956519335508347, + 0.010838069953024387, + 0.01508655771613121, + 0.006863664835691452, + 0.009441726841032505, + 0.013101387768983841, + 0.008259894326329231, + 0.0018671642756089568, + 0.0025711979251354933, + 0.0028490708209574223, + 0.003748118644580245, + 0.0022289822809398174, + 0.0016808283980935812, + 0.003585863159969449, + 0.001796799129806459, + 0.011133678257465363, + 0.02879127860069275, + 0.008831292390823364, + 0.012353114783763885, + 0.008451148867607117, + 0.007154862396419048, + 0.03129059076309204, + 0.010500206612050533, + 0.008337014354765415, + 0.007736533414572477, + 0.005931625608354807, + 0.005843943450599909, + 0.009029361419379711, + 0.00508182542398572, + 0.007051756605505943, + 0.00496730487793684, + 0.02512817084789276, + 0.03244282305240631, + 0.0350487194955349, + 0.04402840510010719, + 0.026053691282868385, + 0.042618561536073685, + 0.0576290525496006, + 0.08050537109375 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 7, + 5 + ], + "selected_reward": -0.014193772338330746 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "Unfolding in an ultra-detailed artistic style, this mesmerizing time-lapse, captured from static, ground-level macro footage, immerses the viewer in the scene. The camera closely follows a gnarled ginseng root, shaped like a sleeping figure with twisted limbs, nestled in dark, rich, fertile soil dotted with tiny pebbles. A soft beam of light penetrates the darkness from above, creating a dramatic black-and-white contrast effect that meticulously highlights the root's woody, fibrous skin, deep cracks, and tiny root hairs. Crystal-like water droplets begin to fall in a rhythmic, deliberate sequence, impacting the soil and roots; each drop glints like a gem in the light before being greedily absorbed, darkening the earth and tracing a shimmering path along the root's contours. As time passes faster, the root visibly plumps and swells with latent energy, its color deepening to a healthier, more saturated brown. The sequence climaxes when the surface of the soil just above the root “head” begins to crack and bulge, and then a tiny, vibrant green sprout miraculously forces its way through the earth, steadily pushing upward and unfolding two tiny, shimmering cotyledons that catch the light, creating an atmosphere of quiet yet profound wonder at the powerful magic of life renewal.", + "seed": 1072, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000071/group_14/index_1504_seed_1072_dp_28_rank_56_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 14, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.11123427227139474, + "vlm_logprob": -0.11123427227139474, + "vlm_yes_no_margin": 3.6, + "vlm_token_credit": -0.11123427227139474, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned at a static, ground-level macro perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a gnarled ginseng root shaped like a sleeping figure with twisted limbs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the scene feature a soft beam of light creating a black-and-white contrast effect?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do crystal-like water droplets fall rhythmically, glint like gems, and darken the soil upon impact?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video show the root swelling and a tiny green sprout pushing upward to unfold two cotyledons?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.027890073135495186 + ], + "opposite_logprob_tokens": [ + -5.402890205383301 + ], + "logprob_reward": -0.027890073135495186, + "yes_no_margin_reward": 5.375, + "token_credit_reward": -0.027890073135495186, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2896728813648224, + 0.11860405653715134, + 0.09418846666812897, + 0.061364881694316864, + 0.0830475240945816, + 0.05819997936487198, + 0.09798455983400345, + 0.038206275552511215, + 0.08363228291273117, + 0.07509906589984894 + ], + "credit_map": [ + 0.02243443951010704, + 0.014701420441269875, + 0.04822736978530884, + 0.016349781304597855, + 0.022861264646053314, + 0.11108848452568054, + 0.010063733905553818, + 0.04394639655947685, + 0.011101052165031433, + 0.00947771966457367, + 0.016772404313087463, + 0.014213815331459045, + 0.008401081897318363, + 0.031235748901963234, + 0.014551637694239616, + 0.012850601226091385, + 0.012861186638474464, + 0.011407166719436646, + 0.008801124058663845, + 0.007848604582250118, + 0.009783204644918442, + 0.026099741458892822, + 0.0070475200191140175, + 0.010339927859604359, + 0.0064704082906246185, + 0.004114488605409861, + 0.0113713713362813, + 0.012827686034142971, + 0.004105524159967899, + 0.006846870761364698, + 0.004778514616191387, + 0.010850022546947002, + 0.022392241284251213, + 0.005044450052082539, + 0.011917910538613796, + 0.006791118998080492, + 0.006754311732947826, + 0.0035636425018310547, + 0.00447044987231493, + 0.022113407030701637, + 0.01604408212006092, + 0.004075170494616032, + 0.010856794193387032, + 0.0045140269212424755, + 0.005685522686690092, + 0.0047316416166722775, + 0.003922535106539726, + 0.00837020855396986, + 0.023914150893688202, + 0.007844926789402962, + 0.006678826175630093, + 0.005227361805737019, + 0.01298463810235262, + 0.005223227199167013, + 0.004583800211548805, + 0.03152763843536377, + 0.009205980226397514, + 0.002569872420281172, + 0.005190997384488583, + 0.0043303342536091805, + 0.00302691338583827, + 0.0019873196724802256, + 0.004153065383434296, + 0.007741795387119055, + 0.010427283123135567, + 0.012644965201616287, + 0.015093172900378704, + 0.009784427471458912, + 0.012585929594933987, + 0.00846156943589449, + 0.00662779388949275, + 0.008007150143384933, + 0.011761358007788658, + 0.013207566924393177, + 0.011222749948501587, + 0.008604061789810658, + 0.00678042508661747, + 0.005526680499315262, + 0.007813559845089912, + 0.010182672180235386 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 6 + ], + "selected_reward": -0.027890073135495186 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned at a static, ground-level macro perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a gnarled ginseng root shaped like a sleeping figure with twisted limbs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the scene feature a soft beam of light creating a black-and-white contrast effect?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do crystal-like water droplets fall rhythmically, glint like gems, and darken the soil upon impact?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video show the root swelling and a tiny green sprout pushing upward to unfold two cotyledons?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.023302050307393074 + ], + "opposite_logprob_tokens": [ + -3.7733020782470703 + ], + "logprob_reward": -0.023302050307393074, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.023302050307393074, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.23426714539527893, + 0.08161278069019318, + 0.06342635303735733, + 0.07430601119995117, + 0.04411821812391281, + 0.05257095769047737, + 0.03220207989215851, + 0.042228370904922485, + 0.19408626854419708, + 0.18118180334568024 + ], + "credit_map": [ + 0.014878764748573303, + 0.014591081067919731, + 0.019276168197393417, + 0.014887536875903606, + 0.020050067454576492, + 0.08784306794404984, + 0.014079377055168152, + 0.04866110160946846, + 0.007003589533269405, + 0.004450090229511261, + 0.018450219184160233, + 0.009653188288211823, + 0.004461402539163828, + 0.019807549193501472, + 0.009203342720866203, + 0.008583403192460537, + 0.006172981578856707, + 0.007415317930281162, + 0.004597478546202183, + 0.006085703615099192, + 0.0068953512236475945, + 0.015062971971929073, + 0.006642183754593134, + 0.010554371401667595, + 0.00853461679071188, + 0.008061797358095646, + 0.00791881512850523, + 0.013393119908869267, + 0.013108666054904461, + 0.008444664999842644, + 0.006572613026946783, + 0.008271723985671997, + 0.00986525509506464, + 0.003841298632323742, + 0.004005675204098225, + 0.0037279464304447174, + 0.004767660982906818, + 0.003318073693662882, + 0.003823102917522192, + 0.010769208893179893, + 0.006044548470526934, + 0.006133140064775944, + 0.01141579169780016, + 0.005181743763387203, + 0.005913970060646534, + 0.005354756023734808, + 0.006341238506138325, + 0.006185770966112614, + 0.006013019010424614, + 0.0036117020063102245, + 0.00405881879851222, + 0.0029513344634324312, + 0.00588965741917491, + 0.0029387264512479305, + 0.002655553864315152, + 0.004083271604031324, + 0.006035344675183296, + 0.0055466401390731335, + 0.005008498672395945, + 0.005152139347046614, + 0.004917957354336977, + 0.004764268174767494, + 0.005143300164490938, + 0.005660225637257099, + 0.03980689495801926, + 0.022060686722397804, + 0.030450953170657158, + 0.01782841421663761, + 0.02138642966747284, + 0.014195808209478855, + 0.015658237040042877, + 0.0326988585293293, + 0.0681682825088501, + 0.014189924113452435, + 0.006846122443675995, + 0.006543416529893875, + 0.008785823360085487, + 0.0059823160991072655, + 0.009055701084434986, + 0.06161022558808327 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 9 + ], + "selected_reward": -0.023302050307393074 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned at a static, ground-level macro perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a gnarled ginseng root shaped like a sleeping figure with twisted limbs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the scene feature a soft beam of light creating a black-and-white contrast effect?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do crystal-like water droplets fall rhythmically, glint like gems, and darken the soil upon impact?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video show the root swelling and a tiny green sprout pushing upward to unfold two cotyledons?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.47425490617752075 + ], + "opposite_logprob_tokens": [ + -0.9742549061775208 + ], + "logprob_reward": -0.47425490617752075, + "yes_no_margin_reward": 0.5, + "token_credit_reward": -0.47425490617752075, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.26894858479499817, + 0.11767596751451492, + 0.030098935589194298, + 0.23147284984588623, + 0.1185450479388237, + 0.02934940718114376, + 0.09877606481313705, + 0.020320937037467957, + 0.026480820029973984, + 0.058331385254859924 + ], + "credit_map": [ + 0.017786087468266487, + 0.01808636263012886, + 0.02828587219119072, + 0.015452981926500797, + 0.018457157537341118, + 0.0825926885008812, + 0.014116373844444752, + 0.07417105883359909, + 0.011926691047847271, + 0.009258786216378212, + 0.015546132810413837, + 0.016439126804471016, + 0.009824429638683796, + 0.025655807927250862, + 0.017310982570052147, + 0.011714008636772633, + 0.006099274381995201, + 0.003017457900568843, + 0.0031022566836327314, + 0.0030471317004412413, + 0.003549128770828247, + 0.002157360315322876, + 0.004324033856391907, + 0.00480229128152132, + 0.04747919365763664, + 0.018817730247974396, + 0.03516370430588722, + 0.03386092185974121, + 0.01822900027036667, + 0.03296026214957237, + 0.017244059592485428, + 0.02771797776222229, + 0.03659651055932045, + 0.006491858046501875, + 0.013648995198309422, + 0.009180314838886261, + 0.00508660264313221, + 0.004589714575558901, + 0.008169599808752537, + 0.034781455993652344, + 0.008428267203271389, + 0.0024142551701515913, + 0.003308675019070506, + 0.0018336041830480099, + 0.0026031811721622944, + 0.0019454677822068334, + 0.0020193681120872498, + 0.006796590052545071, + 0.020469877868890762, + 0.011981015093624592, + 0.013952990993857384, + 0.008034206926822662, + 0.012798615731298923, + 0.005238595884293318, + 0.005831941030919552, + 0.020468819886446, + 0.0021568662486970425, + 0.0025682232808321714, + 0.0032972567714750767, + 0.0023137342650443316, + 0.00340823270380497, + 0.0016404936322942376, + 0.002785029821097851, + 0.0021511020604521036, + 0.0028196857310831547, + 0.003074109787121415, + 0.005069309379905462, + 0.002988953609019518, + 0.0026710270904004574, + 0.0025280839763581753, + 0.0053612119518220425, + 0.001968438969925046, + 0.02189004421234131, + 0.0036378493532538414, + 0.0032688749488443136, + 0.0022078107576817274, + 0.001803674502298236, + 0.0019628959707915783, + 0.003120352514088154, + 0.020439887419342995 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 4 + ], + "selected_reward": -0.47425490617752075 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned at a static, ground-level macro perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a gnarled ginseng root shaped like a sleeping figure with twisted limbs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the scene feature a soft beam of light creating a black-and-white contrast effect?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do crystal-like water droplets fall rhythmically, glint like gems, and darken the soil upon impact?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video show the root swelling and a tiny green sprout pushing upward to unfold two cotyledons?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.012541474774479866 + ], + "opposite_logprob_tokens": [ + -4.3875412940979 + ], + "logprob_reward": -0.012541474774479866, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.012541474774479866, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.18607667088508606, + 0.07846031337976456, + 0.07304800301790237, + 0.16760659217834473, + 0.08685111254453659, + 0.10397536307573318, + 0.11823484301567078, + 0.025835342705249786, + 0.03660549223423004, + 0.12330624461174011 + ], + "credit_map": [ + 0.02106359228491783, + 0.009100225754082203, + 0.0674530565738678, + 0.008774716407060623, + 0.008224748075008392, + 0.008358520455658436, + 0.012915922328829765, + 0.05018588528037071, + 0.006824167910963297, + 0.006334555801004171, + 0.01474955677986145, + 0.008722218684852123, + 0.005850784946233034, + 0.02082020230591297, + 0.008211012929677963, + 0.006947821471840143, + 0.009412773884832859, + 0.007349593099206686, + 0.005984727293252945, + 0.006285075098276138, + 0.008273608051240444, + 0.02058279700577259, + 0.00661615002900362, + 0.00854327343404293, + 0.03276120126247406, + 0.01897578500211239, + 0.019665082916617393, + 0.02121010236442089, + 0.02422410435974598, + 0.021142728626728058, + 0.016977129504084587, + 0.012650465592741966, + 0.020314594730734825, + 0.00593885313719511, + 0.017969980835914612, + 0.006975883152335882, + 0.004440996330231428, + 0.004320062696933746, + 0.004793553613126278, + 0.022097188979387283, + 0.021763278171420097, + 0.009352684020996094, + 0.015229038894176483, + 0.009000466205179691, + 0.014993193559348583, + 0.010397439822554588, + 0.008356833830475807, + 0.014882434159517288, + 0.02285633608698845, + 0.011585635133087635, + 0.011450551450252533, + 0.009756610728800297, + 0.021738240495324135, + 0.007781181018799543, + 0.006885064300149679, + 0.026181228458881378, + 0.004097972065210342, + 0.0037643059622496367, + 0.0033993027172982693, + 0.002741091186180711, + 0.0037572775036096573, + 0.0021811816841363907, + 0.001977235311642289, + 0.003916975110769272, + 0.004263320472091436, + 0.0054936958476901054, + 0.005050132982432842, + 0.004178663715720177, + 0.004030338954180479, + 0.0038352457340806723, + 0.005696599837392569, + 0.0040574935264885426, + 0.04473009705543518, + 0.008950374089181423, + 0.00566174415871501, + 0.0046292198821902275, + 0.0052136764861643314, + 0.004318530671298504, + 0.008579451590776443, + 0.04122314974665642 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 9 + ], + "selected_reward": -0.012541474774479866 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned at a static, ground-level macro perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a gnarled ginseng root shaped like a sleeping figure with twisted limbs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the scene feature a soft beam of light creating a black-and-white contrast effect?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do crystal-like water droplets fall rhythmically, glint like gems, and darken the soil upon impact?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video show the root swelling and a tiny green sprout pushing upward to unfold two cotyledons?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01818285696208477 + ], + "opposite_logprob_tokens": [ + -4.018182754516602 + ], + "logprob_reward": -0.01818285696208477, + "yes_no_margin_reward": 4.0, + "token_credit_reward": -0.01818285696208477, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.058176733553409576, + 0.22426767647266388, + 0.05132772773504257, + 0.07291053235530853, + 0.13386930525302887, + 0.12731605768203735, + 0.04539265111088753, + 0.05645814165472984, + 0.1297304779291153, + 0.10055078566074371 + ], + "credit_map": [ + 0.006491909734904766, + 0.005854305345565081, + 0.014556764625012875, + 0.0067313904874026775, + 0.005233529023826122, + 0.00788359995931387, + 0.006093484815210104, + 0.005331747233867645, + 0.018758872523903847, + 0.013015280477702618, + 0.04601462185382843, + 0.025358885526657104, + 0.013001174665987492, + 0.06618601828813553, + 0.023820703849196434, + 0.018112098798155785, + 0.013610764406621456, + 0.005222144071012735, + 0.004385209642350674, + 0.0033720212522894144, + 0.004005189053714275, + 0.013255739584565163, + 0.0029537349473685026, + 0.0045229229144752026, + 0.007412164006382227, + 0.00605084327980876, + 0.01202603429555893, + 0.016140347346663475, + 0.006938071455806494, + 0.007075607310980558, + 0.006509957369416952, + 0.010757499374449253, + 0.02038181573152542, + 0.01154752541333437, + 0.015653930604457855, + 0.020322361961007118, + 0.023571277037262917, + 0.01586485467851162, + 0.013781492598354816, + 0.012746038846671581, + 0.02903062291443348, + 0.013154852204024792, + 0.019689079374074936, + 0.009647659957408905, + 0.008497633039951324, + 0.01137152872979641, + 0.012157578952610493, + 0.023767083883285522, + 0.006901702843606472, + 0.006016470491886139, + 0.006385790184140205, + 0.004243373870849609, + 0.005667397752404213, + 0.002977578202262521, + 0.003432287834584713, + 0.009768046438694, + 0.010165530256927013, + 0.005023097153753042, + 0.005518228746950626, + 0.004527322482317686, + 0.008424774743616581, + 0.005900138523429632, + 0.007672077044844627, + 0.009226965717971325, + 0.022433212026953697, + 0.00888834148645401, + 0.023055467754602432, + 0.013390567153692245, + 0.013704654760658741, + 0.00864960066974163, + 0.02006925456225872, + 0.019539369270205498, + 0.039386868476867676, + 0.00584065867587924, + 0.004166978877037764, + 0.003558559576049447, + 0.0035079855006188154, + 0.0036432521883398294, + 0.004300466738641262, + 0.03614601492881775 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 4, + 8 + ], + "selected_reward": -0.01818285696208477 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "Unfolding in an ultra-detailed artistic style, this mesmerizing time-lapse, captured from static, ground-level macro footage, immerses the viewer in the scene. The camera closely follows a gnarled ginseng root, shaped like a sleeping figure with twisted limbs, nestled in dark, rich, fertile soil dotted with tiny pebbles. A soft beam of light penetrates the darkness from above, creating a dramatic black-and-white contrast effect that meticulously highlights the root's woody, fibrous skin, deep cracks, and tiny root hairs. Crystal-like water droplets begin to fall in a rhythmic, deliberate sequence, impacting the soil and roots; each drop glints like a gem in the light before being greedily absorbed, darkening the earth and tracing a shimmering path along the root's contours. As time passes faster, the root visibly plumps and swells with latent energy, its color deepening to a healthier, more saturated brown. The sequence climaxes when the surface of the soil just above the root “head” begins to crack and bulge, and then a tiny, vibrant green sprout miraculously forces its way through the earth, steadily pushing upward and unfolding two tiny, shimmering cotyledons that catch the light, creating an atmosphere of quiet yet profound wonder at the powerful magic of life renewal.", + "seed": 1073, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000071/group_14/index_1504_seed_1073_dp_29_rank_58_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 14, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.07700091945007444, + "vlm_logprob": -0.07700091945007444, + "vlm_yes_no_margin": 3.599999952316284, + "vlm_token_credit": -0.07700091945007444, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned at a static, ground-level macro perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a gnarled ginseng root shaped like a sleeping figure with twisted limbs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the scene feature a soft beam of light creating a black-and-white contrast effect?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do crystal-like water droplets fall rhythmically, glint like gems, and darken the soil upon impact?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video show the root swelling and a tiny green sprout pushing upward to unfold two cotyledons?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03500382602214813 + ], + "opposite_logprob_tokens": [ + -5.285003662109375 + ], + "logprob_reward": -0.03500382602214813, + "yes_no_margin_reward": 5.25, + "token_credit_reward": -0.03500382602214813, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.24019958078861237, + 0.025431307032704353, + 0.3451383411884308, + 0.07583758980035782, + 0.08969029784202576, + 0.04454342648386955, + 0.018274718895554543, + 0.03135484457015991, + 0.07219937443733215, + 0.05733054131269455 + ], + "credit_map": [ + 0.023439114913344383, + 0.02423885092139244, + 0.032859113067388535, + 0.026468336582183838, + 0.021919742226600647, + 0.01854141429066658, + 0.060917891561985016, + 0.03181511536240578, + 0.0038536922074854374, + 0.0033543503377586603, + 0.004504385869950056, + 0.0017991425702348351, + 0.0038297700230032206, + 0.0019718001130968332, + 0.0023346447851508856, + 0.003783521708101034, + 0.14183875918388367, + 0.011546224355697632, + 0.013732731342315674, + 0.006641419138759375, + 0.0078868567943573, + 0.010270172730088234, + 0.01352654304355383, + 0.1396956592798233, + 0.012670586816966534, + 0.007357244845479727, + 0.014835570938885212, + 0.003945613745599985, + 0.011413074098527431, + 0.007867823354899883, + 0.005960556212812662, + 0.011787112802267075, + 0.0121776657178998, + 0.008194097317755222, + 0.031089430674910545, + 0.009003630839288235, + 0.006254967302083969, + 0.005256555043160915, + 0.007762001361697912, + 0.009951951913535595, + 0.004902829881757498, + 0.0071951849386096, + 0.007733485195785761, + 0.004608340561389923, + 0.004923322703689337, + 0.004967115819454193, + 0.005804811138659716, + 0.004408337641507387, + 0.003595026209950447, + 0.003223602892830968, + 0.003153638681396842, + 0.0012625199742615223, + 0.0015641212230548263, + 0.0010416716104373336, + 0.0018308177823200822, + 0.0026033218018710613, + 0.005572343710809946, + 0.003649772610515356, + 0.005065823905169964, + 0.0022984715178608894, + 0.00410419749096036, + 0.002433235989883542, + 0.0030971709638834, + 0.005133828613907099, + 0.009222950786352158, + 0.01361691765487194, + 0.010518272407352924, + 0.007608287036418915, + 0.008304407820105553, + 0.007642448879778385, + 0.007272555027157068, + 0.008013532496988773, + 0.010438107885420322, + 0.004089726135134697, + 0.013542738743126392, + 0.003249089466407895, + 0.006721421144902706, + 0.004332836717367172, + 0.005894334055483341, + 0.009062284603714943 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 0, + 4 + ], + "selected_reward": -0.03500382602214813 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned at a static, ground-level macro perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a gnarled ginseng root shaped like a sleeping figure with twisted limbs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the scene feature a soft beam of light creating a black-and-white contrast effect?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do crystal-like water droplets fall rhythmically, glint like gems, and darken the soil upon impact?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video show the root swelling and a tiny green sprout pushing upward to unfold two cotyledons?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.018197722733020782 + ], + "opposite_logprob_tokens": [ + -4.018197536468506 + ], + "logprob_reward": -0.018197722733020782, + "yes_no_margin_reward": 3.999999761581421, + "token_credit_reward": -0.018197722733020782, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.25472554564476013, + 0.05692176893353462, + 0.10245966911315918, + 0.03897344321012497, + 0.07020430266857147, + 0.06504616141319275, + 0.05258236825466156, + 0.17131125926971436, + 0.12101446837186813, + 0.06676101684570312 + ], + "credit_map": [ + 0.02413569949567318, + 0.021304795518517494, + 0.042180951684713364, + 0.02660573646426201, + 0.03287866339087486, + 0.021338723599910736, + 0.057919200509786606, + 0.028361767530441284, + 0.009904579259455204, + 0.008770897053182125, + 0.009668943472206593, + 0.004619054961949587, + 0.005164599511772394, + 0.004766969010233879, + 0.005704393610358238, + 0.008322334848344326, + 0.03994426876306534, + 0.0038910324219614267, + 0.004561782814562321, + 0.002820112742483616, + 0.0035542757250368595, + 0.004120043944567442, + 0.00431393925100565, + 0.03925422206521034, + 0.005672740750014782, + 0.004340456798672676, + 0.008822889067232609, + 0.003565483959391713, + 0.0047604781575500965, + 0.003007151186466217, + 0.004080434329807758, + 0.004723809659481049, + 0.009630462154746056, + 0.014162111096084118, + 0.009094462729990482, + 0.006775077898055315, + 0.008879859000444412, + 0.004391551949083805, + 0.010208025574684143, + 0.007062750402837992, + 0.008583320304751396, + 0.013826068490743637, + 0.011447032913565636, + 0.004628471098840237, + 0.008397066034376621, + 0.003675489453598857, + 0.00901744980365038, + 0.005471263080835342, + 0.007707187440246344, + 0.0037234679330140352, + 0.006525807082653046, + 0.007460198365151882, + 0.006650262046605349, + 0.006098336074501276, + 0.00659151142463088, + 0.007825598120689392, + 0.03271327540278435, + 0.013860343024134636, + 0.021064849570393562, + 0.011678178794682026, + 0.028181858360767365, + 0.012770882807672024, + 0.016087165102362633, + 0.034954704344272614, + 0.021991197019815445, + 0.010952532291412354, + 0.01364950556308031, + 0.010950184427201748, + 0.01801065169274807, + 0.011379352770745754, + 0.013264144770801067, + 0.020816899836063385, + 0.009073076769709587, + 0.006499195471405983, + 0.020126936957240105, + 0.005246814806014299, + 0.006087301764637232, + 0.003939132206141949, + 0.008833746425807476, + 0.006954815238714218 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 7, + 8 + ], + "selected_reward": -0.018197722733020782 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned at a static, ground-level macro perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a gnarled ginseng root shaped like a sleeping figure with twisted limbs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the scene feature a soft beam of light creating a black-and-white contrast effect?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do crystal-like water droplets fall rhythmically, glint like gems, and darken the soil upon impact?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video show the root swelling and a tiny green sprout pushing upward to unfold two cotyledons?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.22555790841579437 + ], + "opposite_logprob_tokens": [ + -1.6005579233169556 + ], + "logprob_reward": -0.22555790841579437, + "yes_no_margin_reward": 1.375, + "token_credit_reward": -0.22555790841579437, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09989432245492935, + 0.027091169729828835, + 0.45792046189308167, + 0.03047811985015869, + 0.061778582632541656, + 0.07115280628204346, + 0.020123785361647606, + 0.1454572230577469, + 0.0679808035492897, + 0.018122663721442223 + ], + "credit_map": [ + 0.012317026033997536, + 0.012810018844902515, + 0.012320108711719513, + 0.010963807813823223, + 0.009000108577311039, + 0.009528632275760174, + 0.0171931404620409, + 0.0157614778727293, + 0.005221682135015726, + 0.0032891505397856236, + 0.005565281957387924, + 0.0017602822044864297, + 0.002474659588187933, + 0.0024762535467743874, + 0.0026252800598740578, + 0.003678580978885293, + 0.17690840363502502, + 0.014956945553421974, + 0.020062588155269623, + 0.009627469815313816, + 0.01814841851592064, + 0.018961165100336075, + 0.019315112382173538, + 0.1799403578042984, + 0.008614358492195606, + 0.0022557575721293688, + 0.004014082718640566, + 0.0015219022752717137, + 0.002491515362635255, + 0.0021532594691962004, + 0.0019586377311497927, + 0.007468605879694223, + 0.008460898883640766, + 0.006811977364122868, + 0.019742436707019806, + 0.006290064658969641, + 0.004081436898559332, + 0.0031769846100360155, + 0.0066225165501236916, + 0.0065922727808356285, + 0.007626725360751152, + 0.011654303409159184, + 0.008879073895514011, + 0.007745749782770872, + 0.00959047395735979, + 0.010483352467417717, + 0.00774144334718585, + 0.007431684993207455, + 0.0029902313835918903, + 0.002777265850454569, + 0.004932713694870472, + 0.0017914061900228262, + 0.002269905526190996, + 0.0014294878346845508, + 0.002057976322248578, + 0.0018747977446764708, + 0.015062067657709122, + 0.018508892506361008, + 0.021810950711369514, + 0.012586846016347408, + 0.03360339626669884, + 0.01654229126870632, + 0.012773379683494568, + 0.014569394290447235, + 0.0077411336824297905, + 0.010216491296887398, + 0.012457609176635742, + 0.0055090817622840405, + 0.014803843572735786, + 0.004946254659444094, + 0.00516513641923666, + 0.007141253910958767, + 0.002079571830108762, + 0.0016885624499991536, + 0.00377063499763608, + 0.001507813693024218, + 0.002954502124339342, + 0.0016268643084913492, + 0.0027878221590071917, + 0.0017068921588361263 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 7, + 0 + ], + "selected_reward": -0.22555790841579437 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned at a static, ground-level macro perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a gnarled ginseng root shaped like a sleeping figure with twisted limbs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the scene feature a soft beam of light creating a black-and-white contrast effect?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do crystal-like water droplets fall rhythmically, glint like gems, and darken the soil upon impact?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video show the root swelling and a tiny green sprout pushing upward to unfold two cotyledons?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0059514218010008335 + ], + "opposite_logprob_tokens": [ + -5.130951404571533 + ], + "logprob_reward": -0.0059514218010008335, + "yes_no_margin_reward": 5.125, + "token_credit_reward": -0.0059514218010008335, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14599113166332245, + 0.06672418862581253, + 0.04454921558499336, + 0.2226625233888626, + 0.06235511600971222, + 0.10652752965688705, + 0.019349824637174606, + 0.09683380275964737, + 0.1725998818874359, + 0.062406737357378006 + ], + "credit_map": [ + 0.017181674018502235, + 0.013878482393920422, + 0.02291354350745678, + 0.01928163878619671, + 0.011056057177484035, + 0.011390206404030323, + 0.03305094316601753, + 0.017238589003682137, + 0.010972420684993267, + 0.010239270515739918, + 0.012318119406700134, + 0.0059666442684829235, + 0.005461720284074545, + 0.0051469323225319386, + 0.006757466588169336, + 0.009861619211733341, + 0.006436960771679878, + 0.00493840454146266, + 0.005309342406690121, + 0.004708270076662302, + 0.005261522252112627, + 0.006016868632286787, + 0.0032027126289904118, + 0.008675138466060162, + 0.034257061779499054, + 0.03058844804763794, + 0.032353099435567856, + 0.016559818759560585, + 0.0324050672352314, + 0.01920941472053528, + 0.021303780376911163, + 0.035985853523015976, + 0.0075095887295901775, + 0.005420506000518799, + 0.013384750112891197, + 0.009194795042276382, + 0.007341240532696247, + 0.005470547825098038, + 0.0076886373572051525, + 0.006345055997371674, + 0.015551429241895676, + 0.013351612724363804, + 0.0171027984470129, + 0.011026536114513874, + 0.01588897407054901, + 0.007166081108152866, + 0.0170288048684597, + 0.009411303326487541, + 0.0023178556002676487, + 0.0022229368332773447, + 0.0033856919035315514, + 0.0018543120240792632, + 0.001680681249126792, + 0.001898020738735795, + 0.003329192753881216, + 0.002661134349182248, + 0.011412151157855988, + 0.008818158879876137, + 0.014029734767973423, + 0.007406451273709536, + 0.024057364091277122, + 0.006781446747481823, + 0.01129046268761158, + 0.013038048520684242, + 0.021123575046658516, + 0.02152971737086773, + 0.0285979975014925, + 0.01306703221052885, + 0.03789718821644783, + 0.011528818868100643, + 0.01706787571310997, + 0.021787673234939575, + 0.006834619212895632, + 0.005753870587795973, + 0.012198769487440586, + 0.005958834197372198, + 0.012539590708911419, + 0.006362132262438536, + 0.00798235647380352, + 0.00477657001465559 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 8, + 0 + ], + "selected_reward": -0.0059514218010008335 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned at a static, ground-level macro perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a gnarled ginseng root shaped like a sleeping figure with twisted limbs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the scene feature a soft beam of light creating a black-and-white contrast effect?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do crystal-like water droplets fall rhythmically, glint like gems, and darken the soil upon impact?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video show the root swelling and a tiny green sprout pushing upward to unfold two cotyledons?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.10029371827840805 + ], + "opposite_logprob_tokens": [ + -2.3502936363220215 + ], + "logprob_reward": -0.10029371827840805, + "yes_no_margin_reward": 2.25, + "token_credit_reward": -0.10029371827840805, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07531318813562393, + 0.07667018473148346, + 0.07125405967235565, + 0.10928069800138474, + 0.08115540444850922, + 0.1355004757642746, + 0.036156635731458664, + 0.1415300816297531, + 0.19937317073345184, + 0.07376603037118912 + ], + "credit_map": [ + 0.010452753864228725, + 0.00934260617941618, + 0.008057776838541031, + 0.007631172891706228, + 0.007581711746752262, + 0.011852540075778961, + 0.008205842226743698, + 0.012188788503408432, + 0.014181157574057579, + 0.01103707030415535, + 0.013705275021493435, + 0.006038763094693422, + 0.005973720457404852, + 0.006626646034419537, + 0.007879707030951977, + 0.011227844282984734, + 0.020128758624196053, + 0.00570940924808383, + 0.008209137246012688, + 0.005302548874169588, + 0.006878951098769903, + 0.005372243467718363, + 0.004327389877289534, + 0.01532562542706728, + 0.025342289358377457, + 0.009219788946211338, + 0.02187126688659191, + 0.005931854713708162, + 0.012619586661458015, + 0.0075205047614872456, + 0.006559732835739851, + 0.02021569013595581, + 0.01095464825630188, + 0.006848074961453676, + 0.009449574165046215, + 0.011092591099441051, + 0.010262351483106613, + 0.009565981104969978, + 0.013695434667170048, + 0.009286751970648766, + 0.016388103365898132, + 0.01776704378426075, + 0.023641031235456467, + 0.015175621025264263, + 0.017580077052116394, + 0.013657650910317898, + 0.017885863780975342, + 0.013405091129243374, + 0.0033507568296045065, + 0.003102552378550172, + 0.00815113540738821, + 0.002478585112839937, + 0.008062775246798992, + 0.002898051170632243, + 0.003932895604521036, + 0.004179883282631636, + 0.01699599251151085, + 0.012773408554494381, + 0.024460166692733765, + 0.007419911213219166, + 0.04054592549800873, + 0.006488045211881399, + 0.01523144543170929, + 0.017615197226405144, + 0.026753019541502, + 0.022261062636971474, + 0.0322292298078537, + 0.01398550346493721, + 0.042163997888565063, + 0.013205497525632381, + 0.01823635958135128, + 0.030538519844412804, + 0.010465381667017937, + 0.0073315599001944065, + 0.010068043135106564, + 0.005678567569702864, + 0.014831829816102982, + 0.0056129880249500275, + 0.009478281252086163, + 0.010299384593963623 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 7, + 5 + ], + "selected_reward": -0.10029371827840805 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "Unfolding in an ultra-detailed artistic style, this mesmerizing time-lapse, captured from static, ground-level macro footage, immerses the viewer in the scene. The camera closely follows a gnarled ginseng root, shaped like a sleeping figure with twisted limbs, nestled in dark, rich, fertile soil dotted with tiny pebbles. A soft beam of light penetrates the darkness from above, creating a dramatic black-and-white contrast effect that meticulously highlights the root's woody, fibrous skin, deep cracks, and tiny root hairs. Crystal-like water droplets begin to fall in a rhythmic, deliberate sequence, impacting the soil and roots; each drop glints like a gem in the light before being greedily absorbed, darkening the earth and tracing a shimmering path along the root's contours. As time passes faster, the root visibly plumps and swells with latent energy, its color deepening to a healthier, more saturated brown. The sequence climaxes when the surface of the soil just above the root “head” begins to crack and bulge, and then a tiny, vibrant green sprout miraculously forces its way through the earth, steadily pushing upward and unfolding two tiny, shimmering cotyledons that catch the light, creating an atmosphere of quiet yet profound wonder at the powerful magic of life renewal.", + "seed": 1074, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000071/group_14/index_1504_seed_1074_dp_29_rank_58_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 14, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.16146824285387992, + "vlm_logprob": -0.16146824285387992, + "vlm_yes_no_margin": 2.4, + "vlm_token_credit": -0.16146824285387992, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned at a static, ground-level macro perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a gnarled ginseng root shaped like a sleeping figure with twisted limbs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the scene feature a soft beam of light creating a black-and-white contrast effect?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do crystal-like water droplets fall rhythmically, glint like gems, and darken the soil upon impact?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video show the root swelling and a tiny green sprout pushing upward to unfold two cotyledons?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0303322933614254 + ], + "opposite_logprob_tokens": [ + -4.655332088470459 + ], + "logprob_reward": -0.0303322933614254, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.0303322933614254, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12544719874858856, + 0.12185356020927429, + 0.07178454101085663, + 0.09194223582744598, + 0.05305584520101547, + 0.15029920637607574, + 0.21511177718639374, + 0.06349293887615204, + 0.05487700179219246, + 0.05213572457432747 + ], + "credit_map": [ + 0.018546726554632187, + 0.010581393726170063, + 0.02761826664209366, + 0.012124964967370033, + 0.011560647748410702, + 0.011724922806024551, + 0.011479608714580536, + 0.02181066945195198, + 0.01650175079703331, + 0.013291936367750168, + 0.015309819020330906, + 0.02283155918121338, + 0.010383985005319118, + 0.015024770982563496, + 0.011760745197534561, + 0.016748985275626183, + 0.011227410286664963, + 0.006806833203881979, + 0.010286015458405018, + 0.009566998109221458, + 0.004436078481376171, + 0.009063390083611012, + 0.00827168207615614, + 0.012126129120588303, + 0.014509468339383602, + 0.0072011156007647514, + 0.016225753352046013, + 0.010794994421303272, + 0.009800846688449383, + 0.007704768795520067, + 0.011925316415727139, + 0.013779964298009872, + 0.007923949509859085, + 0.0038934797048568726, + 0.01026973593980074, + 0.007153166923671961, + 0.0040132515132427216, + 0.005851208698004484, + 0.005366976372897625, + 0.008584071882069111, + 0.021341629326343536, + 0.020528603345155716, + 0.027954116463661194, + 0.01366674154996872, + 0.01840028539299965, + 0.012309716083109379, + 0.012176897376775742, + 0.023921219632029533, + 0.05204279348254204, + 0.00772307813167572, + 0.025285959243774414, + 0.00904244463890791, + 0.00912062730640173, + 0.027113499119877815, + 0.017449041828513145, + 0.06733433157205582, + 0.015011229552328587, + 0.0038083321414887905, + 0.009801291860640049, + 0.004439962562173605, + 0.0055885459296405315, + 0.007745260838419199, + 0.00446114270016551, + 0.012637171894311905, + 0.0086975721642375, + 0.0043872082605957985, + 0.00898000132292509, + 0.005107757169753313, + 0.004723047371953726, + 0.006478224415332079, + 0.005942609626799822, + 0.010560576803982258, + 0.00842536985874176, + 0.0036868113093078136, + 0.009746714495122433, + 0.0052563827484846115, + 0.0044953119941055775, + 0.005415305029600859, + 0.005667504388839006, + 0.00944232102483511 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 5, + 0 + ], + "selected_reward": -0.0303322933614254 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned at a static, ground-level macro perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a gnarled ginseng root shaped like a sleeping figure with twisted limbs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the scene feature a soft beam of light creating a black-and-white contrast effect?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do crystal-like water droplets fall rhythmically, glint like gems, and darken the soil upon impact?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video show the root swelling and a tiny green sprout pushing upward to unfold two cotyledons?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.048699744045734406 + ], + "opposite_logprob_tokens": [ + -3.0486998558044434 + ], + "logprob_reward": -0.048699744045734406, + "yes_no_margin_reward": 3.0, + "token_credit_reward": -0.048699744045734406, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.24336326122283936, + 0.11580950021743774, + 0.06063879281282425, + 0.04562550038099289, + 0.17853455245494843, + 0.11782453209161758, + 0.07182250916957855, + 0.08096829801797867, + 0.028485527262091637, + 0.05692756921052933 + ], + "credit_map": [ + 0.03282243013381958, + 0.03663799911737442, + 0.03365891054272652, + 0.03418228030204773, + 0.01893726922571659, + 0.022305063903331757, + 0.029828611761331558, + 0.034990694373846054, + 0.013903397135436535, + 0.011084347032010555, + 0.015354890376329422, + 0.0299193374812603, + 0.009569196030497551, + 0.010892266407608986, + 0.011241724714636803, + 0.013844339177012444, + 0.01028649602085352, + 0.005646773148328066, + 0.004542824346572161, + 0.003940077032893896, + 0.0038716725539416075, + 0.009135986678302288, + 0.006696418393403292, + 0.016518542543053627, + 0.006644639186561108, + 0.004875997081398964, + 0.006096578203141689, + 0.006335421930998564, + 0.004628392867743969, + 0.003444298868998885, + 0.006952972617000341, + 0.006647196132689714, + 0.025512613356113434, + 0.012389793992042542, + 0.0245454553514719, + 0.043823789805173874, + 0.013333109207451344, + 0.01341967098414898, + 0.019298480823636055, + 0.026211652904748917, + 0.019751569256186485, + 0.008910163305699825, + 0.022137444466352463, + 0.009875582531094551, + 0.011932114139199257, + 0.011615118943154812, + 0.007899529300630093, + 0.02570301480591297, + 0.01107829250395298, + 0.006104114931076765, + 0.00932234339416027, + 0.007400910835713148, + 0.004954289644956589, + 0.00585137540474534, + 0.013484534807503223, + 0.013626637868583202, + 0.018890393897891045, + 0.005399954505264759, + 0.011355380527675152, + 0.005355990491807461, + 0.0072844685055315495, + 0.0105441864579916, + 0.0052680629305541515, + 0.016869856044650078, + 0.0049332319758832455, + 0.002137170173227787, + 0.004772183485329151, + 0.0027851529885083437, + 0.0020476425997912884, + 0.003507176646962762, + 0.0034781470894813538, + 0.004824822302907705, + 0.008082631044089794, + 0.006241431459784508, + 0.009802096523344517, + 0.004718613810837269, + 0.0089929960668087, + 0.005184201058000326, + 0.0051202974282205105, + 0.008785300888121128 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 5 + ], + "selected_reward": -0.048699744045734406 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned at a static, ground-level macro perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a gnarled ginseng root shaped like a sleeping figure with twisted limbs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the scene feature a soft beam of light creating a black-and-white contrast effect?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do crystal-like water droplets fall rhythmically, glint like gems, and darken the soil upon impact?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video show the root swelling and a tiny green sprout pushing upward to unfold two cotyledons?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.25213295221328735 + ], + "opposite_logprob_tokens": [ + -1.5021328926086426 + ], + "logprob_reward": -0.25213295221328735, + "yes_no_margin_reward": 1.25, + "token_credit_reward": -0.25213295221328735, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1428086906671524, + 0.13316687941551208, + 0.10196840018033981, + 0.047693658620119095, + 0.10264579206705093, + 0.25374260544776917, + 0.07893317937850952, + 0.09031691402196884, + 0.014487103559076786, + 0.03423688933253288 + ], + "credit_map": [ + 0.018150538206100464, + 0.01997537538409233, + 0.02626563236117363, + 0.016852522268891335, + 0.008524631150066853, + 0.01647970639169216, + 0.014697168953716755, + 0.021863097324967384, + 0.017765110358595848, + 0.014961495995521545, + 0.02108917385339737, + 0.016124336048960686, + 0.014840352348983288, + 0.015321188606321812, + 0.013871797360479832, + 0.019193395972251892, + 0.018180515617132187, + 0.005853976123034954, + 0.008483120240271091, + 0.006302844267338514, + 0.00538401585072279, + 0.01701042428612709, + 0.010464671067893505, + 0.0302888173609972, + 0.007346696220338345, + 0.0034305148292332888, + 0.00793512538075447, + 0.006221096497029066, + 0.0030755249317735434, + 0.004570410121232271, + 0.008250316604971886, + 0.006863968446850777, + 0.016788840293884277, + 0.006220945622771978, + 0.016393564641475677, + 0.013929265551269054, + 0.005931580439209938, + 0.008921210654079914, + 0.008784584701061249, + 0.025675786659121513, + 0.03675910830497742, + 0.022609522566199303, + 0.02892429009079933, + 0.07077732682228088, + 0.01745818927884102, + 0.01833989843726158, + 0.03431021049618721, + 0.024564044550061226, + 0.019272444769740105, + 0.003980053588747978, + 0.007964776828885078, + 0.0038156535010784864, + 0.0053036510944366455, + 0.009965622797608376, + 0.005072097294032574, + 0.02355886995792389, + 0.01481590699404478, + 0.007634457666426897, + 0.013965773396193981, + 0.007998519577085972, + 0.0063555133529007435, + 0.007716781925410032, + 0.008552199229598045, + 0.023277748376131058, + 0.0027841953560709953, + 0.001167191774584353, + 0.00212254305370152, + 0.0016277929535135627, + 0.0008840896771289408, + 0.0017065968131646514, + 0.0014411992160603404, + 0.0027534926775842905, + 0.005426901858299971, + 0.002649767091497779, + 0.005883974954485893, + 0.003459970699623227, + 0.003049085382372141, + 0.0032494801562279463, + 0.004251497332006693, + 0.006266208831220865 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 0, + 1 + ], + "selected_reward": -0.25213295221328735 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned at a static, ground-level macro perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a gnarled ginseng root shaped like a sleeping figure with twisted limbs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the scene feature a soft beam of light creating a black-and-white contrast effect?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do crystal-like water droplets fall rhythmically, glint like gems, and darken the soil upon impact?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video show the root swelling and a tiny green sprout pushing upward to unfold two cotyledons?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.08907075971364975 + ], + "opposite_logprob_tokens": [ + -2.4640707969665527 + ], + "logprob_reward": -0.08907075971364975, + "yes_no_margin_reward": 2.375, + "token_credit_reward": -0.08907075971364975, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.05930369347333908, + 0.16853713989257812, + 0.07548241317272186, + 0.07687267661094666, + 0.22039204835891724, + 0.11345629394054413, + 0.1764758676290512, + 0.04806484654545784, + 0.019507424905896187, + 0.04190764203667641 + ], + "credit_map": [ + 0.00769734475761652, + 0.008701196871697903, + 0.009530733339488506, + 0.00664593605324626, + 0.006140566430985928, + 0.005671975202858448, + 0.0056272512301802635, + 0.009288686327636242, + 0.021059351041913033, + 0.016386745497584343, + 0.020166844129562378, + 0.03948066383600235, + 0.02144244685769081, + 0.015569443814456463, + 0.017196645960211754, + 0.017234979197382927, + 0.011540732346475124, + 0.007045284379273653, + 0.01332950685173273, + 0.008641362190246582, + 0.006139319855719805, + 0.006923176348209381, + 0.01058401633054018, + 0.011278999969363213, + 0.011854050680994987, + 0.006747706793248653, + 0.014772836118936539, + 0.008414540439844131, + 0.009104103781282902, + 0.006282857619225979, + 0.010385923087596893, + 0.009310651570558548, + 0.034905314445495605, + 0.014294520951807499, + 0.03013158217072487, + 0.05255252867937088, + 0.013064193539321423, + 0.0163290835916996, + 0.023832248523831367, + 0.035282544791698456, + 0.01457163318991661, + 0.01256032008677721, + 0.011527970433235168, + 0.017526859417557716, + 0.017998671159148216, + 0.009670240804553032, + 0.009242234751582146, + 0.02035835199058056, + 0.03176409751176834, + 0.010690857656300068, + 0.022200649604201317, + 0.013595713302493095, + 0.009804310277104378, + 0.01679818518459797, + 0.025989940389990807, + 0.045632097870111465, + 0.00728224404156208, + 0.0033432708587497473, + 0.006888538599014282, + 0.004963977728039026, + 0.006683976389467716, + 0.007103560492396355, + 0.004504327196627855, + 0.007294948678463697, + 0.0041569797322154045, + 0.0016170950839295983, + 0.0025569016579538584, + 0.001911953091621399, + 0.0009860468562692404, + 0.0018732299795374274, + 0.001803796854801476, + 0.004601419437676668, + 0.005704354029148817, + 0.004081360530108213, + 0.005198695231229067, + 0.0039252606220543385, + 0.007110603619366884, + 0.0056044054217636585, + 0.004053586628288031, + 0.0062293726950883865 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 6, + 1 + ], + "selected_reward": -0.08907075971364975 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned at a static, ground-level macro perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a gnarled ginseng root shaped like a sleeping figure with twisted limbs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the scene feature a soft beam of light creating a black-and-white contrast effect?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do crystal-like water droplets fall rhythmically, glint like gems, and darken the soil upon impact?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video show the root swelling and a tiny green sprout pushing upward to unfold two cotyledons?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.38710546493530273 + ], + "opposite_logprob_tokens": [ + -1.1371054649353027 + ], + "logprob_reward": -0.38710546493530273, + "yes_no_margin_reward": 0.75, + "token_credit_reward": -0.38710546493530273, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0417296402156353, + 0.1668042689561844, + 0.06334982067346573, + 0.06653670966625214, + 0.1669033169746399, + 0.16708233952522278, + 0.19752146303653717, + 0.054217174649238586, + 0.012199502438306808, + 0.0636558011174202 + ], + "credit_map": [ + 0.0056213149800896645, + 0.0050977421924471855, + 0.005656118504703045, + 0.0064813499338924885, + 0.0036613158881664276, + 0.0038000179920345545, + 0.006473003886640072, + 0.004938780330121517, + 0.019509946927428246, + 0.01669529639184475, + 0.017325323075056076, + 0.04365907609462738, + 0.01873280107975006, + 0.014776983298361301, + 0.018872307613492012, + 0.017232529819011688, + 0.009853120893239975, + 0.004443993791937828, + 0.007172135170549154, + 0.0059915403835475445, + 0.005242132116109133, + 0.009358864277601242, + 0.00788877997547388, + 0.013399261049926281, + 0.01008070819079876, + 0.005724645685404539, + 0.010913008823990822, + 0.008387693203985691, + 0.007877313531935215, + 0.005609604530036449, + 0.009703340008854866, + 0.008240396156907082, + 0.025394368916749954, + 0.01107692625373602, + 0.023525897413492203, + 0.04128650203347206, + 0.010697560384869576, + 0.011739537119865417, + 0.01812276616692543, + 0.025059755891561508, + 0.025484411045908928, + 0.018559755757451057, + 0.03244192153215408, + 0.01275264285504818, + 0.02085844613611698, + 0.013111752457916737, + 0.0125151127576828, + 0.03135831281542778, + 0.04400693252682686, + 0.009764367714524269, + 0.022480914369225502, + 0.010963688604533672, + 0.01024673506617546, + 0.020751439034938812, + 0.020642930641770363, + 0.058664463460445404, + 0.007942568510770798, + 0.00355503405444324, + 0.007688285782933235, + 0.004942357074469328, + 0.00816643051803112, + 0.008238054811954498, + 0.004992225207388401, + 0.008692217990756035, + 0.001980731263756752, + 0.0005755355232395232, + 0.0022389658261090517, + 0.0016002270858734846, + 0.0007962021045386791, + 0.0012391010532155633, + 0.0009681040537543595, + 0.0028006352949887514, + 0.009750253520905972, + 0.005718574393540621, + 0.010164707899093628, + 0.006352961994707584, + 0.00575067475438118, + 0.006719517987221479, + 0.006206554360687733, + 0.01299255806952715 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 5, + 4 + ], + "selected_reward": -0.38710546493530273 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000071/group_15/index_tensor([1530, 1530])_group_15_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000071/group_15/index_tensor([1530, 1530])_group_15_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..871de2ce18378936e37df800c1aadeba55ee7e41 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000071/group_15/index_tensor([1530, 1530])_group_15_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "static, low-angle, wide shot, rendered in hyper-realistic 8K high-resolution CGI animation, establishes the immensity of a desolate alien world under the harsh, overexposed glare of twin suns suspended in a pale, wispy sky. At the center of the frame stands a hulking, bipedal primitive, its form a terrifying fusion of sinew and armor, its body covered in overlapping, shifting bony plates with a chalky, weathered texture that delicately rub against one another. As it raises a massive claw, the muscles beneath its leathery, dark-gray skin strain with immense force. A vibrant green essence begins to pulse rhythmically from its limbs, and light seeps through a network of cracks, casting an unnatural glow on its obsidian claws. With a swift, decisive downward arc, the creature slams its energetic fist into a discarded piece of machinery half-buried in ochre dust. The incredible force of the impact instantly shatters the rusted, angular metal frame, creating a realistic deformation of twisted steel. Dazzling metallic sparks erupt violently from the point of contact, each individual particle arcing in blinding light while green energy is released in dazzling flashes, creating a tense and powerful atmosphere that emphasizes the raw, untamed power of this alien behemoth.", + "seed": 11219, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000071/group_15/index_1530_seed_11219_dp_30_rank_60_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 15, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.03022417090833187, + "vlm_logprob": -0.03022417090833187, + "vlm_yes_no_margin": 4.05, + "vlm_token_credit": -0.03022417090833187, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static, low-angle, wide shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the environment a desolate alien world under the glare of twin suns in a pale, wispy sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the central subject a hulking bipedal creature with chalky, weathered bony plates and dark-gray skin?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the creature slam its fist into a discarded piece of machinery, causing it to shatter and emit metallic sparks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is there a vibrant green essence pulsing from the creature's limbs and casting an unnatural glow?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.058967556804418564 + ], + "opposite_logprob_tokens": [ + -3.1839675903320312 + ], + "logprob_reward": -0.058967556804418564, + "yes_no_margin_reward": 3.125, + "token_credit_reward": -0.058967556804418564, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1717548370361328, + 0.0826222226023674, + 0.034443482756614685, + 0.07339255511760712, + 0.1065259575843811, + 0.218626007437706, + 0.044249534606933594, + 0.14335831999778748, + 0.06292321532964706, + 0.06210392713546753 + ], + "credit_map": [ + 0.0316871702671051, + 0.02153940312564373, + 0.017897766083478928, + 0.013434401713311672, + 0.016357773914933205, + 0.022403797134757042, + 0.016929246485233307, + 0.031505268067121506, + 0.01148208323866129, + 0.007467371411621571, + 0.010786431841552258, + 0.009966602548956871, + 0.008672310970723629, + 0.010951437056064606, + 0.010564579628407955, + 0.01273140124976635, + 0.0030364743433892727, + 0.004890961106866598, + 0.0031037458684295416, + 0.006205461919307709, + 0.004580404609441757, + 0.003668247489258647, + 0.0038556167855858803, + 0.005102566443383694, + 0.00563226779922843, + 0.013226198963820934, + 0.00567067414522171, + 0.007306077517569065, + 0.012066256254911423, + 0.008864184841513634, + 0.008779439143836498, + 0.011847451329231262, + 0.009765478782355785, + 0.009947049431502819, + 0.01651008613407612, + 0.015201671048998833, + 0.015494123101234436, + 0.008862332440912724, + 0.019848311319947243, + 0.010896893218159676, + 0.027288345620036125, + 0.0330706387758255, + 0.027085060253739357, + 0.02384377270936966, + 0.030385084450244904, + 0.023447617888450623, + 0.026211796328425407, + 0.02729368768632412, + 0.0054018632508814335, + 0.004473005887120962, + 0.006492706015706062, + 0.0035632699728012085, + 0.0037669141311198473, + 0.00900247786194086, + 0.005221518687903881, + 0.006327776703983545, + 0.008774719201028347, + 0.021878555417060852, + 0.01473336759954691, + 0.018006736412644386, + 0.010909234173595905, + 0.03009789250791073, + 0.022391214966773987, + 0.01656658761203289, + 0.005283173639327288, + 0.003288246225565672, + 0.006338411942124367, + 0.004532538820058107, + 0.0069989836774766445, + 0.019765933975577354, + 0.006549933459609747, + 0.010165989398956299, + 0.006283085327595472, + 0.008106770925223827, + 0.006111346185207367, + 0.005967935547232628, + 0.01204073429107666, + 0.009065024554729462, + 0.007025182247161865, + 0.007503844797611237 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 0, + 7 + ], + "selected_reward": -0.058967556804418564 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static, low-angle, wide shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the environment a desolate alien world under the glare of twin suns in a pale, wispy sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the central subject a hulking bipedal creature with chalky, weathered bony plates and dark-gray skin?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the creature slam its fist into a discarded piece of machinery, causing it to shatter and emit metallic sparks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is there a vibrant green essence pulsing from the creature's limbs and casting an unnatural glow?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.05498848482966423 + ], + "opposite_logprob_tokens": [ + -2.929988384246826 + ], + "logprob_reward": -0.05498848482966423, + "yes_no_margin_reward": 2.875, + "token_credit_reward": -0.05498848482966423, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06705805659294128, + 0.1152578741312027, + 0.14949634671211243, + 0.10145705193281174, + 0.028958760201931, + 0.184149831533432, + 0.07346490025520325, + 0.16859067976474762, + 0.06545993685722351, + 0.04610655829310417 + ], + "credit_map": [ + 0.014591454528272152, + 0.00707578007131815, + 0.00591095769777894, + 0.008198142983019352, + 0.00529008312150836, + 0.008493408560752869, + 0.005432215053588152, + 0.012066016905009747, + 0.021589022129774094, + 0.011088402941823006, + 0.009893436916172504, + 0.012712598778307438, + 0.01161140389740467, + 0.016406746581196785, + 0.009155401960015297, + 0.022800855338573456, + 0.01418648473918438, + 0.016550404950976372, + 0.015887141227722168, + 0.02149816043674946, + 0.025314221158623695, + 0.020813269540667534, + 0.016593020409345627, + 0.018653640523552895, + 0.00842619501054287, + 0.017925089225172997, + 0.010517481714487076, + 0.010869203135371208, + 0.016634562984108925, + 0.012476805597543716, + 0.013301825150847435, + 0.01130589097738266, + 0.003361672628670931, + 0.0030181335750967264, + 0.005563278216868639, + 0.004013494122773409, + 0.0038939842488616705, + 0.0026353602297604084, + 0.0033047932665795088, + 0.0031680434476584196, + 0.018975794315338135, + 0.024256641045212746, + 0.0345500186085701, + 0.026056379079818726, + 0.015183540992438793, + 0.028599904850125313, + 0.018867462873458862, + 0.017660100013017654, + 0.007367510814219713, + 0.011991475708782673, + 0.011582030914723873, + 0.009710490703582764, + 0.007709695026278496, + 0.008445815183222294, + 0.008359473198652267, + 0.008298410102725029, + 0.010458170436322689, + 0.04075021296739578, + 0.014557390473783016, + 0.010637795552611351, + 0.015768107026815414, + 0.04684656858444214, + 0.010970107279717922, + 0.018602333962917328, + 0.005914269480854273, + 0.004301524255424738, + 0.01016678474843502, + 0.006771827582269907, + 0.005723126698285341, + 0.01544769387692213, + 0.006750039290636778, + 0.010384669527411461, + 0.0056839631870388985, + 0.004305657465010881, + 0.008912661112844944, + 0.00388211477547884, + 0.00646943598985672, + 0.005707607604563236, + 0.004102869424968958, + 0.007042250595986843 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 7, + 2 + ], + "selected_reward": -0.05498848482966423 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static, low-angle, wide shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the environment a desolate alien world under the glare of twin suns in a pale, wispy sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the central subject a hulking bipedal creature with chalky, weathered bony plates and dark-gray skin?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the creature slam its fist into a discarded piece of machinery, causing it to shatter and emit metallic sparks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is there a vibrant green essence pulsing from the creature's limbs and casting an unnatural glow?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.007630014326423407 + ], + "opposite_logprob_tokens": [ + -4.882629871368408 + ], + "logprob_reward": -0.007630014326423407, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.007630014326423407, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10021424293518066, + 0.061618730425834656, + 0.14778220653533936, + 0.12529870867729187, + 0.07783554494380951, + 0.1293029487133026, + 0.08014276623725891, + 0.1863144189119339, + 0.05798881873488426, + 0.03350157290697098 + ], + "credit_map": [ + 0.012108377180993557, + 0.017346903681755066, + 0.01046240795403719, + 0.011361989192664623, + 0.009822985157370567, + 0.012916367501020432, + 0.012171711772680283, + 0.01402350515127182, + 0.008190370164811611, + 0.007091321516782045, + 0.006528924684971571, + 0.010387899354100227, + 0.00552193121984601, + 0.008412866853177547, + 0.007574604358524084, + 0.007910814136266708, + 0.01381597202271223, + 0.017019497230648994, + 0.013909141533076763, + 0.021371228620409966, + 0.026091836392879486, + 0.0209856815636158, + 0.0165222380310297, + 0.01806662417948246, + 0.01022198423743248, + 0.02517755888402462, + 0.010956091806292534, + 0.014119550585746765, + 0.022985415533185005, + 0.014300611801445484, + 0.015516169369220734, + 0.012021324597299099, + 0.007096481043845415, + 0.011207601986825466, + 0.007272241171449423, + 0.009526357986032963, + 0.016350938007235527, + 0.0066755651496350765, + 0.013916267082095146, + 0.0057900999672710896, + 0.014323296956717968, + 0.019859137013554573, + 0.014124084264039993, + 0.012440435588359833, + 0.02238226681947708, + 0.012713616713881493, + 0.015976544469594955, + 0.017483584582805634, + 0.007962798699736595, + 0.012496293522417545, + 0.01462177187204361, + 0.00943608209490776, + 0.007071459665894508, + 0.008721336722373962, + 0.009498249739408493, + 0.010334785096347332, + 0.01113301794975996, + 0.031138520687818527, + 0.012613017112016678, + 0.022378014400601387, + 0.027716169133782387, + 0.021546347066760063, + 0.03913217782974243, + 0.02065715380012989, + 0.005953881423920393, + 0.0038599988911300898, + 0.006723917089402676, + 0.004324460867792368, + 0.005649308208376169, + 0.015474514104425907, + 0.0069584413431584835, + 0.009044302627444267, + 0.0037721022963523865, + 0.004955763462930918, + 0.005463010631501675, + 0.0031613728497177362, + 0.003643942065536976, + 0.004671860486268997, + 0.002834044862538576, + 0.00499948114156723 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 2, + 5 + ], + "selected_reward": -0.007630014326423407 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static, low-angle, wide shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the environment a desolate alien world under the glare of twin suns in a pale, wispy sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the central subject a hulking bipedal creature with chalky, weathered bony plates and dark-gray skin?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the creature slam its fist into a discarded piece of machinery, causing it to shatter and emit metallic sparks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is there a vibrant green essence pulsing from the creature's limbs and casting an unnatural glow?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02634294331073761 + ], + "opposite_logprob_tokens": [ + -3.6513428688049316 + ], + "logprob_reward": -0.02634294331073761, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.02634294331073761, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08052828907966614, + 0.04752020165324211, + 0.07353535294532776, + 0.035462893545627594, + 0.07402276247739792, + 0.04493346065282822, + 0.04979516193270683, + 0.1529768407344818, + 0.37703418731689453, + 0.06419084966182709 + ], + "credit_map": [ + 0.012442313134670258, + 0.01196384523063898, + 0.008681322447955608, + 0.009547572582960129, + 0.007702236529439688, + 0.010692212730646133, + 0.008774385787546635, + 0.010724405758082867, + 0.00791481789201498, + 0.004403306171298027, + 0.004801732953637838, + 0.005731082987040281, + 0.0049240258522331715, + 0.006492276675999165, + 0.004809387028217316, + 0.008443575352430344, + 0.0072373319417238235, + 0.008808677084743977, + 0.006298482418060303, + 0.010166208259761333, + 0.013538006693124771, + 0.009492048993706703, + 0.007495659403502941, + 0.010498940013349056, + 0.003304811893031001, + 0.004785723052918911, + 0.005413886159658432, + 0.004351454321295023, + 0.004413203336298466, + 0.004493587650358677, + 0.004990581423044205, + 0.0037096478044986725, + 0.007149655371904373, + 0.010893143713474274, + 0.006084475200623274, + 0.009422537870705128, + 0.01627950184047222, + 0.006435307674109936, + 0.011845403350889683, + 0.0059127360582351685, + 0.0033067865297198296, + 0.0089567257091403, + 0.006193375680595636, + 0.0062683518044650555, + 0.005016950890421867, + 0.004401054698973894, + 0.004527950659394264, + 0.006262264214456081, + 0.004379173740744591, + 0.009113455191254616, + 0.008306712843477726, + 0.005764992907643318, + 0.004856619518250227, + 0.00629468634724617, + 0.005426912102848291, + 0.005652613937854767, + 0.009723350405693054, + 0.03463083133101463, + 0.014763723127543926, + 0.013350103050470352, + 0.016726449131965637, + 0.03487200289964676, + 0.012602540664374828, + 0.01630786433815956, + 0.03354564681649208, + 0.03730713203549385, + 0.04798797145485878, + 0.029380645602941513, + 0.03565302863717079, + 0.06240606680512428, + 0.06751339137554169, + 0.06324032694101334, + 0.008065223693847656, + 0.007392058148980141, + 0.009376940317451954, + 0.005783780012279749, + 0.006463396362960339, + 0.012507723644375801, + 0.005739700049161911, + 0.008862026035785675 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 7, + 0 + ], + "selected_reward": -0.02634294331073761 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static, low-angle, wide shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the environment a desolate alien world under the glare of twin suns in a pale, wispy sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the central subject a hulking bipedal creature with chalky, weathered bony plates and dark-gray skin?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the creature slam its fist into a discarded piece of machinery, causing it to shatter and emit metallic sparks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is there a vibrant green essence pulsing from the creature's limbs and casting an unnatural glow?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0031918552704155445 + ], + "opposite_logprob_tokens": [ + -5.753191947937012 + ], + "logprob_reward": -0.0031918552704155445, + "yes_no_margin_reward": 5.75, + "token_credit_reward": -0.0031918552704155445, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11039081960916519, + 0.030409803614020348, + 0.025700341910123825, + 0.10273140668869019, + 0.07128465175628662, + 0.06533078104257584, + 0.11096370220184326, + 0.1588204950094223, + 0.26362451910972595, + 0.060743533074855804 + ], + "credit_map": [ + 0.017802860587835312, + 0.013085446320474148, + 0.011762176640331745, + 0.013953880406916142, + 0.011486527509987354, + 0.013386065140366554, + 0.012852567248046398, + 0.01606130041182041, + 0.004194563254714012, + 0.0033258069306612015, + 0.003213062882423401, + 0.0033642554190009832, + 0.002861644607037306, + 0.005110064055770636, + 0.0027353449258953333, + 0.0056050606071949005, + 0.003068910213187337, + 0.0029773961286991835, + 0.001996269216760993, + 0.0029426543042063713, + 0.0031923120841383934, + 0.003195783356204629, + 0.002253112616017461, + 0.0060739037580788136, + 0.00864699948579073, + 0.022206686437129974, + 0.009753627702593803, + 0.014167648740112782, + 0.019268518313765526, + 0.011147890239953995, + 0.007851241156458855, + 0.009688795544207096, + 0.008866796270012856, + 0.007916253060102463, + 0.011158348061144352, + 0.009204587899148464, + 0.010388623923063278, + 0.005914719775319099, + 0.010754202492535114, + 0.007081124931573868, + 0.010648501105606556, + 0.008087490685284138, + 0.008747728541493416, + 0.005492715630680323, + 0.007693287450820208, + 0.010580534115433693, + 0.005964566487818956, + 0.008115957491099834, + 0.014096615836024284, + 0.018207374960184097, + 0.013034837320446968, + 0.010289441794157028, + 0.008475965820252895, + 0.020009703934192657, + 0.013636757619678974, + 0.013213000260293484, + 0.010133559815585613, + 0.017507225275039673, + 0.011661834083497524, + 0.027523282915353775, + 0.021190328523516655, + 0.02121713012456894, + 0.03130314499139786, + 0.018283994868397713, + 0.022752394899725914, + 0.021437156945466995, + 0.021803436800837517, + 0.016373787075281143, + 0.026400133967399597, + 0.06992383301258087, + 0.04253227636218071, + 0.0424015074968338, + 0.006413796916604042, + 0.012005725875496864, + 0.005942121613770723, + 0.006239639595150948, + 0.007491616532206535, + 0.008922038599848747, + 0.0059703863225877285, + 0.007758211810141802 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 7, + 6 + ], + "selected_reward": -0.0031918552704155445 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "static, low-angle, wide shot, rendered in hyper-realistic 8K high-resolution CGI animation, establishes the immensity of a desolate alien world under the harsh, overexposed glare of twin suns suspended in a pale, wispy sky. At the center of the frame stands a hulking, bipedal primitive, its form a terrifying fusion of sinew and armor, its body covered in overlapping, shifting bony plates with a chalky, weathered texture that delicately rub against one another. As it raises a massive claw, the muscles beneath its leathery, dark-gray skin strain with immense force. A vibrant green essence begins to pulse rhythmically from its limbs, and light seeps through a network of cracks, casting an unnatural glow on its obsidian claws. With a swift, decisive downward arc, the creature slams its energetic fist into a discarded piece of machinery half-buried in ochre dust. The incredible force of the impact instantly shatters the rusted, angular metal frame, creating a realistic deformation of twisted steel. Dazzling metallic sparks erupt violently from the point of contact, each individual particle arcing in blinding light while green energy is released in dazzling flashes, creating a tense and powerful atmosphere that emphasizes the raw, untamed power of this alien behemoth.", + "seed": 11220, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000071/group_15/index_1530_seed_11220_dp_30_rank_60_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 15, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.04263142324052751, + "vlm_logprob": -0.04263142324052751, + "vlm_yes_no_margin": 3.85, + "vlm_token_credit": -0.04263142324052751, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static, low-angle, wide shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the environment a desolate alien world under the glare of twin suns in a pale, wispy sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the central subject a hulking bipedal creature with chalky, weathered bony plates and dark-gray skin?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the creature slam its fist into a discarded piece of machinery, causing it to shatter and emit metallic sparks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is there a vibrant green essence pulsing from the creature's limbs and casting an unnatural glow?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.1307242214679718 + ], + "opposite_logprob_tokens": [ + -2.5057241916656494 + ], + "logprob_reward": -0.1307242214679718, + "yes_no_margin_reward": 2.375, + "token_credit_reward": -0.1307242214679718, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.20562823116779327, + 0.08753059059381485, + 0.06385313719511032, + 0.08407275378704071, + 0.07149545103311539, + 0.24001410603523254, + 0.054919030517339706, + 0.1207096129655838, + 0.0460311695933342, + 0.025745848193764687 + ], + "credit_map": [ + 0.030105343088507652, + 0.04476957768201828, + 0.02847636118531227, + 0.013079964555799961, + 0.01696605421602726, + 0.030972639098763466, + 0.024209948256611824, + 0.017048342153429985, + 0.014134163968265057, + 0.02195613458752632, + 0.006732144858688116, + 0.0045695835724473, + 0.008072393015027046, + 0.012289589270949364, + 0.00861536618322134, + 0.011161228641867638, + 0.006788905709981918, + 0.00921917799860239, + 0.008871516212821007, + 0.005074201617389917, + 0.008914729580283165, + 0.01090994942933321, + 0.008175109513103962, + 0.005899551324546337, + 0.008397010155022144, + 0.007358967326581478, + 0.009743976406753063, + 0.009570375084877014, + 0.01645796000957489, + 0.012421726249158382, + 0.011499270796775818, + 0.008623477071523666, + 0.018983418121933937, + 0.006397181190550327, + 0.005585324950516224, + 0.0052743712440133095, + 0.004529311787337065, + 0.011705871671438217, + 0.007306202780455351, + 0.011713776737451553, + 0.010772951878607273, + 0.008030520752072334, + 0.02090074121952057, + 0.01779930107295513, + 0.02157483622431755, + 0.0352339930832386, + 0.087870754301548, + 0.03783103823661804, + 0.009321318008005619, + 0.0028308744076639414, + 0.010806497186422348, + 0.0039885337464511395, + 0.007251759525388479, + 0.004708437714725733, + 0.0112366434186697, + 0.004774970002472401, + 0.007897973991930485, + 0.009081642143428326, + 0.024805301800370216, + 0.01409758161753416, + 0.010675447061657906, + 0.011302335187792778, + 0.030400173738598824, + 0.012449157424271107, + 0.005696212872862816, + 0.004081516992300749, + 0.005908419378101826, + 0.006790671497583389, + 0.005402331706136465, + 0.005209648050367832, + 0.006709006614983082, + 0.006233368068933487, + 0.003516147844493389, + 0.002872898941859603, + 0.0031490640249103308, + 0.0035540845710784197, + 0.0035997279919683933, + 0.002462700242176652, + 0.0035631039645522833, + 0.0030281238723546267 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 0, + 7 + ], + "selected_reward": -0.1307242214679718 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static, low-angle, wide shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the environment a desolate alien world under the glare of twin suns in a pale, wispy sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the central subject a hulking bipedal creature with chalky, weathered bony plates and dark-gray skin?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the creature slam its fist into a discarded piece of machinery, causing it to shatter and emit metallic sparks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is there a vibrant green essence pulsing from the creature's limbs and casting an unnatural glow?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.023317305371165276 + ], + "opposite_logprob_tokens": [ + -3.773317337036133 + ], + "logprob_reward": -0.023317305371165276, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.023317305371165276, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.16762401163578033, + 0.08491961658000946, + 0.10832784324884415, + 0.04405410587787628, + 0.10983598977327347, + 0.03411206230521202, + 0.09754984825849533, + 0.28111371397972107, + 0.036465875804424286, + 0.035996872931718826 + ], + "credit_map": [ + 0.02647816203534603, + 0.042513664811849594, + 0.022621437907218933, + 0.01095276977866888, + 0.014141993597149849, + 0.024943390861153603, + 0.011487565003335476, + 0.014485049061477184, + 0.012347211129963398, + 0.021981284022331238, + 0.008892202749848366, + 0.0064547243528068066, + 0.00838907714933157, + 0.010152810253202915, + 0.0092439791187644, + 0.007458340842276812, + 0.011555694974958897, + 0.016376202926039696, + 0.01659187488257885, + 0.011174618266522884, + 0.013568875379860401, + 0.015580574050545692, + 0.013735766522586346, + 0.009744248352944851, + 0.00457442132756114, + 0.006472139619290829, + 0.005631774198263884, + 0.0067031653597950935, + 0.005916677415370941, + 0.004240829963237047, + 0.006494354456663132, + 0.004020749591290951, + 0.03086577169597149, + 0.010382932610809803, + 0.009161191992461681, + 0.005584148224443197, + 0.007329949643462896, + 0.020414119586348534, + 0.009843976236879826, + 0.016253914684057236, + 0.002702048048377037, + 0.0029086859431117773, + 0.004454358015209436, + 0.0032538496889173985, + 0.0031760174315422773, + 0.008264548145234585, + 0.005957276560366154, + 0.0033952819649130106, + 0.01785849779844284, + 0.004751137457787991, + 0.017810611054301262, + 0.006511025130748749, + 0.012933566235005856, + 0.009108159691095352, + 0.020526345819234848, + 0.00805051438510418, + 0.038842830806970596, + 0.028058171272277832, + 0.03899409994482994, + 0.04072274640202522, + 0.02820645458996296, + 0.02660709246993065, + 0.03438416123390198, + 0.04529818147420883, + 0.004248916171491146, + 0.004400697071105242, + 0.00399158475920558, + 0.006442003417760134, + 0.0039152526296675205, + 0.004148371983319521, + 0.005198146216571331, + 0.004120906349271536, + 0.0038179054390639067, + 0.003535030409693718, + 0.0047444687224924564, + 0.006996502168476582, + 0.00442669028416276, + 0.0035646252799779177, + 0.004629353526979685, + 0.004282302223145962 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 0, + 4 + ], + "selected_reward": -0.023317305371165276 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static, low-angle, wide shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the environment a desolate alien world under the glare of twin suns in a pale, wispy sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the central subject a hulking bipedal creature with chalky, weathered bony plates and dark-gray skin?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the creature slam its fist into a discarded piece of machinery, causing it to shatter and emit metallic sparks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is there a vibrant green essence pulsing from the creature's limbs and casting an unnatural glow?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.018193157389760017 + ], + "opposite_logprob_tokens": [ + -4.018193244934082 + ], + "logprob_reward": -0.018193157389760017, + "yes_no_margin_reward": 4.0, + "token_credit_reward": -0.018193157389760017, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.23247785866260529, + 0.03566412627696991, + 0.0858536958694458, + 0.12073177099227905, + 0.13706853985786438, + 0.14366625249385834, + 0.09496794641017914, + 0.08449540287256241, + 0.040352676063776016, + 0.024721719324588776 + ], + "credit_map": [ + 0.030695509165525436, + 0.03413502499461174, + 0.02963842637836933, + 0.014738191850483418, + 0.021922389045357704, + 0.04149944707751274, + 0.031109191477298737, + 0.028739668428897858, + 0.005073020234704018, + 0.008550480008125305, + 0.0031426791101694107, + 0.002780672162771225, + 0.0035515509080141783, + 0.004437768831849098, + 0.004732244648039341, + 0.0033957106061279774, + 0.011313839815557003, + 0.02457035519182682, + 0.00883607380092144, + 0.007690639700740576, + 0.010698555037379265, + 0.009732860140502453, + 0.007673912215977907, + 0.005337458569556475, + 0.012955216690897942, + 0.010605593211948872, + 0.02154059149324894, + 0.015982571989297867, + 0.017057018354535103, + 0.013786988332867622, + 0.01807219348847866, + 0.010731595568358898, + 0.03289109095931053, + 0.01213075965642929, + 0.013095894828438759, + 0.007480451837182045, + 0.010652073659002781, + 0.03165536746382713, + 0.014818516559898853, + 0.014344380237162113, + 0.010027097538113594, + 0.005934031214565039, + 0.01757749170064926, + 0.008642716333270073, + 0.009692313149571419, + 0.02572803944349289, + 0.03163568302989006, + 0.03442889451980591, + 0.010635633021593094, + 0.003879972267895937, + 0.014230659231543541, + 0.007258327677845955, + 0.010891937650740147, + 0.004731578752398491, + 0.03508986532688141, + 0.008249971084296703, + 0.00700739398598671, + 0.010500687174499035, + 0.014335104264318943, + 0.007505510468035936, + 0.00975838117301464, + 0.009097729809582233, + 0.012794404290616512, + 0.013496192172169685, + 0.003945167176425457, + 0.004139280412346125, + 0.005099468398839235, + 0.005097957793623209, + 0.005402156617492437, + 0.004045653622597456, + 0.005542306695133448, + 0.0070806825533509254, + 0.0033666533417999744, + 0.003209368558600545, + 0.0026146937161684036, + 0.00385567219927907, + 0.0034124432131648064, + 0.002379784593358636, + 0.003564810613170266, + 0.0023182914592325687 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 5, + 4 + ], + "selected_reward": -0.018193157389760017 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static, low-angle, wide shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the environment a desolate alien world under the glare of twin suns in a pale, wispy sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the central subject a hulking bipedal creature with chalky, weathered bony plates and dark-gray skin?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the creature slam its fist into a discarded piece of machinery, causing it to shatter and emit metallic sparks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is there a vibrant green essence pulsing from the creature's limbs and casting an unnatural glow?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03810246288776398 + ], + "opposite_logprob_tokens": [ + -3.288102388381958 + ], + "logprob_reward": -0.03810246288776398, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.03810246288776398, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07996882498264313, + 0.030003491789102554, + 0.05198473110795021, + 0.13232013583183289, + 0.08099937438964844, + 0.17355892062187195, + 0.04542890936136246, + 0.11108365654945374, + 0.19734883308410645, + 0.09730315208435059 + ], + "credit_map": [ + 0.012850349768996239, + 0.020628532394766808, + 0.007198868319392204, + 0.005525497253984213, + 0.00842161476612091, + 0.011485281400382519, + 0.008070364594459534, + 0.005788309499621391, + 0.005125103984028101, + 0.005899578798562288, + 0.0023353113792836666, + 0.002982878591865301, + 0.0031490440014749765, + 0.003736973274499178, + 0.003226496744900942, + 0.003548105014488101, + 0.005967331118881702, + 0.011167583055794239, + 0.00696924701333046, + 0.005488048307597637, + 0.005187229253351688, + 0.005890475120395422, + 0.006784757133573294, + 0.004530057311058044, + 0.012370062060654163, + 0.015828248113393784, + 0.011480063199996948, + 0.009300801903009415, + 0.027971310541033745, + 0.021397767588496208, + 0.022963322699069977, + 0.011008565314114094, + 0.00963290873914957, + 0.007250382099300623, + 0.008624372072517872, + 0.00447568716481328, + 0.006594066508114338, + 0.023232344537973404, + 0.007629314437508583, + 0.013560297898948193, + 0.013050204142928123, + 0.009726966731250286, + 0.014083142392337322, + 0.009823673404753208, + 0.014798790216445923, + 0.05285428464412689, + 0.027606375515460968, + 0.0316154882311821, + 0.0038894221652299166, + 0.002271606121212244, + 0.007196114398539066, + 0.003805183107033372, + 0.004796156659722328, + 0.0032823095098137856, + 0.014503302052617073, + 0.005684815347194672, + 0.012893369421362877, + 0.012341703288257122, + 0.014156116172671318, + 0.02084195427596569, + 0.01223073247820139, + 0.011451156809926033, + 0.013816644437611103, + 0.013351984322071075, + 0.01747567020356655, + 0.017736200243234634, + 0.022546976804733276, + 0.02995971031486988, + 0.028900763019919395, + 0.01715976744890213, + 0.04087917134165764, + 0.022690579295158386, + 0.01022693794220686, + 0.00892175268381834, + 0.011268964037299156, + 0.014973269775509834, + 0.014477064833045006, + 0.009842544794082642, + 0.016500486060976982, + 0.011092128232121468 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 5, + 3 + ], + "selected_reward": -0.03810246288776398 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static, low-angle, wide shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the environment a desolate alien world under the glare of twin suns in a pale, wispy sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the central subject a hulking bipedal creature with chalky, weathered bony plates and dark-gray skin?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the creature slam its fist into a discarded piece of machinery, causing it to shatter and emit metallic sparks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is there a vibrant green essence pulsing from the creature's limbs and casting an unnatural glow?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0028199690859764814 + ], + "opposite_logprob_tokens": [ + -5.877820014953613 + ], + "logprob_reward": -0.0028199690859764814, + "yes_no_margin_reward": 5.875, + "token_credit_reward": -0.0028199690859764814, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08849342912435532, + 0.027540935203433037, + 0.11403265595436096, + 0.07265714555978775, + 0.12655679881572723, + 0.1621103435754776, + 0.11003674566745758, + 0.18761517107486725, + 0.0720539316534996, + 0.038902830332517624 + ], + "credit_map": [ + 0.01072033029049635, + 0.011064553633332253, + 0.012761388905346394, + 0.007507274858653545, + 0.008586668409407139, + 0.012601700611412525, + 0.009362896904349327, + 0.01588861271739006, + 0.0044175307266414165, + 0.006178372073918581, + 0.0028766568284481764, + 0.001593672321178019, + 0.002586816903203726, + 0.0033831042237579823, + 0.0038470362778753042, + 0.0026577459648251534, + 0.013680129311978817, + 0.017521759495139122, + 0.0188125092536211, + 0.011410088278353214, + 0.014525674283504486, + 0.01562720537185669, + 0.013645246624946594, + 0.00881003774702549, + 0.007466226816177368, + 0.0074500092305243015, + 0.013293205760419369, + 0.009178437292575836, + 0.010584836825728416, + 0.007971663028001785, + 0.00969370175153017, + 0.007019066717475653, + 0.03175085410475731, + 0.013437044806778431, + 0.011554542928934097, + 0.006737085524946451, + 0.00871251430362463, + 0.013544950634241104, + 0.015386912040412426, + 0.025432899594306946, + 0.012119796127080917, + 0.007748860400170088, + 0.018593184649944305, + 0.00733320415019989, + 0.01025188248604536, + 0.039907194674015045, + 0.02440251037478447, + 0.04175371304154396, + 0.008136549964547157, + 0.0037020985037088394, + 0.024171937257051468, + 0.008534226566553116, + 0.010628588497638702, + 0.0052835107780992985, + 0.040036529302597046, + 0.009543301537632942, + 0.016172165051102638, + 0.017350872978568077, + 0.030934860929846764, + 0.022794213145971298, + 0.02256484143435955, + 0.022529173642396927, + 0.023421503603458405, + 0.03184753283858299, + 0.006739392876625061, + 0.005022787023335695, + 0.010338980704545975, + 0.010322757065296173, + 0.008609584532678127, + 0.006401956547051668, + 0.009594234637916088, + 0.015024236403405666, + 0.003820250742137432, + 0.003796954872086644, + 0.005098202731460333, + 0.007079206872731447, + 0.004777560476213694, + 0.004203746560961008, + 0.005327960941940546, + 0.0047989459708333015 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 5, + 4 + ], + "selected_reward": -0.0028199690859764814 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "static, low-angle, wide shot, rendered in hyper-realistic 8K high-resolution CGI animation, establishes the immensity of a desolate alien world under the harsh, overexposed glare of twin suns suspended in a pale, wispy sky. At the center of the frame stands a hulking, bipedal primitive, its form a terrifying fusion of sinew and armor, its body covered in overlapping, shifting bony plates with a chalky, weathered texture that delicately rub against one another. As it raises a massive claw, the muscles beneath its leathery, dark-gray skin strain with immense force. A vibrant green essence begins to pulse rhythmically from its limbs, and light seeps through a network of cracks, casting an unnatural glow on its obsidian claws. With a swift, decisive downward arc, the creature slams its energetic fist into a discarded piece of machinery half-buried in ochre dust. The incredible force of the impact instantly shatters the rusted, angular metal frame, creating a realistic deformation of twisted steel. Dazzling metallic sparks erupt violently from the point of contact, each individual particle arcing in blinding light while green energy is released in dazzling flashes, creating a tense and powerful atmosphere that emphasizes the raw, untamed power of this alien behemoth.", + "seed": 11221, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000071/group_15/index_1530_seed_11221_dp_31_rank_62_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 15, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.04037916373927146, + "vlm_logprob": -0.04037916373927146, + "vlm_yes_no_margin": 4.05, + "vlm_token_credit": -0.04037916373927146, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static, low-angle, wide shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the environment a desolate alien world under the glare of twin suns in a pale, wispy sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the central subject a hulking bipedal creature with chalky, weathered bony plates and dark-gray skin?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the creature slam its fist into a discarded piece of machinery, causing it to shatter and emit metallic sparks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is there a vibrant green essence pulsing from the creature's limbs and casting an unnatural glow?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.10655568540096283 + ], + "opposite_logprob_tokens": [ + -2.481555700302124 + ], + "logprob_reward": -0.10655568540096283, + "yes_no_margin_reward": 2.375, + "token_credit_reward": -0.10655568540096283, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.16059279441833496, + 0.12076768279075623, + 0.09470918029546738, + 0.06329437345266342, + 0.06344984471797943, + 0.0690271332859993, + 0.17316091060638428, + 0.08044973760843277, + 0.1157040148973465, + 0.05884429067373276 + ], + "credit_map": [ + 0.02327430061995983, + 0.0276540145277977, + 0.0146562485024333, + 0.01411697082221508, + 0.02811640501022339, + 0.015373101457953453, + 0.015597134828567505, + 0.021804630756378174, + 0.01632682979106903, + 0.03587622195482254, + 0.011807278729975224, + 0.017992762848734856, + 0.010493279434740543, + 0.010495979338884354, + 0.009541229344904423, + 0.008234092965722084, + 0.016290079802274704, + 0.013792582787573338, + 0.011778694577515125, + 0.013825315982103348, + 0.009957265108823776, + 0.008087819442152977, + 0.012861417606472969, + 0.008116005919873714, + 0.008112162351608276, + 0.008468546904623508, + 0.006548547185957432, + 0.00627087103202939, + 0.0123874107375741, + 0.007249189540743828, + 0.009220576845109463, + 0.005037073045969009, + 0.009957204572856426, + 0.007024018559604883, + 0.015578636899590492, + 0.008686852641403675, + 0.004082285333424807, + 0.005439498461782932, + 0.005686230957508087, + 0.006995119154453278, + 0.009187662973999977, + 0.0106276273727417, + 0.010455934330821037, + 0.00499936006963253, + 0.008118312805891037, + 0.009422323666512966, + 0.008310653269290924, + 0.007905260659754276, + 0.02398943528532982, + 0.02700360119342804, + 0.023060539737343788, + 0.02471810020506382, + 0.01935761794447899, + 0.021173348650336266, + 0.017036017030477524, + 0.01682225614786148, + 0.010123097337782383, + 0.013906256295740604, + 0.009881669655442238, + 0.008283553645014763, + 0.009681177325546741, + 0.0065967314876616, + 0.013049578294157982, + 0.008927667513489723, + 0.014262734912335873, + 0.016032183542847633, + 0.014986037276685238, + 0.01239357516169548, + 0.01381182111799717, + 0.010067446157336235, + 0.021576624363660812, + 0.012573592364788055, + 0.008250940591096878, + 0.007078326772898436, + 0.011897126212716103, + 0.00639899168163538, + 0.005640854127705097, + 0.006446472834795713, + 0.007005582097917795, + 0.006125996354967356 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 0, + 1 + ], + "selected_reward": -0.10655568540096283 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static, low-angle, wide shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the environment a desolate alien world under the glare of twin suns in a pale, wispy sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the central subject a hulking bipedal creature with chalky, weathered bony plates and dark-gray skin?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the creature slam its fist into a discarded piece of machinery, causing it to shatter and emit metallic sparks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is there a vibrant green essence pulsing from the creature's limbs and casting an unnatural glow?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.06207427754998207 + ], + "opposite_logprob_tokens": [ + -2.8120741844177246 + ], + "logprob_reward": -0.06207427754998207, + "yes_no_margin_reward": 2.75, + "token_credit_reward": -0.06207427754998207, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.293027400970459, + 0.07493668049573898, + 0.049247030168771744, + 0.026570184156298637, + 0.15170079469680786, + 0.10023948550224304, + 0.06021184101700783, + 0.08705512434244156, + 0.05675429478287697, + 0.10025715827941895 + ], + "credit_map": [ + 0.04776594415307045, + 0.07238578051328659, + 0.042036570608615875, + 0.02506105788052082, + 0.03213047236204147, + 0.03206262364983559, + 0.020570479333400726, + 0.021014459431171417, + 0.013132843188941479, + 0.010656860657036304, + 0.00785119365900755, + 0.009458120912313461, + 0.014324124902486801, + 0.007958171889185905, + 0.006986023858189583, + 0.004569338634610176, + 0.007884175516664982, + 0.007825714536011219, + 0.00622231001034379, + 0.006182121578603983, + 0.006083914544433355, + 0.005671293940395117, + 0.005280311685055494, + 0.004097187891602516, + 0.003309720428660512, + 0.005595846101641655, + 0.0030095481779426336, + 0.0028537020552903414, + 0.0035617505200207233, + 0.0024136172141879797, + 0.003031224012374878, + 0.0027947761118412018, + 0.023587683215737343, + 0.01866009645164013, + 0.031031746417284012, + 0.024380674585700035, + 0.011013444513082504, + 0.014184333384037018, + 0.01480292808264494, + 0.01403988990932703, + 0.013560779392719269, + 0.014704212546348572, + 0.017723066732287407, + 0.013303817249834538, + 0.0070366403087973595, + 0.014838517643511295, + 0.009017200209200382, + 0.010055254213511944, + 0.006670042872428894, + 0.007577131502330303, + 0.010221927426755428, + 0.008350778371095657, + 0.0068522002547979355, + 0.0065409294329583645, + 0.004448597319424152, + 0.009550236165523529, + 0.009292949922382832, + 0.012882765382528305, + 0.010067125782370567, + 0.016626451164484024, + 0.007942967116832733, + 0.005657119210809469, + 0.01719612255692482, + 0.0073896246030926704, + 0.007027256302535534, + 0.007912839762866497, + 0.006021382752805948, + 0.005403782706707716, + 0.007413796149194241, + 0.005119805224239826, + 0.009917817078530788, + 0.007937614805996418, + 0.01429232582449913, + 0.015182956121861935, + 0.008042212575674057, + 0.00971661414951086, + 0.015639180317521095, + 0.01571066677570343, + 0.010928750969469547, + 0.010744445025920868 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 9 + ], + "selected_reward": -0.06207427754998207 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static, low-angle, wide shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the environment a desolate alien world under the glare of twin suns in a pale, wispy sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the central subject a hulking bipedal creature with chalky, weathered bony plates and dark-gray skin?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the creature slam its fist into a discarded piece of machinery, causing it to shatter and emit metallic sparks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is there a vibrant green essence pulsing from the creature's limbs and casting an unnatural glow?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.00864359550178051 + ], + "opposite_logprob_tokens": [ + -4.758643627166748 + ], + "logprob_reward": -0.00864359550178051, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.00864359550178051, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1133010983467102, + 0.08631646633148193, + 0.051516830921173096, + 0.07430076599121094, + 0.16076746582984924, + 0.05114087089896202, + 0.0486987940967083, + 0.12565653026103973, + 0.24720293283462524, + 0.04109823331236839 + ], + "credit_map": [ + 0.01828869991004467, + 0.02291746996343136, + 0.014762509614229202, + 0.00888106133788824, + 0.016024332493543625, + 0.010086225345730782, + 0.009033522568643093, + 0.013307270593941212, + 0.014104330912232399, + 0.014036167412996292, + 0.008279725909233093, + 0.011279692873358727, + 0.012843119911849499, + 0.009481971152126789, + 0.0098251448944211, + 0.006466306746006012, + 0.007067970000207424, + 0.008189985528588295, + 0.005320104770362377, + 0.005890504457056522, + 0.008883320726454258, + 0.006612458266317844, + 0.005703612230718136, + 0.003848874941468239, + 0.008497907780110836, + 0.013404675759375095, + 0.007068131119012833, + 0.006921632215380669, + 0.014750206843018532, + 0.008316166698932648, + 0.010274704545736313, + 0.005067335907369852, + 0.0219135619699955, + 0.020660292357206345, + 0.026142125949263573, + 0.03477364033460617, + 0.01535762008279562, + 0.00998458918184042, + 0.020227380096912384, + 0.011708253994584084, + 0.007985040545463562, + 0.007657620590180159, + 0.009053746238350868, + 0.004425963386893272, + 0.004542575217783451, + 0.005728563293814659, + 0.005575328599661589, + 0.006172033492475748, + 0.005725053604692221, + 0.004763081204146147, + 0.007939395494759083, + 0.006909267511218786, + 0.005154755432158709, + 0.005887357518076897, + 0.004533874336630106, + 0.007786009926348925, + 0.014981538988649845, + 0.02174203470349312, + 0.013702496886253357, + 0.021204130724072456, + 0.0162981990724802, + 0.00936937890946865, + 0.015223507769405842, + 0.013135244138538837, + 0.030967332422733307, + 0.03121481090784073, + 0.035919737070798874, + 0.02509506791830063, + 0.02825029566884041, + 0.022585561498999596, + 0.04520094767212868, + 0.027969177812337875, + 0.005755038000643253, + 0.0063340747728943825, + 0.0038248205091804266, + 0.00462222658097744, + 0.00644038338214159, + 0.006595708895474672, + 0.004184856079518795, + 0.0033411288168281317 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 4, + 7 + ], + "selected_reward": -0.00864359550178051 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static, low-angle, wide shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the environment a desolate alien world under the glare of twin suns in a pale, wispy sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the central subject a hulking bipedal creature with chalky, weathered bony plates and dark-gray skin?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the creature slam its fist into a discarded piece of machinery, causing it to shatter and emit metallic sparks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is there a vibrant green essence pulsing from the creature's limbs and casting an unnatural glow?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.023285744711756706 + ], + "opposite_logprob_tokens": [ + -3.7732856273651123 + ], + "logprob_reward": -0.023285744711756706, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.023285744711756706, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.220469132065773, + 0.07638731598854065, + 0.04513023793697357, + 0.03255835920572281, + 0.057245995849370956, + 0.036932285875082016, + 0.08104324340820312, + 0.10952062904834747, + 0.24060140550136566, + 0.10011133551597595 + ], + "credit_map": [ + 0.03885750100016594, + 0.06194325163960457, + 0.03877658396959305, + 0.016039161011576653, + 0.018553566187620163, + 0.01581217721104622, + 0.009407516568899155, + 0.021079378202557564, + 0.01044009905308485, + 0.021997889503836632, + 0.006413946859538555, + 0.008386928588151932, + 0.008921067230403423, + 0.008400770835578442, + 0.00615777587518096, + 0.005668835248798132, + 0.006040595471858978, + 0.0058135115541517735, + 0.0052917711436748505, + 0.004895165096968412, + 0.008933532051742077, + 0.0043691834434866905, + 0.0063537065871059895, + 0.0034327711910009384, + 0.0036329468712210655, + 0.004816514439880848, + 0.0034249054733663797, + 0.0037937662564218044, + 0.00721860071644187, + 0.0032840792555361986, + 0.0036286204122006893, + 0.002758928807452321, + 0.00832811463624239, + 0.007644938305020332, + 0.009414693340659142, + 0.010480904020369053, + 0.005473035853356123, + 0.0036886504385620356, + 0.007696384564042091, + 0.004519276320934296, + 0.005318147130310535, + 0.005810208152979612, + 0.005279615521430969, + 0.002823749091476202, + 0.003721805289387703, + 0.0053167263977229595, + 0.004594297613948584, + 0.004067741334438324, + 0.010964450426399708, + 0.010751625522971153, + 0.011289622634649277, + 0.009426325559616089, + 0.008985077030956745, + 0.010140279307961464, + 0.009137949906289577, + 0.010347912088036537, + 0.013393324799835682, + 0.01583925075829029, + 0.014924412593245506, + 0.017193978652358055, + 0.011695556342601776, + 0.008941472508013248, + 0.01579541712999344, + 0.01173721719533205, + 0.02943962626159191, + 0.030521977692842484, + 0.03335258737206459, + 0.025350922718644142, + 0.027739141136407852, + 0.023289045318961143, + 0.04401998221874237, + 0.026888128370046616, + 0.012596839107573032, + 0.01267209742218256, + 0.011982480064034462, + 0.011826536618173122, + 0.012266350910067558, + 0.015551851131021976, + 0.01390126720070839, + 0.009313911199569702 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 0, + 7 + ], + "selected_reward": -0.023285744711756706 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static, low-angle, wide shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the environment a desolate alien world under the glare of twin suns in a pale, wispy sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the central subject a hulking bipedal creature with chalky, weathered bony plates and dark-gray skin?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the creature slam its fist into a discarded piece of machinery, causing it to shatter and emit metallic sparks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is there a vibrant green essence pulsing from the creature's limbs and casting an unnatural glow?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0013365155318751931 + ], + "opposite_logprob_tokens": [ + -6.626336574554443 + ], + "logprob_reward": -0.0013365155318751931, + "yes_no_margin_reward": 6.625, + "token_credit_reward": -0.0013365155318751931, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.22467567026615143, + 0.03917388617992401, + 0.06411950290203094, + 0.07113154232501984, + 0.07040753960609436, + 0.04566749557852745, + 0.07327093183994293, + 0.1168331578373909, + 0.26146748661994934, + 0.03325274586677551 + ], + "credit_map": [ + 0.03990498557686806, + 0.051921550184488297, + 0.04737890139222145, + 0.01733328588306904, + 0.01908961683511734, + 0.01563090644776821, + 0.010285142809152603, + 0.023131299763917923, + 0.006154233124107122, + 0.006375845056027174, + 0.003795762313529849, + 0.004346081987023354, + 0.007505280897021294, + 0.003884832374751568, + 0.003873581299558282, + 0.003238274483010173, + 0.007121175527572632, + 0.009164871647953987, + 0.00711331982165575, + 0.008245084434747696, + 0.011450760066509247, + 0.006911385338753462, + 0.00871204398572445, + 0.0054008751176297665, + 0.009300239384174347, + 0.009826079942286015, + 0.007631204091012478, + 0.009868615306913853, + 0.011759407818317413, + 0.0071318382397294044, + 0.008898413740098476, + 0.0067157503217458725, + 0.009853419847786427, + 0.009188265539705753, + 0.011190257035195827, + 0.014912269078195095, + 0.006298399064689875, + 0.0043999310582876205, + 0.009102445095777512, + 0.005462564527988434, + 0.006560357753187418, + 0.006328762508928776, + 0.0086865508928895, + 0.005415992811322212, + 0.0043074567802250385, + 0.004569802898913622, + 0.004184642340987921, + 0.0056139277294278145, + 0.00847721192985773, + 0.00993182323873043, + 0.01255494449287653, + 0.007470252458006144, + 0.009088777005672455, + 0.008375606499612331, + 0.007541986182332039, + 0.009830335155129433, + 0.014823831617832184, + 0.020078204572200775, + 0.015812579542398453, + 0.01845579594373703, + 0.013898317702114582, + 0.007605112157762051, + 0.014168198220431805, + 0.011991122737526894, + 0.033136844635009766, + 0.03209734708070755, + 0.03473074361681938, + 0.026895714923739433, + 0.029815498739480972, + 0.025035522878170013, + 0.0490499772131443, + 0.03070586547255516, + 0.004542442038655281, + 0.004163785371929407, + 0.0047016507014632225, + 0.0045639039017260075, + 0.003147892188280821, + 0.004424032289534807, + 0.0039991214871406555, + 0.0037099223118275404 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 0, + 7 + ], + "selected_reward": -0.0013365155318751931 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "static, low-angle, wide shot, rendered in hyper-realistic 8K high-resolution CGI animation, establishes the immensity of a desolate alien world under the harsh, overexposed glare of twin suns suspended in a pale, wispy sky. At the center of the frame stands a hulking, bipedal primitive, its form a terrifying fusion of sinew and armor, its body covered in overlapping, shifting bony plates with a chalky, weathered texture that delicately rub against one another. As it raises a massive claw, the muscles beneath its leathery, dark-gray skin strain with immense force. A vibrant green essence begins to pulse rhythmically from its limbs, and light seeps through a network of cracks, casting an unnatural glow on its obsidian claws. With a swift, decisive downward arc, the creature slams its energetic fist into a discarded piece of machinery half-buried in ochre dust. The incredible force of the impact instantly shatters the rusted, angular metal frame, creating a realistic deformation of twisted steel. Dazzling metallic sparks erupt violently from the point of contact, each individual particle arcing in blinding light while green energy is released in dazzling flashes, creating a tense and powerful atmosphere that emphasizes the raw, untamed power of this alien behemoth.", + "seed": 11222, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000071/group_15/index_1530_seed_11222_dp_31_rank_62_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 15, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.03209242383018136, + "vlm_logprob": -0.03209242383018136, + "vlm_yes_no_margin": 4.15, + "vlm_token_credit": -0.03209242383018136, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static, low-angle, wide shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the environment a desolate alien world under the glare of twin suns in a pale, wispy sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the central subject a hulking bipedal creature with chalky, weathered bony plates and dark-gray skin?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the creature slam its fist into a discarded piece of machinery, causing it to shatter and emit metallic sparks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is there a vibrant green essence pulsing from the creature's limbs and casting an unnatural glow?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0927346795797348 + ], + "opposite_logprob_tokens": [ + -2.7177345752716064 + ], + "logprob_reward": -0.0927346795797348, + "yes_no_margin_reward": 2.625, + "token_credit_reward": -0.0927346795797348, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13811121881008148, + 0.07722628861665726, + 0.04411604255437851, + 0.06225864216685295, + 0.19912035763263702, + 0.04085000604391098, + 0.03713095933198929, + 0.35998207330703735, + 0.018178192898631096, + 0.023026183247566223 + ], + "credit_map": [ + 0.024868817999958992, + 0.015884792432188988, + 0.010916676372289658, + 0.008442803286015987, + 0.012548325583338737, + 0.048030950129032135, + 0.007658216170966625, + 0.009760643355548382, + 0.013819034211337566, + 0.005651453044265509, + 0.0075864954851567745, + 0.007462284993380308, + 0.005616715643554926, + 0.01815645769238472, + 0.005146370735019445, + 0.013787471689283848, + 0.008309368044137955, + 0.004250051453709602, + 0.004090331960469484, + 0.003216272220015526, + 0.005767724476754665, + 0.011854416690766811, + 0.002568902214989066, + 0.0040589747950434685, + 0.00898982584476471, + 0.005566386040300131, + 0.0068021975457668304, + 0.005613177083432674, + 0.007834689691662788, + 0.013808821327984333, + 0.0061697447672486305, + 0.007473805453628302, + 0.021653540432453156, + 0.012690745294094086, + 0.007891344837844372, + 0.008228144608438015, + 0.03955935314297676, + 0.018406206741929054, + 0.07884684950113297, + 0.011844181455671787, + 0.007550935260951519, + 0.0037720324471592903, + 0.002086289692670107, + 0.004256852902472019, + 0.014116186648607254, + 0.0027832896448671818, + 0.002443142468109727, + 0.003841276280581951, + 0.003592614782974124, + 0.005391778890043497, + 0.005897132214158773, + 0.0033251687418669462, + 0.005164720583707094, + 0.004848509561270475, + 0.005005189683288336, + 0.003905845806002617, + 0.02129235304892063, + 0.05968041718006134, + 0.014737529680132866, + 0.02108190953731537, + 0.0239524208009243, + 0.152668297290802, + 0.030747711658477783, + 0.03582143411040306, + 0.0017379133496433496, + 0.001842894940637052, + 0.0023582270368933678, + 0.0026199305430054665, + 0.002408452797681093, + 0.002222569892182946, + 0.002555594779551029, + 0.002432607114315033, + 0.0027857020031660795, + 0.0028832871466875076, + 0.004219927359372377, + 0.00335521693341434, + 0.0021802783012390137, + 0.0030219336040318012, + 0.002250976860523224, + 0.0023288612719625235 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 4, + 0 + ], + "selected_reward": -0.0927346795797348 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static, low-angle, wide shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the environment a desolate alien world under the glare of twin suns in a pale, wispy sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the central subject a hulking bipedal creature with chalky, weathered bony plates and dark-gray skin?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the creature slam its fist into a discarded piece of machinery, causing it to shatter and emit metallic sparks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is there a vibrant green essence pulsing from the creature's limbs and casting an unnatural glow?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04308806359767914 + ], + "opposite_logprob_tokens": [ + -3.168087959289551 + ], + "logprob_reward": -0.04308806359767914, + "yes_no_margin_reward": 3.125, + "token_credit_reward": -0.04308806359767914, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.17482233047485352, + 0.04185016453266144, + 0.04353484883904457, + 0.05833215266466141, + 0.24132682383060455, + 0.009738316759467125, + 0.024858666583895683, + 0.37191903591156006, + 0.013793951831758022, + 0.019823729991912842 + ], + "credit_map": [ + 0.035709988325834274, + 0.015061839483678341, + 0.011078983545303345, + 0.007075068540871143, + 0.012054011225700378, + 0.07416368275880814, + 0.008016562089323997, + 0.01166219636797905, + 0.009564767591655254, + 0.002493161940947175, + 0.003136380109935999, + 0.0019050263799726963, + 0.0023510283790528774, + 0.014058932662010193, + 0.0020352222491055727, + 0.0063056438229978085, + 0.01069039385765791, + 0.0028324106242507696, + 0.0024310993030667305, + 0.0020789150148630142, + 0.004459597170352936, + 0.015597835183143616, + 0.002222169889137149, + 0.003222423605620861, + 0.011085089296102524, + 0.004123103339225054, + 0.005458001513034105, + 0.0034238107036799192, + 0.006551103666424751, + 0.019108226522803307, + 0.004109223373234272, + 0.004473594017326832, + 0.027221551164984703, + 0.01839824765920639, + 0.012077957391738892, + 0.023245390504598618, + 0.05102868750691414, + 0.015302632004022598, + 0.08188872039318085, + 0.012163618579506874, + 0.000943193503189832, + 0.0020320091862231493, + 0.0012763281119987369, + 0.0016813615802675486, + 0.0016442469786852598, + 0.0006883578025735915, + 0.0006565768853761256, + 0.0008162418962456286, + 0.003306333441287279, + 0.002948086941614747, + 0.004414533264935017, + 0.002363308798521757, + 0.0024938674177974463, + 0.0025294888764619827, + 0.004286505281925201, + 0.0025165409315377474, + 0.022020451724529266, + 0.060869041830301285, + 0.015170138329267502, + 0.02239464409649372, + 0.024006661027669907, + 0.15898501873016357, + 0.031755901873111725, + 0.03671715408563614, + 0.001419601496309042, + 0.0012537286384031177, + 0.0025660248938947916, + 0.002166010672226548, + 0.0016340380534529686, + 0.0012617009924724698, + 0.001803044113330543, + 0.0016898025060072541, + 0.0028379273135215044, + 0.002480460098013282, + 0.0031809054780751467, + 0.002751185093075037, + 0.002054711105301976, + 0.002344501903280616, + 0.0020206074696034193, + 0.0021534294355660677 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 4, + 0 + ], + "selected_reward": -0.04308806359767914 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static, low-angle, wide shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the environment a desolate alien world under the glare of twin suns in a pale, wispy sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the central subject a hulking bipedal creature with chalky, weathered bony plates and dark-gray skin?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the creature slam its fist into a discarded piece of machinery, causing it to shatter and emit metallic sparks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is there a vibrant green essence pulsing from the creature's limbs and casting an unnatural glow?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.007628594525158405 + ], + "opposite_logprob_tokens": [ + -4.882628440856934 + ], + "logprob_reward": -0.007628594525158405, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.007628594525158405, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09162779152393341, + 0.09917011111974716, + 0.07875274121761322, + 0.07409963011741638, + 0.1954616755247116, + 0.036127686500549316, + 0.06303876638412476, + 0.2991146445274353, + 0.026093047112226486, + 0.03651391714811325 + ], + "credit_map": [ + 0.01086814422160387, + 0.010435816831886768, + 0.010086676105856895, + 0.017438603565096855, + 0.010352502577006817, + 0.013831419870257378, + 0.012401994317770004, + 0.006212630309164524, + 0.019418850541114807, + 0.008316408842802048, + 0.010118174366652966, + 0.00796317309141159, + 0.009235911071300507, + 0.026228873059153557, + 0.006103638093918562, + 0.011785087175667286, + 0.011264191940426826, + 0.009618019685149193, + 0.009179185144603252, + 0.01015335414558649, + 0.01001898106187582, + 0.012539702467620373, + 0.009187060408294201, + 0.0067922500893473625, + 0.013625537976622581, + 0.00537498015910387, + 0.007002177648246288, + 0.004801827482879162, + 0.00866717379540205, + 0.023795397952198982, + 0.005525495857000351, + 0.005307045299559832, + 0.020968087017536163, + 0.012397442944347858, + 0.00766706420108676, + 0.007733275182545185, + 0.03970283269882202, + 0.017983993515372276, + 0.0771714299917221, + 0.011837545782327652, + 0.00269007938914001, + 0.008904844522476196, + 0.004170344211161137, + 0.00899411365389824, + 0.0043027400970458984, + 0.0019258264219388366, + 0.0026387376710772514, + 0.0025010020472109318, + 0.0041273003444075584, + 0.012030276469886303, + 0.007806228939443827, + 0.005492950789630413, + 0.006761278957128525, + 0.008504978381097317, + 0.010252240113914013, + 0.008063515648245811, + 0.0178415235131979, + 0.04906559735536575, + 0.01163991168141365, + 0.018019672483205795, + 0.018846947699785233, + 0.12650823593139648, + 0.025985343381762505, + 0.031207401305437088, + 0.0022002942860126495, + 0.0026639054995030165, + 0.0034448548685759306, + 0.0038696727715432644, + 0.0045632291585206985, + 0.0020699037704616785, + 0.00375126744620502, + 0.00352991814725101, + 0.004942245781421661, + 0.004326409660279751, + 0.0060356189496815205, + 0.0060063921846449375, + 0.0027110897935926914, + 0.004784082528203726, + 0.003491849871352315, + 0.004216231871396303 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 4, + 1 + ], + "selected_reward": -0.007628594525158405 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static, low-angle, wide shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the environment a desolate alien world under the glare of twin suns in a pale, wispy sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the central subject a hulking bipedal creature with chalky, weathered bony plates and dark-gray skin?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the creature slam its fist into a discarded piece of machinery, causing it to shatter and emit metallic sparks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is there a vibrant green essence pulsing from the creature's limbs and casting an unnatural glow?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.014192714355885983 + ], + "opposite_logprob_tokens": [ + -4.264192581176758 + ], + "logprob_reward": -0.014192714355885983, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.014192714355885983, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10961510241031647, + 0.10385154187679291, + 0.03161533549427986, + 0.05589008331298828, + 0.2290186882019043, + 0.07523955404758453, + 0.038483429700136185, + 0.10775315761566162, + 0.18007783591747284, + 0.0684552788734436 + ], + "credit_map": [ + 0.01359078474342823, + 0.011018430814146996, + 0.011779867112636566, + 0.017107967287302017, + 0.007790793664753437, + 0.02643950656056404, + 0.013321803882718086, + 0.008565941825509071, + 0.020282013341784477, + 0.008921082131564617, + 0.012016992084681988, + 0.006563947070389986, + 0.01019587554037571, + 0.027951402589678764, + 0.00730684632435441, + 0.010613375343382359, + 0.0069809830747544765, + 0.0024690295103937387, + 0.0029972123447805643, + 0.0024183569476008415, + 0.0035904741380363703, + 0.008750589564442635, + 0.0020518284291028976, + 0.002356859389692545, + 0.008201677352190018, + 0.005447952076792717, + 0.006508483551442623, + 0.005056495312601328, + 0.008618677034974098, + 0.010957402177155018, + 0.005362340714782476, + 0.005737051367759705, + 0.02887236885726452, + 0.02012120559811592, + 0.012878624722361565, + 0.028044503182172775, + 0.046888064593076706, + 0.015131705440580845, + 0.06616751104593277, + 0.010914690792560577, + 0.013067738153040409, + 0.009186886250972748, + 0.005950689781457186, + 0.008357858285307884, + 0.017520947381854057, + 0.006837606895714998, + 0.004541654139757156, + 0.00977617222815752, + 0.003979978617280722, + 0.005122472532093525, + 0.006112534087151289, + 0.004141638055443764, + 0.0038267611525952816, + 0.003978859633207321, + 0.007167670410126448, + 0.0041535100899636745, + 0.0055886805057525635, + 0.019343508407473564, + 0.006135425064712763, + 0.012962550856173038, + 0.007637200877070427, + 0.03445040434598923, + 0.010044816881418228, + 0.011590573005378246, + 0.013446686789393425, + 0.02474341168999672, + 0.018465038388967514, + 0.03865790739655495, + 0.025465726852416992, + 0.018841933459043503, + 0.025282924994826317, + 0.015174192376434803, + 0.007328662555664778, + 0.007724720984697342, + 0.010957063175737858, + 0.01009384449571371, + 0.005357205867767334, + 0.007268586661666632, + 0.008943257853388786, + 0.010781927965581417 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 8, + 0 + ], + "selected_reward": -0.014192714355885983 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static, low-angle, wide shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the environment a desolate alien world under the glare of twin suns in a pale, wispy sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the central subject a hulking bipedal creature with chalky, weathered bony plates and dark-gray skin?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the creature slam its fist into a discarded piece of machinery, causing it to shatter and emit metallic sparks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is there a vibrant green essence pulsing from the creature's limbs and casting an unnatural glow?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.002818067092448473 + ], + "opposite_logprob_tokens": [ + -5.8778181076049805 + ], + "logprob_reward": -0.002818067092448473, + "yes_no_margin_reward": 5.875, + "token_credit_reward": -0.002818067092448473, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10510938614606857, + 0.08167736977338791, + 0.02797194942831993, + 0.027676565572619438, + 0.19658754765987396, + 0.09400373697280884, + 0.03342423588037491, + 0.2796400785446167, + 0.07027997821569443, + 0.08362909406423569 + ], + "credit_map": [ + 0.014169343747198582, + 0.010690856724977493, + 0.007810227572917938, + 0.017441287636756897, + 0.008287272416055202, + 0.026681844145059586, + 0.013004892505705357, + 0.007023659534752369, + 0.015594408847391605, + 0.007198837585747242, + 0.00937279500067234, + 0.008689326234161854, + 0.008018018677830696, + 0.01995447650551796, + 0.005565317813307047, + 0.007284187711775303, + 0.0068955253809690475, + 0.0024840806145220995, + 0.002057563280686736, + 0.0012263511307537556, + 0.0027884780429303646, + 0.008580638095736504, + 0.0017688433872535825, + 0.0021704696118831635, + 0.004539247136563063, + 0.002341627376154065, + 0.0028501462657004595, + 0.0020673342514783144, + 0.003728078678250313, + 0.007619498763233423, + 0.0026106855366379023, + 0.0019199475646018982, + 0.02305571921169758, + 0.016706673428416252, + 0.010855765081942081, + 0.022688796743750572, + 0.0405697226524353, + 0.01262175478041172, + 0.06059809401631355, + 0.009491016156971455, + 0.017551986500620842, + 0.008694554679095745, + 0.004436272196471691, + 0.008713193237781525, + 0.035136979073286057, + 0.00626227306202054, + 0.00574877904728055, + 0.007459697313606739, + 0.004021934699267149, + 0.004381283186376095, + 0.006785741075873375, + 0.002892422955483198, + 0.0033047308679670095, + 0.004255477339029312, + 0.004307173192501068, + 0.0034754702355712652, + 0.016071107238531113, + 0.03569438308477402, + 0.013681254349648952, + 0.033776819705963135, + 0.01723545789718628, + 0.11709822714328766, + 0.022654403001070023, + 0.023428428918123245, + 0.004816895816475153, + 0.009260598570108414, + 0.012375900521874428, + 0.015960579738020897, + 0.007452154066413641, + 0.006887913215905428, + 0.007373299915343523, + 0.006152634043246508, + 0.012186668813228607, + 0.008503423072397709, + 0.014139987528324127, + 0.01003088429570198, + 0.008576088584959507, + 0.007799971848726273, + 0.012489156797528267, + 0.009902914054691792 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 4, + 0 + ], + "selected_reward": -0.002818067092448473 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000071/group_2/index_tensor([9759, 9759])_group_2_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000071/group_2/index_tensor([9759, 9759])_group_2_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..49aaad930ffb6ecd80c6ede98f5009c9998d3c91 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000071/group_2/index_tensor([9759, 9759])_group_2_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "subsequent low-angle tracking shot captures a streamlined, deep cherry red sports car hurtling forward at breakneck speed through a deep, straight, modern concrete tunnel. The car's mirror-smooth paint reflects the surrounding light, and its signature LED taillights illuminate, trailing two bright red streaks of light in their rapid motion. On either side, rows of cool white LEDs on the tunnel walls are visually stretched and merged by their rapid relative speed, ultimately converging into two continuous, dazzling horizontal streaks of light that sweep backwards, interplaying with the reflections on the wet, dark asphalt. Because the camera moves in sync with the car, the body remains relatively stable in the frame, while the light in the background transforms into dynamic lines. This striking contrast between movement and stillness creates an adrenaline-pumping visual impact. The entire scene, rendered in cold tones, is imbued with a futuristic, austere beauty and a sense of pure speed and power.", + "seed": 27453, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000071/group_2/index_9759_seed_27453_dp_4_rank_8_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 2, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.005658799572847784, + "vlm_logprob": -0.005658799572847784, + "vlm_yes_no_margin": 5.625, + "vlm_token_credit": -0.005658799572847784, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a streamlined, deep cherry red sports car?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the car moving at high speed through a deep, straight, modern concrete tunnel?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the car's LED taillights appear as bright red streaks trailing behind due to motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the rows of cool white LEDs on the tunnel walls visually stretched into continuous horizontal streaks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot captured from a low-angle perspective with the camera tracking in sync with the car?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.006847131997346878 + ], + "opposite_logprob_tokens": [ + -6.131846904754639 + ], + "logprob_reward": -0.006847131997346878, + "yes_no_margin_reward": 6.125, + "token_credit_reward": -0.006847131997346878, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2067384272813797, + 0.19501392543315887, + 0.056658387184143066, + 0.04871882498264313, + 0.024362333118915558, + 0.06193278729915619, + 0.12198103964328766, + 0.17351993918418884, + 0.03482207655906677, + 0.0762522742152214 + ], + "credit_map": [ + 0.041054558008909225, + 0.027058249339461327, + 0.03854548931121826, + 0.024564828723669052, + 0.02087165042757988, + 0.029664570465683937, + 0.014232388697564602, + 0.01074669137597084, + 0.028492655605077744, + 0.033226270228624344, + 0.06959725171327591, + 0.01880684867501259, + 0.0145955141633749, + 0.014832040295004845, + 0.007790169678628445, + 0.007673180196434259, + 0.011822197586297989, + 0.006658333353698254, + 0.0077169062569737434, + 0.00548512814566493, + 0.007825661450624466, + 0.009280377067625523, + 0.004554152023047209, + 0.003315636655315757, + 0.008559032343327999, + 0.007285519037395716, + 0.007120836526155472, + 0.005883248057216406, + 0.0050431787967681885, + 0.006671701557934284, + 0.004449551459401846, + 0.0037057637237012386, + 0.0029800180345773697, + 0.003111568745225668, + 0.005401790142059326, + 0.002582714194431901, + 0.0033341313246637583, + 0.002883774694055319, + 0.0020866275299340487, + 0.001981709385290742, + 0.009937972761690617, + 0.01066726166754961, + 0.008910507895052433, + 0.00784493237733841, + 0.007143624592572451, + 0.009793560951948166, + 0.004253980703651905, + 0.003380945883691311, + 0.026768891140818596, + 0.01123734936118126, + 0.024991510435938835, + 0.010069879703223705, + 0.01599164679646492, + 0.018422061577439308, + 0.006989591289311647, + 0.007510103285312653, + 0.02539345808327198, + 0.03308703750371933, + 0.030856221914291382, + 0.015987996011972427, + 0.023744825273752213, + 0.023646390065550804, + 0.009526258334517479, + 0.011277756653726101, + 0.004435522947460413, + 0.00409637251868844, + 0.00617810245603323, + 0.004065229557454586, + 0.005025544669479132, + 0.004840503446757793, + 0.003396435407921672, + 0.002784362528473139, + 0.006832802202552557, + 0.006397457793354988, + 0.008325343020260334, + 0.005131294019520283, + 0.011947400867938995, + 0.005692531820386648, + 0.011431822553277016, + 0.020493624731898308 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 7 + ], + "selected_reward": -0.006847131997346878 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a streamlined, deep cherry red sports car?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the car moving at high speed through a deep, straight, modern concrete tunnel?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the car's LED taillights appear as bright red streaks trailing behind due to motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the rows of cool white LEDs on the tunnel walls visually stretched into continuous horizontal streaks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot captured from a low-angle perspective with the camera tracking in sync with the car?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005272530950605869 + ], + "opposite_logprob_tokens": [ + -5.255272388458252 + ], + "logprob_reward": -0.005272530950605869, + "yes_no_margin_reward": 5.25, + "token_credit_reward": -0.005272530950605869, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3449588119983673, + 0.07233349233865738, + 0.04438094049692154, + 0.05041038990020752, + 0.08675979822874069, + 0.04535193741321564, + 0.048733633011579514, + 0.2091207206249237, + 0.05122626945376396, + 0.04672403633594513 + ], + "credit_map": [ + 0.07012969255447388, + 0.047160178422927856, + 0.04180976003408432, + 0.027799181640148163, + 0.038516558706760406, + 0.06114489212632179, + 0.02688160538673401, + 0.031516991555690765, + 0.007465999573469162, + 0.021379128098487854, + 0.01105870958417654, + 0.007058954797685146, + 0.01031146664172411, + 0.004315701778978109, + 0.006075090728700161, + 0.004668444395065308, + 0.004864299204200506, + 0.0045666866935789585, + 0.004835776053369045, + 0.006496400106698275, + 0.005485274363309145, + 0.004970637150108814, + 0.006300696171820164, + 0.006861172150820494, + 0.0037271291948854923, + 0.007218791171908379, + 0.0067201717756688595, + 0.006080773659050465, + 0.006374824792146683, + 0.004962077364325523, + 0.008437004871666431, + 0.0068896166048944, + 0.008766645565629005, + 0.011070206761360168, + 0.013328813947737217, + 0.008611251600086689, + 0.011599087156355381, + 0.008904414251446724, + 0.010672731325030327, + 0.013806641101837158, + 0.009010874666273594, + 0.0053812782280147076, + 0.007141966838389635, + 0.005136560183018446, + 0.004846565891057253, + 0.007490848191082478, + 0.002974819391965866, + 0.00336902285926044, + 0.007195994723588228, + 0.005164588335901499, + 0.008424133062362671, + 0.004394644871354103, + 0.006540588103234768, + 0.008917690254747868, + 0.0028343021403998137, + 0.005261693615466356, + 0.027997232973575592, + 0.03653416410088539, + 0.0418429858982563, + 0.022272540256381035, + 0.028122849762439728, + 0.028546836227178574, + 0.010467255488038063, + 0.013336863368749619, + 0.006333203986287117, + 0.009061494842171669, + 0.00871866662055254, + 0.005255526397377253, + 0.007241145242005587, + 0.005948783829808235, + 0.0038516977801918983, + 0.0048157526180148125, + 0.0036674728617072105, + 0.005747886840254068, + 0.006215122062712908, + 0.006060340907424688, + 0.007474229671061039, + 0.003665066557005048, + 0.0047231107018888, + 0.009170805104076862 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 7, + 4 + ], + "selected_reward": -0.005272530950605869 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a streamlined, deep cherry red sports car?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the car moving at high speed through a deep, straight, modern concrete tunnel?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the car's LED taillights appear as bright red streaks trailing behind due to motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the rows of cool white LEDs on the tunnel walls visually stretched into continuous horizontal streaks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot captured from a low-angle perspective with the camera tracking in sync with the car?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.007631551939994097 + ], + "opposite_logprob_tokens": [ + -4.882631778717041 + ], + "logprob_reward": -0.007631551939994097, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.007631551939994097, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.18636998534202576, + 0.09262078255414963, + 0.0927126407623291, + 0.11079250276088715, + 0.061145152896642685, + 0.05505428835749626, + 0.06599437445402145, + 0.11541567742824554, + 0.11823908984661102, + 0.1016555055975914 + ], + "credit_map": [ + 0.025343233719468117, + 0.04703480005264282, + 0.02443794347345829, + 0.015763111412525177, + 0.020745432004332542, + 0.01852029375731945, + 0.015388749539852142, + 0.019136417657136917, + 0.015472440049052238, + 0.00967122707515955, + 0.02119467593729496, + 0.011569265276193619, + 0.010173222981393337, + 0.010468767024576664, + 0.008003327064216137, + 0.006067853886634111, + 0.014784160070121288, + 0.013029691763222218, + 0.014513922855257988, + 0.008817365393042564, + 0.012115996330976486, + 0.01074373722076416, + 0.009887231513857841, + 0.008820532821118832, + 0.01741999015212059, + 0.013937316834926605, + 0.02328023873269558, + 0.01150359958410263, + 0.012916252948343754, + 0.013992945663630962, + 0.009974216111004353, + 0.007767941802740097, + 0.006197092588990927, + 0.006203958299010992, + 0.006410016678273678, + 0.005851926747709513, + 0.008118360303342342, + 0.005730421748012304, + 0.013069909065961838, + 0.009563472121953964, + 0.0053068227134644985, + 0.006976924370974302, + 0.011593499220907688, + 0.006212982814759016, + 0.005544451996684074, + 0.005081410985440016, + 0.00620874622836709, + 0.008129446767270565, + 0.010067294351756573, + 0.005964092910289764, + 0.011630657128989697, + 0.006479052826762199, + 0.008385876193642616, + 0.009497599676251411, + 0.005609022453427315, + 0.008360779844224453, + 0.01282467320561409, + 0.017348960041999817, + 0.02312507852911949, + 0.009664728306233883, + 0.016630636528134346, + 0.010680385865271091, + 0.011564071290194988, + 0.013577142730355263, + 0.01904662325978279, + 0.01714586839079857, + 0.01843642070889473, + 0.014510276727378368, + 0.017492106184363365, + 0.015130712650716305, + 0.008348880335688591, + 0.008128196001052856, + 0.00955157820135355, + 0.011905727908015251, + 0.013622498139739037, + 0.015025751665234566, + 0.015571566298604012, + 0.010510388761758804, + 0.009417418390512466, + 0.016050567850470543 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 7 + ], + "selected_reward": -0.007631551939994097 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a streamlined, deep cherry red sports car?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the car moving at high speed through a deep, straight, modern concrete tunnel?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the car's LED taillights appear as bright red streaks trailing behind due to motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the rows of cool white LEDs on the tunnel walls visually stretched into continuous horizontal streaks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot captured from a low-angle perspective with the camera tracking in sync with the car?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0009190387791022658 + ], + "opposite_logprob_tokens": [ + -7.000918865203857 + ], + "logprob_reward": -0.0009190387791022658, + "yes_no_margin_reward": 7.0, + "token_credit_reward": -0.0009190387791022658, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1976437121629715, + 0.11827000975608826, + 0.037904947996139526, + 0.17013762891292572, + 0.0440322645008564, + 0.07383754104375839, + 0.056657448410987854, + 0.18206652998924255, + 0.06465011090040207, + 0.05479978770017624 + ], + "credit_map": [ + 0.024166513234376907, + 0.030667850747704506, + 0.02738877199590206, + 0.020055729895830154, + 0.021716639399528503, + 0.028685539960861206, + 0.021979698911309242, + 0.022982964292168617, + 0.01473874133080244, + 0.033159103244543076, + 0.014332907274365425, + 0.011976713314652443, + 0.0181821770966053, + 0.010825126431882381, + 0.007989626377820969, + 0.007065618876367807, + 0.0044641681015491486, + 0.004528640769422054, + 0.006015090271830559, + 0.006557227578014135, + 0.004221028182655573, + 0.005200307350605726, + 0.003759452374652028, + 0.0031590317375957966, + 0.027660146355628967, + 0.023590557277202606, + 0.03156473860144615, + 0.01780792511999607, + 0.01932164654135704, + 0.0193727258592844, + 0.018439913168549538, + 0.012379969470202923, + 0.0040328227914869785, + 0.004901694133877754, + 0.009583481587469578, + 0.0048769316636025906, + 0.006386046297848225, + 0.004334838595241308, + 0.005179197993129492, + 0.004737249109894037, + 0.008453111164271832, + 0.01487194187939167, + 0.013706022873520851, + 0.007974956184625626, + 0.009098808281123638, + 0.009556520730257034, + 0.005845916923135519, + 0.004330258816480637, + 0.009249769151210785, + 0.006108418107032776, + 0.008937196806073189, + 0.0061658830381929874, + 0.006999582517892122, + 0.010248220525681973, + 0.004896997939795256, + 0.004051380790770054, + 0.031695276498794556, + 0.029466666281223297, + 0.03457949310541153, + 0.014952982775866985, + 0.02456037700176239, + 0.022094234824180603, + 0.013013431802392006, + 0.011704068630933762, + 0.007250321097671986, + 0.009837710298597813, + 0.011349483393132687, + 0.007369210943579674, + 0.00898986216634512, + 0.008768468163907528, + 0.0043436074629426, + 0.006741445045918226, + 0.006827807519584894, + 0.0059703560546040535, + 0.00952150672674179, + 0.006647690199315548, + 0.007176490966230631, + 0.005466397386044264, + 0.006325935013592243, + 0.006863606162369251 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 7, + 3 + ], + "selected_reward": -0.0009190387791022658 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a streamlined, deep cherry red sports car?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the car moving at high speed through a deep, straight, modern concrete tunnel?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the car's LED taillights appear as bright red streaks trailing behind due to motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the rows of cool white LEDs on the tunnel walls visually stretched into continuous horizontal streaks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot captured from a low-angle perspective with the camera tracking in sync with the car?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.007623744197189808 + ], + "opposite_logprob_tokens": [ + -4.882623672485352 + ], + "logprob_reward": -0.007623744197189808, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.007623744197189808, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.16699250042438507, + 0.14325553178787231, + 0.07861366868019104, + 0.10371693968772888, + 0.06850547343492508, + 0.05952375382184982, + 0.06026872247457504, + 0.19377218186855316, + 0.06133894994854927, + 0.064012311398983 + ], + "credit_map": [ + 0.026813486590981483, + 0.01963081583380699, + 0.020826850086450577, + 0.014394300058484077, + 0.015454297885298729, + 0.026441583409905434, + 0.018732499331235886, + 0.02469867467880249, + 0.026848269626498222, + 0.014176099561154842, + 0.03801659122109413, + 0.013206989504396915, + 0.01681945100426674, + 0.011850479058921337, + 0.010041913948953152, + 0.012295749969780445, + 0.007534509990364313, + 0.009528697468340397, + 0.009749784134328365, + 0.007735949009656906, + 0.011566960252821445, + 0.009821927174925804, + 0.011084072291851044, + 0.011591771617531776, + 0.0183456689119339, + 0.010226727463304996, + 0.010274309664964676, + 0.008966776542365551, + 0.0138771440833807, + 0.010355773381888866, + 0.015252770856022835, + 0.016417769715189934, + 0.006707820110023022, + 0.007447468116879463, + 0.01020753476768732, + 0.010648488998413086, + 0.008565125055611134, + 0.00582404900342226, + 0.010975304991006851, + 0.008129679597914219, + 0.010916960425674915, + 0.006600182503461838, + 0.008847171440720558, + 0.008123631589114666, + 0.006089587230235338, + 0.009132562205195427, + 0.005336236208677292, + 0.004477422218769789, + 0.006716282572597265, + 0.006005392409861088, + 0.01011352613568306, + 0.005205568391829729, + 0.008374599739909172, + 0.006277395877987146, + 0.005670040380209684, + 0.011905918829143047, + 0.031251147389411926, + 0.030284307897090912, + 0.03556222841143608, + 0.016043661162257195, + 0.02566811814904213, + 0.025084467604756355, + 0.01605439931154251, + 0.013823836110532284, + 0.006457704585045576, + 0.0070119029842317104, + 0.011200038716197014, + 0.008065726608037949, + 0.009586223401129246, + 0.006257941480726004, + 0.004812072962522507, + 0.007947339676320553, + 0.0068132588639855385, + 0.009303082711994648, + 0.009705213829874992, + 0.008488320745527744, + 0.008503656834363937, + 0.00788160040974617, + 0.006750423461198807, + 0.006566748954355717 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 0, + 1 + ], + "selected_reward": -0.007623744197189808 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "subsequent low-angle tracking shot captures a streamlined, deep cherry red sports car hurtling forward at breakneck speed through a deep, straight, modern concrete tunnel. The car's mirror-smooth paint reflects the surrounding light, and its signature LED taillights illuminate, trailing two bright red streaks of light in their rapid motion. On either side, rows of cool white LEDs on the tunnel walls are visually stretched and merged by their rapid relative speed, ultimately converging into two continuous, dazzling horizontal streaks of light that sweep backwards, interplaying with the reflections on the wet, dark asphalt. Because the camera moves in sync with the car, the body remains relatively stable in the frame, while the light in the background transforms into dynamic lines. This striking contrast between movement and stillness creates an adrenaline-pumping visual impact. The entire scene, rendered in cold tones, is imbued with a futuristic, austere beauty and a sense of pure speed and power.", + "seed": 27454, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000071/group_2/index_9759_seed_27454_dp_4_rank_8_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 2, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.003740993235260248, + "vlm_logprob": -0.003740993235260248, + "vlm_yes_no_margin": 5.975, + "vlm_token_credit": -0.003740993235260248, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a streamlined, deep cherry red sports car?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the car moving at high speed through a deep, straight, modern concrete tunnel?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the car's LED taillights appear as bright red streaks trailing behind due to motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the rows of cool white LEDs on the tunnel walls visually stretched into continuous horizontal streaks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot captured from a low-angle perspective with the camera tracking in sync with the car?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005563489627093077 + ], + "opposite_logprob_tokens": [ + -6.255563259124756 + ], + "logprob_reward": -0.005563489627093077, + "yes_no_margin_reward": 6.25, + "token_credit_reward": -0.005563489627093077, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1461329460144043, + 0.08185704797506332, + 0.1656285673379898, + 0.17824551463127136, + 0.06770843267440796, + 0.04030345007777214, + 0.08243193477392197, + 0.0946727767586708, + 0.06024475023150444, + 0.0827745571732521 + ], + "credit_map": [ + 0.016844529658555984, + 0.013147417455911636, + 0.023302383720874786, + 0.022714143618941307, + 0.009080447256565094, + 0.013154184445738792, + 0.02507963590323925, + 0.02281019650399685, + 0.017629092559218407, + 0.00640886789187789, + 0.014375331811606884, + 0.010669187642633915, + 0.007277844473719597, + 0.009183565154671669, + 0.006986408960074186, + 0.00932675413787365, + 0.032722096890211105, + 0.010965782217681408, + 0.018739929422736168, + 0.020210841670632362, + 0.01966892182826996, + 0.01634550653398037, + 0.023973967880010605, + 0.02300151437520981, + 0.02621149644255638, + 0.013851748779416084, + 0.018018238246440887, + 0.028680935502052307, + 0.021254412829875946, + 0.016263585537672043, + 0.029365139082074165, + 0.024599960073828697, + 0.009473346173763275, + 0.00593955535441637, + 0.008301075547933578, + 0.008738287724554539, + 0.009968041442334652, + 0.007102601695805788, + 0.00968682486563921, + 0.008498696610331535, + 0.0034651365131139755, + 0.0031589814461767673, + 0.006366661284118891, + 0.007328665815293789, + 0.004379313439130783, + 0.0037261536344885826, + 0.006979693192988634, + 0.0048988452181220055, + 0.00929370429366827, + 0.008257435634732246, + 0.01340459007769823, + 0.01626189984381199, + 0.00701395608484745, + 0.006168967578560114, + 0.01243918389081955, + 0.009592199698090553, + 0.021305251866579056, + 0.016107726842164993, + 0.008386324159801006, + 0.007663315627723932, + 0.010387338697910309, + 0.009373530745506287, + 0.007400008849799633, + 0.014049291610717773, + 0.011625520884990692, + 0.005786910653114319, + 0.007570771966129541, + 0.0064194081351161, + 0.009748995304107666, + 0.005985687952488661, + 0.007370111998170614, + 0.005737344268709421, + 0.009106600657105446, + 0.01445390097796917, + 0.010272903367877007, + 0.014226404018700123, + 0.005250300746411085, + 0.005956001579761505, + 0.01185618806630373, + 0.011652258224785328 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 2, + 0 + ], + "selected_reward": -0.005563489627093077 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a streamlined, deep cherry red sports car?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the car moving at high speed through a deep, straight, modern concrete tunnel?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the car's LED taillights appear as bright red streaks trailing behind due to motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the rows of cool white LEDs on the tunnel walls visually stretched into continuous horizontal streaks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot captured from a low-angle perspective with the camera tracking in sync with the car?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0022109129931777716 + ], + "opposite_logprob_tokens": [ + -6.127211093902588 + ], + "logprob_reward": -0.0022109129931777716, + "yes_no_margin_reward": 6.125, + "token_credit_reward": -0.0022109129931777716, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0991717055439949, + 0.07704323530197144, + 0.35407087206840515, + 0.11103793978691101, + 0.08708342164754868, + 0.05294771119952202, + 0.017578328028321266, + 0.03791068121790886, + 0.016794981434941292, + 0.14636120200157166 + ], + "credit_map": [ + 0.011633049696683884, + 0.007540998049080372, + 0.02322820946574211, + 0.012852869927883148, + 0.004528687335550785, + 0.009577016346156597, + 0.010657954029738903, + 0.019152915105223656, + 0.014748231507837772, + 0.005712841637432575, + 0.012512284331023693, + 0.009128228761255741, + 0.011069846339523792, + 0.008422085084021091, + 0.006418774835765362, + 0.009030942805111408, + 0.06682507693767548, + 0.025554651394486427, + 0.03520731255412102, + 0.06054344028234482, + 0.05063186585903168, + 0.03212398663163185, + 0.0508519783616066, + 0.03233254700899124, + 0.01665756292641163, + 0.009071053005754948, + 0.009687545709311962, + 0.018992096185684204, + 0.01335563138127327, + 0.010701409541070461, + 0.0190690066665411, + 0.013503625057637691, + 0.012539788149297237, + 0.0079512232914567, + 0.01179223507642746, + 0.014414460398256779, + 0.00836782343685627, + 0.006804059259593487, + 0.013238883577287197, + 0.01197495311498642, + 0.007435098756104708, + 0.00381270470097661, + 0.007286683190613985, + 0.005210473667830229, + 0.007730874698609114, + 0.005270465277135372, + 0.007428424432873726, + 0.00877298228442669, + 0.0020078455563634634, + 0.0019157517235726118, + 0.002911975607275963, + 0.002815528539940715, + 0.002082000719383359, + 0.0016377149149775505, + 0.0022422191686928272, + 0.0019652899354696274, + 0.006574505474418402, + 0.004091115668416023, + 0.0038929933216422796, + 0.003767716698348522, + 0.006896183826029301, + 0.004332014359533787, + 0.004073718097060919, + 0.004282434470951557, + 0.0025587796699255705, + 0.0012232124572619796, + 0.0019047685200348496, + 0.002087769564241171, + 0.0026038342621177435, + 0.0018656211905181408, + 0.002245664829388261, + 0.0023053311742842197, + 0.014924867078661919, + 0.013190245255827904, + 0.017653649672865868, + 0.031837571412324905, + 0.011613260954618454, + 0.01247834786772728, + 0.027560055255889893, + 0.017103195190429688 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 9, + 3 + ], + "selected_reward": -0.0022109129931777716 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a streamlined, deep cherry red sports car?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the car moving at high speed through a deep, straight, modern concrete tunnel?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the car's LED taillights appear as bright red streaks trailing behind due to motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the rows of cool white LEDs on the tunnel walls visually stretched into continuous horizontal streaks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot captured from a low-angle perspective with the camera tracking in sync with the car?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.00463993102312088 + ], + "opposite_logprob_tokens": [ + -5.379640102386475 + ], + "logprob_reward": -0.00463993102312088, + "yes_no_margin_reward": 5.375, + "token_credit_reward": -0.00463993102312088, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14742347598075867, + 0.06908617913722992, + 0.17914369702339172, + 0.09561803191900253, + 0.05987543240189552, + 0.07306569814682007, + 0.09959471970796585, + 0.06044522672891617, + 0.07069718092679977, + 0.1450503170490265 + ], + "credit_map": [ + 0.014603343792259693, + 0.01609056629240513, + 0.02423660270869732, + 0.027232682332396507, + 0.006418408360332251, + 0.015246574766933918, + 0.019426368176937103, + 0.024168940261006355, + 0.013986741192638874, + 0.006366862449795008, + 0.010343095287680626, + 0.006675137672573328, + 0.00910856667906046, + 0.007669032551348209, + 0.007170989643782377, + 0.00776576716452837, + 0.03067382238805294, + 0.012620664201676846, + 0.017563901841640472, + 0.0320177786052227, + 0.02621869556605816, + 0.01668165996670723, + 0.026241112500429153, + 0.01712607592344284, + 0.013975469395518303, + 0.0073965867049992085, + 0.01314997673034668, + 0.00987141765654087, + 0.014077134430408478, + 0.009454149752855301, + 0.01076631061732769, + 0.016926996409893036, + 0.005823611281812191, + 0.006443357095122337, + 0.011039244011044502, + 0.010902001522481441, + 0.004299681168049574, + 0.004601524211466312, + 0.008474834263324738, + 0.008291186764836311, + 0.009149412624537945, + 0.005111381411552429, + 0.013873714953660965, + 0.011271687224507332, + 0.007580445613712072, + 0.005718829110264778, + 0.010893200524151325, + 0.00946703739464283, + 0.013420095667243004, + 0.009359309449791908, + 0.01346613559871912, + 0.01836390048265457, + 0.008856955915689468, + 0.007696943357586861, + 0.014955773949623108, + 0.01347561739385128, + 0.008112694136798382, + 0.0061071570962667465, + 0.006002391688525677, + 0.008986985310912132, + 0.0077486238442361355, + 0.0055340067483484745, + 0.012457292526960373, + 0.005496087018400431, + 0.011716158129274845, + 0.00835128128528595, + 0.007726478390395641, + 0.008471976965665817, + 0.010041057132184505, + 0.007444822695106268, + 0.008964147418737411, + 0.007981265895068645, + 0.015238756313920021, + 0.01297612302005291, + 0.01771646924316883, + 0.029073642566800117, + 0.01097667496651411, + 0.011827877722680569, + 0.027719952166080475, + 0.0195208378136158 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 0, + 9 + ], + "selected_reward": -0.00463993102312088 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a streamlined, deep cherry red sports car?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the car moving at high speed through a deep, straight, modern concrete tunnel?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the car's LED taillights appear as bright red streaks trailing behind due to motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the rows of cool white LEDs on the tunnel walls visually stretched into continuous horizontal streaks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot captured from a low-angle perspective with the camera tracking in sync with the car?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0010407513473182917 + ], + "opposite_logprob_tokens": [ + -6.876040935516357 + ], + "logprob_reward": -0.0010407513473182917, + "yes_no_margin_reward": 6.875, + "token_credit_reward": -0.0010407513473182917, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11108594387769699, + 0.05918734893202782, + 0.16503562033176422, + 0.19116058945655823, + 0.16266873478889465, + 0.13759206235408783, + 0.04359843209385872, + 0.026948897168040276, + 0.020826751366257668, + 0.08189564198255539 + ], + "credit_map": [ + 0.013563721440732479, + 0.006806535180658102, + 0.027944941073656082, + 0.01232326589524746, + 0.004576018545776606, + 0.009889663197100163, + 0.012421539053320885, + 0.023560266941785812, + 0.01140572689473629, + 0.004880176857113838, + 0.011237907223403454, + 0.005530291236937046, + 0.007024458143860102, + 0.005908428691327572, + 0.0058147720992565155, + 0.007385586388409138, + 0.030600393190979958, + 0.011489261873066425, + 0.021593481302261353, + 0.018722346052527428, + 0.024671645835042, + 0.018147602677345276, + 0.016886018216609955, + 0.022924864664673805, + 0.03228737786412239, + 0.015118190087378025, + 0.01793321780860424, + 0.031771235167980194, + 0.025958003476262093, + 0.01971202902495861, + 0.02742747962474823, + 0.020953072234988213, + 0.019573811441659927, + 0.011395455338060856, + 0.02137693017721176, + 0.028636611998081207, + 0.014893515966832638, + 0.01251951977610588, + 0.030698638409376144, + 0.023574257269501686, + 0.01620899699628353, + 0.008462844416499138, + 0.021061532199382782, + 0.017033493146300316, + 0.016530567780137062, + 0.01349124126136303, + 0.020963400602340698, + 0.023839978501200676, + 0.004395856522023678, + 0.004410114139318466, + 0.00840147677809, + 0.005666801240295172, + 0.005480467341840267, + 0.0045120855793356895, + 0.004403086844831705, + 0.006328545045107603, + 0.00462803291156888, + 0.002528527984395623, + 0.002327834954485297, + 0.00321890483610332, + 0.004464218858629465, + 0.002907518297433853, + 0.0038807177916169167, + 0.0029931431636214256, + 0.0033155239652842283, + 0.0026011362206190825, + 0.0020741033367812634, + 0.0024350606836378574, + 0.003470180556178093, + 0.002349117072299123, + 0.0023356154561042786, + 0.002246016403660178, + 0.00857518706470728, + 0.008912978693842888, + 0.009533264674246311, + 0.017178812995553017, + 0.006312328856438398, + 0.006568402983248234, + 0.014976443722844124, + 0.009838225319981575 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 2, + 4 + ], + "selected_reward": -0.0010407513473182917 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a streamlined, deep cherry red sports car?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the car moving at high speed through a deep, straight, modern concrete tunnel?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the car's LED taillights appear as bright red streaks trailing behind due to motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the rows of cool white LEDs on the tunnel walls visually stretched into continuous horizontal streaks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot captured from a low-angle perspective with the camera tracking in sync with the car?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005249881185591221 + ], + "opposite_logprob_tokens": [ + -5.255249977111816 + ], + "logprob_reward": -0.005249881185591221, + "yes_no_margin_reward": 5.25, + "token_credit_reward": -0.005249881185591221, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10769909620285034, + 0.05515710636973381, + 0.14823952317237854, + 0.033365458250045776, + 0.04857028275728226, + 0.06839168071746826, + 0.09691280871629715, + 0.034597501158714294, + 0.0399494394659996, + 0.36711710691452026 + ], + "credit_map": [ + 0.011251334100961685, + 0.009789920412003994, + 0.019040923565626144, + 0.01529628224670887, + 0.008035649545490742, + 0.010556585155427456, + 0.014824164099991322, + 0.018904227763414383, + 0.010654378682374954, + 0.006232692394405603, + 0.007454800885170698, + 0.005674996878951788, + 0.006520633120089769, + 0.0058820657432079315, + 0.006213885731995106, + 0.006523648276925087, + 0.02707255817949772, + 0.010243666358292103, + 0.015359936282038689, + 0.021527118980884552, + 0.026226921007037163, + 0.01549546793103218, + 0.016575563699007034, + 0.015738267451524734, + 0.0044996365904808044, + 0.001876385067589581, + 0.005619904492050409, + 0.0052202255465090275, + 0.0026273452676832676, + 0.0026981448754668236, + 0.005769206210970879, + 0.00505460612475872, + 0.005589839071035385, + 0.004429117310792208, + 0.00593773415312171, + 0.008829678408801556, + 0.003973035141825676, + 0.0030211806297302246, + 0.010211911052465439, + 0.006577780470252037, + 0.0075628506019711494, + 0.004525312222540379, + 0.012160777114331722, + 0.009548943489789963, + 0.009093254804611206, + 0.006227951962500811, + 0.009578236378729343, + 0.009694352746009827, + 0.012959444895386696, + 0.008694814518094063, + 0.01810516230762005, + 0.016559358686208725, + 0.008351166732609272, + 0.007935980334877968, + 0.011197257786989212, + 0.013109607622027397, + 0.006331633310765028, + 0.005671271588653326, + 0.0029018432833254337, + 0.004246315453201532, + 0.003953424282371998, + 0.0032231647055596113, + 0.003833698108792305, + 0.0044361441396176815, + 0.006799245253205299, + 0.003663169452920556, + 0.0034069838002324104, + 0.0029868592973798513, + 0.009545153006911278, + 0.005362054333090782, + 0.0033002712298184633, + 0.0048856972716748714, + 0.03904154524207115, + 0.03191166743636131, + 0.04262283444404602, + 0.07779578864574432, + 0.02901625819504261, + 0.03149848058819771, + 0.07027758657932281, + 0.04495290666818619 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 2, + 0 + ], + "selected_reward": -0.005249881185591221 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "subsequent low-angle tracking shot captures a streamlined, deep cherry red sports car hurtling forward at breakneck speed through a deep, straight, modern concrete tunnel. The car's mirror-smooth paint reflects the surrounding light, and its signature LED taillights illuminate, trailing two bright red streaks of light in their rapid motion. On either side, rows of cool white LEDs on the tunnel walls are visually stretched and merged by their rapid relative speed, ultimately converging into two continuous, dazzling horizontal streaks of light that sweep backwards, interplaying with the reflections on the wet, dark asphalt. Because the camera moves in sync with the car, the body remains relatively stable in the frame, while the light in the background transforms into dynamic lines. This striking contrast between movement and stillness creates an adrenaline-pumping visual impact. The entire scene, rendered in cold tones, is imbued with a futuristic, austere beauty and a sense of pure speed and power.", + "seed": 27455, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000071/group_2/index_9759_seed_27455_dp_5_rank_10_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 2, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.009057027660310269, + "vlm_logprob": -0.009057027660310269, + "vlm_yes_no_margin": 5.6, + "vlm_token_credit": -0.009057027660310269, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a streamlined, deep cherry red sports car?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the car moving at high speed through a deep, straight, modern concrete tunnel?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the car's LED taillights appear as bright red streaks trailing behind due to motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the rows of cool white LEDs on the tunnel walls visually stretched into continuous horizontal streaks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot captured from a low-angle perspective with the camera tracking in sync with the car?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.007365098223090172 + ], + "opposite_logprob_tokens": [ + -7.3823652267456055 + ], + "logprob_reward": -0.007365098223090172, + "yes_no_margin_reward": 7.375, + "token_credit_reward": -0.007365098223090172, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07278085500001907, + 0.30625224113464355, + 0.08644837886095047, + 0.13575489819049835, + 0.05128095671534538, + 0.06230014190077782, + 0.0775759369134903, + 0.08266688138246536, + 0.07525617629289627, + 0.04968353360891342 + ], + "credit_map": [ + 0.011710683815181255, + 0.006215917877852917, + 0.01046953909099102, + 0.008363582193851471, + 0.008085175417363644, + 0.009353450499475002, + 0.009760407730937004, + 0.008822101168334484, + 0.02888897992670536, + 0.014247316867113113, + 0.10505824536085129, + 0.010064399801194668, + 0.01497384998947382, + 0.11804159730672836, + 0.0060431137681007385, + 0.008934727869927883, + 0.014012148603796959, + 0.013850158080458641, + 0.011013785377144814, + 0.010312222875654697, + 0.009659389033913612, + 0.01237128209322691, + 0.006898514926433563, + 0.008330874145030975, + 0.028399089351296425, + 0.011421910487115383, + 0.026682134717702866, + 0.011291748844087124, + 0.016399014741182327, + 0.024292850866913795, + 0.007476270664483309, + 0.009791876189410686, + 0.009032081812620163, + 0.004030624404549599, + 0.006568323355168104, + 0.005230261478573084, + 0.0047730025835335255, + 0.011070952750742435, + 0.00501665286719799, + 0.00555905606597662, + 0.00487477658316493, + 0.002971776993945241, + 0.01691523753106594, + 0.004695606883615255, + 0.0023775750305503607, + 0.021434759721159935, + 0.004108206834644079, + 0.004922200925648212, + 0.009012360125780106, + 0.009422454051673412, + 0.009406164288520813, + 0.011968863196671009, + 0.007034554611891508, + 0.008937820792198181, + 0.013124899938702583, + 0.008668816648423672, + 0.012027154676616192, + 0.01636318676173687, + 0.014120807871222496, + 0.008651836775243282, + 0.0076368991285562515, + 0.009920681826770306, + 0.006913547869771719, + 0.007032762281596661, + 0.008786005899310112, + 0.00874991063028574, + 0.016863737255334854, + 0.011320710182189941, + 0.005875857081264257, + 0.007594508584588766, + 0.010042373090982437, + 0.00602307403460145, + 0.0071237701922655106, + 0.006569667719304562, + 0.011922881007194519, + 0.00518416240811348, + 0.003601853968575597, + 0.009159539826214314, + 0.003166675567626953, + 0.002954985713586211 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 3, + 2 + ], + "selected_reward": -0.007365098223090172 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a streamlined, deep cherry red sports car?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the car moving at high speed through a deep, straight, modern concrete tunnel?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the car's LED taillights appear as bright red streaks trailing behind due to motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the rows of cool white LEDs on the tunnel walls visually stretched into continuous horizontal streaks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot captured from a low-angle perspective with the camera tracking in sync with the car?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.020593751221895218 + ], + "opposite_logprob_tokens": [ + -3.8955936431884766 + ], + "logprob_reward": -0.020593751221895218, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.020593751221895218, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.04067665711045265, + 0.49340030550956726, + 0.022658204659819603, + 0.09075994789600372, + 0.05422187224030495, + 0.011151603423058987, + 0.02178965136408806, + 0.02383820153772831, + 0.09651828557252884, + 0.14498522877693176 + ], + "credit_map": [ + 0.005023872014135122, + 0.002804335206747055, + 0.006555531639605761, + 0.0038172847125679255, + 0.002590357093140483, + 0.005700963549315929, + 0.005293136928230524, + 0.008891178295016289, + 0.04731552675366402, + 0.023062311112880707, + 0.1683310717344284, + 0.015711115673184395, + 0.02480323612689972, + 0.19061091542243958, + 0.009597034193575382, + 0.013969109393656254, + 0.0035112451296299696, + 0.002940875943750143, + 0.0030018009711056948, + 0.002446624916046858, + 0.0027351605240255594, + 0.0032355005387216806, + 0.0020296606235206127, + 0.0027573355473577976, + 0.019779860973358154, + 0.008206167258322239, + 0.017336949706077576, + 0.007767997216433287, + 0.010244286619126797, + 0.015938449651002884, + 0.005503889173269272, + 0.005982345435768366, + 0.0061780717223882675, + 0.004005379043519497, + 0.008903197944164276, + 0.008025065064430237, + 0.0035449822898954153, + 0.010500027798116207, + 0.007791516371071339, + 0.005273628514260054, + 0.0010509983403608203, + 0.0012403022265061736, + 0.0011309348046779633, + 0.0015545348869636655, + 0.0015670446446165442, + 0.0010500685311853886, + 0.0021095238626003265, + 0.0014481963589787483, + 0.0017904572887346148, + 0.002355196513235569, + 0.0026609571650624275, + 0.004026948474347591, + 0.001500377431511879, + 0.0022265964653342962, + 0.0037916514556854963, + 0.0034374664537608624, + 0.002019955310970545, + 0.002619087463244796, + 0.0052000307478010654, + 0.002162336139008403, + 0.002726449863985181, + 0.003395208390429616, + 0.0027378553058952093, + 0.0029772804118692875, + 0.016524428501725197, + 0.009228014387190342, + 0.025075672194361687, + 0.010035092942416668, + 0.009483195841312408, + 0.00989373680204153, + 0.00763153750449419, + 0.008646615780889988, + 0.02214185893535614, + 0.015231424942612648, + 0.04226592183113098, + 0.01175347063690424, + 0.010123283602297306, + 0.020077913999557495, + 0.007699396461248398, + 0.01569196581840515 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 9, + 8 + ], + "selected_reward": -0.020593751221895218 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a streamlined, deep cherry red sports car?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the car moving at high speed through a deep, straight, modern concrete tunnel?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the car's LED taillights appear as bright red streaks trailing behind due to motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the rows of cool white LEDs on the tunnel walls visually stretched into continuous horizontal streaks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot captured from a low-angle perspective with the camera tracking in sync with the car?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.012536413036286831 + ], + "opposite_logprob_tokens": [ + -4.387536525726318 + ], + "logprob_reward": -0.012536413036286831, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.012536413036286831, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.026360198855400085, + 0.72190260887146, + 0.044681988656520844, + 0.022939732298254967, + 0.014998417347669601, + 0.027872005477547646, + 0.0816536471247673, + 0.013528929091989994, + 0.02880457602441311, + 0.017257943749427795 + ], + "credit_map": [ + 0.0029181481804698706, + 0.0020902298856526613, + 0.0035750858951359987, + 0.002752403961494565, + 0.002186410827562213, + 0.0034296896774321795, + 0.003870507003739476, + 0.00553772272542119, + 0.06823991239070892, + 0.03373098745942116, + 0.24792198836803436, + 0.023268960416316986, + 0.036820922046899796, + 0.277513712644577, + 0.014018182642757893, + 0.020387955009937286, + 0.00474758492782712, + 0.0033300842624157667, + 0.009867079555988312, + 0.0035624823067337275, + 0.005502027925103903, + 0.007743620313704014, + 0.005630321800708771, + 0.004298784304410219, + 0.003813348710536957, + 0.001737446989864111, + 0.0046994686126708984, + 0.002408687025308609, + 0.0020714348647743464, + 0.00450056092813611, + 0.002136792754754424, + 0.0015719917137175798, + 0.001602978096343577, + 0.0012190226698294282, + 0.0022201715037226677, + 0.0023960433900356293, + 0.0009591338457539678, + 0.00209173490293324, + 0.0026575271040201187, + 0.0018518054857850075, + 0.001935335574671626, + 0.0011583482846617699, + 0.009078774601221085, + 0.001983762253075838, + 0.0011396785266697407, + 0.009664413519203663, + 0.0011695880675688386, + 0.0017421046504750848, + 0.0071459487080574036, + 0.010593998245894909, + 0.009661818854510784, + 0.009874679148197174, + 0.008081654086709023, + 0.0074581727385520935, + 0.015438303351402283, + 0.013399074785411358, + 0.0011496563674882054, + 0.0011500605614855886, + 0.0021597195882350206, + 0.0025237291119992733, + 0.0007642586715519428, + 0.001348231453448534, + 0.003049445804208517, + 0.0013838261365890503, + 0.004010748118162155, + 0.0025097376201301813, + 0.007202304899692535, + 0.0034860987216234207, + 0.002200647722929716, + 0.004434553906321526, + 0.0020605947356671095, + 0.0028998900670558214, + 0.0026165388990193605, + 0.0029242646414786577, + 0.003432187717407942, + 0.0018138069426640868, + 0.0012708524009212852, + 0.00189467694144696, + 0.0018211712595075369, + 0.0014844441320747137 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 6, + 2 + ], + "selected_reward": -0.012536413036286831 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a streamlined, deep cherry red sports car?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the car moving at high speed through a deep, straight, modern concrete tunnel?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the car's LED taillights appear as bright red streaks trailing behind due to motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the rows of cool white LEDs on the tunnel walls visually stretched into continuous horizontal streaks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot captured from a low-angle perspective with the camera tracking in sync with the car?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.001177809201180935 + ], + "opposite_logprob_tokens": [ + -6.751177787780762 + ], + "logprob_reward": -0.001177809201180935, + "yes_no_margin_reward": 6.75, + "token_credit_reward": -0.001177809201180935, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.050873517990112305, + 0.690360963344574, + 0.020085612311959267, + 0.05587667599320412, + 0.025187930092215538, + 0.037996143102645874, + 0.05362434685230255, + 0.018313396722078323, + 0.01932685263454914, + 0.0283545833081007 + ], + "credit_map": [ + 0.007706996984779835, + 0.004178302828222513, + 0.008127672597765923, + 0.004160246811807156, + 0.003127309260889888, + 0.0069326357915997505, + 0.005441757384687662, + 0.01119859516620636, + 0.06515049189329147, + 0.032070111483335495, + 0.23831035196781158, + 0.022045699879527092, + 0.034066230058670044, + 0.26541218161582947, + 0.01406058482825756, + 0.019245224073529243, + 0.0020074460189789534, + 0.0016308737685903907, + 0.0036198787856847048, + 0.0022459442261606455, + 0.0021761965472251177, + 0.003283491125330329, + 0.002759008901193738, + 0.002362772822380066, + 0.01219816692173481, + 0.00537722185254097, + 0.009665253572165966, + 0.005106380209326744, + 0.006938644219189882, + 0.009019983001053333, + 0.00361881242133677, + 0.00395221496000886, + 0.0035770018585026264, + 0.001682204194366932, + 0.004146975465118885, + 0.002699886914342642, + 0.002089192159473896, + 0.006176614202558994, + 0.0019030614057555795, + 0.0029129942413419485, + 0.0027316538617014885, + 0.0011683569755405188, + 0.012707686983048916, + 0.0024970388039946556, + 0.0011941849952563643, + 0.013933137059211731, + 0.0014698068844154477, + 0.002294279867783189, + 0.0058553945273160934, + 0.00448840344324708, + 0.009928349405527115, + 0.005592801608145237, + 0.00652598449960351, + 0.007208002265542746, + 0.007139380089938641, + 0.00688603101298213, + 0.002049572067335248, + 0.0017710771644487977, + 0.0031878179870545864, + 0.0025247347075492144, + 0.0018946595955640078, + 0.0017456469358876348, + 0.0027953481767326593, + 0.002344538690522313, + 0.00327594974078238, + 0.00167950673494488, + 0.00497354194521904, + 0.0019942850340157747, + 0.0012648982228711247, + 0.0023973823990672827, + 0.0018997351871803403, + 0.0018415532540529966, + 0.004128202795982361, + 0.00446513295173645, + 0.006629474926739931, + 0.0029266118071973324, + 0.0019093978917226195, + 0.0037356598768383265, + 0.0020803383085876703, + 0.0024797634687274694 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 3, + 6 + ], + "selected_reward": -0.001177809201180935 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a streamlined, deep cherry red sports car?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the car moving at high speed through a deep, straight, modern concrete tunnel?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the car's LED taillights appear as bright red streaks trailing behind due to motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the rows of cool white LEDs on the tunnel walls visually stretched into continuous horizontal streaks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot captured from a low-angle perspective with the camera tracking in sync with the car?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0036120666190981865 + ], + "opposite_logprob_tokens": [ + -5.628612041473389 + ], + "logprob_reward": -0.0036120666190981865, + "yes_no_margin_reward": 5.625, + "token_credit_reward": -0.0036120666190981865, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.02043360285460949, + 0.7258173227310181, + 0.017942607402801514, + 0.018955493345856667, + 0.024907615035772324, + 0.010903337970376015, + 0.049188852310180664, + 0.043654076755046844, + 0.06422637403011322, + 0.023970680311322212 + ], + "credit_map": [ + 0.0024049468338489532, + 0.0014254713896661997, + 0.003760947845876217, + 0.002311210846528411, + 0.001968638738617301, + 0.0026804348453879356, + 0.0026345192454755306, + 0.00324743683449924, + 0.06890738010406494, + 0.03356754407286644, + 0.24959495663642883, + 0.023308435454964638, + 0.03589652106165886, + 0.27962905168533325, + 0.014401687309145927, + 0.020511748269200325, + 0.002158842049539089, + 0.0016214771894738078, + 0.004165643360465765, + 0.0020585970487445593, + 0.0014669430674985051, + 0.00236751651391387, + 0.0018601869232952595, + 0.002243401948362589, + 0.0036168138030916452, + 0.0021217484027147293, + 0.0038811408448964357, + 0.0017544678412377834, + 0.001752115087583661, + 0.002938502700999379, + 0.0012905530165880919, + 0.0016001518815755844, + 0.003098398679867387, + 0.0024863313883543015, + 0.004113601520657539, + 0.003201065817847848, + 0.0016876115696504712, + 0.004941871389746666, + 0.002417217707261443, + 0.0029615210369229317, + 0.0010540251387283206, + 0.0010726151522248983, + 0.001290694810450077, + 0.0012504832120612264, + 0.0013254131190478802, + 0.0013772646198049188, + 0.0017550522461533546, + 0.0017777911853045225, + 0.005723137874156237, + 0.006713644601404667, + 0.00629440788179636, + 0.00508874049410224, + 0.0048954845406115055, + 0.00471647921949625, + 0.00776654714718461, + 0.00799041148275137, + 0.004659679718315601, + 0.005925055127590895, + 0.011609597131609917, + 0.003966270945966244, + 0.003681657835841179, + 0.007330537773668766, + 0.00369498273357749, + 0.0027862980496138334, + 0.010880609042942524, + 0.009649219922721386, + 0.014622453600168228, + 0.006338353268802166, + 0.005532093811780214, + 0.005965813063085079, + 0.004937565419822931, + 0.006300271023064852, + 0.003456170205026865, + 0.0044473386369645596, + 0.005064689554274082, + 0.002563069574534893, + 0.001785743748769164, + 0.003108250442892313, + 0.001763680367730558, + 0.0017817399930208921 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 8, + 6 + ], + "selected_reward": -0.0036120666190981865 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "subsequent low-angle tracking shot captures a streamlined, deep cherry red sports car hurtling forward at breakneck speed through a deep, straight, modern concrete tunnel. The car's mirror-smooth paint reflects the surrounding light, and its signature LED taillights illuminate, trailing two bright red streaks of light in their rapid motion. On either side, rows of cool white LEDs on the tunnel walls are visually stretched and merged by their rapid relative speed, ultimately converging into two continuous, dazzling horizontal streaks of light that sweep backwards, interplaying with the reflections on the wet, dark asphalt. Because the camera moves in sync with the car, the body remains relatively stable in the frame, while the light in the background transforms into dynamic lines. This striking contrast between movement and stillness creates an adrenaline-pumping visual impact. The entire scene, rendered in cold tones, is imbued with a futuristic, austere beauty and a sense of pure speed and power.", + "seed": 27456, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000071/group_2/index_9759_seed_27456_dp_5_rank_10_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 2, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.0044108892790973185, + "vlm_logprob": -0.0044108892790973185, + "vlm_yes_no_margin": 5.8, + "vlm_token_credit": -0.0044108892790973185, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a streamlined, deep cherry red sports car?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the car moving at high speed through a deep, straight, modern concrete tunnel?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the car's LED taillights appear as bright red streaks trailing behind due to motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the rows of cool white LEDs on the tunnel walls visually stretched into continuous horizontal streaks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot captured from a low-angle perspective with the camera tracking in sync with the car?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.006112216040492058 + ], + "opposite_logprob_tokens": [ + -6.006112098693848 + ], + "logprob_reward": -0.006112216040492058, + "yes_no_margin_reward": 6.0, + "token_credit_reward": -0.006112216040492058, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1300407499074936, + 0.052347030490636826, + 0.4057992100715637, + 0.11249291151762009, + 0.023587988689541817, + 0.04542473703622818, + 0.04515788331627846, + 0.08751717954874039, + 0.07872679829597473, + 0.018905505537986755 + ], + "credit_map": [ + 0.011156686581671238, + 0.007619758602231741, + 0.017895540222525597, + 0.01884838566184044, + 0.0172283872961998, + 0.02044430561363697, + 0.02198098599910736, + 0.014866692945361137, + 0.008304283954203129, + 0.005292384885251522, + 0.006836825516074896, + 0.01091853342950344, + 0.004200823605060577, + 0.005252812057733536, + 0.006025880575180054, + 0.005515485070645809, + 0.017234398052096367, + 0.015798212960362434, + 0.1483030766248703, + 0.08402223885059357, + 0.008752545341849327, + 0.09859350323677063, + 0.022655446082353592, + 0.010439797304570675, + 0.009194396436214447, + 0.006726715248078108, + 0.034031856805086136, + 0.013514786027371883, + 0.0034197233617305756, + 0.030156433582305908, + 0.011325028724968433, + 0.004123969469219446, + 0.003162690671160817, + 0.0015142139745876193, + 0.0029209775384515524, + 0.003699303837493062, + 0.0036242599599063396, + 0.002365734660997987, + 0.00381678924895823, + 0.0024840182159096003, + 0.004986859392374754, + 0.0037051374092698097, + 0.004824293777346611, + 0.006485465914011002, + 0.006112001370638609, + 0.00563200656324625, + 0.006555435247719288, + 0.007123538292944431, + 0.0031201988458633423, + 0.002914902288466692, + 0.012654153630137444, + 0.016875334084033966, + 0.0016525635728612542, + 0.0020529290195554495, + 0.0036007948219776154, + 0.0022870053071528673, + 0.00443998584523797, + 0.006101705599576235, + 0.022141188383102417, + 0.031001852825284004, + 0.005297043826431036, + 0.004409686662256718, + 0.01009325310587883, + 0.004032467026263475, + 0.009550664573907852, + 0.007187914103269577, + 0.012098216451704502, + 0.026469649747014046, + 0.0033049804624170065, + 0.006559089757502079, + 0.007783909793943167, + 0.005772377364337444, + 0.002628989750519395, + 0.0019642887637019157, + 0.003107207827270031, + 0.00300102517940104, + 0.0019698909018188715, + 0.001740520354360342, + 0.002614443888887763, + 0.0018791388720273972 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 0, + 3 + ], + "selected_reward": -0.006112216040492058 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a streamlined, deep cherry red sports car?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the car moving at high speed through a deep, straight, modern concrete tunnel?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the car's LED taillights appear as bright red streaks trailing behind due to motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the rows of cool white LEDs on the tunnel walls visually stretched into continuous horizontal streaks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot captured from a low-angle perspective with the camera tracking in sync with the car?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.002502407180145383 + ], + "opposite_logprob_tokens": [ + -6.00250244140625 + ], + "logprob_reward": -0.002502407180145383, + "yes_no_margin_reward": 6.0, + "token_credit_reward": -0.002502407180145383, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.30370965600013733, + 0.10346679389476776, + 0.06382506340742111, + 0.3156304657459259, + 0.0223398394882679, + 0.023958779871463776, + 0.017510080710053444, + 0.04474375396966934, + 0.07268906384706497, + 0.03212643787264824 + ], + "credit_map": [ + 0.033817414194345474, + 0.01850823126733303, + 0.04210206866264343, + 0.051086872816085815, + 0.034530580043792725, + 0.058853693306446075, + 0.03548320010304451, + 0.029327601194381714, + 0.010893088765442371, + 0.006208544131368399, + 0.02014344185590744, + 0.030669333413243294, + 0.004234427586197853, + 0.014381051994860172, + 0.00672111427411437, + 0.010215790010988712, + 0.005859288852661848, + 0.0061694560572505, + 0.006064645946025848, + 0.011772673577070236, + 0.003278602147474885, + 0.020039111375808716, + 0.006049599498510361, + 0.0045916917733848095, + 0.02470175363123417, + 0.025380605831742287, + 0.09019727259874344, + 0.043121252208948135, + 0.011978685855865479, + 0.08029970526695251, + 0.028303904458880424, + 0.011647289618849754, + 0.0031066611409187317, + 0.0017077941447496414, + 0.002535624196752906, + 0.003878382034599781, + 0.0023868095595389605, + 0.0019583525136113167, + 0.002978898584842682, + 0.0037873173132538795, + 0.002835524035617709, + 0.0017042745603248477, + 0.004442069213837385, + 0.0031658087391406298, + 0.0025098235346376896, + 0.00410845410078764, + 0.0023722820915281773, + 0.002820543246343732, + 0.0031298415269702673, + 0.001478175981901586, + 0.0024036054965108633, + 0.0028461755719035864, + 0.0020763843785971403, + 0.0015581630868837237, + 0.002132367342710495, + 0.0018853654619306326, + 0.0044495658949017525, + 0.0036217316519469023, + 0.010739821940660477, + 0.0075173452496528625, + 0.0033159004524350166, + 0.004416339565068483, + 0.005821475759148598, + 0.004861571826040745, + 0.00959094613790512, + 0.006943848915398121, + 0.010955535806715488, + 0.023378154262900352, + 0.0037764471489936113, + 0.005348881706595421, + 0.006989740300923586, + 0.005705506075173616, + 0.003025263547897339, + 0.0022678321693092585, + 0.007933607324957848, + 0.007156574632972479, + 0.0018231563735753298, + 0.003458783496171236, + 0.003142046043649316, + 0.0033191752154380083 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 0, + 1 + ], + "selected_reward": -0.002502407180145383 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a streamlined, deep cherry red sports car?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the car moving at high speed through a deep, straight, modern concrete tunnel?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the car's LED taillights appear as bright red streaks trailing behind due to motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the rows of cool white LEDs on the tunnel walls visually stretched into continuous horizontal streaks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot captured from a low-angle perspective with the camera tracking in sync with the car?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.00463898153975606 + ], + "opposite_logprob_tokens": [ + -5.379639148712158 + ], + "logprob_reward": -0.00463898153975606, + "yes_no_margin_reward": 5.375, + "token_credit_reward": -0.00463898153975606, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15310455858707428, + 0.20558470487594604, + 0.32456719875335693, + 0.050195153802633286, + 0.04113153740763664, + 0.01942947879433632, + 0.03805367648601532, + 0.10939129441976547, + 0.03386707976460457, + 0.02467522583901882 + ], + "credit_map": [ + 0.017931358888745308, + 0.01058520283550024, + 0.023583970963954926, + 0.025758787989616394, + 0.01670723780989647, + 0.02406824193894863, + 0.018479321151971817, + 0.015990449115633965, + 0.017067067325115204, + 0.01350907701998949, + 0.02654414437711239, + 0.04191732779145241, + 0.010672175325453281, + 0.05647341161966324, + 0.016528282314538956, + 0.022873226553201675, + 0.012106584385037422, + 0.015472723171114922, + 0.12324535101652145, + 0.08190469443798065, + 0.010538180358707905, + 0.0509825125336647, + 0.018821489065885544, + 0.011495684273540974, + 0.004437205381691456, + 0.0057051642797887325, + 0.012574436143040657, + 0.006746422033756971, + 0.0032112840563058853, + 0.007934702560305595, + 0.005345708224922419, + 0.004240233451128006, + 0.0056469496339559555, + 0.002924403641372919, + 0.008649122901260853, + 0.007330774329602718, + 0.0037098440807312727, + 0.0038510628510266542, + 0.004156194161623716, + 0.004863192327320576, + 0.0016079790657386184, + 0.0014297992456704378, + 0.003928057383745909, + 0.004508169833570719, + 0.0024503320455551147, + 0.001363318762741983, + 0.002357830060645938, + 0.0017839949578046799, + 0.007914296351373196, + 0.004614927340298891, + 0.009794189594686031, + 0.003636779263615608, + 0.0026385916862636805, + 0.0031252887565642595, + 0.0035620725248008966, + 0.0027675346937030554, + 0.006968631874769926, + 0.009183253161609173, + 0.028304172679781914, + 0.034915611147880554, + 0.004592319950461388, + 0.0075203608721494675, + 0.011813926510512829, + 0.006093037314713001, + 0.0025842166505753994, + 0.003604486119002104, + 0.004649212583899498, + 0.006387574598193169, + 0.004579650238156319, + 0.003966392949223518, + 0.004619639832526445, + 0.0034759126137942076, + 0.003233666531741619, + 0.002684616483747959, + 0.0036571586970239878, + 0.0037894579581916332, + 0.0031707945745438337, + 0.0017945083091035485, + 0.0035101831890642643, + 0.0028348432388156652 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 1, + 0 + ], + "selected_reward": -0.00463898153975606 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a streamlined, deep cherry red sports car?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the car moving at high speed through a deep, straight, modern concrete tunnel?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the car's LED taillights appear as bright red streaks trailing behind due to motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the rows of cool white LEDs on the tunnel walls visually stretched into continuous horizontal streaks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot captured from a low-angle perspective with the camera tracking in sync with the car?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.001177571015432477 + ], + "opposite_logprob_tokens": [ + -6.751177787780762 + ], + "logprob_reward": -0.001177571015432477, + "yes_no_margin_reward": 6.75, + "token_credit_reward": -0.001177571015432477, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.05713757127523422, + 0.244858518242836, + 0.2068941295146942, + 0.13098640739917755, + 0.027926025912165642, + 0.01403775718063116, + 0.08737131208181381, + 0.13813892006874084, + 0.07901331037282944, + 0.013635988347232342 + ], + "credit_map": [ + 0.006087059620767832, + 0.004615522921085358, + 0.008829191327095032, + 0.010323649272322655, + 0.005725748836994171, + 0.008560873568058014, + 0.0057480973191559315, + 0.007247437257319689, + 0.025188427418470383, + 0.014442430809140205, + 0.04768978804349899, + 0.06859329342842102, + 0.010792355053126812, + 0.04127797484397888, + 0.012912257574498653, + 0.023962030187249184, + 0.009172403253614902, + 0.008146376349031925, + 0.07324467599391937, + 0.04446244239807129, + 0.005258326418697834, + 0.04763403534889221, + 0.011704201810061932, + 0.007271692156791687, + 0.010089113377034664, + 0.008948056027293205, + 0.03915872797369957, + 0.017980873584747314, + 0.003856897121295333, + 0.03371056914329529, + 0.011825472116470337, + 0.005416711792349815, + 0.0034024366177618504, + 0.0021853072103112936, + 0.004690234549343586, + 0.0059996601194143295, + 0.0018496908014640212, + 0.0032456887420266867, + 0.003278529504314065, + 0.0032744810450822115, + 0.0013248231261968613, + 0.0016640366520732641, + 0.0012702990788966417, + 0.0025556038599461317, + 0.001672342885285616, + 0.0018287886632606387, + 0.0015366403385996819, + 0.002185223624110222, + 0.012460366822779179, + 0.008200420066714287, + 0.031102119013667107, + 0.01742202416062355, + 0.0028869598172605038, + 0.0064176395535469055, + 0.00527025293558836, + 0.003611539490520954, + 0.007533584721386433, + 0.01275723334401846, + 0.03574060648679733, + 0.04709990322589874, + 0.0059120855294167995, + 0.008486230857670307, + 0.013738702982664108, + 0.006870584096759558, + 0.010264279320836067, + 0.006878747139126062, + 0.012035450898110867, + 0.027015822008252144, + 0.0033755903132259846, + 0.005894214380532503, + 0.007875660434365273, + 0.005673554260283709, + 0.001068426645360887, + 0.0009682485251687467, + 0.0037346561439335346, + 0.0034278621897101402, + 0.0008330249111168087, + 0.0008687616791576147, + 0.0015149262035265565, + 0.0012200834462419152 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 2, + 7 + ], + "selected_reward": -0.001177571015432477 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a streamlined, deep cherry red sports car?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the car moving at high speed through a deep, straight, modern concrete tunnel?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the car's LED taillights appear as bright red streaks trailing behind due to motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the rows of cool white LEDs on the tunnel walls visually stretched into continuous horizontal streaks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot captured from a low-angle perspective with the camera tracking in sync with the car?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.007623270619660616 + ], + "opposite_logprob_tokens": [ + -4.882623195648193 + ], + "logprob_reward": -0.007623270619660616, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.007623270619660616, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.16880697011947632, + 0.14492538571357727, + 0.21423731744289398, + 0.11514420062303543, + 0.03647700697183609, + 0.07127276808023453, + 0.047164320945739746, + 0.10139773786067963, + 0.0760108008980751, + 0.024563482031226158 + ], + "credit_map": [ + 0.020537735894322395, + 0.014529933221638203, + 0.02627132646739483, + 0.022604892030358315, + 0.019618341699242592, + 0.020010931417346, + 0.025774789974093437, + 0.019459012895822525, + 0.020510459318757057, + 0.012025231495499611, + 0.0161995030939579, + 0.015274368226528168, + 0.009319577366113663, + 0.04336977377533913, + 0.01196217443794012, + 0.01626429706811905, + 0.009157399646937847, + 0.009427336975932121, + 0.057781871408224106, + 0.07291293889284134, + 0.00800421554595232, + 0.03301488980650902, + 0.01392038818448782, + 0.010018275119364262, + 0.010012269951403141, + 0.010597730986773968, + 0.029003247618675232, + 0.013197153806686401, + 0.005388432648032904, + 0.027282163500785828, + 0.012625632807612419, + 0.007037564646452665, + 0.004690859001129866, + 0.0026280414313077927, + 0.004725062754005194, + 0.005108874291181564, + 0.00403016060590744, + 0.003659698413684964, + 0.006239558104425669, + 0.00539475167170167, + 0.009991860017180443, + 0.00457419641315937, + 0.015239004045724869, + 0.008710226975381374, + 0.006754687987267971, + 0.007215625140815973, + 0.007200343534350395, + 0.011586821638047695, + 0.0027661700733006, + 0.0032353047281503677, + 0.009989812038838863, + 0.014106767252087593, + 0.004074901342391968, + 0.0024227993562817574, + 0.005967700853943825, + 0.004600865766406059, + 0.005424020811915398, + 0.006946445908397436, + 0.025949880480766296, + 0.031241953372955322, + 0.00812303926795721, + 0.005792052485048771, + 0.011295896954834461, + 0.00662445230409503, + 0.008393890224397182, + 0.0060298852622509, + 0.01150519773364067, + 0.027307698503136635, + 0.0038013060111552477, + 0.005739007145166397, + 0.008298083208501339, + 0.0049357302486896515, + 0.0023797382600605488, + 0.003704504342749715, + 0.003877445589751005, + 0.0027834095526486635, + 0.001943616895005107, + 0.0032813241705298424, + 0.0026587846223264933, + 0.0039346590638160706 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 0, + 1 + ], + "selected_reward": -0.007623270619660616 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000071/group_3/index_tensor([7642, 7642])_group_3_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000071/group_3/index_tensor([7642, 7642])_group_3_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..2c6d40fa59cdc4a3384973798640b1b9a577ffb5 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000071/group_3/index_tensor([7642, 7642])_group_3_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "static, extreme close-up captures a moment of stored kinetic energy within a garage workshop. Bright, indirect afternoon light streams in from a tall, unseen window, illuminating a small section of a weathered wooden workbench, its surface pockmarked with scratches and faint oil stains. A galvanized screw with a Phillips head protrudes slightly from the grain. The end ring of a metal steel coil spring, precariously hooked on the edge of this screw, is stretched taut against the frame, its individual rings visibly elongated, gleaming in the soft light, a faint, oily sheen. The scene is still until a subtle, low-frequency vibration travels through the workbench, causing the entire surface to tremble for a fraction of a second. This tiny disturbance is enough to dislodge the spring ring; it immediately slips off the screw head with a sharp snap. Released from the tension, the coil contracts violently and rapidly, creating a blur of metallic movement, collapsing in on itself before falling onto the wooden surface with a soft click. It bounced twice, the remaining energy dissipating before finally forming into a tight, relaxed spiral, its metallic form now resting quietly in the warm, dusty afternoon sun, creating an atmosphere of tranquility after the sudden burst of action.", + "seed": 19502, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000071/group_3/index_7642_seed_19502_dp_6_rank_12_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 3, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.031440154556185006, + "vlm_logprob": -0.031440154556185006, + "vlm_yes_no_margin": 4.175, + "vlm_token_credit": -0.031440154556185006, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot framed as a static extreme close-up of a weathered wooden workbench in a garage workshop?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene illuminated by bright, indirect afternoon light coming from an unseen window?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a galvanized Phillips head screw visible with a metal steel coil spring end ring hooked on it?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the spring ring slip off the screw and contract violently into a tight spiral after a vibration?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the final state show the relaxed spring resting quietly on the workbench in warm, dusty afternoon light?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07134826481342316 + ], + "opposite_logprob_tokens": [ + -3.446348190307617 + ], + "logprob_reward": -0.07134826481342316, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.07134826481342316, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13970047235488892, + 0.04211808368563652, + 0.047259390354156494, + 0.08514365553855896, + 0.06470243632793427, + 0.19864143431186676, + 0.05853364244103432, + 0.1424315869808197, + 0.055378686636686325, + 0.16609060764312744 + ], + "credit_map": [ + 0.01887008175253868, + 0.01750282756984234, + 0.007297789212316275, + 0.01634659618139267, + 0.030780550092458725, + 0.018533699214458466, + 0.01864042691886425, + 0.011728501878678799, + 0.006110441870987415, + 0.003793376497924328, + 0.0034536411985754967, + 0.0035298303700983524, + 0.0096999891102314, + 0.00536168459802866, + 0.006657029036432505, + 0.0035120935644954443, + 0.005732835736125708, + 0.00418948894366622, + 0.003476259531453252, + 0.00915836263448, + 0.0058981385082006454, + 0.006095096934586763, + 0.0062830182723701, + 0.006426186300814152, + 0.010869987308979034, + 0.011647617444396019, + 0.0045935604721307755, + 0.0067037479020655155, + 0.012357868254184723, + 0.01602570340037346, + 0.012299095280468464, + 0.010646078735589981, + 0.010413452982902527, + 0.010283155366778374, + 0.00511673791334033, + 0.005254472140222788, + 0.007328578270971775, + 0.010522371158003807, + 0.006275086663663387, + 0.009508578106760979, + 0.05690566077828407, + 0.012469364330172539, + 0.005404703319072723, + 0.006025366950780153, + 0.011493154801428318, + 0.07636690139770508, + 0.012545083649456501, + 0.01743120327591896, + 0.017595866695046425, + 0.005669598933309317, + 0.002260708948597312, + 0.003027299651876092, + 0.012690775096416473, + 0.009697663597762585, + 0.0031450879760086536, + 0.0044466424733400345, + 0.03968482464551926, + 0.009699631482362747, + 0.0029533561319112778, + 0.0033941895235329866, + 0.013447558507323265, + 0.06086980551481247, + 0.0039711263962090015, + 0.008411101065576077, + 0.015725305303931236, + 0.004598428029567003, + 0.0022942584473639727, + 0.0022665862925350666, + 0.003212254261597991, + 0.02110142633318901, + 0.0020489429589360952, + 0.00413148058578372, + 0.056409433484077454, + 0.010207549668848515, + 0.0043938602320849895, + 0.004804545547813177, + 0.0069524324499070644, + 0.06767584383487701, + 0.005141116678714752, + 0.010505828075110912 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 9, + 7 + ], + "selected_reward": -0.07134826481342316 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot framed as a static extreme close-up of a weathered wooden workbench in a garage workshop?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene illuminated by bright, indirect afternoon light coming from an unseen window?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a galvanized Phillips head screw visible with a metal steel coil spring end ring hooked on it?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the spring ring slip off the screw and contract violently into a tight spiral after a vibration?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the final state show the relaxed spring resting quietly on the workbench in warm, dusty afternoon light?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0059962146915495396 + ], + "opposite_logprob_tokens": [ + -5.130996227264404 + ], + "logprob_reward": -0.0059962146915495396, + "yes_no_margin_reward": 5.125, + "token_credit_reward": -0.0059962146915495396, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13839589059352875, + 0.025372084230184555, + 0.01954602263867855, + 0.0497293695807457, + 0.030007224529981613, + 0.23785147070884705, + 0.1030486598610878, + 0.1970670521259308, + 0.1342119574546814, + 0.06477019935846329 + ], + "credit_map": [ + 0.022842999547719955, + 0.02325550466775894, + 0.006623735651373863, + 0.012001420371234417, + 0.02304527722299099, + 0.01677262783050537, + 0.01770828850567341, + 0.016146056354045868, + 0.004149413201957941, + 0.004268369637429714, + 0.0012826231541112065, + 0.0016697453102096915, + 0.005168354604393244, + 0.0035457704216241837, + 0.0032014886382967234, + 0.0020863215904682875, + 0.003234913107007742, + 0.002164082136005163, + 0.0012314384803175926, + 0.0017715900903567672, + 0.002483394229784608, + 0.0025386265479028225, + 0.004471317399293184, + 0.0016506633255630732, + 0.005787977017462254, + 0.007095493376255035, + 0.00237765209749341, + 0.003801364451646805, + 0.007358623668551445, + 0.007869046181440353, + 0.007979899644851685, + 0.007459317333996296, + 0.0038052883464843035, + 0.0026328435633331537, + 0.0018237895565107465, + 0.0028813001699745655, + 0.002440778072923422, + 0.004418512340635061, + 0.00549642788246274, + 0.006508287973701954, + 0.08093172311782837, + 0.012537876144051552, + 0.005944353993982077, + 0.006117229349911213, + 0.008009753189980984, + 0.10331551730632782, + 0.0076684607192873955, + 0.013326574116945267, + 0.03563182055950165, + 0.009791788645088673, + 0.003549971617758274, + 0.0037156606558710337, + 0.018705030903220177, + 0.019634250551462173, + 0.0058216373436152935, + 0.006198510993272066, + 0.052027296274900436, + 0.014961643144488335, + 0.004307238385081291, + 0.0055809011682868, + 0.01698930747807026, + 0.0840633288025856, + 0.006442629732191563, + 0.012694736942648888, + 0.040244121104478836, + 0.0176895372569561, + 0.0063448622822761536, + 0.007170412689447403, + 0.016287438571453094, + 0.028790732845664024, + 0.007146592251956463, + 0.010538289323449135, + 0.019254788756370544, + 0.0060461643151938915, + 0.001666359486989677, + 0.0027912643272429705, + 0.005911886692047119, + 0.022375013679265976, + 0.002349680056795478, + 0.004375053104013205 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 7, + 0 + ], + "selected_reward": -0.0059962146915495396 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot framed as a static extreme close-up of a weathered wooden workbench in a garage workshop?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene illuminated by bright, indirect afternoon light coming from an unseen window?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a galvanized Phillips head screw visible with a metal steel coil spring end ring hooked on it?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the spring ring slip off the screw and contract violently into a tight spiral after a vibration?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the final state show the relaxed spring resting quietly on the workbench in warm, dusty afternoon light?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.012557956390082836 + ], + "opposite_logprob_tokens": [ + -4.3875579833984375 + ], + "logprob_reward": -0.012557956390082836, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.012557956390082836, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.21136754751205444, + 0.08845850080251694, + 0.08058557659387589, + 0.2123076468706131, + 0.0624193474650383, + 0.10320333391427994, + 0.06002536043524742, + 0.10011358559131622, + 0.029648704454302788, + 0.05187038332223892 + ], + "credit_map": [ + 0.02696843445301056, + 0.022383563220500946, + 0.023135805502533913, + 0.01933145709335804, + 0.048608411103487015, + 0.020904140546917915, + 0.028738491237163544, + 0.021297238767147064, + 0.01146575529128313, + 0.012648663483560085, + 0.004406319931149483, + 0.00838746689260006, + 0.021413665264844894, + 0.01134809385985136, + 0.009804937988519669, + 0.00898358877748251, + 0.010867079719901085, + 0.012382498942315578, + 0.004338973201811314, + 0.008714604191482067, + 0.009783572517335415, + 0.011198940686881542, + 0.016071898862719536, + 0.007228001952171326, + 0.025900129228830338, + 0.0321299210190773, + 0.009173215366899967, + 0.014313661493360996, + 0.03145379573106766, + 0.03743802383542061, + 0.03155781328678131, + 0.03034108504652977, + 0.011579847894608974, + 0.00964653305709362, + 0.003945500589907169, + 0.004489316139370203, + 0.004810713697224855, + 0.009520451538264751, + 0.009323904290795326, + 0.009103079326450825, + 0.03390596807003021, + 0.005658129695802927, + 0.0024286494590342045, + 0.0023092778865247965, + 0.0045815398916602135, + 0.04321295768022537, + 0.003964853473007679, + 0.007141964975744486, + 0.016476787626743317, + 0.005325655452907085, + 0.0025109846610575914, + 0.0029204864986240864, + 0.012469073757529259, + 0.008532346226274967, + 0.00412818044424057, + 0.007661845069378614, + 0.03410295769572258, + 0.005936299916356802, + 0.002473751548677683, + 0.002236856846138835, + 0.00441368343308568, + 0.044196709990501404, + 0.0025876136496663094, + 0.004165711812674999, + 0.004174661822617054, + 0.005336338188499212, + 0.0023556710220873356, + 0.0028689897153526545, + 0.004398943856358528, + 0.004568285308778286, + 0.002154802205041051, + 0.003791011171415448, + 0.011942739598453045, + 0.005064200609922409, + 0.0024615442380309105, + 0.0032367268577218056, + 0.0032751101534813643, + 0.0181802399456501, + 0.0025456773582845926, + 0.0051641445606946945 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 0, + 5 + ], + "selected_reward": -0.012557956390082836 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot framed as a static extreme close-up of a weathered wooden workbench in a garage workshop?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene illuminated by bright, indirect afternoon light coming from an unseen window?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a galvanized Phillips head screw visible with a metal steel coil spring end ring hooked on it?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the spring ring slip off the screw and contract violently into a tight spiral after a vibration?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the final state show the relaxed spring resting quietly on the workbench in warm, dusty afternoon light?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.06204738840460777 + ], + "opposite_logprob_tokens": [ + -2.8120474815368652 + ], + "logprob_reward": -0.06204738840460777, + "yes_no_margin_reward": 2.75, + "token_credit_reward": -0.06204738840460777, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2552347183227539, + 0.043247997760772705, + 0.08731433004140854, + 0.19804739952087402, + 0.030085638165473938, + 0.035240210592746735, + 0.09363694489002228, + 0.07462479919195175, + 0.05939088761806488, + 0.12317705899477005 + ], + "credit_map": [ + 0.026153473183512688, + 0.03726227954030037, + 0.009518256410956383, + 0.022898368537425995, + 0.06121402978897095, + 0.031110186129808426, + 0.039807166904211044, + 0.02727096900343895, + 0.006390181835740805, + 0.005547752603888512, + 0.0020070078317075968, + 0.0037886849604547024, + 0.009081008844077587, + 0.005442288238555193, + 0.0069640749134123325, + 0.004027002956718206, + 0.008595499210059643, + 0.010589812882244587, + 0.006275534629821777, + 0.009958281181752682, + 0.011356771923601627, + 0.010936164297163486, + 0.018195027485489845, + 0.011407248675823212, + 0.027333712205290794, + 0.021665003150701523, + 0.012309630401432514, + 0.016954177990555763, + 0.024112895131111145, + 0.03258688747882843, + 0.032387834042310715, + 0.030697250738739967, + 0.0037009858060628176, + 0.0031704872380942106, + 0.001787472632713616, + 0.0018141836626455188, + 0.0021757367067039013, + 0.004470152780413628, + 0.004395489115267992, + 0.008571132086217403, + 0.004885651636868715, + 0.003435153281316161, + 0.0017671356908977032, + 0.0029017722699791193, + 0.0037087020464241505, + 0.004899458959698677, + 0.006406820844858885, + 0.007235516794025898, + 0.03267707675695419, + 0.010081865824759007, + 0.0033045646268874407, + 0.00361397722736001, + 0.019248230382800102, + 0.01563590206205845, + 0.004701658152043819, + 0.004373672418296337, + 0.027102498337626457, + 0.00579313887283206, + 0.0021822198759764433, + 0.001985053764656186, + 0.00692150741815567, + 0.02557523176074028, + 0.002303989138454199, + 0.002761167474091053, + 0.01703728176653385, + 0.004536574240773916, + 0.002402787795290351, + 0.0026006537955254316, + 0.003948485013097525, + 0.02416931837797165, + 0.0016103022499009967, + 0.003085487289354205, + 0.041162095963954926, + 0.008041860535740852, + 0.0028943170327693224, + 0.00307318102568388, + 0.005230640061199665, + 0.054586224257946014, + 0.002526591531932354, + 0.005662151146680117 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 9 + ], + "selected_reward": -0.06204738840460777 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot framed as a static extreme close-up of a weathered wooden workbench in a garage workshop?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene illuminated by bright, indirect afternoon light coming from an unseen window?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a galvanized Phillips head screw visible with a metal steel coil spring end ring hooked on it?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the spring ring slip off the screw and contract violently into a tight spiral after a vibration?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the final state show the relaxed spring resting quietly on the workbench in warm, dusty afternoon light?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.00525094848126173 + ], + "opposite_logprob_tokens": [ + -5.255250930786133 + ], + "logprob_reward": -0.00525094848126173, + "yes_no_margin_reward": 5.25, + "token_credit_reward": -0.00525094848126173, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3152485489845276, + 0.03754381090402603, + 0.05385006591677666, + 0.10701480507850647, + 0.04202950745820999, + 0.09673317521810532, + 0.03667181357741356, + 0.14576466381549835, + 0.08769373595714569, + 0.07744982093572617 + ], + "credit_map": [ + 0.03837500885128975, + 0.048143308609724045, + 0.012064122594892979, + 0.027933815494179726, + 0.06909915804862976, + 0.037953510880470276, + 0.048519935458898544, + 0.033159706741571426, + 0.006390985567122698, + 0.0056275371462106705, + 0.0020994183141738176, + 0.00387627468444407, + 0.005692282225936651, + 0.004586773458868265, + 0.00471015740185976, + 0.004560382105410099, + 0.005351342260837555, + 0.0067009772174060345, + 0.00409293919801712, + 0.006255666725337505, + 0.006822945550084114, + 0.006462964229285717, + 0.01105588860809803, + 0.0071073430590331554, + 0.013347235508263111, + 0.013606092892587185, + 0.00581766664981842, + 0.008887868374586105, + 0.01248437725007534, + 0.018516838550567627, + 0.018107326701283455, + 0.016247402876615524, + 0.00616878317669034, + 0.0035661482252180576, + 0.002910246839746833, + 0.004925245884805918, + 0.004053938202559948, + 0.006587065290659666, + 0.00439491355791688, + 0.009423165582120419, + 0.02828829735517502, + 0.0073148771189153194, + 0.0026331970002502203, + 0.0028934376314282417, + 0.0045139966532588005, + 0.037999626249074936, + 0.004189563915133476, + 0.008900180459022522, + 0.005174456629902124, + 0.004189715720713139, + 0.002272383077070117, + 0.0033052668441087008, + 0.004601982422173023, + 0.0056121982634067535, + 0.005082357674837112, + 0.006433454807847738, + 0.03685160353779793, + 0.012181666679680347, + 0.0030754467006772757, + 0.003683856688439846, + 0.016452955082058907, + 0.06237628310918808, + 0.004610926378518343, + 0.006531908642500639, + 0.028506413102149963, + 0.007429201155900955, + 0.002373133087530732, + 0.0030749281868338585, + 0.004017921630293131, + 0.035779885947704315, + 0.0024635123554617167, + 0.004048739094287157, + 0.026003971695899963, + 0.005272545386105776, + 0.0022763346787542105, + 0.002261274727061391, + 0.002800772665068507, + 0.03341233357787132, + 0.001965161180123687, + 0.0034574312157928944 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 7, + 3 + ], + "selected_reward": -0.00525094848126173 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "static, extreme close-up captures a moment of stored kinetic energy within a garage workshop. Bright, indirect afternoon light streams in from a tall, unseen window, illuminating a small section of a weathered wooden workbench, its surface pockmarked with scratches and faint oil stains. A galvanized screw with a Phillips head protrudes slightly from the grain. The end ring of a metal steel coil spring, precariously hooked on the edge of this screw, is stretched taut against the frame, its individual rings visibly elongated, gleaming in the soft light, a faint, oily sheen. The scene is still until a subtle, low-frequency vibration travels through the workbench, causing the entire surface to tremble for a fraction of a second. This tiny disturbance is enough to dislodge the spring ring; it immediately slips off the screw head with a sharp snap. Released from the tension, the coil contracts violently and rapidly, creating a blur of metallic movement, collapsing in on itself before falling onto the wooden surface with a soft click. It bounced twice, the remaining energy dissipating before finally forming into a tight, relaxed spiral, its metallic form now resting quietly in the warm, dusty afternoon sun, creating an atmosphere of tranquility after the sudden burst of action.", + "seed": 19503, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000071/group_3/index_7642_seed_19503_dp_6_rank_12_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 3, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.03376356437802315, + "vlm_logprob": -0.03376356437802315, + "vlm_yes_no_margin": 4.05, + "vlm_token_credit": -0.03376356437802315, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot framed as a static extreme close-up of a weathered wooden workbench in a garage workshop?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene illuminated by bright, indirect afternoon light coming from an unseen window?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a galvanized Phillips head screw visible with a metal steel coil spring end ring hooked on it?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the spring ring slip off the screw and contract violently into a tight spiral after a vibration?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the final state show the relaxed spring resting quietly on the workbench in warm, dusty afternoon light?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.11398859322071075 + ], + "opposite_logprob_tokens": [ + -2.3639886379241943 + ], + "logprob_reward": -0.11398859322071075, + "yes_no_margin_reward": 2.25, + "token_credit_reward": -0.11398859322071075, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12234485149383545, + 0.050729215145111084, + 0.036268968135118484, + 0.3690646290779114, + 0.12189088761806488, + 0.09342911094427109, + 0.12274205684661865, + 0.045943308621644974, + 0.018374862149357796, + 0.01921214908361435 + ], + "credit_map": [ + 0.017320886254310608, + 0.010957392863929272, + 0.020375685766339302, + 0.01925128884613514, + 0.00862632691860199, + 0.01775512285530567, + 0.010186717845499516, + 0.017871426418423653, + 0.008768768981099129, + 0.0032746668439358473, + 0.008078210055828094, + 0.006488174665719271, + 0.003555241972208023, + 0.00834143627434969, + 0.0052417609840631485, + 0.006980953272432089, + 0.006873093079775572, + 0.0027895704843103886, + 0.005740694236010313, + 0.0038202255964279175, + 0.0021534054540097713, + 0.006966323591768742, + 0.0032425797544419765, + 0.004683075472712517, + 0.027117332443594933, + 0.015272615477442741, + 0.12738396227359772, + 0.01712651364505291, + 0.019039221107959747, + 0.03466792404651642, + 0.025241076946258545, + 0.10321596264839172, + 0.024422569200396538, + 0.011720249429345131, + 0.016955044120550156, + 0.015431712381541729, + 0.01005083043128252, + 0.013069553300738335, + 0.014321594499051571, + 0.01591932773590088, + 0.009209411218762398, + 0.005637687630951405, + 0.008271632716059685, + 0.00808637123554945, + 0.017085179686546326, + 0.025831792503595352, + 0.009560258127748966, + 0.009746781550347805, + 0.014096288941800594, + 0.00814646016806364, + 0.012159092351794243, + 0.006292694713920355, + 0.031912438571453094, + 0.01634584367275238, + 0.006233717314898968, + 0.027555521577596664, + 0.0035710856318473816, + 0.0025790107902139425, + 0.004341727122664452, + 0.002252287231385708, + 0.013473055325448513, + 0.004481082316488028, + 0.0019311209907755256, + 0.013313938863575459, + 0.0026217475533485413, + 0.001943004666827619, + 0.0022426957730203867, + 0.002383324084803462, + 0.0020485215354710817, + 0.0026934482157230377, + 0.0019644959829747677, + 0.002477623289451003, + 0.003809475339949131, + 0.0023423684760928154, + 0.002238356973975897, + 0.0017333707073703408, + 0.0028923186473548412, + 0.0017700525932013988, + 0.0018081453163176775, + 0.0026180611457675695 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 6, + 0 + ], + "selected_reward": -0.11398859322071075 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot framed as a static extreme close-up of a weathered wooden workbench in a garage workshop?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene illuminated by bright, indirect afternoon light coming from an unseen window?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a galvanized Phillips head screw visible with a metal steel coil spring end ring hooked on it?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the spring ring slip off the screw and contract violently into a tight spiral after a vibration?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the final state show the relaxed spring resting quietly on the workbench in warm, dusty afternoon light?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.00868814904242754 + ], + "opposite_logprob_tokens": [ + -4.758687973022461 + ], + "logprob_reward": -0.00868814904242754, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.00868814904242754, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08186360448598862, + 0.033024031668901443, + 0.01598823070526123, + 0.32690003514289856, + 0.09722434729337692, + 0.06017903611063957, + 0.31681621074676514, + 0.03145807981491089, + 0.01721358671784401, + 0.019332874566316605 + ], + "credit_map": [ + 0.011431756429374218, + 0.0060728308744728565, + 0.014604507945477962, + 0.009451470337808132, + 0.009487412869930267, + 0.009727107360959053, + 0.008982228115200996, + 0.012106293812394142, + 0.003928841091692448, + 0.0019506514072418213, + 0.0038414373993873596, + 0.0038876840844750404, + 0.005079143214970827, + 0.00493319146335125, + 0.005272955633699894, + 0.004130127839744091, + 0.002070238348096609, + 0.001862370758317411, + 0.0023913055192679167, + 0.0022436436265707016, + 0.0012112855911254883, + 0.0030566477216780186, + 0.0012732179602608085, + 0.0018795204814523458, + 0.024565335363149643, + 0.008710411377251148, + 0.10884539783000946, + 0.01594696007668972, + 0.02155410125851631, + 0.03689929470419884, + 0.021585527807474136, + 0.08879300951957703, + 0.010833602398633957, + 0.007491788361221552, + 0.015404319390654564, + 0.008390732109546661, + 0.008875024504959583, + 0.020456012338399887, + 0.01027709897607565, + 0.01549577061086893, + 0.003808150766417384, + 0.004016823135316372, + 0.00506984256207943, + 0.0047853742726147175, + 0.012982094660401344, + 0.007890951819717884, + 0.005089974030852318, + 0.01653582602739334, + 0.01523368526250124, + 0.01621675305068493, + 0.02277420647442341, + 0.007039437536150217, + 0.11587796360254288, + 0.036008648574352264, + 0.00825597532093525, + 0.0954095721244812, + 0.003922541160136461, + 0.003435993567109108, + 0.0033717979677021503, + 0.002332423347979784, + 0.004645953886210918, + 0.006571483798325062, + 0.002066163346171379, + 0.005111724138259888, + 0.0017202207818627357, + 0.001970811514183879, + 0.0022211733739823103, + 0.002383404178544879, + 0.0027638287283480167, + 0.0018364816205576062, + 0.002003215253353119, + 0.0023144513834267855, + 0.0023992727510631084, + 0.0019358393037691712, + 0.002286032075062394, + 0.0019787079654634, + 0.003025103360414505, + 0.0034063365310430527, + 0.0017981510609388351, + 0.0025034311693161726 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 6, + 4 + ], + "selected_reward": -0.00868814904242754 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot framed as a static extreme close-up of a weathered wooden workbench in a garage workshop?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene illuminated by bright, indirect afternoon light coming from an unseen window?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a galvanized Phillips head screw visible with a metal steel coil spring end ring hooked on it?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the spring ring slip off the screw and contract violently into a tight spiral after a vibration?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the final state show the relaxed spring resting quietly on the workbench in warm, dusty afternoon light?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.011089595966041088 + ], + "opposite_logprob_tokens": [ + -4.51108980178833 + ], + "logprob_reward": -0.011089595966041088, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.011089595966041088, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12095752358436584, + 0.03768407180905342, + 0.017761847004294395, + 0.20900045335292816, + 0.07397403568029404, + 0.042017802596092224, + 0.16736897826194763, + 0.233637735247612, + 0.061510901898145676, + 0.03608660772442818 + ], + "credit_map": [ + 0.012565255165100098, + 0.01049934420734644, + 0.015188291668891907, + 0.01197113562375307, + 0.0182097926735878, + 0.017749221995472908, + 0.018160272389650345, + 0.016614217311143875, + 0.007448254618793726, + 0.003383397124707699, + 0.004198481794446707, + 0.004413554444909096, + 0.004289223812520504, + 0.004538286942988634, + 0.004626178648322821, + 0.004786693025380373, + 0.004187246784567833, + 0.0017998702824115753, + 0.0022313271183520555, + 0.001842327183112502, + 0.0014536913949996233, + 0.0020828016567975283, + 0.0016252355417236686, + 0.0025393457617610693, + 0.021340087056159973, + 0.007185978814959526, + 0.06829540431499481, + 0.010739483870565891, + 0.012959886342287064, + 0.01987585984170437, + 0.013193484395742416, + 0.05541026219725609, + 0.01243668794631958, + 0.00834712479263544, + 0.011478164233267307, + 0.008114163763821125, + 0.007065415848046541, + 0.007592040114104748, + 0.0077700745314359665, + 0.011170364916324615, + 0.0035844375379383564, + 0.0023670855443924665, + 0.004441678058356047, + 0.0070650274865329266, + 0.005894128233194351, + 0.006569099612534046, + 0.006196916569024324, + 0.005899430252611637, + 0.010758957825601101, + 0.010050186887383461, + 0.012042978778481483, + 0.008518936112523079, + 0.05614958703517914, + 0.020074164494872093, + 0.008345411159098148, + 0.04142875224351883, + 0.012579786591231823, + 0.012977116741240025, + 0.019827723503112793, + 0.007798445411026478, + 0.07941249758005142, + 0.021143920719623566, + 0.007287854794412851, + 0.07261040806770325, + 0.007052922621369362, + 0.004006497096270323, + 0.006139259319752455, + 0.003594273468479514, + 0.01671041175723076, + 0.004315391648560762, + 0.0037232821341603994, + 0.015968862920999527, + 0.005952394567430019, + 0.0031627858988940716, + 0.004519293550401926, + 0.003885690588504076, + 0.0045301481150090694, + 0.005061478819698095, + 0.004467073827981949, + 0.00450774235650897 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 3, + 6 + ], + "selected_reward": -0.011089595966041088 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot framed as a static extreme close-up of a weathered wooden workbench in a garage workshop?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene illuminated by bright, indirect afternoon light coming from an unseen window?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a galvanized Phillips head screw visible with a metal steel coil spring end ring hooked on it?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the spring ring slip off the screw and contract violently into a tight spiral after a vibration?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the final state show the relaxed spring resting quietly on the workbench in warm, dusty afternoon light?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.029799941927194595 + ], + "opposite_logprob_tokens": [ + -3.5297999382019043 + ], + "logprob_reward": -0.029799941927194595, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.029799941927194595, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07608530670404434, + 0.04321696236729622, + 0.04307177662849426, + 0.3343457281589508, + 0.20737868547439575, + 0.09048732370138168, + 0.09399228543043137, + 0.04623887315392494, + 0.02686542458832264, + 0.03831758350133896 + ], + "credit_map": [ + 0.016929836943745613, + 0.009272762574255466, + 0.009664258919656277, + 0.008065128698945045, + 0.00675541814416647, + 0.0076071349903941154, + 0.008312324993312359, + 0.009478439576923847, + 0.004314785357564688, + 0.003444577567279339, + 0.006011822260916233, + 0.0037029010709375143, + 0.004404212813824415, + 0.010215165093541145, + 0.004330978263169527, + 0.006792518775910139, + 0.007300274446606636, + 0.005005196668207645, + 0.006625139154493809, + 0.006435994524508715, + 0.0035207250621169806, + 0.00440041022375226, + 0.0047842669300735, + 0.0049997661262750626, + 0.036519359797239304, + 0.016845431178808212, + 0.10282823443412781, + 0.016848284751176834, + 0.02037489041686058, + 0.03074057213962078, + 0.023596059530973434, + 0.08659292012453079, + 0.03422459959983826, + 0.018246185034513474, + 0.030321190133690834, + 0.02507142350077629, + 0.018449341878294945, + 0.018395541235804558, + 0.03203756734728813, + 0.030632851645350456, + 0.006202710792422295, + 0.00545914564281702, + 0.00818279106169939, + 0.013935927301645279, + 0.01383963879197836, + 0.016719378530979156, + 0.01620253175497055, + 0.009945198893547058, + 0.007061275187879801, + 0.006181147415190935, + 0.008917181752622128, + 0.005906559061259031, + 0.027139488607645035, + 0.011336239986121655, + 0.005037762224674225, + 0.022412635385990143, + 0.0033527479972690344, + 0.002689894288778305, + 0.0050838543102145195, + 0.0029969147872179747, + 0.011890558525919914, + 0.004768703132867813, + 0.0030818565282970667, + 0.012374342419207096, + 0.002628421876579523, + 0.0020689554512500763, + 0.003482281696051359, + 0.0024905346799641848, + 0.005673546809703112, + 0.0025224308483302593, + 0.0026084159035235643, + 0.00539083918556571, + 0.002743490505963564, + 0.002337571932002902, + 0.004304183647036552, + 0.004473952576518059, + 0.006921410094946623, + 0.006128272507339716, + 0.004839298781007528, + 0.006569403689354658 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 4, + 6 + ], + "selected_reward": -0.029799941927194595 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot framed as a static extreme close-up of a weathered wooden workbench in a garage workshop?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene illuminated by bright, indirect afternoon light coming from an unseen window?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a galvanized Phillips head screw visible with a metal steel coil spring end ring hooked on it?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the spring ring slip off the screw and contract violently into a tight spiral after a vibration?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the final state show the relaxed spring resting quietly on the workbench in warm, dusty afternoon light?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.00525154173374176 + ], + "opposite_logprob_tokens": [ + -5.255251407623291 + ], + "logprob_reward": -0.00525154173374176, + "yes_no_margin_reward": 5.25, + "token_credit_reward": -0.00525154173374176, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07685531675815582, + 0.03309788554906845, + 0.02590418979525566, + 0.12393170595169067, + 0.11312465369701385, + 0.11547313630580902, + 0.3190973401069641, + 0.13208572566509247, + 0.032801490277051926, + 0.027628591284155846 + ], + "credit_map": [ + 0.014804041013121605, + 0.010129708796739578, + 0.011031830683350563, + 0.010227321647107601, + 0.005197193939238787, + 0.0083609065040946, + 0.00653242040425539, + 0.010571891441941261, + 0.0055437348783016205, + 0.0041178325191140175, + 0.004704028367996216, + 0.003961009904742241, + 0.0020124989096075296, + 0.005644544027745724, + 0.002759820083156228, + 0.004354412667453289, + 0.004241765942424536, + 0.0026379215996712446, + 0.003816439537331462, + 0.0024089831858873367, + 0.0017244461923837662, + 0.005144184455275536, + 0.0022509570699185133, + 0.003679487155750394, + 0.029824838042259216, + 0.0066094729118049145, + 0.02028690278530121, + 0.008689903654158115, + 0.009855661541223526, + 0.01317704375833273, + 0.017251815646886826, + 0.018236057832837105, + 0.011721514165401459, + 0.010821792297065258, + 0.018277611583471298, + 0.011346390470862389, + 0.012117817997932434, + 0.007897668518126011, + 0.02226964756846428, + 0.01867220550775528, + 0.010345127433538437, + 0.009192065335810184, + 0.012372906319797039, + 0.02358926646411419, + 0.012698465026915073, + 0.007929845713078976, + 0.02227327600121498, + 0.017072169110178947, + 0.019854063168168068, + 0.01844584196805954, + 0.023229168727993965, + 0.01538023166358471, + 0.10889720171689987, + 0.03520432487130165, + 0.013267876580357552, + 0.08481861650943756, + 0.010253220796585083, + 0.007351587060838938, + 0.011973997578024864, + 0.00486862612888217, + 0.04131374880671501, + 0.011928727850317955, + 0.004764624871313572, + 0.03963117673993111, + 0.0029689730145037174, + 0.002443233272060752, + 0.003241542726755142, + 0.003722638124600053, + 0.00750178936868906, + 0.002877604914829135, + 0.003373738145455718, + 0.006671968847513199, + 0.003695206716656685, + 0.002481035888195038, + 0.00388379511423409, + 0.003274294314906001, + 0.0029804534278810024, + 0.004006235394626856, + 0.003642596071586013, + 0.0036649699322879314 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 7, + 3 + ], + "selected_reward": -0.00525154173374176 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "static, extreme close-up captures a moment of stored kinetic energy within a garage workshop. Bright, indirect afternoon light streams in from a tall, unseen window, illuminating a small section of a weathered wooden workbench, its surface pockmarked with scratches and faint oil stains. A galvanized screw with a Phillips head protrudes slightly from the grain. The end ring of a metal steel coil spring, precariously hooked on the edge of this screw, is stretched taut against the frame, its individual rings visibly elongated, gleaming in the soft light, a faint, oily sheen. The scene is still until a subtle, low-frequency vibration travels through the workbench, causing the entire surface to tremble for a fraction of a second. This tiny disturbance is enough to dislodge the spring ring; it immediately slips off the screw head with a sharp snap. Released from the tension, the coil contracts violently and rapidly, creating a blur of metallic movement, collapsing in on itself before falling onto the wooden surface with a soft click. It bounced twice, the remaining energy dissipating before finally forming into a tight, relaxed spiral, its metallic form now resting quietly in the warm, dusty afternoon sun, creating an atmosphere of tranquility after the sudden burst of action.", + "seed": 19504, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000071/group_3/index_7642_seed_19504_dp_7_rank_14_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 3, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.016324066556990147, + "vlm_logprob": -0.016324066556990147, + "vlm_yes_no_margin": 4.575, + "vlm_token_credit": -0.016324066556990147, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot framed as a static extreme close-up of a weathered wooden workbench in a garage workshop?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene illuminated by bright, indirect afternoon light coming from an unseen window?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a galvanized Phillips head screw visible with a metal steel coil spring end ring hooked on it?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the spring ring slip off the screw and contract violently into a tight spiral after a vibration?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the final state show the relaxed spring resting quietly on the workbench in warm, dusty afternoon light?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03744758665561676 + ], + "opposite_logprob_tokens": [ + -3.787447690963745 + ], + "logprob_reward": -0.03744758665561676, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.03744758665561676, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09940514713525772, + 0.06926880776882172, + 0.07558122277259827, + 0.10381168872117996, + 0.07769158482551575, + 0.11892927438020706, + 0.06984364241361618, + 0.10563503950834274, + 0.09800353646278381, + 0.18183010816574097 + ], + "credit_map": [ + 0.011578998528420925, + 0.012854408472776413, + 0.012669624760746956, + 0.019142113626003265, + 0.012283115647733212, + 0.010252776555716991, + 0.008384987711906433, + 0.0122391227632761, + 0.008333674632012844, + 0.008007683791220188, + 0.012709391303360462, + 0.009619781747460365, + 0.007512240204960108, + 0.005803239531815052, + 0.008174083195626736, + 0.009108706377446651, + 0.008099360391497612, + 0.011866254732012749, + 0.011174703016877174, + 0.01102948747575283, + 0.01199630182236433, + 0.007513499818742275, + 0.0062649124301970005, + 0.00763670215383172, + 0.012666063383221626, + 0.009174134582281113, + 0.014034928753972054, + 0.013420740142464638, + 0.015309331007301807, + 0.01169395912438631, + 0.017239097505807877, + 0.01027342677116394, + 0.00946697685867548, + 0.013183102011680603, + 0.008794115856289864, + 0.01166005339473486, + 0.009001578204333782, + 0.007710303645581007, + 0.010051453486084938, + 0.007823998108506203, + 0.014102065935730934, + 0.010657770559191704, + 0.011609053239226341, + 0.015985136851668358, + 0.01935017667710781, + 0.02233939617872238, + 0.013553339056670666, + 0.011332337744534016, + 0.008675940334796906, + 0.010694875381886959, + 0.008738161064684391, + 0.008913061581552029, + 0.008016717620193958, + 0.008731944486498833, + 0.008221938274800777, + 0.007851002737879753, + 0.01315596979111433, + 0.01660110242664814, + 0.014180880971252918, + 0.015663567930459976, + 0.010790208354592323, + 0.013013840653002262, + 0.010914968326687813, + 0.011314509436488152, + 0.013851211406290531, + 0.017173616215586662, + 0.015261598862707615, + 0.019646286964416504, + 0.00877697579562664, + 0.007175680249929428, + 0.008893020451068878, + 0.007225147448480129, + 0.02534732036292553, + 0.03851296752691269, + 0.02457786537706852, + 0.031078213825821877, + 0.02037038654088974, + 0.012992475181818008, + 0.016206936910748482, + 0.012743952684104443 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 5, + 7 + ], + "selected_reward": -0.03744758665561676 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot framed as a static extreme close-up of a weathered wooden workbench in a garage workshop?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene illuminated by bright, indirect afternoon light coming from an unseen window?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a galvanized Phillips head screw visible with a metal steel coil spring end ring hooked on it?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the spring ring slip off the screw and contract violently into a tight spiral after a vibration?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the final state show the relaxed spring resting quietly on the workbench in warm, dusty afternoon light?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.018256016075611115 + ], + "opposite_logprob_tokens": [ + -4.018256187438965 + ], + "logprob_reward": -0.018256016075611115, + "yes_no_margin_reward": 4.0, + "token_credit_reward": -0.018256016075611115, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09189882129430771, + 0.08749265968799591, + 0.26614436507225037, + 0.05591806024312973, + 0.08521450310945511, + 0.14711183309555054, + 0.16922153532505035, + 0.030420156195759773, + 0.032574038952589035, + 0.03400407359004021 + ], + "credit_map": [ + 0.013585281558334827, + 0.012768331915140152, + 0.012013189494609833, + 0.0133872264996171, + 0.010759119875729084, + 0.010348143987357616, + 0.00798751786351204, + 0.011050009168684483, + 0.009303076192736626, + 0.010142671875655651, + 0.012856646440923214, + 0.012620966881513596, + 0.0066194087266922, + 0.014819709584116936, + 0.006431441754102707, + 0.01469874195754528, + 0.03414645418524742, + 0.03582041710615158, + 0.04342518746852875, + 0.03829347714781761, + 0.032860998064279556, + 0.024238083511590958, + 0.025845646858215332, + 0.03151409700512886, + 0.00800560973584652, + 0.0056384773924946785, + 0.013333276845514774, + 0.0062009128741919994, + 0.0057251001708209515, + 0.005689309909939766, + 0.006928576622158289, + 0.004396792966872454, + 0.011412317864596844, + 0.011217471212148666, + 0.014609876088798046, + 0.012354450300335884, + 0.009991137310862541, + 0.0097187589854002, + 0.008321025408804417, + 0.007589464075863361, + 0.01981138437986374, + 0.013279694132506847, + 0.007286990527063608, + 0.01050621923059225, + 0.023281823843717575, + 0.0432758666574955, + 0.01550639420747757, + 0.01416347362101078, + 0.022690871730446815, + 0.03319902718067169, + 0.026192929595708847, + 0.028290389105677605, + 0.018752776086330414, + 0.012939817272126675, + 0.013670196756720543, + 0.013485531322658062, + 0.004077682737261057, + 0.004113858565688133, + 0.004679051693528891, + 0.004361784551292658, + 0.003506737994030118, + 0.0034985539969056845, + 0.002958485623821616, + 0.003223999636247754, + 0.004052572418004274, + 0.007032020948827267, + 0.004113069269806147, + 0.0050354269333183765, + 0.003305267309769988, + 0.002637453144416213, + 0.003058325033634901, + 0.0033399031963199377, + 0.004284610506147146, + 0.005821842700242996, + 0.004883443936705589, + 0.005336294416338205, + 0.0027677484322339296, + 0.0033711802680045366, + 0.0032065713312476873, + 0.004332384094595909 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 6, + 5 + ], + "selected_reward": -0.018256016075611115 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot framed as a static extreme close-up of a weathered wooden workbench in a garage workshop?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene illuminated by bright, indirect afternoon light coming from an unseen window?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a galvanized Phillips head screw visible with a metal steel coil spring end ring hooked on it?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the spring ring slip off the screw and contract violently into a tight spiral after a vibration?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the final state show the relaxed spring resting quietly on the workbench in warm, dusty afternoon light?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004105354659259319 + ], + "opposite_logprob_tokens": [ + -5.504105567932129 + ], + "logprob_reward": -0.004105354659259319, + "yes_no_margin_reward": 5.5, + "token_credit_reward": -0.004105354659259319, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10335197299718857, + 0.07503967732191086, + 0.10305905342102051, + 0.18040797114372253, + 0.13028311729431152, + 0.08951783180236816, + 0.09656177461147308, + 0.12922972440719604, + 0.0704863891005516, + 0.022062502801418304 + ], + "credit_map": [ + 0.012142540886998177, + 0.011828992515802383, + 0.011258304119110107, + 0.015641501173377037, + 0.013215523213148117, + 0.010494967922568321, + 0.01355055719614029, + 0.015219585970044136, + 0.008758400566875935, + 0.007684359792619944, + 0.009760402143001556, + 0.009150505997240543, + 0.013374129310250282, + 0.006571505218744278, + 0.011957860551774502, + 0.007782514672726393, + 0.010626915842294693, + 0.01325224805623293, + 0.013275849632918835, + 0.015244790352880955, + 0.015851516276597977, + 0.01120157539844513, + 0.011840918101370335, + 0.011765236034989357, + 0.020109862089157104, + 0.018417512997984886, + 0.019137831404805183, + 0.027017982676625252, + 0.023005323484539986, + 0.020360056310892105, + 0.034763190895318985, + 0.017596211284399033, + 0.01822180487215519, + 0.020536139607429504, + 0.014857795089483261, + 0.01678122766315937, + 0.016998542472720146, + 0.01603618450462818, + 0.016614174470305443, + 0.010237249545753002, + 0.014637541025876999, + 0.009482361376285553, + 0.0087940888479352, + 0.009221630170941353, + 0.013968109153211117, + 0.012616240419447422, + 0.007756643462926149, + 0.013041213154792786, + 0.010145667009055614, + 0.010903690941631794, + 0.011154357343912125, + 0.00931420549750328, + 0.016150986775755882, + 0.015010448172688484, + 0.011706061661243439, + 0.012176349759101868, + 0.015442042611539364, + 0.013898732140660286, + 0.013989138416945934, + 0.01765591837465763, + 0.012187372893095016, + 0.017660653218626976, + 0.014753448776900768, + 0.023642415180802345, + 0.009063978679478168, + 0.010212598368525505, + 0.010457889176905155, + 0.010428765788674355, + 0.0074169086292386055, + 0.007974368520081043, + 0.007228711619973183, + 0.007703165523707867, + 0.0026705542113631964, + 0.002224783645942807, + 0.004306866321712732, + 0.003948850091546774, + 0.0023384445812553167, + 0.0024547732900828123, + 0.002426902297884226, + 0.0016913294093683362 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 4, + 7 + ], + "selected_reward": -0.004105354659259319 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot framed as a static extreme close-up of a weathered wooden workbench in a garage workshop?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene illuminated by bright, indirect afternoon light coming from an unseen window?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a galvanized Phillips head screw visible with a metal steel coil spring end ring hooked on it?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the spring ring slip off the screw and contract violently into a tight spiral after a vibration?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the final state show the relaxed spring resting quietly on the workbench in warm, dusty afternoon light?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.018194913864135742 + ], + "opposite_logprob_tokens": [ + -4.018195152282715 + ], + "logprob_reward": -0.018194913864135742, + "yes_no_margin_reward": 4.0, + "token_credit_reward": -0.018194913864135742, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10993783175945282, + 0.09250399470329285, + 0.18668551743030548, + 0.07792990654706955, + 0.06113315001130104, + 0.0923876091837883, + 0.12091171741485596, + 0.05133160576224327, + 0.1513894498348236, + 0.05578922852873802 + ], + "credit_map": [ + 0.01417695451527834, + 0.012916442938148975, + 0.009018328040838242, + 0.010337202809751034, + 0.018842320889234543, + 0.016160765662789345, + 0.011766367591917515, + 0.016719451174139977, + 0.00967947393655777, + 0.011281535029411316, + 0.009225446730852127, + 0.0122196851298213, + 0.014586235396564007, + 0.009884676896035671, + 0.013206259347498417, + 0.012420684099197388, + 0.0203570444136858, + 0.01974515989422798, + 0.0259017962962389, + 0.03493114560842514, + 0.024910839274525642, + 0.021201157942414284, + 0.02115124836564064, + 0.01848711259663105, + 0.00946902297437191, + 0.006532532162964344, + 0.01011292077600956, + 0.008844077587127686, + 0.011904082261025906, + 0.012338235042989254, + 0.01156389806419611, + 0.00716513441875577, + 0.007434516679495573, + 0.007882751524448395, + 0.0095650190487504, + 0.008463140577077866, + 0.007884194143116474, + 0.0066068596206605434, + 0.007626726292073727, + 0.005669944453984499, + 0.012058555148541927, + 0.011361242271959782, + 0.010840852744877338, + 0.009493053890764713, + 0.01724056527018547, + 0.00914710108190775, + 0.012315474450588226, + 0.009930760599672794, + 0.013191953301429749, + 0.010010245256125927, + 0.011880497448146343, + 0.01163513958454132, + 0.01766733080148697, + 0.021968362852931023, + 0.015275418758392334, + 0.019282763823866844, + 0.005748151335865259, + 0.004611518699675798, + 0.007656916044652462, + 0.005730340722948313, + 0.006931066978722811, + 0.006549362558871508, + 0.006405347026884556, + 0.007698902860283852, + 0.01705453172326088, + 0.015339777804911137, + 0.015323622152209282, + 0.01468080934137106, + 0.02323218807578087, + 0.022806895896792412, + 0.021673033013939857, + 0.021278588101267815, + 0.006787375081330538, + 0.006668676156550646, + 0.00653595058247447, + 0.0067851608619093895, + 0.007949539460241795, + 0.0071419719606637955, + 0.007547562010586262, + 0.00637299008667469 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 8, + 6 + ], + "selected_reward": -0.018194913864135742 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot framed as a static extreme close-up of a weathered wooden workbench in a garage workshop?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene illuminated by bright, indirect afternoon light coming from an unseen window?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a galvanized Phillips head screw visible with a metal steel coil spring end ring hooked on it?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the spring ring slip off the screw and contract violently into a tight spiral after a vibration?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the final state show the relaxed spring resting quietly on the workbench in warm, dusty afternoon light?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.003616461530327797 + ], + "opposite_logprob_tokens": [ + -5.6286163330078125 + ], + "logprob_reward": -0.003616461530327797, + "yes_no_margin_reward": 5.625, + "token_credit_reward": -0.003616461530327797, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07689360529184341, + 0.05836901068687439, + 0.07516936212778091, + 0.06044778600335121, + 0.046348609030246735, + 0.14290761947631836, + 0.1880277544260025, + 0.07192466408014297, + 0.24484103918075562, + 0.035070568323135376 + ], + "credit_map": [ + 0.008818057365715504, + 0.008930923417210579, + 0.008308234624564648, + 0.010657497681677341, + 0.00970788486301899, + 0.012228863313794136, + 0.006959384307265282, + 0.011282755993306637, + 0.005985719617456198, + 0.006299864035099745, + 0.0069878012873232365, + 0.007386848330497742, + 0.008870737627148628, + 0.008544973097741604, + 0.00681734224781394, + 0.007475726772099733, + 0.008410139009356499, + 0.009345944039523602, + 0.008680939674377441, + 0.01148749329149723, + 0.010488292202353477, + 0.008379733189940453, + 0.009949438273906708, + 0.008427387103438377, + 0.007461033295840025, + 0.006799318827688694, + 0.006164161022752523, + 0.00794370099902153, + 0.0070067401975393295, + 0.006487532984465361, + 0.012403404340147972, + 0.006181895267218351, + 0.00702762883156538, + 0.006764844059944153, + 0.008212904445827007, + 0.005024709273129702, + 0.00513867475092411, + 0.006213261745870113, + 0.0037121493369340897, + 0.004254435654729605, + 0.01690473034977913, + 0.012031071819365025, + 0.00846368819475174, + 0.008429130539298058, + 0.025947729125618935, + 0.04253959655761719, + 0.013116946443915367, + 0.015474727377295494, + 0.02431543916463852, + 0.02214600145816803, + 0.027431096881628036, + 0.01963370107114315, + 0.021435316652059555, + 0.029476182535290718, + 0.018383581191301346, + 0.025206435471773148, + 0.008948723785579205, + 0.009973680600523949, + 0.008527587167918682, + 0.009125886484980583, + 0.010574303567409515, + 0.007933370769023895, + 0.008960187435150146, + 0.007880925200879574, + 0.02723483368754387, + 0.02503283880650997, + 0.024258282035589218, + 0.023719526827335358, + 0.038168977946043015, + 0.036794278770685196, + 0.03478828817605972, + 0.034844014793634415, + 0.003969729878008366, + 0.004174098838120699, + 0.0034989153500646353, + 0.0035344413481652737, + 0.0055039431899785995, + 0.005024800542742014, + 0.004743767436593771, + 0.004620875231921673 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 6, + 5 + ], + "selected_reward": -0.003616461530327797 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "static, extreme close-up captures a moment of stored kinetic energy within a garage workshop. Bright, indirect afternoon light streams in from a tall, unseen window, illuminating a small section of a weathered wooden workbench, its surface pockmarked with scratches and faint oil stains. A galvanized screw with a Phillips head protrudes slightly from the grain. The end ring of a metal steel coil spring, precariously hooked on the edge of this screw, is stretched taut against the frame, its individual rings visibly elongated, gleaming in the soft light, a faint, oily sheen. The scene is still until a subtle, low-frequency vibration travels through the workbench, causing the entire surface to tremble for a fraction of a second. This tiny disturbance is enough to dislodge the spring ring; it immediately slips off the screw head with a sharp snap. Released from the tension, the coil contracts violently and rapidly, creating a blur of metallic movement, collapsing in on itself before falling onto the wooden surface with a soft click. It bounced twice, the remaining energy dissipating before finally forming into a tight, relaxed spiral, its metallic form now resting quietly in the warm, dusty afternoon sun, creating an atmosphere of tranquility after the sudden burst of action.", + "seed": 19505, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000071/group_3/index_7642_seed_19505_dp_7_rank_14_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 3, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.04897040314972401, + "vlm_logprob": -0.04897040314972401, + "vlm_yes_no_margin": 3.675, + "vlm_token_credit": -0.04897040314972401, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot framed as a static extreme close-up of a weathered wooden workbench in a garage workshop?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene illuminated by bright, indirect afternoon light coming from an unseen window?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a galvanized Phillips head screw visible with a metal steel coil spring end ring hooked on it?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the spring ring slip off the screw and contract violently into a tight spiral after a vibration?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the final state show the relaxed spring resting quietly on the workbench in warm, dusty afternoon light?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.17733117938041687 + ], + "opposite_logprob_tokens": [ + -1.9273312091827393 + ], + "logprob_reward": -0.17733117938041687, + "yes_no_margin_reward": 1.75, + "token_credit_reward": -0.17733117938041687, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.6300422549247742, + 0.0769043043255806, + 0.0737958699464798, + 0.057408399879932404, + 0.030127782374620438, + 0.06124620884656906, + 0.0120199304074049, + 0.014094931073486805, + 0.02976543828845024, + 0.0145949088037014 + ], + "credit_map": [ + 0.04007553309202194, + 0.04875088110566139, + 0.10101990401744843, + 0.06173111870884895, + 0.10176538676023483, + 0.1388094425201416, + 0.05321922153234482, + 0.08467075973749161, + 0.004680037498474121, + 0.009629419073462486, + 0.015000972896814346, + 0.005701679270714521, + 0.010890173725783825, + 0.017762096598744392, + 0.0067330775782465935, + 0.006506850477308035, + 0.003914147615432739, + 0.004803681280463934, + 0.018020756542682648, + 0.0030234784353524446, + 0.010138568468391895, + 0.028359897434711456, + 0.0029239675495773554, + 0.002611377974972129, + 0.005294546484947205, + 0.007912003435194492, + 0.005584807600826025, + 0.008698482066392899, + 0.006689244415611029, + 0.009650647640228271, + 0.00840437039732933, + 0.0051742978394031525, + 0.0022245158907026052, + 0.004072005394846201, + 0.002672918140888214, + 0.00433645024895668, + 0.0049149831756949425, + 0.004817630164325237, + 0.0045934515073895454, + 0.002495827851817012, + 0.005656086374074221, + 0.0071935104206204414, + 0.005437841638922691, + 0.005483223125338554, + 0.007396455388516188, + 0.014285169541835785, + 0.007070673163980246, + 0.008723249658942223, + 0.00111014349386096, + 0.0015493211103603244, + 0.0009560141479596496, + 0.0011914706556126475, + 0.0018134030979126692, + 0.0027539480943232775, + 0.0014792467700317502, + 0.001166383270174265, + 0.0010177356889471412, + 0.0015160480979830027, + 0.0012338928645476699, + 0.002019398845732212, + 0.0015806220471858978, + 0.0028440668247640133, + 0.0023858333006501198, + 0.0014973327051848173, + 0.0021407143212854862, + 0.0031470260582864285, + 0.0029266057536005974, + 0.00466554332524538, + 0.0032602930441498756, + 0.005688441917300224, + 0.005114489700645208, + 0.002822324400767684, + 0.001380330417305231, + 0.0016230650944635272, + 0.0013490431010723114, + 0.0017313315765932202, + 0.001973854610696435, + 0.003319923998787999, + 0.001636494416743517, + 0.0015808651223778725 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 2 + ], + "selected_reward": -0.17733117938041687 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot framed as a static extreme close-up of a weathered wooden workbench in a garage workshop?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene illuminated by bright, indirect afternoon light coming from an unseen window?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a galvanized Phillips head screw visible with a metal steel coil spring end ring hooked on it?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the spring ring slip off the screw and contract violently into a tight spiral after a vibration?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the final state show the relaxed spring resting quietly on the workbench in warm, dusty afternoon light?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.023357950150966644 + ], + "opposite_logprob_tokens": [ + -3.77335786819458 + ], + "logprob_reward": -0.023357950150966644, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.023357950150966644, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.7136140465736389, + 0.13352040946483612, + 0.019281897693872452, + 0.040903422981500626, + 0.021545862779021263, + 0.02041606232523918, + 0.014554193243384361, + 0.008905502036213875, + 0.01985437609255314, + 0.007404316682368517 + ], + "credit_map": [ + 0.04595686495304108, + 0.04650261998176575, + 0.18515358865261078, + 0.041228678077459335, + 0.07147973775863647, + 0.23622527718544006, + 0.02952239103615284, + 0.057544879615306854, + 0.009059291332960129, + 0.013982073403894901, + 0.031162254512310028, + 0.008480362594127655, + 0.01501635741442442, + 0.0394020713865757, + 0.006874424871057272, + 0.009543578140437603, + 0.0027681065257638693, + 0.0027464095037430525, + 0.0019990517757833004, + 0.0016749185742810369, + 0.0029018153436481953, + 0.003007129766047001, + 0.0020910294260829687, + 0.0020934375934302807, + 0.002973753958940506, + 0.007278610952198505, + 0.0035543593112379313, + 0.0036090263165533543, + 0.005737964995205402, + 0.009768174029886723, + 0.0043990639969706535, + 0.0035824680235236883, + 0.0014251712709665298, + 0.0023471140302717686, + 0.0023129640612751245, + 0.0019288895418867469, + 0.003803237807005644, + 0.0063761756755411625, + 0.0017961738631129265, + 0.0015561366453766823, + 0.0014154696837067604, + 0.0027121142484247684, + 0.0018310168525204062, + 0.002250388264656067, + 0.002532968996092677, + 0.0035932622849941254, + 0.0031069666147232056, + 0.002973873633891344, + 0.001302835764363408, + 0.0015861323336139321, + 0.0014177735429257154, + 0.0014119548723101616, + 0.002111179521307349, + 0.003167215036228299, + 0.0018568409141153097, + 0.0017002610256895423, + 0.0007908863481134176, + 0.0011657299473881721, + 0.0007438927423208952, + 0.0012877278495579958, + 0.000966749619692564, + 0.001764045562595129, + 0.0010812662076205015, + 0.0011052038753405213, + 0.0015260846121236682, + 0.0020549106411635876, + 0.0019387095235288143, + 0.0031354017555713654, + 0.0023331232368946075, + 0.00406672665849328, + 0.003326225792989135, + 0.0014731932897120714, + 0.0005870567983947694, + 0.0008516086963936687, + 0.0006732179317623377, + 0.0011482000118121505, + 0.000940834463108331, + 0.0012556517031043768, + 0.001149187097325921, + 0.0007985600386746228 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 3 + ], + "selected_reward": -0.023357950150966644 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot framed as a static extreme close-up of a weathered wooden workbench in a garage workshop?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene illuminated by bright, indirect afternoon light coming from an unseen window?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a galvanized Phillips head screw visible with a metal steel coil spring end ring hooked on it?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the spring ring slip off the screw and contract violently into a tight spiral after a vibration?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the final state show the relaxed spring resting quietly on the workbench in warm, dusty afternoon light?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.011089478619396687 + ], + "opposite_logprob_tokens": [ + -4.511089324951172 + ], + "logprob_reward": -0.011089478619396687, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.011089478619396687, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.7351118922233582, + 0.04694610834121704, + 0.015057616867125034, + 0.049224790185689926, + 0.03943249210715294, + 0.04113759845495224, + 0.015266839414834976, + 0.01145930215716362, + 0.024486569687724113, + 0.021876810118556023 + ], + "credit_map": [ + 0.04562244564294815, + 0.05224272981286049, + 0.1633559614419937, + 0.048893917351961136, + 0.09574514627456665, + 0.21340562403202057, + 0.04716234281659126, + 0.0686836913228035, + 0.003973820246756077, + 0.009678431786596775, + 0.005917719099670649, + 0.0032897323835641146, + 0.007414806634187698, + 0.010585341602563858, + 0.003542395308613777, + 0.002543858252465725, + 0.0016554940957576036, + 0.002595241880044341, + 0.0012552288826555014, + 0.0011563739972189069, + 0.003229396417737007, + 0.00220847618766129, + 0.0016138589708134532, + 0.0013435462024062872, + 0.0036143010947853327, + 0.006885800044983625, + 0.005213391501456499, + 0.007061888463795185, + 0.007400865666568279, + 0.00872786995023489, + 0.006299840752035379, + 0.004020834341645241, + 0.003011515596881509, + 0.005891136825084686, + 0.0031446171924471855, + 0.004811480641365051, + 0.007976425811648369, + 0.007713934872299433, + 0.004116564057767391, + 0.0027668187394738197, + 0.0031587257981300354, + 0.004757878370583057, + 0.0033403015695512295, + 0.007155721541494131, + 0.005463638808578253, + 0.0060545154847204685, + 0.008136161603033543, + 0.0030706555116921663, + 0.0010623530251905322, + 0.0016267152968794107, + 0.0015758192166686058, + 0.002882832894101739, + 0.0015126323560252786, + 0.002456668298691511, + 0.0027708890847861767, + 0.0013789294753223658, + 0.0008985106251202524, + 0.001118863350711763, + 0.0009968533413484693, + 0.0018564342753961682, + 0.0013634450733661652, + 0.0020421906374394894, + 0.002143141580745578, + 0.001039862516336143, + 0.002399036893621087, + 0.0027211506385356188, + 0.0022541119251400232, + 0.0028258783277124166, + 0.002794020576402545, + 0.00533125689253211, + 0.0032461679074913263, + 0.002914946526288986, + 0.0016200385289266706, + 0.002165701240301132, + 0.002281984779983759, + 0.004245198797434568, + 0.002455675508826971, + 0.0037573387380689383, + 0.003413405502215028, + 0.0019374676048755646 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 1 + ], + "selected_reward": -0.011089478619396687 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot framed as a static extreme close-up of a weathered wooden workbench in a garage workshop?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene illuminated by bright, indirect afternoon light coming from an unseen window?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a galvanized Phillips head screw visible with a metal steel coil spring end ring hooked on it?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the spring ring slip off the screw and contract violently into a tight spiral after a vibration?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the final state show the relaxed spring resting quietly on the workbench in warm, dusty afternoon light?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02329447865486145 + ], + "opposite_logprob_tokens": [ + -3.773294448852539 + ], + "logprob_reward": -0.02329447865486145, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.02329447865486145, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.6464594602584839, + 0.04886665567755699, + 0.056164201349020004, + 0.055556487292051315, + 0.10990093648433685, + 0.043853238224983215, + 0.013286433182656765, + 0.00783213134855032, + 0.00815736036747694, + 0.009923039004206657 + ], + "credit_map": [ + 0.047527190297842026, + 0.05743928998708725, + 0.14988453686237335, + 0.04241486266255379, + 0.06996597349643707, + 0.19417352974414825, + 0.0354633703827858, + 0.049590736627578735, + 0.0032498568762093782, + 0.008335098624229431, + 0.008411969989538193, + 0.002796536311507225, + 0.006706725340336561, + 0.01268662791699171, + 0.0038736748974770308, + 0.002806171076372266, + 0.003807713743299246, + 0.00672603864222765, + 0.005537510849535465, + 0.005485037341713905, + 0.010093273594975471, + 0.011657197959721088, + 0.008114555850625038, + 0.004742879420518875, + 0.0038613451179116964, + 0.007268773857504129, + 0.005660857539623976, + 0.008065594360232353, + 0.007219682447612286, + 0.0093484316021204, + 0.008392095565795898, + 0.00573970889672637, + 0.0069567118771374226, + 0.014050507918000221, + 0.00970454327762127, + 0.013206737115979195, + 0.021143609657883644, + 0.027417637407779694, + 0.009968787431716919, + 0.007452402263879776, + 0.0032179427798837423, + 0.005840755999088287, + 0.00372053449973464, + 0.006849855184555054, + 0.005878165364265442, + 0.006914720870554447, + 0.007943202741444111, + 0.003488066140562296, + 0.0011513368226587772, + 0.0013956644106656313, + 0.0017342212377116084, + 0.0019264043075963855, + 0.0012062537716701627, + 0.0022484068758785725, + 0.0022830315865576267, + 0.001341115334071219, + 0.0007700325222685933, + 0.000853799981996417, + 0.0007493252633139491, + 0.000919689831789583, + 0.0011489252792671323, + 0.0014406576519832015, + 0.000916278746444732, + 0.001033423119224608, + 0.0005977771361358464, + 0.0010756191331893206, + 0.0006789330509491265, + 0.000918451463803649, + 0.0011883864644914865, + 0.0016222010599449277, + 0.001041398267261684, + 0.001034594257362187, + 0.0006194122834131122, + 0.0009647558908909559, + 0.001082138274796307, + 0.002025175141170621, + 0.0010580028174445033, + 0.0013929948909208179, + 0.0020155354868620634, + 0.0007650245679542422 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 2 + ], + "selected_reward": -0.02329447865486145 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot framed as a static extreme close-up of a weathered wooden workbench in a garage workshop?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene illuminated by bright, indirect afternoon light coming from an unseen window?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a galvanized Phillips head screw visible with a metal steel coil spring end ring hooked on it?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the spring ring slip off the screw and contract violently into a tight spiral after a vibration?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the final state show the relaxed spring resting quietly on the workbench in warm, dusty afternoon light?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.009778928942978382 + ], + "opposite_logprob_tokens": [ + -4.63477897644043 + ], + "logprob_reward": -0.009778928942978382, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.009778928942978382, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.6938121914863586, + 0.08268669247627258, + 0.0356144942343235, + 0.053722258657217026, + 0.049269791692495346, + 0.030178895220160484, + 0.01119194645434618, + 0.008862543851137161, + 0.021460525691509247, + 0.013200624845921993 + ], + "credit_map": [ + 0.04188649356365204, + 0.05001860857009888, + 0.13482382893562317, + 0.0654391422867775, + 0.08220507204532623, + 0.16909201443195343, + 0.045712895691394806, + 0.10463422536849976, + 0.007114127743989229, + 0.012911257334053516, + 0.017200719565153122, + 0.0052397889085114, + 0.008524364791810513, + 0.021141747012734413, + 0.0042678844183683395, + 0.006286812014877796, + 0.0028639889787882566, + 0.004300006199628115, + 0.005609571933746338, + 0.0029342216439545155, + 0.005245779175311327, + 0.009788496419787407, + 0.002383865648880601, + 0.0024885693565011024, + 0.003941358998417854, + 0.008053701370954514, + 0.004849414341151714, + 0.007246180437505245, + 0.008344657719135284, + 0.008854618296027184, + 0.00770588219165802, + 0.004726449027657509, + 0.003444298403337598, + 0.00617521395906806, + 0.005640464834868908, + 0.004061790648847818, + 0.008847794495522976, + 0.01454965304583311, + 0.0033413772471249104, + 0.0032092020846903324, + 0.0025834653060883284, + 0.0035706365015357733, + 0.002808827906847, + 0.004796776920557022, + 0.0030687276739627123, + 0.003725921967998147, + 0.005987934302538633, + 0.003636604407802224, + 0.0009049183572642505, + 0.0014879731461405754, + 0.0010289672063663602, + 0.0015571446856483817, + 0.0016228765016421676, + 0.0016751113580539823, + 0.001588482758961618, + 0.00132647342979908, + 0.0008958203834481537, + 0.0011921967379748821, + 0.0006739574600942433, + 0.001312573323957622, + 0.0009714834159240127, + 0.0012829197803512216, + 0.001367504126392305, + 0.0011660895543172956, + 0.0019939059857279062, + 0.0025238420348614454, + 0.0018391212215647101, + 0.0030671334825456142, + 0.0023767510429024696, + 0.00407223729416728, + 0.003196073928847909, + 0.002391463378444314, + 0.001014367793686688, + 0.0015526991337537766, + 0.0009662997326813638, + 0.002453582128509879, + 0.0014176052063703537, + 0.0017959708347916603, + 0.0026101984549313784, + 0.001389902550727129 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 3 + ], + "selected_reward": -0.009778928942978382 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000071/group_4/index_tensor([2717, 2717])_group_4_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000071/group_4/index_tensor([2717, 2717])_group_4_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..19d93f8a77bb52c687a8ba929aaca4e5c5183736 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000071/group_4/index_tensor([2717, 2717])_group_4_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "static medium shot, shot from a low angle, focuses on a towering, dark, solid wood bookshelf. Its edges, carved with classical patterns, take on a solemn quality under a warm light streaming in from outside the frame. Each shelf sags slightly under its weight, cluttered with a variety of heavy hardcover books and aged paperbacks. The entire scene is permeated with a quiet yet weighty atmosphere. Suddenly, accompanied by a dull groan of wood under pressure, a tiny crack quietly appears on the vertical support at the center of the bookshelf. Then, like a flash of black lightning, it spreads upward and downward at astonishing speed, and the wood fibers visibly burst apart with a sharp, tearing sound. As the crack widened, the entire bookshelf structure began to bend catastrophically inward. First, several books near the crack tilted precariously forward, then fell one after another, triggering an avalanche of books. Books of all sizes and colors cascaded down like a waterfall, tumbling and colliding in the air, and finally smashing to the ground with a heavy, muffled thud. The thick dust raised rolled in the beam of light, instantly replacing the original tranquility with a dynamic scene of chaos and destruction.", + "seed": 13883, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000071/group_4/index_2717_seed_13883_dp_8_rank_16_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 4, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.08161822482943534, + "vlm_logprob": -0.08161822482943534, + "vlm_yes_no_margin": 2.6500000238418577, + "vlm_token_credit": -0.08161822482943534, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static medium shot taken from a low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a towering, dark, solid wood bookshelf with classical carved patterns?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the bookshelf appear cluttered with heavy hardcover books and aged paperbacks on sagging shelves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a crack appear on the central vertical support and spread rapidly causing the bookshelf to bend inward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do books cascade down like a waterfall and smash to the ground creating dust in a beam of light?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.17860999703407288 + ], + "opposite_logprob_tokens": [ + -2.05361008644104 + ], + "logprob_reward": -0.17860999703407288, + "yes_no_margin_reward": 1.8750001192092896, + "token_credit_reward": -0.17860999703407288, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11497663706541061, + 0.059859681874513626, + 0.06134907528758049, + 0.06699389964342117, + 0.027489691972732544, + 0.08541364222764969, + 0.14666515588760376, + 0.1265658140182495, + 0.17790772020816803, + 0.13277862966060638 + ], + "credit_map": [ + 0.016924060881137848, + 0.012617060914635658, + 0.014204167760908604, + 0.010982401669025421, + 0.016613544896245003, + 0.014841185882687569, + 0.012833325192332268, + 0.015960894525051117, + 0.006404391489923, + 0.006744634825736284, + 0.0069051277823746204, + 0.004633545875549316, + 0.009544840082526207, + 0.009004226885735989, + 0.004633428994566202, + 0.011989493854343891, + 0.006567591801285744, + 0.0052750688046216965, + 0.006213614717125893, + 0.005132281221449375, + 0.007992665283381939, + 0.00862254947423935, + 0.009940839372575283, + 0.01160446833819151, + 0.0075866179540753365, + 0.006616733502596617, + 0.01753312163054943, + 0.004678952973335981, + 0.004186022561043501, + 0.01609087735414505, + 0.0035741517785936594, + 0.0067274258472025394, + 0.003375230822712183, + 0.004125535953789949, + 0.003058487316593528, + 0.0035125641152262688, + 0.002998405834659934, + 0.0029371343553066254, + 0.0043959952890872955, + 0.003086341544985771, + 0.009334100410342216, + 0.009144305251538754, + 0.005938921123743057, + 0.010003876872360706, + 0.011313069611787796, + 0.02152898535132408, + 0.0067938161082565784, + 0.011356567032635212, + 0.014609447680413723, + 0.011978480033576488, + 0.03587494045495987, + 0.011471434496343136, + 0.013781663030385971, + 0.020028285682201385, + 0.011787213385105133, + 0.027133706957101822, + 0.00980064645409584, + 0.012348704971373081, + 0.012631634250283241, + 0.03898746520280838, + 0.010935373604297638, + 0.01989811658859253, + 0.008696755394339561, + 0.01326712965965271, + 0.016685429960489273, + 0.026965593919157982, + 0.01728053018450737, + 0.018613357096910477, + 0.022137606516480446, + 0.02287798933684826, + 0.020589441061019897, + 0.032757777720689774, + 0.017852993682026863, + 0.012188483960926533, + 0.014163988642394543, + 0.01519195269793272, + 0.010819796472787857, + 0.02325320802628994, + 0.006991296075284481, + 0.03231692314147949 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 6, + 9 + ], + "selected_reward": -0.17860999703407288 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static medium shot taken from a low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a towering, dark, solid wood bookshelf with classical carved patterns?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the bookshelf appear cluttered with heavy hardcover books and aged paperbacks on sagging shelves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a crack appear on the central vertical support and spread rapidly causing the bookshelf to bend inward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do books cascade down like a waterfall and smash to the ground creating dust in a beam of light?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.062094107270240784 + ], + "opposite_logprob_tokens": [ + -2.812094211578369 + ], + "logprob_reward": -0.062094107270240784, + "yes_no_margin_reward": 2.75, + "token_credit_reward": -0.062094107270240784, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09511857479810715, + 0.06463617086410522, + 0.041142214089632034, + 0.07420460134744644, + 0.12924858927726746, + 0.03604700788855553, + 0.23946306109428406, + 0.08578328043222427, + 0.0876036062836647, + 0.14675284922122955 + ], + "credit_map": [ + 0.01297887321561575, + 0.010991699993610382, + 0.017568301409482956, + 0.006762728560715914, + 0.011888512410223484, + 0.013272416777908802, + 0.009713172912597656, + 0.011942869983613491, + 0.00855809822678566, + 0.006723389029502869, + 0.009778497740626335, + 0.007880468852818012, + 0.008828489109873772, + 0.007558910641819239, + 0.008576112799346447, + 0.006732207257300615, + 0.005557925906032324, + 0.005440377630293369, + 0.004472058732062578, + 0.0038917651399970055, + 0.005074707791209221, + 0.0049059465527534485, + 0.007324995007365942, + 0.004474439658224583, + 0.0077419159933924675, + 0.0069813779555261135, + 0.01527758315205574, + 0.005518336780369282, + 0.006710395682603121, + 0.017943082377314568, + 0.005909680388867855, + 0.008122232742607594, + 0.012996410951018333, + 0.02015192247927189, + 0.012236577458679676, + 0.012156050652265549, + 0.017140911892056465, + 0.01985880732536316, + 0.014397352933883667, + 0.02031056024134159, + 0.004544122610241175, + 0.006905728951096535, + 0.0033125211484730244, + 0.00461606727913022, + 0.003375745378434658, + 0.005472172982990742, + 0.00429478520527482, + 0.0035258610732853413, + 0.027358204126358032, + 0.01199214905500412, + 0.09699094295501709, + 0.008370848372578621, + 0.009337862022221088, + 0.017389100044965744, + 0.008868559263646603, + 0.05915542319417, + 0.0066355373710393906, + 0.011115660890936852, + 0.006324934773147106, + 0.012324006296694279, + 0.018065471202135086, + 0.012675334699451923, + 0.007657800801098347, + 0.010984535329043865, + 0.00872938521206379, + 0.0109814228489995, + 0.010291013866662979, + 0.009243037551641464, + 0.012734703719615936, + 0.009406509809195995, + 0.013593575917184353, + 0.012623962014913559, + 0.01507694460451603, + 0.013883363455533981, + 0.017730582505464554, + 0.018118444830179214, + 0.016989143565297127, + 0.022540472447872162, + 0.01338100153952837, + 0.029032891616225243 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 9, + 4 + ], + "selected_reward": -0.062094107270240784 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static medium shot taken from a low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a towering, dark, solid wood bookshelf with classical carved patterns?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the bookshelf appear cluttered with heavy hardcover books and aged paperbacks on sagging shelves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a crack appear on the central vertical support and spread rapidly causing the bookshelf to bend inward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do books cascade down like a waterfall and smash to the ground creating dust in a beam of light?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04868543520569801 + ], + "opposite_logprob_tokens": [ + -3.0486855506896973 + ], + "logprob_reward": -0.04868543520569801, + "yes_no_margin_reward": 3.0, + "token_credit_reward": -0.04868543520569801, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.157037615776062, + 0.02681356482207775, + 0.0794302225112915, + 0.10417450964450836, + 0.25604063272476196, + 0.051255304366350174, + 0.05448688566684723, + 0.0405561663210392, + 0.062304794788360596, + 0.16790027916431427 + ], + "credit_map": [ + 0.02201690338551998, + 0.01982005126774311, + 0.01875132881104946, + 0.01413362380117178, + 0.022424399852752686, + 0.020044635981321335, + 0.01860715262591839, + 0.021239522844552994, + 0.0033548364881426096, + 0.00247172056697309, + 0.0029658875428140163, + 0.0032315487042069435, + 0.0038080483209341764, + 0.0031609339639544487, + 0.0035101268440485, + 0.004310463555157185, + 0.00965859368443489, + 0.008306351490318775, + 0.009740638546645641, + 0.008358321152627468, + 0.00990790594369173, + 0.008757815696299076, + 0.016220759600400925, + 0.008479829877614975, + 0.010535575449466705, + 0.00993954949080944, + 0.018528195098042488, + 0.008636513724923134, + 0.010803122073411942, + 0.021676409989595413, + 0.010233938694000244, + 0.013821197673678398, + 0.030779246240854263, + 0.029916157945990562, + 0.030696596950292587, + 0.020528877153992653, + 0.03515750542283058, + 0.046019457280635834, + 0.029335113242268562, + 0.033607665449380875, + 0.007108689751476049, + 0.010616983287036419, + 0.004863781854510307, + 0.007004025857895613, + 0.005128686316311359, + 0.0062139155343174934, + 0.004980079829692841, + 0.005339141469448805, + 0.004845387302339077, + 0.004593420308083296, + 0.017452968284487724, + 0.003197558457031846, + 0.004127909429371357, + 0.005569495726376772, + 0.0033654014114290476, + 0.01133474800735712, + 0.0028542482759803534, + 0.003721395740285516, + 0.004101558588445187, + 0.008239027112722397, + 0.007420697715133429, + 0.005412454716861248, + 0.004160340875387192, + 0.004646446090191603, + 0.004365012515336275, + 0.005941807292401791, + 0.006090288050472736, + 0.0063323997892439365, + 0.011535558849573135, + 0.005792001727968454, + 0.008898693136870861, + 0.013349035754799843, + 0.01127896923571825, + 0.011422309093177319, + 0.009268013760447502, + 0.008387895300984383, + 0.01133042573928833, + 0.04808363690972328, + 0.0068978299386799335, + 0.06123121455311775 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 9, + 0 + ], + "selected_reward": -0.04868543520569801 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static medium shot taken from a low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a towering, dark, solid wood bookshelf with classical carved patterns?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the bookshelf appear cluttered with heavy hardcover books and aged paperbacks on sagging shelves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a crack appear on the central vertical support and spread rapidly causing the bookshelf to bend inward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do books cascade down like a waterfall and smash to the ground creating dust in a beam of light?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07002682238817215 + ], + "opposite_logprob_tokens": [ + -2.6950268745422363 + ], + "logprob_reward": -0.07002682238817215, + "yes_no_margin_reward": 2.625, + "token_credit_reward": -0.07002682238817215, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08373335003852844, + 0.027162302285432816, + 0.016879595816135406, + 0.04977680742740631, + 0.029005900025367737, + 0.016864538192749023, + 0.08137767016887665, + 0.051350705325603485, + 0.18310776352882385, + 0.46074140071868896 + ], + "credit_map": [ + 0.011422594077885151, + 0.01035910751670599, + 0.01009342446923256, + 0.00754526536911726, + 0.0112036457285285, + 0.010807262733578682, + 0.010430975817143917, + 0.011871076188981533, + 0.0030536334961652756, + 0.0032336502335965633, + 0.0031320913694798946, + 0.0029711814131587744, + 0.004130162764340639, + 0.003964718896895647, + 0.0030951255466789007, + 0.0035817353054881096, + 0.0024986437056213617, + 0.002124666003510356, + 0.0022707844618707895, + 0.002295759506523609, + 0.001679070177488029, + 0.0017834301106631756, + 0.0019043001811951399, + 0.002322942018508911, + 0.005120733752846718, + 0.005095434375107288, + 0.011914880946278572, + 0.0033340221270918846, + 0.004966920707374811, + 0.010212146677076817, + 0.003458179999142885, + 0.005674492567777634, + 0.0025302800349891186, + 0.0037256614305078983, + 0.0037953329738229513, + 0.003602795535698533, + 0.002858708379790187, + 0.005367732606828213, + 0.0028948578983545303, + 0.0042305318638682365, + 0.002415247028693557, + 0.0020055752247571945, + 0.0021760445088148117, + 0.001922635710798204, + 0.0017204644391313195, + 0.00218245224095881, + 0.0015291833551600575, + 0.0029129351023584604, + 0.008909991011023521, + 0.004706319887191057, + 0.03167319297790527, + 0.0034138031769543886, + 0.003843769198283553, + 0.006021188106387854, + 0.00449787313118577, + 0.01831153593957424, + 0.004199358634650707, + 0.005643800832331181, + 0.0053220512345433235, + 0.00835409015417099, + 0.006998845376074314, + 0.009277467615902424, + 0.003932114224880934, + 0.007622977253049612, + 0.014864268712699413, + 0.023333510383963585, + 0.017519833520054817, + 0.017116272822022438, + 0.027891164645552635, + 0.019664857536554337, + 0.02452664077281952, + 0.038191214203834534, + 0.03194784000515938, + 0.03634278103709221, + 0.02967645786702633, + 0.02518714964389801, + 0.061122339218854904, + 0.12210137397050858, + 0.022632399573922157, + 0.1317310482263565 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 8, + 0 + ], + "selected_reward": -0.07002682238817215 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static medium shot taken from a low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a towering, dark, solid wood bookshelf with classical carved patterns?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the bookshelf appear cluttered with heavy hardcover books and aged paperbacks on sagging shelves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a crack appear on the central vertical support and spread rapidly causing the bookshelf to bend inward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do books cascade down like a waterfall and smash to the ground creating dust in a beam of light?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04867476224899292 + ], + "opposite_logprob_tokens": [ + -3.0486748218536377 + ], + "logprob_reward": -0.04867476224899292, + "yes_no_margin_reward": 3.0, + "token_credit_reward": -0.04867476224899292, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.05973232537508011, + 0.02428075298666954, + 0.01938505284488201, + 0.09267383068799973, + 0.028061769902706146, + 0.023775950074195862, + 0.09650418907403946, + 0.04271838814020157, + 0.14063222706317902, + 0.4722355902194977 + ], + "credit_map": [ + 0.007879198528826237, + 0.007480151019990444, + 0.007845193147659302, + 0.0047224340960383415, + 0.008014623075723648, + 0.007813218981027603, + 0.006657095160335302, + 0.009320409037172794, + 0.0027844090946018696, + 0.002409401349723339, + 0.0029994091019034386, + 0.002597263315692544, + 0.003482233267277479, + 0.0035843253135681152, + 0.0028101210482418537, + 0.0036135907284915447, + 0.002629118040204048, + 0.0022551335860043764, + 0.0023999586701393127, + 0.002876236801967025, + 0.0018910822691395879, + 0.001683825277723372, + 0.0025338565465062857, + 0.003115839557722211, + 0.00939503125846386, + 0.010079916566610336, + 0.021291371434926987, + 0.006108195520937443, + 0.008813909254968166, + 0.020010115578770638, + 0.006572452839463949, + 0.010402821004390717, + 0.003132845973595977, + 0.004527858458459377, + 0.002407453954219818, + 0.00277989124879241, + 0.003116829786449671, + 0.003994973376393318, + 0.003788683796301484, + 0.004313231445848942, + 0.0034314109943807125, + 0.002661935519427061, + 0.0028808368369936943, + 0.0025528708938509226, + 0.0023592356592416763, + 0.002796582644805312, + 0.0026543710846453905, + 0.004438704811036587, + 0.010819059796631336, + 0.004965292755514383, + 0.03879920393228531, + 0.003571370616555214, + 0.004323901608586311, + 0.006625316571444273, + 0.004333082120865583, + 0.02306695096194744, + 0.003835960989817977, + 0.00376781914383173, + 0.0037722315173596144, + 0.008319242857396603, + 0.005632342770695686, + 0.006060738116502762, + 0.0051407599821686745, + 0.00618929136544466, + 0.011163347400724888, + 0.019395975396037102, + 0.011558890342712402, + 0.013199781067669392, + 0.02458726055920124, + 0.014878756366670132, + 0.02006002701818943, + 0.02578817307949066, + 0.03710676357150078, + 0.0379776656627655, + 0.03219582885503769, + 0.03213602676987648, + 0.03588675707578659, + 0.12207189947366714, + 0.019266512244939804, + 0.155594140291214 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 8, + 6 + ], + "selected_reward": -0.04867476224899292 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "static medium shot, shot from a low angle, focuses on a towering, dark, solid wood bookshelf. Its edges, carved with classical patterns, take on a solemn quality under a warm light streaming in from outside the frame. Each shelf sags slightly under its weight, cluttered with a variety of heavy hardcover books and aged paperbacks. The entire scene is permeated with a quiet yet weighty atmosphere. Suddenly, accompanied by a dull groan of wood under pressure, a tiny crack quietly appears on the vertical support at the center of the bookshelf. Then, like a flash of black lightning, it spreads upward and downward at astonishing speed, and the wood fibers visibly burst apart with a sharp, tearing sound. As the crack widened, the entire bookshelf structure began to bend catastrophically inward. First, several books near the crack tilted precariously forward, then fell one after another, triggering an avalanche of books. Books of all sizes and colors cascaded down like a waterfall, tumbling and colliding in the air, and finally smashing to the ground with a heavy, muffled thud. The thick dust raised rolled in the beam of light, instantly replacing the original tranquility with a dynamic scene of chaos and destruction.", + "seed": 13884, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000071/group_4/index_2717_seed_13884_dp_8_rank_16_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 4, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.05241585988551378, + "vlm_logprob": -0.05241585988551378, + "vlm_yes_no_margin": 3.625, + "vlm_token_credit": -0.05241585988551378, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static medium shot taken from a low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a towering, dark, solid wood bookshelf with classical carved patterns?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the bookshelf appear cluttered with heavy hardcover books and aged paperbacks on sagging shelves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a crack appear on the central vertical support and spread rapidly causing the bookshelf to bend inward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do books cascade down like a waterfall and smash to the ground creating dust in a beam of light?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.16745954751968384 + ], + "opposite_logprob_tokens": [ + -2.167459487915039 + ], + "logprob_reward": -0.16745954751968384, + "yes_no_margin_reward": 2.0, + "token_credit_reward": -0.16745954751968384, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1532578319311142, + 0.047380998730659485, + 0.13575229048728943, + 0.04676574468612671, + 0.06950246542692184, + 0.07795362919569016, + 0.16116127371788025, + 0.07755166292190552, + 0.16633014380931854, + 0.06434398144483566 + ], + "credit_map": [ + 0.02779083512723446, + 0.04099053889513016, + 0.031319454312324524, + 0.023282969370484352, + 0.0056053102016448975, + 0.010673262178897858, + 0.006472462322562933, + 0.007123008370399475, + 0.009267440997064114, + 0.009796427562832832, + 0.008425552397966385, + 0.01033813413232565, + 0.002268919488415122, + 0.0021869575139135122, + 0.002640624064952135, + 0.0024569458328187466, + 0.042617738246917725, + 0.014496801421046257, + 0.02971266210079193, + 0.00789774302393198, + 0.0027186961378902197, + 0.011733855120837688, + 0.0024462146684527397, + 0.024128563702106476, + 0.00894160009920597, + 0.007340404205024242, + 0.011174511164426804, + 0.007968472316861153, + 0.0024926913902163506, + 0.0030956256669014692, + 0.002734891837462783, + 0.0030175463762134314, + 0.010966742411255836, + 0.012288594618439674, + 0.010341457091271877, + 0.01620974950492382, + 0.004568008705973625, + 0.004264133982360363, + 0.003480487270280719, + 0.007383288349956274, + 0.008650746196508408, + 0.01524245273321867, + 0.012719213962554932, + 0.014947380870580673, + 0.007459193002432585, + 0.006558852270245552, + 0.007015554700046778, + 0.005360236391425133, + 0.020485449582338333, + 0.03195906803011894, + 0.03298598527908325, + 0.044309090822935104, + 0.006800312548875809, + 0.00750366784632206, + 0.007620250806212425, + 0.009497449733316898, + 0.009315297938883305, + 0.0070818145759403706, + 0.012934337370097637, + 0.015874631702899933, + 0.0071885595098137856, + 0.006737802177667618, + 0.009290458634495735, + 0.009128760546445847, + 0.01735997386276722, + 0.017722349613904953, + 0.028187952935695648, + 0.02353406324982643, + 0.013630040921270847, + 0.022673722356557846, + 0.015386337414383888, + 0.027835696935653687, + 0.0066240206360816956, + 0.008323068730533123, + 0.005474488250911236, + 0.0074409605003893375, + 0.007882609963417053, + 0.009029101580381393, + 0.007792377844452858, + 0.011777348816394806 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 6, + 0 + ], + "selected_reward": -0.16745954751968384 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static medium shot taken from a low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a towering, dark, solid wood bookshelf with classical carved patterns?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the bookshelf appear cluttered with heavy hardcover books and aged paperbacks on sagging shelves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a crack appear on the central vertical support and spread rapidly causing the bookshelf to bend inward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do books cascade down like a waterfall and smash to the ground creating dust in a beam of light?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03374408930540085 + ], + "opposite_logprob_tokens": [ + -3.4087440967559814 + ], + "logprob_reward": -0.03374408930540085, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.03374408930540085, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2158149927854538, + 0.09848844259977341, + 0.05061749741435051, + 0.14908893406391144, + 0.04241158813238144, + 0.08791469782590866, + 0.14796501398086548, + 0.08638792484998703, + 0.04448419436812401, + 0.07682671397924423 + ], + "credit_map": [ + 0.042482733726501465, + 0.056638021022081375, + 0.0355849526822567, + 0.03872915357351303, + 0.007909760810434818, + 0.015359081327915192, + 0.010151654481887817, + 0.008959626778960228, + 0.01804332807660103, + 0.024321462959051132, + 0.017967713996767998, + 0.018335137516260147, + 0.004080801270902157, + 0.00565507635474205, + 0.004556300584226847, + 0.0055286213755607605, + 0.008973873220384121, + 0.009010186418890953, + 0.008176238276064396, + 0.008451013825833797, + 0.0020972390193492174, + 0.004004261456429958, + 0.001910308375954628, + 0.007994377985596657, + 0.03329462930560112, + 0.018285078927874565, + 0.04201783984899521, + 0.030695220455527306, + 0.0048768892884254456, + 0.007651050109416246, + 0.0046780407428741455, + 0.007590196095407009, + 0.004764258395880461, + 0.008649828843772411, + 0.006981139536947012, + 0.009959972463548183, + 0.0036998544819653034, + 0.002953484421595931, + 0.0027714986354112625, + 0.002631549723446369, + 0.010707548819482327, + 0.021503722295165062, + 0.01580384001135826, + 0.019764883443713188, + 0.004532685503363609, + 0.004427213221788406, + 0.004762016702443361, + 0.006412786897271872, + 0.01835123635828495, + 0.02700134925544262, + 0.026258599013090134, + 0.04708855226635933, + 0.007469684351235628, + 0.005363927688449621, + 0.009029165841639042, + 0.00740249315276742, + 0.009324910119175911, + 0.00811023823916912, + 0.020041244104504585, + 0.014003527350723743, + 0.005601736716926098, + 0.004368666559457779, + 0.0071531981229782104, + 0.01778440736234188, + 0.005756421014666557, + 0.0056407093070447445, + 0.00494499783962965, + 0.005315049085766077, + 0.0037427120842039585, + 0.007020591292530298, + 0.0061884354799985886, + 0.005875277332961559, + 0.008622816763818264, + 0.008217096328735352, + 0.009277775883674622, + 0.007424355484545231, + 0.013731850311160088, + 0.010177863761782646, + 0.011408223770558834, + 0.007966733537614346 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 6 + ], + "selected_reward": -0.03374408930540085 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static medium shot taken from a low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a towering, dark, solid wood bookshelf with classical carved patterns?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the bookshelf appear cluttered with heavy hardcover books and aged paperbacks on sagging shelves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a crack appear on the central vertical support and spread rapidly causing the bookshelf to bend inward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do books cascade down like a waterfall and smash to the ground creating dust in a beam of light?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.012557603418827057 + ], + "opposite_logprob_tokens": [ + -4.387557506561279 + ], + "logprob_reward": -0.012557603418827057, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.012557603418827057, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3005436360836029, + 0.2216973453760147, + 0.10932322591543198, + 0.047474682331085205, + 0.03075186349451542, + 0.03702949360013008, + 0.11311506479978561, + 0.03635942190885544, + 0.05134760960936546, + 0.05235766991972923 + ], + "credit_map": [ + 0.059276457875967026, + 0.07646041363477707, + 0.04907160624861717, + 0.0605766661465168, + 0.013585886918008327, + 0.010669652372598648, + 0.018848996609449387, + 0.012053954415023327, + 0.036776527762413025, + 0.058162473142147064, + 0.039553623646497726, + 0.0509372241795063, + 0.008243265561759472, + 0.00795926433056593, + 0.009661861695349216, + 0.010403103195130825, + 0.024123236536979675, + 0.012914232909679413, + 0.01587454415857792, + 0.014061158522963524, + 0.003371034748852253, + 0.011829216964542866, + 0.003158233128488064, + 0.023991569876670837, + 0.010358872823417187, + 0.006990345194935799, + 0.010655239224433899, + 0.008451930247247219, + 0.0027407307643443346, + 0.0028052073903381824, + 0.0025451912079006433, + 0.002927166875451803, + 0.006462129298597574, + 0.004858029540628195, + 0.006016352213919163, + 0.0049332622438669205, + 0.0014340662164613605, + 0.0021432775538414717, + 0.0014028438599780202, + 0.0035019025672227144, + 0.004267093725502491, + 0.007258425932377577, + 0.006188767496496439, + 0.007388245314359665, + 0.0029235389083623886, + 0.0029497218783944845, + 0.003071988234296441, + 0.0029817118775099516, + 0.01557169109582901, + 0.01697302982211113, + 0.021401768550276756, + 0.03212587162852287, + 0.005687697324901819, + 0.0064935204572975636, + 0.006668783724308014, + 0.008192697539925575, + 0.004109739791601896, + 0.0022344382014125586, + 0.007791064213961363, + 0.007248271722346544, + 0.0017632675589993596, + 0.0032970220781862736, + 0.002651248825713992, + 0.007264366373419762, + 0.006897360552102327, + 0.005711654666811228, + 0.008979881182312965, + 0.006094911135733128, + 0.0032292415853589773, + 0.005967981182038784, + 0.005543211940675974, + 0.008923369459807873, + 0.006078116130083799, + 0.0076414188370108604, + 0.005928603932261467, + 0.007560274563729763, + 0.005032259970903397, + 0.006657015066593885, + 0.008185154758393764, + 0.005274826195091009 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 6 + ], + "selected_reward": -0.012557603418827057 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static medium shot taken from a low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a towering, dark, solid wood bookshelf with classical carved patterns?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the bookshelf appear cluttered with heavy hardcover books and aged paperbacks on sagging shelves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a crack appear on the central vertical support and spread rapidly causing the bookshelf to bend inward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do books cascade down like a waterfall and smash to the ground creating dust in a beam of light?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0430624857544899 + ], + "opposite_logprob_tokens": [ + -3.168062448501587 + ], + "logprob_reward": -0.0430624857544899, + "yes_no_margin_reward": 3.125, + "token_credit_reward": -0.0430624857544899, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1638415902853012, + 0.05529411509633064, + 0.03124340809881687, + 0.06261242181062698, + 0.0980929434299469, + 0.0635320246219635, + 0.04676638916134834, + 0.13951590657234192, + 0.10407856106758118, + 0.2350226640701294 + ], + "credit_map": [ + 0.030574409291148186, + 0.042978037148714066, + 0.0341632179915905, + 0.030745062977075577, + 0.006502941716462374, + 0.005463081412017345, + 0.007284468039870262, + 0.006130378693342209, + 0.011438916437327862, + 0.010540071874856949, + 0.008737650699913502, + 0.012660891748964787, + 0.0028294732328504324, + 0.0032412447035312653, + 0.0024837455712258816, + 0.003362120594829321, + 0.006526524666696787, + 0.006347522605210543, + 0.006700967438519001, + 0.005236911587417126, + 0.0014526480808854103, + 0.0015614378498867154, + 0.0013000445906072855, + 0.002117351396009326, + 0.009747608564794064, + 0.009618562646210194, + 0.012454641051590443, + 0.00963005144149065, + 0.005497393198311329, + 0.005961847025901079, + 0.003571566892787814, + 0.006130749825388193, + 0.01528950035572052, + 0.01937577873468399, + 0.016852328553795815, + 0.02350393310189247, + 0.004573740065097809, + 0.005846097134053707, + 0.004412179347127676, + 0.008239390328526497, + 0.00830757338553667, + 0.013361141085624695, + 0.01173360738903284, + 0.01234697736799717, + 0.004952174611389637, + 0.003990863915532827, + 0.004038478247821331, + 0.004801204893738031, + 0.006917198654264212, + 0.006224927492439747, + 0.008683719672262669, + 0.014985083602368832, + 0.001657884451560676, + 0.002658728277310729, + 0.002613151678815484, + 0.003025697311386466, + 0.019759614020586014, + 0.022119292989373207, + 0.02082516811788082, + 0.017589375376701355, + 0.017184119671583176, + 0.0095209376886487, + 0.014336003921926022, + 0.018181396648287773, + 0.011009622365236282, + 0.013627257198095322, + 0.015632301568984985, + 0.015046504326164722, + 0.007817570120096207, + 0.01360257063060999, + 0.0105216596275568, + 0.016821077093482018, + 0.030158042907714844, + 0.024654297158122063, + 0.02537904493510723, + 0.02593202143907547, + 0.019934838637709618, + 0.029627244919538498, + 0.039894767105579376, + 0.03944242745637894 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 0, + 7 + ], + "selected_reward": -0.0430624857544899 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static medium shot taken from a low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a towering, dark, solid wood bookshelf with classical carved patterns?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the bookshelf appear cluttered with heavy hardcover books and aged paperbacks on sagging shelves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a crack appear on the central vertical support and spread rapidly causing the bookshelf to bend inward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do books cascade down like a waterfall and smash to the ground creating dust in a beam of light?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005255573429167271 + ], + "opposite_logprob_tokens": [ + -5.255255699157715 + ], + "logprob_reward": -0.005255573429167271, + "yes_no_margin_reward": 5.25, + "token_credit_reward": -0.005255573429167271, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06383076310157776, + 0.08545146882534027, + 0.09837515652179718, + 0.047809626907110214, + 0.04106905311346054, + 0.024846632033586502, + 0.05104149132966995, + 0.14776289463043213, + 0.11852554976940155, + 0.3212873041629791 + ], + "credit_map": [ + 0.012830247171223164, + 0.013855671510100365, + 0.011332106776535511, + 0.010678743943572044, + 0.0030843925196677446, + 0.0035027763806283474, + 0.003985742572695017, + 0.0045610880479216576, + 0.014869882725179195, + 0.021529005840420723, + 0.016160238534212112, + 0.018377237021923065, + 0.0034709852188825607, + 0.0033875086810439825, + 0.0038923078682273626, + 0.0037643059622496367, + 0.0278945229947567, + 0.009657446295022964, + 0.016835888847708702, + 0.00892745703458786, + 0.0022340284194797277, + 0.01014264952391386, + 0.0023552270140498877, + 0.020327948033809662, + 0.011248578317463398, + 0.00692031579092145, + 0.012090467847883701, + 0.00805707648396492, + 0.002970800269395113, + 0.0017874403856694698, + 0.002472053049132228, + 0.002262899884954095, + 0.007666486315429211, + 0.007998611778020859, + 0.00722274836152792, + 0.00858779065310955, + 0.0027754048351198435, + 0.0021580958273261786, + 0.002083081752061844, + 0.0025768394116312265, + 0.003090713871642947, + 0.006201455835253, + 0.004553784616291523, + 0.004293523728847504, + 0.0015950258821249008, + 0.0013073058798909187, + 0.0018788866000249982, + 0.0019259386463090777, + 0.007368681952357292, + 0.006823141127824783, + 0.010056481696665287, + 0.01609276980161667, + 0.002033813390880823, + 0.0027727512642741203, + 0.00256005534902215, + 0.0033338034991174936, + 0.021164581179618835, + 0.020467856898903847, + 0.022116539999842644, + 0.028139961883425713, + 0.013762068934738636, + 0.009099030867218971, + 0.017336232587695122, + 0.015676649287343025, + 0.011457527056336403, + 0.01708333007991314, + 0.016126349568367004, + 0.017714330926537514, + 0.00948764942586422, + 0.015418934635818005, + 0.010954288765788078, + 0.020283153280615807, + 0.03967278078198433, + 0.04550740122795105, + 0.03678697720170021, + 0.03677530586719513, + 0.028974968940019608, + 0.04301345348358154, + 0.0505111925303936, + 0.04004526883363724 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 7, + 8 + ], + "selected_reward": -0.005255573429167271 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "static medium shot, shot from a low angle, focuses on a towering, dark, solid wood bookshelf. Its edges, carved with classical patterns, take on a solemn quality under a warm light streaming in from outside the frame. Each shelf sags slightly under its weight, cluttered with a variety of heavy hardcover books and aged paperbacks. The entire scene is permeated with a quiet yet weighty atmosphere. Suddenly, accompanied by a dull groan of wood under pressure, a tiny crack quietly appears on the vertical support at the center of the bookshelf. Then, like a flash of black lightning, it spreads upward and downward at astonishing speed, and the wood fibers visibly burst apart with a sharp, tearing sound. As the crack widened, the entire bookshelf structure began to bend catastrophically inward. First, several books near the crack tilted precariously forward, then fell one after another, triggering an avalanche of books. Books of all sizes and colors cascaded down like a waterfall, tumbling and colliding in the air, and finally smashing to the ground with a heavy, muffled thud. The thick dust raised rolled in the beam of light, instantly replacing the original tranquility with a dynamic scene of chaos and destruction.", + "seed": 13885, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000071/group_4/index_2717_seed_13885_dp_9_rank_18_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 4, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.10815260782837868, + "vlm_logprob": -0.10815260782837868, + "vlm_yes_no_margin": 2.274999976158142, + "vlm_token_credit": -0.10815260782837868, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static medium shot taken from a low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a towering, dark, solid wood bookshelf with classical carved patterns?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the bookshelf appear cluttered with heavy hardcover books and aged paperbacks on sagging shelves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a crack appear on the central vertical support and spread rapidly causing the bookshelf to bend inward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do books cascade down like a waterfall and smash to the ground creating dust in a beam of light?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.12567441165447235 + ], + "opposite_logprob_tokens": [ + -2.5006744861602783 + ], + "logprob_reward": -0.12567441165447235, + "yes_no_margin_reward": 2.375, + "token_credit_reward": -0.12567441165447235, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0497976578772068, + 0.05497996509075165, + 0.052480779588222504, + 0.050397008657455444, + 0.029895957559347153, + 0.03867677226662636, + 0.10968495160341263, + 0.42514485120773315, + 0.07679599523544312, + 0.11214608699083328 + ], + "credit_map": [ + 0.005896026734262705, + 0.00450498890131712, + 0.005507600959390402, + 0.005443755071610212, + 0.010361389257013798, + 0.007374327629804611, + 0.00609422056004405, + 0.004615348763763905, + 0.004991491790860891, + 0.005188605282455683, + 0.007503310218453407, + 0.005270610097795725, + 0.013117015361785889, + 0.008008217439055443, + 0.006783070042729378, + 0.004117646254599094, + 0.005798297468572855, + 0.005042340140789747, + 0.006446201354265213, + 0.006293982267379761, + 0.010895706713199615, + 0.006490516941994429, + 0.00624831672757864, + 0.005265411920845509, + 0.004248125944286585, + 0.004913405980914831, + 0.004743767902255058, + 0.0043320865370333195, + 0.011202836409211159, + 0.0067228227853775024, + 0.005556722171604633, + 0.00867723673582077, + 0.0028472919948399067, + 0.003676347667351365, + 0.00231313519179821, + 0.004333856049925089, + 0.004523155279457569, + 0.0042954860255122185, + 0.0049593267031013966, + 0.0029473581817001104, + 0.0027323050890117884, + 0.003917156718671322, + 0.0036913363728672266, + 0.00457513565197587, + 0.005304187070578337, + 0.005790663883090019, + 0.007769648917019367, + 0.004896339029073715, + 0.007897758856415749, + 0.011317572556436062, + 0.015212022699415684, + 0.014896396547555923, + 0.007019787561148405, + 0.025347476825118065, + 0.008556999266147614, + 0.019436931237578392, + 0.013860523700714111, + 0.015446081757545471, + 0.029045376926660538, + 0.012470846064388752, + 0.007112510502338409, + 0.15032486617565155, + 0.01792147010564804, + 0.17896315455436707, + 0.009622554294764996, + 0.008204896934330463, + 0.01180536299943924, + 0.011512506753206253, + 0.0065980516374111176, + 0.011667984537780285, + 0.007711357902735472, + 0.009673280641436577, + 0.010655529797077179, + 0.015628337860107422, + 0.01584762893617153, + 0.021311437711119652, + 0.008093305863440037, + 0.011166070587933064, + 0.01742810755968094, + 0.01201566867530346 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 9, + 6 + ], + "selected_reward": -0.12567441165447235 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static medium shot taken from a low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a towering, dark, solid wood bookshelf with classical carved patterns?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the bookshelf appear cluttered with heavy hardcover books and aged paperbacks on sagging shelves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a crack appear on the central vertical support and spread rapidly causing the bookshelf to bend inward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do books cascade down like a waterfall and smash to the ground creating dust in a beam of light?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.11297770589590073 + ], + "opposite_logprob_tokens": [ + -2.2379777431488037 + ], + "logprob_reward": -0.11297770589590073, + "yes_no_margin_reward": 2.125, + "token_credit_reward": -0.11297770589590073, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07589313387870789, + 0.034622952342033386, + 0.020052438601851463, + 0.017780205234885216, + 0.017776453867554665, + 0.02479906938970089, + 0.06959283351898193, + 0.6279338598251343, + 0.03380119800567627, + 0.07774780690670013 + ], + "credit_map": [ + 0.010686634108424187, + 0.006624917965382338, + 0.007492418400943279, + 0.013681999407708645, + 0.011321176774799824, + 0.012319136410951614, + 0.005618923343718052, + 0.008147930726408958, + 0.0032999373506754637, + 0.0028553383890539408, + 0.00298402551561594, + 0.0046968767419457436, + 0.007195653393864632, + 0.005907493643462658, + 0.003910720348358154, + 0.003772907657548785, + 0.0022195507772266865, + 0.001657773507758975, + 0.001958054257556796, + 0.002256502630189061, + 0.003393390914425254, + 0.0032252033706754446, + 0.0025453995913267136, + 0.002796564484015107, + 0.0013736473629251122, + 0.0013418744783848524, + 0.001265264581888914, + 0.0023680394515395164, + 0.0033225915394723415, + 0.002608848735690117, + 0.002407389460131526, + 0.0030925488099455833, + 0.0021514124237000942, + 0.0018269220599904656, + 0.0017421608790755272, + 0.0019032724667340517, + 0.003083377843722701, + 0.002217006403952837, + 0.002367710927501321, + 0.002484592143446207, + 0.001457122270949185, + 0.002112398389726877, + 0.002822357229888439, + 0.0019426598446443677, + 0.00556740490719676, + 0.003365124110132456, + 0.004005215596407652, + 0.0035267877392470837, + 0.005751731339842081, + 0.008300014771521091, + 0.010077015496790409, + 0.0069862729869782925, + 0.00581364007666707, + 0.012547831051051617, + 0.007613733876496553, + 0.012502597644925117, + 0.021919237449765205, + 0.02366657741367817, + 0.04364955052733421, + 0.020077630877494812, + 0.016335448250174522, + 0.2164321392774582, + 0.02985405921936035, + 0.25599926710128784, + 0.003493491094559431, + 0.005770474672317505, + 0.00408052047714591, + 0.00408706022426486, + 0.0034305083099752665, + 0.005624587647616863, + 0.0029435339383780956, + 0.004371022805571556, + 0.007689670193940401, + 0.012363489717245102, + 0.011387201026082039, + 0.015173343941569328, + 0.006063017528504133, + 0.006868495140224695, + 0.011912927962839603, + 0.0062896572053432465 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 9, + 0 + ], + "selected_reward": -0.11297770589590073 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static medium shot taken from a low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a towering, dark, solid wood bookshelf with classical carved patterns?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the bookshelf appear cluttered with heavy hardcover books and aged paperbacks on sagging shelves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a crack appear on the central vertical support and spread rapidly causing the bookshelf to bend inward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do books cascade down like a waterfall and smash to the ground creating dust in a beam of light?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.062086377292871475 + ], + "opposite_logprob_tokens": [ + -2.812086343765259 + ], + "logprob_reward": -0.062086377292871475, + "yes_no_margin_reward": 2.75, + "token_credit_reward": -0.062086377292871475, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11548777669668198, + 0.06403191387653351, + 0.04003123566508293, + 0.02648802660405636, + 0.04223615303635597, + 0.07301414757966995, + 0.0722300112247467, + 0.2997245788574219, + 0.045996248722076416, + 0.22075989842414856 + ], + "credit_map": [ + 0.01197019126266241, + 0.012597789987921715, + 0.012707282789051533, + 0.016587670892477036, + 0.015432732179760933, + 0.020438244566321373, + 0.010209991596639156, + 0.015543877147138119, + 0.006485914345830679, + 0.006439710035920143, + 0.006796436849981546, + 0.010106478817760944, + 0.00831468403339386, + 0.011778339743614197, + 0.005325783044099808, + 0.00878456886857748, + 0.003127725562080741, + 0.0042000520043075085, + 0.004104267340153456, + 0.006230311933904886, + 0.006774113979190588, + 0.005776002537459135, + 0.004340576007962227, + 0.00547819072380662, + 0.002171522006392479, + 0.0025110100395977497, + 0.0022260474506765604, + 0.003460032632574439, + 0.004208182916045189, + 0.0038123552221804857, + 0.0040025170892477036, + 0.004096361342817545, + 0.0050500924699008465, + 0.0041644275188446045, + 0.0034887997899204493, + 0.0032515430357307196, + 0.006368231028318405, + 0.00716654397547245, + 0.0072306678630411625, + 0.005515851080417633, + 0.004924143198877573, + 0.007392180152237415, + 0.00864220131188631, + 0.005372321233153343, + 0.015912843868136406, + 0.009675351902842522, + 0.012060883454978466, + 0.009034226648509502, + 0.005892766173928976, + 0.008647003211081028, + 0.009416444227099419, + 0.010975201614201069, + 0.005706241820007563, + 0.01361843291670084, + 0.0064503783360123634, + 0.011523543857038021, + 0.010606505908071995, + 0.010698060505092144, + 0.02119576558470726, + 0.01200931891798973, + 0.007726688403636217, + 0.09982265532016754, + 0.016992244869470596, + 0.12067335098981857, + 0.004637317266315222, + 0.006026098504662514, + 0.006012673489749432, + 0.007774959318339825, + 0.004235584754496813, + 0.0076311868615448475, + 0.0033377516083419323, + 0.006340677849948406, + 0.01948946714401245, + 0.03285878896713257, + 0.03550048917531967, + 0.04800495505332947, + 0.01998746395111084, + 0.020169051364064217, + 0.028437668457627296, + 0.01631203666329384 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 9, + 0 + ], + "selected_reward": -0.062086377292871475 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static medium shot taken from a low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a towering, dark, solid wood bookshelf with classical carved patterns?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the bookshelf appear cluttered with heavy hardcover books and aged paperbacks on sagging shelves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a crack appear on the central vertical support and spread rapidly causing the bookshelf to bend inward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do books cascade down like a waterfall and smash to the ground creating dust in a beam of light?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.1270512044429779 + ], + "opposite_logprob_tokens": [ + -2.1270511150360107 + ], + "logprob_reward": -0.1270512044429779, + "yes_no_margin_reward": 1.9999998807907104, + "token_credit_reward": -0.1270512044429779, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06990690529346466, + 0.05933649092912674, + 0.04382964223623276, + 0.023085150867700577, + 0.033665817230939865, + 0.16677716374397278, + 0.029302900657057762, + 0.1325685828924179, + 0.15743163228034973, + 0.28409573435783386 + ], + "credit_map": [ + 0.006145463790744543, + 0.007963628508150578, + 0.0063088154420256615, + 0.008742266334593296, + 0.016116229817271233, + 0.01014255452901125, + 0.008505361154675484, + 0.005982582923024893, + 0.004933562129735947, + 0.005397696979343891, + 0.004413301590830088, + 0.006856183055788279, + 0.011942149139940739, + 0.008688547648489475, + 0.010222128592431545, + 0.006882925517857075, + 0.0028956872411072254, + 0.004492643754929304, + 0.0032109261956065893, + 0.005071305204182863, + 0.010036228224635124, + 0.005577778443694115, + 0.005548556335270405, + 0.0069965156726539135, + 0.001854349859058857, + 0.002460696268826723, + 0.0019029333489015698, + 0.0025331194046884775, + 0.0033294714521616697, + 0.0030512241646647453, + 0.003216359531506896, + 0.0047369967214763165, + 0.0018758930964395404, + 0.0027135482523590326, + 0.0022494636941701174, + 0.00481282128021121, + 0.006114035379141569, + 0.004372974392026663, + 0.005585948936641216, + 0.005941133946180344, + 0.009280217811465263, + 0.017028532922267914, + 0.01470116525888443, + 0.014327192679047585, + 0.043290045112371445, + 0.022578997537493706, + 0.03018450178205967, + 0.015386522747576237, + 0.003032916923984885, + 0.003542098216712475, + 0.0044455030001699924, + 0.0032239132560789585, + 0.0028367717750370502, + 0.004998120944947004, + 0.003380961949005723, + 0.003842615755274892, + 0.006007134914398193, + 0.005709046497941017, + 0.009409221820533276, + 0.007642754353582859, + 0.003697470761835575, + 0.042352400720119476, + 0.0072176516987383366, + 0.050532903522253036, + 0.014227368868887424, + 0.02033337950706482, + 0.020446257665753365, + 0.018026573583483696, + 0.02302808128297329, + 0.024201104417443275, + 0.018148133531212807, + 0.019020726904273033, + 0.025611838325858116, + 0.04946484416723251, + 0.04150301218032837, + 0.05585039034485817, + 0.02803574874997139, + 0.022762557491660118, + 0.041145291179418564, + 0.01972205378115177 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 5, + 8 + ], + "selected_reward": -0.1270512044429779 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static medium shot taken from a low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a towering, dark, solid wood bookshelf with classical carved patterns?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the bookshelf appear cluttered with heavy hardcover books and aged paperbacks on sagging shelves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a crack appear on the central vertical support and spread rapidly causing the bookshelf to bend inward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do books cascade down like a waterfall and smash to the ground creating dust in a beam of light?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.11297333985567093 + ], + "opposite_logprob_tokens": [ + -2.23797345161438 + ], + "logprob_reward": -0.11297333985567093, + "yes_no_margin_reward": 2.125, + "token_credit_reward": -0.11297333985567093, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.05385639891028404, + 0.019643442705273628, + 0.013836339116096497, + 0.020185042172670364, + 0.017508581280708313, + 0.030183229595422745, + 0.0527828224003315, + 0.2151302993297577, + 0.05709657445549965, + 0.5197772979736328 + ], + "credit_map": [ + 0.003591783344745636, + 0.006449164357036352, + 0.005209291819483042, + 0.00762658566236496, + 0.009401491843163967, + 0.007899840362370014, + 0.007464757189154625, + 0.006213483400642872, + 0.0016298923874273896, + 0.002006903290748596, + 0.00196643965318799, + 0.0024928508792072535, + 0.003776926314458251, + 0.0030930936336517334, + 0.0028162607923150063, + 0.00186107715126127, + 0.0013393813278526068, + 0.0015145899960771203, + 0.0015412858920171857, + 0.0017508298624306917, + 0.002336833393201232, + 0.001982662593945861, + 0.001974712125957012, + 0.0013960448559373617, + 0.002021896420046687, + 0.0018684855895116925, + 0.0017091810004785657, + 0.0020872869063168764, + 0.005222087260335684, + 0.0027900966815650463, + 0.0027321549132466316, + 0.00175385270267725, + 0.0015523005276918411, + 0.0011773896403610706, + 0.0010172983165830374, + 0.002012120094150305, + 0.0035324583295732737, + 0.0022561762016266584, + 0.002914035227149725, + 0.0030468024779111147, + 0.001558314193971455, + 0.0036637093871831894, + 0.0018319584196433425, + 0.00516963517293334, + 0.006499912589788437, + 0.004433064721524715, + 0.004627226386219263, + 0.002399411518126726, + 0.004432381130754948, + 0.005479684565216303, + 0.006747397128492594, + 0.008344593457877636, + 0.003402221016585827, + 0.011685218662023544, + 0.004333969205617905, + 0.008357357233762741, + 0.00697490107268095, + 0.00879229512065649, + 0.014679974876344204, + 0.007660498842597008, + 0.004069191869348288, + 0.073433056473732, + 0.012022598646581173, + 0.087497778236866, + 0.0058332220651209354, + 0.009900148026645184, + 0.00693586515262723, + 0.011238513514399529, + 0.004807638004422188, + 0.007185277994722128, + 0.003949746023863554, + 0.007246166933327913, + 0.048488739877939224, + 0.06522170454263687, + 0.08382997661828995, + 0.12433905154466629, + 0.039503298699855804, + 0.05137523263692856, + 0.06840730458498001, + 0.038611989468336105 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 7, + 8 + ], + "selected_reward": -0.11297333985567093 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "static medium shot, shot from a low angle, focuses on a towering, dark, solid wood bookshelf. Its edges, carved with classical patterns, take on a solemn quality under a warm light streaming in from outside the frame. Each shelf sags slightly under its weight, cluttered with a variety of heavy hardcover books and aged paperbacks. The entire scene is permeated with a quiet yet weighty atmosphere. Suddenly, accompanied by a dull groan of wood under pressure, a tiny crack quietly appears on the vertical support at the center of the bookshelf. Then, like a flash of black lightning, it spreads upward and downward at astonishing speed, and the wood fibers visibly burst apart with a sharp, tearing sound. As the crack widened, the entire bookshelf structure began to bend catastrophically inward. First, several books near the crack tilted precariously forward, then fell one after another, triggering an avalanche of books. Books of all sizes and colors cascaded down like a waterfall, tumbling and colliding in the air, and finally smashing to the ground with a heavy, muffled thud. The thick dust raised rolled in the beam of light, instantly replacing the original tranquility with a dynamic scene of chaos and destruction.", + "seed": 13886, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000071/group_4/index_2717_seed_13886_dp_9_rank_18_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 4, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.11685550827533006, + "vlm_logprob": -0.11685550827533006, + "vlm_yes_no_margin": 2.649999976158142, + "vlm_token_credit": -0.11685550827533006, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static medium shot taken from a low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a towering, dark, solid wood bookshelf with classical carved patterns?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the bookshelf appear cluttered with heavy hardcover books and aged paperbacks on sagging shelves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a crack appear on the central vertical support and spread rapidly causing the bookshelf to bend inward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do books cascade down like a waterfall and smash to the ground creating dust in a beam of light?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.17828737199306488 + ], + "opposite_logprob_tokens": [ + -2.0532872676849365 + ], + "logprob_reward": -0.17828737199306488, + "yes_no_margin_reward": 1.8749998807907104, + "token_credit_reward": -0.17828737199306488, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10692428797483444, + 0.0363522469997406, + 0.034754373133182526, + 0.09976416826248169, + 0.04258599132299423, + 0.09922751784324646, + 0.17173226177692413, + 0.1217915415763855, + 0.13786865770816803, + 0.1489989459514618 + ], + "credit_map": [ + 0.01909482106566429, + 0.012657244689762592, + 0.013159404508769512, + 0.01365464087575674, + 0.015138078480958939, + 0.012842574156820774, + 0.011111045256257057, + 0.009266475215554237, + 0.008539386093616486, + 0.003790547838434577, + 0.006233387626707554, + 0.0029114484786987305, + 0.0042094080708920956, + 0.004470399115234613, + 0.003012207103893161, + 0.0031854636035859585, + 0.004263195674866438, + 0.004518262110650539, + 0.005235840100795031, + 0.0029213973321020603, + 0.005838441662490368, + 0.003779415972530842, + 0.004733201116323471, + 0.003464615670964122, + 0.010116370394825935, + 0.009913085959851742, + 0.010421695187687874, + 0.011430625803768635, + 0.019076284021139145, + 0.0084232147783041, + 0.02089233323931694, + 0.009490560740232468, + 0.005970877595245838, + 0.004885056056082249, + 0.0037670438177883625, + 0.005225437227636576, + 0.008342978544533253, + 0.003685014322400093, + 0.0069250850938260555, + 0.003784500528126955, + 0.01857784576714039, + 0.011816169135272503, + 0.007421494461596012, + 0.00938267633318901, + 0.014411751180887222, + 0.008586869575083256, + 0.014532163739204407, + 0.014498542994260788, + 0.015947867184877396, + 0.023894786834716797, + 0.012534786947071552, + 0.023794136941432953, + 0.04538963735103607, + 0.013009389862418175, + 0.021816611289978027, + 0.015345037914812565, + 0.015543192625045776, + 0.017183419317007065, + 0.012341057881712914, + 0.013637036085128784, + 0.02078322507441044, + 0.014403682202100754, + 0.017216697335243225, + 0.010683225467801094, + 0.013251280412077904, + 0.012528831139206886, + 0.016391897574067116, + 0.02247999608516693, + 0.014798911288380623, + 0.016473542898893356, + 0.02269473671913147, + 0.019249457865953445, + 0.020914750173687935, + 0.016572844237089157, + 0.022218069061636925, + 0.018532002344727516, + 0.009688988327980042, + 0.006022518966346979, + 0.03439640253782272, + 0.020653367042541504 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 9, + 8 + ], + "selected_reward": -0.17828737199306488 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static medium shot taken from a low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a towering, dark, solid wood bookshelf with classical carved patterns?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the bookshelf appear cluttered with heavy hardcover books and aged paperbacks on sagging shelves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a crack appear on the central vertical support and spread rapidly causing the bookshelf to bend inward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do books cascade down like a waterfall and smash to the ground creating dust in a beam of light?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.11298675835132599 + ], + "opposite_logprob_tokens": [ + -2.2379868030548096 + ], + "logprob_reward": -0.11298675835132599, + "yes_no_margin_reward": 2.125, + "token_credit_reward": -0.11298675835132599, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15702657401561737, + 0.062230370938777924, + 0.06230498477816582, + 0.10909051448106766, + 0.07571374624967575, + 0.1235533058643341, + 0.10197316110134125, + 0.054575201123952866, + 0.08603330701589584, + 0.16749879717826843 + ], + "credit_map": [ + 0.02338145300745964, + 0.022671813145279884, + 0.02664274349808693, + 0.01977286860346794, + 0.02355826459825039, + 0.01796809583902359, + 0.012138055637478828, + 0.010893266648054123, + 0.010381054133176804, + 0.008945303969085217, + 0.00820942036807537, + 0.006853204686194658, + 0.007683085277676582, + 0.006760114803910255, + 0.006967849563807249, + 0.006430340465158224, + 0.01068909652531147, + 0.007309230975806713, + 0.011847943998873234, + 0.005994236096739769, + 0.008325600065290928, + 0.005713427439332008, + 0.006015930324792862, + 0.006409517023712397, + 0.01308070495724678, + 0.01441043522208929, + 0.014525717124342918, + 0.010937691666185856, + 0.016857629641890526, + 0.01036310475319624, + 0.019830968230962753, + 0.009084261022508144, + 0.009733990766108036, + 0.007800941355526447, + 0.005646970588713884, + 0.007445723284035921, + 0.021949872374534607, + 0.0045482865534722805, + 0.010778038762509823, + 0.007809920236468315, + 0.021106770262122154, + 0.01719692535698414, + 0.009505177848041058, + 0.013243083842098713, + 0.018457891419529915, + 0.010505033656954765, + 0.01713596098124981, + 0.016402464359998703, + 0.013769624754786491, + 0.012746801599860191, + 0.007752746343612671, + 0.010763289406895638, + 0.02204364910721779, + 0.011592723429203033, + 0.01306917518377304, + 0.010235151275992393, + 0.010983755812048912, + 0.005823210813105106, + 0.006471719592809677, + 0.004449436906725168, + 0.008979149162769318, + 0.006233694963157177, + 0.0069960979744791985, + 0.004638138692826033, + 0.007991671562194824, + 0.006781520787626505, + 0.009935352019965649, + 0.012697196565568447, + 0.009095762856304646, + 0.010936292819678783, + 0.0115054277703166, + 0.01709008775651455, + 0.020207712426781654, + 0.019480058923363686, + 0.019854461774230003, + 0.021659089252352715, + 0.012164671905338764, + 0.009451690129935741, + 0.039498958736658096, + 0.025182155892252922 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 0, + 5 + ], + "selected_reward": -0.11298675835132599 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static medium shot taken from a low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a towering, dark, solid wood bookshelf with classical carved patterns?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the bookshelf appear cluttered with heavy hardcover books and aged paperbacks on sagging shelves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a crack appear on the central vertical support and spread rapidly causing the bookshelf to bend inward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do books cascade down like a waterfall and smash to the ground creating dust in a beam of light?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.252118319272995 + ], + "opposite_logprob_tokens": [ + -1.5021183490753174 + ], + "logprob_reward": -0.252118319272995, + "yes_no_margin_reward": 1.25, + "token_credit_reward": -0.252118319272995, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1603599488735199, + 0.09457413107156754, + 0.03402205556631088, + 0.0671611875295639, + 0.044800762087106705, + 0.09818808734416962, + 0.04680822789669037, + 0.04370027035474777, + 0.07409460097551346, + 0.33629080653190613 + ], + "credit_map": [ + 0.0240174513310194, + 0.02201267145574093, + 0.02207537554204464, + 0.01802721805870533, + 0.02362511307001114, + 0.01784750260412693, + 0.015459805727005005, + 0.017294812947511673, + 0.01641867123544216, + 0.01607383042573929, + 0.014252789318561554, + 0.011720815673470497, + 0.010452095419168472, + 0.009904615581035614, + 0.009894945658743382, + 0.005856366828083992, + 0.0035864338278770447, + 0.006789702456444502, + 0.004173979628831148, + 0.0028324180748313665, + 0.0050086830742657185, + 0.00527083408087492, + 0.003483549226075411, + 0.0028764568269252777, + 0.007490345276892185, + 0.008571693673729897, + 0.009778245352208614, + 0.007636064663529396, + 0.010440700687468052, + 0.005547741428017616, + 0.01165465172380209, + 0.006041740998625755, + 0.004661067388951778, + 0.004724982660263777, + 0.0032685003243386745, + 0.004224363714456558, + 0.011967821046710014, + 0.004515525884926319, + 0.006654921919107437, + 0.004783580079674721, + 0.01141422986984253, + 0.01471821777522564, + 0.006711607798933983, + 0.010050456039607525, + 0.0161235760897398, + 0.009942728094756603, + 0.01227064523845911, + 0.016956616193056107, + 0.0037893778644502163, + 0.005094858817756176, + 0.004288502968847752, + 0.005004056263715029, + 0.009041312150657177, + 0.0047761620953679085, + 0.00628023874014616, + 0.008533718064427376, + 0.004566618707031012, + 0.00731850229203701, + 0.003254543524235487, + 0.003066536970436573, + 0.010017847642302513, + 0.005237885285168886, + 0.007183116860687733, + 0.003055219305679202, + 0.00844394788146019, + 0.00628116587176919, + 0.009475942701101303, + 0.010131467133760452, + 0.008263212628662586, + 0.006359175778925419, + 0.014250117354094982, + 0.010889576748013496, + 0.06046440452337265, + 0.040196917951107025, + 0.042848486453294754, + 0.04676133766770363, + 0.019360389560461044, + 0.013047750107944012, + 0.07773765921592712, + 0.03587385639548302 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 0, + 5 + ], + "selected_reward": -0.252118319272995 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static medium shot taken from a low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a towering, dark, solid wood bookshelf with classical carved patterns?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the bookshelf appear cluttered with heavy hardcover books and aged paperbacks on sagging shelves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a crack appear on the central vertical support and spread rapidly causing the bookshelf to bend inward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do books cascade down like a waterfall and smash to the ground creating dust in a beam of light?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02980433776974678 + ], + "opposite_logprob_tokens": [ + -3.529804229736328 + ], + "logprob_reward": -0.02980433776974678, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.02980433776974678, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15863589942455292, + 0.05100400373339653, + 0.024837790057063103, + 0.03487103059887886, + 0.11266893893480301, + 0.04707686975598335, + 0.13905136287212372, + 0.05026651546359062, + 0.11513206362724304, + 0.2664555311203003 + ], + "credit_map": [ + 0.020307840779423714, + 0.025651084259152412, + 0.02055743709206581, + 0.019083816558122635, + 0.025300677865743637, + 0.018828673288226128, + 0.01575540378689766, + 0.013150968588888645, + 0.007003907114267349, + 0.010116844438016415, + 0.005549611058086157, + 0.0058556292206048965, + 0.0056914594024419785, + 0.006221428979188204, + 0.006006223615258932, + 0.004558900836855173, + 0.0040791938081383705, + 0.002703535370528698, + 0.00281893159262836, + 0.0026025644037872553, + 0.002787235891446471, + 0.0038956415373831987, + 0.0029196583200246096, + 0.0030310286674648523, + 0.004568536765873432, + 0.002927510067820549, + 0.004464420955628157, + 0.002954917261376977, + 0.006640083156526089, + 0.0029682626482099295, + 0.006971538998186588, + 0.003375760745257139, + 0.009565280750393867, + 0.007817323319613934, + 0.008265397511422634, + 0.006923857610672712, + 0.04529614374041557, + 0.009108572266995907, + 0.017715703696012497, + 0.007976650260388851, + 0.007832261733710766, + 0.00694661308079958, + 0.0035193387884646654, + 0.004590457305312157, + 0.006824841257184744, + 0.003931695595383644, + 0.006698929704725742, + 0.006732727400958538, + 0.00912878755480051, + 0.01463324949145317, + 0.013322563841938972, + 0.013575676828622818, + 0.03020806983113289, + 0.014155065640807152, + 0.015161502175033092, + 0.02886645868420601, + 0.004850675817579031, + 0.003832032438367605, + 0.006101278122514486, + 0.0032718901056796312, + 0.013691303320229053, + 0.005996870342642069, + 0.00840585958212614, + 0.004116603173315525, + 0.016907550394535065, + 0.011531081050634384, + 0.013481443747878075, + 0.01609381102025509, + 0.012639300897717476, + 0.008757729083299637, + 0.023133108392357826, + 0.012588035315275192, + 0.040617723017930984, + 0.02773289754986763, + 0.04748116433620453, + 0.040204863995313644, + 0.02153898775577545, + 0.009375523775815964, + 0.04820268973708153, + 0.03130169212818146 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 0, + 6 + ], + "selected_reward": -0.02980433776974678 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static medium shot taken from a low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a towering, dark, solid wood bookshelf with classical carved patterns?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the bookshelf appear cluttered with heavy hardcover books and aged paperbacks on sagging shelves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a crack appear on the central vertical support and spread rapidly causing the bookshelf to bend inward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do books cascade down like a waterfall and smash to the ground creating dust in a beam of light?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.011080753989517689 + ], + "opposite_logprob_tokens": [ + -4.511080741882324 + ], + "logprob_reward": -0.011080753989517689, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.011080753989517689, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07295950502157211, + 0.02765655517578125, + 0.007566411048173904, + 0.021984880790114403, + 0.012676346115767956, + 0.028862720355391502, + 0.03595158830285072, + 0.038409601897001266, + 0.07252845168113708, + 0.6814039349555969 + ], + "credit_map": [ + 0.010307889431715012, + 0.007200406398624182, + 0.010643777437508106, + 0.009717846289277077, + 0.014548113569617271, + 0.006855469197034836, + 0.008029062300920486, + 0.0056569394655525684, + 0.004872975405305624, + 0.004087772686034441, + 0.003241854254156351, + 0.003679882735013962, + 0.0030533268582075834, + 0.00299365958198905, + 0.003104341449216008, + 0.0026227417401969433, + 0.0013369264779612422, + 0.0008125329622998834, + 0.0009621056960895658, + 0.00066709506791085, + 0.0008179868455044925, + 0.0011405078694224358, + 0.0009427200420759618, + 0.0008865352137945592, + 0.0024664669763296843, + 0.0018674202729016542, + 0.002905631437897682, + 0.0024125692434608936, + 0.0037595625035464764, + 0.001854876521974802, + 0.004797053523361683, + 0.0019213006598874927, + 0.00178567657712847, + 0.001388323144055903, + 0.0013623831328004599, + 0.0009866702603176236, + 0.0026826849207282066, + 0.0010298383422195911, + 0.0023022382520139217, + 0.0011385309044271708, + 0.00488275196403265, + 0.004755224566906691, + 0.0026429814752191305, + 0.0027481303550302982, + 0.004262517672032118, + 0.0022370931692421436, + 0.003856112016364932, + 0.0034779116977006197, + 0.0038271520752459764, + 0.0040208930149674416, + 0.003974901977926493, + 0.004214202519506216, + 0.004877871368080378, + 0.004303110297769308, + 0.0038417698815464973, + 0.00689168693497777, + 0.004328215029090643, + 0.006548217032104731, + 0.0036757062189280987, + 0.003679040353745222, + 0.007511443458497524, + 0.0045456490479409695, + 0.004688560962677002, + 0.003432769328355789, + 0.009038612246513367, + 0.006387826520949602, + 0.010408693924546242, + 0.012544660829007626, + 0.008264413103461266, + 0.005275796167552471, + 0.01162354089319706, + 0.00898490846157074, + 0.11670125275850296, + 0.0768810510635376, + 0.0872180312871933, + 0.09021443873643875, + 0.03701396286487579, + 0.025365546345710754, + 0.16860148310661316, + 0.07940810173749924 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 0, + 8 + ], + "selected_reward": -0.011080753989517689 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000071/group_5/index_tensor([7328, 7328])_group_5_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000071/group_5/index_tensor([7328, 7328])_group_5_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..ca811b8001f5555e358920aadb05368bab08bc39 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000071/group_5/index_tensor([7328, 7328])_group_5_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "medium shot, at eye level, initially depicts an elderly man with kindly East Asian features, neatly combed silver hair, and a face etched with the gentle wrinkles of a happy life. He wears a soft beige cashmere cardigan over a crisp light blue button-down shirt and dark gray trousers, his arms folded, his posture straight yet relaxed. He walks along a weathered stone path with a deliberate, rhythmic pace, his dark brown leather shoes making almost no sound on the scattered, dry autumn leaves. As he walks, the camera begins a slow, fluid pullback, steadily widening the perspective. This movement gradually reveals the tranquil, leafy setting of Serenity Park in late afternoon. Low, golden sunlight filters through the vibrant canopies of towering ginkgo and maple trees, casting long, soft shadows and making the crimson and amber leaves on the ground sparkle. The ever-expanding view reveals a vast, manicured lawn to one side and the soft glow of a pond beyond, creating an atmosphere of deep tranquility and contemplation.", + "seed": 4774, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000071/group_5/index_7328_seed_4774_dp_10_rank_20_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 5, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.02231482192873955, + "vlm_logprob": -0.02231482192873955, + "vlm_yes_no_margin": 3.975, + "vlm_token_credit": -0.02231482192873955, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an elderly man with East Asian features and silver hair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man wearing a beige cashmere cardigan over a light blue button-down shirt and dark gray trousers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the man walk along a weathered stone path with a deliberate, rhythmic pace?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera begin a slow, fluid pullback that widens the perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the setting a park with ginkgo and maple trees, a manicured lawn, and a pond in late afternoon light?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02191423811018467 + ], + "opposite_logprob_tokens": [ + -3.89691424369812 + ], + "logprob_reward": -0.02191423811018467, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.02191423811018467, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10602559894323349, + 0.09294397383928299, + 0.09416497498750687, + 0.1043572723865509, + 0.0660293847322464, + 0.0632452517747879, + 0.0509258396923542, + 0.17543946206569672, + 0.1307021975517273, + 0.11616609990596771 + ], + "credit_map": [ + 0.0064199864864349365, + 0.005922319367527962, + 0.018935924395918846, + 0.01655750349164009, + 0.007186247501522303, + 0.01556750200688839, + 0.018076756969094276, + 0.017359357327222824, + 0.008700083941221237, + 0.007289918605238199, + 0.007548080757260323, + 0.00813237763941288, + 0.01958581432700157, + 0.015980908647179604, + 0.018072782084345818, + 0.0076340120285749435, + 0.010020303539931774, + 0.008511618711054325, + 0.009627309627830982, + 0.00829508900642395, + 0.01874048076570034, + 0.010176005773246288, + 0.01855674386024475, + 0.01023742277175188, + 0.009164716117084026, + 0.00828211847692728, + 0.009079248644411564, + 0.009101972915232182, + 0.02036842703819275, + 0.01743241585791111, + 0.018995443359017372, + 0.011932924389839172, + 0.005692596547305584, + 0.006683874875307083, + 0.008617788553237915, + 0.01188361831009388, + 0.007866062223911285, + 0.005513710435479879, + 0.009705575183033943, + 0.010066153481602669, + 0.0033928663469851017, + 0.0028236613143235445, + 0.006571490317583084, + 0.004151046741753817, + 0.004005027934908867, + 0.016666386276483536, + 0.004637159872800112, + 0.020997611805796623, + 0.007134672719985247, + 0.004751122090965509, + 0.0058335307985544205, + 0.004646423738449812, + 0.009432957507669926, + 0.005559146404266357, + 0.008454953320324421, + 0.005113030783832073, + 0.01169226597994566, + 0.016035046428442, + 0.02293069288134575, + 0.036341000348329544, + 0.022205853834748268, + 0.014194449409842491, + 0.025477634742856026, + 0.02656252309679985, + 0.009906182996928692, + 0.020421279594302177, + 0.014812381006777287, + 0.02052329108119011, + 0.008418252691626549, + 0.006302433088421822, + 0.030294077470898628, + 0.020024308934807777, + 0.01373552717268467, + 0.005949081387370825, + 0.00786865409463644, + 0.012980353087186813, + 0.0068364073522388935, + 0.007005712948739529, + 0.01871293969452381, + 0.043077435344457626 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 8, + 9 + ], + "selected_reward": -0.02191423811018467 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an elderly man with East Asian features and silver hair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man wearing a beige cashmere cardigan over a light blue button-down shirt and dark gray trousers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the man walk along a weathered stone path with a deliberate, rhythmic pace?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera begin a slow, fluid pullback that widens the perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the setting a park with ginkgo and maple trees, a manicured lawn, and a pond in late afternoon light?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.014211048372089863 + ], + "opposite_logprob_tokens": [ + -4.264211177825928 + ], + "logprob_reward": -0.014211048372089863, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.014211048372089863, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1976921409368515, + 0.09450530260801315, + 0.11901950091123581, + 0.06196320056915283, + 0.06613628566265106, + 0.04512632265686989, + 0.09672868996858597, + 0.108780138194561, + 0.12236091494560242, + 0.08768749982118607 + ], + "credit_map": [ + 0.014569500461220741, + 0.009858844801783562, + 0.018920602276921272, + 0.017458021640777588, + 0.015835123136639595, + 0.03067765198647976, + 0.05220436304807663, + 0.0381680391728878, + 0.006302819121629, + 0.005304103717207909, + 0.008133742026984692, + 0.009076916612684727, + 0.01855071261525154, + 0.014143363572657108, + 0.021251467987895012, + 0.011742179282009602, + 0.00812257919460535, + 0.007207122631371021, + 0.011598820798099041, + 0.013973528519272804, + 0.013682412914931774, + 0.014771337620913982, + 0.03056071698665619, + 0.019102979451417923, + 0.004038621671497822, + 0.00415274640545249, + 0.006384396925568581, + 0.006146462634205818, + 0.008910845033824444, + 0.009367208927869797, + 0.012965003028512001, + 0.00999792292714119, + 0.006730543915182352, + 0.00658854003995657, + 0.00837075524032116, + 0.009944095276296139, + 0.010796102695167065, + 0.006868378259241581, + 0.009185967966914177, + 0.007651900872588158, + 0.0028679934330284595, + 0.003913406748324633, + 0.004336548503488302, + 0.004865666385740042, + 0.008001736365258694, + 0.009722195565700531, + 0.004521622788161039, + 0.00689715426415205, + 0.00828677136451006, + 0.009196379221975803, + 0.00990312173962593, + 0.010986200533807278, + 0.020413795486092567, + 0.009255640208721161, + 0.021199611946940422, + 0.007487171329557896, + 0.008128470741212368, + 0.00938964169472456, + 0.01103910617530346, + 0.013179208151996136, + 0.02430148981511593, + 0.009577988646924496, + 0.016467634588479996, + 0.01669660024344921, + 0.008396027609705925, + 0.017805645242333412, + 0.01647823676466942, + 0.02082785777747631, + 0.014310990460216999, + 0.0075441645458340645, + 0.020596405491232872, + 0.016401585191488266, + 0.009249920956790447, + 0.006861869711428881, + 0.012470677495002747, + 0.007955363020300865, + 0.006561361253261566, + 0.006071656942367554, + 0.028313979506492615, + 0.010202664881944656 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 2 + ], + "selected_reward": -0.014211048372089863 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an elderly man with East Asian features and silver hair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man wearing a beige cashmere cardigan over a light blue button-down shirt and dark gray trousers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the man walk along a weathered stone path with a deliberate, rhythmic pace?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera begin a slow, fluid pullback that widens the perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the setting a park with ginkgo and maple trees, a manicured lawn, and a pond in late afternoon light?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.018182270228862762 + ], + "opposite_logprob_tokens": [ + -4.018182277679443 + ], + "logprob_reward": -0.018182270228862762, + "yes_no_margin_reward": 4.0, + "token_credit_reward": -0.018182270228862762, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.22680547833442688, + 0.10592595487833023, + 0.09717545658349991, + 0.07080037891864777, + 0.11314015090465546, + 0.07715681195259094, + 0.10664263367652893, + 0.07632613182067871, + 0.09045634418725967, + 0.03557073324918747 + ], + "credit_map": [ + 0.011639543808996677, + 0.008377966471016407, + 0.025092436000704765, + 0.022395577281713486, + 0.01458351593464613, + 0.03968021646142006, + 0.05111918970942497, + 0.05391703546047211, + 0.00602813670411706, + 0.007144314702600241, + 0.007496556267142296, + 0.008390550501644611, + 0.022899484261870384, + 0.015464916825294495, + 0.025722652673721313, + 0.01277934294193983, + 0.007747234310954809, + 0.007761907763779163, + 0.009095768444240093, + 0.009614542126655579, + 0.019889188930392265, + 0.011290558613836765, + 0.02112962305545807, + 0.010646634735167027, + 0.005993260070681572, + 0.005816818214952946, + 0.007719229906797409, + 0.005770071875303984, + 0.009240101091563702, + 0.01398809440433979, + 0.011098295450210571, + 0.011174513958394527, + 0.00986003689467907, + 0.011173861101269722, + 0.012205573730170727, + 0.016803592443466187, + 0.017484469339251518, + 0.01397525891661644, + 0.017604218795895576, + 0.014033148065209389, + 0.004935726057738066, + 0.0063790068961679935, + 0.007680184673517942, + 0.0066506750881671906, + 0.006867583841085434, + 0.017769932746887207, + 0.007531574461609125, + 0.019342126324772835, + 0.006106185261160135, + 0.0049603949300944805, + 0.031210551038384438, + 0.00735008530318737, + 0.008106350898742676, + 0.007407394703477621, + 0.00872405618429184, + 0.032777611166238785, + 0.004579311236739159, + 0.007285574451088905, + 0.007877247408032417, + 0.013658968731760979, + 0.015176028944551945, + 0.006252508144825697, + 0.010915756225585938, + 0.010580737143754959, + 0.0058429427444934845, + 0.014657709747552872, + 0.010269323363900185, + 0.013923567719757557, + 0.0060584633611142635, + 0.003968247678130865, + 0.019725771620869637, + 0.01601031981408596, + 0.002965474734082818, + 0.0036048227921128273, + 0.004013293422758579, + 0.0039922501891851425, + 0.004754774272441864, + 0.004830764140933752, + 0.007139420136809349, + 0.004269934259355068 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 6 + ], + "selected_reward": -0.018182270228862762 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an elderly man with East Asian features and silver hair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man wearing a beige cashmere cardigan over a light blue button-down shirt and dark gray trousers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the man walk along a weathered stone path with a deliberate, rhythmic pace?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera begin a slow, fluid pullback that widens the perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the setting a park with ginkgo and maple trees, a manicured lawn, and a pond in late afternoon light?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.048632293939590454 + ], + "opposite_logprob_tokens": [ + -3.0486323833465576 + ], + "logprob_reward": -0.048632293939590454, + "yes_no_margin_reward": 3.0, + "token_credit_reward": -0.048632293939590454, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12347409129142761, + 0.1183597519993782, + 0.02681133709847927, + 0.07071147859096527, + 0.0954892560839653, + 0.06814637780189514, + 0.031807489693164825, + 0.22689352929592133, + 0.14128783345222473, + 0.09701883047819138 + ], + "credit_map": [ + 0.007137048989534378, + 0.009231463074684143, + 0.01764451153576374, + 0.013502978719770908, + 0.01021764986217022, + 0.01935279183089733, + 0.024280155077576637, + 0.02210749313235283, + 0.007951403968036175, + 0.008205155842006207, + 0.008203920908272266, + 0.00840961653739214, + 0.027949724346399307, + 0.02127899043262005, + 0.02323669008910656, + 0.01312424149364233, + 0.0034785037860274315, + 0.002437946852296591, + 0.004444517195224762, + 0.0033408370800316334, + 0.0032246243208646774, + 0.0028216198552399874, + 0.0041769216768443584, + 0.002886365167796612, + 0.006767059210687876, + 0.00474350294098258, + 0.012044091708958149, + 0.008936729282140732, + 0.009485054761171341, + 0.01135169342160225, + 0.009282873012125492, + 0.008100466802716255, + 0.006746239494532347, + 0.00704588508233428, + 0.010862081311643124, + 0.015383241698145866, + 0.016781644895672798, + 0.011517325416207314, + 0.013245775364339352, + 0.013907067477703094, + 0.003366699442267418, + 0.003080436261370778, + 0.006915112491697073, + 0.0036216634325683117, + 0.0036170624662190676, + 0.018419736996293068, + 0.0045370738953351974, + 0.02458859793841839, + 0.0037865168415009975, + 0.0021494929678738117, + 0.004154365975409746, + 0.004129799082875252, + 0.005353109445422888, + 0.0029887203127145767, + 0.005189580377191305, + 0.004055903758853674, + 0.016203023493289948, + 0.01840072125196457, + 0.03585943207144737, + 0.038990579545497894, + 0.01648728922009468, + 0.01795821078121662, + 0.03302740678191185, + 0.049966856837272644, + 0.010281065478920937, + 0.019911881536245346, + 0.013592252507805824, + 0.020472818985581398, + 0.011335107497870922, + 0.006129458546638489, + 0.03498569503426552, + 0.02457955852150917, + 0.010015083476901054, + 0.009179257787764072, + 0.01883411593735218, + 0.013912797905504704, + 0.011452080681920052, + 0.0072839949280023575, + 0.014713943004608154, + 0.011627556756138802 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 8, + 0 + ], + "selected_reward": -0.048632293939590454 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an elderly man with East Asian features and silver hair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man wearing a beige cashmere cardigan over a light blue button-down shirt and dark gray trousers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the man walk along a weathered stone path with a deliberate, rhythmic pace?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera begin a slow, fluid pullback that widens the perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the setting a park with ginkgo and maple trees, a manicured lawn, and a pond in late afternoon light?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.00863425899296999 + ], + "opposite_logprob_tokens": [ + -4.758634090423584 + ], + "logprob_reward": -0.00863425899296999, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.00863425899296999, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2710868716239929, + 0.06722448021173477, + 0.055215492844581604, + 0.03367903456091881, + 0.0436762310564518, + 0.04985084757208824, + 0.09556569904088974, + 0.12309624999761581, + 0.20740336179733276, + 0.05320172384381294 + ], + "credit_map": [ + 0.014465449377894402, + 0.013431613333523273, + 0.02471063658595085, + 0.02436172589659691, + 0.026856277137994766, + 0.04273749887943268, + 0.08453411608934402, + 0.03998955339193344, + 0.0064635202288627625, + 0.004607240669429302, + 0.013837117701768875, + 0.009006693959236145, + 0.007280399557203054, + 0.009812102653086185, + 0.008313787169754505, + 0.007903624325990677, + 0.008193302899599075, + 0.004491331521421671, + 0.007170404773205519, + 0.006377249024808407, + 0.009172631427645683, + 0.006684991996735334, + 0.008443204686045647, + 0.004682376980781555, + 0.0030188418459147215, + 0.002994747366756201, + 0.0037429207004606724, + 0.0031902550254017115, + 0.005512137431651354, + 0.005598242394626141, + 0.004531023092567921, + 0.005090868566185236, + 0.008717406541109085, + 0.004811026621609926, + 0.005394072737544775, + 0.0036822485271841288, + 0.004136360250413418, + 0.004470148589462042, + 0.004916682839393616, + 0.00754828006029129, + 0.005095214582979679, + 0.003897827584296465, + 0.006381595972925425, + 0.004465694539248943, + 0.005429304204881191, + 0.010977263562381268, + 0.0043953657150268555, + 0.009208579547703266, + 0.012696512043476105, + 0.009845680557191372, + 0.011766788549721241, + 0.009304370731115341, + 0.016320521011948586, + 0.00932967010885477, + 0.015384664759039879, + 0.010917487554252148, + 0.006608197931200266, + 0.010804460383951664, + 0.014168627560138702, + 0.010840055532753468, + 0.027968712151050568, + 0.0124594047665596, + 0.017730388790369034, + 0.0225164033472538, + 0.01668889820575714, + 0.0299257542937994, + 0.021003644913434982, + 0.03385236859321594, + 0.024914132431149483, + 0.009327465668320656, + 0.046465564519166946, + 0.02522554062306881, + 0.0036469257902354, + 0.004418679978698492, + 0.004201991483569145, + 0.004708337597548962, + 0.00655353395268321, + 0.004786570090800524, + 0.012742995284497738, + 0.012142688035964966 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 7 + ], + "selected_reward": -0.00863425899296999 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "medium shot, at eye level, initially depicts an elderly man with kindly East Asian features, neatly combed silver hair, and a face etched with the gentle wrinkles of a happy life. He wears a soft beige cashmere cardigan over a crisp light blue button-down shirt and dark gray trousers, his arms folded, his posture straight yet relaxed. He walks along a weathered stone path with a deliberate, rhythmic pace, his dark brown leather shoes making almost no sound on the scattered, dry autumn leaves. As he walks, the camera begins a slow, fluid pullback, steadily widening the perspective. This movement gradually reveals the tranquil, leafy setting of Serenity Park in late afternoon. Low, golden sunlight filters through the vibrant canopies of towering ginkgo and maple trees, casting long, soft shadows and making the crimson and amber leaves on the ground sparkle. The ever-expanding view reveals a vast, manicured lawn to one side and the soft glow of a pond beyond, creating an atmosphere of deep tranquility and contemplation.", + "seed": 4775, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000071/group_5/index_7328_seed_4775_dp_10_rank_20_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 5, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.03171015437692404, + "vlm_logprob": -0.03171015437692404, + "vlm_yes_no_margin": 3.6, + "vlm_token_credit": -0.03171015437692404, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an elderly man with East Asian features and silver hair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man wearing a beige cashmere cardigan over a light blue button-down shirt and dark gray trousers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the man walk along a weathered stone path with a deliberate, rhythmic pace?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera begin a slow, fluid pullback that widens the perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the setting a park with ginkgo and maple trees, a manicured lawn, and a pond in late afternoon light?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.027679869905114174 + ], + "opposite_logprob_tokens": [ + -3.652679920196533 + ], + "logprob_reward": -0.027679869905114174, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.027679869905114174, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.16197027266025543, + 0.15016980469226837, + 0.06714177131652832, + 0.15295164287090302, + 0.06566671282052994, + 0.042988233268260956, + 0.032975196838378906, + 0.08068054914474487, + 0.08794211596250534, + 0.15751373767852783 + ], + "credit_map": [ + 0.02470005676150322, + 0.016929810866713524, + 0.015277322381734848, + 0.011467540636658669, + 0.026735477149486542, + 0.018422113731503487, + 0.0323830284178257, + 0.016054922714829445, + 0.018670829012989998, + 0.014847340062260628, + 0.010526709258556366, + 0.010966397821903229, + 0.033285465091466904, + 0.014364928007125854, + 0.03631339222192764, + 0.011194749735295773, + 0.011492854915559292, + 0.00649990513920784, + 0.007915993221104145, + 0.004278667736798525, + 0.011591870337724686, + 0.0075888861902058125, + 0.009666260331869125, + 0.008107336238026619, + 0.01994921825826168, + 0.008929921314120293, + 0.04092816635966301, + 0.008467701263725758, + 0.010150694288313389, + 0.01234035287052393, + 0.014627965167164803, + 0.037557609379291534, + 0.018368931487202644, + 0.008792919106781483, + 0.007857912220060825, + 0.0031889777164906263, + 0.005635205656290054, + 0.007695446722209454, + 0.006411304231733084, + 0.007716010790318251, + 0.005570762790739536, + 0.003471068572252989, + 0.005585032515227795, + 0.004822198301553726, + 0.0068970019929111, + 0.005383713636547327, + 0.005609097424894571, + 0.005649358965456486, + 0.002900260267779231, + 0.0032141939736902714, + 0.004543509799987078, + 0.0021909307688474655, + 0.005337547045201063, + 0.0033436676021665335, + 0.005603959783911705, + 0.005841126665472984, + 0.0065189506858587265, + 0.00736090587452054, + 0.008512010797858238, + 0.010171730071306229, + 0.0089718634262681, + 0.008601902984082699, + 0.019699135795235634, + 0.010844050906598568, + 0.010457676835358143, + 0.008560975082218647, + 0.010893530212342739, + 0.012417859397828579, + 0.014271775260567665, + 0.01019285898655653, + 0.011722614988684654, + 0.009424831718206406, + 0.019391467794775963, + 0.012760963290929794, + 0.016270307824015617, + 0.016032949090003967, + 0.03198453411459923, + 0.01935567334294319, + 0.02118237130343914, + 0.020535465329885483 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 3 + ], + "selected_reward": -0.027679869905114174 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an elderly man with East Asian features and silver hair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man wearing a beige cashmere cardigan over a light blue button-down shirt and dark gray trousers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the man walk along a weathered stone path with a deliberate, rhythmic pace?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera begin a slow, fluid pullback that widens the perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the setting a park with ginkgo and maple trees, a manicured lawn, and a pond in late afternoon light?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01820732094347477 + ], + "opposite_logprob_tokens": [ + -4.018207550048828 + ], + "logprob_reward": -0.01820732094347477, + "yes_no_margin_reward": 4.0, + "token_credit_reward": -0.01820732094347477, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.20990075170993805, + 0.07403162121772766, + 0.0461353063583374, + 0.1861562877893448, + 0.06046131253242493, + 0.07738439738750458, + 0.07732465118169785, + 0.07761290669441223, + 0.10063790529966354, + 0.09035491198301315 + ], + "credit_map": [ + 0.02801409922540188, + 0.026311341673135757, + 0.02312440052628517, + 0.015332420356571674, + 0.020789986476302147, + 0.02531607076525688, + 0.030458979308605194, + 0.04055343195796013, + 0.010642542503774166, + 0.011398538947105408, + 0.005145499482750893, + 0.007220685947686434, + 0.00988117977976799, + 0.010384934023022652, + 0.012733976356685162, + 0.0066242581233382225, + 0.00775811355561018, + 0.004598929546773434, + 0.0040382398292422295, + 0.003904173616319895, + 0.006631950847804546, + 0.005942698568105698, + 0.007583544589579105, + 0.005677653476595879, + 0.016101958230137825, + 0.011365576647222042, + 0.045308757573366165, + 0.00816156342625618, + 0.01642904244363308, + 0.015622790902853012, + 0.017465760931372643, + 0.05570082366466522, + 0.012832761742174625, + 0.0061606490053236485, + 0.009510370902717113, + 0.004122913349419832, + 0.00479467585682869, + 0.008367047645151615, + 0.006814818829298019, + 0.007858074270188808, + 0.013042519800364971, + 0.006220465991646051, + 0.007524931337684393, + 0.007052956614643335, + 0.008606739342212677, + 0.010967821814119816, + 0.014643516391515732, + 0.009325439110398293, + 0.007082546129822731, + 0.007759390398859978, + 0.007187122479081154, + 0.008116739802062511, + 0.016396205872297287, + 0.00766829214990139, + 0.01334664598107338, + 0.009767702780663967, + 0.009533669799566269, + 0.008625579066574574, + 0.010225856676697731, + 0.008701146580278873, + 0.006254651118069887, + 0.00984227005392313, + 0.01040270458906889, + 0.014027023687958717, + 0.011704696342349052, + 0.008297470398247242, + 0.02267909236252308, + 0.014415177516639233, + 0.008522380143404007, + 0.00945008173584938, + 0.010481194593012333, + 0.015087797306478024, + 0.008648463524878025, + 0.008821509778499603, + 0.009285838343203068, + 0.014612817205488682, + 0.013694497756659985, + 0.011339515447616577, + 0.014095296151936054, + 0.009856967255473137 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 8 + ], + "selected_reward": -0.01820732094347477 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an elderly man with East Asian features and silver hair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man wearing a beige cashmere cardigan over a light blue button-down shirt and dark gray trousers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the man walk along a weathered stone path with a deliberate, rhythmic pace?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera begin a slow, fluid pullback that widens the perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the setting a park with ginkgo and maple trees, a manicured lawn, and a pond in late afternoon light?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.062038760632276535 + ], + "opposite_logprob_tokens": [ + -2.8120386600494385 + ], + "logprob_reward": -0.062038760632276535, + "yes_no_margin_reward": 2.75, + "token_credit_reward": -0.062038760632276535, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.19524133205413818, + 0.05814630910754204, + 0.05250965431332588, + 0.23664012551307678, + 0.05302226170897484, + 0.05726688355207443, + 0.055508408695459366, + 0.10846656560897827, + 0.09299023449420929, + 0.0902082696557045 + ], + "credit_map": [ + 0.0315994918346405, + 0.027939366176724434, + 0.017364995554089546, + 0.01581229269504547, + 0.018932942301034927, + 0.02477642148733139, + 0.018518945202231407, + 0.04029686376452446, + 0.007790549658238888, + 0.007606284227222204, + 0.004320759791880846, + 0.005331693682819605, + 0.010373744182288647, + 0.006976904347538948, + 0.011049735359847546, + 0.004696633201092482, + 0.010125406086444855, + 0.005148748401552439, + 0.007878907024860382, + 0.003661432070657611, + 0.0077161844819784164, + 0.0060773142613470554, + 0.006372385658323765, + 0.005529272835701704, + 0.017632117494940758, + 0.01246778666973114, + 0.06491230428218842, + 0.01083009410649538, + 0.022445881739258766, + 0.01905759423971176, + 0.021976260468363762, + 0.06731806695461273, + 0.011713016778230667, + 0.00847967341542244, + 0.006255342625081539, + 0.003346968675032258, + 0.005355802830308676, + 0.006991435773670673, + 0.0047903950326144695, + 0.006089620292186737, + 0.010956170968711376, + 0.0042571984231472015, + 0.006635864265263081, + 0.005423877388238907, + 0.007664654869586229, + 0.006259443704038858, + 0.007674879860132933, + 0.008394789882004261, + 0.004748946521431208, + 0.003995687700808048, + 0.0070883953012526035, + 0.005189075134694576, + 0.007830881513655186, + 0.006178742740303278, + 0.012839074246585369, + 0.0076376041397452354, + 0.013060949742794037, + 0.011827603913843632, + 0.013069313950836658, + 0.010625598952174187, + 0.009361115284264088, + 0.015846313908696175, + 0.017537809908390045, + 0.01713785156607628, + 0.011263222433626652, + 0.007329881656914949, + 0.02200634405016899, + 0.014092489145696163, + 0.005261573009192944, + 0.010539502836763859, + 0.007159127853810787, + 0.015338088385760784, + 0.008704515174031258, + 0.0072234198451042175, + 0.010000105015933514, + 0.011641093529760838, + 0.018433012068271637, + 0.011639682576060295, + 0.012563122436404228, + 0.010003304108977318 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 0, + 7 + ], + "selected_reward": -0.062038760632276535 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an elderly man with East Asian features and silver hair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man wearing a beige cashmere cardigan over a light blue button-down shirt and dark gray trousers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the man walk along a weathered stone path with a deliberate, rhythmic pace?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera begin a slow, fluid pullback that widens the perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the setting a park with ginkgo and maple trees, a manicured lawn, and a pond in late afternoon light?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03808605298399925 + ], + "opposite_logprob_tokens": [ + -3.2880859375 + ], + "logprob_reward": -0.03808605298399925, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.03808605298399925, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.16126814484596252, + 0.06797102838754654, + 0.06482406705617905, + 0.10275706648826599, + 0.08659736812114716, + 0.06319747120141983, + 0.14222998917102814, + 0.1422991305589676, + 0.1027875691652298, + 0.0660681501030922 + ], + "credit_map": [ + 0.030598875135183334, + 0.01703067682683468, + 0.01648341678082943, + 0.014924931339919567, + 0.022722115740180016, + 0.01711813546717167, + 0.03131648525595665, + 0.011073519475758076, + 0.00849703699350357, + 0.00971116404980421, + 0.007211914751678705, + 0.010602236725389957, + 0.007289041765034199, + 0.006853140890598297, + 0.009274289011955261, + 0.008532206527888775, + 0.013385406695306301, + 0.00709698349237442, + 0.007581408601254225, + 0.0044587780721485615, + 0.007833398878574371, + 0.006616299506276846, + 0.008336538448929787, + 0.00951525941491127, + 0.013725277036428452, + 0.00759714562445879, + 0.018821461126208305, + 0.007938736118376255, + 0.008684064261615276, + 0.011118700727820396, + 0.01109174732118845, + 0.023779943585395813, + 0.015369294211268425, + 0.01613490842282772, + 0.010450080037117004, + 0.006967546883970499, + 0.010134465992450714, + 0.00897168181836605, + 0.008479716256260872, + 0.010089687071740627, + 0.015648508444428444, + 0.006279753055423498, + 0.006355364341288805, + 0.004042890854179859, + 0.007642796263098717, + 0.007239026948809624, + 0.008622672408819199, + 0.007366465870290995, + 0.018401747569441795, + 0.019201459363102913, + 0.02994830347597599, + 0.007046414073556662, + 0.014795479364693165, + 0.01814166270196438, + 0.011228619143366814, + 0.023466311395168304, + 0.0208839550614357, + 0.020240306854248047, + 0.01699845865368843, + 0.012094900012016296, + 0.009060273878276348, + 0.022732893005013466, + 0.018853837624192238, + 0.021434523165225983, + 0.011426553130149841, + 0.00812446791678667, + 0.024988476186990738, + 0.012307334691286087, + 0.0067329611629247665, + 0.009609430097043514, + 0.00646122545003891, + 0.023137126117944717, + 0.008853647857904434, + 0.00597486412152648, + 0.010722086764872074, + 0.010751794092357159, + 0.006032784469425678, + 0.007928859442472458, + 0.0061590108089149, + 0.009645112790167332 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 7, + 6 + ], + "selected_reward": -0.03808605298399925 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an elderly man with East Asian features and silver hair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man wearing a beige cashmere cardigan over a light blue button-down shirt and dark gray trousers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the man walk along a weathered stone path with a deliberate, rhythmic pace?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera begin a slow, fluid pullback that widens the perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the setting a park with ginkgo and maple trees, a manicured lawn, and a pond in late afternoon light?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.012538767419755459 + ], + "opposite_logprob_tokens": [ + -4.387538909912109 + ], + "logprob_reward": -0.012538767419755459, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.012538767419755459, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.18400691449642181, + 0.06446782499551773, + 0.11628211289644241, + 0.11289006471633911, + 0.08520025759935379, + 0.07228531688451767, + 0.08899262547492981, + 0.0648333877325058, + 0.09945870190858841, + 0.11158274114131927 + ], + "credit_map": [ + 0.033557597547769547, + 0.019672710448503494, + 0.02118493989109993, + 0.014376956969499588, + 0.016603702679276466, + 0.019681600853800774, + 0.018069040030241013, + 0.040860388427972794, + 0.013420993462204933, + 0.010525655932724476, + 0.004857014864683151, + 0.008268077857792377, + 0.007289310451596975, + 0.00687501672655344, + 0.007213377859443426, + 0.006018383894115686, + 0.02110966295003891, + 0.016687918454408646, + 0.012570716440677643, + 0.006429463624954224, + 0.015430103056132793, + 0.015612023882567883, + 0.012041443027555943, + 0.016400791704654694, + 0.018974723294377327, + 0.013152588158845901, + 0.015684882178902626, + 0.009958868846297264, + 0.013561658561229706, + 0.011272924020886421, + 0.016394929960370064, + 0.013889499008655548, + 0.01827598549425602, + 0.007350014988332987, + 0.01556651946157217, + 0.006129745859652758, + 0.006887755822390318, + 0.014683744870126247, + 0.00576420733705163, + 0.01054229773581028, + 0.017419863492250443, + 0.00685031246393919, + 0.009667688049376011, + 0.004188614431768656, + 0.007197930011898279, + 0.010934760794043541, + 0.005120060406625271, + 0.010906090028584003, + 0.011333736591041088, + 0.012809240259230137, + 0.015058466233313084, + 0.004663644824177027, + 0.006536988075822592, + 0.01593489572405815, + 0.006277790758758783, + 0.01637786254286766, + 0.010399658232927322, + 0.008517187088727951, + 0.008718172088265419, + 0.006218917202204466, + 0.004285610746592283, + 0.008931690827012062, + 0.00792750809341669, + 0.009834649041295052, + 0.009537247940897942, + 0.006346369627863169, + 0.027225680649280548, + 0.008158581331372261, + 0.006713885348290205, + 0.009608509950339794, + 0.005403019953519106, + 0.02646542154252529, + 0.015094339847564697, + 0.010541285388171673, + 0.015494169667363167, + 0.024840498343110085, + 0.008284419775009155, + 0.013574645854532719, + 0.009775537997484207, + 0.01397784985601902 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 3 + ], + "selected_reward": -0.012538767419755459 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "medium shot, at eye level, initially depicts an elderly man with kindly East Asian features, neatly combed silver hair, and a face etched with the gentle wrinkles of a happy life. He wears a soft beige cashmere cardigan over a crisp light blue button-down shirt and dark gray trousers, his arms folded, his posture straight yet relaxed. He walks along a weathered stone path with a deliberate, rhythmic pace, his dark brown leather shoes making almost no sound on the scattered, dry autumn leaves. As he walks, the camera begins a slow, fluid pullback, steadily widening the perspective. This movement gradually reveals the tranquil, leafy setting of Serenity Park in late afternoon. Low, golden sunlight filters through the vibrant canopies of towering ginkgo and maple trees, casting long, soft shadows and making the crimson and amber leaves on the ground sparkle. The ever-expanding view reveals a vast, manicured lawn to one side and the soft glow of a pond beyond, creating an atmosphere of deep tranquility and contemplation.", + "seed": 4776, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000071/group_5/index_7328_seed_4776_dp_11_rank_22_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 5, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.024976956099271773, + "vlm_logprob": -0.024976956099271773, + "vlm_yes_no_margin": 3.825, + "vlm_token_credit": -0.024976956099271773, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an elderly man with East Asian features and silver hair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man wearing a beige cashmere cardigan over a light blue button-down shirt and dark gray trousers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the man walk along a weathered stone path with a deliberate, rhythmic pace?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera begin a slow, fluid pullback that widens the perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the setting a park with ginkgo and maple trees, a manicured lawn, and a pond in late afternoon light?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.031309038400650024 + ], + "opposite_logprob_tokens": [ + -3.531309127807617 + ], + "logprob_reward": -0.031309038400650024, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.031309038400650024, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.20223607122898102, + 0.12135309725999832, + 0.05894337221980095, + 0.0482814721763134, + 0.16694624722003937, + 0.07406396418809891, + 0.09592565149068832, + 0.050286971032619476, + 0.11380106210708618, + 0.06816206127405167 + ], + "credit_map": [ + 0.02310107834637165, + 0.023083951324224472, + 0.028279505670070648, + 0.032500311732292175, + 0.020187530666589737, + 0.022251002490520477, + 0.03646869584918022, + 0.016363998875021935, + 0.03661145642399788, + 0.016700774431228638, + 0.02969330921769142, + 0.0077136484906077385, + 0.007565483916550875, + 0.00724514527246356, + 0.009879635646939278, + 0.0059436471201479435, + 0.00759123545140028, + 0.006021300330758095, + 0.006931637413799763, + 0.007048868108540773, + 0.009738277643918991, + 0.003972165286540985, + 0.011284612119197845, + 0.006355280056595802, + 0.01124344952404499, + 0.006558982655405998, + 0.0056467317044734955, + 0.009861002676188946, + 0.003909280523657799, + 0.0029411064460873604, + 0.0037997851613909006, + 0.004321138374507427, + 0.012976564466953278, + 0.02002522349357605, + 0.015620219521224499, + 0.030882520601153374, + 0.021791553124785423, + 0.016300732269883156, + 0.03586001694202423, + 0.013489439152181149, + 0.008710501715540886, + 0.006543525028973818, + 0.00909650418907404, + 0.015519771724939346, + 0.009054461494088173, + 0.0076941982842981815, + 0.008888034150004387, + 0.008556966669857502, + 0.008679850026965141, + 0.011354636400938034, + 0.006797443609684706, + 0.00851735845208168, + 0.016815761104226112, + 0.015336219221353531, + 0.014627208933234215, + 0.01379717793315649, + 0.0034486460499465466, + 0.004742361605167389, + 0.00587252713739872, + 0.005941674578934908, + 0.007347578182816505, + 0.005372262094169855, + 0.01030385959893465, + 0.007258063647896051, + 0.006119467318058014, + 0.009210944175720215, + 0.010588062927126884, + 0.009239302948117256, + 0.017834456637501717, + 0.007477990817278624, + 0.037996578961610794, + 0.01533426158130169, + 0.005147705320268869, + 0.006133163347840309, + 0.005059005226939917, + 0.007129713427275419, + 0.009249509312212467, + 0.007886417210102081, + 0.013632867485284805, + 0.013923688791692257 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 1 + ], + "selected_reward": -0.031309038400650024 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an elderly man with East Asian features and silver hair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man wearing a beige cashmere cardigan over a light blue button-down shirt and dark gray trousers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the man walk along a weathered stone path with a deliberate, rhythmic pace?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera begin a slow, fluid pullback that widens the perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the setting a park with ginkgo and maple trees, a manicured lawn, and a pond in late afternoon light?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.012561016716063023 + ], + "opposite_logprob_tokens": [ + -4.387560844421387 + ], + "logprob_reward": -0.012561016716063023, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.012561016716063023, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1261809915304184, + 0.10871664434671402, + 0.02549075148999691, + 0.07829026132822037, + 0.0280062984675169, + 0.10093611478805542, + 0.09530995786190033, + 0.03867526352405548, + 0.12746384739875793, + 0.2709299325942993 + ], + "credit_map": [ + 0.01595577597618103, + 0.01630711741745472, + 0.017088521271944046, + 0.017812618985772133, + 0.01561802625656128, + 0.012743839994072914, + 0.016809841617941856, + 0.01384525466710329, + 0.027350254356861115, + 0.013972712680697441, + 0.02298998460173607, + 0.008042541332542896, + 0.009435796178877354, + 0.0061953868716955185, + 0.014284661039710045, + 0.0064453077502548695, + 0.003984234761446714, + 0.003265316365286708, + 0.0033279976341873407, + 0.0027599683962762356, + 0.003072221064940095, + 0.0018438635161146522, + 0.003334237029775977, + 0.0039029137697070837, + 0.018082238733768463, + 0.008429713547229767, + 0.011346861720085144, + 0.022410336881875992, + 0.003473074873909354, + 0.0035192284267395735, + 0.004550494719296694, + 0.006478318478912115, + 0.0026377972681075335, + 0.0030298507772386074, + 0.002903843531385064, + 0.005398912820965052, + 0.002755556022748351, + 0.0024369102902710438, + 0.006325079593807459, + 0.0025183476973325014, + 0.010473066940903664, + 0.009152816608548164, + 0.012380555272102356, + 0.016085276380181313, + 0.01813078671693802, + 0.010638484731316566, + 0.015771565958857536, + 0.008303559385240078, + 0.010464172810316086, + 0.010358283296227455, + 0.007925322279334068, + 0.008290798403322697, + 0.012752505019307137, + 0.017614053562283516, + 0.010510999709367752, + 0.017393818125128746, + 0.0029962656553834677, + 0.0033182217739522457, + 0.003876719856634736, + 0.005562606267631054, + 0.007642372045665979, + 0.0036668265238404274, + 0.007799957413226366, + 0.0038122956175357103, + 0.00812615267932415, + 0.01108790934085846, + 0.012511336244642735, + 0.015721473842859268, + 0.014695892110466957, + 0.011715881526470184, + 0.036590754985809326, + 0.017014434561133385, + 0.015254395082592964, + 0.020259367302060127, + 0.015483088791370392, + 0.032292891293764114, + 0.03382725268602371, + 0.03814389184117317, + 0.06753676384687424, + 0.0481322817504406 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 8, + 0 + ], + "selected_reward": -0.012561016716063023 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an elderly man with East Asian features and silver hair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man wearing a beige cashmere cardigan over a light blue button-down shirt and dark gray trousers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the man walk along a weathered stone path with a deliberate, rhythmic pace?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera begin a slow, fluid pullback that widens the perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the setting a park with ginkgo and maple trees, a manicured lawn, and a pond in late afternoon light?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.018181685358285904 + ], + "opposite_logprob_tokens": [ + -4.018181800842285 + ], + "logprob_reward": -0.018181685358285904, + "yes_no_margin_reward": 4.0, + "token_credit_reward": -0.018181685358285904, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.18898572027683258, + 0.33666738867759705, + 0.0736600011587143, + 0.10165797173976898, + 0.05884137377142906, + 0.06556718051433563, + 0.02669673040509224, + 0.043102409690618515, + 0.04473572596907616, + 0.060085441917181015 + ], + "credit_map": [ + 0.017086124047636986, + 0.023031210526823997, + 0.0226916316896677, + 0.03399553894996643, + 0.02057749405503273, + 0.019192393869161606, + 0.027217814698815346, + 0.025193532928824425, + 0.09192916005849838, + 0.0497540682554245, + 0.07300221174955368, + 0.02040555328130722, + 0.026381753385066986, + 0.0195002518594265, + 0.03451606631278992, + 0.02117832377552986, + 0.009886536747217178, + 0.006967546418309212, + 0.013929089531302452, + 0.009833596646785736, + 0.005965400021523237, + 0.004629347939044237, + 0.008222849108278751, + 0.014225630089640617, + 0.020941441878676414, + 0.012179029174149036, + 0.01824415661394596, + 0.027650417760014534, + 0.005484446417540312, + 0.0050955647602677345, + 0.0068045430816709995, + 0.005258375313133001, + 0.004752065986394882, + 0.007789922412484884, + 0.006555372383445501, + 0.01317848451435566, + 0.005882492754608393, + 0.007443935610353947, + 0.008960948325693607, + 0.004278157372027636, + 0.006772185210138559, + 0.00524241803213954, + 0.008479921147227287, + 0.011053626425564289, + 0.012517259456217289, + 0.00660160044208169, + 0.009426034986972809, + 0.005474136210978031, + 0.0036447811871767044, + 0.0037892353720963, + 0.003106617834419012, + 0.002976898569613695, + 0.004236150532960892, + 0.002755227265879512, + 0.003392336890101433, + 0.0027954853139817715, + 0.0028520997148007154, + 0.0038634049706161022, + 0.00672987150028348, + 0.006475236266851425, + 0.006861114874482155, + 0.0033800348173826933, + 0.007331924047321081, + 0.005608729552477598, + 0.002722168806940317, + 0.004631380550563335, + 0.00520606292411685, + 0.0037460802122950554, + 0.009282948449254036, + 0.0030279613565653563, + 0.011592199094593525, + 0.004526927135884762, + 0.006210165098309517, + 0.012161852791905403, + 0.004801077302545309, + 0.009916197508573532, + 0.006670278497040272, + 0.006494798697531223, + 0.00808437168598175, + 0.005746703594923019 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 3 + ], + "selected_reward": -0.018181685358285904 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an elderly man with East Asian features and silver hair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man wearing a beige cashmere cardigan over a light blue button-down shirt and dark gray trousers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the man walk along a weathered stone path with a deliberate, rhythmic pace?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera begin a slow, fluid pullback that widens the perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the setting a park with ginkgo and maple trees, a manicured lawn, and a pond in late afternoon light?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04863797500729561 + ], + "opposite_logprob_tokens": [ + -3.048637866973877 + ], + "logprob_reward": -0.04863797500729561, + "yes_no_margin_reward": 3.0, + "token_credit_reward": -0.04863797500729561, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1242198571562767, + 0.06613626331090927, + 0.03329959884285927, + 0.07853367179632187, + 0.05511860549449921, + 0.22662779688835144, + 0.07369339466094971, + 0.11311076581478119, + 0.1574353575706482, + 0.07182468473911285 + ], + "credit_map": [ + 0.022965656593441963, + 0.02114567905664444, + 0.017264019697904587, + 0.013233378529548645, + 0.012655310332775116, + 0.01058335229754448, + 0.016864590346813202, + 0.009507871232926846, + 0.016891570761799812, + 0.007602619472891092, + 0.01265579555183649, + 0.004586975090205669, + 0.005781186278909445, + 0.004715373273938894, + 0.008581657893955708, + 0.005321081727743149, + 0.004353970289230347, + 0.0034108608961105347, + 0.00482986168935895, + 0.0060841781087219715, + 0.0037948968820273876, + 0.0020217816345393658, + 0.004512100480496883, + 0.004291947465389967, + 0.019179536029696465, + 0.008243062533438206, + 0.006311827339231968, + 0.027492688968777657, + 0.004651885014027357, + 0.003582037752494216, + 0.00591833982616663, + 0.0031542975921183825, + 0.005275691393762827, + 0.0062991962768137455, + 0.006255212239921093, + 0.00669228658080101, + 0.006395950447767973, + 0.00913604162633419, + 0.010495711117982864, + 0.00456851813942194, + 0.02548862062394619, + 0.0194346085190773, + 0.030508805066347122, + 0.037208933383226395, + 0.02718569152057171, + 0.031086022034287453, + 0.030021678656339645, + 0.025693431496620178, + 0.00599229522049427, + 0.005284567829221487, + 0.006898419000208378, + 0.010127033106982708, + 0.010455718263983727, + 0.010736765339970589, + 0.014277773909270763, + 0.009920821525156498, + 0.007181166205555201, + 0.012238558381795883, + 0.018315326422452927, + 0.018314532935619354, + 0.0121360644698143, + 0.01280460599809885, + 0.017624519765377045, + 0.014496000483632088, + 0.009202308021485806, + 0.014889052137732506, + 0.011606798507273197, + 0.022465351969003677, + 0.02347812056541443, + 0.011617191135883331, + 0.04492348060011864, + 0.019253047183156013, + 0.006439451593905687, + 0.009437432512640953, + 0.006799447350203991, + 0.013100782409310341, + 0.008035772480070591, + 0.006683375220745802, + 0.011519307270646095, + 0.009809117764234543 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 8, + 0 + ], + "selected_reward": -0.04863797500729561 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an elderly man with East Asian features and silver hair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man wearing a beige cashmere cardigan over a light blue button-down shirt and dark gray trousers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the man walk along a weathered stone path with a deliberate, rhythmic pace?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera begin a slow, fluid pullback that widens the perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the setting a park with ginkgo and maple trees, a manicured lawn, and a pond in late afternoon light?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.014195065014064312 + ], + "opposite_logprob_tokens": [ + -4.264194965362549 + ], + "logprob_reward": -0.014195065014064312, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.014195065014064312, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08778738230466843, + 0.11614130437374115, + 0.05190449580550194, + 0.10557775944471359, + 0.05314049497246742, + 0.1328396499156952, + 0.07560966163873672, + 0.09898889064788818, + 0.2166413515806198, + 0.06136898696422577 + ], + "credit_map": [ + 0.00925107765942812, + 0.011169783771038055, + 0.011867680586874485, + 0.01411465648561716, + 0.013936043716967106, + 0.006065245717763901, + 0.012717465870082378, + 0.008665432222187519, + 0.03334325551986694, + 0.01483775395900011, + 0.02663491480052471, + 0.01128250453621149, + 0.004832693375647068, + 0.00816430989652872, + 0.007224606350064278, + 0.009821266867220402, + 0.007833818905055523, + 0.004948596004396677, + 0.006698428653180599, + 0.005230588838458061, + 0.004404708743095398, + 0.003762793028727174, + 0.00548239890486002, + 0.013543158769607544, + 0.02787684090435505, + 0.010262997820973396, + 0.012783904559910297, + 0.037484414875507355, + 0.0027524500619620085, + 0.004129654262214899, + 0.00451734708622098, + 0.005770151969045401, + 0.005196696612983942, + 0.007453949190676212, + 0.0061065517365932465, + 0.00936987902969122, + 0.004847531206905842, + 0.0063412850722670555, + 0.009437380358576775, + 0.004387218039482832, + 0.014516966417431831, + 0.014097404666244984, + 0.017644166946411133, + 0.013657784089446068, + 0.0273591335862875, + 0.015935510396957397, + 0.018807480111718178, + 0.010821206495165825, + 0.006655707024037838, + 0.008022421970963478, + 0.00963455718010664, + 0.011096764355897903, + 0.008895481005311012, + 0.008862141519784927, + 0.013376728631556034, + 0.009065857157111168, + 0.006308667827397585, + 0.009857848286628723, + 0.016713548451662064, + 0.020517537370324135, + 0.009588196873664856, + 0.010210023261606693, + 0.012394160963594913, + 0.013398908078670502, + 0.014561522752046585, + 0.02225954458117485, + 0.029511751607060432, + 0.04934060201048851, + 0.020842131227254868, + 0.019074171781539917, + 0.032720234245061874, + 0.02833140827715397, + 0.005106288939714432, + 0.007207563146948814, + 0.004489973187446594, + 0.008484995923936367, + 0.00684488145634532, + 0.00825598556548357, + 0.012429426424205303, + 0.008549871854484081 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 5, + 1 + ], + "selected_reward": -0.014195065014064312 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "medium shot, at eye level, initially depicts an elderly man with kindly East Asian features, neatly combed silver hair, and a face etched with the gentle wrinkles of a happy life. He wears a soft beige cashmere cardigan over a crisp light blue button-down shirt and dark gray trousers, his arms folded, his posture straight yet relaxed. He walks along a weathered stone path with a deliberate, rhythmic pace, his dark brown leather shoes making almost no sound on the scattered, dry autumn leaves. As he walks, the camera begins a slow, fluid pullback, steadily widening the perspective. This movement gradually reveals the tranquil, leafy setting of Serenity Park in late afternoon. Low, golden sunlight filters through the vibrant canopies of towering ginkgo and maple trees, casting long, soft shadows and making the crimson and amber leaves on the ground sparkle. The ever-expanding view reveals a vast, manicured lawn to one side and the soft glow of a pond beyond, creating an atmosphere of deep tranquility and contemplation.", + "seed": 4777, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000071/group_5/index_7328_seed_4777_dp_11_rank_22_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 5, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.01584737338125706, + "vlm_logprob": -0.01584737338125706, + "vlm_yes_no_margin": 4.3, + "vlm_token_credit": -0.01584737338125706, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an elderly man with East Asian features and silver hair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man wearing a beige cashmere cardigan over a light blue button-down shirt and dark gray trousers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the man walk along a weathered stone path with a deliberate, rhythmic pace?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera begin a slow, fluid pullback that widens the perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the setting a park with ginkgo and maple trees, a manicured lawn, and a pond in late afternoon light?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.009146456606686115 + ], + "opposite_logprob_tokens": [ + -4.884146690368652 + ], + "logprob_reward": -0.009146456606686115, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.009146456606686115, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09209803491830826, + 0.05965335667133331, + 0.08850082755088806, + 0.035304781049489975, + 0.060351334512233734, + 0.05772702395915985, + 0.061420418322086334, + 0.43781474232673645, + 0.03467120602726936, + 0.07245825231075287 + ], + "credit_map": [ + 0.010833844542503357, + 0.013136894442141056, + 0.010663102380931377, + 0.012604644522070885, + 0.009288999252021313, + 0.01584375463426113, + 0.009483219124376774, + 0.010243576020002365, + 0.005379608832299709, + 0.006811297032982111, + 0.008058544248342514, + 0.006685434840619564, + 0.013063114136457443, + 0.005948356352746487, + 0.00785838533192873, + 0.005848615430295467, + 0.007284040562808514, + 0.011444381438195705, + 0.009621109813451767, + 0.007789680268615484, + 0.01942255161702633, + 0.012173576280474663, + 0.011774509213864803, + 0.008990976959466934, + 0.0029483274556696415, + 0.005605138838291168, + 0.0038528258446604013, + 0.0035789988469332457, + 0.006496633868664503, + 0.00489633996039629, + 0.00445963628590107, + 0.003466880414634943, + 0.004181760363280773, + 0.004380187951028347, + 0.005763817112892866, + 0.017240814864635468, + 0.015695054084062576, + 0.004155599977821112, + 0.004524311050772667, + 0.00440979516133666, + 0.003951183054596186, + 0.004262632690370083, + 0.0021977522410452366, + 0.005946350749582052, + 0.011590455658733845, + 0.006128310225903988, + 0.007345144636929035, + 0.016305197030305862, + 0.004881854634732008, + 0.012104272842407227, + 0.002819242887198925, + 0.004028575494885445, + 0.009299607016146183, + 0.017679113894701004, + 0.0055550094693899155, + 0.005052739288657904, + 0.052766092121601105, + 0.02493389882147312, + 0.028042571619153023, + 0.022930676117539406, + 0.024464137852191925, + 0.1776936948299408, + 0.028800424188375473, + 0.07818320393562317, + 0.0032345184590667486, + 0.001848446438089013, + 0.0013097929768264294, + 0.0011726965894922614, + 0.001731536234728992, + 0.008736111223697662, + 0.0026133600622415543, + 0.014024743810296059, + 0.012720011174678802, + 0.006651369854807854, + 0.0036579719744622707, + 0.00188646768219769, + 0.004327770322561264, + 0.016768869012594223, + 0.005292287562042475, + 0.021153509616851807 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 0, + 2 + ], + "selected_reward": -0.009146456606686115 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an elderly man with East Asian features and silver hair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man wearing a beige cashmere cardigan over a light blue button-down shirt and dark gray trousers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the man walk along a weathered stone path with a deliberate, rhythmic pace?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera begin a slow, fluid pullback that widens the perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the setting a park with ginkgo and maple trees, a manicured lawn, and a pond in late afternoon light?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.011094076558947563 + ], + "opposite_logprob_tokens": [ + -4.511094093322754 + ], + "logprob_reward": -0.011094076558947563, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.011094076558947563, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08989062160253525, + 0.0400085523724556, + 0.042196981608867645, + 0.0324721597135067, + 0.08817234635353088, + 0.08697355538606644, + 0.049687519669532776, + 0.3891550898551941, + 0.08445217460393906, + 0.09699095785617828 + ], + "credit_map": [ + 0.010026397183537483, + 0.01176474429666996, + 0.01593136042356491, + 0.015578270889818668, + 0.00996406003832817, + 0.0121924364939332, + 0.007708837278187275, + 0.006724521052092314, + 0.003963522147387266, + 0.00456279469653964, + 0.0053144218400120735, + 0.00621041702106595, + 0.005229541566222906, + 0.0051369452849030495, + 0.004444019868969917, + 0.0051468899473547935, + 0.0039651114493608475, + 0.005204689223319292, + 0.005756565369665623, + 0.0043963054195046425, + 0.007119541056454182, + 0.004638196900486946, + 0.00618756515905261, + 0.004929007962346077, + 0.0039877379313111305, + 0.003901500254869461, + 0.002495130989700556, + 0.005572878755629063, + 0.004325853660702705, + 0.003138536587357521, + 0.002858181018382311, + 0.006192340049892664, + 0.005505907814949751, + 0.006143711507320404, + 0.014800062403082848, + 0.03505251556634903, + 0.008278869092464447, + 0.008135034702718258, + 0.004885678179562092, + 0.005370567087084055, + 0.0056496779434382915, + 0.00806550495326519, + 0.0036800268571823835, + 0.009713192470371723, + 0.01535369548946619, + 0.00998132023960352, + 0.011275563389062881, + 0.023254575207829475, + 0.008687742054462433, + 0.010204602032899857, + 0.0029223584569990635, + 0.0025802035816013813, + 0.003563077887520194, + 0.013296311721205711, + 0.003994956612586975, + 0.004438268486410379, + 0.035775572061538696, + 0.016636639833450317, + 0.05662995204329491, + 0.0244308952242136, + 0.020317398011684418, + 0.14852026104927063, + 0.023952528834342957, + 0.0628918707370758, + 0.008062037639319897, + 0.004378841258585453, + 0.006653485354036093, + 0.0036696370225399733, + 0.0033734613098204136, + 0.020507484674453735, + 0.006148016080260277, + 0.031659211963415146, + 0.018630363047122955, + 0.005700608249753714, + 0.004269816912710667, + 0.002710925415158272, + 0.004645457491278648, + 0.021501094102859497, + 0.004984024912118912, + 0.03454865887761116 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 9, + 0 + ], + "selected_reward": -0.011094076558947563 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an elderly man with East Asian features and silver hair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man wearing a beige cashmere cardigan over a light blue button-down shirt and dark gray trousers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the man walk along a weathered stone path with a deliberate, rhythmic pace?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera begin a slow, fluid pullback that widens the perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the setting a park with ginkgo and maple trees, a manicured lawn, and a pond in late afternoon light?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.020575417205691338 + ], + "opposite_logprob_tokens": [ + -3.895575523376465 + ], + "logprob_reward": -0.020575417205691338, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.020575417205691338, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.05067746341228485, + 0.06310755759477615, + 0.06328493356704712, + 0.04574473202228546, + 0.13784673810005188, + 0.06311184912919998, + 0.037361957132816315, + 0.29145172238349915, + 0.1874963641166687, + 0.05991670489311218 + ], + "credit_map": [ + 0.005575002171099186, + 0.009850190952420235, + 0.009366276673972607, + 0.00888764951378107, + 0.0038438146002590656, + 0.005980027839541435, + 0.003772203577682376, + 0.0034022917971014977, + 0.007337766233831644, + 0.007295851595699787, + 0.013131261803209782, + 0.01220180094242096, + 0.006195031106472015, + 0.00597662478685379, + 0.004134679213166237, + 0.006834534928202629, + 0.005245603621006012, + 0.009571476839482784, + 0.01189748477190733, + 0.008806268684566021, + 0.006874072831124067, + 0.007817637175321579, + 0.006487448234111071, + 0.006584938149899244, + 0.005505579989403486, + 0.006674892734736204, + 0.008662644773721695, + 0.0053100548684597015, + 0.004593833815306425, + 0.007217471953481436, + 0.0036662486381828785, + 0.004114003852009773, + 0.01150534488260746, + 0.011144991032779217, + 0.023219715803861618, + 0.04769532009959221, + 0.013487845659255981, + 0.013136287219822407, + 0.008812510408461094, + 0.008844698779284954, + 0.006854887120425701, + 0.007990588434040546, + 0.003212531330063939, + 0.006702382583171129, + 0.008309205062687397, + 0.00602675462141633, + 0.013239448890089989, + 0.010776051320135593, + 0.0035860829520970583, + 0.004808054771274328, + 0.0031250768806785345, + 0.0026206844486296177, + 0.008431830443441868, + 0.004431150853633881, + 0.006187465041875839, + 0.004171608481556177, + 0.02656111866235733, + 0.012713762000203133, + 0.045418623834848404, + 0.019366709515452385, + 0.015390572138130665, + 0.10835084319114685, + 0.01773475855588913, + 0.04591531306505203, + 0.017321035265922546, + 0.00908080767840147, + 0.005362310446798801, + 0.004630569368600845, + 0.00838802196085453, + 0.05167786031961441, + 0.009303989820182323, + 0.08173175156116486, + 0.009898279793560505, + 0.004062718246132135, + 0.003274414921179414, + 0.0020140367560088634, + 0.003305535763502121, + 0.013840512372553349, + 0.0037686077412217855, + 0.019752593711018562 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 8, + 4 + ], + "selected_reward": -0.020575417205691338 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an elderly man with East Asian features and silver hair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man wearing a beige cashmere cardigan over a light blue button-down shirt and dark gray trousers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the man walk along a weathered stone path with a deliberate, rhythmic pace?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera begin a slow, fluid pullback that widens the perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the setting a park with ginkgo and maple trees, a manicured lawn, and a pond in late afternoon light?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02978559397161007 + ], + "opposite_logprob_tokens": [ + -3.529785633087158 + ], + "logprob_reward": -0.02978559397161007, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.02978559397161007, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11487783491611481, + 0.10534823685884476, + 0.06595133990049362, + 0.05407707020640373, + 0.11843094229698181, + 0.17822037637233734, + 0.039135564118623734, + 0.1011262834072113, + 0.10765878856182098, + 0.11517353355884552 + ], + "credit_map": [ + 0.01684775948524475, + 0.020434416830539703, + 0.02188720367848873, + 0.016672654077410698, + 0.006437722593545914, + 0.015404350124299526, + 0.0070993308909237385, + 0.01009438931941986, + 0.009794400073587894, + 0.011183350346982479, + 0.01808510534465313, + 0.02676529996097088, + 0.007656205911189318, + 0.013683885335922241, + 0.006163246463984251, + 0.01201674435287714, + 0.005798839498311281, + 0.009379725903272629, + 0.009275381453335285, + 0.009837774559855461, + 0.007674312684684992, + 0.008788478560745716, + 0.005578696262091398, + 0.009618131443858147, + 0.008524701930582523, + 0.008874577470123768, + 0.006780894007533789, + 0.006837469059973955, + 0.0064390795305371284, + 0.005966901313513517, + 0.005440028849989176, + 0.005213416181504726, + 0.00823044404387474, + 0.010355615988373756, + 0.01888948492705822, + 0.0451187938451767, + 0.005640837829560041, + 0.011759920045733452, + 0.008332952857017517, + 0.010102889500558376, + 0.015278209932148457, + 0.01617446541786194, + 0.008373173885047436, + 0.018003007397055626, + 0.030871406197547913, + 0.01898513361811638, + 0.024367833510041237, + 0.04616713896393776, + 0.0035079678054898977, + 0.005892913322895765, + 0.002749491250142455, + 0.0026969527825713158, + 0.008681191131472588, + 0.0065592085011303425, + 0.005484545603394508, + 0.003563291160389781, + 0.009336430579423904, + 0.004220848437398672, + 0.012933658435940742, + 0.005698392633348703, + 0.006090966518968344, + 0.03832827880978584, + 0.008025988936424255, + 0.01649172604084015, + 0.00978763122111559, + 0.006964983884245157, + 0.003629142651334405, + 0.003418760374188423, + 0.004928073845803738, + 0.028057657182216644, + 0.005034990608692169, + 0.04583754390478134, + 0.020287619903683662, + 0.0064400313422083855, + 0.004706071224063635, + 0.003355748485773802, + 0.006137277465313673, + 0.028282122686505318, + 0.005968871060758829, + 0.03999580070376396 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 4, + 9 + ], + "selected_reward": -0.02978559397161007 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an elderly man with East Asian features and silver hair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man wearing a beige cashmere cardigan over a light blue button-down shirt and dark gray trousers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the man walk along a weathered stone path with a deliberate, rhythmic pace?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera begin a slow, fluid pullback that widens the perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the setting a park with ginkgo and maple trees, a manicured lawn, and a pond in late afternoon light?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0086353225633502 + ], + "opposite_logprob_tokens": [ + -4.758635520935059 + ], + "logprob_reward": -0.0086353225633502, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.0086353225633502, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11854714155197144, + 0.08508013933897018, + 0.024802006781101227, + 0.02526651695370674, + 0.09684799611568451, + 0.07444730401039124, + 0.036055613309144974, + 0.4661765694618225, + 0.022547978907823563, + 0.050228774547576904 + ], + "credit_map": [ + 0.01494850404560566, + 0.025252029299736023, + 0.02068030647933483, + 0.019105246290564537, + 0.0072227902710437775, + 0.013617459684610367, + 0.00805681012570858, + 0.009663984179496765, + 0.008821483701467514, + 0.010161131620407104, + 0.01546600740402937, + 0.023075321689248085, + 0.004275127779692411, + 0.010456325486302376, + 0.003984824288636446, + 0.008839909918606281, + 0.0018159234896302223, + 0.003556738141924143, + 0.0034747100435197353, + 0.004068720620125532, + 0.002799754496663809, + 0.003984234295785427, + 0.002049082890152931, + 0.0030528404749929905, + 0.003085358999669552, + 0.003842779202386737, + 0.002403759164735675, + 0.004420246928930283, + 0.002981250872835517, + 0.003498408244922757, + 0.00194528978317976, + 0.003089422360062599, + 0.0075722043402493, + 0.006526889745146036, + 0.017003098502755165, + 0.025836355984210968, + 0.016530053690075874, + 0.010679093189537525, + 0.006340216379612684, + 0.006360076833516359, + 0.004809345118701458, + 0.007080476265400648, + 0.0031373349484056234, + 0.008225617930293083, + 0.014469154179096222, + 0.008418850600719452, + 0.009071591310203075, + 0.019234927371144295, + 0.0037724741268903017, + 0.0032877575140446424, + 0.002468937309458852, + 0.0022712936624884605, + 0.004988973960280418, + 0.009344608522951603, + 0.0056325760670006275, + 0.004288994707167149, + 0.04728331416845322, + 0.021480737254023552, + 0.053036462515592575, + 0.02542705275118351, + 0.025402681902050972, + 0.18484964966773987, + 0.029506465420126915, + 0.07919018715620041, + 0.0018190107075497508, + 0.002708936110138893, + 0.0038924370892345905, + 0.0018293720204383135, + 0.0013670242624357343, + 0.003739170031622052, + 0.001926792785525322, + 0.005265234969556332, + 0.009980512782931328, + 0.0043636951595544815, + 0.002915291115641594, + 0.0014058093074709177, + 0.003193995449692011, + 0.01033447403460741, + 0.0019656463991850615, + 0.01606934331357479 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 0, + 4 + ], + "selected_reward": -0.0086353225633502 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000071/group_6/index_tensor([8653, 8653])_group_6_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000071/group_6/index_tensor([8653, 8653])_group_6_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..77f3d86a01bd3bf2d7071f22617f872a1e81906b --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000071/group_6/index_tensor([8653, 8653])_group_6_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "An eye-level, static, medium shot captures a man in his 30s, with neatly styled dark brown hair and light stubble, sitting at an outdoor table. He wears a dark blue merino wool sweater over a crisp white collared shirt. Initially appearing slightly off-camera, he smoothly turns his head to face the camera, his expression transforming into a friendly, engaging expression as he begins to speak. The setting is a sun-drenched modern square, where a multi-tiered stone fountain provides a tranquil backdrop; water cascades down its weathered surface, and the pool at the bottom creates a canvas of gently shimmering ripples in the bright daylight. As he speaks, the man reaches for a closed, ultra-thin laptop with a silver aluminum finish on the table. Placing both hands on it, he opens the lid with a steady, deliberate movement. The cool light from the screen subtly illuminates his focused features and creates a soft reflection in his eyes. The entire scene is imbued with a calm, professional atmosphere, the man's focused movements harmoniously balanced with the fountain's steady, tranquil flow.", + "seed": 10895, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000071/group_6/index_8653_seed_10895_dp_12_rank_24_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 6, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.05941354762762785, + "vlm_logprob": -0.05941354762762785, + "vlm_yes_no_margin": 3.5, + "vlm_token_credit": -0.05941354762762785, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot framed as an eye-level, static, medium shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man wearing a dark blue merino wool sweater over a white collared shirt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the man turn his head from off-camera to face the camera while opening a silver aluminum laptop?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a sun-drenched modern square featuring a multi-tiered stone fountain with cascading water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the cool light from the laptop screen illuminate the man's face and create a reflection in his eyes?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.09297563880681992 + ], + "opposite_logprob_tokens": [ + -3.842975616455078 + ], + "logprob_reward": -0.09297563880681992, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.09297563880681992, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.22646693885326385, + 0.1076788455247879, + 0.09411440044641495, + 0.058435894548892975, + 0.09866322576999664, + 0.06790798157453537, + 0.06129821017384529, + 0.07293695211410522, + 0.14609737694263458, + 0.0664001852273941 + ], + "credit_map": [ + 0.026058513671159744, + 0.02979109063744545, + 0.02603220008313656, + 0.022318631410598755, + 0.01827777549624443, + 0.0319410115480423, + 0.025043552741408348, + 0.047004155814647675, + 0.011611711233854294, + 0.010253770276904106, + 0.010854979045689106, + 0.010357174091041088, + 0.008032863028347492, + 0.024031564593315125, + 0.01141448225826025, + 0.02112230844795704, + 0.01610695756971836, + 0.010911790654063225, + 0.011180139146745205, + 0.009842434898018837, + 0.012759274803102016, + 0.009600560180842876, + 0.011714253574609756, + 0.011998984962701797, + 0.006978357210755348, + 0.010195702314376831, + 0.006141501013189554, + 0.00447230227291584, + 0.006899585947394371, + 0.007894021458923817, + 0.006198588293045759, + 0.009655836969614029, + 0.019949115812778473, + 0.010289977304637432, + 0.009311363101005554, + 0.007080578710883856, + 0.01028788834810257, + 0.01299424096941948, + 0.013836483471095562, + 0.014913582243025303, + 0.01321906503289938, + 0.006492553744465113, + 0.007259577978402376, + 0.005259898491203785, + 0.005768005270510912, + 0.010465274564921856, + 0.009103856980800629, + 0.01033974438905716, + 0.007997849024832249, + 0.004939741920679808, + 0.006337450351566076, + 0.007491293363273144, + 0.004285623785108328, + 0.005581727251410484, + 0.015325246378779411, + 0.009339280426502228, + 0.007810044568032026, + 0.00802856869995594, + 0.012003820389509201, + 0.00942764338105917, + 0.006478963419795036, + 0.007722170557826757, + 0.010150579735636711, + 0.01131515670567751, + 0.015143701806664467, + 0.010685804300010204, + 0.019371401518583298, + 0.011264670640230179, + 0.01227374654263258, + 0.023551849648356438, + 0.019349446520209312, + 0.034456755965948105, + 0.009020364843308926, + 0.006898561492562294, + 0.00782752688974142, + 0.004953836556524038, + 0.008209213614463806, + 0.009518132545053959, + 0.012007100507616997, + 0.00796544924378395 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 1 + ], + "selected_reward": -0.09297563880681992 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot framed as an eye-level, static, medium shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man wearing a dark blue merino wool sweater over a white collared shirt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the man turn his head from off-camera to face the camera while opening a silver aluminum laptop?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a sun-drenched modern square featuring a multi-tiered stone fountain with cascading water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the cool light from the laptop screen illuminate the man's face and create a reflection in his eyes?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04876968264579773 + ], + "opposite_logprob_tokens": [ + -3.04876971244812 + ], + "logprob_reward": -0.04876968264579773, + "yes_no_margin_reward": 3.0, + "token_credit_reward": -0.04876968264579773, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2766181528568268, + 0.12211309373378754, + 0.12634903192520142, + 0.03100038319826126, + 0.09386876970529556, + 0.06842619925737381, + 0.050152551382780075, + 0.10771778225898743, + 0.09162825345993042, + 0.03212578594684601 + ], + "credit_map": [ + 0.022122368216514587, + 0.03384937345981598, + 0.036919232457876205, + 0.06055661290884018, + 0.022954294458031654, + 0.028953831642866135, + 0.03255937248468399, + 0.0387030653655529, + 0.01296214573085308, + 0.015992620959877968, + 0.017774445936083794, + 0.025232266634702682, + 0.013663345947861671, + 0.011770126409828663, + 0.01220650877803564, + 0.012511623091995716, + 0.01316151488572359, + 0.01803862862288952, + 0.015650611370801926, + 0.028889266774058342, + 0.013949488289654255, + 0.010895494371652603, + 0.012786726467311382, + 0.012977302074432373, + 0.0037269964814186096, + 0.005271571222692728, + 0.0037156292237341404, + 0.0033937785774469376, + 0.004176102578639984, + 0.003671142039820552, + 0.0030972796957939863, + 0.003947883378714323, + 0.007574751973152161, + 0.00750776706263423, + 0.01057550311088562, + 0.00937317218631506, + 0.007455955725163221, + 0.014949615113437176, + 0.011596819385886192, + 0.024835187941789627, + 0.008707270957529545, + 0.006278068758547306, + 0.00771564245223999, + 0.0038931630551815033, + 0.006634143181145191, + 0.009661934338510036, + 0.009284337982535362, + 0.016251640394330025, + 0.004282006993889809, + 0.004520394839346409, + 0.0058242096565663815, + 0.002875781850889325, + 0.00338649470359087, + 0.007536801975220442, + 0.009718856774270535, + 0.012008002027869225, + 0.014107346534729004, + 0.013070197775959969, + 0.012103411369025707, + 0.00978170894086361, + 0.013067559339106083, + 0.011530710384249687, + 0.01901363581418991, + 0.015043209306895733, + 0.007676783017814159, + 0.005740403197705746, + 0.014945692382752895, + 0.005820997059345245, + 0.0065011149272322655, + 0.013725507073104382, + 0.013395177200436592, + 0.02382258139550686, + 0.0031790612265467644, + 0.0022614605259150267, + 0.004666728433221579, + 0.002880815649405122, + 0.002979598706588149, + 0.003736684564501047, + 0.006471965927630663, + 0.005949471145868301 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 1 + ], + "selected_reward": -0.04876968264579773 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot framed as an eye-level, static, medium shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man wearing a dark blue merino wool sweater over a white collared shirt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the man turn his head from off-camera to face the camera while opening a silver aluminum laptop?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a sun-drenched modern square featuring a multi-tiered stone fountain with cascading water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the cool light from the laptop screen illuminate the man's face and create a reflection in his eyes?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.12709130346775055 + ], + "opposite_logprob_tokens": [ + -2.127091407775879 + ], + "logprob_reward": -0.12709130346775055, + "yes_no_margin_reward": 2.0, + "token_credit_reward": -0.12709130346775055, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11405642330646515, + 0.07245232164859772, + 0.07400829344987869, + 0.11241653561592102, + 0.1485600620508194, + 0.0856274738907814, + 0.06841053068637848, + 0.09783700108528137, + 0.17750413715839386, + 0.049127280712127686 + ], + "credit_map": [ + 0.011521321721374989, + 0.014711204916238785, + 0.013080690056085587, + 0.01884656772017479, + 0.012709446251392365, + 0.010006214492022991, + 0.013942808844149113, + 0.019238173961639404, + 0.011311034671962261, + 0.008856645785272121, + 0.008240452967584133, + 0.006259332410991192, + 0.007321234792470932, + 0.010478689335286617, + 0.007498280145227909, + 0.012486650608479977, + 0.008598600514233112, + 0.008061639033257961, + 0.006777926813811064, + 0.008733561262488365, + 0.011418611742556095, + 0.011644495651125908, + 0.008222595788538456, + 0.010550855658948421, + 0.008311467245221138, + 0.016900580376386642, + 0.00996699184179306, + 0.01106309611350298, + 0.013456173241138458, + 0.019760282710194588, + 0.010981260798871517, + 0.021976688876748085, + 0.013763842172920704, + 0.018641283735632896, + 0.016333987936377525, + 0.019681477919220924, + 0.011749276891350746, + 0.020612144842743874, + 0.01859492063522339, + 0.029183126986026764, + 0.008354315534234047, + 0.008884119801223278, + 0.009007810615003109, + 0.004291819408535957, + 0.006631974130868912, + 0.014598151668906212, + 0.01015289407223463, + 0.02370638959109783, + 0.00617180485278368, + 0.005298165138810873, + 0.007761161774396896, + 0.005145634990185499, + 0.005049142986536026, + 0.009576420299708843, + 0.013473297469317913, + 0.01593489572405815, + 0.008633000776171684, + 0.00992147158831358, + 0.014908527955412865, + 0.005995367653667927, + 0.010048838332295418, + 0.012271077372133732, + 0.02495165541768074, + 0.01110706478357315, + 0.015444478020071983, + 0.012828740291297436, + 0.025249402970075607, + 0.013046729378402233, + 0.012031491845846176, + 0.024942541494965553, + 0.029951803386211395, + 0.04400894045829773, + 0.004738427232950926, + 0.003756403224542737, + 0.007456272374838591, + 0.00481870211660862, + 0.004015299491584301, + 0.007670113816857338, + 0.007499407045543194, + 0.009172660298645496 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 4, + 0 + ], + "selected_reward": -0.12709130346775055 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot framed as an eye-level, static, medium shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man wearing a dark blue merino wool sweater over a white collared shirt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the man turn his head from off-camera to face the camera while opening a silver aluminum laptop?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a sun-drenched modern square featuring a multi-tiered stone fountain with cascading water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the cool light from the laptop screen illuminate the man's face and create a reflection in his eyes?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0076382951810956 + ], + "opposite_logprob_tokens": [ + -4.882638454437256 + ], + "logprob_reward": -0.0076382951810956, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.0076382951810956, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3816829323768616, + 0.08429118245840073, + 0.11676495522260666, + 0.05538615584373474, + 0.07499603182077408, + 0.08556433022022247, + 0.0336868055164814, + 0.07977251708507538, + 0.05333871394395828, + 0.03451637551188469 + ], + "credit_map": [ + 0.04611852020025253, + 0.034737132489681244, + 0.04234743490815163, + 0.026462292298674583, + 0.03510349616408348, + 0.06637270003557205, + 0.04701206833124161, + 0.0835292711853981, + 0.008467619307339191, + 0.008997934870421886, + 0.008728807792067528, + 0.004829269368201494, + 0.00889680813997984, + 0.0149907898157835, + 0.00940623227506876, + 0.019973725080490112, + 0.011579446494579315, + 0.009084131568670273, + 0.010730516165494919, + 0.00766380550339818, + 0.011069556698203087, + 0.022757334634661674, + 0.015208860859274864, + 0.028671303763985634, + 0.005330884363502264, + 0.006101912818849087, + 0.005331935826689005, + 0.0055307140573859215, + 0.009296688251197338, + 0.008605103008449078, + 0.005708679091185331, + 0.009480242617428303, + 0.00777486152946949, + 0.007653263863176107, + 0.007852318696677685, + 0.007245176006108522, + 0.005971837788820267, + 0.011081993579864502, + 0.008624152280390263, + 0.01879243366420269, + 0.009904537349939346, + 0.006173364352434874, + 0.009170438162982464, + 0.0039409310556948185, + 0.0066662319004535675, + 0.017795534804463387, + 0.01128776092082262, + 0.020625539124011993, + 0.004011738114058971, + 0.0026206839829683304, + 0.003273193957284093, + 0.002058894606307149, + 0.0028063280042260885, + 0.005371288862079382, + 0.006104167550802231, + 0.007440513465553522, + 0.0068456572480499744, + 0.008020468056201935, + 0.012047884985804558, + 0.008120317943394184, + 0.007513887714594603, + 0.011236836202442646, + 0.014136983081698418, + 0.011850487440824509, + 0.006764231249690056, + 0.005689783953130245, + 0.007700328249484301, + 0.0038690022192895412, + 0.004131433088332415, + 0.010501147247850895, + 0.006179040763527155, + 0.008503747172653675, + 0.004054541699588299, + 0.0025507407262921333, + 0.00501948082819581, + 0.0031351984944194555, + 0.003778720973059535, + 0.004028366878628731, + 0.006415015086531639, + 0.005534315947443247 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 5 + ], + "selected_reward": -0.0076382951810956 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot framed as an eye-level, static, medium shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man wearing a dark blue merino wool sweater over a white collared shirt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the man turn his head from off-camera to face the camera while opening a silver aluminum laptop?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a sun-drenched modern square featuring a multi-tiered stone fountain with cascading water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the cool light from the laptop screen illuminate the man's face and create a reflection in his eyes?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.020592818036675453 + ], + "opposite_logprob_tokens": [ + -3.8955929279327393 + ], + "logprob_reward": -0.020592818036675453, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.020592818036675453, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.17477712035179138, + 0.046703387051820755, + 0.04218447953462601, + 0.034933511167764664, + 0.07335502654314041, + 0.09946132451295853, + 0.04556792974472046, + 0.11710219830274582, + 0.21388624608516693, + 0.15202873945236206 + ], + "credit_map": [ + 0.021724846214056015, + 0.019210949540138245, + 0.0189394298940897, + 0.027374880388379097, + 0.01641947589814663, + 0.02057385817170143, + 0.02072741650044918, + 0.02980627492070198, + 0.005213139578700066, + 0.0055429148487746716, + 0.005769359413534403, + 0.0051559144631028175, + 0.005145695526152849, + 0.007165359798818827, + 0.005374512635171413, + 0.007336494978517294, + 0.003966734744608402, + 0.004068541806191206, + 0.0043862368911504745, + 0.004956235643476248, + 0.0032886872068047523, + 0.006503895856440067, + 0.004865216091275215, + 0.010148933157324791, + 0.0036395431961864233, + 0.0065732309594750404, + 0.0032925636041909456, + 0.003956764470785856, + 0.0036729543935507536, + 0.004836829379200935, + 0.003452543169260025, + 0.005509085021913052, + 0.006081010214984417, + 0.008859519846737385, + 0.007383761927485466, + 0.008102131076157093, + 0.007668854668736458, + 0.011845010332763195, + 0.008701840415596962, + 0.014712907373905182, + 0.011658329516649246, + 0.00930311530828476, + 0.010561368428170681, + 0.006262727081775665, + 0.00901602953672409, + 0.016333796083927155, + 0.011077976785600185, + 0.02524799294769764, + 0.002919042482972145, + 0.004913740791380405, + 0.005390248727053404, + 0.0024327870924025774, + 0.003777969628572464, + 0.008947093971073627, + 0.006024342030286789, + 0.01116271037608385, + 0.017923247069120407, + 0.012437509372830391, + 0.01810021884739399, + 0.00938478484749794, + 0.010358069092035294, + 0.015793506056070328, + 0.014267072081565857, + 0.018837803974747658, + 0.019353067502379417, + 0.014085276983678341, + 0.028373422101140022, + 0.01610703580081463, + 0.015895219519734383, + 0.032699450850486755, + 0.03369854763150215, + 0.05367424339056015, + 0.012348729185760021, + 0.016496289521455765, + 0.020373757928609848, + 0.0219266377389431, + 0.020996278151869774, + 0.014416792429983616, + 0.023781808093190193, + 0.02168845757842064 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 0, + 9 + ], + "selected_reward": -0.020592818036675453 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "An eye-level, static, medium shot captures a man in his 30s, with neatly styled dark brown hair and light stubble, sitting at an outdoor table. He wears a dark blue merino wool sweater over a crisp white collared shirt. Initially appearing slightly off-camera, he smoothly turns his head to face the camera, his expression transforming into a friendly, engaging expression as he begins to speak. The setting is a sun-drenched modern square, where a multi-tiered stone fountain provides a tranquil backdrop; water cascades down its weathered surface, and the pool at the bottom creates a canvas of gently shimmering ripples in the bright daylight. As he speaks, the man reaches for a closed, ultra-thin laptop with a silver aluminum finish on the table. Placing both hands on it, he opens the lid with a steady, deliberate movement. The cool light from the screen subtly illuminates his focused features and creates a soft reflection in his eyes. The entire scene is imbued with a calm, professional atmosphere, the man's focused movements harmoniously balanced with the fountain's steady, tranquil flow.", + "seed": 10896, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000071/group_6/index_8653_seed_10896_dp_12_rank_24_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 6, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.06740649193525314, + "vlm_logprob": -0.06740649193525314, + "vlm_yes_no_margin": 3.149999976158142, + "vlm_token_credit": -0.06740649193525314, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot framed as an eye-level, static, medium shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man wearing a dark blue merino wool sweater over a white collared shirt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the man turn his head from off-camera to face the camera while opening a silver aluminum laptop?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a sun-drenched modern square featuring a multi-tiered stone fountain with cascading water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the cool light from the laptop screen illuminate the man's face and create a reflection in his eyes?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0911700651049614 + ], + "opposite_logprob_tokens": [ + -3.591170072555542 + ], + "logprob_reward": -0.0911700651049614, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.0911700651049614, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.17544786632061005, + 0.08112466335296631, + 0.07468295097351074, + 0.10523004829883575, + 0.0617796964943409, + 0.03579552099108696, + 0.0870031788945198, + 0.16146431863307953, + 0.07650494575500488, + 0.14096686244010925 + ], + "credit_map": [ + 0.01790003478527069, + 0.030171744525432587, + 0.013734865933656693, + 0.019999852403998375, + 0.023046664893627167, + 0.01982070319354534, + 0.031986650079488754, + 0.018787333741784096, + 0.006873837672173977, + 0.00787685438990593, + 0.009112206287682056, + 0.006932717282325029, + 0.014930187724530697, + 0.00750893447548151, + 0.01994224824011326, + 0.007947671227157116, + 0.006681967992335558, + 0.014080447144806385, + 0.00788409449160099, + 0.010304117575287819, + 0.013248148374259472, + 0.00762207480147481, + 0.007262059021741152, + 0.007600036915391684, + 0.009400766342878342, + 0.02431231178343296, + 0.00998293049633503, + 0.013575255870819092, + 0.013076932169497013, + 0.012924326583743095, + 0.012074033729732037, + 0.009883481077849865, + 0.006749222986400127, + 0.010365315712988377, + 0.006872406229376793, + 0.008375539444386959, + 0.010468403808772564, + 0.004986998159438372, + 0.007926447317004204, + 0.00603535445407033, + 0.004891265649348497, + 0.007189848460257053, + 0.0032080733217298985, + 0.005019183270633221, + 0.005097232758998871, + 0.004063400439918041, + 0.0035845679230988026, + 0.002741945907473564, + 0.01098556537181139, + 0.02082383818924427, + 0.007419339846819639, + 0.007267439737915993, + 0.010601479560136795, + 0.01028643362224102, + 0.011569195427000523, + 0.008049879223108292, + 0.012195931747555733, + 0.02720450609922409, + 0.020091498270630836, + 0.04779985547065735, + 0.012524479068815708, + 0.012316401116549969, + 0.015037034638226032, + 0.014294592663645744, + 0.008148934692144394, + 0.011379462666809559, + 0.009073605760931969, + 0.01051492802798748, + 0.00887205172330141, + 0.00771403918042779, + 0.014476951211690903, + 0.006324968300759792, + 0.013087017461657524, + 0.01952308416366577, + 0.013984641060233116, + 0.020980950444936752, + 0.014454318210482597, + 0.016133273020386696, + 0.02700023725628853, + 0.015803324058651924 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 7, + 9 + ], + "selected_reward": -0.0911700651049614 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot framed as an eye-level, static, medium shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man wearing a dark blue merino wool sweater over a white collared shirt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the man turn his head from off-camera to face the camera while opening a silver aluminum laptop?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a sun-drenched modern square featuring a multi-tiered stone fountain with cascading water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the cool light from the laptop screen illuminate the man's face and create a reflection in his eyes?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.055026959627866745 + ], + "opposite_logprob_tokens": [ + -2.9300270080566406 + ], + "logprob_reward": -0.055026959627866745, + "yes_no_margin_reward": 2.875, + "token_credit_reward": -0.055026959627866745, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.20720338821411133, + 0.22933080792427063, + 0.041703421622514725, + 0.06141481176018715, + 0.07963716238737106, + 0.057563889771699905, + 0.04472317546606064, + 0.07857750356197357, + 0.0843333825469017, + 0.11551240086555481 + ], + "credit_map": [ + 0.018875595182180405, + 0.03750242665410042, + 0.016747452318668365, + 0.02764303795993328, + 0.029087303206324577, + 0.021473176777362823, + 0.03436896950006485, + 0.02150544337928295, + 0.019952408969402313, + 0.03616805747151375, + 0.020118677988648415, + 0.020234087482094765, + 0.0483194924890995, + 0.016028601676225662, + 0.0465664304792881, + 0.02194308675825596, + 0.0038474025204777718, + 0.005328415427356958, + 0.005110546480864286, + 0.004826326854526997, + 0.010427714325487614, + 0.003480485873296857, + 0.004182266537100077, + 0.004500268027186394, + 0.004795977380126715, + 0.011260007508099079, + 0.006029988639056683, + 0.0050778076983988285, + 0.008575097657740116, + 0.005544107872992754, + 0.013668539933860302, + 0.0064632948487997055, + 0.005930774379521608, + 0.013154465705156326, + 0.008320053108036518, + 0.010843019932508469, + 0.010829747654497623, + 0.006242765113711357, + 0.015677541494369507, + 0.008638804778456688, + 0.005589481443166733, + 0.006660394836217165, + 0.005782003980129957, + 0.006590874399989843, + 0.007707853335887194, + 0.004153437912464142, + 0.014370150864124298, + 0.006709698587656021, + 0.004597747232764959, + 0.00941446702927351, + 0.003930376842617989, + 0.004708039108663797, + 0.005490828305482864, + 0.004584782756865025, + 0.007623506709933281, + 0.004373433534055948, + 0.005870728753507137, + 0.016343258321285248, + 0.009470268152654171, + 0.02049620822072029, + 0.005099343601614237, + 0.006676063407212496, + 0.00685665151104331, + 0.007764989044517279, + 0.008645311929285526, + 0.012612832710146904, + 0.01002634596079588, + 0.013761759735643864, + 0.008758305571973324, + 0.0064595164731144905, + 0.015750568360090256, + 0.00831875391304493, + 0.011545148678123951, + 0.011924096383154392, + 0.012420007027685642, + 0.024790139868855476, + 0.012422574684023857, + 0.008021587505936623, + 0.02117803692817688, + 0.013210820034146309 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 9 + ], + "selected_reward": -0.055026959627866745 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot framed as an eye-level, static, medium shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man wearing a dark blue merino wool sweater over a white collared shirt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the man turn his head from off-camera to face the camera while opening a silver aluminum laptop?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a sun-drenched modern square featuring a multi-tiered stone fountain with cascading water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the cool light from the laptop screen illuminate the man's face and create a reflection in his eyes?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.14283621311187744 + ], + "opposite_logprob_tokens": [ + -2.017836093902588 + ], + "logprob_reward": -0.14283621311187744, + "yes_no_margin_reward": 1.8749998807907104, + "token_credit_reward": -0.14283621311187744, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10440286993980408, + 0.046255841851234436, + 0.10684884339570999, + 0.09265144914388657, + 0.027406184002757072, + 0.036255210638046265, + 0.06720910221338272, + 0.12292681634426117, + 0.188985213637352, + 0.20705850422382355 + ], + "credit_map": [ + 0.006453083362430334, + 0.013198157772421837, + 0.010629800148308277, + 0.014699730090796947, + 0.022989116609096527, + 0.00907965935766697, + 0.017671925947070122, + 0.009681398048996925, + 0.004871597047895193, + 0.005181111395359039, + 0.004650157410651445, + 0.0047676763497292995, + 0.010000411421060562, + 0.004261962603777647, + 0.00792219303548336, + 0.004600732587277889, + 0.008244108408689499, + 0.018056191504001617, + 0.010160073637962341, + 0.016383923590183258, + 0.016650043427944183, + 0.01104196161031723, + 0.015449529513716698, + 0.010863015428185463, + 0.007720648311078548, + 0.021485915407538414, + 0.0076552690006792545, + 0.011052153073251247, + 0.01232944242656231, + 0.010220867581665516, + 0.012873134575784206, + 0.009314018301665783, + 0.002301144879311323, + 0.0053344350308179855, + 0.002813794184476137, + 0.002330103423446417, + 0.003463710192590952, + 0.0032671208027750254, + 0.004768991842865944, + 0.0031268836464732885, + 0.003793536452576518, + 0.004972977098077536, + 0.0035704176407307386, + 0.0038561183027923107, + 0.005806447938084602, + 0.0032533009070903063, + 0.007167037110775709, + 0.003835374256595969, + 0.008076343685388565, + 0.01729002222418785, + 0.00552338408306241, + 0.005368588026612997, + 0.008153576403856277, + 0.007153023965656757, + 0.009349915198981762, + 0.006294248625636101, + 0.009737823158502579, + 0.026036350056529045, + 0.013329091481864452, + 0.026507670059800148, + 0.009585529565811157, + 0.011806362308561802, + 0.01423562504351139, + 0.011688359081745148, + 0.01664257049560547, + 0.022534294053912163, + 0.019967926666140556, + 0.024018431082367897, + 0.020580396056175232, + 0.015238559804856777, + 0.05008222907781601, + 0.019920796155929565, + 0.01618042215704918, + 0.03029787540435791, + 0.03053945116698742, + 0.07403475046157837, + 0.014432908967137337, + 0.01617249846458435, + 0.014101720415055752, + 0.01129886880517006 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 8, + 7 + ], + "selected_reward": -0.14283621311187744 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot framed as an eye-level, static, medium shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man wearing a dark blue merino wool sweater over a white collared shirt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the man turn his head from off-camera to face the camera while opening a silver aluminum laptop?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a sun-drenched modern square featuring a multi-tiered stone fountain with cascading water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the cool light from the laptop screen illuminate the man's face and create a reflection in his eyes?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.018192337825894356 + ], + "opposite_logprob_tokens": [ + -4.018192291259766 + ], + "logprob_reward": -0.018192337825894356, + "yes_no_margin_reward": 4.0, + "token_credit_reward": -0.018192337825894356, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.20837096869945526, + 0.3369132876396179, + 0.051633596420288086, + 0.04139173403382301, + 0.06997967511415482, + 0.060225896537303925, + 0.019107263535261154, + 0.08262840658426285, + 0.07388416677713394, + 0.05586506053805351 + ], + "credit_map": [ + 0.020674943923950195, + 0.03581540286540985, + 0.01894421875476837, + 0.02889185957610607, + 0.025857020169496536, + 0.024382825940847397, + 0.03386398032307625, + 0.01994071528315544, + 0.04095732420682907, + 0.06480974704027176, + 0.03397209942340851, + 0.03522934392094612, + 0.04979128763079643, + 0.05131277069449425, + 0.0342135950922966, + 0.026627101004123688, + 0.005958779249340296, + 0.008957577869296074, + 0.005585800390690565, + 0.005461936350911856, + 0.008086600340902805, + 0.005479617975652218, + 0.006134763825684786, + 0.005968515761196613, + 0.004440909717231989, + 0.006811162922531366, + 0.003652826650068164, + 0.004015362821519375, + 0.004548624623566866, + 0.0037434205878525972, + 0.009998762980103493, + 0.004180658608675003, + 0.008821493946015835, + 0.01483976747840643, + 0.006915388163179159, + 0.008570458739995956, + 0.008690017275512218, + 0.007546676322817802, + 0.008171998895704746, + 0.006423866376280785, + 0.005687880329787731, + 0.016505854204297066, + 0.004987908992916346, + 0.007299214135855436, + 0.0070527829229831696, + 0.007485759444534779, + 0.00524915661662817, + 0.005957338493317366, + 0.002117609605193138, + 0.0037730245385318995, + 0.0016957910265773535, + 0.0023916789796203375, + 0.002469519851729274, + 0.0019147205166518688, + 0.0026670859660953283, + 0.0020778316538780928, + 0.008231665939092636, + 0.015981489792466164, + 0.008523951284587383, + 0.015095707029104233, + 0.007736882660537958, + 0.0084649957716465, + 0.009599543176591396, + 0.008994163013994694, + 0.007179402280598879, + 0.012530364096164703, + 0.009707366116344929, + 0.012635242193937302, + 0.005227313842624426, + 0.006803350988775492, + 0.011832991614937782, + 0.007968125864863396, + 0.006804395001381636, + 0.006896061822772026, + 0.007120457012206316, + 0.011990672908723354, + 0.005831541959196329, + 0.004663677420467138, + 0.006807900033891201, + 0.005750349722802639 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 7 + ], + "selected_reward": -0.018192337825894356 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot framed as an eye-level, static, medium shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man wearing a dark blue merino wool sweater over a white collared shirt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the man turn his head from off-camera to face the camera while opening a silver aluminum laptop?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a sun-drenched modern square featuring a multi-tiered stone fountain with cascading water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the cool light from the laptop screen illuminate the man's face and create a reflection in his eyes?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02980688400566578 + ], + "opposite_logprob_tokens": [ + -3.5298068523406982 + ], + "logprob_reward": -0.02980688400566578, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.02980688400566578, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11139988899230957, + 0.06699539721012115, + 0.08425400406122208, + 0.10829188674688339, + 0.07025422900915146, + 0.05717307701706886, + 0.06537524610757828, + 0.17952263355255127, + 0.12136620283126831, + 0.13536743819713593 + ], + "credit_map": [ + 0.011243038810789585, + 0.01959729939699173, + 0.008526806719601154, + 0.013758965767920017, + 0.02006864733994007, + 0.014823831617832184, + 0.01341874897480011, + 0.009962552227079868, + 0.005832458380609751, + 0.010262002237141132, + 0.006075466517359018, + 0.006280946079641581, + 0.015218276530504227, + 0.005966659635305405, + 0.009734220802783966, + 0.007625366561114788, + 0.006796760484576225, + 0.014274637214839458, + 0.008269845508038998, + 0.011527344584465027, + 0.014098310843110085, + 0.008461655117571354, + 0.012094711884856224, + 0.008730735629796982, + 0.009560170583426952, + 0.022927727550268173, + 0.009012752212584019, + 0.013247792609035969, + 0.013879927806556225, + 0.011129660531878471, + 0.018321368843317032, + 0.01021249033510685, + 0.0061549236997962, + 0.01214574370533228, + 0.007386866956949234, + 0.009960003197193146, + 0.010318187065422535, + 0.0061172060668468475, + 0.010526749305427074, + 0.00764455134049058, + 0.005705737508833408, + 0.008378105238080025, + 0.005613827612251043, + 0.00717449514195323, + 0.007208456285297871, + 0.005138492677360773, + 0.011958052404224873, + 0.005995907820761204, + 0.008298587054014206, + 0.017137674614787102, + 0.004892088007181883, + 0.005920603405684233, + 0.007705556694418192, + 0.008021329529583454, + 0.006858962122350931, + 0.0065404437482357025, + 0.014336802065372467, + 0.03709829971194267, + 0.02195560745894909, + 0.04853658005595207, + 0.01231128629297018, + 0.016665996983647346, + 0.013882497325539589, + 0.014735568314790726, + 0.010631066747009754, + 0.019017308950424194, + 0.012522289529442787, + 0.016545291990041733, + 0.013375031761825085, + 0.009748464450240135, + 0.027462385594844818, + 0.01206436287611723, + 0.011104559525847435, + 0.0189617108553648, + 0.019118353724479675, + 0.0405474491417408, + 0.011463632807135582, + 0.012648316100239754, + 0.011227831244468689, + 0.01029558852314949 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 9, + 8 + ], + "selected_reward": -0.02980688400566578 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "An eye-level, static, medium shot captures a man in his 30s, with neatly styled dark brown hair and light stubble, sitting at an outdoor table. He wears a dark blue merino wool sweater over a crisp white collared shirt. Initially appearing slightly off-camera, he smoothly turns his head to face the camera, his expression transforming into a friendly, engaging expression as he begins to speak. The setting is a sun-drenched modern square, where a multi-tiered stone fountain provides a tranquil backdrop; water cascades down its weathered surface, and the pool at the bottom creates a canvas of gently shimmering ripples in the bright daylight. As he speaks, the man reaches for a closed, ultra-thin laptop with a silver aluminum finish on the table. Placing both hands on it, he opens the lid with a steady, deliberate movement. The cool light from the screen subtly illuminates his focused features and creates a soft reflection in his eyes. The entire scene is imbued with a calm, professional atmosphere, the man's focused movements harmoniously balanced with the fountain's steady, tranquil flow.", + "seed": 10897, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000071/group_6/index_8653_seed_10897_dp_13_rank_26_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 6, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.2323529962450266, + "vlm_logprob": -0.2323529962450266, + "vlm_yes_no_margin": 2.675, + "vlm_token_credit": -0.2323529962450266, + "vlm_yes_ratio": 0.8, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot framed as an eye-level, static, medium shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man wearing a dark blue merino wool sweater over a white collared shirt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the man turn his head from off-camera to face the camera while opening a silver aluminum laptop?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a sun-drenched modern square featuring a multi-tiered stone fountain with cascading water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the cool light from the laptop screen illuminate the man's face and create a reflection in his eyes?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.05934543162584305 + ], + "opposite_logprob_tokens": [ + -3.5593454837799072 + ], + "logprob_reward": -0.05934543162584305, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.05934543162584305, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.167250856757164, + 0.10675674676895142, + 0.047270145267248154, + 0.06331578642129898, + 0.11338462680578232, + 0.07495979964733124, + 0.16013579070568085, + 0.041215021163225174, + 0.16034506261348724, + 0.06536617130041122 + ], + "credit_map": [ + 0.014752683229744434, + 0.02844020538032055, + 0.016949838027358055, + 0.019256489351391792, + 0.03506414219737053, + 0.021922964602708817, + 0.01617608405649662, + 0.014688464812934399, + 0.011172776110470295, + 0.010585340671241283, + 0.016110314056277275, + 0.009357952512800694, + 0.013032435439527035, + 0.02302470989525318, + 0.01054803840816021, + 0.012925183400511742, + 0.003963971044868231, + 0.005847551394253969, + 0.007829890586435795, + 0.007420134264975786, + 0.006922521162778139, + 0.0040441518649458885, + 0.006193730980157852, + 0.0050481949001550674, + 0.006469652522355318, + 0.010261908173561096, + 0.00960770808160305, + 0.01496990118175745, + 0.0063128480687737465, + 0.004798635374754667, + 0.005230546463280916, + 0.005664587486535311, + 0.01494295708835125, + 0.013242367655038834, + 0.020783690735697746, + 0.011996646411716938, + 0.016180643811821938, + 0.01183050312101841, + 0.01156988088041544, + 0.012837938033044338, + 0.014364220201969147, + 0.008200759999454021, + 0.010203635320067406, + 0.008427123539149761, + 0.0101714376360178, + 0.007509638089686632, + 0.008113456889986992, + 0.007969537749886513, + 0.019561687484383583, + 0.018436972051858902, + 0.03617182374000549, + 0.028481528162956238, + 0.014908943325281143, + 0.01228033285588026, + 0.012635483406484127, + 0.0176590234041214, + 0.004845726303756237, + 0.003916995599865913, + 0.006345669273287058, + 0.008357713930308819, + 0.004039216786623001, + 0.005014700349420309, + 0.004345102701336145, + 0.004349894355982542, + 0.01943732611835003, + 0.020446788519620895, + 0.03286164626479149, + 0.0365547239780426, + 0.016145940870046616, + 0.009787296876311302, + 0.012399575673043728, + 0.012711779214441776, + 0.006418953649699688, + 0.005814180243760347, + 0.016038980334997177, + 0.01105711329728365, + 0.007504023611545563, + 0.006241366267204285, + 0.005786897614598274, + 0.0065046572126448154 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 6 + ], + "selected_reward": -0.05934543162584305 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot framed as an eye-level, static, medium shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man wearing a dark blue merino wool sweater over a white collared shirt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the man turn his head from off-camera to face the camera while opening a silver aluminum laptop?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a sun-drenched modern square featuring a multi-tiered stone fountain with cascading water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the cool light from the laptop screen illuminate the man's face and create a reflection in his eyes?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.08916006982326508 + ], + "opposite_logprob_tokens": [ + -2.4641599655151367 + ], + "logprob_reward": -0.08916006982326508, + "yes_no_margin_reward": 2.375, + "token_credit_reward": -0.08916006982326508, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.20467059314250946, + 0.07545814663171768, + 0.06532636284828186, + 0.08424336463212967, + 0.10764127224683762, + 0.10038281977176666, + 0.15212972462177277, + 0.02824322134256363, + 0.14719267189502716, + 0.03471189737319946 + ], + "credit_map": [ + 0.016146577894687653, + 0.03674323484301567, + 0.021816549822688103, + 0.019182898104190826, + 0.04960276186466217, + 0.021571367979049683, + 0.022081881761550903, + 0.017525307834148407, + 0.00988412369042635, + 0.006994528230279684, + 0.013844281435012817, + 0.007473268546164036, + 0.011186040937900543, + 0.007971423678100109, + 0.008631955832242966, + 0.009472521021962166, + 0.013745254836976528, + 0.006973386742174625, + 0.007593677379190922, + 0.006807500030845404, + 0.008347626775503159, + 0.008963659405708313, + 0.005966731812804937, + 0.006928528659045696, + 0.014002475887537003, + 0.008012287318706512, + 0.021767186000943184, + 0.008879994042217731, + 0.007788403425365686, + 0.007359941955655813, + 0.005539383739233017, + 0.010893692262470722, + 0.009304896928369999, + 0.009922302328050137, + 0.01214631274342537, + 0.015271564014256, + 0.016723979264497757, + 0.012547902762889862, + 0.018314387649297714, + 0.013409911654889584, + 0.009995711036026478, + 0.011907421983778477, + 0.013702458702027798, + 0.014664755202829838, + 0.013596837408840656, + 0.007121982518583536, + 0.01720600388944149, + 0.012187641113996506, + 0.02187463268637657, + 0.01619105041027069, + 0.031264398247003555, + 0.02813061885535717, + 0.014883275143802166, + 0.012675231322646141, + 0.011338399723172188, + 0.015772119164466858, + 0.0019353767856955528, + 0.0029790340922772884, + 0.00482491310685873, + 0.005031040869653225, + 0.004273530095815659, + 0.0028697520028799772, + 0.0033609173260629177, + 0.0029686533380299807, + 0.014762897975742817, + 0.017468582838773727, + 0.024830767884850502, + 0.024572189897298813, + 0.018112793564796448, + 0.009460489265620708, + 0.024350527673959732, + 0.013634406961500645, + 0.002825249684974551, + 0.0034552745055407286, + 0.007134613581001759, + 0.004646558314561844, + 0.006638725753873587, + 0.0031706364825367928, + 0.0033965834882110357, + 0.003444253234192729 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 6, + 8 + ], + "selected_reward": -0.08916006982326508 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot framed as an eye-level, static, medium shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man wearing a dark blue merino wool sweater over a white collared shirt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the man turn his head from off-camera to face the camera while opening a silver aluminum laptop?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a sun-drenched modern square featuring a multi-tiered stone fountain with cascading water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the cool light from the laptop screen illuminate the man's face and create a reflection in his eyes?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -0.9743437767028809 + ], + "opposite_logprob_tokens": [ + -0.47434377670288086 + ], + "logprob_reward": -0.9743437767028809, + "yes_no_margin_reward": -0.5, + "token_credit_reward": -0.9743437767028809, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.18559981882572174, + 0.09839842468500137, + 0.06377972662448883, + 0.03988775983452797, + 0.03367357701063156, + 0.07140196114778519, + 0.0674436092376709, + 0.06254711747169495, + 0.3603975474834442, + 0.016870521008968353 + ], + "credit_map": [ + 0.017574286088347435, + 0.022744523361325264, + 0.018101869150996208, + 0.020286638289690018, + 0.04110301285982132, + 0.02993842214345932, + 0.01901835761964321, + 0.016832726076245308, + 0.012735811062157154, + 0.009852859191596508, + 0.01337452419102192, + 0.013874479569494724, + 0.016001660376787186, + 0.01072086300700903, + 0.011608208529651165, + 0.010230015031993389, + 0.004515595268458128, + 0.0047792671248316765, + 0.020677248015999794, + 0.007708511780947447, + 0.006753005087375641, + 0.004486283287405968, + 0.00466473214328289, + 0.01019508671015501, + 0.005537775345146656, + 0.006726650986820459, + 0.0068984865210950375, + 0.003920954652130604, + 0.0052735586650669575, + 0.0033808674197643995, + 0.0037978123873472214, + 0.004351654089987278, + 0.0039886897429823875, + 0.003954702988266945, + 0.005556780844926834, + 0.004829973913729191, + 0.005124792456626892, + 0.00289947004057467, + 0.0035314715933054686, + 0.0037876968272030354, + 0.008615972474217415, + 0.007609753403812647, + 0.015598876401782036, + 0.010026246309280396, + 0.008145860396325588, + 0.005816649179905653, + 0.008012468926608562, + 0.007576135452836752, + 0.013063767924904823, + 0.0064422148279845715, + 0.013153373263776302, + 0.005631168372929096, + 0.008554941974580288, + 0.00615634536370635, + 0.006938444450497627, + 0.007503348868340254, + 0.006838694214820862, + 0.00784815102815628, + 0.007795457728207111, + 0.010926200076937675, + 0.00885917991399765, + 0.007626943290233612, + 0.006549676414579153, + 0.0061028185300529, + 0.03078017197549343, + 0.035720936954021454, + 0.05098589509725571, + 0.05264152213931084, + 0.05886716768145561, + 0.026180174201726913, + 0.07041355222463608, + 0.03480809926986694, + 0.0019813606049865484, + 0.002613968215882778, + 0.0024892452638596296, + 0.0029499405063688755, + 0.0023762208875268698, + 0.001410474767908454, + 0.0017514670034870505, + 0.0012978423619642854 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 0, + 1 + ], + "selected_reward": -0.9743437767028809 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot framed as an eye-level, static, medium shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man wearing a dark blue merino wool sweater over a white collared shirt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the man turn his head from off-camera to face the camera while opening a silver aluminum laptop?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a sun-drenched modern square featuring a multi-tiered stone fountain with cascading water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the cool light from the laptop screen illuminate the man's face and create a reflection in his eyes?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01255360059440136 + ], + "opposite_logprob_tokens": [ + -4.387553691864014 + ], + "logprob_reward": -0.01255360059440136, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.01255360059440136, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.25359266996383667, + 0.1345251202583313, + 0.09257607907056808, + 0.10809298604726791, + 0.07207585126161575, + 0.04022533819079399, + 0.031547579914331436, + 0.052690550684928894, + 0.1644342839717865, + 0.05023949220776558 + ], + "credit_map": [ + 0.024860188364982605, + 0.04673009738326073, + 0.031119532883167267, + 0.026598837226629257, + 0.05547671765089035, + 0.02442610077559948, + 0.02543409913778305, + 0.018947118893265724, + 0.020208092406392097, + 0.013416203670203686, + 0.021901950240135193, + 0.025355270132422447, + 0.012573422864079475, + 0.013978247530758381, + 0.013790980912744999, + 0.013300947844982147, + 0.007965621538460255, + 0.007005374878644943, + 0.03278876841068268, + 0.012655912898480892, + 0.007137193810194731, + 0.006632208824157715, + 0.00525016151368618, + 0.013140839524567127, + 0.014151486568152905, + 0.013442051596939564, + 0.027727745473384857, + 0.018977662548422813, + 0.006554529536515474, + 0.00827407743781805, + 0.007177126128226519, + 0.011788306757807732, + 0.011454506777226925, + 0.008932642638683319, + 0.01854669861495495, + 0.01051284745335579, + 0.004167236387729645, + 0.0056517054326832294, + 0.004340342711657286, + 0.008469868451356888, + 0.0055163842625916, + 0.004710068926215172, + 0.008667899295687675, + 0.006207749713212252, + 0.003638418158516288, + 0.003727068891748786, + 0.0031087868846952915, + 0.004648961592465639, + 0.005464234855026007, + 0.0032683874014765024, + 0.006462224293500185, + 0.004039929248392582, + 0.003376457141712308, + 0.0025458820164203644, + 0.002751016290858388, + 0.003639448434114456, + 0.0076715475879609585, + 0.00762274069711566, + 0.010135577991604805, + 0.008426858112215996, + 0.005046146921813488, + 0.004922755062580109, + 0.003966224379837513, + 0.004898697603493929, + 0.014798779040575027, + 0.01715056598186493, + 0.025173965841531754, + 0.0251348614692688, + 0.02350246161222458, + 0.01222971174865961, + 0.030714645981788635, + 0.015729285776615143, + 0.003756181104108691, + 0.003855862421914935, + 0.015337607823312283, + 0.008134732022881508, + 0.0046935006976127625, + 0.00435289554297924, + 0.0039252168498933315, + 0.006183497607707977 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 1 + ], + "selected_reward": -0.01255360059440136 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot framed as an eye-level, static, medium shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man wearing a dark blue merino wool sweater over a white collared shirt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the man turn his head from off-camera to face the camera while opening a silver aluminum laptop?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a sun-drenched modern square featuring a multi-tiered stone fountain with cascading water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the cool light from the laptop screen illuminate the man's face and create a reflection in his eyes?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0263621024787426 + ], + "opposite_logprob_tokens": [ + -3.651362180709839 + ], + "logprob_reward": -0.0263621024787426, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.0263621024787426, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.27939316630363464, + 0.09904260188341141, + 0.057676076889038086, + 0.04509815201163292, + 0.03999301418662071, + 0.06315691024065018, + 0.06976497173309326, + 0.04582527279853821, + 0.27130430936813354, + 0.028745543211698532 + ], + "credit_map": [ + 0.016782600432634354, + 0.07987598329782486, + 0.023678414523601532, + 0.01810881681740284, + 0.07687398046255112, + 0.026668593287467957, + 0.01886044815182686, + 0.01854434236884117, + 0.009080510586500168, + 0.00874846801161766, + 0.020116878673434258, + 0.014627635478973389, + 0.015189848840236664, + 0.008570999838411808, + 0.009819325059652328, + 0.012888941913843155, + 0.006331086624413729, + 0.005141479894518852, + 0.014941563829779625, + 0.007364907767623663, + 0.006029898300766945, + 0.00445386441424489, + 0.0046738856472074986, + 0.008739395998418331, + 0.007074110209941864, + 0.005451293662190437, + 0.009859939105808735, + 0.004204357508569956, + 0.005014581605792046, + 0.004031538963317871, + 0.003771638497710228, + 0.005690695717930794, + 0.0030568146612495184, + 0.00421593664214015, + 0.00455981632694602, + 0.004497117828577757, + 0.008679009974002838, + 0.0031879304442554712, + 0.006281893700361252, + 0.005514495074748993, + 0.007642908953130245, + 0.00807874370366335, + 0.010496502742171288, + 0.012687505222856998, + 0.00723686907440424, + 0.004941421560943127, + 0.005536537617444992, + 0.006536422297358513, + 0.011127219535410404, + 0.007156337145715952, + 0.013384087011218071, + 0.011614801362156868, + 0.006853404454886913, + 0.0056181359104812145, + 0.006662067491561174, + 0.007348919287323952, + 0.004092932678759098, + 0.00576254166662693, + 0.008068237453699112, + 0.008945421315729618, + 0.005955618340522051, + 0.003474291879683733, + 0.004870621487498283, + 0.004655610304325819, + 0.023360125720500946, + 0.030072538182139397, + 0.03805527836084366, + 0.03944573923945427, + 0.04032997414469719, + 0.01813559979200363, + 0.053641024976968765, + 0.028264062479138374, + 0.002525902818888426, + 0.0035939544904977083, + 0.0045147547498345375, + 0.005840099882334471, + 0.0050206598825752735, + 0.001955060986801982, + 0.0027340699452906847, + 0.0025610418524593115 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 1 + ], + "selected_reward": -0.0263621024787426 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "No", + "Yes", + "Yes" + ] + }, + { + "prompt": "An eye-level, static, medium shot captures a man in his 30s, with neatly styled dark brown hair and light stubble, sitting at an outdoor table. He wears a dark blue merino wool sweater over a crisp white collared shirt. Initially appearing slightly off-camera, he smoothly turns his head to face the camera, his expression transforming into a friendly, engaging expression as he begins to speak. The setting is a sun-drenched modern square, where a multi-tiered stone fountain provides a tranquil backdrop; water cascades down its weathered surface, and the pool at the bottom creates a canvas of gently shimmering ripples in the bright daylight. As he speaks, the man reaches for a closed, ultra-thin laptop with a silver aluminum finish on the table. Placing both hands on it, he opens the lid with a steady, deliberate movement. The cool light from the screen subtly illuminates his focused features and creates a soft reflection in his eyes. The entire scene is imbued with a calm, professional atmosphere, the man's focused movements harmoniously balanced with the fountain's steady, tranquil flow.", + "seed": 10898, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000071/group_6/index_8653_seed_10898_dp_13_rank_26_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 6, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.05886700181290507, + "vlm_logprob": -0.05886700181290507, + "vlm_yes_no_margin": 3.525, + "vlm_token_credit": -0.05886700181290507, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot framed as an eye-level, static, medium shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man wearing a dark blue merino wool sweater over a white collared shirt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the man turn his head from off-camera to face the camera while opening a silver aluminum laptop?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a sun-drenched modern square featuring a multi-tiered stone fountain with cascading water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the cool light from the laptop screen illuminate the man's face and create a reflection in his eyes?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.09904886782169342 + ], + "opposite_logprob_tokens": [ + -3.5990488529205322 + ], + "logprob_reward": -0.09904886782169342, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.09904886782169342, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13471099734306335, + 0.12616108357906342, + 0.1542440950870514, + 0.035632673650979996, + 0.044885724782943726, + 0.04369596391916275, + 0.04048769176006317, + 0.31158074736595154, + 0.06709661334753036, + 0.04150444269180298 + ], + "credit_map": [ + 0.021594127640128136, + 0.025295602157711983, + 0.025128018110990524, + 0.016073426231741905, + 0.015132580883800983, + 0.011037266813218594, + 0.009680520743131638, + 0.010769469663500786, + 0.006942082196474075, + 0.026353217661380768, + 0.036436181515455246, + 0.013005398213863373, + 0.01221186388283968, + 0.009087211452424526, + 0.013441193848848343, + 0.00868392363190651, + 0.029895830899477005, + 0.02324589341878891, + 0.05611114576458931, + 0.01570749096572399, + 0.00842402596026659, + 0.00897239614278078, + 0.0049095978029072285, + 0.0069777099415659904, + 0.0045127542689442635, + 0.004223896190524101, + 0.006076912861317396, + 0.005172334611415863, + 0.003878695657476783, + 0.00532298069447279, + 0.0030597816221415997, + 0.003385320072993636, + 0.006739666685461998, + 0.0077874488197267056, + 0.008744141086935997, + 0.008249291218817234, + 0.0028416565619409084, + 0.003499071579426527, + 0.0024851872585713863, + 0.0045392620377242565, + 0.0044441563077270985, + 0.0067421915009617805, + 0.007699011359363794, + 0.00812418945133686, + 0.006373689975589514, + 0.003046247176826, + 0.003691730322316289, + 0.0035747478250414133, + 0.0038835974410176277, + 0.006425398401916027, + 0.006142421625554562, + 0.009402303025126457, + 0.004308067727833986, + 0.003551115747541189, + 0.0032135893125087023, + 0.003561197081580758, + 0.06644736975431442, + 0.12598659098148346, + 0.026450766250491142, + 0.0197896808385849, + 0.02216046303510666, + 0.010167434811592102, + 0.010266179218888283, + 0.03031226620078087, + 0.009021691046655178, + 0.005485227797180414, + 0.007036246359348297, + 0.0066901626996695995, + 0.007953647524118423, + 0.017594415694475174, + 0.005100512411445379, + 0.008214704692363739, + 0.00413224333897233, + 0.005580204539000988, + 0.005605889018625021, + 0.007360228803008795, + 0.005169393960386515, + 0.005494366865605116, + 0.0044345310889184475, + 0.003727586707100272 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 2, + 0 + ], + "selected_reward": -0.09904886782169342 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot framed as an eye-level, static, medium shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man wearing a dark blue merino wool sweater over a white collared shirt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the man turn his head from off-camera to face the camera while opening a silver aluminum laptop?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a sun-drenched modern square featuring a multi-tiered stone fountain with cascading water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the cool light from the laptop screen illuminate the man's face and create a reflection in his eyes?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.06213174760341644 + ], + "opposite_logprob_tokens": [ + -2.812131643295288 + ], + "logprob_reward": -0.06213174760341644, + "yes_no_margin_reward": 2.75, + "token_credit_reward": -0.06213174760341644, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2006046324968338, + 0.130087748169899, + 0.07392619550228119, + 0.037078607827425, + 0.042196400463581085, + 0.04059207811951637, + 0.023497000336647034, + 0.4083191752433777, + 0.0211503766477108, + 0.02254779264330864 + ], + "credit_map": [ + 0.011972208507359028, + 0.05381209775805473, + 0.06323631852865219, + 0.02292107790708542, + 0.011924984864890575, + 0.009650287218391895, + 0.01452233549207449, + 0.012565317563712597, + 0.007362391334027052, + 0.022410281002521515, + 0.027013439685106277, + 0.0125174131244421, + 0.024273810908198357, + 0.009693940170109272, + 0.016172591596841812, + 0.010643880814313889, + 0.012845591641962528, + 0.02073775790631771, + 0.013390565291047096, + 0.007402390707284212, + 0.005021674558520317, + 0.004827507771551609, + 0.003930285573005676, + 0.0057704211212694645, + 0.0046081566251814365, + 0.0045081619173288345, + 0.008666940964758396, + 0.005186573136597872, + 0.003029695013538003, + 0.005064796656370163, + 0.00249549956060946, + 0.0035187804605811834, + 0.007638916373252869, + 0.008679802529513836, + 0.006095634773373604, + 0.004580752458423376, + 0.004692339804023504, + 0.003346784505993128, + 0.0026653569657355547, + 0.004496814217418432, + 0.004399615805596113, + 0.004983989521861076, + 0.007590864319354296, + 0.006195145659148693, + 0.003975683357566595, + 0.004247170872986317, + 0.0032632052898406982, + 0.005936404224485159, + 0.002560354070737958, + 0.0035894548054784536, + 0.002630820730701089, + 0.004811514168977737, + 0.002722258446738124, + 0.002176980720832944, + 0.0026955422945320606, + 0.002310076728463173, + 0.09185050427913666, + 0.16678105294704437, + 0.03333635255694389, + 0.025244547054171562, + 0.027383457869291306, + 0.010175155475735664, + 0.013868547044694424, + 0.03967957943677902, + 0.0021886778995394707, + 0.0037611888255923986, + 0.006124441046267748, + 0.0024391168262809515, + 0.001444938126951456, + 0.0012286114506423473, + 0.0015781174879521132, + 0.0023852859158068895, + 0.0018462628358975053, + 0.0016978293424472213, + 0.0021960376761853695, + 0.002557585248723626, + 0.005267122760415077, + 0.0021931088995188475, + 0.0044197035022079945, + 0.0023701421450823545 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 0, + 1 + ], + "selected_reward": -0.06213174760341644 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot framed as an eye-level, static, medium shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man wearing a dark blue merino wool sweater over a white collared shirt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the man turn his head from off-camera to face the camera while opening a silver aluminum laptop?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a sun-drenched modern square featuring a multi-tiered stone fountain with cascading water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the cool light from the laptop screen illuminate the man's face and create a reflection in his eyes?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.11297557502985 + ], + "opposite_logprob_tokens": [ + -2.237975597381592 + ], + "logprob_reward": -0.11297557502985, + "yes_no_margin_reward": 2.125, + "token_credit_reward": -0.11297557502985, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.20194925367832184, + 0.07253962755203247, + 0.18071246147155762, + 0.04658142477273941, + 0.03295546770095825, + 0.1247943788766861, + 0.0401608869433403, + 0.2348892241716385, + 0.04370266944169998, + 0.021714594215154648 + ], + "credit_map": [ + 0.014679884538054466, + 0.028042295947670937, + 0.04468269646167755, + 0.022376444190740585, + 0.0374324731528759, + 0.02314027026295662, + 0.016093460842967033, + 0.015501735731959343, + 0.007644271943718195, + 0.005583604332059622, + 0.010354827158153057, + 0.009371698834002018, + 0.01095797959715128, + 0.010409507900476456, + 0.011441641487181187, + 0.006776099093258381, + 0.012562641873955727, + 0.04934443533420563, + 0.06505045294761658, + 0.016274360939860344, + 0.011405883356928825, + 0.008776530623435974, + 0.006668602582067251, + 0.010629557073116302, + 0.0054941908456385136, + 0.005482773296535015, + 0.013285145163536072, + 0.005273293238133192, + 0.004325182177126408, + 0.005933626089245081, + 0.0029404202941805124, + 0.003846794366836548, + 0.003098815679550171, + 0.0041124774143099785, + 0.011608659289777279, + 0.0036642078775912523, + 0.0028020073659718037, + 0.0018077356508001685, + 0.001553559210151434, + 0.004308005329221487, + 0.008622678928077221, + 0.013346437364816666, + 0.04260578006505966, + 0.016668308526277542, + 0.008357786573469639, + 0.005017181858420372, + 0.006695288233458996, + 0.023480908945202827, + 0.003648285521194339, + 0.003478279337286949, + 0.005752017255872488, + 0.007969004102051258, + 0.005738688167184591, + 0.004667609464377165, + 0.004874781239777803, + 0.004032221157103777, + 0.05676187947392464, + 0.08731923997402191, + 0.02109864167869091, + 0.013754121959209442, + 0.015107466839253902, + 0.007006557192653418, + 0.010058877989649773, + 0.02378244698047638, + 0.006778951268643141, + 0.005635588429868221, + 0.005624727346003056, + 0.0068739233538508415, + 0.006712003611028194, + 0.003486504778265953, + 0.0038838540203869343, + 0.004707114305347204, + 0.0019863976631313562, + 0.001694448641501367, + 0.003485444700345397, + 0.0035047661513090134, + 0.0036065869498997927, + 0.0028936455491930246, + 0.0026605483144521713, + 0.0018827570602297783 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 0, + 2 + ], + "selected_reward": -0.11297557502985 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot framed as an eye-level, static, medium shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man wearing a dark blue merino wool sweater over a white collared shirt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the man turn his head from off-camera to face the camera while opening a silver aluminum laptop?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a sun-drenched modern square featuring a multi-tiered stone fountain with cascading water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the cool light from the laptop screen illuminate the man's face and create a reflection in his eyes?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.007636283989995718 + ], + "opposite_logprob_tokens": [ + -4.882636070251465 + ], + "logprob_reward": -0.007636283989995718, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.007636283989995718, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.24852286279201508, + 0.10540136694908142, + 0.1499439924955368, + 0.11912573873996735, + 0.1565455049276352, + 0.08262915909290314, + 0.03164307028055191, + 0.05188879743218422, + 0.020929085090756416, + 0.03337040916085243 + ], + "credit_map": [ + 0.028562970459461212, + 0.07179895788431168, + 0.05853335186839104, + 0.034751418977975845, + 0.014154571108520031, + 0.011046181432902813, + 0.014420971274375916, + 0.015254458412528038, + 0.009190254844725132, + 0.014278015121817589, + 0.026484860107302666, + 0.02187744714319706, + 0.008927516639232635, + 0.009813118726015091, + 0.007281553000211716, + 0.007548597641289234, + 0.04727557301521301, + 0.021423377096652985, + 0.036703020334243774, + 0.015132283791899681, + 0.006506594829261303, + 0.009439586661756039, + 0.005311311688274145, + 0.008152240887284279, + 0.015461316332221031, + 0.01572917029261589, + 0.026036908850073814, + 0.02063026651740074, + 0.009327403269708157, + 0.01716681942343712, + 0.006684069987386465, + 0.008089780807495117, + 0.02122754603624344, + 0.029984086751937866, + 0.019558671861886978, + 0.03086281567811966, + 0.012148837558925152, + 0.023188894614577293, + 0.008743222802877426, + 0.010831434279680252, + 0.00817448552697897, + 0.010902704671025276, + 0.029756471514701843, + 0.010190132074058056, + 0.005631631705909967, + 0.004928037989884615, + 0.004003388341516256, + 0.00904231145977974, + 0.0033486553002148867, + 0.004959562327712774, + 0.006212117616087198, + 0.007077432237565517, + 0.0031276110094040632, + 0.0022192394826561213, + 0.0023201950825750828, + 0.0023782553616911173, + 0.007124875672161579, + 0.01959678716957569, + 0.005804060027003288, + 0.004175314214080572, + 0.003800125326961279, + 0.002077596029266715, + 0.0029358025640249252, + 0.006374237593263388, + 0.001618224661797285, + 0.00279452302493155, + 0.004646902903914452, + 0.002608792157843709, + 0.0021708866115659475, + 0.0029822317883372307, + 0.0021548622753471136, + 0.0019526617834344506, + 0.004645431414246559, + 0.004906035028398037, + 0.004786858335137367, + 0.007426535710692406, + 0.003233378054574132, + 0.002541791182011366, + 0.003035556524991989, + 0.0027948252391070127 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 2 + ], + "selected_reward": -0.007636283989995718 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot framed as an eye-level, static, medium shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man wearing a dark blue merino wool sweater over a white collared shirt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the man turn his head from off-camera to face the camera while opening a silver aluminum laptop?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a sun-drenched modern square featuring a multi-tiered stone fountain with cascading water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the cool light from the laptop screen illuminate the man's face and create a reflection in his eyes?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.012542534619569778 + ], + "opposite_logprob_tokens": [ + -4.387542724609375 + ], + "logprob_reward": -0.012542534619569778, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.012542534619569778, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.16825900971889496, + 0.03029596246778965, + 0.19311386346817017, + 0.05100111663341522, + 0.07351607829332352, + 0.03514638543128967, + 0.022676564753055573, + 0.3648979365825653, + 0.033950887620449066, + 0.027142176404595375 + ], + "credit_map": [ + 0.01242996659129858, + 0.030528061091899872, + 0.04210200533270836, + 0.026447949931025505, + 0.01828967034816742, + 0.012701444327831268, + 0.015506232157349586, + 0.010253689251840115, + 0.0034412911627441645, + 0.002417138312011957, + 0.0051781777292490005, + 0.005505212116986513, + 0.003447273513302207, + 0.00479725981131196, + 0.0028202272951602936, + 0.0026893841568380594, + 0.016867313534021378, + 0.052751123905181885, + 0.059251852333545685, + 0.02134605310857296, + 0.011261017993092537, + 0.012681850232183933, + 0.007178314030170441, + 0.011776350438594818, + 0.00614248076453805, + 0.005542923230677843, + 0.01192409172654152, + 0.009040641598403454, + 0.003897294169291854, + 0.007721292786300182, + 0.00296444958075881, + 0.003767947666347027, + 0.013046528212726116, + 0.009304611012339592, + 0.02931315079331398, + 0.006489069666713476, + 0.005120755173265934, + 0.0030123102478682995, + 0.0029865987598896027, + 0.004243060480803251, + 0.005537876859307289, + 0.003709714859724045, + 0.0045929523184895515, + 0.00426417076960206, + 0.005144862458109856, + 0.002919387072324753, + 0.0023452099412679672, + 0.0066322144120931625, + 0.0022581620141863823, + 0.0023839890491217375, + 0.003309813793748617, + 0.004992422182112932, + 0.002716055838391185, + 0.0023209359496831894, + 0.002854076214134693, + 0.0018411112250760198, + 0.11548411101102829, + 0.1105375736951828, + 0.021962786093354225, + 0.015254505909979343, + 0.018611470237374306, + 0.0148897934705019, + 0.014240466058254242, + 0.05391722172498703, + 0.003323256503790617, + 0.0037890884559601545, + 0.011258476413786411, + 0.0048766061663627625, + 0.0032260033767670393, + 0.0020557644311338663, + 0.002403853228315711, + 0.0030178395099937916, + 0.0023712010588496923, + 0.0021589978132396936, + 0.004045616835355759, + 0.003999345004558563, + 0.0037190807051956654, + 0.002934894757345319, + 0.004079666920006275, + 0.0038333754055202007 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 2, + 0 + ], + "selected_reward": -0.012542534619569778 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000071/group_7/index_tensor([9153, 9153])_group_7_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000071/group_7/index_tensor([9153, 9153])_group_7_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..a52ac47b5f6e69fda3999c726a671a061a04e284 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000071/group_7/index_tensor([9153, 9153])_group_7_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "static close-up shot from a slightly upward angle focuses on a serene practitioner seated before a heavy wooden Go board, lost in thought. His head is shaved, his features thin and angular, his eyes downcast as he gazes intently at the endgame on the board, his lips pursed in a look of intense concentration. He wears a simple, dark gray cotton and linen robe, its collar neatly tied, its folds naturally flowing as he leans forward. A strong sidelight streams in from outside the frame, sharply sculpting the curve of his head, the high bridge of his nose, and the contours of his cheeks in classic black and white, while the other side of his face is completely shrouded in deep shadow. His right hand gracefully hovers above the board, his index and middle fingers firmly holding a round, smooth black Go piece. The piece reflects a subtle luster in the light, as if about to fall, yet suspended in the air. The background is a blurred dark wooden grille, and the entire picture is extremely restrained in tone, presenting a high-contrast black and white film texture, creating a tense atmosphere of frozen time, far-reaching tranquility, and full of oriental philosophical wisdom.", + "seed": 3716, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000071/group_7/index_9153_seed_3716_dp_14_rank_28_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 7, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.023037057369947433, + "vlm_logprob": -0.023037057369947433, + "vlm_yes_no_margin": 3.925, + "vlm_token_credit": -0.023037057369947433, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a practitioner with a shaved head wearing a dark gray cotton and linen robe?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the practitioner seated before a heavy wooden Go board with his right hand hovering above it holding a black Go piece?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene lit by a strong sidelight from outside the frame that creates high contrast with deep shadows on one side of the face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera positioned in a static close-up shot from a slightly upward angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video feature a blurred dark wooden grille background with a high-contrast black and white film texture?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0242171511054039 + ], + "opposite_logprob_tokens": [ + -3.774217128753662 + ], + "logprob_reward": -0.0242171511054039, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.0242171511054039, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.051597945392131805, + 0.025525309145450592, + 0.033556945621967316, + 0.033439699560403824, + 0.05840197578072548, + 0.10709585249423981, + 0.04207417741417885, + 0.04691101983189583, + 0.05476279556751251, + 0.5466343760490417 + ], + "credit_map": [ + 0.004337313584983349, + 0.004703432787209749, + 0.00514032831415534, + 0.0076645296066999435, + 0.009089051745831966, + 0.007481940556317568, + 0.00702481996268034, + 0.006156531628221273, + 0.002154936082661152, + 0.0033418305683881044, + 0.002481587463989854, + 0.00244907452724874, + 0.004111014772206545, + 0.003828529268503189, + 0.00441397912800312, + 0.0027443578001111746, + 0.003249288769438863, + 0.0025266828015446663, + 0.008657648228108883, + 0.0019166965503245592, + 0.0031179385259747505, + 0.007593047805130482, + 0.002938411431387067, + 0.0035572308115661144, + 0.0071671693585813046, + 0.0020903770346194506, + 0.001833887305110693, + 0.0018579947063699365, + 0.002867767820134759, + 0.005433198995888233, + 0.002667579799890518, + 0.009521722793579102, + 0.007034273352473974, + 0.007217299658805132, + 0.0036793590988963842, + 0.004303467459976673, + 0.013259835541248322, + 0.009020286612212658, + 0.006867788266390562, + 0.007019667886197567, + 0.010515446774661541, + 0.013958906754851341, + 0.005067826714366674, + 0.0064244018867611885, + 0.022936193272471428, + 0.024363767355680466, + 0.011036355048418045, + 0.012792956084012985, + 0.004844421520829201, + 0.004788024816662073, + 0.0022173363249748945, + 0.002612774260342121, + 0.008707774803042412, + 0.00864238291978836, + 0.004538619890809059, + 0.005722840782254934, + 0.004260326270014048, + 0.004557049833238125, + 0.0020273991394788027, + 0.0024251737631857395, + 0.010976654477417469, + 0.010795524343848228, + 0.004756119102239609, + 0.00711276987567544, + 0.005454840138554573, + 0.007813850417733192, + 0.002584993140771985, + 0.0028212256729602814, + 0.006243388168513775, + 0.014400700107216835, + 0.0047735245898365974, + 0.010670273564755917, + 0.06518537551164627, + 0.10551387816667557, + 0.023151613771915436, + 0.02974526956677437, + 0.06888865679502487, + 0.08797786384820938, + 0.039271384477615356, + 0.1269003450870514 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 5, + 4 + ], + "selected_reward": -0.0242171511054039 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a practitioner with a shaved head wearing a dark gray cotton and linen robe?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the practitioner seated before a heavy wooden Go board with his right hand hovering above it holding a black Go piece?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene lit by a strong sidelight from outside the frame that creates high contrast with deep shadows on one side of the face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera positioned in a static close-up shot from a slightly upward angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video feature a blurred dark wooden grille background with a high-contrast black and white film texture?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0381077416241169 + ], + "opposite_logprob_tokens": [ + -3.2881076335906982 + ], + "logprob_reward": -0.0381077416241169, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.0381077416241169, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07684608548879623, + 0.030773647129535675, + 0.5245429277420044, + 0.09054634720087051, + 0.07086536288261414, + 0.03456586226820946, + 0.017706794664263725, + 0.020942000672221184, + 0.09597443044185638, + 0.037236444652080536 + ], + "credit_map": [ + 0.009503018110990524, + 0.005725071299821138, + 0.006580303888767958, + 0.009275258518755436, + 0.01369707752019167, + 0.012646786868572235, + 0.011965491808950901, + 0.007453083526343107, + 0.0028119152411818504, + 0.0038000354543328285, + 0.003209028160199523, + 0.004381158854812384, + 0.005575447343289852, + 0.0036494347732514143, + 0.0036663166247308254, + 0.0036803148686885834, + 0.04445725306868553, + 0.025266170501708984, + 0.1604168564081192, + 0.03056267648935318, + 0.027076557278633118, + 0.14181533455848694, + 0.03908655419945717, + 0.05586158484220505, + 0.025832224637269974, + 0.006977557670325041, + 0.005590067710727453, + 0.004177233669906855, + 0.0058386679738759995, + 0.010201860219240189, + 0.006734749767929316, + 0.025193991139531136, + 0.0076506114564836025, + 0.007592757232487202, + 0.004031407181173563, + 0.0040763746947050095, + 0.0142595786601305, + 0.013403707183897495, + 0.007942844182252884, + 0.011908096261322498, + 0.002318461425602436, + 0.004249468445777893, + 0.0014987573958933353, + 0.001793957082554698, + 0.010059554129838943, + 0.005211336072534323, + 0.0047192950733006, + 0.004715037997812033, + 0.0018731545424088836, + 0.002403587568551302, + 0.0013177882647141814, + 0.0013199941022321582, + 0.003921582363545895, + 0.0026619601994752884, + 0.0018950686790049076, + 0.002313661389052868, + 0.001936277374625206, + 0.0025251538027077913, + 0.001168027869425714, + 0.001576500479131937, + 0.003523200284689665, + 0.004615310579538345, + 0.002965575084090233, + 0.0026319585740566254, + 0.010132134892046452, + 0.014561373740434647, + 0.004540377762168646, + 0.005356257315725088, + 0.010511834174394608, + 0.023935992270708084, + 0.007544221822172403, + 0.01939225383102894, + 0.004842362366616726, + 0.005969981662929058, + 0.0018445876194164157, + 0.0023237348068505526, + 0.007772912736982107, + 0.004381032194942236, + 0.0032310483511537313, + 0.006870787125080824 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 8, + 3 + ], + "selected_reward": -0.0381077416241169 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a practitioner with a shaved head wearing a dark gray cotton and linen robe?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the practitioner seated before a heavy wooden Go board with his right hand hovering above it holding a black Go piece?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene lit by a strong sidelight from outside the frame that creates high contrast with deep shadows on one side of the face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera positioned in a static close-up shot from a slightly upward angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video feature a blurred dark wooden grille background with a high-contrast black and white film texture?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.026340274140238762 + ], + "opposite_logprob_tokens": [ + -3.6513402462005615 + ], + "logprob_reward": -0.026340274140238762, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.026340274140238762, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06301715224981308, + 0.02398158609867096, + 0.7243151664733887, + 0.05145859345793724, + 0.03263537585735321, + 0.028242405503988266, + 0.012587363831698895, + 0.008559051901102066, + 0.013117903843522072, + 0.0420854277908802 + ], + "credit_map": [ + 0.007454650476574898, + 0.00451356265693903, + 0.006549187004566193, + 0.009140538983047009, + 0.010492374189198017, + 0.010401846840977669, + 0.008553262799978256, + 0.005911722779273987, + 0.003261514939367771, + 0.002500873990356922, + 0.002913119737058878, + 0.0030366696882992983, + 0.003613658482208848, + 0.002670586109161377, + 0.0027751759625971317, + 0.0032099876552820206, + 0.06169874966144562, + 0.03499439358711243, + 0.2091817706823349, + 0.043040405958890915, + 0.039855290204286575, + 0.20374910533428192, + 0.051657140254974365, + 0.08013832569122314, + 0.013593260198831558, + 0.0030522989109158516, + 0.002847128314897418, + 0.002407608786597848, + 0.005377172492444515, + 0.005312473978847265, + 0.004815876483917236, + 0.014052776619791985, + 0.003881890559569001, + 0.004934457130730152, + 0.002186605939641595, + 0.002301323926076293, + 0.006100753787904978, + 0.0053075384348630905, + 0.0036446822341531515, + 0.004278122913092375, + 0.002862790832296014, + 0.0031446637585759163, + 0.0014703241176903248, + 0.0012584531214088202, + 0.007306208834052086, + 0.00495704123750329, + 0.003134671598672867, + 0.004108252469450235, + 0.0012637058971449733, + 0.0014004192780703306, + 0.001012102235108614, + 0.0008161134319379926, + 0.002109164372086525, + 0.0024760866072028875, + 0.00137006351724267, + 0.002139708260074258, + 0.0007628711173310876, + 0.0008793296874500811, + 0.0005617430433630943, + 0.000687145977281034, + 0.002171826548874378, + 0.0012427535839378834, + 0.0010559658985584974, + 0.0011974158696830273, + 0.0014194183750078082, + 0.0017002490349113941, + 0.0006987327942624688, + 0.0007182596600614488, + 0.0017018825747072697, + 0.0033832413610070944, + 0.001150819705799222, + 0.002345300279557705, + 0.004991388879716396, + 0.00820101611316204, + 0.0019185146084055305, + 0.0024027011822909117, + 0.004583388101309538, + 0.006673465017229319, + 0.0031031924299895763, + 0.010211761109530926 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 0, + 3 + ], + "selected_reward": -0.026340274140238762 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a practitioner with a shaved head wearing a dark gray cotton and linen robe?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the practitioner seated before a heavy wooden Go board with his right hand hovering above it holding a black Go piece?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene lit by a strong sidelight from outside the frame that creates high contrast with deep shadows on one side of the face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera positioned in a static close-up shot from a slightly upward angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video feature a blurred dark wooden grille background with a high-contrast black and white film texture?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0205729641020298 + ], + "opposite_logprob_tokens": [ + -3.8955729007720947 + ], + "logprob_reward": -0.0205729641020298, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.0205729641020298, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1205790787935257, + 0.0269614290446043, + 0.17327378690242767, + 0.1441170871257782, + 0.23574139177799225, + 0.13139835000038147, + 0.034512851387262344, + 0.0245524775236845, + 0.040503621101379395, + 0.06835990399122238 + ], + "credit_map": [ + 0.014113672077655792, + 0.011149823665618896, + 0.00969678908586502, + 0.017208311706781387, + 0.02357099950313568, + 0.01932942122220993, + 0.01464845146983862, + 0.010861607268452644, + 0.002380759920924902, + 0.0031410525552928448, + 0.0031228982843458652, + 0.0034821578301489353, + 0.0047346255742013454, + 0.003394262632355094, + 0.003469337010756135, + 0.0032363352365791798, + 0.017285022884607315, + 0.009698462672531605, + 0.04585537314414978, + 0.0099330460652709, + 0.011822979897260666, + 0.046942178159952164, + 0.012909499928355217, + 0.018827220425009727, + 0.03644289821386337, + 0.008877292275428772, + 0.009071663953363895, + 0.007162970490753651, + 0.012027272023260593, + 0.016678348183631897, + 0.011745119467377663, + 0.04211151972413063, + 0.029908297583460808, + 0.027658628299832344, + 0.015435220673680305, + 0.01664775423705578, + 0.035578690469264984, + 0.04972696676850319, + 0.024885429069399834, + 0.03590041399002075, + 0.012375568971037865, + 0.022161925211548805, + 0.006339848507195711, + 0.009728909470140934, + 0.02312077395617962, + 0.028809839859604836, + 0.012698467820882797, + 0.01616300828754902, + 0.0034503706265240908, + 0.004556827712804079, + 0.0024390474427491426, + 0.0029413041193038225, + 0.0047854776494205, + 0.0071080513298511505, + 0.003805315587669611, + 0.005426458548754454, + 0.0029056943021714687, + 0.0028398644644767046, + 0.0012294907355681062, + 0.0022175395861268044, + 0.0038216542452573776, + 0.005310878623276949, + 0.002518567955121398, + 0.003708786331117153, + 0.005337505601346493, + 0.005179579835385084, + 0.0020797618199139833, + 0.0033170178066939116, + 0.004558652173727751, + 0.008553970605134964, + 0.0035011155996471643, + 0.007976016961038113, + 0.008389883674681187, + 0.012183520942926407, + 0.003087370889261365, + 0.0041563245467841625, + 0.00833108276128769, + 0.012125073932111263, + 0.005244813393801451, + 0.014841832220554352 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 2, + 3 + ], + "selected_reward": -0.0205729641020298 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a practitioner with a shaved head wearing a dark gray cotton and linen robe?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the practitioner seated before a heavy wooden Go board with his right hand hovering above it holding a black Go piece?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene lit by a strong sidelight from outside the frame that creates high contrast with deep shadows on one side of the face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera positioned in a static close-up shot from a slightly upward angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video feature a blurred dark wooden grille background with a high-contrast black and white film texture?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005947155877947807 + ], + "opposite_logprob_tokens": [ + -5.130947113037109 + ], + "logprob_reward": -0.005947155877947807, + "yes_no_margin_reward": 5.125, + "token_credit_reward": -0.005947155877947807, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09282784909009933, + 0.035225752741098404, + 0.3749307096004486, + 0.09782605618238449, + 0.051181692630052567, + 0.04921076446771622, + 0.02720210887491703, + 0.06734716147184372, + 0.029279284179210663, + 0.17496857047080994 + ], + "credit_map": [ + 0.009049776010215282, + 0.008119256235659122, + 0.008082623593509197, + 0.014714443124830723, + 0.019082583487033844, + 0.01421782374382019, + 0.008919629268348217, + 0.010641718283295631, + 0.004046260844916105, + 0.004302666522562504, + 0.0034574149176478386, + 0.003959811758249998, + 0.005571015644818544, + 0.0046118139289319515, + 0.005269330460578203, + 0.004007438663393259, + 0.0330020897090435, + 0.01738414727151394, + 0.10624364763498306, + 0.020354552194476128, + 0.02449825592339039, + 0.09649594873189926, + 0.031664613634347916, + 0.04528745263814926, + 0.024762889370322227, + 0.006321745924651623, + 0.005788256414234638, + 0.004382378421723843, + 0.008716611191630363, + 0.011572489514946938, + 0.008620726875960827, + 0.027660969644784927, + 0.006875311955809593, + 0.008828610181808472, + 0.003612650791183114, + 0.004121033009141684, + 0.008270141668617725, + 0.007994238287210464, + 0.005205002147704363, + 0.006274711340665817, + 0.005402632523328066, + 0.005833074450492859, + 0.0027388217858970165, + 0.002694932511076331, + 0.013399294577538967, + 0.0067486087791621685, + 0.0055469549261033535, + 0.00684644840657711, + 0.002239230088889599, + 0.0030498364940285683, + 0.001954395091161132, + 0.0018371540354564786, + 0.006147114560008049, + 0.004653578158468008, + 0.0029379550833255053, + 0.0043828473426401615, + 0.011317441239953041, + 0.010264348238706589, + 0.003671195125207305, + 0.005500914063304663, + 0.010424157604575157, + 0.011530105955898762, + 0.005634409841150045, + 0.009004597552120686, + 0.0033905126620084047, + 0.004213591106235981, + 0.0016702709253877401, + 0.0022003164049237967, + 0.00623165862634778, + 0.004622506909072399, + 0.0028271167539060116, + 0.004123313818126917, + 0.019592564553022385, + 0.030913591384887695, + 0.008742458187043667, + 0.010756183415651321, + 0.019233250990509987, + 0.0342443510890007, + 0.013279793784022331, + 0.03820639103651047 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 9, + 3 + ], + "selected_reward": -0.005947155877947807 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "static close-up shot from a slightly upward angle focuses on a serene practitioner seated before a heavy wooden Go board, lost in thought. His head is shaved, his features thin and angular, his eyes downcast as he gazes intently at the endgame on the board, his lips pursed in a look of intense concentration. He wears a simple, dark gray cotton and linen robe, its collar neatly tied, its folds naturally flowing as he leans forward. A strong sidelight streams in from outside the frame, sharply sculpting the curve of his head, the high bridge of his nose, and the contours of his cheeks in classic black and white, while the other side of his face is completely shrouded in deep shadow. His right hand gracefully hovers above the board, his index and middle fingers firmly holding a round, smooth black Go piece. The piece reflects a subtle luster in the light, as if about to fall, yet suspended in the air. The background is a blurred dark wooden grille, and the entire picture is extremely restrained in tone, presenting a high-contrast black and white film texture, creating a tense atmosphere of frozen time, far-reaching tranquility, and full of oriental philosophical wisdom.", + "seed": 3717, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000071/group_7/index_9153_seed_3717_dp_14_rank_28_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 7, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.06419069366529584, + "vlm_logprob": -0.06419069366529584, + "vlm_yes_no_margin": 3.525, + "vlm_token_credit": -0.06419069366529584, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a practitioner with a shaved head wearing a dark gray cotton and linen robe?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the practitioner seated before a heavy wooden Go board with his right hand hovering above it holding a black Go piece?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene lit by a strong sidelight from outside the frame that creates high contrast with deep shadows on one side of the face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera positioned in a static close-up shot from a slightly upward angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video feature a blurred dark wooden grille background with a high-contrast black and white film texture?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.12981241941452026 + ], + "opposite_logprob_tokens": [ + -2.129812479019165 + ], + "logprob_reward": -0.12981241941452026, + "yes_no_margin_reward": 2.0, + "token_credit_reward": -0.12981241941452026, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.331539124250412, + 0.1650487184524536, + 0.09020750224590302, + 0.05888253077864647, + 0.035544272512197495, + 0.07270970940589905, + 0.1116868257522583, + 0.04585638269782066, + 0.0540575310587883, + 0.0344674326479435 + ], + "credit_map": [ + 0.04960041493177414, + 0.06577374786138535, + 0.04820268228650093, + 0.040121424943208694, + 0.03431185707449913, + 0.03500118479132652, + 0.02610764093697071, + 0.03242017328739166, + 0.02449950948357582, + 0.03508970886468887, + 0.018963467329740524, + 0.027141956612467766, + 0.01350779365748167, + 0.012124237604439259, + 0.01737281307578087, + 0.016349226236343384, + 0.01411539874970913, + 0.021586839109659195, + 0.009352785535156727, + 0.011764795519411564, + 0.00831583235412836, + 0.014283064752817154, + 0.005529406014829874, + 0.005259372293949127, + 0.009729364886879921, + 0.012375087477266788, + 0.009126232005655766, + 0.004805422853678465, + 0.005512774921953678, + 0.008287065662443638, + 0.0052741155959665775, + 0.003772469237446785, + 0.004963413346558809, + 0.007966740056872368, + 0.004305969458073378, + 0.0046701086685061455, + 0.0034126616083085537, + 0.005247477442026138, + 0.002780264476314187, + 0.002197636989876628, + 0.011823452077805996, + 0.013748176395893097, + 0.008353101089596748, + 0.006200350821018219, + 0.006353709846735001, + 0.015987714752554893, + 0.006625432055443525, + 0.0036177728325128555, + 0.0160971749573946, + 0.023709215223789215, + 0.012003367766737938, + 0.011684834025800228, + 0.014227408915758133, + 0.017236802726984024, + 0.009694871492683887, + 0.007033152040094137, + 0.007648401893675327, + 0.009593609720468521, + 0.0066622234880924225, + 0.006070853676646948, + 0.00399680994451046, + 0.006318449042737484, + 0.003301402786746621, + 0.002264633309096098, + 0.009487656876444817, + 0.010241088457405567, + 0.007772854994982481, + 0.006484131794422865, + 0.004110750276595354, + 0.009350646287202835, + 0.003488909685984254, + 0.003121490590274334, + 0.007837226614356041, + 0.0062057324685156345, + 0.004499164409935474, + 0.0032559665851294994, + 0.002365682041272521, + 0.0065124728716909885, + 0.0022907014936208725, + 0.0015004831366240978 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 6 + ], + "selected_reward": -0.12981241941452026 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a practitioner with a shaved head wearing a dark gray cotton and linen robe?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the practitioner seated before a heavy wooden Go board with his right hand hovering above it holding a black Go piece?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene lit by a strong sidelight from outside the frame that creates high contrast with deep shadows on one side of the face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera positioned in a static close-up shot from a slightly upward angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video feature a blurred dark wooden grille background with a high-contrast black and white film texture?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.160425066947937 + ], + "opposite_logprob_tokens": [ + -1.910425066947937 + ], + "logprob_reward": -0.160425066947937, + "yes_no_margin_reward": 1.75, + "token_credit_reward": -0.160425066947937, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.36615777015686035, + 0.15343454480171204, + 0.1191248670220375, + 0.05243176594376564, + 0.05257638543844223, + 0.055668264627456665, + 0.09996013343334198, + 0.05643749237060547, + 0.021097198128700256, + 0.023111529648303986 + ], + "credit_map": [ + 0.04927099123597145, + 0.08093713223934174, + 0.05160379409790039, + 0.03795621916651726, + 0.029074829071760178, + 0.0386827290058136, + 0.03794672340154648, + 0.040685370564460754, + 0.023533301427960396, + 0.0305151529610157, + 0.01858687959611416, + 0.014431615360081196, + 0.016516167670488358, + 0.02004631981253624, + 0.01781209371984005, + 0.011993027292191982, + 0.016618717461824417, + 0.022035853937268257, + 0.014051945880055428, + 0.01876670867204666, + 0.010467160493135452, + 0.01843217760324478, + 0.010825841687619686, + 0.007926461286842823, + 0.009604896418750286, + 0.009069159626960754, + 0.007833600975573063, + 0.004454385954886675, + 0.004032833967357874, + 0.008522208780050278, + 0.004182848613709211, + 0.00473182974383235, + 0.007360092364251614, + 0.01056758500635624, + 0.006162751000374556, + 0.007652448955923319, + 0.005147164221853018, + 0.007213166914880276, + 0.00518491817638278, + 0.00328825949691236, + 0.009825431741774082, + 0.010287255980074406, + 0.0074684196151793, + 0.00402393564581871, + 0.0056124497205019, + 0.010941652581095695, + 0.004488703329116106, + 0.00302041694521904, + 0.014233035035431385, + 0.022107278928160667, + 0.010810213163495064, + 0.008932535536587238, + 0.01228165253996849, + 0.016762664541602135, + 0.008439339697360992, + 0.006393413990736008, + 0.008111526258289814, + 0.010176734067499638, + 0.0074266092851758, + 0.00906023383140564, + 0.004234478808939457, + 0.009532186202704906, + 0.004982420708984137, + 0.002913299947977066, + 0.003893993329256773, + 0.004039254505187273, + 0.0028325419407337904, + 0.0017510497709736228, + 0.0018824965227395296, + 0.0035020301584154367, + 0.0015463180607184768, + 0.0016495159361511469, + 0.004523777402937412, + 0.0032705296762287617, + 0.003165795700624585, + 0.003000885946676135, + 0.0018092042300850153, + 0.003969909157603979, + 0.0020102171692997217, + 0.0013612094335258007 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 2 + ], + "selected_reward": -0.160425066947937 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a practitioner with a shaved head wearing a dark gray cotton and linen robe?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the practitioner seated before a heavy wooden Go board with his right hand hovering above it holding a black Go piece?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene lit by a strong sidelight from outside the frame that creates high contrast with deep shadows on one side of the face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera positioned in a static close-up shot from a slightly upward angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video feature a blurred dark wooden grille background with a high-contrast black and white film texture?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.006738912779837847 + ], + "opposite_logprob_tokens": [ + -5.006739139556885 + ], + "logprob_reward": -0.006738912779837847, + "yes_no_margin_reward": 5.0, + "token_credit_reward": -0.006738912779837847, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09372305124998093, + 0.10776253789663315, + 0.07783401012420654, + 0.0481841117143631, + 0.039183586835861206, + 0.38610318303108215, + 0.10028126835823059, + 0.08445508033037186, + 0.03514045104384422, + 0.02733268216252327 + ], + "credit_map": [ + 0.01490716077387333, + 0.02244720235466957, + 0.011859382502734661, + 0.01032735500484705, + 0.009485282935202122, + 0.008538800291717052, + 0.008210072293877602, + 0.007947803474962711, + 0.014421967789530754, + 0.02308550290763378, + 0.010733114555478096, + 0.016734087839722633, + 0.009198332205414772, + 0.014055825769901276, + 0.00799640640616417, + 0.011537297628819942, + 0.013520962558686733, + 0.012399142608046532, + 0.01289793848991394, + 0.004804898053407669, + 0.005043818615376949, + 0.01660754531621933, + 0.006287852302193642, + 0.006271853111684322, + 0.007930396124720573, + 0.008532811887562275, + 0.008414733223617077, + 0.0041246600449085236, + 0.0032074693590402603, + 0.00916881300508976, + 0.0036432219203561544, + 0.0031620028894394636, + 0.006951515097171068, + 0.006442300975322723, + 0.0038978373631834984, + 0.004200263414531946, + 0.003531579626724124, + 0.008174783550202847, + 0.002752735512331128, + 0.0032325724605470896, + 0.06554095447063446, + 0.07228261977434158, + 0.06661400943994522, + 0.021921377629041672, + 0.027736498042941093, + 0.09560727328062057, + 0.01880098134279251, + 0.017599506303668022, + 0.016642659902572632, + 0.014807650819420815, + 0.012885155156254768, + 0.007384793367236853, + 0.010467165149748325, + 0.02322474494576454, + 0.007818741723895073, + 0.007050355896353722, + 0.015862202271819115, + 0.01565869338810444, + 0.011959289200603962, + 0.005859183147549629, + 0.006881861016154289, + 0.018786098808050156, + 0.00527431583032012, + 0.0041734385304152966, + 0.005699594039469957, + 0.006468528416007757, + 0.005347210448235273, + 0.004619483835995197, + 0.0026434152387082577, + 0.005818277131766081, + 0.002582625951617956, + 0.001961318776011467, + 0.005687386728823185, + 0.005244201514869928, + 0.0036996558774262667, + 0.002057230332866311, + 0.0019879210740327835, + 0.004956218414008617, + 0.0018846230814233422, + 0.0018154445569962263 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 1, + 6 + ], + "selected_reward": -0.006738912779837847 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a practitioner with a shaved head wearing a dark gray cotton and linen robe?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the practitioner seated before a heavy wooden Go board with his right hand hovering above it holding a black Go piece?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene lit by a strong sidelight from outside the frame that creates high contrast with deep shadows on one side of the face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera positioned in a static close-up shot from a slightly upward angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video feature a blurred dark wooden grille background with a high-contrast black and white film texture?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.014193654991686344 + ], + "opposite_logprob_tokens": [ + -4.264193534851074 + ], + "logprob_reward": -0.014193654991686344, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.014193654991686344, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15353479981422424, + 0.13487888872623444, + 0.14353643357753754, + 0.10515566915273666, + 0.10211371630430222, + 0.11859741061925888, + 0.07038340717554092, + 0.04240266978740692, + 0.06090647727251053, + 0.06849048286676407 + ], + "credit_map": [ + 0.023244788870215416, + 0.04377932474017143, + 0.019516639411449432, + 0.014786917716264725, + 0.018857205286622047, + 0.01171969622373581, + 0.014594992622733116, + 0.007035252172499895, + 0.023278212174773216, + 0.02122143656015396, + 0.022639930248260498, + 0.010348733514547348, + 0.011151236481964588, + 0.02334502339363098, + 0.010556473396718502, + 0.012337857857346535, + 0.022808045148849487, + 0.029080217704176903, + 0.025026651099324226, + 0.013926470652222633, + 0.008873164653778076, + 0.02100505121052265, + 0.013511024415493011, + 0.009305821731686592, + 0.017233876511454582, + 0.02295372448861599, + 0.015498568303883076, + 0.0097054960206151, + 0.00795629620552063, + 0.016864048317074776, + 0.008329197764396667, + 0.0066144708544015884, + 0.01633310876786709, + 0.019817223772406578, + 0.01614062301814556, + 0.009808927774429321, + 0.007500109728425741, + 0.016316590830683708, + 0.009337428025901318, + 0.006859711837023497, + 0.019507169723510742, + 0.01904180273413658, + 0.01998456008732319, + 0.009314622730016708, + 0.008268361911177635, + 0.0256943516433239, + 0.009912285953760147, + 0.006874259561300278, + 0.013439333066344261, + 0.010415516793727875, + 0.009202878922224045, + 0.004424340091645718, + 0.0045788828283548355, + 0.019167574122548103, + 0.004577923100441694, + 0.00457695871591568, + 0.008348679170012474, + 0.006568758748471737, + 0.0047855256125330925, + 0.0027119978331029415, + 0.0038466122932732105, + 0.010474937036633492, + 0.0028669778257608414, + 0.0027991877868771553, + 0.009646378457546234, + 0.014835648238658905, + 0.008718526922166348, + 0.006113634444773197, + 0.0059361946769058704, + 0.007410083897411823, + 0.004677547141909599, + 0.003568470012396574, + 0.011795955710113049, + 0.015114787966012955, + 0.010305747389793396, + 0.006082320120185614, + 0.005088465288281441, + 0.010830573737621307, + 0.004828082397580147, + 0.004444555379450321 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 1 + ], + "selected_reward": -0.014193654991686344 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a practitioner with a shaved head wearing a dark gray cotton and linen robe?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the practitioner seated before a heavy wooden Go board with his right hand hovering above it holding a black Go piece?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene lit by a strong sidelight from outside the frame that creates high contrast with deep shadows on one side of the face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera positioned in a static close-up shot from a slightly upward angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video feature a blurred dark wooden grille background with a high-contrast black and white film texture?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.00978341419249773 + ], + "opposite_logprob_tokens": [ + -4.6347832679748535 + ], + "logprob_reward": -0.00978341419249773, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.00978341419249773, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14433015882968903, + 0.19456088542938232, + 0.07825031876564026, + 0.04787101969122887, + 0.06363319605588913, + 0.1535795032978058, + 0.09205500036478043, + 0.15371577441692352, + 0.02891922928392887, + 0.04308490827679634 + ], + "credit_map": [ + 0.021252162754535675, + 0.02912549488246441, + 0.021474255248904228, + 0.01738002337515354, + 0.013387673534452915, + 0.012231685221195221, + 0.01871459186077118, + 0.010764271020889282, + 0.02988898754119873, + 0.03260141983628273, + 0.029726136475801468, + 0.027195021510124207, + 0.011699151247739792, + 0.03185540810227394, + 0.015737459063529968, + 0.01585729978978634, + 0.011591502465307713, + 0.01436605304479599, + 0.00942237675189972, + 0.010122428648173809, + 0.006782409735023975, + 0.012081421911716461, + 0.006571886595338583, + 0.007312242407351732, + 0.00739861698821187, + 0.00921840500086546, + 0.006798218935728073, + 0.00510245468467474, + 0.004319263156503439, + 0.007889281958341599, + 0.0032006290275603533, + 0.003944148775190115, + 0.009134919382631779, + 0.013941354118287563, + 0.0075564514845609665, + 0.007991161197423935, + 0.006438605021685362, + 0.008352828212082386, + 0.006821746937930584, + 0.0033961287699639797, + 0.02766457572579384, + 0.027623815461993217, + 0.024946894496679306, + 0.008480680175125599, + 0.01063220202922821, + 0.03759931027889252, + 0.009773604571819305, + 0.006858409848064184, + 0.014598974026739597, + 0.023816736415028572, + 0.01174849271774292, + 0.008952285163104534, + 0.01075056754052639, + 0.01014701183885336, + 0.0060087996535003185, + 0.00603213906288147, + 0.026912705972790718, + 0.033429041504859924, + 0.021617740392684937, + 0.012600858695805073, + 0.013478188775479794, + 0.028565797954797745, + 0.009788763709366322, + 0.007322684396058321, + 0.004786108620464802, + 0.00710652768611908, + 0.003407139563933015, + 0.003554705763235688, + 0.0021269377321004868, + 0.004434244241565466, + 0.001735606580041349, + 0.0017679608426988125, + 0.010047360323369503, + 0.0066343736834824085, + 0.0050268168561160564, + 0.003502986393868923, + 0.0028251023031771183, + 0.01025377493351698, + 0.002339758677408099, + 0.002454733941704035 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 7, + 5 + ], + "selected_reward": -0.00978341419249773 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "static close-up shot from a slightly upward angle focuses on a serene practitioner seated before a heavy wooden Go board, lost in thought. His head is shaved, his features thin and angular, his eyes downcast as he gazes intently at the endgame on the board, his lips pursed in a look of intense concentration. He wears a simple, dark gray cotton and linen robe, its collar neatly tied, its folds naturally flowing as he leans forward. A strong sidelight streams in from outside the frame, sharply sculpting the curve of his head, the high bridge of his nose, and the contours of his cheeks in classic black and white, while the other side of his face is completely shrouded in deep shadow. His right hand gracefully hovers above the board, his index and middle fingers firmly holding a round, smooth black Go piece. The piece reflects a subtle luster in the light, as if about to fall, yet suspended in the air. The background is a blurred dark wooden grille, and the entire picture is extremely restrained in tone, presenting a high-contrast black and white film texture, creating a tense atmosphere of frozen time, far-reaching tranquility, and full of oriental philosophical wisdom.", + "seed": 3718, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000071/group_7/index_9153_seed_3718_dp_15_rank_30_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 7, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.02952299462631345, + "vlm_logprob": -0.02952299462631345, + "vlm_yes_no_margin": 3.95, + "vlm_token_credit": -0.02952299462631345, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a practitioner with a shaved head wearing a dark gray cotton and linen robe?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the practitioner seated before a heavy wooden Go board with his right hand hovering above it holding a black Go piece?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene lit by a strong sidelight from outside the frame that creates high contrast with deep shadows on one side of the face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera positioned in a static close-up shot from a slightly upward angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video feature a blurred dark wooden grille background with a high-contrast black and white film texture?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.08144311606884003 + ], + "opposite_logprob_tokens": [ + -2.5814430713653564 + ], + "logprob_reward": -0.08144311606884003, + "yes_no_margin_reward": 2.5, + "token_credit_reward": -0.08144311606884003, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3574672341346741, + 0.109113909304142, + 0.05233597010374069, + 0.058944687247276306, + 0.03802606463432312, + 0.02216041274368763, + 0.023464716970920563, + 0.03207239508628845, + 0.10762998461723328, + 0.19878466427326202 + ], + "credit_map": [ + 0.05199918895959854, + 0.052135758101940155, + 0.044706929475069046, + 0.04240832477807999, + 0.05526013672351837, + 0.05111461877822876, + 0.028344305232167244, + 0.03149798512458801, + 0.012860007584095001, + 0.014409525319933891, + 0.011977382004261017, + 0.009149093180894852, + 0.017736075446009636, + 0.016563300043344498, + 0.013517539948225021, + 0.012900984846055508, + 0.004546380136162043, + 0.006227629259228706, + 0.006773320026695728, + 0.009312067180871964, + 0.00719458470121026, + 0.005317771807312965, + 0.008277243003249168, + 0.004686971660703421, + 0.007501689251512289, + 0.012166870757937431, + 0.00411418080329895, + 0.007117099594324827, + 0.00820909533649683, + 0.006075229961425066, + 0.007912077009677887, + 0.005848447326570749, + 0.0021099457517266273, + 0.002550640143454075, + 0.004080777522176504, + 0.005119650159031153, + 0.009172371588647366, + 0.003848063061013818, + 0.005922261159867048, + 0.005222356412559748, + 0.0019203542033210397, + 0.004082975443452597, + 0.002670770976692438, + 0.0027310841251164675, + 0.0031048848759382963, + 0.0030619127210229635, + 0.0021430053748190403, + 0.0024454237427562475, + 0.0018251209985464811, + 0.003736847545951605, + 0.002135711954906583, + 0.0021154717542231083, + 0.0048394398763775826, + 0.0031713948119431734, + 0.002732550958171487, + 0.002908179769292474, + 0.001906400779262185, + 0.00319901155307889, + 0.0032445634715259075, + 0.004120105877518654, + 0.008229122497141361, + 0.0033320256043225527, + 0.004657264333218336, + 0.003383901435881853, + 0.00583707494661212, + 0.016875656321644783, + 0.0102712232619524, + 0.008781353943049908, + 0.028708184137940407, + 0.011106492951512337, + 0.013757430948317051, + 0.012292568571865559, + 0.012606938369572163, + 0.034060779958963394, + 0.026115192100405693, + 0.015153621323406696, + 0.04432559385895729, + 0.022092659026384354, + 0.024483077228069305, + 0.019946793094277382 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 1 + ], + "selected_reward": -0.08144311606884003 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a practitioner with a shaved head wearing a dark gray cotton and linen robe?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the practitioner seated before a heavy wooden Go board with his right hand hovering above it holding a black Go piece?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene lit by a strong sidelight from outside the frame that creates high contrast with deep shadows on one side of the face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera positioned in a static close-up shot from a slightly upward angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video feature a blurred dark wooden grille background with a high-contrast black and white film texture?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.026363147422671318 + ], + "opposite_logprob_tokens": [ + -3.6513631343841553 + ], + "logprob_reward": -0.026363147422671318, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.026363147422671318, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.17097382247447968, + 0.14299367368221283, + 0.11246990412473679, + 0.1726655513048172, + 0.06542349606752396, + 0.07844018936157227, + 0.0403643473982811, + 0.10441768914461136, + 0.06039822846651077, + 0.05185309797525406 + ], + "credit_map": [ + 0.018254723399877548, + 0.022284623235464096, + 0.03793618455529213, + 0.014489109627902508, + 0.028213758021593094, + 0.017959775403141975, + 0.01694478467106819, + 0.014890872873365879, + 0.014314116910099983, + 0.01512480340898037, + 0.0238057728856802, + 0.013656345196068287, + 0.01863834634423256, + 0.021689806133508682, + 0.020525775849819183, + 0.015238713473081589, + 0.01665552705526352, + 0.017868682742118835, + 0.010368693619966507, + 0.009924524463713169, + 0.020550819113850594, + 0.012170130386948586, + 0.010387018322944641, + 0.014544503763318062, + 0.01975325122475624, + 0.02282637730240822, + 0.02369323931634426, + 0.02537032775580883, + 0.02473568171262741, + 0.016421308740973473, + 0.018818330019712448, + 0.021047044545412064, + 0.003675200045108795, + 0.005388943012803793, + 0.009299412369728088, + 0.012223545461893082, + 0.012042373418807983, + 0.007553427945822477, + 0.008027683943510056, + 0.007212912198156118, + 0.0054070185869932175, + 0.010982807725667953, + 0.008379113860428333, + 0.011554139666259289, + 0.016649823635816574, + 0.009321295656263828, + 0.008184341713786125, + 0.007961655966937542, + 0.005475526209920645, + 0.007117924280464649, + 0.004572639241814613, + 0.0035938292276114225, + 0.006336649879813194, + 0.004807725548744202, + 0.003920640330761671, + 0.0045394087210297585, + 0.009408089332282543, + 0.008402224630117416, + 0.014080812223255634, + 0.016739362850785255, + 0.021916644647717476, + 0.01131905522197485, + 0.011754658073186874, + 0.010796844027936459, + 0.0047460393980145454, + 0.0070747616700828075, + 0.008166716434061527, + 0.008895075879991055, + 0.008392047137022018, + 0.007763028144836426, + 0.008066069334745407, + 0.007294494658708572, + 0.003071039915084839, + 0.006467348430305719, + 0.007552696391940117, + 0.004691465757787228, + 0.012371903285384178, + 0.005713915918022394, + 0.006427735555917025, + 0.005556995514780283 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 0, + 1 + ], + "selected_reward": -0.026363147422671318 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a practitioner with a shaved head wearing a dark gray cotton and linen robe?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the practitioner seated before a heavy wooden Go board with his right hand hovering above it holding a black Go piece?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene lit by a strong sidelight from outside the frame that creates high contrast with deep shadows on one side of the face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera positioned in a static close-up shot from a slightly upward angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video feature a blurred dark wooden grille background with a high-contrast black and white film texture?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.006738557480275631 + ], + "opposite_logprob_tokens": [ + -5.006738662719727 + ], + "logprob_reward": -0.006738557480275631, + "yes_no_margin_reward": 5.0, + "token_credit_reward": -0.006738557480275631, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2758423089981079, + 0.09838534146547318, + 0.1687978059053421, + 0.22357720136642456, + 0.06245743855834007, + 0.058111924678087234, + 0.0516078844666481, + 0.022047335281968117, + 0.017226943746209145, + 0.021945776417851448 + ], + "credit_map": [ + 0.033325210213661194, + 0.032463718205690384, + 0.03367939591407776, + 0.023656800389289856, + 0.06378424912691116, + 0.03617877513170242, + 0.023122716695070267, + 0.029631463810801506, + 0.01227849442511797, + 0.017607377842068672, + 0.01391392108052969, + 0.009825441986322403, + 0.015720799565315247, + 0.01094027329236269, + 0.00791492685675621, + 0.01018410176038742, + 0.01501315925270319, + 0.03073761984705925, + 0.017593756318092346, + 0.016598664224147797, + 0.03516524285078049, + 0.0193759948015213, + 0.01759898103773594, + 0.016714390367269516, + 0.01896783523261547, + 0.03727909177541733, + 0.022094640880823135, + 0.023851880803704262, + 0.04644095152616501, + 0.023684987798333168, + 0.027207331731915474, + 0.024050477892160416, + 0.004668359644711018, + 0.007901346310973167, + 0.008105635643005371, + 0.0112750930711627, + 0.008955202996730804, + 0.00625171884894371, + 0.007643990218639374, + 0.0076560936868190765, + 0.0038833788130432367, + 0.004567414056509733, + 0.007326039951294661, + 0.00925432238727808, + 0.008492708206176758, + 0.006379854399710894, + 0.009681628085672855, + 0.008526576682925224, + 0.005912229418754578, + 0.005956416483968496, + 0.006479879375547171, + 0.006847729906439781, + 0.008145051077008247, + 0.006510379258543253, + 0.00515835452824831, + 0.006597844418138266, + 0.0027441312558948994, + 0.0030216360464692116, + 0.0030225508380681276, + 0.0022107551340013742, + 0.002719506621360779, + 0.002720488468185067, + 0.00258716382086277, + 0.0030211026314646006, + 0.002052140422165394, + 0.0017121017444878817, + 0.0017785093514248729, + 0.0019543860107660294, + 0.003429937642067671, + 0.0017845260445028543, + 0.0027341675013303757, + 0.0017811742145568132, + 0.002483314834535122, + 0.003954099956899881, + 0.0028912778943777084, + 0.002331939060240984, + 0.003441590815782547, + 0.002647346118465066, + 0.001898697461001575, + 0.0022975103929638863 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 2 + ], + "selected_reward": -0.006738557480275631 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a practitioner with a shaved head wearing a dark gray cotton and linen robe?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the practitioner seated before a heavy wooden Go board with his right hand hovering above it holding a black Go piece?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene lit by a strong sidelight from outside the frame that creates high contrast with deep shadows on one side of the face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera positioned in a static close-up shot from a slightly upward angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video feature a blurred dark wooden grille background with a high-contrast black and white film texture?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.026336094364523888 + ], + "opposite_logprob_tokens": [ + -3.651336193084717 + ], + "logprob_reward": -0.026336094364523888, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.026336094364523888, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15484952926635742, + 0.11774298548698425, + 0.14484748244285583, + 0.15473662316799164, + 0.15354900062084198, + 0.07125242799520493, + 0.07888869941234589, + 0.05494236573576927, + 0.049674537032842636, + 0.019516337662935257 + ], + "credit_map": [ + 0.014722511172294617, + 0.016188202425837517, + 0.03351428359746933, + 0.01576228253543377, + 0.024131182581186295, + 0.020779047161340714, + 0.013510597869753838, + 0.016241416335105896, + 0.007494078949093819, + 0.010884652845561504, + 0.007368098478764296, + 0.010514851659536362, + 0.030300738289952278, + 0.0173515472561121, + 0.017903435975313187, + 0.015925582498311996, + 0.007868703454732895, + 0.020533936098217964, + 0.021615823730826378, + 0.02048596180975437, + 0.026420921087265015, + 0.014925358816981316, + 0.019438913092017174, + 0.01355786994099617, + 0.012895331718027592, + 0.02271067537367344, + 0.012286216951906681, + 0.018556812778115273, + 0.03513192757964134, + 0.015327648259699345, + 0.0211984571069479, + 0.01662956364452839, + 0.010702081955969334, + 0.020709361881017685, + 0.012073738500475883, + 0.017652591690421104, + 0.03895042464137077, + 0.01815214194357395, + 0.01886555552482605, + 0.016443097963929176, + 0.0040494403801858425, + 0.0056710923090577126, + 0.01093997247517109, + 0.014968441799283028, + 0.012060300447046757, + 0.006642760243266821, + 0.007749106269329786, + 0.009171314537525177, + 0.008549700491130352, + 0.014496581628918648, + 0.006726033985614777, + 0.007803245447576046, + 0.014780152589082718, + 0.008775947615504265, + 0.008666638284921646, + 0.009090402163565159, + 0.006722663063555956, + 0.007078441325575113, + 0.00558315496891737, + 0.005059679504483938, + 0.00826281402260065, + 0.007282686419785023, + 0.007179608102887869, + 0.00777331693097949, + 0.0031630906742066145, + 0.007116814609616995, + 0.005390886217355728, + 0.00442401971668005, + 0.012761738151311874, + 0.005531057249754667, + 0.006512950174510479, + 0.00477397907525301, + 0.002799921901896596, + 0.001610103528946638, + 0.0031617358326911926, + 0.002282763598486781, + 0.002964845858514309, + 0.0021086372435092926, + 0.0025951662100851536, + 0.001993163488805294 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 4 + ], + "selected_reward": -0.026336094364523888 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a practitioner with a shaved head wearing a dark gray cotton and linen robe?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the practitioner seated before a heavy wooden Go board with his right hand hovering above it holding a black Go piece?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene lit by a strong sidelight from outside the frame that creates high contrast with deep shadows on one side of the face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera positioned in a static close-up shot from a slightly upward angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video feature a blurred dark wooden grille background with a high-contrast black and white film texture?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.006734057795256376 + ], + "opposite_logprob_tokens": [ + -5.0067338943481445 + ], + "logprob_reward": -0.006734057795256376, + "yes_no_margin_reward": 5.0, + "token_credit_reward": -0.006734057795256376, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15661007165908813, + 0.15906061232089996, + 0.0770544782280922, + 0.08672337979078293, + 0.052450165152549744, + 0.09629737585783005, + 0.15867651998996735, + 0.12295909225940704, + 0.024050986394286156, + 0.0661173164844513 + ], + "credit_map": [ + 0.024077516049146652, + 0.013267581351101398, + 0.020215824246406555, + 0.020201820880174637, + 0.024589113891124725, + 0.019450565800070763, + 0.018795683979988098, + 0.016011977568268776, + 0.022958016023039818, + 0.025356022641062737, + 0.020637081936001778, + 0.01650898903608322, + 0.021644730120897293, + 0.021244633942842484, + 0.01518448255956173, + 0.015526656061410904, + 0.005712129641324282, + 0.012137720361351967, + 0.0079354802146554, + 0.00946537870913744, + 0.014387326315045357, + 0.0085952403023839, + 0.009985349141061306, + 0.008835854940116405, + 0.00857686810195446, + 0.014680932275950909, + 0.00880366750061512, + 0.008996328338980675, + 0.018121959641575813, + 0.008566715754568577, + 0.01010593306273222, + 0.008870975114405155, + 0.007404264993965626, + 0.006774041801691055, + 0.00725861731916666, + 0.005784871056675911, + 0.008667698130011559, + 0.006032225675880909, + 0.004945798777043819, + 0.005582645535469055, + 0.011414836160838604, + 0.016258765012025833, + 0.0060143549926579, + 0.009394842199981213, + 0.016867980360984802, + 0.014101659879088402, + 0.009943562559783459, + 0.012301379814743996, + 0.01766609214246273, + 0.02637975849211216, + 0.014585345983505249, + 0.014124001376330853, + 0.03430692106485367, + 0.018279101699590683, + 0.016471372917294502, + 0.016863925382494926, + 0.01347088161855936, + 0.017207004129886627, + 0.011893038637936115, + 0.011897452175617218, + 0.026741856709122658, + 0.013923048041760921, + 0.014027808792889118, + 0.013797999359667301, + 0.00304445275105536, + 0.001656325999647379, + 0.0035739329177886248, + 0.003977365326136351, + 0.0036028572358191013, + 0.002566651673987508, + 0.002643087413161993, + 0.002986311912536621, + 0.004244115669280291, + 0.010693666525185108, + 0.010780402459204197, + 0.007451628800481558, + 0.012395327910780907, + 0.007346866186708212, + 0.006389488000422716, + 0.006815820001065731 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 6, + 0 + ], + "selected_reward": -0.006734057795256376 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "static close-up shot from a slightly upward angle focuses on a serene practitioner seated before a heavy wooden Go board, lost in thought. His head is shaved, his features thin and angular, his eyes downcast as he gazes intently at the endgame on the board, his lips pursed in a look of intense concentration. He wears a simple, dark gray cotton and linen robe, its collar neatly tied, its folds naturally flowing as he leans forward. A strong sidelight streams in from outside the frame, sharply sculpting the curve of his head, the high bridge of his nose, and the contours of his cheeks in classic black and white, while the other side of his face is completely shrouded in deep shadow. His right hand gracefully hovers above the board, his index and middle fingers firmly holding a round, smooth black Go piece. The piece reflects a subtle luster in the light, as if about to fall, yet suspended in the air. The background is a blurred dark wooden grille, and the entire picture is extremely restrained in tone, presenting a high-contrast black and white film texture, creating a tense atmosphere of frozen time, far-reaching tranquility, and full of oriental philosophical wisdom.", + "seed": 3719, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000071/group_7/index_9153_seed_3719_dp_15_rank_30_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 7, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.021615790389478206, + "vlm_logprob": -0.021615790389478206, + "vlm_yes_no_margin": 3.95, + "vlm_token_credit": -0.021615790389478206, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a practitioner with a shaved head wearing a dark gray cotton and linen robe?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the practitioner seated before a heavy wooden Go board with his right hand hovering above it holding a black Go piece?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene lit by a strong sidelight from outside the frame that creates high contrast with deep shadows on one side of the face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera positioned in a static close-up shot from a slightly upward angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video feature a blurred dark wooden grille background with a high-contrast black and white film texture?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03540075570344925 + ], + "opposite_logprob_tokens": [ + -3.410400867462158 + ], + "logprob_reward": -0.03540075570344925, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.03540075570344925, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.30824118852615356, + 0.057883407920598984, + 0.07880773395299911, + 0.02519315481185913, + 0.06004207953810692, + 0.031005728989839554, + 0.04964200407266617, + 0.09093581885099411, + 0.2861643433570862, + 0.012084612622857094 + ], + "credit_map": [ + 0.011081530712544918, + 0.010322876274585724, + 0.015920883044600487, + 0.02047397941350937, + 0.030705614015460014, + 0.030667588114738464, + 0.11621541529893875, + 0.07285327464342117, + 0.00581572437658906, + 0.0030745205003768206, + 0.003497119527310133, + 0.004748447332531214, + 0.006339179817587137, + 0.008466657251119614, + 0.011001614853739738, + 0.014940140768885612, + 0.006456472910940647, + 0.004998861812055111, + 0.004931539762765169, + 0.007037722505629063, + 0.010579446330666542, + 0.006459504831582308, + 0.0180835984647274, + 0.020260576158761978, + 0.0019597704522311687, + 0.0011471317848190665, + 0.0016685276059433818, + 0.0016898504691198468, + 0.0037359450943768024, + 0.0037464944180101156, + 0.004845757503062487, + 0.006399673875421286, + 0.006875521503388882, + 0.003424170194193721, + 0.0034674149937927723, + 0.002585154492408037, + 0.004048840142786503, + 0.005085131153464317, + 0.014396598562598228, + 0.020159244537353516, + 0.001594873028807342, + 0.0020164046436548233, + 0.003123590024188161, + 0.003000810742378235, + 0.004249905701726675, + 0.003225839463993907, + 0.007403775583952665, + 0.0063905268907547, + 0.00509432191029191, + 0.0027872996870428324, + 0.0025250406470149755, + 0.005200015846639872, + 0.006078758276998997, + 0.005490821320563555, + 0.012400884181261063, + 0.01006485614925623, + 0.010787509381771088, + 0.007136008236557245, + 0.00590480724349618, + 0.008912176825106144, + 0.01746421307325363, + 0.008430907502770424, + 0.014154564589262009, + 0.018145622685551643, + 0.024515390396118164, + 0.014685283415019512, + 0.015529421158134937, + 0.02558516524732113, + 0.07016058266162872, + 0.031036820262670517, + 0.04953058063983917, + 0.055121053010225296, + 0.0009112592088058591, + 0.0006326206494122744, + 0.000746783975046128, + 0.001013358705677092, + 0.00348107167519629, + 0.0008898490341380239, + 0.0021910744253546, + 0.0022185936104506254 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 7 + ], + "selected_reward": -0.03540075570344925 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a practitioner with a shaved head wearing a dark gray cotton and linen robe?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the practitioner seated before a heavy wooden Go board with his right hand hovering above it holding a black Go piece?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene lit by a strong sidelight from outside the frame that creates high contrast with deep shadows on one side of the face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera positioned in a static close-up shot from a slightly upward angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video feature a blurred dark wooden grille background with a high-contrast black and white film texture?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.029805611819028854 + ], + "opposite_logprob_tokens": [ + -3.5298056602478027 + ], + "logprob_reward": -0.029805611819028854, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.029805611819028854, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.4861304759979248, + 0.09528736025094986, + 0.06087952107191086, + 0.03395109996199608, + 0.05491276830434799, + 0.08990608155727386, + 0.05141991004347801, + 0.026943175122141838, + 0.04074019193649292, + 0.059829454869031906 + ], + "credit_map": [ + 0.014546611346304417, + 0.01730393059551716, + 0.022129708901047707, + 0.03204364329576492, + 0.04770486056804657, + 0.04626516252756119, + 0.1874357908964157, + 0.11870077252388, + 0.010940619744360447, + 0.0044961776584386826, + 0.009199606254696846, + 0.010932695120573044, + 0.015080752782523632, + 0.010914265178143978, + 0.016341913491487503, + 0.0173813309520483, + 0.003956967499107122, + 0.004049768205732107, + 0.004048631060868502, + 0.004520393908023834, + 0.009013465605676174, + 0.006718486547470093, + 0.014711673371493816, + 0.013860132545232773, + 0.0018399746622890234, + 0.0015832510543987155, + 0.0029449528083205223, + 0.0029235605616122484, + 0.007677045185118914, + 0.003311715554445982, + 0.007597676943987608, + 0.0060729216784238815, + 0.005747233983129263, + 0.0035199904814362526, + 0.0042661456391215324, + 0.002970437053591013, + 0.004500190727412701, + 0.004424452316015959, + 0.012591274455189705, + 0.016893042251467705, + 0.005093846004456282, + 0.00519738020375371, + 0.01143072359263897, + 0.009370915591716766, + 0.013167217373847961, + 0.010985691100358963, + 0.02004106342792511, + 0.014619244262576103, + 0.006588761694729328, + 0.002756361151114106, + 0.0031561003997921944, + 0.0031039256136864424, + 0.005112315993756056, + 0.0028908781241625547, + 0.011954272165894508, + 0.01585729420185089, + 0.003384919837117195, + 0.0015384325524792075, + 0.0015552568947896361, + 0.002105914754793048, + 0.0029308449011296034, + 0.0018486505141481757, + 0.00586526095867157, + 0.007713897153735161, + 0.003998656757175922, + 0.0018077994463965297, + 0.002746875397861004, + 0.0032099976669996977, + 0.007158850785344839, + 0.0029154527001082897, + 0.008366340771317482, + 0.010536217130720615, + 0.004156854469329119, + 0.0030479563865810633, + 0.004040750674903393, + 0.01042240485548973, + 0.009619541466236115, + 0.003627532860264182, + 0.013552513904869556, + 0.011361897923052311 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 5 + ], + "selected_reward": -0.029805611819028854 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a practitioner with a shaved head wearing a dark gray cotton and linen robe?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the practitioner seated before a heavy wooden Go board with his right hand hovering above it holding a black Go piece?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene lit by a strong sidelight from outside the frame that creates high contrast with deep shadows on one side of the face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera positioned in a static close-up shot from a slightly upward angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video feature a blurred dark wooden grille background with a high-contrast black and white film texture?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.016061680391430855 + ], + "opposite_logprob_tokens": [ + -4.141061782836914 + ], + "logprob_reward": -0.016061680391430855, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.016061680391430855, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3494178056716919, + 0.14706696569919586, + 0.11791199445724487, + 0.04973584786057472, + 0.03808803856372833, + 0.033527713268995285, + 0.06262906640768051, + 0.04697469249367714, + 0.0640794187784195, + 0.09056850522756577 + ], + "credit_map": [ + 0.013394623063504696, + 0.01395296398550272, + 0.018581615760922432, + 0.02368905395269394, + 0.04509345069527626, + 0.04763183370232582, + 0.11078197509050369, + 0.07629227638244629, + 0.01800900138914585, + 0.007050402462482452, + 0.011655071750283241, + 0.011387206614017487, + 0.018321963027119637, + 0.028060538694262505, + 0.025694619864225388, + 0.02688813954591751, + 0.0062319934368133545, + 0.0061670891009271145, + 0.009535863995552063, + 0.009196191094815731, + 0.017913803458213806, + 0.011439699679613113, + 0.029760826379060745, + 0.02766651287674904, + 0.006444266065955162, + 0.0027340769302099943, + 0.004058484919369221, + 0.004251665901392698, + 0.006100013852119446, + 0.0059434352442622185, + 0.009030644781887531, + 0.011173255741596222, + 0.0030240651685744524, + 0.0029834629967808723, + 0.003057195572182536, + 0.00297999894246459, + 0.005220013204962015, + 0.008629374206066132, + 0.006098253186792135, + 0.006095671560615301, + 0.002038731472566724, + 0.002888350747525692, + 0.0034748713951557875, + 0.00411266228184104, + 0.003179360181093216, + 0.003721792483702302, + 0.007296601310372353, + 0.006815338972955942, + 0.0077882930636405945, + 0.003890681080520153, + 0.0037567296531051397, + 0.0038613276556134224, + 0.006435479037463665, + 0.003948679193854332, + 0.013906771317124367, + 0.019041098654270172, + 0.006311577279120684, + 0.004065642599016428, + 0.0037959939800202847, + 0.005978316068649292, + 0.006672104354947805, + 0.004311087541282177, + 0.008919388987123966, + 0.006920573767274618, + 0.004892288241535425, + 0.003225010586902499, + 0.0033485309686511755, + 0.00850594975054264, + 0.01469501107931137, + 0.004918507765978575, + 0.012210876680910587, + 0.01228323858231306, + 0.011577417142689228, + 0.004959478974342346, + 0.004972432740032673, + 0.01106235757470131, + 0.011179357767105103, + 0.005785600747913122, + 0.020855683833360672, + 0.020176172256469727 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 2 + ], + "selected_reward": -0.016061680391430855 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a practitioner with a shaved head wearing a dark gray cotton and linen robe?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the practitioner seated before a heavy wooden Go board with his right hand hovering above it holding a black Go piece?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene lit by a strong sidelight from outside the frame that creates high contrast with deep shadows on one side of the face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera positioned in a static close-up shot from a slightly upward angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video feature a blurred dark wooden grille background with a high-contrast black and white film texture?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.018177354708313942 + ], + "opposite_logprob_tokens": [ + -4.018177509307861 + ], + "logprob_reward": -0.018177354708313942, + "yes_no_margin_reward": 4.0, + "token_credit_reward": -0.018177354708313942, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3054676055908203, + 0.1073012724518776, + 0.07551160454750061, + 0.06307142972946167, + 0.16880103945732117, + 0.04853687807917595, + 0.08163127303123474, + 0.04880597069859505, + 0.05623951181769371, + 0.04463343322277069 + ], + "credit_map": [ + 0.01052402425557375, + 0.0171195175498724, + 0.016234269365668297, + 0.02556649036705494, + 0.03626561909914017, + 0.03487187251448631, + 0.09915772825479507, + 0.0657280832529068, + 0.010353375226259232, + 0.008486179634928703, + 0.00908090639859438, + 0.008594968356192112, + 0.016919918358325958, + 0.015650063753128052, + 0.019901171326637268, + 0.01831468939781189, + 0.008092709816992283, + 0.004328451119363308, + 0.004114903509616852, + 0.005407420918345451, + 0.012683256529271603, + 0.007894800044596195, + 0.016370641067624092, + 0.016619417816400528, + 0.0039783199317753315, + 0.004577233921736479, + 0.004502373747527599, + 0.0047139995731413364, + 0.010639575310051441, + 0.00960498210042715, + 0.012364935129880905, + 0.012690005823969841, + 0.010970573872327805, + 0.012593074701726437, + 0.013285831548273563, + 0.012890494428575039, + 0.02485726773738861, + 0.020836930721998215, + 0.03263803943991661, + 0.04072882980108261, + 0.0035233786329627037, + 0.003714130725711584, + 0.0057910410687327385, + 0.0054791159927845, + 0.005704173352569342, + 0.005307934712618589, + 0.00988042913377285, + 0.009136677719652653, + 0.008013041689991951, + 0.0036648926325142384, + 0.0040661245584487915, + 0.005244085565209389, + 0.011170043610036373, + 0.004942276980727911, + 0.020059462636709213, + 0.024471350014209747, + 0.0026690950617194176, + 0.004477557726204395, + 0.0028812051750719547, + 0.005151880439370871, + 0.01081792637705803, + 0.007404747419059277, + 0.007456949446350336, + 0.007946610450744629, + 0.006838040892034769, + 0.003317491849884391, + 0.0038674853276461363, + 0.004817062523216009, + 0.009951887652277946, + 0.004784982651472092, + 0.011061187833547592, + 0.011601371690630913, + 0.005371119827032089, + 0.00223142234608531, + 0.002229232806712389, + 0.004902917426079512, + 0.007676994893699884, + 0.003108767094090581, + 0.010074520483613014, + 0.009038456715643406 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 1 + ], + "selected_reward": -0.018177354708313942 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a practitioner with a shaved head wearing a dark gray cotton and linen robe?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the practitioner seated before a heavy wooden Go board with his right hand hovering above it holding a black Go piece?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene lit by a strong sidelight from outside the frame that creates high contrast with deep shadows on one side of the face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera positioned in a static close-up shot from a slightly upward angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video feature a blurred dark wooden grille background with a high-contrast black and white film texture?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.008633549325168133 + ], + "opposite_logprob_tokens": [ + -4.758633613586426 + ], + "logprob_reward": -0.008633549325168133, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.008633549325168133, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1558944284915924, + 0.10711343586444855, + 0.06874044984579086, + 0.03464815765619278, + 0.10717348754405975, + 0.11770763993263245, + 0.06696399301290512, + 0.13864655792713165, + 0.060683269053697586, + 0.1424284726381302 + ], + "credit_map": [ + 0.018346479162573814, + 0.013412436470389366, + 0.010274720378220081, + 0.010874440893530846, + 0.0178520567715168, + 0.025099726393818855, + 0.03595742583274841, + 0.0240771621465683, + 0.01662033423781395, + 0.0054711890406906605, + 0.010447870008647442, + 0.009725592099130154, + 0.01107542309910059, + 0.016158347949385643, + 0.01882202737033367, + 0.018792664632201195, + 0.006934401113539934, + 0.00399739108979702, + 0.005165349692106247, + 0.006164097227156162, + 0.00939958542585373, + 0.007016908843070269, + 0.016932498663663864, + 0.013130228035151958, + 0.004951003938913345, + 0.0022765498142689466, + 0.002231704071164131, + 0.003245995147153735, + 0.005479850340634584, + 0.004543972667306662, + 0.006376190111041069, + 0.005542892962694168, + 0.011052149347960949, + 0.007190093398094177, + 0.004586704540997744, + 0.003853030502796173, + 0.006252006161957979, + 0.009956831112504005, + 0.02718067541718483, + 0.03710200637578964, + 0.012485024519264698, + 0.004931304138153791, + 0.007057882845401764, + 0.006825231947004795, + 0.009580257348716259, + 0.008236930705606937, + 0.03069128841161728, + 0.03789972886443138, + 0.008908487856388092, + 0.003868312807753682, + 0.004293442703783512, + 0.0068650334142148495, + 0.008038299158215523, + 0.005637279246002436, + 0.015419828705489635, + 0.013933317735791206, + 0.017884789034724236, + 0.009221941232681274, + 0.009346903301775455, + 0.01838766224682331, + 0.019947689026594162, + 0.008791189640760422, + 0.029846834018826485, + 0.02521957829594612, + 0.00678299693390727, + 0.004047947470098734, + 0.003526558168232441, + 0.005222947336733341, + 0.014734240248799324, + 0.005847295746207237, + 0.009837367571890354, + 0.010683920234441757, + 0.022626196965575218, + 0.011329240165650845, + 0.007934330962598324, + 0.018628448247909546, + 0.016315309330821037, + 0.01083201915025711, + 0.02607096917927265, + 0.028691979125142097 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 7 + ], + "selected_reward": -0.008633549325168133 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000071/group_8/index_tensor([6659, 6659])_group_8_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000071/group_8/index_tensor([6659, 6659])_group_8_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..69fce4d9cfb4e3f9b90e6eb9e91d361b8101b073 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000071/group_8/index_tensor([6659, 6659])_group_8_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "stunning close-up, captured with a shallow depth of field, captures a single, shimmering teardrop. Soft, diffuse light emanates as it slowly and heavily slides down the delicate skin, leaving a shimmering, wet trail across the cheek. Following the tear's path, the camera tilts upward with an incredibly slow and smooth motion, closely tracking the path of this sorrowful drop, finally bringing its focus to its source: a deep brown eye, brimming with tears and slightly tinged with red. Wet eyelashes are distinctly woven together, reflecting an endless sadness and vulnerability. The entire scene, shrouded in a tranquil yet emotionally charged atmosphere and rendered in soft, natural tones, creates a powerful visual and emotional impact.", + "seed": 31296, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000071/group_8/index_6659_seed_31296_dp_16_rank_32_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 8, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.036183634866029026, + "vlm_logprob": -0.036183634866029026, + "vlm_yes_no_margin": 3.925, + "vlm_token_credit": -0.036183634866029026, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the primary subject a single, shimmering teardrop sliding down skin?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the teardrop leave a visible shimmering, wet trail across the cheek?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the camera tilting upward to track the tear's path and focus on a deep brown eye?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the eyelashes visible as wet and distinctly woven together?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene rendered with a shallow depth of field and soft, diffuse lighting?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.038595203310251236 + ], + "opposite_logprob_tokens": [ + -3.288595199584961 + ], + "logprob_reward": -0.038595203310251236, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.038595203310251236, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15824316442012787, + 0.17981767654418945, + 0.08127535134553909, + 0.051949940621852875, + 0.02829771675169468, + 0.013997398316860199, + 0.05233033001422882, + 0.258402943611145, + 0.07309510558843613, + 0.10259032249450684 + ], + "credit_map": [ + 0.01174376904964447, + 0.021384861320257187, + 0.010264263488352299, + 0.02132863737642765, + 0.02184440940618515, + 0.017153169959783554, + 0.02745438553392887, + 0.02706965245306492, + 0.008541682735085487, + 0.048284854739904404, + 0.010953878983855247, + 0.05758575722575188, + 0.01628824695944786, + 0.012612082064151764, + 0.017757894471287727, + 0.007793275639414787, + 0.006408844143152237, + 0.012069784104824066, + 0.008453586138784885, + 0.02416176348924637, + 0.007964665070176125, + 0.006999308243393898, + 0.009052191860973835, + 0.006165208760648966, + 0.004263447131961584, + 0.006050082854926586, + 0.005390666425228119, + 0.007214875426143408, + 0.01043718308210373, + 0.005985700991004705, + 0.006553547456860542, + 0.006054435390979052, + 0.0024421080015599728, + 0.0038202626165002584, + 0.002683930564671755, + 0.008118568919599056, + 0.0024056266993284225, + 0.0020944057032465935, + 0.003920873627066612, + 0.0028119401540607214, + 0.0011846530251204967, + 0.0015591499395668507, + 0.0015864706365391612, + 0.002536690328270197, + 0.0017771933926269412, + 0.0012032657396048307, + 0.0022123330272734165, + 0.0019376426935195923, + 0.003984255250543356, + 0.007036426570266485, + 0.012482232414186, + 0.010354249738156796, + 0.0071216668002307415, + 0.0034509259276092052, + 0.005033681634813547, + 0.002866892609745264, + 0.023297881707549095, + 0.04206129163503647, + 0.025980155915021896, + 0.051736027002334595, + 0.023877574130892754, + 0.04073987901210785, + 0.028744494542479515, + 0.02196563594043255, + 0.003948331344872713, + 0.003552222391590476, + 0.014703520573675632, + 0.016665421426296234, + 0.00392760057002306, + 0.006252992432564497, + 0.01782785728573799, + 0.006217161659151316, + 0.009796669706702232, + 0.004876826889812946, + 0.03463773429393768, + 0.0126196863129735, + 0.003639081260189414, + 0.006008410360664129, + 0.01226380467414856, + 0.018748106434941292 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 1, + 0 + ], + "selected_reward": -0.038595203310251236 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the primary subject a single, shimmering teardrop sliding down skin?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the teardrop leave a visible shimmering, wet trail across the cheek?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the camera tilting upward to track the tear's path and focus on a deep brown eye?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the eyelashes visible as wet and distinctly woven together?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene rendered with a shallow depth of field and soft, diffuse lighting?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01608314923942089 + ], + "opposite_logprob_tokens": [ + -4.141083240509033 + ], + "logprob_reward": -0.01608314923942089, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.01608314923942089, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.21311137080192566, + 0.2048415243625641, + 0.0635785162448883, + 0.028602970764040947, + 0.030821245163679123, + 0.01986769028007984, + 0.10747978091239929, + 0.17232078313827515, + 0.06605330854654312, + 0.09332272410392761 + ], + "credit_map": [ + 0.01713130995631218, + 0.03216832876205444, + 0.013570606708526611, + 0.027838053181767464, + 0.0306166410446167, + 0.026371661573648453, + 0.02994038350880146, + 0.035474419593811035, + 0.01159729715436697, + 0.05314759910106659, + 0.014048861339688301, + 0.06257060170173645, + 0.017952056601643562, + 0.01342079509049654, + 0.02252788469195366, + 0.009576459415256977, + 0.006088638212531805, + 0.006676753982901573, + 0.005790043156594038, + 0.01637304574251175, + 0.0071641686372458935, + 0.004962685517966747, + 0.010824916884303093, + 0.00569827388972044, + 0.002222129376605153, + 0.0030826323200017214, + 0.003773333039134741, + 0.0036086176987737417, + 0.005834548268467188, + 0.003919071517884731, + 0.0032851416617631912, + 0.0028774989768862724, + 0.002658010460436344, + 0.004042788874357939, + 0.0032374172005802393, + 0.0063133928924798965, + 0.002874396974220872, + 0.0028675170615315437, + 0.005364852957427502, + 0.0034628724679350853, + 0.0015706818085163832, + 0.0021811083424836397, + 0.002352684736251831, + 0.0031149336136877537, + 0.0021064348984509706, + 0.0021927114576101303, + 0.0026456841733306646, + 0.003703455440700054, + 0.007956583984196186, + 0.01423262245953083, + 0.02585708536207676, + 0.02330591157078743, + 0.014861018396914005, + 0.007239852100610733, + 0.008254961110651493, + 0.005771765485405922, + 0.01345651876181364, + 0.026909200474619865, + 0.01624898985028267, + 0.03738592937588692, + 0.02050887979567051, + 0.020895909518003464, + 0.019105233252048492, + 0.017810141667723656, + 0.005584734957665205, + 0.005867388099431992, + 0.008123103529214859, + 0.011058053001761436, + 0.004827625583857298, + 0.0073751164600253105, + 0.015765056014060974, + 0.007452238351106644, + 0.008890085853636265, + 0.005125665105879307, + 0.03097815066576004, + 0.011441133916378021, + 0.0030934764072299004, + 0.005470662843436003, + 0.010504373349249363, + 0.017819181084632874 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 7 + ], + "selected_reward": -0.01608314923942089 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the primary subject a single, shimmering teardrop sliding down skin?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the teardrop leave a visible shimmering, wet trail across the cheek?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the camera tilting upward to track the tear's path and focus on a deep brown eye?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the eyelashes visible as wet and distinctly woven together?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene rendered with a shallow depth of field and soft, diffuse lighting?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.008649149909615517 + ], + "opposite_logprob_tokens": [ + -4.7586493492126465 + ], + "logprob_reward": -0.008649149909615517, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.008649149909615517, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.18939919769763947, + 0.10596944391727448, + 0.16291867196559906, + 0.04891540855169296, + 0.037757281213998795, + 0.021690169349312782, + 0.07221255451440811, + 0.1327488273382187, + 0.14645877480506897, + 0.08192964643239975 + ], + "credit_map": [ + 0.01704176887869835, + 0.02706199884414673, + 0.011633027344942093, + 0.026678437367081642, + 0.029264619573950768, + 0.02083119936287403, + 0.027551140636205673, + 0.029337020590901375, + 0.009182177484035492, + 0.013274826109409332, + 0.008817318826913834, + 0.012564899399876595, + 0.01821465604007244, + 0.019860057160258293, + 0.014136334881186485, + 0.00991919357329607, + 0.015969501808285713, + 0.02346842549741268, + 0.013654442504048347, + 0.0238717719912529, + 0.025782082229852676, + 0.015780631452798843, + 0.025822404772043228, + 0.018569424748420715, + 0.004568027798086405, + 0.00541729386895895, + 0.00547550106421113, + 0.007236655335873365, + 0.007080872543156147, + 0.006010506767779589, + 0.005281842779368162, + 0.0078447125852108, + 0.003261927282437682, + 0.004420022014528513, + 0.0028327691834419966, + 0.006697606761008501, + 0.003872854635119438, + 0.003657161956652999, + 0.008071857504546642, + 0.004943085368722677, + 0.0025554297026246786, + 0.0032503774855285883, + 0.0029652609955519438, + 0.0034061011392623186, + 0.0022539659403264523, + 0.0014876035274937749, + 0.0037363239098340273, + 0.0020351074635982513, + 0.006139727309346199, + 0.016174795106053352, + 0.009897969663143158, + 0.013476584106683731, + 0.010456484742462635, + 0.0054603926837444305, + 0.004560816567391157, + 0.006045794114470482, + 0.011129973456263542, + 0.02332586795091629, + 0.015452519990503788, + 0.018626462668180466, + 0.011728635057806969, + 0.02358713373541832, + 0.015111083164811134, + 0.013787158764898777, + 0.007675024680793285, + 0.007210311014205217, + 0.03886906057596207, + 0.028811754658818245, + 0.005858088377863169, + 0.008991130627691746, + 0.02919968031346798, + 0.01984373666346073, + 0.007629259955137968, + 0.005054863169789314, + 0.026167791336774826, + 0.011666319333016872, + 0.004783932585269213, + 0.0059497603215277195, + 0.004411216359585524, + 0.016266509890556335 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 8 + ], + "selected_reward": -0.008649149909615517 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the primary subject a single, shimmering teardrop sliding down skin?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the teardrop leave a visible shimmering, wet trail across the cheek?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the camera tilting upward to track the tear's path and focus on a deep brown eye?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the eyelashes visible as wet and distinctly woven together?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene rendered with a shallow depth of field and soft, diffuse lighting?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.11295204609632492 + ], + "opposite_logprob_tokens": [ + -2.2379519939422607 + ], + "logprob_reward": -0.11295204609632492, + "yes_no_margin_reward": 2.125, + "token_credit_reward": -0.11295204609632492, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12269021570682526, + 0.11642738431692123, + 0.04504147917032242, + 0.06677503138780594, + 0.028292063623666763, + 0.014161772094666958, + 0.1175897940993309, + 0.05816938355565071, + 0.22314880788326263, + 0.20770405232906342 + ], + "credit_map": [ + 0.01377082522958517, + 0.02067553624510765, + 0.008331412449479103, + 0.019065456464886665, + 0.013980189338326454, + 0.012942672707140446, + 0.01650487445294857, + 0.017419248819351196, + 0.009464585222303867, + 0.018644750118255615, + 0.008680685423314571, + 0.019463729113340378, + 0.01789885386824608, + 0.01636836864054203, + 0.01717987284064293, + 0.008726540952920914, + 0.0048577855341136456, + 0.005884206388145685, + 0.005778297316282988, + 0.006495020817965269, + 0.007864033803343773, + 0.0043560657650232315, + 0.005895392969250679, + 0.003910674247890711, + 0.005650182254612446, + 0.00696380203589797, + 0.011733124032616615, + 0.010473656468093395, + 0.010577970184385777, + 0.006997296586632729, + 0.006862984504550695, + 0.007516023237258196, + 0.002320745261386037, + 0.0032316287979483604, + 0.002708602463826537, + 0.004631904885172844, + 0.0029168950859457254, + 0.0019558609928935766, + 0.007321698125451803, + 0.0032047266140580177, + 0.001161528518423438, + 0.0018050629878416657, + 0.0015369936591014266, + 0.0022781144361943007, + 0.0016706816386431456, + 0.0016273792134597898, + 0.0014297559391707182, + 0.0026522569824010134, + 0.009759342297911644, + 0.01952105015516281, + 0.01902364380657673, + 0.025346754118800163, + 0.01790483109652996, + 0.0088737728074193, + 0.009622708894312382, + 0.007537702564150095, + 0.004654424265027046, + 0.007916484028100967, + 0.006209272425621748, + 0.012986039742827415, + 0.005782238207757473, + 0.0075964187271893024, + 0.007109291385859251, + 0.005915218032896519, + 0.012290248647332191, + 0.012820608913898468, + 0.055908456444740295, + 0.0442052036523819, + 0.009439491666853428, + 0.013045824132859707, + 0.04761214554309845, + 0.027826843783259392, + 0.01836312748491764, + 0.01919255591928959, + 0.05333241447806358, + 0.027778344228863716, + 0.014986122027039528, + 0.014806795865297318, + 0.029233958572149277, + 0.03001072071492672 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 9, + 0 + ], + "selected_reward": -0.11295204609632492 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the primary subject a single, shimmering teardrop sliding down skin?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the teardrop leave a visible shimmering, wet trail across the cheek?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the camera tilting upward to track the tear's path and focus on a deep brown eye?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the eyelashes visible as wet and distinctly woven together?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene rendered with a shallow depth of field and soft, diffuse lighting?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0046386257745325565 + ], + "opposite_logprob_tokens": [ + -5.379638671875 + ], + "logprob_reward": -0.0046386257745325565, + "yes_no_margin_reward": 5.375, + "token_credit_reward": -0.0046386257745325565, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1564469337463379, + 0.3533618152141571, + 0.07971546053886414, + 0.058036766946315765, + 0.03303178399801254, + 0.031645890325307846, + 0.0582168772816658, + 0.05505114048719406, + 0.1046610102057457, + 0.06983227282762527 + ], + "credit_map": [ + 0.01410901453346014, + 0.032472722232341766, + 0.014274464920163155, + 0.018948428332805634, + 0.017999349161982536, + 0.016531987115740776, + 0.024036334827542305, + 0.01807464472949505, + 0.021555624902248383, + 0.10004206746816635, + 0.01927274279296398, + 0.12012844532728195, + 0.028171300888061523, + 0.02124481089413166, + 0.028376920148730278, + 0.014569949358701706, + 0.008181150071322918, + 0.011169495061039925, + 0.009338866919279099, + 0.014759628102183342, + 0.009082824923098087, + 0.0067315297201275826, + 0.012282945215702057, + 0.008169029839336872, + 0.005066746845841408, + 0.0065263607539236546, + 0.00612158328294754, + 0.007589427754282951, + 0.013645240105688572, + 0.006806186866015196, + 0.005852320697158575, + 0.006428903434425592, + 0.003045175224542618, + 0.004625817760825157, + 0.004070611670613289, + 0.006611546967178583, + 0.002966108964756131, + 0.0029219884891062975, + 0.005007451865822077, + 0.0037830909714102745, + 0.002821579808369279, + 0.0035138821694999933, + 0.0030745011754333973, + 0.006239434704184532, + 0.004305262118577957, + 0.0023645253386348486, + 0.006130578462034464, + 0.0031961314380168915, + 0.005804297048598528, + 0.008678487502038479, + 0.007650025654584169, + 0.012637445703148842, + 0.009181917645037174, + 0.004925894550979137, + 0.0045027779415249825, + 0.004836040548980236, + 0.004229987971484661, + 0.006635639350861311, + 0.006221843417733908, + 0.013376313261687756, + 0.005781973712146282, + 0.005747034680098295, + 0.007586614694446325, + 0.005471739452332258, + 0.0063885897397994995, + 0.006770254112780094, + 0.020008554682135582, + 0.023960743099451065, + 0.004831253085285425, + 0.0075125498697161674, + 0.024696283042430878, + 0.01049279235303402, + 0.007785431109368801, + 0.007834009826183319, + 0.009728017263114452, + 0.010635393671691418, + 0.005066285841166973, + 0.008059803396463394, + 0.011512490920722485, + 0.009210851043462753 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 8 + ], + "selected_reward": -0.0046386257745325565 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "stunning close-up, captured with a shallow depth of field, captures a single, shimmering teardrop. Soft, diffuse light emanates as it slowly and heavily slides down the delicate skin, leaving a shimmering, wet trail across the cheek. Following the tear's path, the camera tilts upward with an incredibly slow and smooth motion, closely tracking the path of this sorrowful drop, finally bringing its focus to its source: a deep brown eye, brimming with tears and slightly tinged with red. Wet eyelashes are distinctly woven together, reflecting an endless sadness and vulnerability. The entire scene, shrouded in a tranquil yet emotionally charged atmosphere and rendered in soft, natural tones, creates a powerful visual and emotional impact.", + "seed": 31297, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000071/group_8/index_6659_seed_31297_dp_16_rank_32_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 8, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.02365670190192759, + "vlm_logprob": -0.02365670190192759, + "vlm_yes_no_margin": 4.25, + "vlm_token_credit": -0.02365670190192759, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the primary subject a single, shimmering teardrop sliding down skin?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the teardrop leave a visible shimmering, wet trail across the cheek?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the camera tilting upward to track the tear's path and focus on a deep brown eye?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the eyelashes visible as wet and distinctly woven together?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene rendered with a shallow depth of field and soft, diffuse lighting?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02385789155960083 + ], + "opposite_logprob_tokens": [ + -3.773857831954956 + ], + "logprob_reward": -0.02385789155960083, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.02385789155960083, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12827572226524353, + 0.17449641227722168, + 0.02862437069416046, + 0.13382847607135773, + 0.09188824146986008, + 0.04109355807304382, + 0.0776251032948494, + 0.08931678533554077, + 0.19059394299983978, + 0.044257376343011856 + ], + "credit_map": [ + 0.011674010194838047, + 0.019087884575128555, + 0.013864940032362938, + 0.02868199534714222, + 0.024196043610572815, + 0.006160242483019829, + 0.013273137621581554, + 0.011337477713823318, + 0.03310780227184296, + 0.03156322240829468, + 0.03675665706396103, + 0.024367567151784897, + 0.016253216192126274, + 0.0060129654593765736, + 0.017809569835662842, + 0.008625411428511143, + 0.002698336960747838, + 0.0037730811163783073, + 0.004495562985539436, + 0.005799251142889261, + 0.0030592873226851225, + 0.002232800703495741, + 0.005194890778511763, + 0.0013711581705138087, + 0.006654708180576563, + 0.04562759026885033, + 0.007795550860464573, + 0.04220789670944214, + 0.011898013763129711, + 0.006253417115658522, + 0.008907608687877655, + 0.004483695141971111, + 0.011262461543083191, + 0.01810244470834732, + 0.011628231965005398, + 0.014971457421779633, + 0.014851447194814682, + 0.005988242570310831, + 0.009659267961978912, + 0.005424692761152983, + 0.006432306487113237, + 0.0065645985305309296, + 0.005241981241852045, + 0.0051271491684019566, + 0.0045040808618068695, + 0.003642559051513672, + 0.0039055675733834505, + 0.005675317719578743, + 0.0035228414926677942, + 0.0034628657158464193, + 0.02439006231725216, + 0.00914050079882145, + 0.0025895999278873205, + 0.004390760790556669, + 0.02344445139169693, + 0.0066840206272900105, + 0.009211546741425991, + 0.008759113028645515, + 0.007824772968888283, + 0.01969989947974682, + 0.007810134440660477, + 0.006695075426250696, + 0.019247910007834435, + 0.010068332776427269, + 0.011557122692465782, + 0.014878255315124989, + 0.016806919127702713, + 0.04685645177960396, + 0.013108006678521633, + 0.01586643047630787, + 0.05210302397608757, + 0.019417738541960716, + 0.0032960865646600723, + 0.0029104610439389944, + 0.00575635302811861, + 0.006436686031520367, + 0.0037995222955942154, + 0.003927665762603283, + 0.013080856762826443, + 0.005049741826951504 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 1, + 3 + ], + "selected_reward": -0.02385789155960083 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the primary subject a single, shimmering teardrop sliding down skin?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the teardrop leave a visible shimmering, wet trail across the cheek?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the camera tilting upward to track the tear's path and focus on a deep brown eye?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the eyelashes visible as wet and distinctly woven together?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene rendered with a shallow depth of field and soft, diffuse lighting?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.012558309361338615 + ], + "opposite_logprob_tokens": [ + -4.387558460235596 + ], + "logprob_reward": -0.012558309361338615, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.012558309361338615, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11861375719308853, + 0.07439915090799332, + 0.03131324425339699, + 0.09132958203554153, + 0.17112408578395844, + 0.017810344696044922, + 0.3329380750656128, + 0.07426021993160248, + 0.02354513853788376, + 0.06466645747423172 + ], + "credit_map": [ + 0.016103599220514297, + 0.019153183326125145, + 0.02341059409081936, + 0.02332443743944168, + 0.014875179156661034, + 0.004216892644762993, + 0.01080674584954977, + 0.0067231310531497, + 0.01433075312525034, + 0.010347547940909863, + 0.01613854430615902, + 0.010611548088490963, + 0.007817470468580723, + 0.0028111750725656748, + 0.008050886914134026, + 0.004291232209652662, + 0.00226722308434546, + 0.004304787609726191, + 0.005625047720968723, + 0.00617195526137948, + 0.0030218122992664576, + 0.0033292174339294434, + 0.0045484681613743305, + 0.002044732216745615, + 0.004358882550150156, + 0.03288891166448593, + 0.005079029127955437, + 0.029212702065706253, + 0.00837837252765894, + 0.004603138193488121, + 0.00485900416970253, + 0.001949545694515109, + 0.01595134101808071, + 0.05708196014165878, + 0.015418203547596931, + 0.040382105857133865, + 0.01595383882522583, + 0.007271705660969019, + 0.01156087126582861, + 0.007504068780690432, + 0.0020820824429392815, + 0.004132578149437904, + 0.0018785380525514483, + 0.0027606997173279524, + 0.0015938851283863187, + 0.001811376539990306, + 0.002035197103396058, + 0.0015159879112616181, + 0.013119739480316639, + 0.00789218582212925, + 0.13484813272953033, + 0.03214874118566513, + 0.010889123193919659, + 0.0068918196484446526, + 0.11645529419183731, + 0.010693016462028027, + 0.007940894924104214, + 0.004984775558114052, + 0.00933548342436552, + 0.015045530162751675, + 0.006736602634191513, + 0.006507278885692358, + 0.01705685257911682, + 0.006652808748185635, + 0.0017577363178133965, + 0.002460027812048793, + 0.002476037247106433, + 0.003616300178691745, + 0.002784470794722438, + 0.0030617795418947935, + 0.004779145587235689, + 0.0026096412912011147, + 0.007601757999509573, + 0.007589380722492933, + 0.01181122101843357, + 0.0068390099331736565, + 0.007840471342206001, + 0.006554633844643831, + 0.006919567938894033, + 0.009510417468845844 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 4, + 0 + ], + "selected_reward": -0.012558309361338615 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the primary subject a single, shimmering teardrop sliding down skin?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the teardrop leave a visible shimmering, wet trail across the cheek?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the camera tilting upward to track the tear's path and focus on a deep brown eye?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the eyelashes visible as wet and distinctly woven together?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene rendered with a shallow depth of field and soft, diffuse lighting?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.008648676797747612 + ], + "opposite_logprob_tokens": [ + -4.758648872375488 + ], + "logprob_reward": -0.008648676797747612, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.008648676797747612, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14057941734790802, + 0.06854727864265442, + 0.04953283816576004, + 0.11208900064229965, + 0.18064233660697937, + 0.04588519036769867, + 0.1884320080280304, + 0.13131023943424225, + 0.04740818962454796, + 0.03557352349162102 + ], + "credit_map": [ + 0.021408425644040108, + 0.024224968627095222, + 0.018678827211260796, + 0.023863250389695168, + 0.021740896627306938, + 0.006146844942122698, + 0.01258869655430317, + 0.011927506886422634, + 0.008314835838973522, + 0.009147701784968376, + 0.011318172328174114, + 0.0097905108705163, + 0.00889683235436678, + 0.005184490699321032, + 0.009652254171669483, + 0.006242482922971249, + 0.004472363740205765, + 0.010199733078479767, + 0.005831167567521334, + 0.00849093683063984, + 0.005119369365274906, + 0.0036528564523905516, + 0.009270891547203064, + 0.002495525637641549, + 0.008094177581369877, + 0.03303145617246628, + 0.010880016721785069, + 0.029716838151216507, + 0.008522452786564827, + 0.00812210701406002, + 0.009391291067004204, + 0.004330663941800594, + 0.01265859417617321, + 0.06538242101669312, + 0.013025172986090183, + 0.05124356225132942, + 0.015113488771021366, + 0.006629131268709898, + 0.009670178405940533, + 0.006919774692505598, + 0.006027203984558582, + 0.008766389451920986, + 0.0044870986603200436, + 0.0061234235763549805, + 0.005740870721638203, + 0.004163919482380152, + 0.004943906795233488, + 0.005632378160953522, + 0.012952346354722977, + 0.01207085233181715, + 0.06395001709461212, + 0.03212752938270569, + 0.01334717683494091, + 0.012897013686597347, + 0.02401665598154068, + 0.01707041636109352, + 0.011336298659443855, + 0.00801467802375555, + 0.011967452242970467, + 0.031390536576509476, + 0.010140527971088886, + 0.006902862340211868, + 0.03584446758031845, + 0.015713414177298546, + 0.004344042390584946, + 0.006049735937267542, + 0.004702677018940449, + 0.008689642883837223, + 0.005709945224225521, + 0.004209992941468954, + 0.008534104563295841, + 0.005168050527572632, + 0.004222424700856209, + 0.0036263163201510906, + 0.008078677579760551, + 0.0038023735396564007, + 0.003465113462880254, + 0.0032221837900578976, + 0.0043017566204071045, + 0.004854677245020866 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 4, + 0 + ], + "selected_reward": -0.008648676797747612 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the primary subject a single, shimmering teardrop sliding down skin?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the teardrop leave a visible shimmering, wet trail across the cheek?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the camera tilting upward to track the tear's path and focus on a deep brown eye?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the eyelashes visible as wet and distinctly woven together?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene rendered with a shallow depth of field and soft, diffuse lighting?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0700240433216095 + ], + "opposite_logprob_tokens": [ + -2.695024013519287 + ], + "logprob_reward": -0.0700240433216095, + "yes_no_margin_reward": 2.625, + "token_credit_reward": -0.0700240433216095, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09575487673282623, + 0.022819196805357933, + 0.015310767106711864, + 0.0308237224817276, + 0.0988617092370987, + 0.025384189561009407, + 0.15674003958702087, + 0.15365177392959595, + 0.33567798137664795, + 0.06497571617364883 + ], + "credit_map": [ + 0.010278897359967232, + 0.02192646451294422, + 0.011407510377466679, + 0.020908156409859657, + 0.013331000693142414, + 0.002947354456409812, + 0.008378402329981327, + 0.00657708989456296, + 0.0032254199031740427, + 0.0033290546853095293, + 0.0034503673668950796, + 0.003539314027875662, + 0.002680162899196148, + 0.0012858861591666937, + 0.0034631111193448305, + 0.0018458812264725566, + 0.0014488663291558623, + 0.002162643475458026, + 0.0021379452664405107, + 0.0027968951035290956, + 0.0017313878051936626, + 0.001565502374432981, + 0.0025664069689810276, + 0.0009011199581436813, + 0.0029789418913424015, + 0.007062568329274654, + 0.0033049832563847303, + 0.0057779839262366295, + 0.00341107533313334, + 0.0031187802087515593, + 0.003456710372120142, + 0.001712678698822856, + 0.005364319309592247, + 0.03922915831208229, + 0.009441131725907326, + 0.023368343710899353, + 0.007843481376767159, + 0.0032435308676213026, + 0.0070718564093112946, + 0.003299886826425791, + 0.002476828172802925, + 0.002608759328722954, + 0.0029250739607959986, + 0.0071226381696760654, + 0.0024765136186033487, + 0.0023790490813553333, + 0.0029805509839206934, + 0.002414778573438525, + 0.006723028142005205, + 0.003559043165296316, + 0.060193851590156555, + 0.012847461737692356, + 0.004635048098862171, + 0.0030488648917526007, + 0.06032262742519379, + 0.005410119891166687, + 0.011054507456719875, + 0.01050956454128027, + 0.014481252059340477, + 0.04294352978467941, + 0.012214256450533867, + 0.008873233571648598, + 0.038243770599365234, + 0.015331650152802467, + 0.021899374201893806, + 0.023452892899513245, + 0.031585175544023514, + 0.08191095292568207, + 0.025235917419195175, + 0.024711113423109055, + 0.09047595411539078, + 0.03640660271048546, + 0.0064295497722923756, + 0.006181744858622551, + 0.013789219781756401, + 0.0068682050332427025, + 0.0062289186753332615, + 0.0063195014372467995, + 0.010064205154776573, + 0.00909436959773302 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 6, + 7 + ], + "selected_reward": -0.0700240433216095 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the primary subject a single, shimmering teardrop sliding down skin?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the teardrop leave a visible shimmering, wet trail across the cheek?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the camera tilting upward to track the tear's path and focus on a deep brown eye?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the eyelashes visible as wet and distinctly woven together?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene rendered with a shallow depth of field and soft, diffuse lighting?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0031945884693413973 + ], + "opposite_logprob_tokens": [ + -5.753194808959961 + ], + "logprob_reward": -0.0031945884693413973, + "yes_no_margin_reward": 5.75, + "token_credit_reward": -0.0031945884693413973, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1250290721654892, + 0.05024837702512741, + 0.032878149300813675, + 0.1404312402009964, + 0.2685973048210144, + 0.02340645343065262, + 0.034425895661115646, + 0.07293802499771118, + 0.16157427430152893, + 0.09047122299671173 + ], + "credit_map": [ + 0.015369395725429058, + 0.018296340480446815, + 0.020660825073719025, + 0.019952474161982536, + 0.02164778672158718, + 0.005595935042947531, + 0.012309958226978779, + 0.011196358129382133, + 0.005733412224799395, + 0.006697166245430708, + 0.006716358941048384, + 0.00748950382694602, + 0.007704231422394514, + 0.004442818928509951, + 0.006059282459318638, + 0.005405613221228123, + 0.0038275106344372034, + 0.005516578909009695, + 0.004701347555965185, + 0.007103078067302704, + 0.0030858172103762627, + 0.002856014296412468, + 0.004130169283598661, + 0.0016576357884332538, + 0.005897520575672388, + 0.04896331578493118, + 0.008646388538181782, + 0.044114816933870316, + 0.011747783049941063, + 0.006585137918591499, + 0.01016521081328392, + 0.0043110717087984085, + 0.017274625599384308, + 0.09745696187019348, + 0.016590943560004234, + 0.06837236881256104, + 0.0360795222222805, + 0.00870472565293312, + 0.016255570575594902, + 0.007862605154514313, + 0.0028209600131958723, + 0.0021582660265266895, + 0.0028484633658081293, + 0.00493265874683857, + 0.002623627195134759, + 0.0029138182289898396, + 0.0027834747452288866, + 0.0023251858074218035, + 0.0021157898008823395, + 0.0021397422533482313, + 0.007680284790694714, + 0.008941546082496643, + 0.002078442834317684, + 0.002461290918290615, + 0.0065369075164198875, + 0.002471894258633256, + 0.006543555296957493, + 0.005456125363707542, + 0.006777235306799412, + 0.014921422116458416, + 0.00700482539832592, + 0.006236642133444548, + 0.018585005775094032, + 0.007413223851472139, + 0.01026171538978815, + 0.009331554174423218, + 0.014223732985556126, + 0.04189978539943695, + 0.010074889287352562, + 0.011606204323470592, + 0.04812086746096611, + 0.016055531799793243, + 0.009736436419188976, + 0.009604660794138908, + 0.02129020169377327, + 0.009189937263727188, + 0.010269152000546455, + 0.008485851809382439, + 0.009392943233251572, + 0.012502042576670647 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 8, + 3 + ], + "selected_reward": -0.0031945884693413973 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "stunning close-up, captured with a shallow depth of field, captures a single, shimmering teardrop. Soft, diffuse light emanates as it slowly and heavily slides down the delicate skin, leaving a shimmering, wet trail across the cheek. Following the tear's path, the camera tilts upward with an incredibly slow and smooth motion, closely tracking the path of this sorrowful drop, finally bringing its focus to its source: a deep brown eye, brimming with tears and slightly tinged with red. Wet eyelashes are distinctly woven together, reflecting an endless sadness and vulnerability. The entire scene, shrouded in a tranquil yet emotionally charged atmosphere and rendered in soft, natural tones, creates a powerful visual and emotional impact.", + "seed": 31298, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000071/group_8/index_6659_seed_31298_dp_17_rank_34_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 8, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.024482604768127202, + "vlm_logprob": -0.024482604768127202, + "vlm_yes_no_margin": 4.2, + "vlm_token_credit": -0.024482604768127202, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the primary subject a single, shimmering teardrop sliding down skin?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the teardrop leave a visible shimmering, wet trail across the cheek?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the camera tilting upward to track the tear's path and focus on a deep brown eye?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the eyelashes visible as wet and distinctly woven together?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene rendered with a shallow depth of field and soft, diffuse lighting?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.014640401117503643 + ], + "opposite_logprob_tokens": [ + -4.2646403312683105 + ], + "logprob_reward": -0.014640401117503643, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.014640401117503643, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.39917629957199097, + 0.11598266661167145, + 0.09737992286682129, + 0.1268705576658249, + 0.05454721301794052, + 0.0404079370200634, + 0.05447906628251076, + 0.05072933807969093, + 0.035792771726846695, + 0.024634210392832756 + ], + "credit_map": [ + 0.013376706279814243, + 0.020229604095220566, + 0.018363265320658684, + 0.015473870560526848, + 0.03582209721207619, + 0.13949555158615112, + 0.029311906546354294, + 0.1271033138036728, + 0.0077803535386919975, + 0.01065631490200758, + 0.01244485005736351, + 0.012202671729028225, + 0.015102925710380077, + 0.014823401346802711, + 0.03019811399281025, + 0.012774032540619373, + 0.004635592456907034, + 0.009605664759874344, + 0.007836638949811459, + 0.007670812774449587, + 0.01122557744383812, + 0.019848672673106194, + 0.014389951713383198, + 0.022167008370161057, + 0.006443903315812349, + 0.012480919249355793, + 0.01499759778380394, + 0.02147717960178852, + 0.015268515795469284, + 0.016344554722309113, + 0.027934858575463295, + 0.011923031881451607, + 0.006137428339570761, + 0.005169495474547148, + 0.007017548196017742, + 0.007438673172146082, + 0.006344342604279518, + 0.006429777480661869, + 0.0107564777135849, + 0.005253470502793789, + 0.003383204573765397, + 0.0029392861761152744, + 0.004555283579975367, + 0.005714386701583862, + 0.004125053528696299, + 0.0038097179494798183, + 0.01102896686643362, + 0.004852037411183119, + 0.007381466683000326, + 0.004149341490119696, + 0.0054136160761117935, + 0.005806228145956993, + 0.0061422414146363735, + 0.0055673676542937756, + 0.009704978205263615, + 0.010313823819160461, + 0.0063175177201628685, + 0.005149336997419596, + 0.008605729788541794, + 0.007634889334440231, + 0.0049580419436097145, + 0.004993441980332136, + 0.005822540260851383, + 0.0072478423826396465, + 0.00487649068236351, + 0.003784064669162035, + 0.0065239365212619305, + 0.005309932865202427, + 0.0035697261337190866, + 0.003102292539551854, + 0.004789373371750116, + 0.003836953779682517, + 0.0028828077483922243, + 0.0022433046251535416, + 0.006154065486043692, + 0.002072668867185712, + 0.002221980830654502, + 0.0019572684541344643, + 0.0018855857197195292, + 0.0052165258675813675 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 1 + ], + "selected_reward": -0.014640401117503643 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the primary subject a single, shimmering teardrop sliding down skin?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the teardrop leave a visible shimmering, wet trail across the cheek?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the camera tilting upward to track the tear's path and focus on a deep brown eye?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the eyelashes visible as wet and distinctly woven together?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene rendered with a shallow depth of field and soft, diffuse lighting?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01608467288315296 + ], + "opposite_logprob_tokens": [ + -4.141084671020508 + ], + "logprob_reward": -0.01608467288315296, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.01608467288315296, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.28381744027137756, + 0.1461792141199112, + 0.1185990497469902, + 0.14519654214382172, + 0.053394000977277756, + 0.0662783831357956, + 0.0249672532081604, + 0.0450747050344944, + 0.0627421960234642, + 0.0537511445581913 + ], + "credit_map": [ + 0.013464091345667839, + 0.01614396460354328, + 0.01810474880039692, + 0.015004453249275684, + 0.036625731736421585, + 0.0800345242023468, + 0.03097323328256607, + 0.07346668839454651, + 0.008977125398814678, + 0.011642146855592728, + 0.013573174364864826, + 0.012623853050172329, + 0.02014215849339962, + 0.025983981788158417, + 0.027656013146042824, + 0.02558075450360775, + 0.007508232723921537, + 0.01263637188822031, + 0.01131532248109579, + 0.012206208892166615, + 0.015945514664053917, + 0.021586140617728233, + 0.01747419498860836, + 0.019927067682147026, + 0.007159863132983446, + 0.015396852977573872, + 0.0286160409450531, + 0.01772703416645527, + 0.01996425725519657, + 0.015380856581032276, + 0.025275196880102158, + 0.015676438808441162, + 0.003847596002742648, + 0.0048586162738502026, + 0.005867233499884605, + 0.00919266790151596, + 0.0059356363490223885, + 0.004532589111477137, + 0.013596265576779842, + 0.00556339742615819, + 0.007294731214642525, + 0.004988981410861015, + 0.006467817816883326, + 0.011806400492787361, + 0.007968466728925705, + 0.007079711649566889, + 0.014124187640845776, + 0.006548088975250721, + 0.002984480233862996, + 0.0019449219107627869, + 0.0022925317753106356, + 0.004123905673623085, + 0.0032334867864847183, + 0.002528505166992545, + 0.0043604373931884766, + 0.0034989845007658005, + 0.005324923899024725, + 0.004382513929158449, + 0.005739940330386162, + 0.007093122694641352, + 0.00673716189339757, + 0.005206086672842503, + 0.005751938093453646, + 0.004839017987251282, + 0.009078044444322586, + 0.005009226500988007, + 0.01960618607699871, + 0.005045321304351091, + 0.003342576092109084, + 0.004426100756973028, + 0.004575725644826889, + 0.011659014038741589, + 0.00739023694768548, + 0.002685345010831952, + 0.019710339605808258, + 0.004341032821685076, + 0.002445729449391365, + 0.003508468857035041, + 0.001956640277057886, + 0.011713351123034954 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 3 + ], + "selected_reward": -0.01608467288315296 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the primary subject a single, shimmering teardrop sliding down skin?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the teardrop leave a visible shimmering, wet trail across the cheek?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the camera tilting upward to track the tear's path and focus on a deep brown eye?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the eyelashes visible as wet and distinctly woven together?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene rendered with a shallow depth of field and soft, diffuse lighting?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.006748148240149021 + ], + "opposite_logprob_tokens": [ + -5.006748199462891 + ], + "logprob_reward": -0.006748148240149021, + "yes_no_margin_reward": 5.0, + "token_credit_reward": -0.006748148240149021, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.4617860019207001, + 0.07652638852596283, + 0.0907064750790596, + 0.11227080225944519, + 0.08091376721858978, + 0.05850628763437271, + 0.031787171959877014, + 0.029763607308268547, + 0.042466115206480026, + 0.015273360535502434 + ], + "credit_map": [ + 0.016009127721190453, + 0.02115773782134056, + 0.019386950880289078, + 0.02095717564225197, + 0.03516089916229248, + 0.16615568101406097, + 0.026612471789121628, + 0.1563459187746048, + 0.005032897926867008, + 0.007170878350734711, + 0.009848800487816334, + 0.007626262027770281, + 0.00737023027613759, + 0.011155907064676285, + 0.019651787355542183, + 0.008669612929224968, + 0.004242005757987499, + 0.012807572260499, + 0.009963784366846085, + 0.011950313113629818, + 0.008416752330958843, + 0.016978759318590164, + 0.010342447087168694, + 0.016004830598831177, + 0.0060232230462133884, + 0.00914480909705162, + 0.027741847559809685, + 0.012304022908210754, + 0.010177088901400566, + 0.013261684216558933, + 0.0185067318379879, + 0.01511138305068016, + 0.008504251018166542, + 0.008852800354361534, + 0.011224046349525452, + 0.013114835135638714, + 0.010005098767578602, + 0.009381427429616451, + 0.011430672369897366, + 0.008400634862482548, + 0.00555043900385499, + 0.004378067795187235, + 0.006857929285615683, + 0.0066881184466183186, + 0.007076593581587076, + 0.00705177104100585, + 0.014682735316455364, + 0.006220633629709482, + 0.004524105228483677, + 0.0030672044958919287, + 0.0036895121447741985, + 0.004184265621006489, + 0.0038322852924466133, + 0.003332193475216627, + 0.004291764926165342, + 0.004865838214755058, + 0.002912208903580904, + 0.0026084608398377895, + 0.0033962447196245193, + 0.004959600046277046, + 0.0033951550722122192, + 0.0029707506764680147, + 0.004693557973951101, + 0.004827626049518585, + 0.00560368038713932, + 0.005035095848143101, + 0.012667757458984852, + 0.004431636072695255, + 0.003907959908246994, + 0.0031308962497860193, + 0.002637288998812437, + 0.005051797721534967, + 0.001998095540329814, + 0.0013630345929414034, + 0.002145994920283556, + 0.0015551070682704449, + 0.0018746318528428674, + 0.001483559375628829, + 0.001354697160422802, + 0.003498238045722246 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 2 + ], + "selected_reward": -0.006748148240149021 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the primary subject a single, shimmering teardrop sliding down skin?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the teardrop leave a visible shimmering, wet trail across the cheek?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the camera tilting upward to track the tear's path and focus on a deep brown eye?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the eyelashes visible as wet and distinctly woven together?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene rendered with a shallow depth of field and soft, diffuse lighting?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07898944616317749 + ], + "opposite_logprob_tokens": [ + -2.5789895057678223 + ], + "logprob_reward": -0.07898944616317749, + "yes_no_margin_reward": 2.5, + "token_credit_reward": -0.07898944616317749, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14407268166542053, + 0.11560181528329849, + 0.06779041886329651, + 0.09883930534124374, + 0.05351323261857033, + 0.053747233003377914, + 0.042458850890398026, + 0.07559269666671753, + 0.04820852726697922, + 0.300175279378891 + ], + "credit_map": [ + 0.010333088226616383, + 0.016786491498351097, + 0.018854497000575066, + 0.011815063655376434, + 0.030753737315535545, + 0.021023506298661232, + 0.020065991207957268, + 0.01444029901176691, + 0.009593234397470951, + 0.009474905207753181, + 0.0153754698112607, + 0.01638941466808319, + 0.013041152618825436, + 0.019754569977521896, + 0.01149736251682043, + 0.02047569677233696, + 0.003712693229317665, + 0.006124873645603657, + 0.00814986601471901, + 0.00744126969948411, + 0.008921115659177303, + 0.011916433461010456, + 0.010678782127797604, + 0.010845381766557693, + 0.0056740534491837025, + 0.011374999769032001, + 0.01741393841803074, + 0.013099158182740211, + 0.014609293080866337, + 0.01076694205403328, + 0.017572617158293724, + 0.008328292518854141, + 0.006157432217150927, + 0.007112573366612196, + 0.007903783582150936, + 0.007495260797441006, + 0.005787334870547056, + 0.005921141244471073, + 0.007140570320188999, + 0.005995131563395262, + 0.005432008765637875, + 0.003593660891056061, + 0.006245170719921589, + 0.011012724600732327, + 0.005550757981836796, + 0.005494699347764254, + 0.010684989392757416, + 0.005733218044042587, + 0.004626389127224684, + 0.00362044433131814, + 0.005209033377468586, + 0.0070266807451844215, + 0.006128842476755381, + 0.0055660526268184185, + 0.006577362306416035, + 0.0037040458992123604, + 0.00625969935208559, + 0.006494281347841024, + 0.010877210646867752, + 0.006173230241984129, + 0.004873993340879679, + 0.010508468374609947, + 0.012227519415318966, + 0.018178287893533707, + 0.005693389568477869, + 0.004459862131625414, + 0.011296144686639309, + 0.006572427693754435, + 0.0048841084353625774, + 0.0029544862918555737, + 0.006191786378622055, + 0.006156319752335548, + 0.04081592336297035, + 0.020323392003774643, + 0.10941240936517715, + 0.02561616152524948, + 0.022376535460352898, + 0.0306846983730793, + 0.013950128108263016, + 0.03699599951505661 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 0, + 1 + ], + "selected_reward": -0.07898944616317749 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the primary subject a single, shimmering teardrop sliding down skin?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the teardrop leave a visible shimmering, wet trail across the cheek?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the camera tilting upward to track the tear's path and focus on a deep brown eye?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the eyelashes visible as wet and distinctly woven together?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene rendered with a shallow depth of field and soft, diffuse lighting?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005950355436652899 + ], + "opposite_logprob_tokens": [ + -5.130950450897217 + ], + "logprob_reward": -0.005950355436652899, + "yes_no_margin_reward": 5.125, + "token_credit_reward": -0.005950355436652899, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.613650918006897, + 0.0552140511572361, + 0.07081401348114014, + 0.024989008903503418, + 0.031126594170928, + 0.037783313542604446, + 0.020345881581306458, + 0.03893839567899704, + 0.07324907183647156, + 0.033888768404722214 + ], + "credit_map": [ + 0.020745201036334038, + 0.026358334347605705, + 0.02951682358980179, + 0.024532219395041466, + 0.046261344105005264, + 0.22145457565784454, + 0.03655646741390228, + 0.20822595059871674, + 0.0037528201937675476, + 0.004814936313778162, + 0.005964942742139101, + 0.005711160134524107, + 0.0064553385600447655, + 0.007608962245285511, + 0.013864640146493912, + 0.0070412494242191315, + 0.0037337620742619038, + 0.007970253005623817, + 0.009167156182229519, + 0.009185337461531162, + 0.009358666837215424, + 0.008078782819211483, + 0.013224744237959385, + 0.010095309466123581, + 0.0015716382768005133, + 0.0025506652891635895, + 0.0027069442439824343, + 0.0030648165848106146, + 0.00434578163549304, + 0.002987408312037587, + 0.004761857446283102, + 0.002999895717948675, + 0.002997478935867548, + 0.0032867740374058485, + 0.0046671354211866856, + 0.004110297188162804, + 0.003546136897057295, + 0.0042933207005262375, + 0.004776684567332268, + 0.003448766190558672, + 0.004172052722424269, + 0.002714010886847973, + 0.003369719721376896, + 0.005895389709621668, + 0.003320508636534214, + 0.0034679793752729893, + 0.010235313326120377, + 0.004608338698744774, + 0.0027671458665281534, + 0.0018500675214454532, + 0.002244045725092292, + 0.002643835498020053, + 0.0028652141336351633, + 0.0022498629987239838, + 0.0028420903254300356, + 0.0028836182318627834, + 0.005840696860104799, + 0.004345300607383251, + 0.0057107796892523766, + 0.006125273182988167, + 0.0036154896952211857, + 0.004407441709190607, + 0.003806006396189332, + 0.005087406374514103, + 0.008969521149992943, + 0.005542758386582136, + 0.028512120246887207, + 0.004442581441253424, + 0.002809335244819522, + 0.0038301090244203806, + 0.0029239205177873373, + 0.01621873304247856, + 0.005176007281988859, + 0.00408910121768713, + 0.005052116233855486, + 0.003808335866779089, + 0.004970996640622616, + 0.003985204268246889, + 0.002667198423296213, + 0.004139807540923357 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 2 + ], + "selected_reward": -0.005950355436652899 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "stunning close-up, captured with a shallow depth of field, captures a single, shimmering teardrop. Soft, diffuse light emanates as it slowly and heavily slides down the delicate skin, leaving a shimmering, wet trail across the cheek. Following the tear's path, the camera tilts upward with an incredibly slow and smooth motion, closely tracking the path of this sorrowful drop, finally bringing its focus to its source: a deep brown eye, brimming with tears and slightly tinged with red. Wet eyelashes are distinctly woven together, reflecting an endless sadness and vulnerability. The entire scene, shrouded in a tranquil yet emotionally charged atmosphere and rendered in soft, natural tones, creates a powerful visual and emotional impact.", + "seed": 31299, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000071/group_8/index_6659_seed_31299_dp_17_rank_34_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 8, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.019505565892904995, + "vlm_logprob": -0.019505565892904995, + "vlm_yes_no_margin": 4.4, + "vlm_token_credit": -0.019505565892904995, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the primary subject a single, shimmering teardrop sliding down skin?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the teardrop leave a visible shimmering, wet trail across the cheek?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the camera tilting upward to track the tear's path and focus on a deep brown eye?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the eyelashes visible as wet and distinctly woven together?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene rendered with a shallow depth of field and soft, diffuse lighting?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01454582717269659 + ], + "opposite_logprob_tokens": [ + -4.264545917510986 + ], + "logprob_reward": -0.01454582717269659, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.01454582717269659, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09078782796859741, + 0.03221055492758751, + 0.04804440215229988, + 0.09682875126600266, + 0.13613459467887878, + 0.08300045132637024, + 0.1049862951040268, + 0.0709698349237442, + 0.2413255125284195, + 0.09571180492639542 + ], + "credit_map": [ + 0.01097686868160963, + 0.011727302335202694, + 0.01473300438374281, + 0.009846954606473446, + 0.015110312029719353, + 0.007064925506711006, + 0.01593000814318657, + 0.005398437846451998, + 0.003819454461336136, + 0.004872970748692751, + 0.005078518763184547, + 0.005118776112794876, + 0.005190689116716385, + 0.0018466649344190955, + 0.00496514979749918, + 0.0013183283153921366, + 0.004481955897063017, + 0.00810189452022314, + 0.006984917912632227, + 0.012226449325680733, + 0.005019251722842455, + 0.00188837933819741, + 0.0070093534886837006, + 0.0023321970365941525, + 0.007948662154376507, + 0.016500655561685562, + 0.015119330026209354, + 0.01534030307084322, + 0.0144647266715765, + 0.0041125123389065266, + 0.01740582287311554, + 0.005936725530773401, + 0.013090716674923897, + 0.018209047615528107, + 0.021382128819823265, + 0.016265517100691795, + 0.018209263682365417, + 0.0101698013022542, + 0.03180307149887085, + 0.007005041465163231, + 0.009500776417553425, + 0.009051845408976078, + 0.011740623973309994, + 0.014236195012927055, + 0.011302401311695576, + 0.007689835503697395, + 0.010438529774546623, + 0.009040231816470623, + 0.009989387355744839, + 0.009770900942385197, + 0.014232787303626537, + 0.014247794635593891, + 0.019163034856319427, + 0.01539695356041193, + 0.015264305286109447, + 0.00692112697288394, + 0.008005673997104168, + 0.007390332408249378, + 0.009015975520014763, + 0.007673397660255432, + 0.0128109036013484, + 0.008632993325591087, + 0.00920187495648861, + 0.00823867879807949, + 0.01762721873819828, + 0.020267676562070847, + 0.01943741738796234, + 0.04173288121819496, + 0.017718173563480377, + 0.060026343911886215, + 0.02855489030480385, + 0.035960886627435684, + 0.018848508596420288, + 0.008306073024868965, + 0.010023456998169422, + 0.013933880254626274, + 0.005811419803649187, + 0.015210909768939018, + 0.005165680777281523, + 0.018411867320537567 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 4, + 6 + ], + "selected_reward": -0.01454582717269659 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the primary subject a single, shimmering teardrop sliding down skin?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the teardrop leave a visible shimmering, wet trail across the cheek?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the camera tilting upward to track the tear's path and focus on a deep brown eye?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the eyelashes visible as wet and distinctly woven together?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene rendered with a shallow depth of field and soft, diffuse lighting?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.008651395328342915 + ], + "opposite_logprob_tokens": [ + -4.758651256561279 + ], + "logprob_reward": -0.008651395328342915, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.008651395328342915, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06919518113136292, + 0.04937655106186867, + 0.06609086692333221, + 0.24472661316394806, + 0.09830235689878464, + 0.11255902051925659, + 0.07701762020587921, + 0.15456317365169525, + 0.04733913764357567, + 0.08082950860261917 + ], + "credit_map": [ + 0.009198159910738468, + 0.008288202807307243, + 0.011239035055041313, + 0.007274027913808823, + 0.01382063515484333, + 0.004724615253508091, + 0.011196688748896122, + 0.0034538167528808117, + 0.004857438616454601, + 0.011355278082191944, + 0.0075484090484678745, + 0.005594715476036072, + 0.010030245408415794, + 0.00244918093085289, + 0.005592923145741224, + 0.0019483590731397271, + 0.006406001280993223, + 0.013940895907580853, + 0.007258985657244921, + 0.01332010980695486, + 0.01000635139644146, + 0.002705399878323078, + 0.01006059069186449, + 0.002392533700913191, + 0.017383307218551636, + 0.028421854600310326, + 0.049941983073949814, + 0.03637804463505745, + 0.0417623408138752, + 0.008243920281529427, + 0.05052460357546806, + 0.01207054778933525, + 0.009794364683330059, + 0.009891785681247711, + 0.017101356759667397, + 0.013356502167880535, + 0.016139019280672073, + 0.007454355247318745, + 0.016988635063171387, + 0.007576332427561283, + 0.009758060798048973, + 0.00833752192556858, + 0.015364476479589939, + 0.031482815742492676, + 0.013628182001411915, + 0.008796978741884232, + 0.01609824411571026, + 0.009092742577195168, + 0.0076376888900995255, + 0.007850917056202888, + 0.008169780485332012, + 0.015351436100900173, + 0.012867567129433155, + 0.008142918348312378, + 0.009718808345496655, + 0.007278506178408861, + 0.012312008999288082, + 0.012273367494344711, + 0.012231537140905857, + 0.02745937742292881, + 0.010237546637654305, + 0.04148036986589432, + 0.025218766182661057, + 0.013350219465792179, + 0.004700120072811842, + 0.0041227275505661964, + 0.0038578796666115522, + 0.0038618054240942, + 0.0056130364537239075, + 0.012474006041884422, + 0.006799688097089529, + 0.00590987503528595, + 0.01645469479262829, + 0.008107873611152172, + 0.009096928872168064, + 0.010540480725467205, + 0.006148761138319969, + 0.011038452386856079, + 0.005203444045037031, + 0.014238867908716202 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 7, + 5 + ], + "selected_reward": -0.008651395328342915 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the primary subject a single, shimmering teardrop sliding down skin?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the teardrop leave a visible shimmering, wet trail across the cheek?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the camera tilting upward to track the tear's path and focus on a deep brown eye?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the eyelashes visible as wet and distinctly woven together?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene rendered with a shallow depth of field and soft, diffuse lighting?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0076417261734604836 + ], + "opposite_logprob_tokens": [ + -4.882641792297363 + ], + "logprob_reward": -0.0076417261734604836, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.0076417261734604836, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.05384713411331177, + 0.036242928355932236, + 0.040667574852705, + 0.07935240119695663, + 0.1892034262418747, + 0.07601242512464523, + 0.09293626248836517, + 0.11488815397024155, + 0.2697775065898895, + 0.04707229137420654 + ], + "credit_map": [ + 0.005862739868462086, + 0.008961188606917858, + 0.008942319080233574, + 0.006142272148281336, + 0.008573359809815884, + 0.004384520463645458, + 0.007850184105336666, + 0.0031305395532399416, + 0.003815203905105591, + 0.009507113136351109, + 0.004127584397792816, + 0.004822788294404745, + 0.0053958892822265625, + 0.0019092326983809471, + 0.0052382140420377254, + 0.0014268949162214994, + 0.007110702805221081, + 0.006034722551703453, + 0.004725972190499306, + 0.006968261208385229, + 0.005687073338776827, + 0.0022298095282167196, + 0.005566996987909079, + 0.0023440297227352858, + 0.0056200213730335236, + 0.007359080947935581, + 0.015793459489941597, + 0.015211984515190125, + 0.012622365728020668, + 0.0032828787807375193, + 0.016535205766558647, + 0.002927385736256838, + 0.01554664596915245, + 0.01733505353331566, + 0.03938477858901024, + 0.013384739868342876, + 0.018559802323579788, + 0.015021486207842827, + 0.058320026844739914, + 0.011650857515633106, + 0.008914685808122158, + 0.009664908982813358, + 0.007220169994980097, + 0.014778440818190575, + 0.011832168325781822, + 0.0059715574607253075, + 0.01033821702003479, + 0.007292266935110092, + 0.010683341883122921, + 0.011232453398406506, + 0.013402133248746395, + 0.008434093557298183, + 0.014237928204238415, + 0.01284051313996315, + 0.01170960534363985, + 0.010396174155175686, + 0.01051481906324625, + 0.010522969998419285, + 0.015624456107616425, + 0.03489995375275612, + 0.012971429154276848, + 0.010164652019739151, + 0.01119254156947136, + 0.008997317403554916, + 0.026666052639484406, + 0.020253675058484077, + 0.017036736011505127, + 0.035748351365327835, + 0.01940145157277584, + 0.06254133582115173, + 0.06143195182085037, + 0.026697881519794464, + 0.007306961342692375, + 0.005451182369142771, + 0.005372920073568821, + 0.009544494561851025, + 0.0052201044745743275, + 0.005127113778144121, + 0.004204499069601297, + 0.004845006857067347 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 4, + 7 + ], + "selected_reward": -0.0076417261734604836 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the primary subject a single, shimmering teardrop sliding down skin?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the teardrop leave a visible shimmering, wet trail across the cheek?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the camera tilting upward to track the tear's path and focus on a deep brown eye?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the eyelashes visible as wet and distinctly woven together?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene rendered with a shallow depth of field and soft, diffuse lighting?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.06204906851053238 + ], + "opposite_logprob_tokens": [ + -2.812049150466919 + ], + "logprob_reward": -0.06204906851053238, + "yes_no_margin_reward": 2.75, + "token_credit_reward": -0.06204906851053238, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.03886854276061058, + 0.02740643173456192, + 0.03213496506214142, + 0.13231344521045685, + 0.07319079339504242, + 0.04685894772410393, + 0.17384743690490723, + 0.2515694200992584, + 0.08594666421413422, + 0.13786333799362183 + ], + "credit_map": [ + 0.0047294083051383495, + 0.004912272095680237, + 0.0060826693661510944, + 0.005493071861565113, + 0.006254717241972685, + 0.0027270950376987457, + 0.006728933192789555, + 0.0019403783371672034, + 0.0030133938416838646, + 0.004798091482371092, + 0.0037850334774702787, + 0.004594349302351475, + 0.0038914114702492952, + 0.0021950218360871077, + 0.0037498055025935173, + 0.0013793250545859337, + 0.0035448065027594566, + 0.003917597699910402, + 0.005176527425646782, + 0.0077134971506893635, + 0.004409442190080881, + 0.0018462167354300618, + 0.0041066431440413, + 0.0014202329330146313, + 0.013633663766086102, + 0.016839640215039253, + 0.023978417739272118, + 0.024238357320427895, + 0.024248363450169563, + 0.0045996480621397495, + 0.01967800036072731, + 0.005097357090562582, + 0.006406050641089678, + 0.007676951587200165, + 0.01458127610385418, + 0.006138524040579796, + 0.008047648705542088, + 0.004506574012339115, + 0.021545417606830597, + 0.0042883530259132385, + 0.004253383260220289, + 0.0051624407060444355, + 0.005812256596982479, + 0.012025034986436367, + 0.006003077141940594, + 0.0036077017430216074, + 0.00632845563814044, + 0.0036665969528257847, + 0.016317779198288918, + 0.01742270588874817, + 0.029688114300370216, + 0.030334053561091423, + 0.02640925906598568, + 0.019786980003118515, + 0.0225823987275362, + 0.011306142434477806, + 0.01996432989835739, + 0.01696055755019188, + 0.023387296125292778, + 0.05058594420552254, + 0.021082021296024323, + 0.05804920941591263, + 0.04173117130994797, + 0.019808871671557426, + 0.010684737004339695, + 0.0074849845841526985, + 0.014823376201093197, + 0.021043147891759872, + 0.01045219600200653, + 0.006160635966807604, + 0.009351453743875027, + 0.005946136545389891, + 0.02648942545056343, + 0.013772739097476006, + 0.013977304100990295, + 0.01828332245349884, + 0.009052636101841927, + 0.021731704473495483, + 0.007376788649708033, + 0.027179425582289696 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 6, + 9 + ], + "selected_reward": -0.06204906851053238 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the primary subject a single, shimmering teardrop sliding down skin?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the teardrop leave a visible shimmering, wet trail across the cheek?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the camera tilting upward to track the tear's path and focus on a deep brown eye?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the eyelashes visible as wet and distinctly woven together?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene rendered with a shallow depth of field and soft, diffuse lighting?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004639812279492617 + ], + "opposite_logprob_tokens": [ + -5.379639625549316 + ], + "logprob_reward": -0.004639812279492617, + "yes_no_margin_reward": 5.375, + "token_credit_reward": -0.004639812279492617, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1550809144973755, + 0.08641066402196884, + 0.0817565843462944, + 0.14713622629642487, + 0.10902805626392365, + 0.10249730199575424, + 0.09018108248710632, + 0.06998211145401001, + 0.06080480292439461, + 0.09712223708629608 + ], + "credit_map": [ + 0.015290128998458385, + 0.018398813903331757, + 0.03078112006187439, + 0.017635606229305267, + 0.021150877699255943, + 0.013909566216170788, + 0.029949814081192017, + 0.007964983582496643, + 0.009745121002197266, + 0.01884947344660759, + 0.010886945761740208, + 0.012573359534144402, + 0.017019307240843773, + 0.004443812649697065, + 0.009309081360697746, + 0.003583557903766632, + 0.010139605030417442, + 0.00894768163561821, + 0.012515612877905369, + 0.018355410546064377, + 0.011506523936986923, + 0.004640141036361456, + 0.012191087938845158, + 0.003460511565208435, + 0.011247637681663036, + 0.015948519110679626, + 0.030515393242239952, + 0.023044908419251442, + 0.02726627327501774, + 0.005014350637793541, + 0.028510408475995064, + 0.005588718689978123, + 0.009486468508839607, + 0.013521968387067318, + 0.01837853714823723, + 0.014250325970351696, + 0.011206495575606823, + 0.005636208690702915, + 0.02792803756892681, + 0.008619998581707478, + 0.01373887900263071, + 0.008522529155015945, + 0.012544319964945316, + 0.010641535744071007, + 0.017303673550486565, + 0.013113414868712425, + 0.014244814403355122, + 0.012388120405375957, + 0.010433965362608433, + 0.008077418431639671, + 0.010182756930589676, + 0.016797678545117378, + 0.01224546693265438, + 0.010782652534544468, + 0.010376599617302418, + 0.011284532025456429, + 0.0064554777927696705, + 0.0049779401160776615, + 0.006765215191990137, + 0.015769274905323982, + 0.008729908615350723, + 0.012801442295312881, + 0.00923097226768732, + 0.005251878872513771, + 0.0072128381580114365, + 0.00913515780121088, + 0.006428667809814215, + 0.01128486916422844, + 0.006121544167399406, + 0.0062340074218809605, + 0.008394838310778141, + 0.005992875434458256, + 0.009842582046985626, + 0.011096997186541557, + 0.014422430656850338, + 0.02309655211865902, + 0.010362670756876469, + 0.009304113686084747, + 0.00852678157389164, + 0.010470096953213215 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 4 + ], + "selected_reward": -0.004639812279492617 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000071/group_9/index_tensor([2654, 2654])_group_9_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000071/group_9/index_tensor([2654, 2654])_group_9_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..4b915bec6bee4fce2b4d2e8dd39dab96f318cd56 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000071/group_9/index_tensor([2654, 2654])_group_9_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "static, tight medium close-up captures the unshakable poise of an elegant courtesan, Yanggiri Tayyoung, her form a still fortress against a backdrop of pure chaos. Her face, a mask of intense emphasis accented by perfect powder-white skin, precisely painted red lips, and dark, kohl-lined eyes, remains impassive as she gazes slightly past the camera. Her meticulously crafted black hair, sculpted into a traditional, intricate hairstyle secured by ornate tortoiseshell and coral kanzashi hairpins, captures the cold, diffuse light of a stormy sky. She wears a multi-layered, heavy silk brocade kimono, its crimson fabric woven with golden cranes. Behind her, an undersea earthquake causes the ocean to erupt in violent turbulence. Massive deep blue waves, capped by foaming whitecaps overhead, roll and crash, sending plumes of spray high into the air. The entire scene is rendered with the hyper-real, crisp detail of high-speed motion photography; individual water droplets appear suspended in the air, the texture of her silk sharp. Despite the catastrophic chaos unfolding just feet away, her calm demeanor is absolute; her posture remains straight and unwavering, creating a stark contrast that creates a powerful and tense atmosphere and highlights her extraordinary grace under pressure.", + "seed": 21893, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000071/group_9/index_2654_seed_21893_dp_18_rank_36_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 9, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.06383768916130066, + "vlm_logprob": -0.06383768916130066, + "vlm_yes_no_margin": 3.4, + "vlm_token_credit": -0.06383768916130066, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static, tight medium close-up of a woman?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the woman's skin powder-white with precisely painted red lips and dark kohl-lined eyes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the woman wearing a multi-layered heavy silk brocade kimono with crimson fabric and golden cranes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background showing an undersea earthquake causing massive deep blue waves with foaming whitecaps to erupt and crash?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the visual style hyper-real and crisp with individual water droplets appearing suspended in the air?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01442551240324974 + ], + "opposite_logprob_tokens": [ + -5.389425277709961 + ], + "logprob_reward": -0.01442551240324974, + "yes_no_margin_reward": 5.375, + "token_credit_reward": -0.01442551240324974, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09012092649936676, + 0.03155039623379707, + 0.0264339130371809, + 0.12265170365571976, + 0.41176512837409973, + 0.18638116121292114, + 0.03546611964702606, + 0.029627857729792595, + 0.029475877061486244, + 0.03652695193886757 + ], + "credit_map": [ + 0.014325879514217377, + 0.007013589143753052, + 0.02344425581395626, + 0.012203293852508068, + 0.007059717550873756, + 0.00610806979238987, + 0.011587090790271759, + 0.008379032835364342, + 0.004628448747098446, + 0.0025670770555734634, + 0.00544738071039319, + 0.007063327357172966, + 0.002690767403692007, + 0.002544251037761569, + 0.004050224553793669, + 0.002558917971327901, + 0.004357694648206234, + 0.004264493938535452, + 0.0035242701414972544, + 0.00426200358197093, + 0.0025306183379143476, + 0.002471077023074031, + 0.0031771210487931967, + 0.0018466348992660642, + 0.046274732798337936, + 0.042737383395433426, + 0.007955433800816536, + 0.004945711698383093, + 0.004859543405473232, + 0.004341194871813059, + 0.007336691487580538, + 0.004201014060527086, + 0.18827736377716064, + 0.18157435953617096, + 0.006282152608036995, + 0.006328810937702656, + 0.0137430215254426, + 0.003917942754924297, + 0.006922626867890358, + 0.0047188373282551765, + 0.08140784502029419, + 0.07592611759901047, + 0.005781046114861965, + 0.005160992965102196, + 0.006777050904929638, + 0.0035388506948947906, + 0.004974208306521177, + 0.002815059619024396, + 0.005912346299737692, + 0.0058162277564406395, + 0.007423759438097477, + 0.004440346732735634, + 0.003080817172303796, + 0.002081536455079913, + 0.004760307259857655, + 0.0019507801625877619, + 0.005071855615824461, + 0.0043062446638941765, + 0.0026481577660888433, + 0.004035220947116613, + 0.0030966224148869514, + 0.004699846263974905, + 0.004227754659950733, + 0.0015421558637171984, + 0.0035604501608759165, + 0.0038600359112024307, + 0.004134603776037693, + 0.005686544347554445, + 0.0029824564699083567, + 0.0037132776342332363, + 0.004170574247837067, + 0.0013679364928975701, + 0.005516082048416138, + 0.005951818078756332, + 0.007436508312821388, + 0.004153805319219828, + 0.003281366778537631, + 0.003912564367055893, + 0.004372400231659412, + 0.0019024062203243375 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 5, + 3 + ], + "selected_reward": -0.01442551240324974 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static, tight medium close-up of a woman?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the woman's skin powder-white with precisely painted red lips and dark kohl-lined eyes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the woman wearing a multi-layered heavy silk brocade kimono with crimson fabric and golden cranes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background showing an undersea earthquake causing massive deep blue waves with foaming whitecaps to erupt and crash?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the visual style hyper-real and crisp with individual water droplets appearing suspended in the air?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.009810801595449448 + ], + "opposite_logprob_tokens": [ + -4.634810924530029 + ], + "logprob_reward": -0.009810801595449448, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.009810801595449448, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3641287386417389, + 0.1732630580663681, + 0.05698559433221817, + 0.0837046280503273, + 0.051212117075920105, + 0.059164877980947495, + 0.0477769672870636, + 0.061030320823192596, + 0.06075252965092659, + 0.04198117554187775 + ], + "credit_map": [ + 0.051090143620967865, + 0.03832053765654564, + 0.06865763664245605, + 0.04100324213504791, + 0.034528881311416626, + 0.039512671530246735, + 0.05614455044269562, + 0.034871090203523636, + 0.026412734761834145, + 0.01587984710931778, + 0.03498705476522446, + 0.03470306843519211, + 0.01222248561680317, + 0.012313316576182842, + 0.022680463269352913, + 0.014064094983041286, + 0.008551065810024738, + 0.009118780493736267, + 0.011709790676832199, + 0.007621858734637499, + 0.004630180075764656, + 0.0038552836049348116, + 0.007090561557561159, + 0.00440807081758976, + 0.019205257296562195, + 0.016854576766490936, + 0.01110974233597517, + 0.012188120745122433, + 0.004517372231930494, + 0.006303961388766766, + 0.007714357227087021, + 0.005811240058392286, + 0.007774898782372475, + 0.008365530520677567, + 0.009324206039309502, + 0.007483059540390968, + 0.0041313981637358665, + 0.003513180883601308, + 0.007033635396510363, + 0.0035862093791365623, + 0.00895723421126604, + 0.010074570775032043, + 0.006818301975727081, + 0.008928676135838032, + 0.005473138764500618, + 0.006434254348278046, + 0.008077255450189114, + 0.004401443060487509, + 0.006494731176644564, + 0.007023000158369541, + 0.009921434335410595, + 0.007661790121346712, + 0.0038233157247304916, + 0.003633496817201376, + 0.006759021896868944, + 0.0024601740296930075, + 0.008251785300672054, + 0.00842376984655857, + 0.007864396087825298, + 0.008250897750258446, + 0.005850956309586763, + 0.006262659095227718, + 0.011886336840689182, + 0.004239518195390701, + 0.007890548557043076, + 0.009560899809002876, + 0.009683948941528797, + 0.01597404107451439, + 0.004887425806373358, + 0.0031699787359684706, + 0.005664841737598181, + 0.003920842427760363, + 0.007693698164075613, + 0.007109770085662603, + 0.008074491284787655, + 0.005498982034623623, + 0.002827503252774477, + 0.0033803873229771852, + 0.004589345771819353, + 0.0028069932013750076 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 3 + ], + "selected_reward": -0.009810801595449448 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static, tight medium close-up of a woman?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the woman's skin powder-white with precisely painted red lips and dark kohl-lined eyes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the woman wearing a multi-layered heavy silk brocade kimono with crimson fabric and golden cranes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background showing an undersea earthquake causing massive deep blue waves with foaming whitecaps to erupt and crash?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the visual style hyper-real and crisp with individual water droplets appearing suspended in the air?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.05496445298194885 + ], + "opposite_logprob_tokens": [ + -2.929964542388916 + ], + "logprob_reward": -0.05496445298194885, + "yes_no_margin_reward": 2.875, + "token_credit_reward": -0.05496445298194885, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1617431491613388, + 0.05415499210357666, + 0.15931770205497742, + 0.3598223626613617, + 0.09265337139368057, + 0.05866082012653351, + 0.011692766100168228, + 0.014976289123296738, + 0.04057296738028526, + 0.04640556499361992 + ], + "credit_map": [ + 0.02157607302069664, + 0.014957090839743614, + 0.027936965227127075, + 0.024465011432766914, + 0.012889686971902847, + 0.017374487593770027, + 0.026755856350064278, + 0.015787988901138306, + 0.007277477532625198, + 0.005338067188858986, + 0.0074200802482664585, + 0.010236148722469807, + 0.005148013588041067, + 0.0042867232114076614, + 0.009891943074762821, + 0.0045565394684672356, + 0.026370253413915634, + 0.026806706562638283, + 0.036423809826374054, + 0.02063387632369995, + 0.009465285576879978, + 0.010320719331502914, + 0.01719769649207592, + 0.01209938619285822, + 0.11045021563768387, + 0.10926163196563721, + 0.02661762945353985, + 0.051458146423101425, + 0.013912648893892765, + 0.012896157801151276, + 0.02144221030175686, + 0.013783764094114304, + 0.038796715438365936, + 0.03764165937900543, + 0.0039519802667200565, + 0.0028095406014472246, + 0.003121620509773493, + 0.001561561250127852, + 0.002743280027061701, + 0.0020270266104489565, + 0.0180642232298851, + 0.017041832208633423, + 0.004224149044603109, + 0.00589462835341692, + 0.0033427714370191097, + 0.0031180449295789003, + 0.004430202767252922, + 0.002544968156144023, + 0.0020923521369695663, + 0.0020353978034108877, + 0.0014325667871162295, + 0.001341884140856564, + 0.0012259732466191053, + 0.0015813393983989954, + 0.0012878761626780033, + 0.0006953775882720947, + 0.002276861574500799, + 0.001979537308216095, + 0.001945794909261167, + 0.0016777775017544627, + 0.0013044324005022645, + 0.0015019518323242664, + 0.0027570032980293036, + 0.0015329313464462757, + 0.005977306514978409, + 0.0064855231903493404, + 0.006105185952037573, + 0.010088099166750908, + 0.0033013690263032913, + 0.0027866647578775883, + 0.0037026009522378445, + 0.0021262208465486765, + 0.007676809094846249, + 0.00793335773050785, + 0.007796799764037132, + 0.008739258162677288, + 0.00258367951028049, + 0.003749899799004197, + 0.004321787040680647, + 0.0036039778497070074 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 0, + 2 + ], + "selected_reward": -0.05496445298194885 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static, tight medium close-up of a woman?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the woman's skin powder-white with precisely painted red lips and dark kohl-lined eyes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the woman wearing a multi-layered heavy silk brocade kimono with crimson fabric and golden cranes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background showing an undersea earthquake causing massive deep blue waves with foaming whitecaps to erupt and crash?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the visual style hyper-real and crisp with individual water droplets appearing suspended in the air?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.12704521417617798 + ], + "opposite_logprob_tokens": [ + -2.127045154571533 + ], + "logprob_reward": -0.12704521417617798, + "yes_no_margin_reward": 2.0, + "token_credit_reward": -0.12704521417617798, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0461890771985054, + 0.08503446727991104, + 0.09236917644739151, + 0.07698804885149002, + 0.5870677828788757, + 0.042978256940841675, + 0.032492805272340775, + 0.017976270988583565, + 0.011063125915825367, + 0.007840973325073719 + ], + "credit_map": [ + 0.0070448266342282295, + 0.0036646039225161076, + 0.012297563254833221, + 0.004498348105698824, + 0.003689954988658428, + 0.0025891850236803293, + 0.00820178259164095, + 0.004202812444418669, + 0.01487505342811346, + 0.008709684014320374, + 0.02403801865875721, + 0.01134245190769434, + 0.005038638599216938, + 0.0043523870408535, + 0.007422159891575575, + 0.009256073273718357, + 0.02214558608829975, + 0.018499154597520828, + 0.024086805060505867, + 0.006578267551958561, + 0.004174812231212854, + 0.003996696788817644, + 0.006457353942096233, + 0.006430500652641058, + 0.022864999249577522, + 0.02283278852701187, + 0.007266053464263678, + 0.008944693952798843, + 0.003553934395313263, + 0.002597440266981721, + 0.005099951289594173, + 0.0038281860761344433, + 0.25621700286865234, + 0.2490871548652649, + 0.020744256675243378, + 0.011733287014067173, + 0.01993466727435589, + 0.006802673451602459, + 0.011049866676330566, + 0.011498872190713882, + 0.013081871904432774, + 0.012343624606728554, + 0.0028883940540254116, + 0.002908288035541773, + 0.0029501027893275023, + 0.0028058579191565514, + 0.004053641110658646, + 0.0019464809447526932, + 0.004680490121245384, + 0.005091438535600901, + 0.007825124077498913, + 0.004428130108863115, + 0.0022508762776851654, + 0.0014789925189688802, + 0.004978939890861511, + 0.0017588137416169047, + 0.0032038060016930103, + 0.0026748948730528355, + 0.0036385718267410994, + 0.002274107886478305, + 0.0015821533743292093, + 0.0014719724422320724, + 0.00233028014190495, + 0.0008004833944141865, + 0.001760469051077962, + 0.0016613309271633625, + 0.0010536012705415487, + 0.0017790264682844281, + 0.0010680188424885273, + 0.0012874472886323929, + 0.0018288494320586324, + 0.0006243821699172258, + 0.0010039621265605092, + 0.00107312120962888, + 0.001275727991014719, + 0.0011245198547840118, + 0.000750178936868906, + 0.0010332469828426838, + 0.0010328016942366958, + 0.0005474149365909398 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 2, + 1 + ], + "selected_reward": -0.12704521417617798 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static, tight medium close-up of a woman?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the woman's skin powder-white with precisely painted red lips and dark kohl-lined eyes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the woman wearing a multi-layered heavy silk brocade kimono with crimson fabric and golden cranes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background showing an undersea earthquake causing massive deep blue waves with foaming whitecaps to erupt and crash?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the visual style hyper-real and crisp with individual water droplets appearing suspended in the air?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.11294246464967728 + ], + "opposite_logprob_tokens": [ + -2.2379424571990967 + ], + "logprob_reward": -0.11294246464967728, + "yes_no_margin_reward": 2.125, + "token_credit_reward": -0.11294246464967728, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.23991048336029053, + 0.08811120688915253, + 0.031032681465148926, + 0.18471622467041016, + 0.1739983707666397, + 0.14465472102165222, + 0.046073395758867264, + 0.04480304941534996, + 0.0224554892629385, + 0.024244297295808792 + ], + "credit_map": [ + 0.034938231110572815, + 0.019010305404663086, + 0.02937258966267109, + 0.059430427849292755, + 0.022069979459047318, + 0.023436538875102997, + 0.033802151679992676, + 0.01785026304423809, + 0.012734429910779, + 0.007537658791989088, + 0.015205273404717445, + 0.024666592478752136, + 0.0064189983531832695, + 0.004482883494347334, + 0.009672497399151325, + 0.007392884697765112, + 0.004743019584566355, + 0.0047195469960570335, + 0.005161404609680176, + 0.004211329389363527, + 0.002937243552878499, + 0.003032161621376872, + 0.0036058633122593164, + 0.002622115658596158, + 0.07030755281448364, + 0.06609747558832169, + 0.01055984478443861, + 0.013397961854934692, + 0.006767064332962036, + 0.0049980441108345985, + 0.007852067239582539, + 0.004736239090561867, + 0.0670006200671196, + 0.0630890280008316, + 0.014861547388136387, + 0.007533198222517967, + 0.0061088078655302525, + 0.004495578818023205, + 0.00709322327747941, + 0.0038163817953318357, + 0.057566240429878235, + 0.05604807659983635, + 0.006901233922690153, + 0.0057690381072461605, + 0.00561398733407259, + 0.004646980669349432, + 0.0046524167992174625, + 0.0034567585680633783, + 0.005657155066728592, + 0.005751173011958599, + 0.008608168922364712, + 0.00684758834540844, + 0.004683473147451878, + 0.0042809550650417805, + 0.007628857158124447, + 0.002616029465571046, + 0.007116497494280338, + 0.006356893107295036, + 0.004630726296454668, + 0.005951541010290384, + 0.004460712429136038, + 0.006053202319890261, + 0.0071049099788069725, + 0.003128569573163986, + 0.003081568516790867, + 0.00350589700974524, + 0.0026391707360744476, + 0.0045777312479913235, + 0.0019533028826117516, + 0.0017091312911361456, + 0.003118226770311594, + 0.0018704639514908195, + 0.003161497414112091, + 0.003279264084994793, + 0.0029542003758251667, + 0.002213990082964301, + 0.0030708799604326487, + 0.0035015002358704805, + 0.003988837357610464, + 0.002074130577966571 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 4 + ], + "selected_reward": -0.11294246464967728 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "static, tight medium close-up captures the unshakable poise of an elegant courtesan, Yanggiri Tayyoung, her form a still fortress against a backdrop of pure chaos. Her face, a mask of intense emphasis accented by perfect powder-white skin, precisely painted red lips, and dark, kohl-lined eyes, remains impassive as she gazes slightly past the camera. Her meticulously crafted black hair, sculpted into a traditional, intricate hairstyle secured by ornate tortoiseshell and coral kanzashi hairpins, captures the cold, diffuse light of a stormy sky. She wears a multi-layered, heavy silk brocade kimono, its crimson fabric woven with golden cranes. Behind her, an undersea earthquake causes the ocean to erupt in violent turbulence. Massive deep blue waves, capped by foaming whitecaps overhead, roll and crash, sending plumes of spray high into the air. The entire scene is rendered with the hyper-real, crisp detail of high-speed motion photography; individual water droplets appear suspended in the air, the texture of her silk sharp. Despite the catastrophic chaos unfolding just feet away, her calm demeanor is absolute; her posture remains straight and unwavering, creating a stark contrast that creates a powerful and tense atmosphere and highlights her extraordinary grace under pressure.", + "seed": 21894, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000071/group_9/index_2654_seed_21894_dp_18_rank_36_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 9, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.04011767357587814, + "vlm_logprob": -0.04011767357587814, + "vlm_yes_no_margin": 3.775, + "vlm_token_credit": -0.04011767357587814, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static, tight medium close-up of a woman?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the woman's skin powder-white with precisely painted red lips and dark kohl-lined eyes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the woman wearing a multi-layered heavy silk brocade kimono with crimson fabric and golden cranes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background showing an undersea earthquake causing massive deep blue waves with foaming whitecaps to erupt and crash?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the visual style hyper-real and crisp with individual water droplets appearing suspended in the air?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.007843528874218464 + ], + "opposite_logprob_tokens": [ + -5.757843494415283 + ], + "logprob_reward": -0.007843528874218464, + "yes_no_margin_reward": 5.75, + "token_credit_reward": -0.007843528874218464, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.199923574924469, + 0.26534393429756165, + 0.17750787734985352, + 0.028610145673155785, + 0.07900190353393555, + 0.05795158073306084, + 0.04385039210319519, + 0.04944189637899399, + 0.05276705324649811, + 0.04560162499547005 + ], + "credit_map": [ + 0.026815427467226982, + 0.02520318701863289, + 0.016639959067106247, + 0.014657407067716122, + 0.023251278325915337, + 0.02947372943162918, + 0.016611013561487198, + 0.04727157950401306, + 0.047759804874658585, + 0.021550998091697693, + 0.03256271779537201, + 0.039130799472332, + 0.016518281772732735, + 0.036664705723524094, + 0.020884519442915916, + 0.05027211457490921, + 0.02093852497637272, + 0.015922561287879944, + 0.025811120867729187, + 0.024674158543348312, + 0.010609671473503113, + 0.023195620626211166, + 0.027195796370506287, + 0.029160428792238235, + 0.004066151101142168, + 0.0027283206582069397, + 0.003394964151084423, + 0.00380940199829638, + 0.0026071849279105663, + 0.00421467050909996, + 0.003420358058065176, + 0.004369093105196953, + 0.009838961996138096, + 0.011608664877712727, + 0.010724161751568317, + 0.009149034507572651, + 0.006425702478736639, + 0.009426338598132133, + 0.008130335249006748, + 0.013698703609406948, + 0.010999995283782482, + 0.006028163246810436, + 0.011824803426861763, + 0.004327342379838228, + 0.003519205143675208, + 0.008116203360259533, + 0.003959779627621174, + 0.009176090359687805, + 0.007726476993411779, + 0.003283063182607293, + 0.004831604193896055, + 0.005364651791751385, + 0.003367729252204299, + 0.00731714116409421, + 0.005768908187747002, + 0.006190813612192869, + 0.005961224902421236, + 0.005221870262175798, + 0.004186970181763172, + 0.009840707294642925, + 0.0032183262519538403, + 0.006215312983840704, + 0.00879440177232027, + 0.006003083661198616, + 0.008470485918223858, + 0.003741235937923193, + 0.006041042041033506, + 0.0092952074483037, + 0.0031162071973085403, + 0.00789534579962492, + 0.008084427565336227, + 0.006123095750808716, + 0.006741042248904705, + 0.0042260270565748215, + 0.004397563170641661, + 0.008162584155797958, + 0.003649341408163309, + 0.007798215374350548, + 0.006982632912695408, + 0.0036442202981561422 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 2 + ], + "selected_reward": -0.007843528874218464 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static, tight medium close-up of a woman?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the woman's skin powder-white with precisely painted red lips and dark kohl-lined eyes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the woman wearing a multi-layered heavy silk brocade kimono with crimson fabric and golden cranes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background showing an undersea earthquake causing massive deep blue waves with foaming whitecaps to erupt and crash?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the visual style hyper-real and crisp with individual water droplets appearing suspended in the air?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.008661913685500622 + ], + "opposite_logprob_tokens": [ + -4.75866174697876 + ], + "logprob_reward": -0.008661913685500622, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.008661913685500622, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.21027705073356628, + 0.4054338037967682, + 0.09189075976610184, + 0.07842791825532913, + 0.043031152337789536, + 0.0237968098372221, + 0.05903388187289238, + 0.03091461770236492, + 0.04096687212586403, + 0.01622714102268219 + ], + "credit_map": [ + 0.03769340366125107, + 0.018595851957798004, + 0.01589314453303814, + 0.020824648439884186, + 0.015455527231097221, + 0.03254208713769913, + 0.017481595277786255, + 0.051790788769721985, + 0.07870879024267197, + 0.03669673949480057, + 0.06559012830257416, + 0.04675376042723656, + 0.031145019456744194, + 0.05192659795284271, + 0.02847149409353733, + 0.06614126265048981, + 0.012108647264540195, + 0.010177834890782833, + 0.010027823969721794, + 0.012841211631894112, + 0.011678430251777172, + 0.011292176321148872, + 0.014388062991201878, + 0.009376573376357555, + 0.00933974701911211, + 0.013299018144607544, + 0.00872084591537714, + 0.012733398005366325, + 0.007429841905832291, + 0.008958055637776852, + 0.01192266121506691, + 0.006024348083883524, + 0.008199241012334824, + 0.0038834339939057827, + 0.008119774051010609, + 0.003810682101175189, + 0.0030681442003697157, + 0.006181993056088686, + 0.0036376453936100006, + 0.006130238063633442, + 0.002748789032921195, + 0.0042214784771203995, + 0.0035220100544393063, + 0.0030284372624009848, + 0.002273107413202524, + 0.0028088546823710203, + 0.003133869031444192, + 0.0020602643489837646, + 0.009374032728374004, + 0.004046325106173754, + 0.0075686415657401085, + 0.007881175726652145, + 0.0038426450919359922, + 0.009679376147687435, + 0.008980239741504192, + 0.007661445531994104, + 0.0032214440871030092, + 0.0033109516371041536, + 0.0035019596107304096, + 0.005862703081220388, + 0.0024029414635151625, + 0.003521320875734091, + 0.005795659497380257, + 0.0032976381480693817, + 0.006316119339317083, + 0.003282000543549657, + 0.004137236624956131, + 0.007170058321207762, + 0.00270196795463562, + 0.005287196487188339, + 0.007420571520924568, + 0.004651716910302639, + 0.0023320515174418688, + 0.002397920936346054, + 0.001699128421023488, + 0.00200275587849319, + 0.00222438620403409, + 0.0023723577614873648, + 0.0018386727897450328, + 0.00135986867826432 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 2 + ], + "selected_reward": -0.008661913685500622 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static, tight medium close-up of a woman?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the woman's skin powder-white with precisely painted red lips and dark kohl-lined eyes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the woman wearing a multi-layered heavy silk brocade kimono with crimson fabric and golden cranes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background showing an undersea earthquake causing massive deep blue waves with foaming whitecaps to erupt and crash?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the visual style hyper-real and crisp with individual water droplets appearing suspended in the air?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.043068766593933105 + ], + "opposite_logprob_tokens": [ + -3.1680688858032227 + ], + "logprob_reward": -0.043068766593933105, + "yes_no_margin_reward": 3.125, + "token_credit_reward": -0.043068766593933105, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.05780875310301781, + 0.7623329162597656, + 0.04652368649840355, + 0.048826079815626144, + 0.02216344140470028, + 0.012183304876089096, + 0.010279015637934208, + 0.026762252673506737, + 0.006919019855558872, + 0.006201573647558689 + ], + "credit_map": [ + 0.006531564984470606, + 0.007298719137907028, + 0.006120915990322828, + 0.006793019827455282, + 0.0048744515515863895, + 0.00782766006886959, + 0.004854321479797363, + 0.013508101925253868, + 0.13600943982601166, + 0.0703878179192543, + 0.12263749539852142, + 0.06891288608312607, + 0.06366222351789474, + 0.09044715017080307, + 0.08209149539470673, + 0.12818434834480286, + 0.006322933826595545, + 0.0037522355560213327, + 0.004963398445397615, + 0.005114822648465633, + 0.005185690708458424, + 0.0055029187351465225, + 0.006879198364913464, + 0.008802487514913082, + 0.006487058009952307, + 0.0063733323477208614, + 0.005389418452978134, + 0.007724031340330839, + 0.0038971418980509043, + 0.0058493646793067455, + 0.00830812193453312, + 0.004797609988600016, + 0.003008418483659625, + 0.0024172570556402206, + 0.0029390216805040836, + 0.0035782165359705687, + 0.0016476581804454327, + 0.002375526586547494, + 0.004060645122081041, + 0.0021366970613598824, + 0.0018644774099811912, + 0.001765255001373589, + 0.0014724517241120338, + 0.0016529103741049767, + 0.0011713509447872639, + 0.0015048326458781958, + 0.001643864787183702, + 0.0011081615230068564, + 0.0013116460759192705, + 0.0007747695199213922, + 0.0012439677957445383, + 0.0018293391913175583, + 0.0006893485551699996, + 0.001480364939197898, + 0.0020663829054683447, + 0.0008831965387798846, + 0.0033308048732578754, + 0.0023833189625293016, + 0.0030854062642902136, + 0.0030963565222918987, + 0.002341450657695532, + 0.00365982623770833, + 0.0029935177881270647, + 0.00587157066911459, + 0.0009564476786181331, + 0.000725177233107388, + 0.0006178781040944159, + 0.0011085832957178354, + 0.0006368950707837939, + 0.0008378903730772436, + 0.0013002887135371566, + 0.0007358595612458885, + 0.0010203057900071144, + 0.0007994544575922191, + 0.00048141294973902404, + 0.0007374874548986554, + 0.0008850157610140741, + 0.0010080842766910791, + 0.0006308006122708321, + 0.0006390121416188776 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 3 + ], + "selected_reward": -0.043068766593933105 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static, tight medium close-up of a woman?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the woman's skin powder-white with precisely painted red lips and dark kohl-lined eyes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the woman wearing a multi-layered heavy silk brocade kimono with crimson fabric and golden cranes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background showing an undersea earthquake causing massive deep blue waves with foaming whitecaps to erupt and crash?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the visual style hyper-real and crisp with individual water droplets appearing suspended in the air?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07896564900875092 + ], + "opposite_logprob_tokens": [ + -2.578965663909912 + ], + "logprob_reward": -0.07896564900875092, + "yes_no_margin_reward": 2.5, + "token_credit_reward": -0.07896564900875092, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.16051682829856873, + 0.6837940812110901, + 0.0254007950425148, + 0.03579450398683548, + 0.01341469306498766, + 0.011309600435197353, + 0.043526530265808105, + 0.004076176322996616, + 0.012725098058581352, + 0.009441728703677654 + ], + "credit_map": [ + 0.021364040672779083, + 0.011490988545119762, + 0.020035240799188614, + 0.015344981104135513, + 0.01169921737164259, + 0.021882005035877228, + 0.01273268274962902, + 0.04596766456961632, + 0.13080401718616486, + 0.05769171193242073, + 0.11223910003900528, + 0.06036118417978287, + 0.054595571011304855, + 0.08484643697738647, + 0.061265911906957626, + 0.12199006974697113, + 0.0038882510270923376, + 0.0013549491995945573, + 0.004578041844069958, + 0.002362776082009077, + 0.0015500775771215558, + 0.004055908881127834, + 0.002934491029009223, + 0.004676297307014465, + 0.005239497404545546, + 0.0031929260585457087, + 0.005397598259150982, + 0.004517346154898405, + 0.0023611842188984156, + 0.004813018254935741, + 0.0050730532966554165, + 0.00519987940788269, + 0.002028788672760129, + 0.0013348308857530355, + 0.0021698540076613426, + 0.0019429242238402367, + 0.0011926754377782345, + 0.0016947436379268765, + 0.0016272186767309904, + 0.0014236577553674579, + 0.0019486617529764771, + 0.0008822479285299778, + 0.0023151198402047157, + 0.0011508141178637743, + 0.0006008908385410905, + 0.001338602160103619, + 0.0010713384253904223, + 0.002001924207434058, + 0.005957712419331074, + 0.003299358068034053, + 0.003999394830316305, + 0.007902806624770164, + 0.0026041457895189524, + 0.006643719971179962, + 0.00903585460036993, + 0.004083534702658653, + 0.0006533293635584414, + 0.00035987625597044826, + 0.00045830017188563943, + 0.0005215091514401138, + 0.0002846459101419896, + 0.0005531405913643539, + 0.0004578100342769176, + 0.0007875647279433906, + 0.0018397452076897025, + 0.0010389366652816534, + 0.0015493363607674837, + 0.0023456504568457603, + 0.0009124689386226237, + 0.0018706346163526177, + 0.001972524682059884, + 0.001195798977278173, + 0.0024544978514313698, + 0.0005557183176279068, + 0.0006416199030354619, + 0.0006674824981018901, + 0.0007089534192346036, + 0.002195425098761916, + 0.0008217432769015431, + 0.0013962879311293364 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 6 + ], + "selected_reward": -0.07896564900875092 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static, tight medium close-up of a woman?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the woman's skin powder-white with precisely painted red lips and dark kohl-lined eyes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the woman wearing a multi-layered heavy silk brocade kimono with crimson fabric and golden cranes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background showing an undersea earthquake causing massive deep blue waves with foaming whitecaps to erupt and crash?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the visual style hyper-real and crisp with individual water droplets appearing suspended in the air?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.06204850971698761 + ], + "opposite_logprob_tokens": [ + -2.8120484352111816 + ], + "logprob_reward": -0.06204850971698761, + "yes_no_margin_reward": 2.75, + "token_credit_reward": -0.06204850971698761, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.28049859404563904, + 0.3041059076786041, + 0.09591219574213028, + 0.07179249078035355, + 0.06588011234998703, + 0.04446438327431679, + 0.05036502704024315, + 0.023342888802289963, + 0.03310297057032585, + 0.030535463243722916 + ], + "credit_map": [ + 0.03490355610847473, + 0.03412953391671181, + 0.026540396735072136, + 0.025037754327058792, + 0.032052282243967056, + 0.03664272651076317, + 0.026802130043506622, + 0.06439020484685898, + 0.06015146151185036, + 0.020758826285600662, + 0.040824562311172485, + 0.034663885831832886, + 0.01897389069199562, + 0.04230697453022003, + 0.02287771925330162, + 0.06354856491088867, + 0.011876806616783142, + 0.008229974657297134, + 0.015475532039999962, + 0.013387832790613174, + 0.00689076678827405, + 0.012881596572697163, + 0.01365453191101551, + 0.01351514458656311, + 0.00963759794831276, + 0.00509360758587718, + 0.010480186901986599, + 0.009904227219522, + 0.004810765385627747, + 0.009718252345919609, + 0.01165382843464613, + 0.010494020767509937, + 0.009622366167604923, + 0.007707907352596521, + 0.014033674262464046, + 0.008394261822104454, + 0.004491425585001707, + 0.0074416641145944595, + 0.006669799331575632, + 0.0075190067291259766, + 0.0051948437467217445, + 0.0067694904282689095, + 0.005317070987075567, + 0.007729416247457266, + 0.0038430725689977407, + 0.0049575078301131725, + 0.007134136278182268, + 0.0035188382025808096, + 0.00783984363079071, + 0.005256542470306158, + 0.005121628288179636, + 0.006521175615489483, + 0.003760719671845436, + 0.00738215958699584, + 0.006397643126547337, + 0.008085307665169239, + 0.0036408905871212482, + 0.002502196468412876, + 0.002347946399822831, + 0.0032005240209400654, + 0.0019041962223127484, + 0.0035088073927909136, + 0.0025437811855226755, + 0.0036945443134754896, + 0.005357849877327681, + 0.002377893775701523, + 0.0033279876224696636, + 0.005592918023467064, + 0.0024652068968862295, + 0.0058317361399531364, + 0.005341759417206049, + 0.002807616489008069, + 0.005680467467755079, + 0.0018598962342366576, + 0.0024577518925070763, + 0.005003687925636768, + 0.0016123529057949781, + 0.00651696277782321, + 0.004807051736861467, + 0.0025972893927246332 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 2 + ], + "selected_reward": -0.06204850971698761 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "static, tight medium close-up captures the unshakable poise of an elegant courtesan, Yanggiri Tayyoung, her form a still fortress against a backdrop of pure chaos. Her face, a mask of intense emphasis accented by perfect powder-white skin, precisely painted red lips, and dark, kohl-lined eyes, remains impassive as she gazes slightly past the camera. Her meticulously crafted black hair, sculpted into a traditional, intricate hairstyle secured by ornate tortoiseshell and coral kanzashi hairpins, captures the cold, diffuse light of a stormy sky. She wears a multi-layered, heavy silk brocade kimono, its crimson fabric woven with golden cranes. Behind her, an undersea earthquake causes the ocean to erupt in violent turbulence. Massive deep blue waves, capped by foaming whitecaps overhead, roll and crash, sending plumes of spray high into the air. The entire scene is rendered with the hyper-real, crisp detail of high-speed motion photography; individual water droplets appear suspended in the air, the texture of her silk sharp. Despite the catastrophic chaos unfolding just feet away, her calm demeanor is absolute; her posture remains straight and unwavering, creating a stark contrast that creates a powerful and tense atmosphere and highlights her extraordinary grace under pressure.", + "seed": 21895, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000071/group_9/index_2654_seed_21895_dp_19_rank_38_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 9, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.041198995988816024, + "vlm_logprob": -0.041198995988816024, + "vlm_yes_no_margin": 4.175, + "vlm_token_credit": -0.041198995988816024, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static, tight medium close-up of a woman?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the woman's skin powder-white with precisely painted red lips and dark kohl-lined eyes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the woman wearing a multi-layered heavy silk brocade kimono with crimson fabric and golden cranes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background showing an undersea earthquake causing massive deep blue waves with foaming whitecaps to erupt and crash?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the visual style hyper-real and crisp with individual water droplets appearing suspended in the air?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01801122911274433 + ], + "opposite_logprob_tokens": [ + -6.268011093139648 + ], + "logprob_reward": -0.01801122911274433, + "yes_no_margin_reward": 6.25, + "token_credit_reward": -0.01801122911274433, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.23655171692371368, + 0.15171372890472412, + 0.10217804461717606, + 0.08621557801961899, + 0.04494152218103409, + 0.05732659995555878, + 0.06374119222164154, + 0.08118364959955215, + 0.10598387569189072, + 0.0701640248298645 + ], + "credit_map": [ + 0.030713647603988647, + 0.036136019974946976, + 0.03271622583270073, + 0.02374493144452572, + 0.023636480793356895, + 0.029843231663107872, + 0.025976335629820824, + 0.03378484398126602, + 0.012592483311891556, + 0.016486650332808495, + 0.03766200318932533, + 0.01355055533349514, + 0.013441202230751514, + 0.023302171379327774, + 0.014281964860856533, + 0.020396707579493523, + 0.022930413484573364, + 0.012443230487406254, + 0.015517073683440685, + 0.00648884754627943, + 0.007354945410043001, + 0.01479701604694128, + 0.007057329639792442, + 0.015589183196425438, + 0.013447863981127739, + 0.010494240559637547, + 0.008031859993934631, + 0.01229565218091011, + 0.00785466842353344, + 0.009785160422325134, + 0.010165502317249775, + 0.01414063386619091, + 0.006627738010138273, + 0.007447823416441679, + 0.005234073381870985, + 0.004729259293526411, + 0.004520454443991184, + 0.005560430698096752, + 0.004356821067631245, + 0.006464922800660133, + 0.006163925863802433, + 0.005479883402585983, + 0.005574642214924097, + 0.008977334015071392, + 0.005180411506444216, + 0.008030048571527004, + 0.007637144066393375, + 0.010283215902745724, + 0.005920156836509705, + 0.006349857896566391, + 0.007201558910310268, + 0.00964872445911169, + 0.0075986795127391815, + 0.006609449628740549, + 0.011080275289714336, + 0.009332488290965557, + 0.009734763763844967, + 0.01330806314945221, + 0.00626271590590477, + 0.012430679984390736, + 0.007690283004194498, + 0.007557522971183062, + 0.012382304295897484, + 0.011817318387329578, + 0.014914306811988354, + 0.011316115036606789, + 0.007701602298766375, + 0.013570667244493961, + 0.00666462816298008, + 0.018311358988285065, + 0.015049593523144722, + 0.01845559850335121, + 0.01051165908575058, + 0.011850313283503056, + 0.009307116270065308, + 0.008507339283823967, + 0.005465511232614517, + 0.009354300796985626, + 0.006303712725639343, + 0.00886407308280468 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 8 + ], + "selected_reward": -0.01801122911274433 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static, tight medium close-up of a woman?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the woman's skin powder-white with precisely painted red lips and dark kohl-lined eyes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the woman wearing a multi-layered heavy silk brocade kimono with crimson fabric and golden cranes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background showing an undersea earthquake causing massive deep blue waves with foaming whitecaps to erupt and crash?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the visual style hyper-real and crisp with individual water droplets appearing suspended in the air?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004666747059673071 + ], + "opposite_logprob_tokens": [ + -5.379666805267334 + ], + "logprob_reward": -0.004666747059673071, + "yes_no_margin_reward": 5.375, + "token_credit_reward": -0.004666747059673071, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1997593641281128, + 0.1822432577610016, + 0.10680361092090607, + 0.07173769176006317, + 0.06485626846551895, + 0.048744283616542816, + 0.08317074179649353, + 0.1002056822180748, + 0.07250766456127167, + 0.06997144222259521 + ], + "credit_map": [ + 0.026080414652824402, + 0.03925732523202896, + 0.018049435690045357, + 0.019330067560076714, + 0.022808535024523735, + 0.023207809776067734, + 0.021969731897115707, + 0.029056044295430183, + 0.017963282763957977, + 0.03261823579668999, + 0.031110312789678574, + 0.018377766013145447, + 0.014082114212214947, + 0.022954585030674934, + 0.01736370660364628, + 0.027773257344961166, + 0.01085568405687809, + 0.013961187563836575, + 0.01837896928191185, + 0.012566600926220417, + 0.008812816813588142, + 0.014130471274256706, + 0.011618404649198055, + 0.016479460522532463, + 0.007038493640720844, + 0.0149362338706851, + 0.007556614466011524, + 0.009625891223549843, + 0.007660724688321352, + 0.007478878367692232, + 0.007971839047968388, + 0.009469005279242992, + 0.011626985855400562, + 0.006132594309747219, + 0.01062821689993143, + 0.005471004638820887, + 0.004349166061729193, + 0.008075272664427757, + 0.005860601086169481, + 0.01271242555230856, + 0.005958146881312132, + 0.005230470560491085, + 0.00427227234467864, + 0.006594211328774691, + 0.003974242601543665, + 0.006644918583333492, + 0.009079410694539547, + 0.006990605499595404, + 0.007873689755797386, + 0.010717429220676422, + 0.01099605206400156, + 0.008196990936994553, + 0.007344772107899189, + 0.00934546161442995, + 0.011389433406293392, + 0.01730690523982048, + 0.012046252377331257, + 0.014708243310451508, + 0.01797519065439701, + 0.011353827081620693, + 0.007204884197562933, + 0.011956604197621346, + 0.00865660049021244, + 0.016304071992635727, + 0.00861341506242752, + 0.008970933966338634, + 0.006954587530344725, + 0.00901731289923191, + 0.006602128967642784, + 0.009638967923820019, + 0.010624194517731667, + 0.01208611112087965, + 0.008696908131241798, + 0.010970340110361576, + 0.01048195268958807, + 0.007788701448589563, + 0.005748212803155184, + 0.008802180178463459, + 0.006008357275277376, + 0.011474779807031155 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 2 + ], + "selected_reward": -0.004666747059673071 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static, tight medium close-up of a woman?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the woman's skin powder-white with precisely painted red lips and dark kohl-lined eyes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the woman wearing a multi-layered heavy silk brocade kimono with crimson fabric and golden cranes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background showing an undersea earthquake causing massive deep blue waves with foaming whitecaps to erupt and crash?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the visual style hyper-real and crisp with individual water droplets appearing suspended in the air?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.018189528957009315 + ], + "opposite_logprob_tokens": [ + -4.018189430236816 + ], + "logprob_reward": -0.018189528957009315, + "yes_no_margin_reward": 4.0, + "token_credit_reward": -0.018189528957009315, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.22853083908557892, + 0.1647733598947525, + 0.07525821030139923, + 0.07364198565483093, + 0.06985295563936234, + 0.13763563334941864, + 0.0495443120598793, + 0.10376426577568054, + 0.06081513315439224, + 0.036183297634124756 + ], + "credit_map": [ + 0.01959276758134365, + 0.03320985659956932, + 0.03138856962323189, + 0.027426574379205704, + 0.030995840206742287, + 0.02621627040207386, + 0.030887922272086143, + 0.02881304919719696, + 0.014872651547193527, + 0.03590201213955879, + 0.019332630559802055, + 0.016907505691051483, + 0.022257402539253235, + 0.018987424671649933, + 0.015224922448396683, + 0.02128881961107254, + 0.007070924621075392, + 0.009444747120141983, + 0.011234107427299023, + 0.012439143843948841, + 0.004828379023820162, + 0.008926364593207836, + 0.009927131235599518, + 0.011387411504983902, + 0.010703173466026783, + 0.015914881601929665, + 0.00480540469288826, + 0.009157991036772728, + 0.007449842989444733, + 0.007332573179155588, + 0.008646412752568722, + 0.00963171012699604, + 0.005733670201152563, + 0.013458643108606339, + 0.006312559824436903, + 0.013045093975961208, + 0.00806854572147131, + 0.0075405314564704895, + 0.008284018374979496, + 0.007409899495542049, + 0.011056347750127316, + 0.01607162319123745, + 0.011696342378854752, + 0.02724495157599449, + 0.008542567491531372, + 0.01823713630437851, + 0.022954490035772324, + 0.021832166239619255, + 0.005051181185990572, + 0.006850955542176962, + 0.00530729116871953, + 0.004533406812697649, + 0.005089849699288607, + 0.005866061896085739, + 0.006953159347176552, + 0.00989240501075983, + 0.013815143145620823, + 0.010366193950176239, + 0.017154572531580925, + 0.01301290187984705, + 0.006444836501032114, + 0.01058242842555046, + 0.015146778896450996, + 0.017241407185792923, + 0.007964886724948883, + 0.007018599659204483, + 0.010003466159105301, + 0.007819934748113155, + 0.005251213908195496, + 0.006640149746090174, + 0.008574755862355232, + 0.007542125415056944, + 0.004547020886093378, + 0.005771243944764137, + 0.003584789577871561, + 0.004611958283931017, + 0.003032864537090063, + 0.005048112478107214, + 0.0032462873496115208, + 0.006341021042317152 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 5 + ], + "selected_reward": -0.018189528957009315 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static, tight medium close-up of a woman?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the woman's skin powder-white with precisely painted red lips and dark kohl-lined eyes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the woman wearing a multi-layered heavy silk brocade kimono with crimson fabric and golden cranes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background showing an undersea earthquake causing massive deep blue waves with foaming whitecaps to erupt and crash?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the visual style hyper-real and crisp with individual water droplets appearing suspended in the air?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.038087088614702225 + ], + "opposite_logprob_tokens": [ + -3.2880871295928955 + ], + "logprob_reward": -0.038087088614702225, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.038087088614702225, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1280355155467987, + 0.19929882884025574, + 0.25170576572418213, + 0.08861581236124039, + 0.04046441242098808, + 0.04662434011697769, + 0.06349148601293564, + 0.080445297062397, + 0.07305540889501572, + 0.028263166546821594 + ], + "credit_map": [ + 0.017198609188199043, + 0.017265573143959045, + 0.020468510687351227, + 0.012208583764731884, + 0.01212487742304802, + 0.014202992431819439, + 0.012367909774184227, + 0.022198447957634926, + 0.019498512148857117, + 0.018216779455542564, + 0.05287141725420952, + 0.011791599914431572, + 0.012242433615028858, + 0.03225946053862572, + 0.016782686114311218, + 0.03563590720295906, + 0.029638290405273438, + 0.026979947462677956, + 0.050399333238601685, + 0.016210267320275307, + 0.01591881550848484, + 0.035661451518535614, + 0.023439034819602966, + 0.05345858260989189, + 0.01172166783362627, + 0.011703988537192345, + 0.009948216378688812, + 0.01598765328526497, + 0.007950566709041595, + 0.010084979236125946, + 0.009582401253283024, + 0.011636335402727127, + 0.004258072469383478, + 0.00563066778704524, + 0.004740539938211441, + 0.00553505914285779, + 0.0038530402816832066, + 0.004158227238804102, + 0.00493245292454958, + 0.007356348447501659, + 0.004376123659312725, + 0.005353296175599098, + 0.004174833185970783, + 0.006045551039278507, + 0.002609224058687687, + 0.0072139594703912735, + 0.0063196211121976376, + 0.010531727224588394, + 0.006887440569698811, + 0.007172508630901575, + 0.006845731753855944, + 0.00499923387542367, + 0.00532532250508666, + 0.008117670193314552, + 0.009507260285317898, + 0.014636312611401081, + 0.008677911944687366, + 0.010994683019816875, + 0.0070617045275866985, + 0.011105161160230637, + 0.007986695505678654, + 0.008228899911046028, + 0.012805146165192127, + 0.013585088774561882, + 0.007637774106115103, + 0.00822340976446867, + 0.010700643062591553, + 0.008431919850409031, + 0.0061766947619616985, + 0.010519994422793388, + 0.00902477465569973, + 0.012340191751718521, + 0.002695862902328372, + 0.004589596297591925, + 0.004444977268576622, + 0.0031823415774852037, + 0.0023670545779168606, + 0.003539967117831111, + 0.0023174709640443325, + 0.0051258946768939495 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 1, + 0 + ], + "selected_reward": -0.038087088614702225 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static, tight medium close-up of a woman?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the woman's skin powder-white with precisely painted red lips and dark kohl-lined eyes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the woman wearing a multi-layered heavy silk brocade kimono with crimson fabric and golden cranes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background showing an undersea earthquake causing massive deep blue waves with foaming whitecaps to erupt and crash?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the visual style hyper-real and crisp with individual water droplets appearing suspended in the air?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.12704038619995117 + ], + "opposite_logprob_tokens": [ + -2.127040386199951 + ], + "logprob_reward": -0.12704038619995117, + "yes_no_margin_reward": 2.0, + "token_credit_reward": -0.12704038619995117, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3258328437805176, + 0.1124829575419426, + 0.1215066909790039, + 0.05855072662234306, + 0.0891684964299202, + 0.04919266328215599, + 0.07211548089981079, + 0.06241464987397194, + 0.06340641528367996, + 0.04532907158136368 + ], + "credit_map": [ + 0.03816771134734154, + 0.0678938776254654, + 0.030596794560551643, + 0.03048471361398697, + 0.03976618871092796, + 0.035946253687143326, + 0.031220771372318268, + 0.05175652727484703, + 0.015308818779885769, + 0.021530453115701675, + 0.017168663442134857, + 0.012459776364266872, + 0.011197219602763653, + 0.012878639623522758, + 0.011156860738992691, + 0.01078252587467432, + 0.011628815904259682, + 0.015421280637383461, + 0.02177436277270317, + 0.01053751353174448, + 0.007831846363842487, + 0.01715230382978916, + 0.013489654287695885, + 0.02367091178894043, + 0.0070385742001235485, + 0.008846885524690151, + 0.005985682364553213, + 0.010279656387865543, + 0.0052112736739218235, + 0.006198103539645672, + 0.007068148348480463, + 0.00792239885777235, + 0.0097769470885396, + 0.00835430808365345, + 0.008419390767812729, + 0.007846861146390438, + 0.005736473947763443, + 0.009615366347134113, + 0.012350531294941902, + 0.027068618685007095, + 0.004686104133725166, + 0.005631879437714815, + 0.0063001615926623344, + 0.009680748917162418, + 0.0035268221981823444, + 0.005995381623506546, + 0.005943767260760069, + 0.0074277943931519985, + 0.008124874904751778, + 0.006698381155729294, + 0.007852248847484589, + 0.007143670693039894, + 0.005618789233267307, + 0.009628238156437874, + 0.007385945878922939, + 0.01966332644224167, + 0.006823716685175896, + 0.007538503035902977, + 0.006324379239231348, + 0.00977408792823553, + 0.004865337163209915, + 0.006601326167583466, + 0.008842138573527336, + 0.011645160615444183, + 0.005155431106686592, + 0.006867502816021442, + 0.005468812771141529, + 0.0076986136846244335, + 0.003824998391792178, + 0.013135924004018307, + 0.006659850478172302, + 0.014595289714634418, + 0.006042900029569864, + 0.00691116601228714, + 0.006496394518762827, + 0.005302337929606438, + 0.003077310975641012, + 0.006418564356863499, + 0.004695487208664417, + 0.006384911946952343 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 1 + ], + "selected_reward": -0.12704038619995117 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "static, tight medium close-up captures the unshakable poise of an elegant courtesan, Yanggiri Tayyoung, her form a still fortress against a backdrop of pure chaos. Her face, a mask of intense emphasis accented by perfect powder-white skin, precisely painted red lips, and dark, kohl-lined eyes, remains impassive as she gazes slightly past the camera. Her meticulously crafted black hair, sculpted into a traditional, intricate hairstyle secured by ornate tortoiseshell and coral kanzashi hairpins, captures the cold, diffuse light of a stormy sky. She wears a multi-layered, heavy silk brocade kimono, its crimson fabric woven with golden cranes. Behind her, an undersea earthquake causes the ocean to erupt in violent turbulence. Massive deep blue waves, capped by foaming whitecaps overhead, roll and crash, sending plumes of spray high into the air. The entire scene is rendered with the hyper-real, crisp detail of high-speed motion photography; individual water droplets appear suspended in the air, the texture of her silk sharp. Despite the catastrophic chaos unfolding just feet away, her calm demeanor is absolute; her posture remains straight and unwavering, creating a stark contrast that creates a powerful and tense atmosphere and highlights her extraordinary grace under pressure.", + "seed": 21896, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000071/group_9/index_2654_seed_21896_dp_19_rank_38_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 9, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.06151618473231792, + "vlm_logprob": -0.06151618473231792, + "vlm_yes_no_margin": 3.725, + "vlm_token_credit": -0.06151618473231792, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static, tight medium close-up of a woman?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the woman's skin powder-white with precisely painted red lips and dark kohl-lined eyes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the woman wearing a multi-layered heavy silk brocade kimono with crimson fabric and golden cranes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background showing an undersea earthquake causing massive deep blue waves with foaming whitecaps to erupt and crash?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the visual style hyper-real and crisp with individual water droplets appearing suspended in the air?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01083349622786045 + ], + "opposite_logprob_tokens": [ + -5.760833263397217 + ], + "logprob_reward": -0.01083349622786045, + "yes_no_margin_reward": 5.75, + "token_credit_reward": -0.01083349622786045, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07821869105100632, + 0.06324047595262527, + 0.1336059868335724, + 0.05090772733092308, + 0.0384536050260067, + 0.0260933805257082, + 0.031915705651044846, + 0.04201829433441162, + 0.021453315392136574, + 0.5140928626060486 + ], + "credit_map": [ + 0.012553134933114052, + 0.007519164122641087, + 0.009751319885253906, + 0.008241133764386177, + 0.009045729413628578, + 0.013172952458262444, + 0.006977093406021595, + 0.010958165861666203, + 0.009021628648042679, + 0.01135224662721157, + 0.006917163264006376, + 0.005992703605443239, + 0.006887899711728096, + 0.009933863766491413, + 0.005621548742055893, + 0.007513420656323433, + 0.016761111095547676, + 0.038664381951093674, + 0.024058517068624496, + 0.008716979064047337, + 0.009182675741612911, + 0.018253134563565254, + 0.009633947163820267, + 0.008335231803357601, + 0.0046778335236012936, + 0.00823703221976757, + 0.00942796841263771, + 0.006398330442607403, + 0.004447172395884991, + 0.005418272688984871, + 0.006718439981341362, + 0.005582674406468868, + 0.00658718217164278, + 0.008116002194583416, + 0.005246405024081469, + 0.003678264794871211, + 0.0023360098712146282, + 0.005964834243059158, + 0.0022313182707875967, + 0.004293591715395451, + 0.004383638501167297, + 0.003320147283375263, + 0.0027046706527471542, + 0.0020719696767628193, + 0.002716515911743045, + 0.004373004660010338, + 0.003747008042410016, + 0.002776425564661622, + 0.0047135427594184875, + 0.0024856417439877987, + 0.0038585574366152287, + 0.002937574405223131, + 0.003156159305945039, + 0.002934780204668641, + 0.00907809566706419, + 0.002751351799815893, + 0.009112831205129623, + 0.0031139703933149576, + 0.005530229769647121, + 0.00331385200843215, + 0.002760264789685607, + 0.011158507317304611, + 0.0033307326957583427, + 0.0036979052238166332, + 0.0030416271183639765, + 0.002672574482858181, + 0.0022066927049309015, + 0.002034430392086506, + 0.0035908117424696684, + 0.0030744385439902544, + 0.0029849447309970856, + 0.001847796025685966, + 0.07448123395442963, + 0.04132658988237381, + 0.1415177285671234, + 0.03724563866853714, + 0.03811704367399216, + 0.10069455206394196, + 0.03045240044593811, + 0.05025763064622879 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 2, + 0 + ], + "selected_reward": -0.01083349622786045 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static, tight medium close-up of a woman?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the woman's skin powder-white with precisely painted red lips and dark kohl-lined eyes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the woman wearing a multi-layered heavy silk brocade kimono with crimson fabric and golden cranes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background showing an undersea earthquake causing massive deep blue waves with foaming whitecaps to erupt and crash?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the visual style hyper-real and crisp with individual water droplets appearing suspended in the air?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.006759278476238251 + ], + "opposite_logprob_tokens": [ + -5.006759166717529 + ], + "logprob_reward": -0.006759278476238251, + "yes_no_margin_reward": 5.0, + "token_credit_reward": -0.006759278476238251, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08327353745698929, + 0.04823175072669983, + 0.06274992227554321, + 0.04499838873744011, + 0.05162234976887703, + 0.026702137663960457, + 0.061873696744441986, + 0.06695195287466049, + 0.03336677700281143, + 0.5202294588088989 + ], + "credit_map": [ + 0.013185232877731323, + 0.011545765213668346, + 0.011210223659873009, + 0.00962434895336628, + 0.011184091679751873, + 0.010831807740032673, + 0.007840253412723541, + 0.00785182137042284, + 0.007547916378825903, + 0.009417599067091942, + 0.0053454190492630005, + 0.007754873484373093, + 0.00470051309093833, + 0.006120855920016766, + 0.0036822669208049774, + 0.0036623107735067606, + 0.007714600767940283, + 0.01785905845463276, + 0.00955244991928339, + 0.005440978799015284, + 0.004547240678220987, + 0.007937625050544739, + 0.0055086263455450535, + 0.0041893539018929005, + 0.004182970616966486, + 0.007834404706954956, + 0.00874415971338749, + 0.006911810487508774, + 0.0032037447672337294, + 0.005455272272229195, + 0.005456088110804558, + 0.0032099455129355192, + 0.00761543121188879, + 0.006336939986795187, + 0.011286371387541294, + 0.004869254771620035, + 0.00396018335595727, + 0.007444197311997414, + 0.00506059592589736, + 0.00504938093945384, + 0.0044574313797056675, + 0.0031570487190037966, + 0.0030718035995960236, + 0.002964606275781989, + 0.00295213027857244, + 0.0038962268736213446, + 0.003367288038134575, + 0.0028356078546494246, + 0.009731273166835308, + 0.00552735198289156, + 0.012197158299386501, + 0.00458868034183979, + 0.00567612936720252, + 0.004023481160402298, + 0.015310623683035374, + 0.004819005262106657, + 0.013862097635865211, + 0.004027216695249081, + 0.012027270160615444, + 0.005812076386064291, + 0.004438907839357853, + 0.016789505258202553, + 0.005579769611358643, + 0.004415113013237715, + 0.006089411210268736, + 0.004353857133537531, + 0.0025797304697334766, + 0.003044974058866501, + 0.006509906146675348, + 0.0040754457004368305, + 0.003934280481189489, + 0.0027791736647486687, + 0.07591196894645691, + 0.03993688523769379, + 0.14093752205371857, + 0.03668218478560448, + 0.03719765692949295, + 0.10340992361307144, + 0.03016112744808197, + 0.055992260575294495 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 0, + 7 + ], + "selected_reward": -0.006759278476238251 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static, tight medium close-up of a woman?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the woman's skin powder-white with precisely painted red lips and dark kohl-lined eyes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the woman wearing a multi-layered heavy silk brocade kimono with crimson fabric and golden cranes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background showing an undersea earthquake causing massive deep blue waves with foaming whitecaps to erupt and crash?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the visual style hyper-real and crisp with individual water droplets appearing suspended in the air?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02635258063673973 + ], + "opposite_logprob_tokens": [ + -3.651352643966675 + ], + "logprob_reward": -0.02635258063673973, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.02635258063673973, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.04768182337284088, + 0.025001054629683495, + 0.047277480363845825, + 0.03642619028687477, + 0.037443857640028, + 0.05354528874158859, + 0.05150195211172104, + 0.235445037484169, + 0.024389052763581276, + 0.4412882328033447 + ], + "credit_map": [ + 0.006911165546625853, + 0.0054471539333462715, + 0.006131030153483152, + 0.005518394988030195, + 0.005507946480065584, + 0.007720451336354017, + 0.0047540850937366486, + 0.005691598169505596, + 0.004561276640743017, + 0.004250958561897278, + 0.0034060042817145586, + 0.001955978572368622, + 0.002335731638595462, + 0.003396793268620968, + 0.0025485625956207514, + 0.002545749070122838, + 0.005860425997525454, + 0.01196711789816618, + 0.009455057792365551, + 0.0045261383056640625, + 0.0038277360145002604, + 0.004778335336595774, + 0.003990980796515942, + 0.0028716896194964647, + 0.0035076846834272146, + 0.006313460413366556, + 0.00688179163262248, + 0.004300155211240053, + 0.0030132634565234184, + 0.004302681889384985, + 0.0048195840790867805, + 0.0032875691540539265, + 0.00490817753598094, + 0.005069466307759285, + 0.006904631853103638, + 0.0033674901351332664, + 0.002848633797839284, + 0.005970390513539314, + 0.004101634956896305, + 0.004273429978638887, + 0.005477642174810171, + 0.004822934977710247, + 0.005911877378821373, + 0.006376294419169426, + 0.006944482214748859, + 0.006938387174159288, + 0.008009511046111584, + 0.009064159356057644, + 0.00975786056369543, + 0.005556365009397268, + 0.010959292761981487, + 0.00376667152158916, + 0.004652753937989473, + 0.005819600075483322, + 0.006887285970151424, + 0.004102122038602829, + 0.05014249309897423, + 0.015372118912637234, + 0.018665390089154243, + 0.016896169632673264, + 0.017623882740736008, + 0.06022897735238075, + 0.028957946226000786, + 0.02755805291235447, + 0.003780561499297619, + 0.0021802191622555256, + 0.0018559368327260017, + 0.002120222197845578, + 0.0038516756612807512, + 0.003524956526234746, + 0.004927781876176596, + 0.0021476971451193094, + 0.06469988822937012, + 0.03141713887453079, + 0.11852976679801941, + 0.031518638134002686, + 0.029503710567951202, + 0.08826400339603424, + 0.029118264093995094, + 0.048236820846796036 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 7, + 5 + ], + "selected_reward": -0.02635258063673973 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static, tight medium close-up of a woman?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the woman's skin powder-white with precisely painted red lips and dark kohl-lined eyes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the woman wearing a multi-layered heavy silk brocade kimono with crimson fabric and golden cranes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background showing an undersea earthquake causing massive deep blue waves with foaming whitecaps to erupt and crash?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the visual style hyper-real and crisp with individual water droplets appearing suspended in the air?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.06204604357481003 + ], + "opposite_logprob_tokens": [ + -2.8120460510253906 + ], + "logprob_reward": -0.06204604357481003, + "yes_no_margin_reward": 2.75, + "token_credit_reward": -0.06204604357481003, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06763758510351181, + 0.07254341244697571, + 0.2825377881526947, + 0.023214280605316162, + 0.07636521756649017, + 0.03736696019768715, + 0.04649507626891136, + 0.25272154808044434, + 0.03925305977463722, + 0.10186504572629929 + ], + "credit_map": [ + 0.009899944067001343, + 0.006293048150837421, + 0.01164821907877922, + 0.007000999990850687, + 0.007938705384731293, + 0.011407127603888512, + 0.007277328055351973, + 0.006172222085297108, + 0.013409258797764778, + 0.015610069036483765, + 0.009659328497946262, + 0.005171753466129303, + 0.007718468550592661, + 0.010641436092555523, + 0.005120023153722286, + 0.005213075317442417, + 0.026663094758987427, + 0.0768054649233818, + 0.03639555349946022, + 0.017264634370803833, + 0.01197202131152153, + 0.08845841139554977, + 0.012012800201773643, + 0.012965835630893707, + 0.0025364449247717857, + 0.004251408856362104, + 0.005511831026524305, + 0.00270339148119092, + 0.0018091637175530195, + 0.002987266518175602, + 0.001972770318388939, + 0.0014420060906559229, + 0.014936834573745728, + 0.01561504602432251, + 0.012023638002574444, + 0.00721150916069746, + 0.004247751552611589, + 0.011352794244885445, + 0.005701109766960144, + 0.005276550073176622, + 0.00666024349629879, + 0.003979284316301346, + 0.004175903275609016, + 0.004433532245457172, + 0.0027510386426001787, + 0.007029641419649124, + 0.003863405901938677, + 0.004473913460969925, + 0.007661192212253809, + 0.004425395280122757, + 0.007041972130537033, + 0.003475075587630272, + 0.0034933211281895638, + 0.0066297622397542, + 0.009942379780113697, + 0.003825983265414834, + 0.07665351778268814, + 0.013085105456411839, + 0.03488372266292572, + 0.014246960170567036, + 0.01153637282550335, + 0.081165611743927, + 0.01289911475032568, + 0.008251174353063107, + 0.006616675294935703, + 0.004145551938563585, + 0.005934552755206823, + 0.0037801002617925406, + 0.00330948643386364, + 0.006700232159346342, + 0.004530969075858593, + 0.004235491622239351, + 0.014861528761684895, + 0.006316781975328922, + 0.027664121240377426, + 0.0069924634881317616, + 0.0057646483182907104, + 0.021274002268910408, + 0.008513141423463821, + 0.010478373616933823 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 7, + 9 + ], + "selected_reward": -0.06204604357481003 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static, tight medium close-up of a woman?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the woman's skin powder-white with precisely painted red lips and dark kohl-lined eyes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the woman wearing a multi-layered heavy silk brocade kimono with crimson fabric and golden cranes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background showing an undersea earthquake causing massive deep blue waves with foaming whitecaps to erupt and crash?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the visual style hyper-real and crisp with individual water droplets appearing suspended in the air?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.20158952474594116 + ], + "opposite_logprob_tokens": [ + -1.701589584350586 + ], + "logprob_reward": -0.20158952474594116, + "yes_no_margin_reward": 1.5, + "token_credit_reward": -0.20158952474594116, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11144236475229263, + 0.0688849687576294, + 0.09909568727016449, + 0.04275592789053917, + 0.03215758502483368, + 0.04472311586141586, + 0.021055061370134354, + 0.03215200453996658, + 0.0298062302172184, + 0.5179271697998047 + ], + "credit_map": [ + 0.016426702961325645, + 0.01140879187732935, + 0.018905948847532272, + 0.011859371326863766, + 0.010259078815579414, + 0.017353622242808342, + 0.011268098838627338, + 0.013960742391645908, + 0.012870285660028458, + 0.015134134329855442, + 0.007442767731845379, + 0.006231748964637518, + 0.006901106797158718, + 0.011643442325294018, + 0.0036216676235198975, + 0.005039808806031942, + 0.010570822283625603, + 0.030240541324019432, + 0.015584302134811878, + 0.0068705421872437, + 0.007224770728498697, + 0.015140706673264503, + 0.007711385376751423, + 0.005752600729465485, + 0.003650129307061434, + 0.007074104622006416, + 0.010856261476874352, + 0.00765996566042304, + 0.002901381114497781, + 0.0032590802293270826, + 0.003956294618546963, + 0.0033987059723585844, + 0.0037650594022125006, + 0.004174360539764166, + 0.006518273614346981, + 0.003104219911620021, + 0.002837131265550852, + 0.004219709429889917, + 0.002944777486845851, + 0.004594047088176012, + 0.0069831120781600475, + 0.006533534731715918, + 0.004815171472728252, + 0.00361813185736537, + 0.0034715263172984123, + 0.010554383508861065, + 0.005167887080460787, + 0.0035793629940599203, + 0.003181963926181197, + 0.0020641402807086706, + 0.0029661860316991806, + 0.0015525284688919783, + 0.002112317131832242, + 0.002798851113766432, + 0.004644771106541157, + 0.001734300283715129, + 0.005342403892427683, + 0.002726099221035838, + 0.006859935354441404, + 0.0028161064255982637, + 0.00378192076459527, + 0.005123272072523832, + 0.0026887175627052784, + 0.0028135464526712894, + 0.00599875720217824, + 0.002303100423887372, + 0.003920885268598795, + 0.002248159609735012, + 0.004040460102260113, + 0.0050940909422934055, + 0.0036985226906836033, + 0.0025022525805979967, + 0.07606027275323868, + 0.0381513275206089, + 0.14119504392147064, + 0.036134954541921616, + 0.03552663326263428, + 0.1043948084115982, + 0.029728511348366737, + 0.05673554167151451 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 0, + 2 + ], + "selected_reward": -0.20158952474594116 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000095/group_0/index_tensor([6549, 6549])_group_0_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000095/group_0/index_tensor([6549, 6549])_group_0_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..384e7655cd9f410e7b1af02838e01b14fbe4abcd --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000095/group_0/index_tensor([6549, 6549])_group_0_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "stunning close-up, shot with a slow, smooth push-in technique, begins from a slightly downward angle and focuses on a slender wrist gracefully raised. The skin of this wrist, fair and smooth, emerges effortlessly from the cuff of a rich, dark charcoal wool suit. With a slight twist of the wrist, a meticulously crafted stainless steel watch comes fully into view. A soft, directional, warm light illuminates the watch, reflecting a captivating luster from its interlaced polished and brushed metal surfaces. The watch boasts a classic round case and a simple bezel. Beneath the domed sapphire crystal lies a flawless, sun-brushed silver dial, its applied stick-shaped hour markers gleaming sharply in the light. When the wrist is steadily tilted towards the camera, as if at the moment of confirming the time, the focus falls precisely on the dial. A slender blued second hand is silently sweeping across the dial in an absolutely smooth and unpausing posture. Its continuous movement forms a sharp contrast with the stationary sword-shaped hour and minute hands. The whole picture conveys a luxurious atmosphere of concentration, tranquility and high-end texture in the ultimate details.", + "seed": 31222, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000095/group_0/index_6549_seed_31222_dp_0_rank_0_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 0, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.08842806462198496, + "vlm_logprob": -0.08842806462198496, + "vlm_yes_no_margin": 3.3, + "vlm_token_credit": -0.08842806462198496, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up of a slender wrist with fair, smooth skin emerging from a dark charcoal wool suit cuff?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a stainless steel watch with a classic round case, simple bezel, and domed sapphire crystal visible on the wrist?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the watch face feature a sun-brushed silver dial with applied stick-shaped hour markers and a slender blued second hand?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the blued second hand shown sweeping continuously while the sword-shaped hour and minute hands remain stationary?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera execute a slow, smooth push-in starting from a slightly downward angle?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07405067980289459 + ], + "opposite_logprob_tokens": [ + -2.6990506649017334 + ], + "logprob_reward": -0.07405067980289459, + "yes_no_margin_reward": 2.625, + "token_credit_reward": -0.07405067980289459, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1602022647857666, + 0.21398873627185822, + 0.05648769438266754, + 0.10824259370565414, + 0.20550018548965454, + 0.0628860667347908, + 0.04429054632782936, + 0.03791187331080437, + 0.05852934718132019, + 0.05196073651313782 + ], + "credit_map": [ + 0.025121964514255524, + 0.023161068558692932, + 0.021139085292816162, + 0.015880132094025612, + 0.024928752332925797, + 0.018081344664096832, + 0.019034763798117638, + 0.012855143286287785, + 0.016957247629761696, + 0.019461700692772865, + 0.04011208564043045, + 0.02815995365381241, + 0.021969927474856377, + 0.03044111654162407, + 0.021259795874357224, + 0.03562692180275917, + 0.007142729591578245, + 0.00540558947250247, + 0.008846694603562355, + 0.005217067897319794, + 0.005403604824095964, + 0.00919997040182352, + 0.009123111143708229, + 0.00614892365410924, + 0.014532857574522495, + 0.013045130297541618, + 0.017653875052928925, + 0.015172191895544529, + 0.010034927166998386, + 0.016213255003094673, + 0.01026158407330513, + 0.011328775435686111, + 0.02792472392320633, + 0.02055291086435318, + 0.02976584993302822, + 0.02276730351150036, + 0.01344588864594698, + 0.033335186541080475, + 0.027891775593161583, + 0.029816556721925735, + 0.006974911782890558, + 0.005327241960912943, + 0.008565951138734818, + 0.006136191543191671, + 0.004664980340749025, + 0.014885232783854008, + 0.006133888848125935, + 0.010197669267654419, + 0.0060428110882639885, + 0.0040486520156264305, + 0.007623808458447456, + 0.0033643280621618032, + 0.003501781029626727, + 0.008176766335964203, + 0.004935801029205322, + 0.006596599239856005, + 0.005437771789729595, + 0.0032825081143528223, + 0.0034260463435202837, + 0.004875770304352045, + 0.00340607319958508, + 0.00750399986281991, + 0.004146029707044363, + 0.005833671893924475, + 0.007746381219476461, + 0.003303105477243662, + 0.005484808702021837, + 0.00451665511354804, + 0.003000423079356551, + 0.014657758176326752, + 0.0041335709393024445, + 0.0156866442412138, + 0.004815252032130957, + 0.0027598997112363577, + 0.004771487321704626, + 0.004837612621486187, + 0.003201675834134221, + 0.013391292653977871, + 0.0047628325410187244, + 0.013420685194432735 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 4, + 0 + ], + "selected_reward": -0.07405067980289459 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up of a slender wrist with fair, smooth skin emerging from a dark charcoal wool suit cuff?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a stainless steel watch with a classic round case, simple bezel, and domed sapphire crystal visible on the wrist?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the watch face feature a sun-brushed silver dial with applied stick-shaped hour markers and a slender blued second hand?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the blued second hand shown sweeping continuously while the sword-shaped hour and minute hands remain stationary?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera execute a slow, smooth push-in starting from a slightly downward angle?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004126605577766895 + ], + "opposite_logprob_tokens": [ + -5.50412654876709 + ], + "logprob_reward": -0.004126605577766895, + "yes_no_margin_reward": 5.5, + "token_credit_reward": -0.004126605577766895, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2745872735977173, + 0.14445611834526062, + 0.07278958708047867, + 0.09483183175325394, + 0.09443756937980652, + 0.07360523194074631, + 0.09531646221876144, + 0.045916665345430374, + 0.04818199202418327, + 0.05587727576494217 + ], + "credit_map": [ + 0.060570184141397476, + 0.06046710163354874, + 0.03298088163137436, + 0.026531318202614784, + 0.02264791913330555, + 0.025975236669182777, + 0.027320651337504387, + 0.01809399574995041, + 0.012331108562648296, + 0.014727238565683365, + 0.018815945833921432, + 0.014310948550701141, + 0.015740031376481056, + 0.026652539148926735, + 0.02577298879623413, + 0.016105322167277336, + 0.00603499636054039, + 0.007041125558316708, + 0.006483578123152256, + 0.008758879266679287, + 0.013024638406932354, + 0.01153663732111454, + 0.01391549501568079, + 0.00599424634128809, + 0.013820506632328033, + 0.008358634077012539, + 0.0106895100325346, + 0.00860461127012968, + 0.008961810730397701, + 0.017847150564193726, + 0.010493829846382141, + 0.01605577953159809, + 0.010491509921848774, + 0.0074557955376803875, + 0.010147464461624622, + 0.007608167827129364, + 0.017744949087500572, + 0.016021739691495895, + 0.012505018152296543, + 0.01246291771531105, + 0.011256786063313484, + 0.007765579968690872, + 0.007819641381502151, + 0.0055772000923752785, + 0.008937066420912743, + 0.012727534398436546, + 0.00954708643257618, + 0.009974338114261627, + 0.010710977017879486, + 0.01307146530598402, + 0.010393613018095493, + 0.01614721678197384, + 0.006080294959247112, + 0.012017305009067059, + 0.012562110088765621, + 0.014333483763039112, + 0.007151555269956589, + 0.003982692025601864, + 0.0065824119374156, + 0.004003055859357119, + 0.004639391787350178, + 0.008191009983420372, + 0.004545905161648989, + 0.006820644252002239, + 0.004508839454501867, + 0.002989624859765172, + 0.003992882557213306, + 0.004699941724538803, + 0.0035424826201051474, + 0.012817524373531342, + 0.005530123133212328, + 0.010100572369992733, + 0.006629138719290495, + 0.00404971931129694, + 0.006655286997556686, + 0.005325132980942726, + 0.006064402870833874, + 0.009882715530693531, + 0.00827158335596323, + 0.0089992955327034 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 6 + ], + "selected_reward": -0.004126605577766895 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up of a slender wrist with fair, smooth skin emerging from a dark charcoal wool suit cuff?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a stainless steel watch with a classic round case, simple bezel, and domed sapphire crystal visible on the wrist?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the watch face feature a sun-brushed silver dial with applied stick-shaped hour markers and a slender blued second hand?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the blued second hand shown sweeping continuously while the sword-shaped hour and minute hands remain stationary?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera execute a slow, smooth push-in starting from a slightly downward angle?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.020618975162506104 + ], + "opposite_logprob_tokens": [ + -3.8956189155578613 + ], + "logprob_reward": -0.020618975162506104, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.020618975162506104, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2056228369474411, + 0.14027611911296844, + 0.11047377437353134, + 0.15551792085170746, + 0.09151997417211533, + 0.0490570031106472, + 0.08285398036241531, + 0.04331585764884949, + 0.07159115374088287, + 0.04977140948176384 + ], + "credit_map": [ + 0.029395077377557755, + 0.030466249212622643, + 0.030405405908823013, + 0.02213544398546219, + 0.02969384379684925, + 0.016649480909109116, + 0.02435211092233658, + 0.022525237873196602, + 0.013537158258259296, + 0.02192843332886696, + 0.014719747006893158, + 0.015886811539530754, + 0.017323384061455727, + 0.019865741953253746, + 0.022271055728197098, + 0.014743783511221409, + 0.008356638252735138, + 0.009680253453552723, + 0.012038193643093109, + 0.010068287141621113, + 0.014298245310783386, + 0.019856562837958336, + 0.026626165956258774, + 0.009549430571496487, + 0.01764822192490101, + 0.010004234500229359, + 0.011035993695259094, + 0.009335668757557869, + 0.009081603027880192, + 0.0464693084359169, + 0.014191366732120514, + 0.03775152564048767, + 0.017530269920825958, + 0.008942686952650547, + 0.011327742598950863, + 0.006967401597648859, + 0.006868078373372555, + 0.016346100717782974, + 0.010003290139138699, + 0.01353440061211586, + 0.00664478400722146, + 0.004444042686372995, + 0.006172080058604479, + 0.005007321946322918, + 0.006040666718035936, + 0.007694134954363108, + 0.006583272945135832, + 0.006470697000622749, + 0.01070884708315134, + 0.010525910183787346, + 0.011912481859326363, + 0.013365695253014565, + 0.005513852462172508, + 0.008324583992362022, + 0.011288031004369259, + 0.011214579455554485, + 0.0051617915742099285, + 0.003411412937566638, + 0.004947165492922068, + 0.0037703076377511024, + 0.004670276306569576, + 0.008551687933504581, + 0.005453367717564106, + 0.00734985014423728, + 0.00796682108193636, + 0.0039049971383064985, + 0.00624062167480588, + 0.005711409728974104, + 0.0035913554020226, + 0.02005022019147873, + 0.004722285084426403, + 0.01940344274044037, + 0.00450257258489728, + 0.0034050745889544487, + 0.006579900626093149, + 0.0038949744775891304, + 0.0036703781224787235, + 0.012531260959804058, + 0.0045234160497784615, + 0.010663831606507301 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 1 + ], + "selected_reward": -0.020618975162506104 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up of a slender wrist with fair, smooth skin emerging from a dark charcoal wool suit cuff?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a stainless steel watch with a classic round case, simple bezel, and domed sapphire crystal visible on the wrist?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the watch face feature a sun-brushed silver dial with applied stick-shaped hour markers and a slender blued second hand?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the blued second hand shown sweeping continuously while the sword-shaped hour and minute hands remain stationary?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera execute a slow, smooth push-in starting from a slightly downward angle?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.3135499060153961 + ], + "opposite_logprob_tokens": [ + -1.3135498762130737 + ], + "logprob_reward": -0.3135499060153961, + "yes_no_margin_reward": 1.0, + "token_credit_reward": -0.3135499060153961, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.28518062829971313, + 0.14923593401908875, + 0.054914988577365875, + 0.04325989633798599, + 0.11339135468006134, + 0.07661564648151398, + 0.14264655113220215, + 0.029348211362957954, + 0.04980864375829697, + 0.0555981770157814 + ], + "credit_map": [ + 0.050230517983436584, + 0.04902727156877518, + 0.03223857656121254, + 0.02776944451034069, + 0.0473979227244854, + 0.02791510336101055, + 0.027018750086426735, + 0.023583056405186653, + 0.01365178544074297, + 0.013637964613735676, + 0.028832649812102318, + 0.017938457429409027, + 0.021330062299966812, + 0.01920960657298565, + 0.01610969379544258, + 0.018525712192058563, + 0.004549178760498762, + 0.00607794476673007, + 0.006769472733139992, + 0.005792680196464062, + 0.006218072958290577, + 0.00879424624145031, + 0.009517400525510311, + 0.007195993792265654, + 0.009061796590685844, + 0.004990806337445974, + 0.0053861928172409534, + 0.005430692806839943, + 0.004843026865273714, + 0.005143639165908098, + 0.004628004040569067, + 0.003775740507990122, + 0.013343129307031631, + 0.009632705710828304, + 0.008940638974308968, + 0.008509132079780102, + 0.010464807040989399, + 0.026599038392305374, + 0.014532548375427723, + 0.02136934921145439, + 0.014944036491215229, + 0.008586863055825233, + 0.007618866395205259, + 0.006366352085024118, + 0.008457141928374767, + 0.012859580107033253, + 0.006925284396857023, + 0.010857527144253254, + 0.020821183919906616, + 0.011403358541429043, + 0.02287439815700054, + 0.010766643099486828, + 0.01142291259020567, + 0.02849043533205986, + 0.01425392460078001, + 0.022613700479269028, + 0.004224306438118219, + 0.0031098995823413134, + 0.004836641252040863, + 0.0026559862308204174, + 0.0026100315153598785, + 0.004852857440710068, + 0.0027227848768234253, + 0.004335705656558275, + 0.006053649354726076, + 0.0034654836636036634, + 0.005680757574737072, + 0.004419872537255287, + 0.005162075161933899, + 0.01130692195147276, + 0.0038873322773724794, + 0.00983255635946989, + 0.006112613715231419, + 0.003151964396238327, + 0.005827721673995256, + 0.006201683543622494, + 0.003777439473196864, + 0.012852862477302551, + 0.005715970415621996, + 0.01195792481303215 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 6 + ], + "selected_reward": -0.3135499060153961 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up of a slender wrist with fair, smooth skin emerging from a dark charcoal wool suit cuff?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a stainless steel watch with a classic round case, simple bezel, and domed sapphire crystal visible on the wrist?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the watch face feature a sun-brushed silver dial with applied stick-shaped hour markers and a slender blued second hand?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the blued second hand shown sweeping continuously while the sword-shaped hour and minute hands remain stationary?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera execute a slow, smooth push-in starting from a slightly downward angle?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.029794156551361084 + ], + "opposite_logprob_tokens": [ + -3.529794216156006 + ], + "logprob_reward": -0.029794156551361084, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.029794156551361084, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.16994671523571014, + 0.27881893515586853, + 0.03463948890566826, + 0.09934331476688385, + 0.13223013281822205, + 0.07307291030883789, + 0.06890317797660828, + 0.02626211754977703, + 0.07851903140544891, + 0.03826413303613663 + ], + "credit_map": [ + 0.02304263785481453, + 0.02296152338385582, + 0.034637030214071274, + 0.017256595194339752, + 0.022856343537569046, + 0.014391444623470306, + 0.020240947604179382, + 0.014560192823410034, + 0.024435579776763916, + 0.028711579740047455, + 0.06642702966928482, + 0.0417008213698864, + 0.016297530382871628, + 0.03920311480760574, + 0.026571188122034073, + 0.03547211363911629, + 0.004730251617729664, + 0.003113503335043788, + 0.004272296093404293, + 0.0047371587716042995, + 0.004362906329333782, + 0.0049414499662816525, + 0.005237792618572712, + 0.0032441297080367804, + 0.006994258612394333, + 0.005031777545809746, + 0.006505706813186407, + 0.006967044435441494, + 0.007433484774082899, + 0.031035766005516052, + 0.008289285935461521, + 0.027085985988378525, + 0.017251718789339066, + 0.011146510019898415, + 0.014998642727732658, + 0.009793594479560852, + 0.010554526932537556, + 0.029589133337140083, + 0.01377477403730154, + 0.025121234357357025, + 0.010437041521072388, + 0.007378480397164822, + 0.008494879119098186, + 0.004754982423037291, + 0.007511023432016373, + 0.016783859580755234, + 0.006709105335175991, + 0.011003537103533745, + 0.01099228486418724, + 0.0062383320182561874, + 0.009629378095269203, + 0.008520316332578659, + 0.005769289564341307, + 0.00986683089286089, + 0.008049989119172096, + 0.009836754761636257, + 0.0047394814901053905, + 0.002453177236020565, + 0.004579756408929825, + 0.0029760091565549374, + 0.0015674879541620612, + 0.004790975712239742, + 0.001691380050033331, + 0.003463848028331995, + 0.011097106151282787, + 0.005953220184892416, + 0.0061162253841757774, + 0.004504916723817587, + 0.0035262703895568848, + 0.02221357263624668, + 0.0031424181070178747, + 0.021965300664305687, + 0.002325630048289895, + 0.002134984824806452, + 0.0033600821625441313, + 0.003941590432077646, + 0.0025210161693394184, + 0.01142894197255373, + 0.002848696894943714, + 0.009703192859888077 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 4 + ], + "selected_reward": -0.029794156551361084 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "stunning close-up, shot with a slow, smooth push-in technique, begins from a slightly downward angle and focuses on a slender wrist gracefully raised. The skin of this wrist, fair and smooth, emerges effortlessly from the cuff of a rich, dark charcoal wool suit. With a slight twist of the wrist, a meticulously crafted stainless steel watch comes fully into view. A soft, directional, warm light illuminates the watch, reflecting a captivating luster from its interlaced polished and brushed metal surfaces. The watch boasts a classic round case and a simple bezel. Beneath the domed sapphire crystal lies a flawless, sun-brushed silver dial, its applied stick-shaped hour markers gleaming sharply in the light. When the wrist is steadily tilted towards the camera, as if at the moment of confirming the time, the focus falls precisely on the dial. A slender blued second hand is silently sweeping across the dial in an absolutely smooth and unpausing posture. Its continuous movement forms a sharp contrast with the stationary sword-shaped hour and minute hands. The whole picture conveys a luxurious atmosphere of concentration, tranquility and high-end texture in the ultimate details.", + "seed": 31223, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000095/group_0/index_6549_seed_31223_dp_0_rank_0_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 0, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.09195772064849735, + "vlm_logprob": -0.09195772064849735, + "vlm_yes_no_margin": 3.6, + "vlm_token_credit": -0.09195772064849735, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up of a slender wrist with fair, smooth skin emerging from a dark charcoal wool suit cuff?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a stainless steel watch with a classic round case, simple bezel, and domed sapphire crystal visible on the wrist?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the watch face feature a sun-brushed silver dial with applied stick-shaped hour markers and a slender blued second hand?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the blued second hand shown sweeping continuously while the sword-shaped hour and minute hands remain stationary?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera execute a slow, smooth push-in starting from a slightly downward angle?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.036549344658851624 + ], + "opposite_logprob_tokens": [ + -3.4115493297576904 + ], + "logprob_reward": -0.036549344658851624, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.036549344658851624, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06874526292085648, + 0.04769463464617729, + 0.08945735543966293, + 0.08497774600982666, + 0.0654328241944313, + 0.31387874484062195, + 0.11142884939908981, + 0.1715424805879593, + 0.025788625702261925, + 0.021053455770015717 + ], + "credit_map": [ + 0.005421819165349007, + 0.018173476681113243, + 0.010407283902168274, + 0.005966460332274437, + 0.006082417909055948, + 0.00411919504404068, + 0.009511315263807774, + 0.009063301607966423, + 0.0034786744508892298, + 0.005232453811913729, + 0.005593129899352789, + 0.0073658437468111515, + 0.004828281234949827, + 0.00684340437874198, + 0.007363660726696253, + 0.0069891903549432755, + 0.006937528029084206, + 0.013298651203513145, + 0.007764444686472416, + 0.006500293966382742, + 0.01439882256090641, + 0.017334608361124992, + 0.012510829605162144, + 0.010712177492678165, + 0.010851782746613026, + 0.010946718975901604, + 0.008611051365733147, + 0.007185969036072493, + 0.01427318062633276, + 0.013140206225216389, + 0.009335250593721867, + 0.010633590631186962, + 0.0043529304675757885, + 0.0031542207580059767, + 0.0026597471442073584, + 0.0027507729828357697, + 0.006222153082489967, + 0.02118406631052494, + 0.0032085415441542864, + 0.02190040424466133, + 0.01101622823625803, + 0.005756754893809557, + 0.0062452577985823154, + 0.009077232331037521, + 0.009470080956816673, + 0.13302350044250488, + 0.010657910257577896, + 0.12863177061080933, + 0.003193029435351491, + 0.004117025062441826, + 0.0036768175195902586, + 0.004075666423887014, + 0.003094503190368414, + 0.04424305260181427, + 0.006246110424399376, + 0.042782656848430634, + 0.005653274245560169, + 0.011411136947572231, + 0.007669425569474697, + 0.011093148030340672, + 0.008539116010069847, + 0.057791903614997864, + 0.013237476348876953, + 0.0561470165848732, + 0.002146719954907894, + 0.0023225361946970224, + 0.002477681962773204, + 0.003130963770672679, + 0.0020016615744680166, + 0.004954921081662178, + 0.0030518160201609135, + 0.005702327471226454, + 0.0019697302486747503, + 0.001313679269514978, + 0.002208142075687647, + 0.0016905877273529768, + 0.0018245009705424309, + 0.004530538804829121, + 0.00227270252071321, + 0.005243575666099787 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 7, + 6 + ], + "selected_reward": -0.036549344658851624 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up of a slender wrist with fair, smooth skin emerging from a dark charcoal wool suit cuff?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a stainless steel watch with a classic round case, simple bezel, and domed sapphire crystal visible on the wrist?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the watch face feature a sun-brushed silver dial with applied stick-shaped hour markers and a slender blued second hand?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the blued second hand shown sweeping continuously while the sword-shaped hour and minute hands remain stationary?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera execute a slow, smooth push-in starting from a slightly downward angle?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0036482936702668667 + ], + "opposite_logprob_tokens": [ + -5.628648281097412 + ], + "logprob_reward": -0.0036482936702668667, + "yes_no_margin_reward": 5.625, + "token_credit_reward": -0.0036482936702668667, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08509528636932373, + 0.045816969126462936, + 0.04071478545665741, + 0.12871144711971283, + 0.10521586984395981, + 0.23891571164131165, + 0.16978733241558075, + 0.07910936325788498, + 0.051701948046684265, + 0.05493130534887314 + ], + "credit_map": [ + 0.011400021612644196, + 0.013618778437376022, + 0.00999473500996828, + 0.0076208640821278095, + 0.010093554854393005, + 0.007318935822695494, + 0.010317334905266762, + 0.014731061644852161, + 0.004530264064669609, + 0.005908853840082884, + 0.006188228260725737, + 0.006757967174053192, + 0.004210384096950293, + 0.00656858691945672, + 0.005628375802189112, + 0.006024307571351528, + 0.003223912324756384, + 0.008280129171907902, + 0.004317262675613165, + 0.0021031564101576805, + 0.006995484698563814, + 0.006737308111041784, + 0.00372799439355731, + 0.005329534877091646, + 0.015241214074194431, + 0.018057726323604584, + 0.012646756134927273, + 0.014956815168261528, + 0.020266903564333916, + 0.0171918123960495, + 0.01527032908052206, + 0.01507988478988409, + 0.00445643300190568, + 0.0029228203929960728, + 0.0034102543722838163, + 0.004015602637082338, + 0.00572271877899766, + 0.04062473773956299, + 0.004455172922462225, + 0.03960813209414482, + 0.009371187537908554, + 0.008882815949618816, + 0.007198241073638201, + 0.01433818694204092, + 0.014563915319740772, + 0.08458883315324783, + 0.0191507525742054, + 0.08082177489995956, + 0.004807166289538145, + 0.005156238097697496, + 0.005730331409722567, + 0.005696948152035475, + 0.004927916452288628, + 0.06808379292488098, + 0.010111309587955475, + 0.06527362763881683, + 0.003775248536840081, + 0.008610355667769909, + 0.004490102641284466, + 0.007508200593292713, + 0.004992322996258736, + 0.022265374660491943, + 0.006657132878899574, + 0.020810618996620178, + 0.0038281658198684454, + 0.0036151744425296783, + 0.005091633182018995, + 0.006384498905390501, + 0.0032418493647128344, + 0.011375359259545803, + 0.006548293866217136, + 0.011616974137723446, + 0.005174784455448389, + 0.0055340793915092945, + 0.0058881063014268875, + 0.0063035255298018456, + 0.004957685247063637, + 0.010779260657727718, + 0.005796114914119244, + 0.010497747920453548 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 6, + 3 + ], + "selected_reward": -0.0036482936702668667 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up of a slender wrist with fair, smooth skin emerging from a dark charcoal wool suit cuff?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a stainless steel watch with a classic round case, simple bezel, and domed sapphire crystal visible on the wrist?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the watch face feature a sun-brushed silver dial with applied stick-shaped hour markers and a slender blued second hand?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the blued second hand shown sweeping continuously while the sword-shaped hour and minute hands remain stationary?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera execute a slow, smooth push-in starting from a slightly downward angle?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.014221390709280968 + ], + "opposite_logprob_tokens": [ + -4.26422119140625 + ], + "logprob_reward": -0.014221390709280968, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.014221390709280968, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09794528782367706, + 0.0441511794924736, + 0.06523211300373077, + 0.05710036680102348, + 0.05337919667363167, + 0.5089566707611084, + 0.04010970890522003, + 0.07475034892559052, + 0.041735514998435974, + 0.016639715060591698 + ], + "credit_map": [ + 0.00983390398323536, + 0.018157225102186203, + 0.012584604322910309, + 0.011070557869970798, + 0.01309269294142723, + 0.009908897802233696, + 0.01119375228881836, + 0.012103642337024212, + 0.004386906046420336, + 0.00582393491640687, + 0.0056633842177689075, + 0.005928766913712025, + 0.0037692165933549404, + 0.007182493340224028, + 0.006209272891283035, + 0.005187198985368013, + 0.0054822261445224285, + 0.011928290128707886, + 0.0069998023100197315, + 0.004607424605637789, + 0.007615508511662483, + 0.012245980091392994, + 0.007488440256565809, + 0.008864433504641056, + 0.006479976233094931, + 0.009096110239624977, + 0.004369926173239946, + 0.006744962185621262, + 0.009538830257952213, + 0.007964657619595528, + 0.005792493931949139, + 0.00711340457201004, + 0.003237334778532386, + 0.0023270556703209877, + 0.00240559340454638, + 0.0032382227946072817, + 0.005102810449898243, + 0.01752997189760208, + 0.002921774983406067, + 0.016616424545645714, + 0.013462760485708714, + 0.00763819320127368, + 0.0094008669257164, + 0.01224984135478735, + 0.013607962056994438, + 0.2233508825302124, + 0.012922918424010277, + 0.21632321178913116, + 0.0024232962168753147, + 0.0023716336581856012, + 0.0027347267605364323, + 0.002871171338483691, + 0.0016144652618095279, + 0.012470442801713943, + 0.0036775690969079733, + 0.011946398764848709, + 0.0026281403843313456, + 0.00466103246435523, + 0.003670061007142067, + 0.004644234199076891, + 0.003950357902795076, + 0.025518620386719704, + 0.005236483179032803, + 0.024441421031951904, + 0.004610743839293718, + 0.002470993669703603, + 0.004548460710793734, + 0.00413886085152626, + 0.003040575422346592, + 0.008252196945250034, + 0.004954586271196604, + 0.009719091467559338, + 0.0011691927211359143, + 0.001980172237381339, + 0.0017309945542365313, + 0.002492019673809409, + 0.0014548732433468103, + 0.0024215178564190865, + 0.002275839913636446, + 0.003115102183073759 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 0, + 7 + ], + "selected_reward": -0.014221390709280968 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up of a slender wrist with fair, smooth skin emerging from a dark charcoal wool suit cuff?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a stainless steel watch with a classic round case, simple bezel, and domed sapphire crystal visible on the wrist?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the watch face feature a sun-brushed silver dial with applied stick-shaped hour markers and a slender blued second hand?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the blued second hand shown sweeping continuously while the sword-shaped hour and minute hands remain stationary?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera execute a slow, smooth push-in starting from a slightly downward angle?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.38718414306640625 + ], + "opposite_logprob_tokens": [ + -1.1371841430664062 + ], + "logprob_reward": -0.38718414306640625, + "yes_no_margin_reward": 0.75, + "token_credit_reward": -0.38718414306640625, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11277353763580322, + 0.0886017307639122, + 0.0582960769534111, + 0.1687796413898468, + 0.06304647773504257, + 0.19378085434436798, + 0.10206593573093414, + 0.1076887771487236, + 0.04395579546689987, + 0.061011139303445816 + ], + "credit_map": [ + 0.008454855531454086, + 0.013662063516676426, + 0.015910902991890907, + 0.015828195959329605, + 0.012171821668744087, + 0.011816715821623802, + 0.023081419989466667, + 0.011847559362649918, + 0.009740845300257206, + 0.006752984132617712, + 0.00913646537810564, + 0.01424827054142952, + 0.00801358837634325, + 0.011229478754103184, + 0.016726933419704437, + 0.012753157876431942, + 0.005482138600200415, + 0.008735299110412598, + 0.006118342746049166, + 0.005419469438493252, + 0.0076593877747654915, + 0.010069184005260468, + 0.006048433482646942, + 0.008763820864260197, + 0.021296173334121704, + 0.020990755409002304, + 0.01849500834941864, + 0.014364784583449364, + 0.029908282682299614, + 0.023672081530094147, + 0.01980586349964142, + 0.02024669758975506, + 0.009192307479679585, + 0.005388302728533745, + 0.004760715179145336, + 0.006278909742832184, + 0.009640980511903763, + 0.011538909748196602, + 0.007095265667885542, + 0.00915108434855938, + 0.008119923993945122, + 0.008062439039349556, + 0.005796995013952255, + 0.013044099323451519, + 0.011377967894077301, + 0.06433450430631638, + 0.01923859864473343, + 0.06380633264780045, + 0.005306279752403498, + 0.005690888036042452, + 0.005846700631082058, + 0.006429427769035101, + 0.0040879203006625175, + 0.0336163304746151, + 0.009986823424696922, + 0.031101569533348083, + 0.0039687687531113625, + 0.006896934472024441, + 0.005895301699638367, + 0.005239688325673342, + 0.004413214046508074, + 0.03662955388426781, + 0.008135038428008556, + 0.036510270088911057, + 0.0032900944352149963, + 0.0032932774629443884, + 0.0053681666031479836, + 0.005699994042515755, + 0.00322313467040658, + 0.008643589913845062, + 0.0065437001176178455, + 0.007893839851021767, + 0.004679103847593069, + 0.006132941693067551, + 0.0069171590730547905, + 0.0074011352844536304, + 0.0061278799548745155, + 0.009754734113812447, + 0.007935242727398872, + 0.01206294260919094 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 3, + 0 + ], + "selected_reward": -0.38718414306640625 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up of a slender wrist with fair, smooth skin emerging from a dark charcoal wool suit cuff?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a stainless steel watch with a classic round case, simple bezel, and domed sapphire crystal visible on the wrist?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the watch face feature a sun-brushed silver dial with applied stick-shaped hour markers and a slender blued second hand?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the blued second hand shown sweeping continuously while the sword-shaped hour and minute hands remain stationary?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera execute a slow, smooth push-in starting from a slightly downward angle?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.018185431137681007 + ], + "opposite_logprob_tokens": [ + -4.018185615539551 + ], + "logprob_reward": -0.018185431137681007, + "yes_no_margin_reward": 4.0, + "token_credit_reward": -0.018185431137681007, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11665374040603638, + 0.07530946284532547, + 0.09747949242591858, + 0.1773395985364914, + 0.2688906788825989, + 0.10355915874242783, + 0.07050175964832306, + 0.05510587617754936, + 0.021358322352170944, + 0.013801930472254753 + ], + "credit_map": [ + 0.014677818864583969, + 0.013758111745119095, + 0.018405379727482796, + 0.011817903257906437, + 0.011170136742293835, + 0.01067788153886795, + 0.01565932109951973, + 0.020487187430262566, + 0.006083723157644272, + 0.0074548013508319855, + 0.0062877400778234005, + 0.007874321192502975, + 0.008568822406232357, + 0.01161686610430479, + 0.017624033614993095, + 0.009799151681363583, + 0.007811887189745903, + 0.01714160479605198, + 0.010529388673603535, + 0.00728488527238369, + 0.011319883167743683, + 0.018747396767139435, + 0.011036631651222706, + 0.013607818633317947, + 0.019813060760498047, + 0.022625859826803207, + 0.014086147770285606, + 0.02580174244940281, + 0.02975526824593544, + 0.021340524777770042, + 0.022527147084474564, + 0.02138984575867653, + 0.00989893265068531, + 0.005456628743559122, + 0.007576110307127237, + 0.00809139758348465, + 0.015457178466022015, + 0.10710900276899338, + 0.007671406026929617, + 0.10763002932071686, + 0.014443010091781616, + 0.0070866248570382595, + 0.008117134682834148, + 0.006723266094923019, + 0.009789030067622662, + 0.022631172090768814, + 0.012484257109463215, + 0.02228466607630253, + 0.0057042124681174755, + 0.0074356417171657085, + 0.00728345662355423, + 0.007959246635437012, + 0.004850815981626511, + 0.01244710385799408, + 0.011188961565494537, + 0.013632319867610931, + 0.003779249731451273, + 0.0049585262313485146, + 0.004301208071410656, + 0.005783676635473967, + 0.004230240825563669, + 0.013683863915503025, + 0.007666244171559811, + 0.010702869854867458, + 0.0016383855836465955, + 0.001359545742161572, + 0.0019479667535051703, + 0.0022315497044473886, + 0.0015735538909211755, + 0.004850325174629688, + 0.0024116551503539085, + 0.005345340818166733, + 0.001131308265030384, + 0.0015647707041352987, + 0.001522050122730434, + 0.0014123090077191591, + 0.0009576122392900288, + 0.002573680365458131, + 0.001924575655721128, + 0.0027156234718859196 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 3, + 0 + ], + "selected_reward": -0.018185431137681007 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "stunning close-up, shot with a slow, smooth push-in technique, begins from a slightly downward angle and focuses on a slender wrist gracefully raised. The skin of this wrist, fair and smooth, emerges effortlessly from the cuff of a rich, dark charcoal wool suit. With a slight twist of the wrist, a meticulously crafted stainless steel watch comes fully into view. A soft, directional, warm light illuminates the watch, reflecting a captivating luster from its interlaced polished and brushed metal surfaces. The watch boasts a classic round case and a simple bezel. Beneath the domed sapphire crystal lies a flawless, sun-brushed silver dial, its applied stick-shaped hour markers gleaming sharply in the light. When the wrist is steadily tilted towards the camera, as if at the moment of confirming the time, the focus falls precisely on the dial. A slender blued second hand is silently sweeping across the dial in an absolutely smooth and unpausing posture. Its continuous movement forms a sharp contrast with the stationary sword-shaped hour and minute hands. The whole picture conveys a luxurious atmosphere of concentration, tranquility and high-end texture in the ultimate details.", + "seed": 31224, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000095/group_0/index_6549_seed_31224_dp_1_rank_2_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 0, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.08586827367544174, + "vlm_logprob": -0.08586827367544174, + "vlm_yes_no_margin": 3.6999999403953554, + "vlm_token_credit": -0.08586827367544174, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up of a slender wrist with fair, smooth skin emerging from a dark charcoal wool suit cuff?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a stainless steel watch with a classic round case, simple bezel, and domed sapphire crystal visible on the wrist?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the watch face feature a sun-brushed silver dial with applied stick-shaped hour markers and a slender blued second hand?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the blued second hand shown sweeping continuously while the sword-shaped hour and minute hands remain stationary?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera execute a slow, smooth push-in starting from a slightly downward angle?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04627664014697075 + ], + "opposite_logprob_tokens": [ + -3.171276569366455 + ], + "logprob_reward": -0.04627664014697075, + "yes_no_margin_reward": 3.125, + "token_credit_reward": -0.04627664014697075, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12949688732624054, + 0.09278355538845062, + 0.04609975591301918, + 0.027123700827360153, + 0.02217959426343441, + 0.284387469291687, + 0.04849296808242798, + 0.10188675671815872, + 0.043673448264598846, + 0.2038758248090744 + ], + "credit_map": [ + 0.023899929597973824, + 0.021613650023937225, + 0.019125107675790787, + 0.014219366014003754, + 0.011002046056091785, + 0.013823123648762703, + 0.010922297835350037, + 0.014891360886394978, + 0.010402150452136993, + 0.017808822914958, + 0.014665243215858936, + 0.014175769872963428, + 0.0061323014087975025, + 0.008747335523366928, + 0.011162426322698593, + 0.009689509868621826, + 0.007090124301612377, + 0.005316848400980234, + 0.006516225636005402, + 0.006927694194018841, + 0.004171719308942556, + 0.005380452144891024, + 0.006950938142836094, + 0.003745754947885871, + 0.004542837850749493, + 0.004397846758365631, + 0.004163297358900309, + 0.003210274502635002, + 0.002110316650941968, + 0.0023078310769051313, + 0.003899065311998129, + 0.002492228988558054, + 0.0029713003896176815, + 0.0030728878919035196, + 0.0034247825387865305, + 0.0038389740511775017, + 0.002139183459803462, + 0.0019358801655471325, + 0.0027195801958441734, + 0.0020770044066011906, + 0.07766304165124893, + 0.008285308256745338, + 0.11274679005146027, + 0.012079215608537197, + 0.00859280675649643, + 0.016633879393339157, + 0.005784146953374147, + 0.0426022931933403, + 0.00880538672208786, + 0.004704199265688658, + 0.006917961407452822, + 0.0029380631167441607, + 0.005526362452656031, + 0.005122465547174215, + 0.0037832017987966537, + 0.010695328935980797, + 0.008531668223440647, + 0.01188039593398571, + 0.02144477516412735, + 0.010070650838315487, + 0.008929988369345665, + 0.01154654286801815, + 0.014092992059886456, + 0.015389736741781235, + 0.004506315570324659, + 0.0038536989595741034, + 0.008075451478362083, + 0.00336838117800653, + 0.004074946045875549, + 0.008483026176691055, + 0.004000731278210878, + 0.007310896646231413, + 0.02064703032374382, + 0.012319897301495075, + 0.03549054265022278, + 0.014911766164004803, + 0.02037086710333824, + 0.0488661453127861, + 0.017406605184078217, + 0.033862967044115067 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 9, + 0 + ], + "selected_reward": -0.04627664014697075 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up of a slender wrist with fair, smooth skin emerging from a dark charcoal wool suit cuff?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a stainless steel watch with a classic round case, simple bezel, and domed sapphire crystal visible on the wrist?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the watch face feature a sun-brushed silver dial with applied stick-shaped hour markers and a slender blued second hand?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the blued second hand shown sweeping continuously while the sword-shaped hour and minute hands remain stationary?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera execute a slow, smooth push-in starting from a slightly downward angle?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0019685905426740646 + ], + "opposite_logprob_tokens": [ + -6.2519683837890625 + ], + "logprob_reward": -0.0019685905426740646, + "yes_no_margin_reward": 6.25, + "token_credit_reward": -0.0019685905426740646, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14502644538879395, + 0.027936583384871483, + 0.03375713899731636, + 0.027468685060739517, + 0.028494270518422127, + 0.22830958664417267, + 0.05934229865670204, + 0.33101004362106323, + 0.05102893337607384, + 0.06762604415416718 + ], + "credit_map": [ + 0.01365056075155735, + 0.02301890216767788, + 0.021721329540014267, + 0.028911137953400612, + 0.01252018939703703, + 0.014821507036685944, + 0.011250991374254227, + 0.019131824374198914, + 0.003961907234042883, + 0.0038666832260787487, + 0.004540898837149143, + 0.004489168059080839, + 0.0024648020043969154, + 0.0028251258190721273, + 0.0032732440158724785, + 0.0025147534906864166, + 0.0044426871463656425, + 0.0028260929975658655, + 0.004703633952885866, + 0.004347659647464752, + 0.004403129685670137, + 0.0044541009701788425, + 0.0053587425500154495, + 0.003221089718863368, + 0.003286269959062338, + 0.002630398143082857, + 0.004008711781352758, + 0.00465841731056571, + 0.002360736019909382, + 0.0031494544818997383, + 0.005183032713830471, + 0.002191664185374975, + 0.0034262817353010178, + 0.004311021883040667, + 0.004280126187950373, + 0.004341759718954563, + 0.003633543150499463, + 0.0023531841579824686, + 0.0034056506119668484, + 0.002742702141404152, + 0.06384766846895218, + 0.006740324664860964, + 0.09055744856595993, + 0.01037280261516571, + 0.006684713065624237, + 0.01378179807215929, + 0.005396666470915079, + 0.030928155407309532, + 0.009950542822480202, + 0.005159432999789715, + 0.006769781466573477, + 0.004963758401572704, + 0.007226092275232077, + 0.006338461767882109, + 0.005712076090276241, + 0.01322215050458908, + 0.0314723439514637, + 0.03669138625264168, + 0.05672052875161171, + 0.030613726004958153, + 0.031213773414492607, + 0.05462111532688141, + 0.04073013365268707, + 0.048947036266326904, + 0.004818630404770374, + 0.004623164888471365, + 0.010433236137032509, + 0.004259130451828241, + 0.00461594108492136, + 0.00983425509184599, + 0.003959805704653263, + 0.00848476868122816, + 0.006953785195946693, + 0.005353488959372044, + 0.011640367098152637, + 0.004693660885095596, + 0.006541000213474035, + 0.015138630755245686, + 0.0070790695026516914, + 0.01022604014724493 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 5, + 0 + ], + "selected_reward": -0.0019685905426740646 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up of a slender wrist with fair, smooth skin emerging from a dark charcoal wool suit cuff?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a stainless steel watch with a classic round case, simple bezel, and domed sapphire crystal visible on the wrist?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the watch face feature a sun-brushed silver dial with applied stick-shaped hour markers and a slender blued second hand?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the blued second hand shown sweeping continuously while the sword-shaped hour and minute hands remain stationary?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera execute a slow, smooth push-in starting from a slightly downward angle?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.014222448691725731 + ], + "opposite_logprob_tokens": [ + -4.264222621917725 + ], + "logprob_reward": -0.014222448691725731, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.014222448691725731, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10750717669725418, + 0.039217349141836166, + 0.03414700925350189, + 0.030524907633662224, + 0.026219051331281662, + 0.27723947167396545, + 0.024732153862714767, + 0.32931599020957947, + 0.0443587526679039, + 0.08673807978630066 + ], + "credit_map": [ + 0.014474630355834961, + 0.017412647604942322, + 0.01601758599281311, + 0.0137047553434968, + 0.009643388912081718, + 0.013656276278197765, + 0.009211499243974686, + 0.013386392034590244, + 0.005793333984911442, + 0.007014153059571981, + 0.0058700405061244965, + 0.005700823850929737, + 0.002720219548791647, + 0.003561581252142787, + 0.004396237898617983, + 0.0041609639301896095, + 0.004895377904176712, + 0.0035585651639848948, + 0.004768424201756716, + 0.0051806592382490635, + 0.004259701352566481, + 0.003764394437894225, + 0.004497043788433075, + 0.003222845261916518, + 0.0054234704002738, + 0.0048959446139633656, + 0.004928559064865112, + 0.002875777892768383, + 0.0026577217504382133, + 0.0026867962442338467, + 0.005166688002645969, + 0.0018899505957961082, + 0.0029188713524490595, + 0.0030745931435376406, + 0.004080094862729311, + 0.0024683752562850714, + 0.004644887521862984, + 0.002573451492935419, + 0.004285415168851614, + 0.0021733646281063557, + 0.0838346853852272, + 0.008118918165564537, + 0.1151237040758133, + 0.01084381714463234, + 0.008180507458746433, + 0.016505111008882523, + 0.004506614990532398, + 0.03012610599398613, + 0.003570477943867445, + 0.0023524535354226828, + 0.0017151522915810347, + 0.0023598873522132635, + 0.004302994813770056, + 0.00275610969401896, + 0.0028197024948894978, + 0.004855378996580839, + 0.031703948974609375, + 0.036863062530756, + 0.0537390373647213, + 0.02965591289103031, + 0.030802037566900253, + 0.05812991410493851, + 0.04025811702013016, + 0.048163965344429016, + 0.0032715469133108854, + 0.005490531213581562, + 0.007119290996342897, + 0.006939145736396313, + 0.005710880272090435, + 0.004574136342853308, + 0.0051769777201116085, + 0.006076251156628132, + 0.009103966876864433, + 0.006086298730224371, + 0.015428369864821434, + 0.008672922849655151, + 0.008747288957238197, + 0.016090935096144676, + 0.009679782204329967, + 0.012928524985909462 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 5, + 0 + ], + "selected_reward": -0.014222448691725731 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up of a slender wrist with fair, smooth skin emerging from a dark charcoal wool suit cuff?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a stainless steel watch with a classic round case, simple bezel, and domed sapphire crystal visible on the wrist?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the watch face feature a sun-brushed silver dial with applied stick-shaped hour markers and a slender blued second hand?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the blued second hand shown sweeping continuously while the sword-shaped hour and minute hands remain stationary?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera execute a slow, smooth push-in starting from a slightly downward angle?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.34869223833084106 + ], + "opposite_logprob_tokens": [ + -1.2236921787261963 + ], + "logprob_reward": -0.34869223833084106, + "yes_no_margin_reward": 0.8749999403953552, + "token_credit_reward": -0.34869223833084106, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.04610254615545273, + 0.03610154241323471, + 0.04106549546122551, + 0.034937534481287, + 0.02240867167711258, + 0.44602131843566895, + 0.05038711801171303, + 0.24231842160224915, + 0.027678070589900017, + 0.05297929793596268 + ], + "credit_map": [ + 0.007504008710384369, + 0.007437014486640692, + 0.007397993002086878, + 0.00687837228178978, + 0.0034249734599143267, + 0.004473584238439798, + 0.004474945832043886, + 0.00451165484264493, + 0.004751291126012802, + 0.005427844822406769, + 0.00538239162415266, + 0.007116436492651701, + 0.002371731214225292, + 0.0028490826953202486, + 0.003866489278152585, + 0.004336273297667503, + 0.005851952824741602, + 0.003995110280811787, + 0.0065651764161884785, + 0.006631075870245695, + 0.004603269975632429, + 0.0037022458855062723, + 0.00576796056702733, + 0.003948702942579985, + 0.004164055455476046, + 0.004841099958866835, + 0.005131121259182692, + 0.0034223422408103943, + 0.0037283061537891626, + 0.0038490763399749994, + 0.006939778570085764, + 0.0028617563657462597, + 0.0028518964536488056, + 0.0028013535775244236, + 0.0034842458553612232, + 0.0024634210858494043, + 0.003353268373757601, + 0.0022682971321046352, + 0.003420120570808649, + 0.0017660679295659065, + 0.12105485796928406, + 0.012100309133529663, + 0.17589125037193298, + 0.018071118742227554, + 0.012538442388176918, + 0.027818484231829643, + 0.009522583335638046, + 0.06902427226305008, + 0.007197793107479811, + 0.006548685487359762, + 0.007302270736545324, + 0.004278105217963457, + 0.006194896996021271, + 0.006567379459738731, + 0.004718434065580368, + 0.0075795515440404415, + 0.023163752630352974, + 0.026100123301148415, + 0.043678730726242065, + 0.02163625694811344, + 0.02298409305512905, + 0.04127535596489906, + 0.02858692780137062, + 0.034893181174993515, + 0.0026937038637697697, + 0.003054168540984392, + 0.005208332557231188, + 0.0033080028370022774, + 0.002709635067731142, + 0.0038224400486797094, + 0.002835056511685252, + 0.0040467302314937115, + 0.005954125430434942, + 0.00344268255867064, + 0.008554200641810894, + 0.004494135268032551, + 0.00497862184420228, + 0.011816548183560371, + 0.004928609821945429, + 0.0088103748857975 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 7, + 9 + ], + "selected_reward": -0.34869223833084106 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up of a slender wrist with fair, smooth skin emerging from a dark charcoal wool suit cuff?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a stainless steel watch with a classic round case, simple bezel, and domed sapphire crystal visible on the wrist?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the watch face feature a sun-brushed silver dial with applied stick-shaped hour markers and a slender blued second hand?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the blued second hand shown sweeping continuously while the sword-shaped hour and minute hands remain stationary?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera execute a slow, smooth push-in starting from a slightly downward angle?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0181814506649971 + ], + "opposite_logprob_tokens": [ + -4.018181324005127 + ], + "logprob_reward": -0.0181814506649971, + "yes_no_margin_reward": 3.999999761581421, + "token_credit_reward": -0.0181814506649971, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.04737979173660278, + 0.02278275415301323, + 0.017213920131325722, + 0.012586807832121849, + 0.02455141767859459, + 0.6570483446121216, + 0.04131607338786125, + 0.11919158697128296, + 0.03038579225540161, + 0.027543535456061363 + ], + "credit_map": [ + 0.009955259039998055, + 0.008153482340276241, + 0.007335145492106676, + 0.004580750595778227, + 0.00495406799018383, + 0.004325584042817354, + 0.003967353608459234, + 0.004108141642063856, + 0.003572906367480755, + 0.003504292806610465, + 0.003965234383940697, + 0.0039562261663377285, + 0.0015466829063370824, + 0.0022743523586541414, + 0.0019951392896473408, + 0.001967918360605836, + 0.0028097706381231546, + 0.001525570871308446, + 0.002706344937905669, + 0.0022484210785478354, + 0.0018629144178703427, + 0.0022266553714871407, + 0.002131362911313772, + 0.0017028795555233955, + 0.002338702091947198, + 0.0014353350270539522, + 0.002106201136484742, + 0.0017373530426993966, + 0.0011840073857456446, + 0.0014429150614887476, + 0.0012963631888851523, + 0.0010459290351718664, + 0.004513869062066078, + 0.0027106839697808027, + 0.0041082464158535, + 0.004199283663183451, + 0.002179334405809641, + 0.0018574395217001438, + 0.0028170307632535696, + 0.002165529178455472, + 0.19148565828800201, + 0.017780093476176262, + 0.26900655031204224, + 0.025317100808024406, + 0.017918258905410767, + 0.038455743342638016, + 0.010209067724645138, + 0.0868758037686348, + 0.00624610623344779, + 0.004352737683802843, + 0.006201221141964197, + 0.004052979405969381, + 0.005103005096316338, + 0.003843424143269658, + 0.004342550877481699, + 0.007174044381827116, + 0.011042586527764797, + 0.013271158561110497, + 0.022278938442468643, + 0.010161444544792175, + 0.00941961258649826, + 0.018681693822145462, + 0.015906648710370064, + 0.018429497256875038, + 0.0030896065291017294, + 0.0034260067623108625, + 0.004943642765283585, + 0.004968002438545227, + 0.0020908021833747625, + 0.0029117311351001263, + 0.004661459941416979, + 0.004294538404792547, + 0.001916344277560711, + 0.0032125322613865137, + 0.004125732928514481, + 0.0036449695471674204, + 0.003458531340584159, + 0.0041291420347988605, + 0.003153955563902855, + 0.0039023261051625013 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 7, + 0 + ], + "selected_reward": -0.0181814506649971 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "stunning close-up, shot with a slow, smooth push-in technique, begins from a slightly downward angle and focuses on a slender wrist gracefully raised. The skin of this wrist, fair and smooth, emerges effortlessly from the cuff of a rich, dark charcoal wool suit. With a slight twist of the wrist, a meticulously crafted stainless steel watch comes fully into view. A soft, directional, warm light illuminates the watch, reflecting a captivating luster from its interlaced polished and brushed metal surfaces. The watch boasts a classic round case and a simple bezel. Beneath the domed sapphire crystal lies a flawless, sun-brushed silver dial, its applied stick-shaped hour markers gleaming sharply in the light. When the wrist is steadily tilted towards the camera, as if at the moment of confirming the time, the focus falls precisely on the dial. A slender blued second hand is silently sweeping across the dial in an absolutely smooth and unpausing posture. Its continuous movement forms a sharp contrast with the stationary sword-shaped hour and minute hands. The whole picture conveys a luxurious atmosphere of concentration, tranquility and high-end texture in the ultimate details.", + "seed": 31225, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000095/group_0/index_6549_seed_31225_dp_1_rank_2_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 0, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.22570653352886438, + "vlm_logprob": -0.22570653352886438, + "vlm_yes_no_margin": 2.499999988079071, + "vlm_token_credit": -0.22570653352886438, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up of a slender wrist with fair, smooth skin emerging from a dark charcoal wool suit cuff?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a stainless steel watch with a classic round case, simple bezel, and domed sapphire crystal visible on the wrist?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the watch face feature a sun-brushed silver dial with applied stick-shaped hour markers and a slender blued second hand?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the blued second hand shown sweeping continuously while the sword-shaped hour and minute hands remain stationary?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera execute a slow, smooth push-in starting from a slightly downward angle?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.6352698802947998 + ], + "opposite_logprob_tokens": [ + -0.7602698802947998 + ], + "logprob_reward": -0.6352698802947998, + "yes_no_margin_reward": 0.125, + "token_credit_reward": -0.6352698802947998, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.24986153841018677, + 0.03828761354088783, + 0.07714428752660751, + 0.12538090348243713, + 0.0805327519774437, + 0.09470643848180771, + 0.05612219497561455, + 0.06446735560894012, + 0.1481935679912567, + 0.06530335545539856 + ], + "credit_map": [ + 0.02993735857307911, + 0.018751973286271095, + 0.050877053290605545, + 0.025740550830960274, + 0.021050628274679184, + 0.02913353033363819, + 0.023402947932481766, + 0.0509674996137619, + 0.0039301360957324505, + 0.00303221819922328, + 0.006805048324167728, + 0.00372675689868629, + 0.0030034661758691072, + 0.00849415548145771, + 0.003654567990452051, + 0.00564126530662179, + 0.00922178104519844, + 0.0037288330495357513, + 0.01812712475657463, + 0.006349812261760235, + 0.005066667217761278, + 0.01569865457713604, + 0.005663924850523472, + 0.013287489302456379, + 0.01567990705370903, + 0.008122596889734268, + 0.02184571698307991, + 0.013431509956717491, + 0.010312158614397049, + 0.030235575512051582, + 0.009187068790197372, + 0.016566362231969833, + 0.012504658661782742, + 0.010924182832241058, + 0.012332219630479813, + 0.008454293012619019, + 0.006075155921280384, + 0.011283042840659618, + 0.007917594164609909, + 0.01104161236435175, + 0.023688537999987602, + 0.015484217554330826, + 0.01822715811431408, + 0.007564755622297525, + 0.006595165468752384, + 0.008811605162918568, + 0.00445963442325592, + 0.009875361807644367, + 0.010488040745258331, + 0.006153393536806107, + 0.009413044899702072, + 0.003802912775427103, + 0.007263521198183298, + 0.009333977475762367, + 0.004302182700484991, + 0.005365120712667704, + 0.010625828988850117, + 0.009944487363100052, + 0.010779972188174725, + 0.0049484590999782085, + 0.005272998008877039, + 0.010403876192867756, + 0.004569412674754858, + 0.007922321557998657, + 0.04836162552237511, + 0.013825362548232079, + 0.03297964483499527, + 0.010734113864600658, + 0.005712208803743124, + 0.010377858765423298, + 0.006823807023465633, + 0.019378937780857086, + 0.013999354094266891, + 0.004637901671230793, + 0.007310505956411362, + 0.0042896391823887825, + 0.0056926230899989605, + 0.008076057769358158, + 0.005880526266992092, + 0.01541674230247736 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 3 + ], + "selected_reward": -0.6352698802947998 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up of a slender wrist with fair, smooth skin emerging from a dark charcoal wool suit cuff?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a stainless steel watch with a classic round case, simple bezel, and domed sapphire crystal visible on the wrist?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the watch face feature a sun-brushed silver dial with applied stick-shaped hour markers and a slender blued second hand?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the blued second hand shown sweeping continuously while the sword-shaped hour and minute hands remain stationary?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera execute a slow, smooth push-in starting from a slightly downward angle?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.011104804463684559 + ], + "opposite_logprob_tokens": [ + -4.511104583740234 + ], + "logprob_reward": -0.011104804463684559, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.011104804463684559, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.39551785588264465, + 0.10222670435905457, + 0.05483651161193848, + 0.046883173286914825, + 0.08818177878856659, + 0.0567929781973362, + 0.0547008141875267, + 0.021852459758520126, + 0.12331897020339966, + 0.055688705295324326 + ], + "credit_map": [ + 0.05233168601989746, + 0.02845378965139389, + 0.0778203085064888, + 0.04445153847336769, + 0.026315372437238693, + 0.04999128356575966, + 0.037640657275915146, + 0.0785132348537445, + 0.016252415254712105, + 0.016427326947450638, + 0.017257673665881157, + 0.007535248529165983, + 0.01074167713522911, + 0.010824809782207012, + 0.008786692284047604, + 0.014400859363377094, + 0.004852607846260071, + 0.0035476067569106817, + 0.010776704177260399, + 0.004324313253164291, + 0.004038217011839151, + 0.01158109400421381, + 0.004236788954585791, + 0.011479176580905914, + 0.006094734650105238, + 0.00525230448693037, + 0.007441720925271511, + 0.004236176609992981, + 0.004269396886229515, + 0.009925308637320995, + 0.004431325010955334, + 0.005232207011431456, + 0.015081975609064102, + 0.0091177336871624, + 0.014622519724071026, + 0.009453052654862404, + 0.008548683486878872, + 0.010897843167185783, + 0.009153918363153934, + 0.011306044645607471, + 0.008305426687002182, + 0.007640186697244644, + 0.007510406896471977, + 0.005661089904606342, + 0.005042735952883959, + 0.008053077384829521, + 0.005385704338550568, + 0.009194349870085716, + 0.007424117997288704, + 0.008032441139221191, + 0.00808204896748066, + 0.004660181235522032, + 0.0074096620082855225, + 0.009502761997282505, + 0.004153455141931772, + 0.005436144769191742, + 0.0031718292739242315, + 0.0031942701898515224, + 0.0032264215406030416, + 0.0023434667382389307, + 0.0024862140417099, + 0.0025893389247357845, + 0.0027109680231660604, + 0.002129950327798724, + 0.03114340640604496, + 0.012077533639967442, + 0.025420617312192917, + 0.009942234493792057, + 0.007739562075585127, + 0.02132069505751133, + 0.00628901831805706, + 0.00938590057194233, + 0.013002513907849789, + 0.005156742874532938, + 0.008983529172837734, + 0.003973479848355055, + 0.00357685680501163, + 0.006712248548865318, + 0.005246449261903763, + 0.009036880917847157 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 1 + ], + "selected_reward": -0.011104804463684559 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up of a slender wrist with fair, smooth skin emerging from a dark charcoal wool suit cuff?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a stainless steel watch with a classic round case, simple bezel, and domed sapphire crystal visible on the wrist?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the watch face feature a sun-brushed silver dial with applied stick-shaped hour markers and a slender blued second hand?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the blued second hand shown sweeping continuously while the sword-shaped hour and minute hands remain stationary?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera execute a slow, smooth push-in starting from a slightly downward angle?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02332243137061596 + ], + "opposite_logprob_tokens": [ + -3.773322343826294 + ], + "logprob_reward": -0.02332243137061596, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.02332243137061596, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.25798749923706055, + 0.08582469075918198, + 0.05863511934876442, + 0.04394260793924332, + 0.05120425671339035, + 0.04155850410461426, + 0.049888975918293, + 0.027568910270929337, + 0.30857470631599426, + 0.07481470704078674 + ], + "credit_map": [ + 0.03345389664173126, + 0.01950879767537117, + 0.05125713720917702, + 0.02842923253774643, + 0.019046682864427567, + 0.03200002759695053, + 0.02279401756823063, + 0.05149770900607109, + 0.008045404218137264, + 0.007481065113097429, + 0.022785695269703865, + 0.00855544488877058, + 0.005874818656593561, + 0.008160301484167576, + 0.00756964972242713, + 0.017352312803268433, + 0.005851994268596172, + 0.003819867502897978, + 0.012244819663465023, + 0.004002874251455069, + 0.004178242292255163, + 0.008703505620360374, + 0.005289332941174507, + 0.014544486068189144, + 0.006063906010240316, + 0.0031535583548247814, + 0.006920916959643364, + 0.004396307282149792, + 0.0038366075605154037, + 0.01101711206138134, + 0.0032736086286604404, + 0.005280590150505304, + 0.006427842192351818, + 0.0060770269483327866, + 0.006929096765816212, + 0.005531458184123039, + 0.005422309041023254, + 0.007692602928727865, + 0.005941588897258043, + 0.007182335015386343, + 0.006143742240965366, + 0.005145876202732325, + 0.007237544748932123, + 0.0040713706985116005, + 0.0038343342021107674, + 0.007290605455636978, + 0.0028566448017954826, + 0.004978385753929615, + 0.010125670582056046, + 0.003939997870475054, + 0.009111197665333748, + 0.0033032686915248632, + 0.006234342698007822, + 0.008556544780731201, + 0.003112124279141426, + 0.005505828652530909, + 0.00529417535290122, + 0.00281082303263247, + 0.004931342788040638, + 0.0016511823050677776, + 0.002232299419119954, + 0.005239806137979031, + 0.0023875541519373655, + 0.0030217263847589493, + 0.06986600160598755, + 0.015539714135229588, + 0.05597207695245743, + 0.015236818231642246, + 0.022965198382735252, + 0.07937543094158173, + 0.011383882723748684, + 0.038235604763031006, + 0.021681902930140495, + 0.003934570588171482, + 0.014328060671687126, + 0.003963726572692394, + 0.005208220798522234, + 0.00891781970858574, + 0.0036813057959079742, + 0.01309910323470831 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 0, + 1 + ], + "selected_reward": -0.02332243137061596 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up of a slender wrist with fair, smooth skin emerging from a dark charcoal wool suit cuff?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a stainless steel watch with a classic round case, simple bezel, and domed sapphire crystal visible on the wrist?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the watch face feature a sun-brushed silver dial with applied stick-shaped hour markers and a slender blued second hand?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the blued second hand shown sweeping continuously while the sword-shaped hour and minute hands remain stationary?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera execute a slow, smooth push-in starting from a slightly downward angle?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.42903977632522583 + ], + "opposite_logprob_tokens": [ + -1.054039716720581 + ], + "logprob_reward": -0.42903977632522583, + "yes_no_margin_reward": 0.6249999403953552, + "token_credit_reward": -0.42903977632522583, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15357783436775208, + 0.07754085958003998, + 0.046955667436122894, + 0.0842161476612091, + 0.06684701144695282, + 0.06337596476078033, + 0.07853083312511444, + 0.03162870556116104, + 0.3177688419818878, + 0.07955814152956009 + ], + "credit_map": [ + 0.015024282969534397, + 0.013633140362799168, + 0.03138674423098564, + 0.013541917316615582, + 0.010937206447124481, + 0.01679644174873829, + 0.017803004011511803, + 0.034455105662345886, + 0.011578556150197983, + 0.007819371297955513, + 0.011232689023017883, + 0.006546413525938988, + 0.006612950004637241, + 0.013685050420463085, + 0.00987133476883173, + 0.010194501839578152, + 0.005662090610712767, + 0.00449022650718689, + 0.011876399628818035, + 0.003907760605216026, + 0.003116748295724392, + 0.0042261891067028046, + 0.005121943075209856, + 0.008554305881261826, + 0.012280065566301346, + 0.006327994633466005, + 0.016841096803545952, + 0.010267173871397972, + 0.005858035292476416, + 0.015890805050730705, + 0.006879110354930162, + 0.00987186748534441, + 0.009600672870874405, + 0.006417895667254925, + 0.009997361339628696, + 0.006667932029813528, + 0.006573211867362261, + 0.011179689317941666, + 0.007889931090176105, + 0.008520317263901234, + 0.012522002682089806, + 0.009444852359592915, + 0.010933881625533104, + 0.005721327383071184, + 0.006022660527378321, + 0.006505887024104595, + 0.004440089222043753, + 0.007785262539982796, + 0.011129260994493961, + 0.011181395500898361, + 0.013427182100713253, + 0.006294316612184048, + 0.009631898254156113, + 0.012923276983201504, + 0.005442499183118343, + 0.008501015603542328, + 0.004190745297819376, + 0.005291324574500322, + 0.004361495841294527, + 0.002573024481534958, + 0.002602781169116497, + 0.005760088097304106, + 0.002586547052487731, + 0.004262699279934168, + 0.1167214959859848, + 0.019400574266910553, + 0.06910410523414612, + 0.013606635853648186, + 0.011541691608726978, + 0.022402381524443626, + 0.009698777459561825, + 0.05529317259788513, + 0.01660579815506935, + 0.004152277484536171, + 0.014497485011816025, + 0.006029476877301931, + 0.006747541017830372, + 0.011275764554738998, + 0.005623499862849712, + 0.014626299031078815 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 0, + 3 + ], + "selected_reward": -0.42903977632522583 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up of a slender wrist with fair, smooth skin emerging from a dark charcoal wool suit cuff?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a stainless steel watch with a classic round case, simple bezel, and domed sapphire crystal visible on the wrist?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the watch face feature a sun-brushed silver dial with applied stick-shaped hour markers and a slender blued second hand?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the blued second hand shown sweeping continuously while the sword-shaped hour and minute hands remain stationary?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera execute a slow, smooth push-in starting from a slightly downward angle?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.029795775189995766 + ], + "opposite_logprob_tokens": [ + -3.5297958850860596 + ], + "logprob_reward": -0.029795775189995766, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.029795775189995766, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.054881349205970764, + 0.06623121351003647, + 0.13552528619766235, + 0.05844709277153015, + 0.14457622170448303, + 0.19688712060451508, + 0.03117086924612522, + 0.02772919088602066, + 0.2645247280597687, + 0.020026864483952522 + ], + "credit_map": [ + 0.00587364612147212, + 0.003765571629628539, + 0.009438837878406048, + 0.0050268773920834064, + 0.004846422933042049, + 0.008936714380979538, + 0.0067256297916173935, + 0.010267652571201324, + 0.005633887369185686, + 0.006992185954004526, + 0.010537439957261086, + 0.009024850092828274, + 0.006398048717528582, + 0.011830363422632217, + 0.006322016008198261, + 0.009492426179349422, + 0.015335110016167164, + 0.009045673534274101, + 0.021177763119339943, + 0.013899203389883041, + 0.009837278164923191, + 0.03626195713877678, + 0.009975812397897243, + 0.01999248005449772, + 0.008143709041178226, + 0.007120100315660238, + 0.011007651686668396, + 0.005662685725837946, + 0.005343121942132711, + 0.007413783576339483, + 0.006921133492141962, + 0.006834906060248613, + 0.017795978114008904, + 0.01429709978401661, + 0.01828628033399582, + 0.01576848514378071, + 0.013122493401169777, + 0.02669893018901348, + 0.01491275429725647, + 0.02369421347975731, + 0.04420808330178261, + 0.028142744675278664, + 0.03292316198348999, + 0.016009444370865822, + 0.013583151623606682, + 0.02151576429605484, + 0.016217811033129692, + 0.024286961182951927, + 0.0053412956185638905, + 0.005406249780207872, + 0.005481013562530279, + 0.002530792960897088, + 0.0031975333113223314, + 0.0041407071985304356, + 0.002365651074796915, + 0.0027076255064457655, + 0.003924775868654251, + 0.0031118406914174557, + 0.004020403139293194, + 0.002500730799511075, + 0.0027667791582643986, + 0.00499275466427207, + 0.0027319167274981737, + 0.0036799905356019735, + 0.07557276636362076, + 0.024426335468888283, + 0.05138915404677391, + 0.01643730327486992, + 0.009724924340844154, + 0.021158168092370033, + 0.010617549531161785, + 0.05519852787256241, + 0.0036724135279655457, + 0.0014489279128611088, + 0.004701666068285704, + 0.0015421513235196471, + 0.0016098364721983671, + 0.0031980876810848713, + 0.0014071946498006582, + 0.0024465874303132296 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 5, + 4 + ], + "selected_reward": -0.029795775189995766 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000095/group_1/index_tensor([6412, 6412])_group_1_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000095/group_1/index_tensor([6412, 6412])_group_1_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..831f7695198a1099f86b6bfe5c4a0277b120d6a2 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000095/group_1/index_tensor([6412, 6412])_group_1_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "sweeping, sweeping 360-degree panning shot captures the epic birth and evolution of a city core constructed from luminescent voxels. Initially, the frame's center is a deep digital void, with only a few solitary, translucent voxels pulsing in the darkness with a faint electric blue glow, like dormant data seeds. As the camera steadily pans, a breathtaking process of creation unfolds at an exponential rate: these voxels begin to replicate, aggregate, and stack upward, forming complex geometric structures that rise from a barren digital plain. In mere seconds, a futuristic metropolis of skyscrapers, composed of countless voxels shimmering in neon pink, cyber green, and bright orange, takes shape. Its sharp, intricate outlines resemble a giant integrated circuit board. Meanwhile, streams of tiny cars, composed of pure white light, speed across the multi-layered, transparent aerial highway, leaving behind brilliant trails of light and imbuing the self-luminous city with a relentless energy. The entire scene appears magnificent and awe-inspiring under the cool digital light, creating a high-tech, fast-paced, futuristic atmosphere full of infinite possibilities.", + "seed": 4895, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000095/group_1/index_6412_seed_4895_dp_2_rank_4_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 1, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.36365932105109094, + "vlm_logprob": -0.36365932105109094, + "vlm_yes_no_margin": 2.825, + "vlm_token_credit": -0.36365932105109094, + "vlm_yes_ratio": 0.8, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the camera execute a 360-degree panning shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the buildings constructed from luminescent voxels?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the skyscrapers shimmer in neon pink, cyber green, and bright orange colors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are there streams of tiny cars composed of pure white light moving on transparent aerial highways?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the cars leave behind brilliant trails of light as they move?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -1.7048823833465576 + ], + "opposite_logprob_tokens": [ + -0.2048824280500412 + ], + "logprob_reward": -1.7048823833465576, + "yes_no_margin_reward": -1.5, + "token_credit_reward": -1.7048823833465576, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.15531395375728607, + 0.10429293662309647, + 0.11385367065668106, + 0.04077085107564926, + 0.09251108765602112, + 0.08515030890703201, + 0.028117919340729713, + 0.09301158785820007, + 0.2605457305908203, + 0.02643190324306488 + ], + "credit_map": [ + 0.015872590243816376, + 0.019253386184573174, + 0.02011759765446186, + 0.02576075680553913, + 0.02333643287420273, + 0.016786595806479454, + 0.016643421724438667, + 0.017543183639645576, + 0.011206913739442825, + 0.011522091925144196, + 0.016306361183524132, + 0.01590498723089695, + 0.011583142913877964, + 0.014929274097084999, + 0.014320334419608116, + 0.008519834838807583, + 0.009992578998208046, + 0.013708760030567646, + 0.01985182985663414, + 0.0118860499933362, + 0.01112124789506197, + 0.021951239556074142, + 0.015257783234119415, + 0.010084187611937523, + 0.005482015665620565, + 0.0049961325712502, + 0.006395284086465836, + 0.0068687680177390575, + 0.005014803726226091, + 0.004149578511714935, + 0.0044492208398878574, + 0.003415047423914075, + 0.008234528824687004, + 0.022174397483468056, + 0.00644065672531724, + 0.008599581196904182, + 0.01467236876487732, + 0.011272002011537552, + 0.007243141066282988, + 0.013874412514269352, + 0.01621648296713829, + 0.010623163543641567, + 0.0038722942117601633, + 0.005455491598695517, + 0.0133886206895113, + 0.010380196385085583, + 0.006041119806468487, + 0.01917293667793274, + 0.003231445560231805, + 0.00263957935385406, + 0.0021424253936856985, + 0.004953501280397177, + 0.0037344936281442642, + 0.004015597514808178, + 0.0040535153821110725, + 0.00334736006334424, + 0.006257109344005585, + 0.006104675587266684, + 0.007485848851501942, + 0.028766360133886337, + 0.012010442093014717, + 0.006700064986944199, + 0.009235790930688381, + 0.01645130291581154, + 0.07790416479110718, + 0.011489070020616055, + 0.015442850068211555, + 0.010851064696907997, + 0.010332605801522732, + 0.10021665692329407, + 0.011890730820596218, + 0.022418584674596786, + 0.0027858284302055836, + 0.004036851692944765, + 0.0016347516793757677, + 0.002276080194860697, + 0.0019641697872430086, + 0.005225156433880329, + 0.004043817985802889, + 0.004465245641767979 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 0, + 2 + ], + "selected_reward": -1.7048823833465576 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the camera execute a 360-degree panning shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the buildings constructed from luminescent voxels?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the skyscrapers shimmer in neon pink, cyber green, and bright orange colors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are there streams of tiny cars composed of pure white light moving on transparent aerial highways?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the cars leave behind brilliant trails of light as they move?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0206286683678627 + ], + "opposite_logprob_tokens": [ + -3.8956286907196045 + ], + "logprob_reward": -0.0206286683678627, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.0206286683678627, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0695803314447403, + 0.08580098301172256, + 0.1099780797958374, + 0.06881667673587799, + 0.07009956985712051, + 0.07277210056781769, + 0.04106849804520607, + 0.02239331044256687, + 0.3924308717250824, + 0.06705954670906067 + ], + "credit_map": [ + 0.005228367634117603, + 0.0078989053145051, + 0.012658670544624329, + 0.01022223848849535, + 0.009151602163910866, + 0.00983234029263258, + 0.007351136300712824, + 0.007237071171402931, + 0.007288139779120684, + 0.008785008452832699, + 0.015205749310553074, + 0.013318793848156929, + 0.008530445396900177, + 0.014566128142178059, + 0.010519559495151043, + 0.007587155792862177, + 0.009786068461835384, + 0.01540465373545885, + 0.011393544264137745, + 0.01776205562055111, + 0.015490934252738953, + 0.013589759357273579, + 0.017614293843507767, + 0.008936777710914612, + 0.005290750414133072, + 0.005867481231689453, + 0.009763221256434917, + 0.012294381856918335, + 0.011760732159018517, + 0.0100800646468997, + 0.008523637428879738, + 0.005236409138888121, + 0.00685171177610755, + 0.01072329469025135, + 0.005665685050189495, + 0.00547791225835681, + 0.011635017581284046, + 0.012678401544690132, + 0.0073457383550703526, + 0.009721806272864342, + 0.008916030637919903, + 0.012943307869136333, + 0.005302297882735729, + 0.0058676800690591335, + 0.010871865786612034, + 0.008640349842607975, + 0.005911423359066248, + 0.014319142326712608, + 0.004795399494469166, + 0.0048829722218215466, + 0.002427857369184494, + 0.005855212919414043, + 0.005987408105283976, + 0.007442949339747429, + 0.005248932167887688, + 0.004427768290042877, + 0.0024520643055438995, + 0.0022254108916968107, + 0.0018246631370857358, + 0.005103125236928463, + 0.004112336318939924, + 0.0016962062800303102, + 0.0023135633673518896, + 0.002665939275175333, + 0.11161668598651886, + 0.0173331331461668, + 0.02404875122010708, + 0.01784929633140564, + 0.01601232774555683, + 0.15209519863128662, + 0.01978381723165512, + 0.0336916521191597, + 0.004121091682463884, + 0.005461928900331259, + 0.0042581092566251755, + 0.005962691269814968, + 0.0034225371200591326, + 0.014869175851345062, + 0.0212344191968441, + 0.007729591801762581 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 2, + 1 + ], + "selected_reward": -0.0206286683678627 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the camera execute a 360-degree panning shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the buildings constructed from luminescent voxels?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the skyscrapers shimmer in neon pink, cyber green, and bright orange colors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are there streams of tiny cars composed of pure white light moving on transparent aerial highways?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the cars leave behind brilliant trails of light as they move?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004110222216695547 + ], + "opposite_logprob_tokens": [ + -5.504110336303711 + ], + "logprob_reward": -0.004110222216695547, + "yes_no_margin_reward": 5.5, + "token_credit_reward": -0.004110222216695547, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10625887662172318, + 0.04775013029575348, + 0.04223362356424332, + 0.036919619888067245, + 0.04648708924651146, + 0.06007159873843193, + 0.058502502739429474, + 0.48316943645477295, + 0.05458204820752144, + 0.06402510404586792 + ], + "credit_map": [ + 0.00841780286282301, + 0.011168488301336765, + 0.02175096422433853, + 0.01817246526479721, + 0.012078018859028816, + 0.015528089366853237, + 0.010298231616616249, + 0.00884481705725193, + 0.004895963706076145, + 0.005405736621469259, + 0.010199247859418392, + 0.007287765387445688, + 0.0034688308369368315, + 0.007766298949718475, + 0.004640190862119198, + 0.004086094908416271, + 0.0033431686460971832, + 0.004348244518041611, + 0.006294379010796547, + 0.008003519847989082, + 0.003919895272701979, + 0.007696034852415323, + 0.005425820127129555, + 0.003202559659257531, + 0.003922567702829838, + 0.0035640473943203688, + 0.006141092162579298, + 0.005768985487520695, + 0.004680452402681112, + 0.00511395325884223, + 0.004765253979712725, + 0.0029632672667503357, + 0.006142383441329002, + 0.005372635088860989, + 0.0047098612412810326, + 0.004572263918817043, + 0.00752624636515975, + 0.007308422587811947, + 0.005796958226710558, + 0.005058317445218563, + 0.008698602207005024, + 0.012111173942685127, + 0.003367139957845211, + 0.00440012663602829, + 0.00942139234393835, + 0.008256017230451107, + 0.004297301173210144, + 0.009519844315946102, + 0.006707320921123028, + 0.007118786219507456, + 0.004004438407719135, + 0.00760340690612793, + 0.006142054218798876, + 0.010939761996269226, + 0.008873273618519306, + 0.0071134623140096664, + 0.03355187550187111, + 0.03173363581299782, + 0.04042566940188408, + 0.134657084941864, + 0.028579674661159515, + 0.034631289541721344, + 0.05213494971394539, + 0.12745524942874908, + 0.01432554330676794, + 0.0027864219155162573, + 0.003304962068796158, + 0.003475037869066, + 0.0024620124604552984, + 0.020126672461628914, + 0.0034573001321405172, + 0.00464409776031971, + 0.006934639532119036, + 0.00987886730581522, + 0.004328527487814426, + 0.007417036686092615, + 0.0037585627287626266, + 0.012333790771663189, + 0.011617777869105339, + 0.007755897007882595 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 0, + 9 + ], + "selected_reward": -0.004110222216695547 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the camera execute a 360-degree panning shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the buildings constructed from luminescent voxels?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the skyscrapers shimmer in neon pink, cyber green, and bright orange colors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are there streams of tiny cars composed of pure white light moving on transparent aerial highways?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the cars leave behind brilliant trails of light as they move?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.054970432072877884 + ], + "opposite_logprob_tokens": [ + -2.9299705028533936 + ], + "logprob_reward": -0.054970432072877884, + "yes_no_margin_reward": 2.875, + "token_credit_reward": -0.054970432072877884, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12687285244464874, + 0.07460596412420273, + 0.055067919194698334, + 0.08279354125261307, + 0.09148084372282028, + 0.06875208020210266, + 0.08374611288309097, + 0.09642372280359268, + 0.251091867685318, + 0.06916507333517075 + ], + "credit_map": [ + 0.019949093461036682, + 0.014393661171197891, + 0.01965922676026821, + 0.023733433336019516, + 0.013636893592774868, + 0.012953976169228554, + 0.01070956140756607, + 0.011837007477879524, + 0.007360588293522596, + 0.008709321729838848, + 0.01243459340184927, + 0.015584481880068779, + 0.005022943951189518, + 0.013114755973219872, + 0.006133340299129486, + 0.006245940458029509, + 0.0048294817097485065, + 0.007617798168212175, + 0.007443825714290142, + 0.006270173471421003, + 0.008245855569839478, + 0.008501212112605572, + 0.0061245872639119625, + 0.0060349879786372185, + 0.0090198228135705, + 0.009138673543930054, + 0.01617269404232502, + 0.013852754607796669, + 0.00967471580952406, + 0.012030033394694328, + 0.006929047405719757, + 0.005975800100713968, + 0.008532674983143806, + 0.015122572891414165, + 0.009077636525034904, + 0.008584221825003624, + 0.013158325105905533, + 0.013426847755908966, + 0.010771496221423149, + 0.012807072140276432, + 0.008179944008588791, + 0.011308628134429455, + 0.005838998593389988, + 0.0060881697572767735, + 0.009958493523299694, + 0.011242514476180077, + 0.005971329286694527, + 0.010164003819227219, + 0.007678191177546978, + 0.007037780247628689, + 0.00511082261800766, + 0.012675920501351357, + 0.009073164314031601, + 0.015763018280267715, + 0.013964502140879631, + 0.012442715466022491, + 0.007677095476537943, + 0.006806089077144861, + 0.008336232975125313, + 0.023661376908421516, + 0.006629894953221083, + 0.008309081196784973, + 0.009905683808028698, + 0.025098275393247604, + 0.049876075237989426, + 0.014586541801691055, + 0.020980628207325935, + 0.01905660331249237, + 0.013756769709289074, + 0.08868983387947083, + 0.019828977063298225, + 0.024316439405083656, + 0.0065281204879283905, + 0.004640733357518911, + 0.0035671857185661793, + 0.005341485142707825, + 0.0037344908341765404, + 0.01842586323618889, + 0.01991674304008484, + 0.0070104473270475864 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 0, + 7 + ], + "selected_reward": -0.054970432072877884 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the camera execute a 360-degree panning shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the buildings constructed from luminescent voxels?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the skyscrapers shimmer in neon pink, cyber green, and bright orange colors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are there streams of tiny cars composed of pure white light moving on transparent aerial highways?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the cars leave behind brilliant trails of light as they move?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03370489925146103 + ], + "opposite_logprob_tokens": [ + -3.408704996109009 + ], + "logprob_reward": -0.03370489925146103, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.03370489925146103, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.17007380723953247, + 0.09459476917982101, + 0.0806080773472786, + 0.08504600822925568, + 0.10910007357597351, + 0.1717885285615921, + 0.0873376727104187, + 0.05673046410083771, + 0.0732477605342865, + 0.07147287577390671 + ], + "credit_map": [ + 0.015773599967360497, + 0.01899072900414467, + 0.03700312227010727, + 0.024866310879588127, + 0.019264893606305122, + 0.024580156430602074, + 0.01607421413064003, + 0.013520761393010616, + 0.011723603121936321, + 0.012888304889202118, + 0.013161107897758484, + 0.01706286147236824, + 0.0079159215092659, + 0.014002753421664238, + 0.008735385723412037, + 0.009104828350245953, + 0.005919753108173609, + 0.010096915066242218, + 0.013431394472718239, + 0.011383592151105404, + 0.00832117535173893, + 0.015903187915682793, + 0.00932552944868803, + 0.006226537749171257, + 0.006473850458860397, + 0.0077003054320812225, + 0.011249039322137833, + 0.015492318198084831, + 0.015240990556776524, + 0.012001913972198963, + 0.009897394105792046, + 0.00699020316824317, + 0.014623652212321758, + 0.017270343378186226, + 0.00904065277427435, + 0.013153512962162495, + 0.014675726182758808, + 0.01735542342066765, + 0.01215823832899332, + 0.010822520591318607, + 0.019685495644807816, + 0.0312354639172554, + 0.01720176637172699, + 0.018627019599080086, + 0.02323734387755394, + 0.01708841137588024, + 0.025442292913794518, + 0.01927074044942856, + 0.010639830492436886, + 0.009477270767092705, + 0.00537539180368185, + 0.011481906287372112, + 0.010100733488798141, + 0.01827009581029415, + 0.00986211933195591, + 0.012130320072174072, + 0.005823945160955191, + 0.004426905885338783, + 0.006770777050405741, + 0.0072549255564808846, + 0.012123780325055122, + 0.005818620789796114, + 0.008367982693016529, + 0.006143528036773205, + 0.017151476815342903, + 0.003430743236094713, + 0.004282029345631599, + 0.004021735396236181, + 0.0032285931520164013, + 0.026847347617149353, + 0.004995641298592091, + 0.009290196932852268, + 0.008693337440490723, + 0.00597875053063035, + 0.0035758642479777336, + 0.005707003176212311, + 0.0046335309743881226, + 0.015954935923218727, + 0.017032131552696228, + 0.009897315874695778 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 0, + 4 + ], + "selected_reward": -0.03370489925146103 + } + ], + "vlm_decoded_outputs": [ + "No", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "sweeping, sweeping 360-degree panning shot captures the epic birth and evolution of a city core constructed from luminescent voxels. Initially, the frame's center is a deep digital void, with only a few solitary, translucent voxels pulsing in the darkness with a faint electric blue glow, like dormant data seeds. As the camera steadily pans, a breathtaking process of creation unfolds at an exponential rate: these voxels begin to replicate, aggregate, and stack upward, forming complex geometric structures that rise from a barren digital plain. In mere seconds, a futuristic metropolis of skyscrapers, composed of countless voxels shimmering in neon pink, cyber green, and bright orange, takes shape. Its sharp, intricate outlines resemble a giant integrated circuit board. Meanwhile, streams of tiny cars, composed of pure white light, speed across the multi-layered, transparent aerial highway, leaving behind brilliant trails of light and imbuing the self-luminous city with a relentless energy. The entire scene appears magnificent and awe-inspiring under the cool digital light, creating a high-tech, fast-paced, futuristic atmosphere full of infinite possibilities.", + "seed": 4896, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000095/group_1/index_6412_seed_4896_dp_2_rank_4_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 1, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.28360350118018685, + "vlm_logprob": -0.28360350118018685, + "vlm_yes_no_margin": 3.1, + "vlm_token_credit": -0.28360350118018685, + "vlm_yes_ratio": 0.8, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the camera execute a 360-degree panning shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the buildings constructed from luminescent voxels?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the skyscrapers shimmer in neon pink, cyber green, and bright orange colors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are there streams of tiny cars composed of pure white light moving on transparent aerial highways?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the cars leave behind brilliant trails of light as they move?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -1.3198496103286743 + ], + "opposite_logprob_tokens": [ + -0.3198496401309967 + ], + "logprob_reward": -1.3198496103286743, + "yes_no_margin_reward": -1.0, + "token_credit_reward": -1.3198496103286743, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.32141268253326416, + 0.05567781627178192, + 0.17293362319469452, + 0.050246816128492355, + 0.058535218238830566, + 0.07769657671451569, + 0.03617509827017784, + 0.19203415513038635, + 0.021956942975521088, + 0.013331088237464428 + ], + "credit_map": [ + 0.04606557637453079, + 0.05223506689071655, + 0.018306676298379898, + 0.024859795346856117, + 0.0314757339656353, + 0.11201037466526031, + 0.015051147900521755, + 0.0214083194732666, + 0.006628185976296663, + 0.00861690565943718, + 0.006475361529737711, + 0.005065141245722771, + 0.006228047423064709, + 0.006725999526679516, + 0.00602538650855422, + 0.009912789799273014, + 0.009759009815752506, + 0.05611053481698036, + 0.013116900809109211, + 0.011207208968698978, + 0.056327369064092636, + 0.010961621068418026, + 0.006896798498928547, + 0.008554164320230484, + 0.005334437359124422, + 0.006412080954760313, + 0.008136548101902008, + 0.0051094042137265205, + 0.008623871952295303, + 0.009057027287781239, + 0.0036696079187095165, + 0.003903838573023677, + 0.006129249464720488, + 0.006280695553869009, + 0.007385467644780874, + 0.008536193519830704, + 0.007870364002883434, + 0.009160433895885944, + 0.008139519020915031, + 0.0050332932732999325, + 0.012158730067312717, + 0.011711404658854008, + 0.007695784792304039, + 0.00967011321336031, + 0.010160482488572598, + 0.01089868601411581, + 0.008838680572807789, + 0.006562696769833565, + 0.004418825730681419, + 0.0050064316019415855, + 0.003075025510042906, + 0.003339405870065093, + 0.0064593590795993805, + 0.0034714024513959885, + 0.005754444282501936, + 0.004650203976780176, + 0.016817571595311165, + 0.06588078290224075, + 0.06593579053878784, + 0.011918985284864902, + 0.007229038514196873, + 0.008161990903317928, + 0.010805556550621986, + 0.005284443031996489, + 0.004123486578464508, + 0.0022907822858542204, + 0.0026930009480565786, + 0.0028117068577557802, + 0.002919022925198078, + 0.0031697028316557407, + 0.0023634021636098623, + 0.0015858383849263191, + 0.0011830634903162718, + 0.0015360353281721473, + 0.001699355198070407, + 0.0023765156511217356, + 0.0015078949509188533, + 0.0014309082180261612, + 0.0017173256492242217, + 0.0018799902172759175 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 7, + 2 + ], + "selected_reward": -1.3198496103286743 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the camera execute a 360-degree panning shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the buildings constructed from luminescent voxels?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the skyscrapers shimmer in neon pink, cyber green, and bright orange colors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are there streams of tiny cars composed of pure white light moving on transparent aerial highways?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the cars leave behind brilliant trails of light as they move?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03375711292028427 + ], + "opposite_logprob_tokens": [ + -3.408757209777832 + ], + "logprob_reward": -0.03375711292028427, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.03375711292028427, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12205930799245834, + 0.046815209090709686, + 0.504700243473053, + 0.016872743144631386, + 0.02206895500421524, + 0.030967874452471733, + 0.04991322383284569, + 0.12272915244102478, + 0.05216905474662781, + 0.03170425817370415 + ], + "credit_map": [ + 0.012439406476914883, + 0.013738494366407394, + 0.016727231442928314, + 0.022509077563881874, + 0.013544201850891113, + 0.013202554546296597, + 0.017250224947929382, + 0.012648114003241062, + 0.0045720371417701244, + 0.007428667042404413, + 0.005567999556660652, + 0.0034056014847010374, + 0.005424424074590206, + 0.007379619404673576, + 0.004086051136255264, + 0.008950808085501194, + 0.014264734461903572, + 0.21828189492225647, + 0.009349899366497993, + 0.010499312542378902, + 0.21672001481056213, + 0.017487114295363426, + 0.011532776057720184, + 0.0065644956193864346, + 0.0018315702909603715, + 0.002453271998092532, + 0.0027506318874657154, + 0.002063060412183404, + 0.002345802029594779, + 0.002711560809984803, + 0.001484881155192852, + 0.0012319653760641813, + 0.002121696015819907, + 0.0026632051449269056, + 0.002337249694392085, + 0.0030090692453086376, + 0.0036881454288959503, + 0.00364279979839921, + 0.0027639486361294985, + 0.0018428424373269081, + 0.004986993502825499, + 0.004572706762701273, + 0.0031218177173286676, + 0.003698192071169615, + 0.0041949073784053326, + 0.004120220895856619, + 0.003628685139119625, + 0.002644349355250597, + 0.004378166515380144, + 0.0074267820455133915, + 0.0038913812022656202, + 0.00586190540343523, + 0.0047950539737939835, + 0.006063504610210657, + 0.005854887422174215, + 0.011641542427241802, + 0.04199058189988136, + 0.04685070365667343, + 0.007109528407454491, + 0.005709171760827303, + 0.004614355508238077, + 0.005732086952775717, + 0.007986603304743767, + 0.002736113965511322, + 0.005139891989529133, + 0.007969453930854797, + 0.0067137619480490685, + 0.008049151860177517, + 0.006500916089862585, + 0.005513190291821957, + 0.006488043814897537, + 0.005794644821435213, + 0.0033383064437657595, + 0.003524057101458311, + 0.0044197263196110725, + 0.004385027568787336, + 0.004772079177200794, + 0.0033687499817460775, + 0.004007099196314812, + 0.003889210755005479 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 7, + 0 + ], + "selected_reward": -0.03375711292028427 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the camera execute a 360-degree panning shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the buildings constructed from luminescent voxels?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the skyscrapers shimmer in neon pink, cyber green, and bright orange colors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are there streams of tiny cars composed of pure white light moving on transparent aerial highways?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the cars leave behind brilliant trails of light as they move?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.00320504535920918 + ], + "opposite_logprob_tokens": [ + -5.753204822540283 + ], + "logprob_reward": -0.00320504535920918, + "yes_no_margin_reward": 5.75, + "token_credit_reward": -0.00320504535920918, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.29015564918518066, + 0.026308441534638405, + 0.32145753502845764, + 0.051549602299928665, + 0.033287689089775085, + 0.034988850355148315, + 0.047352924942970276, + 0.0997404083609581, + 0.05204102024435997, + 0.04311785101890564 + ], + "credit_map": [ + 0.038136888295412064, + 0.04350965842604637, + 0.019089575856924057, + 0.02950209565460682, + 0.027860619127750397, + 0.0928383320569992, + 0.018126944079995155, + 0.021091528236865997, + 0.0025482922792434692, + 0.0031492779962718487, + 0.0032034043688327074, + 0.0027838817331939936, + 0.003544817678630352, + 0.004064918961375952, + 0.0026727570220828056, + 0.004341091960668564, + 0.009146002121269703, + 0.13641925156116486, + 0.007183772511780262, + 0.007313006557524204, + 0.13695017993450165, + 0.011699513532221317, + 0.00775923952460289, + 0.004986593499779701, + 0.004169994033873081, + 0.004992570262402296, + 0.003773674601688981, + 0.00611224677413702, + 0.01277206838130951, + 0.011219202540814877, + 0.005164240021258593, + 0.0033456047531217337, + 0.003062077099457383, + 0.003926073666661978, + 0.004290820099413395, + 0.005137552972882986, + 0.004018284380435944, + 0.00523257115855813, + 0.004949433729052544, + 0.002670872490853071, + 0.004203345160931349, + 0.0063806818798184395, + 0.003300524316728115, + 0.003711977042257786, + 0.005406237673014402, + 0.004985428415238857, + 0.003938439302146435, + 0.0030622186604887247, + 0.00540575385093689, + 0.007249536924064159, + 0.003930921200662851, + 0.005776930600404739, + 0.004640878178179264, + 0.004529528319835663, + 0.006674355361610651, + 0.009145019575953484, + 0.008759980089962482, + 0.033997781574726105, + 0.03338409960269928, + 0.006206966936588287, + 0.004194171167910099, + 0.0043933275155723095, + 0.0060943723656237125, + 0.0027097126003354788, + 0.006018544547259808, + 0.008353864774107933, + 0.006191765423864126, + 0.006249142810702324, + 0.008923325687646866, + 0.005615037865936756, + 0.005691733211278915, + 0.0049976082518696785, + 0.004499510861933231, + 0.006890425458550453, + 0.003542753402143717, + 0.005890961736440659, + 0.007987194694578648, + 0.004590298049151897, + 0.0031471368856728077, + 0.006569571327418089 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 0, + 7 + ], + "selected_reward": -0.00320504535920918 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the camera execute a 360-degree panning shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the buildings constructed from luminescent voxels?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the skyscrapers shimmer in neon pink, cyber green, and bright orange colors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are there streams of tiny cars composed of pure white light moving on transparent aerial highways?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the cars leave behind brilliant trails of light as they move?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.048662614077329636 + ], + "opposite_logprob_tokens": [ + -3.0486626625061035 + ], + "logprob_reward": -0.048662614077329636, + "yes_no_margin_reward": 3.0, + "token_credit_reward": -0.048662614077329636, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09850046783685684, + 0.012017548084259033, + 0.5703685283660889, + 0.028280355036258698, + 0.022393546998500824, + 0.02318427711725235, + 0.025459034368395805, + 0.15594050288200378, + 0.0059739816933870316, + 0.05788175016641617 + ], + "credit_map": [ + 0.01522129774093628, + 0.01682562194764614, + 0.005853051319718361, + 0.005147059913724661, + 0.009557798504829407, + 0.03660193458199501, + 0.0034727195743471384, + 0.005820984952151775, + 0.001758676371537149, + 0.0019326528999954462, + 0.0012285539414733648, + 0.0010341558372601867, + 0.0012184313964098692, + 0.0014186620246618986, + 0.0010815704008564353, + 0.002344845561310649, + 0.015687867999076843, + 0.250439316034317, + 0.0069847372360527515, + 0.010040301829576492, + 0.24960045516490936, + 0.019152162596583366, + 0.012727894820272923, + 0.005735764279961586, + 0.0032415450550615788, + 0.004438158590346575, + 0.0039635575376451015, + 0.003718207124620676, + 0.003762470791116357, + 0.0035362562630325556, + 0.003029781160876155, + 0.002590375952422619, + 0.0016086301766335964, + 0.0027082962915301323, + 0.0021883423905819654, + 0.004478783812373877, + 0.002419030526652932, + 0.0036020444240421057, + 0.0034569085109978914, + 0.0019315105164423585, + 0.002108281943947077, + 0.004154989495873451, + 0.0024123401381075382, + 0.00290188891813159, + 0.003769409842789173, + 0.0036971108056604862, + 0.0021123101469129324, + 0.002027945825830102, + 0.0020901509560644627, + 0.003009919775649905, + 0.0018887795740738511, + 0.0030291483271867037, + 0.0031397882848978043, + 0.0037564574740827084, + 0.0031253586057573557, + 0.0054194312542676926, + 0.015098615549504757, + 0.06168336793780327, + 0.04357295110821724, + 0.008481408469378948, + 0.006797533016651869, + 0.007353242952376604, + 0.009138704277575016, + 0.00381468771956861, + 0.0006489023217000067, + 0.0007029972039163113, + 0.000884218723513186, + 0.0009888821514323354, + 0.0007922815857455134, + 0.000523559981957078, + 0.0009073294932022691, + 0.0005258101155050099, + 0.00559436297044158, + 0.005713669117540121, + 0.008350534364581108, + 0.008607422932982445, + 0.005854722112417221, + 0.006164186634123325, + 0.008312870748341084, + 0.009283982217311859 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 7, + 0 + ], + "selected_reward": -0.048662614077329636 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the camera execute a 360-degree panning shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the buildings constructed from luminescent voxels?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the skyscrapers shimmer in neon pink, cyber green, and bright orange colors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are there streams of tiny cars composed of pure white light moving on transparent aerial highways?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the cars leave behind brilliant trails of light as they move?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.012543123215436935 + ], + "opposite_logprob_tokens": [ + -4.387543201446533 + ], + "logprob_reward": -0.012543123215436935, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.012543123215436935, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06504389643669128, + 0.018164506182074547, + 0.3734540045261383, + 0.02302929013967514, + 0.012515024282038212, + 0.016560152173042297, + 0.00833516288548708, + 0.32192501425743103, + 0.016045553609728813, + 0.14492741227149963 + ], + "credit_map": [ + 0.008982541039586067, + 0.010300440713763237, + 0.004456586204469204, + 0.0055373720824718475, + 0.006395808421075344, + 0.02239566296339035, + 0.002846234245225787, + 0.0041292388923466206, + 0.0023026438429951668, + 0.00321764312684536, + 0.001777269528247416, + 0.0015107906656339765, + 0.0019358204444870353, + 0.0024804524146020412, + 0.0013991208979859948, + 0.0035407654941082, + 0.010474632494151592, + 0.1635148674249649, + 0.005358780734241009, + 0.007408728823065758, + 0.16126397252082825, + 0.013058125041425228, + 0.008478211238980293, + 0.003896633628755808, + 0.002240590751171112, + 0.003202892607077956, + 0.0025882599875330925, + 0.0023732236586511135, + 0.005038726143538952, + 0.004492907784879208, + 0.001317168353125453, + 0.001775519223883748, + 0.0011024263221770525, + 0.0018200537888333201, + 0.0014913576887920499, + 0.0017335916636511683, + 0.0012080251472070813, + 0.0019922181963920593, + 0.0019664419814944267, + 0.001200907863676548, + 0.002199582289904356, + 0.0027500397991389036, + 0.0017952336929738522, + 0.0018491762457415462, + 0.002324714558199048, + 0.002261865884065628, + 0.001862307428382337, + 0.0015172300627455115, + 0.0008724294602870941, + 0.0008651891839690506, + 0.0010316120460629463, + 0.0010256280656903982, + 0.0013676260132342577, + 0.0013072998262941837, + 0.0009832667419686913, + 0.0008821109076961875, + 0.027179008349776268, + 0.12572728097438812, + 0.0948352962732315, + 0.01808203011751175, + 0.013670380227267742, + 0.014992928132414818, + 0.018822528421878815, + 0.008615544997155666, + 0.002895868383347988, + 0.002565668663010001, + 0.0015559806488454342, + 0.0020766458474099636, + 0.0021070537623018026, + 0.0021738545037806034, + 0.0013922962825745344, + 0.0012781844707205892, + 0.014072403311729431, + 0.0159449502825737, + 0.022686421871185303, + 0.023577777668833733, + 0.014819059520959854, + 0.01363334059715271, + 0.019110025838017464, + 0.0210834089666605 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 7, + 9 + ], + "selected_reward": -0.012543123215436935 + } + ], + "vlm_decoded_outputs": [ + "No", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "sweeping, sweeping 360-degree panning shot captures the epic birth and evolution of a city core constructed from luminescent voxels. Initially, the frame's center is a deep digital void, with only a few solitary, translucent voxels pulsing in the darkness with a faint electric blue glow, like dormant data seeds. As the camera steadily pans, a breathtaking process of creation unfolds at an exponential rate: these voxels begin to replicate, aggregate, and stack upward, forming complex geometric structures that rise from a barren digital plain. In mere seconds, a futuristic metropolis of skyscrapers, composed of countless voxels shimmering in neon pink, cyber green, and bright orange, takes shape. Its sharp, intricate outlines resemble a giant integrated circuit board. Meanwhile, streams of tiny cars, composed of pure white light, speed across the multi-layered, transparent aerial highway, leaving behind brilliant trails of light and imbuing the self-luminous city with a relentless energy. The entire scene appears magnificent and awe-inspiring under the cool digital light, creating a high-tech, fast-paced, futuristic atmosphere full of infinite possibilities.", + "seed": 4897, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000095/group_1/index_6412_seed_4897_dp_3_rank_6_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 1, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.1951891344040632, + "vlm_logprob": -0.1951891344040632, + "vlm_yes_no_margin": 2.85, + "vlm_token_credit": -0.1951891344040632, + "vlm_yes_ratio": 0.8, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the camera execute a 360-degree panning shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the buildings constructed from luminescent voxels?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the skyscrapers shimmer in neon pink, cyber green, and bright orange colors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are there streams of tiny cars composed of pure white light moving on transparent aerial highways?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the cars leave behind brilliant trails of light as they move?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -0.8333093523979187 + ], + "opposite_logprob_tokens": [ + -0.5833093523979187 + ], + "logprob_reward": -0.8333093523979187, + "yes_no_margin_reward": -0.25, + "token_credit_reward": -0.8333093523979187, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.07239929586648941, + 0.12368649989366531, + 0.059937503188848495, + 0.4262838363647461, + 0.05995609611272812, + 0.01444048248231411, + 0.03725103661417961, + 0.1064002513885498, + 0.010831034742295742, + 0.08881395310163498 + ], + "credit_map": [ + 0.00510935066267848, + 0.013651878573000431, + 0.00746252341195941, + 0.008122335188090801, + 0.010040600784122944, + 0.012246696278452873, + 0.00838762242347002, + 0.007378290873020887, + 0.009946298785507679, + 0.026185913011431694, + 0.017478929832577705, + 0.0130350012332201, + 0.011365287937223911, + 0.023742176592350006, + 0.010145737789571285, + 0.011787166818976402, + 0.006011158227920532, + 0.01143945287913084, + 0.007265651598572731, + 0.006510727107524872, + 0.007100456394255161, + 0.010771848261356354, + 0.006237753666937351, + 0.0046004545874893665, + 0.01359346229583025, + 0.017875492572784424, + 0.009774800390005112, + 0.008306808769702911, + 0.17182594537734985, + 0.17975303530693054, + 0.010729262605309486, + 0.014425020664930344, + 0.003524176776409149, + 0.004287185147404671, + 0.0031296508386731148, + 0.0034645963460206985, + 0.010516470298171043, + 0.018737759441137314, + 0.011644500307738781, + 0.004651757422834635, + 0.0017861571395769715, + 0.002027194481343031, + 0.0010684023145586252, + 0.0010993726318702102, + 0.0023247231729328632, + 0.0019779910799115896, + 0.002199323382228613, + 0.001957318279892206, + 0.0033041185233742, + 0.00515323132276535, + 0.004455395508557558, + 0.0038388276007026434, + 0.0060579716227948666, + 0.008141686208546162, + 0.003213243093341589, + 0.00308656576089561, + 0.007567157968878746, + 0.004357084631919861, + 0.010949750430881977, + 0.009766406379640102, + 0.009781328961253166, + 0.022631345316767693, + 0.03452731668949127, + 0.00681986752897501, + 0.0009049289510585368, + 0.0009245096007362008, + 0.001973819686099887, + 0.0011615047696977854, + 0.0009965193457901478, + 0.0008351869182661176, + 0.0016445759683847427, + 0.0023899897933006287, + 0.007363702170550823, + 0.002748561091721058, + 0.023714685812592506, + 0.010419022291898727, + 0.004620336927473545, + 0.0021109171211719513, + 0.030104830861091614, + 0.007731892168521881 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 1, + 7 + ], + "selected_reward": -0.8333093523979187 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the camera execute a 360-degree panning shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the buildings constructed from luminescent voxels?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the skyscrapers shimmer in neon pink, cyber green, and bright orange colors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are there streams of tiny cars composed of pure white light moving on transparent aerial highways?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the cars leave behind brilliant trails of light as they move?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.055007778108119965 + ], + "opposite_logprob_tokens": [ + -2.9300076961517334 + ], + "logprob_reward": -0.055007778108119965, + "yes_no_margin_reward": 2.875, + "token_credit_reward": -0.055007778108119965, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.19837650656700134, + 0.11337129771709442, + 0.08748181909322739, + 0.038887884467840195, + 0.05888864025473595, + 0.04570256173610687, + 0.04239996522665024, + 0.11094164848327637, + 0.0274057537317276, + 0.27654391527175903 + ], + "credit_map": [ + 0.00923138577491045, + 0.06299205124378204, + 0.017031647264957428, + 0.00986451655626297, + 0.017150182276964188, + 0.05495322495698929, + 0.014637845568358898, + 0.012515654787421227, + 0.009066836908459663, + 0.02246822416782379, + 0.020841488614678383, + 0.007291710004210472, + 0.012637424282729626, + 0.02100973017513752, + 0.009286344982683659, + 0.010769547894597054, + 0.008158722892403603, + 0.013351386412978172, + 0.010423381812870502, + 0.012323247268795967, + 0.017803533002734184, + 0.011091087944805622, + 0.00842548068612814, + 0.005904987920075655, + 0.004541985224932432, + 0.004269700963050127, + 0.002903946442529559, + 0.0036975385155528784, + 0.005508927628397942, + 0.009084583260118961, + 0.0034841103479266167, + 0.005397091619670391, + 0.005474272649735212, + 0.00524474959820509, + 0.004930251277983189, + 0.005017037037760019, + 0.009515766985714436, + 0.01581050455570221, + 0.008647062815725803, + 0.0042489999905228615, + 0.004672084469348192, + 0.00632735388353467, + 0.004272328224033117, + 0.004157004412263632, + 0.008094564080238342, + 0.008321469649672508, + 0.004489295184612274, + 0.005368465557694435, + 0.004463379271328449, + 0.005440144799649715, + 0.005099523812532425, + 0.00443141208961606, + 0.0054002972319722176, + 0.007016906980425119, + 0.005850762594491243, + 0.0046975393779575825, + 0.00843121949583292, + 0.005164419300854206, + 0.01690162904560566, + 0.011530599556863308, + 0.006083579268306494, + 0.013437376357614994, + 0.04284520074725151, + 0.00654761865735054, + 0.002664559055119753, + 0.0031249793246388435, + 0.006159117445349693, + 0.003628578269854188, + 0.002371426671743393, + 0.002115237293764949, + 0.003949563018977642, + 0.003392294282093644, + 0.02267374098300934, + 0.008408540859818459, + 0.07022594660520554, + 0.03283253312110901, + 0.015012189745903015, + 0.006831521634012461, + 0.0948142483830452, + 0.025745196267962456 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 0, + 1 + ], + "selected_reward": -0.055007778108119965 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the camera execute a 360-degree panning shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the buildings constructed from luminescent voxels?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the skyscrapers shimmer in neon pink, cyber green, and bright orange colors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are there streams of tiny cars composed of pure white light moving on transparent aerial highways?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the cars leave behind brilliant trails of light as they move?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005264585837721825 + ], + "opposite_logprob_tokens": [ + -5.255264759063721 + ], + "logprob_reward": -0.005264585837721825, + "yes_no_margin_reward": 5.25, + "token_credit_reward": -0.005264585837721825, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15008273720741272, + 0.06328511983156204, + 0.08120919018983841, + 0.18050517141819, + 0.03237790986895561, + 0.06630022823810577, + 0.17329148948192596, + 0.047607697546482086, + 0.07038318365812302, + 0.13495737314224243 + ], + "credit_map": [ + 0.01133133564144373, + 0.024154528975486755, + 0.016139548271894455, + 0.017748070880770683, + 0.020248623564839363, + 0.022928016260266304, + 0.02333715558052063, + 0.014195441268384457, + 0.005360070150345564, + 0.014704208821058273, + 0.009068367071449757, + 0.006141049321740866, + 0.005700463894754648, + 0.011716624721884727, + 0.004526457749307156, + 0.006067868322134018, + 0.0070991069078445435, + 0.014519107528030872, + 0.0070108454674482346, + 0.0077341715805232525, + 0.014015293680131435, + 0.012828906998038292, + 0.010455789044499397, + 0.007545959670096636, + 0.007140222005546093, + 0.012588097713887691, + 0.0038891059812158346, + 0.00567786768078804, + 0.07567377388477325, + 0.06154258921742439, + 0.005834415555000305, + 0.008159088902175426, + 0.003127844538539648, + 0.0037757379468530416, + 0.002273698104545474, + 0.002485936973243952, + 0.005638017784804106, + 0.00841649528592825, + 0.004357452038675547, + 0.0023027227725833654, + 0.009482675231993198, + 0.00902264192700386, + 0.004409970715641975, + 0.005044310819357634, + 0.00932476855814457, + 0.011865809559822083, + 0.009564300067722797, + 0.007585745770484209, + 0.015284253284335136, + 0.020253093913197517, + 0.020767655223608017, + 0.017270397394895554, + 0.026289112865924835, + 0.039466742426157, + 0.01725492998957634, + 0.016705287620425224, + 0.003734760917723179, + 0.002454624744132161, + 0.008555884473025799, + 0.005182321649044752, + 0.0028849211521446705, + 0.0039964220486581326, + 0.01805400848388672, + 0.0027447484899312258, + 0.004435548093169928, + 0.00885541457682848, + 0.007298363372683525, + 0.006921895779669285, + 0.006033208221197128, + 0.006714080460369587, + 0.009065292775630951, + 0.021059373393654823, + 0.010248033329844475, + 0.006064338143914938, + 0.03218035399913788, + 0.014761362224817276, + 0.007361067458987236, + 0.0049416362307965755, + 0.037631262093782425, + 0.02176930196583271 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 6, + 0 + ], + "selected_reward": -0.005264585837721825 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the camera execute a 360-degree panning shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the buildings constructed from luminescent voxels?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the skyscrapers shimmer in neon pink, cyber green, and bright orange colors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are there streams of tiny cars composed of pure white light moving on transparent aerial highways?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the cars leave behind brilliant trails of light as they move?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04866158962249756 + ], + "opposite_logprob_tokens": [ + -3.048661708831787 + ], + "logprob_reward": -0.04866158962249756, + "yes_no_margin_reward": 3.0, + "token_credit_reward": -0.04866158962249756, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.22980792820453644, + 0.1339428573846817, + 0.14559216797351837, + 0.11923026293516159, + 0.07193969935178757, + 0.05154810845851898, + 0.10534989088773727, + 0.05947573855519295, + 0.009681027382612228, + 0.07343234121799469 + ], + "credit_map": [ + 0.01864091120660305, + 0.04898066818714142, + 0.030666407197713852, + 0.023349732160568237, + 0.02001195028424263, + 0.05285529047250748, + 0.01573910191655159, + 0.019563838839530945, + 0.010265166871249676, + 0.02694464847445488, + 0.026058169081807137, + 0.013089381158351898, + 0.01222177129238844, + 0.023567931726574898, + 0.010551009327173233, + 0.011244765482842922, + 0.012145480141043663, + 0.02284189872443676, + 0.014694899320602417, + 0.017160607501864433, + 0.02989046275615692, + 0.022820904850959778, + 0.01613815873861313, + 0.0098997438326478, + 0.006069589871913195, + 0.010578920133411884, + 0.006268039811402559, + 0.010529305785894394, + 0.03025062195956707, + 0.03657439723610878, + 0.01107199676334858, + 0.007887379266321659, + 0.004274875856935978, + 0.003560989862307906, + 0.0031663253903388977, + 0.0024006229359656572, + 0.017381493002176285, + 0.02600478194653988, + 0.011098334565758705, + 0.004052272532135248, + 0.005849639885127544, + 0.006900399457663298, + 0.00415021600201726, + 0.00515406671911478, + 0.0071253604255616665, + 0.008523973636329174, + 0.0067793140187859535, + 0.0070651317946612835, + 0.008494616486132145, + 0.014318053610622883, + 0.013375749811530113, + 0.010992267169058323, + 0.014327503740787506, + 0.02208591438829899, + 0.011926122009754181, + 0.00982965249568224, + 0.004601024091243744, + 0.002728849882259965, + 0.009341335855424404, + 0.006199549417942762, + 0.003177962964400649, + 0.006757298018783331, + 0.023049768060445786, + 0.003619944676756859, + 0.0009292264003306627, + 0.0011313764844089746, + 0.0016406680224463344, + 0.0012406223686411977, + 0.0008671387913636863, + 0.0011911100009456277, + 0.0014055538922548294, + 0.0012753308983519673, + 0.005225632339715958, + 0.002414329908788204, + 0.014469233341515064, + 0.009256171062588692, + 0.0036553821992129087, + 0.0024100495502352715, + 0.02326643280684948, + 0.012735106982290745 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 1 + ], + "selected_reward": -0.04866158962249756 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the camera execute a 360-degree panning shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the buildings constructed from luminescent voxels?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the skyscrapers shimmer in neon pink, cyber green, and bright orange colors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are there streams of tiny cars composed of pure white light moving on transparent aerial highways?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the cars leave behind brilliant trails of light as they move?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.033702366054058075 + ], + "opposite_logprob_tokens": [ + -3.4087023735046387 + ], + "logprob_reward": -0.033702366054058075, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.033702366054058075, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3354058265686035, + 0.08820860832929611, + 0.029856156557798386, + 0.11396975070238113, + 0.02673123963177204, + 0.06286176294088364, + 0.21114026010036469, + 0.022026920691132545, + 0.028629528358578682, + 0.08116994053125381 + ], + "credit_map": [ + 0.028034934774041176, + 0.06561370939016342, + 0.04583076760172844, + 0.025807829573750496, + 0.048744115978479385, + 0.0685553178191185, + 0.028432004153728485, + 0.024387141689658165, + 0.007524612359702587, + 0.01543230190873146, + 0.014764850959181786, + 0.007078626658767462, + 0.01034393347799778, + 0.016806332394480705, + 0.007564974017441273, + 0.008692975156009197, + 0.0033761709928512573, + 0.006299708504229784, + 0.004101935308426619, + 0.0029748694505542517, + 0.003300965763628483, + 0.004138559103012085, + 0.003240142250433564, + 0.0024238047190010548, + 0.010649251751601696, + 0.01368793472647667, + 0.004398589488118887, + 0.006165048107504845, + 0.01663326472043991, + 0.03848789259791374, + 0.005757548846304417, + 0.01819022372364998, + 0.004292531404644251, + 0.00337916798889637, + 0.0034441358875483274, + 0.0032133760396391153, + 0.003627804107964039, + 0.0030785955023020506, + 0.002843099180608988, + 0.0028525295201689005, + 0.005154648330062628, + 0.005109287332743406, + 0.0055697266943752766, + 0.005950851831585169, + 0.012875547632575035, + 0.012064614333212376, + 0.006837469059973955, + 0.00929961446672678, + 0.022146333009004593, + 0.02375628426671028, + 0.034539949148893356, + 0.028051696717739105, + 0.02126969024538994, + 0.030556656420230865, + 0.02998002991080284, + 0.0208396278321743, + 0.0026495452038943768, + 0.0016162717947736382, + 0.002987189218401909, + 0.002766671823337674, + 0.002947071800008416, + 0.0035810868721455336, + 0.0031574072781950235, + 0.0023216777481138706, + 0.0019255731021985412, + 0.003934424836188555, + 0.005227196961641312, + 0.003959597088396549, + 0.0024283535312861204, + 0.0027631912380456924, + 0.004609354771673679, + 0.0037818357814103365, + 0.006652239244431257, + 0.0034909220412373543, + 0.021450409665703773, + 0.009214429184794426, + 0.0044611054472625256, + 0.003076436696574092, + 0.024728327989578247, + 0.008096064440906048 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 6, + 3 + ], + "selected_reward": -0.033702366054058075 + } + ], + "vlm_decoded_outputs": [ + "No", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "sweeping, sweeping 360-degree panning shot captures the epic birth and evolution of a city core constructed from luminescent voxels. Initially, the frame's center is a deep digital void, with only a few solitary, translucent voxels pulsing in the darkness with a faint electric blue glow, like dormant data seeds. As the camera steadily pans, a breathtaking process of creation unfolds at an exponential rate: these voxels begin to replicate, aggregate, and stack upward, forming complex geometric structures that rise from a barren digital plain. In mere seconds, a futuristic metropolis of skyscrapers, composed of countless voxels shimmering in neon pink, cyber green, and bright orange, takes shape. Its sharp, intricate outlines resemble a giant integrated circuit board. Meanwhile, streams of tiny cars, composed of pure white light, speed across the multi-layered, transparent aerial highway, leaving behind brilliant trails of light and imbuing the self-luminous city with a relentless energy. The entire scene appears magnificent and awe-inspiring under the cool digital light, creating a high-tech, fast-paced, futuristic atmosphere full of infinite possibilities.", + "seed": 4898, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000095/group_1/index_6412_seed_4898_dp_3_rank_6_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 1, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.300265933573246, + "vlm_logprob": -0.300265933573246, + "vlm_yes_no_margin": 3.0, + "vlm_token_credit": -0.300265933573246, + "vlm_yes_ratio": 0.8, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the camera execute a 360-degree panning shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the buildings constructed from luminescent voxels?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the skyscrapers shimmer in neon pink, cyber green, and bright orange colors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are there streams of tiny cars composed of pure white light moving on transparent aerial highways?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the cars leave behind brilliant trails of light as they move?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -1.410034418106079 + ], + "opposite_logprob_tokens": [ + -0.2850343883037567 + ], + "logprob_reward": -1.410034418106079, + "yes_no_margin_reward": -1.125, + "token_credit_reward": -1.410034418106079, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.18201352655887604, + 0.04824450984597206, + 0.10520713776350021, + 0.07387139648199081, + 0.047149885445833206, + 0.4236350953578949, + 0.04197030887007713, + 0.04529650881886482, + 0.0163242407143116, + 0.01628732681274414 + ], + "credit_map": [ + 0.017721859738230705, + 0.023215413093566895, + 0.025138692930340767, + 0.03146882727742195, + 0.01518249325454235, + 0.025982100516557693, + 0.02656499855220318, + 0.016739148646593094, + 0.003971538972109556, + 0.006775977555662394, + 0.0054220035672187805, + 0.008257735520601273, + 0.00543324975296855, + 0.0077928537502884865, + 0.006614329293370247, + 0.003976824227720499, + 0.010197220370173454, + 0.011936522088944912, + 0.014101962558925152, + 0.014776667580008507, + 0.012532513588666916, + 0.015454404056072235, + 0.011295015923678875, + 0.014912834390997887, + 0.01305576041340828, + 0.007575191557407379, + 0.00987012404948473, + 0.006775738671422005, + 0.010728789493441582, + 0.010894669219851494, + 0.007250552996993065, + 0.007720578461885452, + 0.008949455805122852, + 0.007198588456958532, + 0.005684433504939079, + 0.0033635159488767385, + 0.006707172840833664, + 0.005155004095286131, + 0.005361293442547321, + 0.004730423912405968, + 0.16601760685443878, + 0.16576479375362396, + 0.016360975801944733, + 0.015626197680830956, + 0.015931278467178345, + 0.018348054960370064, + 0.009488873183727264, + 0.01609734445810318, + 0.005446193739771843, + 0.0060812728479504585, + 0.005752606317400932, + 0.004282939713448286, + 0.006013678852468729, + 0.004773893393576145, + 0.00525146909058094, + 0.004368256777524948, + 0.00669630104675889, + 0.00697097135707736, + 0.004988064058125019, + 0.005774336867034435, + 0.007070580497384071, + 0.004687021020799875, + 0.005434996448457241, + 0.003674239618703723, + 0.0016147141577675939, + 0.0016171266324818134, + 0.0013954504393041134, + 0.00223241257481277, + 0.0012624087976291776, + 0.0014112138887867332, + 0.003913144115358591, + 0.00287777092307806, + 0.0016309592174366117, + 0.002255387604236603, + 0.001306163496337831, + 0.002116104820743203, + 0.0023236582055687904, + 0.002640714170411229, + 0.0026206232141703367, + 0.0013937177136540413 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 0, + 2 + ], + "selected_reward": -1.410034418106079 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the camera execute a 360-degree panning shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the buildings constructed from luminescent voxels?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the skyscrapers shimmer in neon pink, cyber green, and bright orange colors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are there streams of tiny cars composed of pure white light moving on transparent aerial highways?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the cars leave behind brilliant trails of light as they move?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0206122025847435 + ], + "opposite_logprob_tokens": [ + -3.8956122398376465 + ], + "logprob_reward": -0.0206122025847435, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.0206122025847435, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14559459686279297, + 0.0649830773472786, + 0.05046197772026062, + 0.09862475842237473, + 0.29699891805648804, + 0.09301567822694778, + 0.09339369088411331, + 0.037463296204805374, + 0.06227526068687439, + 0.057188719511032104 + ], + "credit_map": [ + 0.012402551248669624, + 0.020464051514863968, + 0.02106356807053089, + 0.03191409632563591, + 0.010042231529951096, + 0.01655675657093525, + 0.023204106837511063, + 0.009947230108082294, + 0.0054734633304178715, + 0.009293011389672756, + 0.007952391169965267, + 0.009805569425225258, + 0.009000808000564575, + 0.010218078270554543, + 0.008515805937349796, + 0.004723949357867241, + 0.0056519778445363045, + 0.005371322389692068, + 0.0052172718569636345, + 0.00771025475114584, + 0.004998120944947004, + 0.006653850898146629, + 0.01047421619296074, + 0.004384962841868401, + 0.02035757713019848, + 0.012328565120697021, + 0.013974583707749844, + 0.011044135317206383, + 0.011008432134985924, + 0.010311665944755077, + 0.009773856960237026, + 0.009825937449932098, + 0.05368998274207115, + 0.0498814582824707, + 0.02787482924759388, + 0.019829226657748222, + 0.049560509622097015, + 0.04412464052438736, + 0.023318424820899963, + 0.02871984802186489, + 0.02116321586072445, + 0.021763602271676064, + 0.008641774766147137, + 0.006197027396410704, + 0.013183691538870335, + 0.009521515108644962, + 0.005210389383137226, + 0.007334460038691759, + 0.01718309335410595, + 0.01358465664088726, + 0.010517111048102379, + 0.0102513013407588, + 0.011930607259273529, + 0.010611450299620628, + 0.013001612387597561, + 0.006313859019428492, + 0.005407941993325949, + 0.00623238505795598, + 0.003922283183783293, + 0.005280893296003342, + 0.004258437547832727, + 0.004020209424197674, + 0.005605940241366625, + 0.0027352040633559227, + 0.007424543146044016, + 0.010155239142477512, + 0.005489132832735777, + 0.0071942913345992565, + 0.006781211122870445, + 0.005415268242359161, + 0.011200768873095512, + 0.008614804595708847, + 0.005148242227733135, + 0.007138359826058149, + 0.003573850728571415, + 0.006372029893100262, + 0.004961080849170685, + 0.008237489499151707, + 0.014293093234300613, + 0.00746456952765584 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 0, + 3 + ], + "selected_reward": -0.0206122025847435 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the camera execute a 360-degree panning shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the buildings constructed from luminescent voxels?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the skyscrapers shimmer in neon pink, cyber green, and bright orange colors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are there streams of tiny cars composed of pure white light moving on transparent aerial highways?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the cars leave behind brilliant trails of light as they move?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005959954112768173 + ], + "opposite_logprob_tokens": [ + -5.130959987640381 + ], + "logprob_reward": -0.005959954112768173, + "yes_no_margin_reward": 5.125, + "token_credit_reward": -0.005959954112768173, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1501902937889099, + 0.058113496750593185, + 0.05915738269686699, + 0.08365271240472794, + 0.10029813647270203, + 0.16564279794692993, + 0.028121288865804672, + 0.04451686143875122, + 0.045984022319316864, + 0.26432302594184875 + ], + "credit_map": [ + 0.015942808240652084, + 0.02029011771082878, + 0.02171100489795208, + 0.021137703210115433, + 0.013697221875190735, + 0.019681328907608986, + 0.02394045703113079, + 0.013789629563689232, + 0.005505514331161976, + 0.006654592230916023, + 0.006376202683895826, + 0.01056523248553276, + 0.008069035597145557, + 0.008829284459352493, + 0.008332710713148117, + 0.003780917963013053, + 0.0061600226908922195, + 0.007125393953174353, + 0.006041817832738161, + 0.01059902273118496, + 0.006184016820043325, + 0.007298233453184366, + 0.007281480357050896, + 0.008467387408018112, + 0.015804776921868324, + 0.01148425042629242, + 0.008512123487889767, + 0.005823218729346991, + 0.014955333434045315, + 0.012550522573292255, + 0.007864110171794891, + 0.006658367346972227, + 0.016349133104085922, + 0.013860694132745266, + 0.013193267397582531, + 0.009740395471453667, + 0.01751868613064289, + 0.01072008814662695, + 0.012238841503858566, + 0.0066770282573997974, + 0.05046689882874489, + 0.0502818264067173, + 0.014447874389588833, + 0.012247158214449883, + 0.009263773448765278, + 0.00916790496557951, + 0.007196258753538132, + 0.012571088969707489, + 0.008100172504782677, + 0.0036494089290499687, + 0.001985988114029169, + 0.002280190587043762, + 0.0024540284648537636, + 0.004136364441365004, + 0.002155813854187727, + 0.003359320340678096, + 0.006474142894148827, + 0.008223402313888073, + 0.0050753322429955006, + 0.004762685392051935, + 0.006442930083721876, + 0.005351123865693808, + 0.005080845206975937, + 0.003106396645307541, + 0.004495081026107073, + 0.0038823930080980062, + 0.005159176420420408, + 0.006786744575947523, + 0.003583085024729371, + 0.00315410690382123, + 0.011039895936846733, + 0.007883534766733646, + 0.013063588179647923, + 0.009477226994931698, + 0.007772246841341257, + 0.03423720970749855, + 0.03149544447660446, + 0.061175424605607986, + 0.07751299440860748, + 0.02958890236914158 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 5, + 0 + ], + "selected_reward": -0.005959954112768173 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the camera execute a 360-degree panning shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the buildings constructed from luminescent voxels?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the skyscrapers shimmer in neon pink, cyber green, and bright orange colors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are there streams of tiny cars composed of pure white light moving on transparent aerial highways?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the cars leave behind brilliant trails of light as they move?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.048656366765499115 + ], + "opposite_logprob_tokens": [ + -3.048656463623047 + ], + "logprob_reward": -0.048656366765499115, + "yes_no_margin_reward": 3.0, + "token_credit_reward": -0.048656366765499115, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.34936484694480896, + 0.12946242094039917, + 0.08607654273509979, + 0.08234213292598724, + 0.10344353318214417, + 0.12338719516992569, + 0.013616069220006466, + 0.026462683454155922, + 0.021679416298866272, + 0.0641651526093483 + ], + "credit_map": [ + 0.02372836135327816, + 0.08943867683410645, + 0.027844129130244255, + 0.06354855746030807, + 0.024525485932826996, + 0.07069677859544754, + 0.02853189967572689, + 0.02105097845196724, + 0.012575640343129635, + 0.014692614786326885, + 0.01237158291041851, + 0.01725138910114765, + 0.025175603106617928, + 0.01848602667450905, + 0.019423561170697212, + 0.009486008435487747, + 0.007348382379859686, + 0.011117498390376568, + 0.006669909227639437, + 0.014137806370854378, + 0.012790380045771599, + 0.008306601084768772, + 0.017750201746821404, + 0.007955770008265972, + 0.014934414997696877, + 0.01206236518919468, + 0.010577855631709099, + 0.009643751196563244, + 0.009275655262172222, + 0.009630989283323288, + 0.008358667604625225, + 0.007858443073928356, + 0.021243955940008163, + 0.014851483516395092, + 0.011865753680467606, + 0.0070822457782924175, + 0.017232414335012436, + 0.01226283609867096, + 0.010229541920125484, + 0.008675305172801018, + 0.048102736473083496, + 0.048851367086172104, + 0.0045557282865047455, + 0.00480114808306098, + 0.004964811261743307, + 0.005244535394012928, + 0.00287999352440238, + 0.003986879251897335, + 0.0017685735365375876, + 0.0014649860095232725, + 0.001895706169307232, + 0.001980900065973401, + 0.0016468206886202097, + 0.001877717673778534, + 0.0016451786505058408, + 0.0013361875899136066, + 0.003751120762899518, + 0.003241426544263959, + 0.0028716977685689926, + 0.003586788661777973, + 0.0035063810646533966, + 0.0035548533778637648, + 0.0041775330901145935, + 0.0017728834645822644, + 0.002091320464387536, + 0.002300748834386468, + 0.003142992500215769, + 0.0025063164066523314, + 0.002195472363382578, + 0.0019590198062360287, + 0.004324262496083975, + 0.0031592852901667356, + 0.005727454088628292, + 0.00516495993360877, + 0.006080925464630127, + 0.00870014913380146, + 0.0054797339253127575, + 0.00969669595360756, + 0.014243052341043949, + 0.009072191081941128 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 5 + ], + "selected_reward": -0.048656366765499115 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the camera execute a 360-degree panning shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the buildings constructed from luminescent voxels?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the skyscrapers shimmer in neon pink, cyber green, and bright orange colors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are there streams of tiny cars composed of pure white light moving on transparent aerial highways?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the cars leave behind brilliant trails of light as they move?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01606672629714012 + ], + "opposite_logprob_tokens": [ + -4.141066551208496 + ], + "logprob_reward": -0.01606672629714012, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.01606672629714012, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.24675698578357697, + 0.09646438807249069, + 0.05227501317858696, + 0.062212876975536346, + 0.07003510743379593, + 0.28616201877593994, + 0.02470557391643524, + 0.01938544772565365, + 0.07281915098428726, + 0.06918343901634216 + ], + "credit_map": [ + 0.015924254432320595, + 0.059290364384651184, + 0.028984244912862778, + 0.03705867752432823, + 0.018302960321307182, + 0.05127948150038719, + 0.02195759303867817, + 0.01395940687507391, + 0.00828865822404623, + 0.01195089053362608, + 0.011033160611987114, + 0.017324794083833694, + 0.01146752666682005, + 0.014217637479305267, + 0.01451462134718895, + 0.00766709866002202, + 0.005119113251566887, + 0.006705497391521931, + 0.00480771204456687, + 0.009201208129525185, + 0.008021757006645203, + 0.005359871778637171, + 0.0099370451644063, + 0.0031228105071932077, + 0.010229323990643024, + 0.009000773541629314, + 0.010822583921253681, + 0.008609921671450138, + 0.007971697486937046, + 0.0050286343321204185, + 0.005951013416051865, + 0.004598925355821848, + 0.012249455787241459, + 0.009164143353700638, + 0.007631789892911911, + 0.004422447178512812, + 0.01319750864058733, + 0.009660367853939533, + 0.007491546683013439, + 0.006217847112566233, + 0.11377643793821335, + 0.11422669887542725, + 0.009710405021905899, + 0.010639536194503307, + 0.011090232990682125, + 0.01113902311772108, + 0.006177978590130806, + 0.00940172653645277, + 0.0029042288661003113, + 0.0027966005727648735, + 0.0036231493577361107, + 0.0030982326716184616, + 0.003358757123351097, + 0.0030724231619387865, + 0.0029967371374368668, + 0.0028554436285048723, + 0.0023272072430700064, + 0.002777118468657136, + 0.0023084513377398252, + 0.002962726866826415, + 0.0026973295025527477, + 0.0020474959164857864, + 0.0027504265308380127, + 0.001514692441560328, + 0.006534687243402004, + 0.004449654370546341, + 0.0092050451785326, + 0.013780947774648666, + 0.005333269014954567, + 0.004607279784977436, + 0.0161861851811409, + 0.012722081504762173, + 0.00703662121668458, + 0.008966583758592606, + 0.007957572117447853, + 0.009717638604342937, + 0.006893368903547525, + 0.008796267211437225, + 0.0077582155354321, + 0.012057172134518623 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 0, + 1 + ], + "selected_reward": -0.01606672629714012 + } + ], + "vlm_decoded_outputs": [ + "No", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000095/group_10/index_tensor([914, 914])_group_10_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000095/group_10/index_tensor([914, 914])_group_10_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..8daf1766883a9c1950721f6f41bf4acf2fbdcef9 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000095/group_10/index_tensor([914, 914])_group_10_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "In the style of a sleek cyberpunk animation, a static, slightly low-angled, extreme close-up meticulously depicts a complex robotic arm on a futuristic automotive assembly line. Its surface, a blend of polished gunmetal and matte carbon fiber, gleams under the chilly factory ambient lighting. Suddenly, as a malfunction triggers the fire suppression system, supercooled liquid cascades down, and shimmering droplets can be seen streaking across the frame. The moment freezing rain strikes the arm, a delicate and swift web of crystalline ice blooms across its intricate hydraulic joints, frost quickly enveloping the vibrant cyan glow emanating from the pressurized tubes. The arm's fluid, deliberate movements are interspersed with violent, shuddering tremors, its motion halting mid-task. Immediately, a bright, chaotic electrical spark erupts from the now-frozen joint, arcing erratically between the frozen metal plates. This dazzling, frenzied yellow and white light contrasts sharply with the scene's cool, dark color palette, creating a tense and dramatic atmosphere of catastrophic technological failure against the soft-focus backdrop of the vast, automated factory floor.", + "seed": 8274, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000095/group_10/index_914_seed_8274_dp_20_rank_40_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 10, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.021726315631531178, + "vlm_logprob": -0.021726315631531178, + "vlm_yes_no_margin": 4.3, + "vlm_token_credit": -0.021726315631531178, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a complex robotic arm located on a futuristic automotive assembly line?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the robotic arm's surface show a blend of polished gunmetal and matte carbon fiber materials?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the video show supercooled liquid cascading down and crystalline ice blooming across the arm's joints?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is there a bright, chaotic yellow and white electrical spark arcing from a frozen joint?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera positioned in a static, slightly low-angled extreme close-up shot?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03833687677979469 + ], + "opposite_logprob_tokens": [ + -3.413336992263794 + ], + "logprob_reward": -0.03833687677979469, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.03833687677979469, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10377020388841629, + 0.07088371366262436, + 0.08032922446727753, + 0.06315454095602036, + 0.16982366144657135, + 0.04055223986506462, + 0.11267831921577454, + 0.016465071588754654, + 0.30939772725105286, + 0.032945286482572556 + ], + "credit_map": [ + 0.014904441311955452, + 0.015227547846734524, + 0.014350274577736855, + 0.02017947845160961, + 0.011075366288423538, + 0.008388560265302658, + 0.009957141242921352, + 0.009687392972409725, + 0.009880601428449154, + 0.008215184323489666, + 0.01016621571034193, + 0.012545482255518436, + 0.009474121034145355, + 0.00911414809525013, + 0.0055958861485123634, + 0.0058920737355947495, + 0.008027096278965473, + 0.010194549337029457, + 0.007523783948272467, + 0.013197910971939564, + 0.009131906554102898, + 0.013159060850739479, + 0.009664985351264477, + 0.009429930709302425, + 0.008073003962635994, + 0.00716662360355258, + 0.010892212390899658, + 0.006824662443250418, + 0.006943491753190756, + 0.007738292217254639, + 0.0058029224164783955, + 0.009713332168757915, + 0.02523166313767433, + 0.032915402203798294, + 0.022417664527893066, + 0.01551017351448536, + 0.02128812111914158, + 0.01697514019906521, + 0.013618574477732182, + 0.021866921335458755, + 0.0048378389328718185, + 0.0048677995800971985, + 0.005378898233175278, + 0.004240382928401232, + 0.005682524759322405, + 0.0051332018338143826, + 0.003233778988942504, + 0.007177816703915596, + 0.015065709128975868, + 0.011728073470294476, + 0.013619320467114449, + 0.014330407604575157, + 0.016181321814656258, + 0.014796392060816288, + 0.018644848838448524, + 0.008312249556183815, + 0.001450856332667172, + 0.0019920156337320805, + 0.001132196863181889, + 0.001734054763801396, + 0.001981999957934022, + 0.0035361330956220627, + 0.001439804444089532, + 0.003198010614141822, + 0.030536804348230362, + 0.1324390023946762, + 0.010230731219053268, + 0.009063884615898132, + 0.07638860493898392, + 0.0313282236456871, + 0.009101653471589088, + 0.010308821685612202, + 0.004201869480311871, + 0.00392988370731473, + 0.0028183874674141407, + 0.0036513276863843203, + 0.004559624474495649, + 0.004352078773081303, + 0.004906017333269119, + 0.004526092205196619 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 4, + 6 + ], + "selected_reward": -0.03833687677979469 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a complex robotic arm located on a futuristic automotive assembly line?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the robotic arm's surface show a blend of polished gunmetal and matte carbon fiber materials?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the video show supercooled liquid cascading down and crystalline ice blooming across the arm's joints?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is there a bright, chaotic yellow and white electrical spark arcing from a frozen joint?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera positioned in a static, slightly low-angled extreme close-up shot?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0430825836956501 + ], + "opposite_logprob_tokens": [ + -3.1680824756622314 + ], + "logprob_reward": -0.0430825836956501, + "yes_no_margin_reward": 3.125, + "token_credit_reward": -0.0430825836956501, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12682446837425232, + 0.05690290406346321, + 0.06935767829418182, + 0.08787994831800461, + 0.08011677861213684, + 0.05410744622349739, + 0.24149592220783234, + 0.05439259484410286, + 0.17125029861927032, + 0.057671960443258286 + ], + "credit_map": [ + 0.015886476263403893, + 0.010558811947703362, + 0.02089729718863964, + 0.022429144009947777, + 0.012396201491355896, + 0.015095720067620277, + 0.016211632639169693, + 0.013349177315831184, + 0.007067715283483267, + 0.006514242384582758, + 0.00643542967736721, + 0.007036061491817236, + 0.008542537689208984, + 0.007282665930688381, + 0.005129043944180012, + 0.008895207196474075, + 0.006857187021523714, + 0.010121033526957035, + 0.006035801954567432, + 0.010064936242997646, + 0.007569786626845598, + 0.012849090620875359, + 0.007889709435403347, + 0.007970135658979416, + 0.012771440669894218, + 0.01392445806413889, + 0.019757358357310295, + 0.009920448064804077, + 0.007921210490167141, + 0.006554654333740473, + 0.006087409798055887, + 0.010942966677248478, + 0.012381407432258129, + 0.009115814231336117, + 0.009090260602533817, + 0.008297053165733814, + 0.011379379779100418, + 0.010455196723341942, + 0.007935711182653904, + 0.011461952701210976, + 0.007327285595238209, + 0.007229767739772797, + 0.005391035228967667, + 0.0051491656340658665, + 0.008044171147048473, + 0.00632115826010704, + 0.006457305978983641, + 0.008187558501958847, + 0.036851268261671066, + 0.02935403399169445, + 0.019375648349523544, + 0.031131044030189514, + 0.03730274736881256, + 0.02686765231192112, + 0.04145706444978714, + 0.019156470894813538, + 0.003638559253886342, + 0.008115335367619991, + 0.007839014753699303, + 0.006864591035991907, + 0.007300776895135641, + 0.007927287369966507, + 0.005824692081660032, + 0.006882335990667343, + 0.01946941204369068, + 0.06702463328838348, + 0.007638962473720312, + 0.0062379674054682255, + 0.036007679998874664, + 0.017681390047073364, + 0.0051431250758469105, + 0.012047119438648224, + 0.007753309328109026, + 0.006392918061465025, + 0.004979553632438183, + 0.006024077534675598, + 0.00739288842305541, + 0.008220282383263111, + 0.011037302203476429, + 0.005871627479791641 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 8, + 0 + ], + "selected_reward": -0.0430825836956501 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a complex robotic arm located on a futuristic automotive assembly line?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the robotic arm's surface show a blend of polished gunmetal and matte carbon fiber materials?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the video show supercooled liquid cascading down and crystalline ice blooming across the arm's joints?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is there a bright, chaotic yellow and white electrical spark arcing from a frozen joint?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera positioned in a static, slightly low-angled extreme close-up shot?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.011080400086939335 + ], + "opposite_logprob_tokens": [ + -4.511080265045166 + ], + "logprob_reward": -0.011080400086939335, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.011080400086939335, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.039345886558294296, + 0.0390845388174057, + 0.03135441616177559, + 0.03054480068385601, + 0.45216113328933716, + 0.02254560962319374, + 0.09248597919940948, + 0.01965220458805561, + 0.24866215884685516, + 0.024163266643881798 + ], + "credit_map": [ + 0.00577932083979249, + 0.002821803791448474, + 0.006678421515971422, + 0.004726191516965628, + 0.004248833749443293, + 0.007023205049335957, + 0.0042524198070168495, + 0.0038156951777637005, + 0.005547499284148216, + 0.004190858919173479, + 0.004524947144091129, + 0.004335555713623762, + 0.003814150346443057, + 0.006492354441434145, + 0.0036230485420674086, + 0.006556128617376089, + 0.004031793214380741, + 0.005855024326592684, + 0.0028251688927412033, + 0.004750597756356001, + 0.0040956842713057995, + 0.003162810578942299, + 0.0037378817796707153, + 0.0028954583685845137, + 0.00539680989459157, + 0.0032635151874274015, + 0.005717166233807802, + 0.0034994587767869234, + 0.002306909067556262, + 0.0031891504768282175, + 0.0033138631843030453, + 0.0038579311221837997, + 0.061690859496593475, + 0.07695025950670242, + 0.06660464406013489, + 0.044524967670440674, + 0.05781681835651398, + 0.04905787110328674, + 0.041707973927259445, + 0.05380774289369583, + 0.0029458566568791866, + 0.0025821984745562077, + 0.0034903930500149727, + 0.0020844056271016598, + 0.0028733296785503626, + 0.0025380542501807213, + 0.0024289204739034176, + 0.0036024521104991436, + 0.01441718265414238, + 0.011296821758151054, + 0.008333290927112103, + 0.011236901395022869, + 0.012251072563230991, + 0.007330169435590506, + 0.019535042345523834, + 0.00808549951761961, + 0.0013567098649218678, + 0.0026142005808651447, + 0.0033671699929982424, + 0.0021499176509678364, + 0.0032797311432659626, + 0.0028761152643710375, + 0.002037034835666418, + 0.001971325371414423, + 0.028901204466819763, + 0.10474297404289246, + 0.007899933494627476, + 0.007399026770144701, + 0.06085620075464249, + 0.023892343044281006, + 0.006813657470047474, + 0.008156838826835155, + 0.0036966342013329268, + 0.0036746531259268522, + 0.0023414078168570995, + 0.0021915659308433533, + 0.00297087081708014, + 0.002915753051638603, + 0.004081807564944029, + 0.002290574600920081 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 8, + 6 + ], + "selected_reward": -0.011080400086939335 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a complex robotic arm located on a futuristic automotive assembly line?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the robotic arm's surface show a blend of polished gunmetal and matte carbon fiber materials?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the video show supercooled liquid cascading down and crystalline ice blooming across the arm's joints?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is there a bright, chaotic yellow and white electrical spark arcing from a frozen joint?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera positioned in a static, slightly low-angled extreme close-up shot?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.001942058908753097 + ], + "opposite_logprob_tokens": [ + -6.251942157745361 + ], + "logprob_reward": -0.001942058908753097, + "yes_no_margin_reward": 6.25, + "token_credit_reward": -0.001942058908753097, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07525268197059631, + 0.07740642130374908, + 0.04909667372703552, + 0.0697939470410347, + 0.11373596638441086, + 0.039706092327833176, + 0.06805648654699326, + 0.06335440278053284, + 0.31339335441589355, + 0.1302039474248886 + ], + "credit_map": [ + 0.00740477629005909, + 0.010517967864871025, + 0.009044111706316471, + 0.013451403006911278, + 0.007764090783894062, + 0.008403562940657139, + 0.009139063768088818, + 0.009527708403766155, + 0.010661536827683449, + 0.006873969454318285, + 0.009214362129569054, + 0.008910922333598137, + 0.007491706404834986, + 0.010001340880990028, + 0.00765861663967371, + 0.016593964770436287, + 0.007057768292725086, + 0.006293819285929203, + 0.004326120484620333, + 0.006380584556609392, + 0.005727720446884632, + 0.007035319693386555, + 0.0076329512521624565, + 0.004642386920750141, + 0.007857332937419415, + 0.007881661877036095, + 0.012364201247692108, + 0.0073516699485480785, + 0.007065558340400457, + 0.007121704984456301, + 0.007327021099627018, + 0.012824795208871365, + 0.016618764027953148, + 0.02976205013692379, + 0.011239022947847843, + 0.006886471528559923, + 0.013342393562197685, + 0.008967896923422813, + 0.007558673154562712, + 0.019360700622200966, + 0.006889671552926302, + 0.005861843470484018, + 0.003443874651566148, + 0.004117421340197325, + 0.004841633141040802, + 0.005876752082258463, + 0.002543538575991988, + 0.006131354719400406, + 0.010881505906581879, + 0.008023821748793125, + 0.008909816853702068, + 0.008933247067034245, + 0.013489934615790844, + 0.008491805754601955, + 0.005658917129039764, + 0.0036674384027719498, + 0.005202391650527716, + 0.010320368222892284, + 0.007264068815857172, + 0.0057256026193499565, + 0.00962807983160019, + 0.010610352270305157, + 0.0051783849485218525, + 0.009425148367881775, + 0.019464852288365364, + 0.10640911012887955, + 0.01754341647028923, + 0.012872575782239437, + 0.09415096789598465, + 0.027046222239732742, + 0.013221891596913338, + 0.022684315219521523, + 0.01274458784610033, + 0.017288336530327797, + 0.01861151121556759, + 0.01165863685309887, + 0.01931692846119404, + 0.021702295169234276, + 0.013443319126963615, + 0.01543833501636982 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 9, + 4 + ], + "selected_reward": -0.001942058908753097 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a complex robotic arm located on a futuristic automotive assembly line?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the robotic arm's surface show a blend of polished gunmetal and matte carbon fiber materials?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the video show supercooled liquid cascading down and crystalline ice blooming across the arm's joints?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is there a bright, chaotic yellow and white electrical spark arcing from a frozen joint?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera positioned in a static, slightly low-angled extreme close-up shot?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.014189658686518669 + ], + "opposite_logprob_tokens": [ + -4.264189720153809 + ], + "logprob_reward": -0.014189658686518669, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.014189658686518669, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09340130537748337, + 0.04612413048744202, + 0.0691017434000969, + 0.08986907452344894, + 0.11325094103813171, + 0.03562270849943161, + 0.12143304198980331, + 0.06769940257072449, + 0.3153582215309143, + 0.048139430582523346 + ], + "credit_map": [ + 0.012501770630478859, + 0.011757295578718185, + 0.015914548188447952, + 0.016131579875946045, + 0.007815791293978691, + 0.00994050595909357, + 0.010220568627119064, + 0.009119240567088127, + 0.007844696752727032, + 0.005557373631745577, + 0.006983980070799589, + 0.007194632198661566, + 0.004599968437105417, + 0.006049579940736294, + 0.004236537031829357, + 0.0036573612596839666, + 0.008389841765165329, + 0.008968107402324677, + 0.006161231081932783, + 0.010455398820340633, + 0.007349981460720301, + 0.01071518287062645, + 0.009031512774527073, + 0.008030489087104797, + 0.011160115711390972, + 0.011015123687684536, + 0.015573081560432911, + 0.009276417084038258, + 0.009592557325959206, + 0.011077315546572208, + 0.00930924154818058, + 0.012865226715803146, + 0.01628844439983368, + 0.010484280064702034, + 0.01459681149572134, + 0.011097881011664867, + 0.019465243443846703, + 0.013685764744877815, + 0.012768814340233803, + 0.01486370712518692, + 0.0049543981440365314, + 0.005831560585647821, + 0.004739250987768173, + 0.0036118088755756617, + 0.004685951396822929, + 0.003608875209465623, + 0.002534209517762065, + 0.0056566535495221615, + 0.018510987982153893, + 0.01451367512345314, + 0.01157868281006813, + 0.014001562260091305, + 0.018206363543868065, + 0.011237533763051033, + 0.023482559248805046, + 0.009901679120957851, + 0.006163226906210184, + 0.00915179867297411, + 0.010279460810124874, + 0.007275171112269163, + 0.011054670438170433, + 0.00893824826925993, + 0.00591250928118825, + 0.00892432127147913, + 0.034923579543828964, + 0.13634616136550903, + 0.009835151955485344, + 0.00856769923120737, + 0.07358141988515854, + 0.03147776052355766, + 0.009435026906430721, + 0.01119141560047865, + 0.006773019675165415, + 0.0067976959981024265, + 0.005844467785209417, + 0.006094445940107107, + 0.005537501536309719, + 0.00426795007660985, + 0.008457537740468979, + 0.0043668122962117195 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 6, + 4 + ], + "selected_reward": -0.014189658686518669 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "In the style of a sleek cyberpunk animation, a static, slightly low-angled, extreme close-up meticulously depicts a complex robotic arm on a futuristic automotive assembly line. Its surface, a blend of polished gunmetal and matte carbon fiber, gleams under the chilly factory ambient lighting. Suddenly, as a malfunction triggers the fire suppression system, supercooled liquid cascades down, and shimmering droplets can be seen streaking across the frame. The moment freezing rain strikes the arm, a delicate and swift web of crystalline ice blooms across its intricate hydraulic joints, frost quickly enveloping the vibrant cyan glow emanating from the pressurized tubes. The arm's fluid, deliberate movements are interspersed with violent, shuddering tremors, its motion halting mid-task. Immediately, a bright, chaotic electrical spark erupts from the now-frozen joint, arcing erratically between the frozen metal plates. This dazzling, frenzied yellow and white light contrasts sharply with the scene's cool, dark color palette, creating a tense and dramatic atmosphere of catastrophic technological failure against the soft-focus backdrop of the vast, automated factory floor.", + "seed": 8275, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000095/group_10/index_914_seed_8275_dp_20_rank_40_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 10, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.028457785584032537, + "vlm_logprob": -0.028457785584032537, + "vlm_yes_no_margin": 3.75, + "vlm_token_credit": -0.028457785584032537, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a complex robotic arm located on a futuristic automotive assembly line?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the robotic arm's surface show a blend of polished gunmetal and matte carbon fiber materials?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the video show supercooled liquid cascading down and crystalline ice blooming across the arm's joints?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is there a bright, chaotic yellow and white electrical spark arcing from a frozen joint?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera positioned in a static, slightly low-angled extreme close-up shot?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.05276527255773544 + ], + "opposite_logprob_tokens": [ + -3.052765369415283 + ], + "logprob_reward": -0.05276527255773544, + "yes_no_margin_reward": 3.0, + "token_credit_reward": -0.05276527255773544, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1344119906425476, + 0.06389530003070831, + 0.04965360462665558, + 0.06546405702829361, + 0.06838668882846832, + 0.09397300332784653, + 0.2531055510044098, + 0.031206205487251282, + 0.06890225410461426, + 0.17100131511688232 + ], + "credit_map": [ + 0.024144429713487625, + 0.015170597471296787, + 0.021604113280773163, + 0.014334852807223797, + 0.018199903890490532, + 0.012278642505407333, + 0.012327688746154308, + 0.016351761296391487, + 0.009602821432054043, + 0.008750922046601772, + 0.006848437245935202, + 0.006110031623393297, + 0.007223229855298996, + 0.009833255782723427, + 0.00745607353746891, + 0.008070527575910091, + 0.007588498760014772, + 0.005977680440992117, + 0.005159496795386076, + 0.004485283512622118, + 0.009173247031867504, + 0.005768627859652042, + 0.005683769937604666, + 0.005816998891532421, + 0.005498216487467289, + 0.014108464121818542, + 0.008654181845486164, + 0.008163992315530777, + 0.007779665291309357, + 0.009315728209912777, + 0.007392237428575754, + 0.004551572725176811, + 0.010573945008218288, + 0.012782545760273933, + 0.007175373379141092, + 0.010950564406812191, + 0.005548254586756229, + 0.006863331887871027, + 0.006675431039184332, + 0.007817238569259644, + 0.007673269137740135, + 0.0091625414788723, + 0.01183814276009798, + 0.02062676101922989, + 0.010590329766273499, + 0.009564436972141266, + 0.01622498221695423, + 0.008292539045214653, + 0.016081487759947777, + 0.029091954231262207, + 0.017050575464963913, + 0.02063405141234398, + 0.02139711193740368, + 0.07656621187925339, + 0.022743824869394302, + 0.04954032599925995, + 0.002681132173165679, + 0.0020491736941039562, + 0.003560936078429222, + 0.003179052611812949, + 0.009973662905395031, + 0.005745964590460062, + 0.0021378921810537577, + 0.0018783918349072337, + 0.007613054476678371, + 0.008157579228281975, + 0.00860779918730259, + 0.00618171039968729, + 0.016887664794921875, + 0.008354648016393185, + 0.00786546990275383, + 0.005234329961240292, + 0.008613320998847485, + 0.008671826682984829, + 0.006652375683188438, + 0.056507211178541183, + 0.009735136292874813, + 0.006486272905021906, + 0.06624193489551544, + 0.008093249052762985 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 9, + 0 + ], + "selected_reward": -0.05276527255773544 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a complex robotic arm located on a futuristic automotive assembly line?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the robotic arm's surface show a blend of polished gunmetal and matte carbon fiber materials?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the video show supercooled liquid cascading down and crystalline ice blooming across the arm's joints?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is there a bright, chaotic yellow and white electrical spark arcing from a frozen joint?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera positioned in a static, slightly low-angled extreme close-up shot?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04308966174721718 + ], + "opposite_logprob_tokens": [ + -3.1680896282196045 + ], + "logprob_reward": -0.04308966174721718, + "yes_no_margin_reward": 3.125, + "token_credit_reward": -0.04308966174721718, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15208078920841217, + 0.11637002229690552, + 0.06582210958003998, + 0.09582643955945969, + 0.04701348394155502, + 0.06713555753231049, + 0.17883418500423431, + 0.07545728981494904, + 0.09447275102138519, + 0.10698740184307098 + ], + "credit_map": [ + 0.028581472113728523, + 0.020758548751473427, + 0.021481461822986603, + 0.016058482229709625, + 0.015342230908572674, + 0.015341981314122677, + 0.018291901797056198, + 0.016224687919020653, + 0.015051080845296383, + 0.015815602615475655, + 0.014600101858377457, + 0.010394305922091007, + 0.01428157091140747, + 0.015674006193876266, + 0.018857361748814583, + 0.011695974506437778, + 0.013480751775205135, + 0.006356740836054087, + 0.006859497167170048, + 0.006834777537733316, + 0.010140031576156616, + 0.007865140214562416, + 0.006672929041087627, + 0.007612232118844986, + 0.009771189652383327, + 0.010710756294429302, + 0.012911856174468994, + 0.01580427773296833, + 0.013424040749669075, + 0.012564617209136486, + 0.011339440010488033, + 0.009300260804593563, + 0.00719168595969677, + 0.005633711814880371, + 0.006506850942969322, + 0.005586110986769199, + 0.005033027846366167, + 0.005053697153925896, + 0.004142463207244873, + 0.007865933701395988, + 0.007263091392815113, + 0.007713687140494585, + 0.007577703334391117, + 0.01061552669852972, + 0.009448094293475151, + 0.006779177580028772, + 0.010442260652780533, + 0.007296012714505196, + 0.011624419130384922, + 0.013186663389205933, + 0.010387889109551907, + 0.011506815440952778, + 0.018022997304797173, + 0.06201218441128731, + 0.012658311054110527, + 0.03943489491939545, + 0.005131853278726339, + 0.005330320913344622, + 0.010716177523136139, + 0.009159455075860023, + 0.017414746806025505, + 0.012183933518826962, + 0.007352215237915516, + 0.008168584667146206, + 0.010019652545452118, + 0.013984151184558868, + 0.009833037853240967, + 0.013821730390191078, + 0.019244616851210594, + 0.008204160258173943, + 0.013589264824986458, + 0.005776129197329283, + 0.007703089155256748, + 0.011479240842163563, + 0.010403177700936794, + 0.025151235982775688, + 0.009079569950699806, + 0.011603346094489098, + 0.021976632997393608, + 0.009591101668775082 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 0, + 1 + ], + "selected_reward": -0.04308966174721718 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a complex robotic arm located on a futuristic automotive assembly line?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the robotic arm's surface show a blend of polished gunmetal and matte carbon fiber materials?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the video show supercooled liquid cascading down and crystalline ice blooming across the arm's joints?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is there a bright, chaotic yellow and white electrical spark arcing from a frozen joint?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera positioned in a static, slightly low-angled extreme close-up shot?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.016073176637291908 + ], + "opposite_logprob_tokens": [ + -4.141073226928711 + ], + "logprob_reward": -0.016073176637291908, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.016073176637291908, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1171470433473587, + 0.0789763480424881, + 0.03720336779952049, + 0.08186627179384232, + 0.0514644980430603, + 0.12813375890254974, + 0.04782925918698311, + 0.07988262921571732, + 0.13559643924236298, + 0.24190038442611694 + ], + "credit_map": [ + 0.018977992236614227, + 0.013528815470635891, + 0.015761082991957664, + 0.008167031221091747, + 0.013634505681693554, + 0.017388051375746727, + 0.009454336948692799, + 0.020235229283571243, + 0.011057466268539429, + 0.00804704986512661, + 0.012330087833106518, + 0.006843315437436104, + 0.008291917853057384, + 0.013852336443960667, + 0.00877781119197607, + 0.009776362217962742, + 0.004311354830861092, + 0.002874005353078246, + 0.003529082518070936, + 0.003390314057469368, + 0.004716936964541674, + 0.004834421444684267, + 0.006043197121471167, + 0.007504056207835674, + 0.007905473001301289, + 0.01565787009894848, + 0.013028900139033794, + 0.008119483478367329, + 0.006137317046523094, + 0.01678437553346157, + 0.008011684753000736, + 0.006221171002835035, + 0.00669135944917798, + 0.009194044396281242, + 0.0050442954525351524, + 0.006892136298120022, + 0.004593311809003353, + 0.005844602826982737, + 0.008021519519388676, + 0.005183226428925991, + 0.011895086616277695, + 0.012280932627618313, + 0.016242820769548416, + 0.02523517981171608, + 0.01476365514099598, + 0.013608219102025032, + 0.02167145349085331, + 0.012436416931450367, + 0.004935835022479296, + 0.006571906618773937, + 0.006266778334975243, + 0.009441658854484558, + 0.004874344449490309, + 0.004738650284707546, + 0.006699072662740946, + 0.004301011096686125, + 0.005783268250524998, + 0.004858520347625017, + 0.014903337694704533, + 0.008298888802528381, + 0.016102902591228485, + 0.011805718764662743, + 0.010621902532875538, + 0.007508086506277323, + 0.013991526328027248, + 0.022771818563342094, + 0.015567120164632797, + 0.015090941451489925, + 0.029419729486107826, + 0.012955431826412678, + 0.016023848205804825, + 0.009776018559932709, + 0.0105601130053401, + 0.011124157346785069, + 0.008784307166934013, + 0.08529169112443924, + 0.01329764537513256, + 0.008317598141729832, + 0.09386131167411804, + 0.010663565248250961 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 8, + 5 + ], + "selected_reward": -0.016073176637291908 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a complex robotic arm located on a futuristic automotive assembly line?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the robotic arm's surface show a blend of polished gunmetal and matte carbon fiber materials?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the video show supercooled liquid cascading down and crystalline ice blooming across the arm's joints?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is there a bright, chaotic yellow and white electrical spark arcing from a frozen joint?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera positioned in a static, slightly low-angled extreme close-up shot?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.009783296845853329 + ], + "opposite_logprob_tokens": [ + -4.6347832679748535 + ], + "logprob_reward": -0.009783296845853329, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.009783296845853329, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07339605689048767, + 0.03045876882970333, + 0.0217958502471447, + 0.028134353458881378, + 0.03577204793691635, + 0.08713503181934357, + 0.09563644975423813, + 0.03706318140029907, + 0.40781205892562866, + 0.1827961802482605 + ], + "credit_map": [ + 0.009925980120897293, + 0.00869783852249384, + 0.007768247742205858, + 0.0055295489728450775, + 0.007935428991913795, + 0.01360312383621931, + 0.00803455151617527, + 0.01190133299678564, + 0.0036364507395774126, + 0.003538789227604866, + 0.003256112104281783, + 0.0028265698347240686, + 0.0038001048378646374, + 0.006480780895799398, + 0.0032963661942631006, + 0.0036235947627574205, + 0.0036487914621829987, + 0.0025876013096421957, + 0.0020985514856874943, + 0.002526302123442292, + 0.003100530942901969, + 0.0024015577509999275, + 0.0022171165328472853, + 0.003215398406609893, + 0.004385488107800484, + 0.006309355143457651, + 0.004319547209888697, + 0.0027965139597654343, + 0.003112322185188532, + 0.003232724731788039, + 0.002183625241741538, + 0.0017947780434042215, + 0.002680279081687331, + 0.00691889738664031, + 0.004337287042289972, + 0.004791939165443182, + 0.0026106499135494232, + 0.004425750579684973, + 0.003923031501471996, + 0.006084213964641094, + 0.0067397672683000565, + 0.00894178356975317, + 0.010027331300079823, + 0.01816725917160511, + 0.010333937592804432, + 0.009172708727419376, + 0.01651637628674507, + 0.007235866971313953, + 0.005399858113378286, + 0.008870084770023823, + 0.004926026798784733, + 0.005830380134284496, + 0.007970872335135937, + 0.0327533558011055, + 0.007334016263484955, + 0.02255185693502426, + 0.0028355601243674755, + 0.004442854784429073, + 0.006694238167256117, + 0.006423260550945997, + 0.005128114949911833, + 0.004084812011569738, + 0.0043057347647845745, + 0.0031486039515584707, + 0.042216189205646515, + 0.05873546749353409, + 0.047149866819381714, + 0.05793873965740204, + 0.08024883270263672, + 0.03895765542984009, + 0.0513683520257473, + 0.0311969555914402, + 0.007925082929432392, + 0.012872266583144665, + 0.009006014093756676, + 0.06192646920681, + 0.010342434979975224, + 0.010199282318353653, + 0.059808652848005295, + 0.01071598194539547 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 9, + 6 + ], + "selected_reward": -0.009783296845853329 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a complex robotic arm located on a futuristic automotive assembly line?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the robotic arm's surface show a blend of polished gunmetal and matte carbon fiber materials?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the video show supercooled liquid cascading down and crystalline ice blooming across the arm's joints?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is there a bright, chaotic yellow and white electrical spark arcing from a frozen joint?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera positioned in a static, slightly low-angled extreme close-up shot?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02057752013206482 + ], + "opposite_logprob_tokens": [ + -3.8955774307250977 + ], + "logprob_reward": -0.02057752013206482, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.02057752013206482, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09418826550245285, + 0.122613824903965, + 0.0990758016705513, + 0.08485788106918335, + 0.11354951560497284, + 0.15070544183254242, + 0.1167391762137413, + 0.061091743409633636, + 0.07703925669193268, + 0.08013916015625 + ], + "credit_map": [ + 0.0171970184892416, + 0.012378860265016556, + 0.011602049693465233, + 0.008501623757183552, + 0.011330085806548595, + 0.010987505316734314, + 0.009696057066321373, + 0.012495054863393307, + 0.022160110995173454, + 0.009217724204063416, + 0.011563176289200783, + 0.010908964090049267, + 0.012748573906719685, + 0.02644968591630459, + 0.014518552459776402, + 0.015047026798129082, + 0.01363253965973854, + 0.009695482440292835, + 0.010546484030783176, + 0.012829146347939968, + 0.010745045728981495, + 0.012318774126470089, + 0.010651242919266224, + 0.018657075241208076, + 0.007188336458057165, + 0.014929918572306633, + 0.01179969497025013, + 0.010090612806379795, + 0.0124935582280159, + 0.013287398032844067, + 0.010377191938459873, + 0.004691167268902063, + 0.012538263574242592, + 0.03113357163965702, + 0.012718562968075275, + 0.013927350752055645, + 0.009811284020543098, + 0.010469312779605389, + 0.010903468355536461, + 0.01204768568277359, + 0.011761782690882683, + 0.01611296646296978, + 0.01691548340022564, + 0.03563817963004112, + 0.017882250249385834, + 0.013461869210004807, + 0.029010703787207603, + 0.009922191500663757, + 0.008117875084280968, + 0.012997008860111237, + 0.008963474072515965, + 0.01276281289756298, + 0.011334115639328957, + 0.025712670758366585, + 0.011074761860072613, + 0.025776444002985954, + 0.0041454993188381195, + 0.0034920629113912582, + 0.010000470094382763, + 0.006143997889012098, + 0.017019277438521385, + 0.010956460610032082, + 0.005677965469658375, + 0.0036560094449669123, + 0.007873867638409138, + 0.010520119220018387, + 0.011172316037118435, + 0.012692988850176334, + 0.01162602100521326, + 0.006943604908883572, + 0.009826786816120148, + 0.006383541505783796, + 0.0030577084980905056, + 0.005541069898754358, + 0.004016597289592028, + 0.028253821656107903, + 0.003911921754479408, + 0.003985577262938023, + 0.022868061438202858, + 0.008504397235810757 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 1, + 6 + ], + "selected_reward": -0.02057752013206482 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "In the style of a sleek cyberpunk animation, a static, slightly low-angled, extreme close-up meticulously depicts a complex robotic arm on a futuristic automotive assembly line. Its surface, a blend of polished gunmetal and matte carbon fiber, gleams under the chilly factory ambient lighting. Suddenly, as a malfunction triggers the fire suppression system, supercooled liquid cascades down, and shimmering droplets can be seen streaking across the frame. The moment freezing rain strikes the arm, a delicate and swift web of crystalline ice blooms across its intricate hydraulic joints, frost quickly enveloping the vibrant cyan glow emanating from the pressurized tubes. The arm's fluid, deliberate movements are interspersed with violent, shuddering tremors, its motion halting mid-task. Immediately, a bright, chaotic electrical spark erupts from the now-frozen joint, arcing erratically between the frozen metal plates. This dazzling, frenzied yellow and white light contrasts sharply with the scene's cool, dark color palette, creating a tense and dramatic atmosphere of catastrophic technological failure against the soft-focus backdrop of the vast, automated factory floor.", + "seed": 8276, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000095/group_10/index_914_seed_8276_dp_21_rank_42_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 10, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.020699455728754403, + "vlm_logprob": -0.020699455728754403, + "vlm_yes_no_margin": 4.275, + "vlm_token_credit": -0.020699455728754403, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a complex robotic arm located on a futuristic automotive assembly line?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the robotic arm's surface show a blend of polished gunmetal and matte carbon fiber materials?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the video show supercooled liquid cascading down and crystalline ice blooming across the arm's joints?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is there a bright, chaotic yellow and white electrical spark arcing from a frozen joint?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera positioned in a static, slightly low-angled extreme close-up shot?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.017808659002184868 + ], + "opposite_logprob_tokens": [ + -4.392808437347412 + ], + "logprob_reward": -0.017808659002184868, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.017808659002184868, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.19628964364528656, + 0.1240573525428772, + 0.12135399878025055, + 0.08695127069950104, + 0.2262216955423355, + 0.0301933903247118, + 0.05984240025281906, + 0.050943028181791306, + 0.02994273602962494, + 0.07420448213815689 + ], + "credit_map": [ + 0.014346274547278881, + 0.04515315592288971, + 0.027297088876366615, + 0.017445597797632217, + 0.025414591655135155, + 0.02675911970436573, + 0.022361960262060165, + 0.017511842772364616, + 0.008211496286094189, + 0.01742599345743656, + 0.017597747966647148, + 0.012400968931615353, + 0.0123722730204463, + 0.01873444952070713, + 0.014768064022064209, + 0.022546367719769478, + 0.024785008281469345, + 0.01238042488694191, + 0.016577988862991333, + 0.018156522884964943, + 0.011412707157433033, + 0.014553199522197247, + 0.012450197711586952, + 0.011037948541343212, + 0.010197751224040985, + 0.006237732246518135, + 0.017997337505221367, + 0.014980057254433632, + 0.010568922385573387, + 0.008925026282668114, + 0.012866687029600143, + 0.005177761428058147, + 0.022903436794877052, + 0.02603336237370968, + 0.02508535049855709, + 0.05170436576008797, + 0.01517974678426981, + 0.022760659456253052, + 0.03728272020816803, + 0.025272056460380554, + 0.0038964631967246532, + 0.005121378228068352, + 0.0032576839439570904, + 0.003430461511015892, + 0.002391596557572484, + 0.003521130420267582, + 0.004268861375749111, + 0.0043058148585259914, + 0.006001520901918411, + 0.007999355904757977, + 0.004479344468563795, + 0.010774928145110607, + 0.0068985349498689175, + 0.00797533243894577, + 0.0112127885222435, + 0.004500597715377808, + 0.008125994354486465, + 0.005176090169698, + 0.005762969609349966, + 0.007107391953468323, + 0.00628270860761404, + 0.004887902177870274, + 0.007803590968251228, + 0.005796382669359446, + 0.002892825985327363, + 0.003694447223097086, + 0.0036734899040311575, + 0.0037022815085947514, + 0.005068499594926834, + 0.0029831496067345142, + 0.0037487035151571035, + 0.004179337061941624, + 0.005634117405861616, + 0.010933968238532543, + 0.0065636117942631245, + 0.008082613348960876, + 0.01968309096992016, + 0.005280435085296631, + 0.012163662351667881, + 0.00586298992857337 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 0, + 1 + ], + "selected_reward": -0.017808659002184868 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a complex robotic arm located on a futuristic automotive assembly line?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the robotic arm's surface show a blend of polished gunmetal and matte carbon fiber materials?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the video show supercooled liquid cascading down and crystalline ice blooming across the arm's joints?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is there a bright, chaotic yellow and white electrical spark arcing from a frozen joint?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera positioned in a static, slightly low-angled extreme close-up shot?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03372034430503845 + ], + "opposite_logprob_tokens": [ + -3.4087202548980713 + ], + "logprob_reward": -0.03372034430503845, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.03372034430503845, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06555482745170593, + 0.05834127217531204, + 0.05922235921025276, + 0.6048887968063354, + 0.05879710242152214, + 0.0437496080994606, + 0.03597049042582512, + 0.03236740455031395, + 0.01444273255765438, + 0.02666533924639225 + ], + "credit_map": [ + 0.004563248250633478, + 0.01025919895619154, + 0.00838066078722477, + 0.008281215094029903, + 0.007692646235227585, + 0.010200493037700653, + 0.009234564378857613, + 0.0069428058341145515, + 0.004085952881723642, + 0.008872272446751595, + 0.00498554389923811, + 0.008736444637179375, + 0.008904744870960712, + 0.007412814535200596, + 0.007049454376101494, + 0.008294048719108105, + 0.010878720320761204, + 0.007540605962276459, + 0.0061798919923603535, + 0.00764729268848896, + 0.00562575738877058, + 0.005967890843749046, + 0.007586094085127115, + 0.007796107325702906, + 0.06379473954439163, + 0.03853346407413483, + 0.12993691861629486, + 0.10984562337398529, + 0.06501494348049164, + 0.06746023148298264, + 0.10374661535024643, + 0.026556331664323807, + 0.005902848672121763, + 0.0056350440718233585, + 0.007325341925024986, + 0.010207968764007092, + 0.004443078767508268, + 0.005640640389174223, + 0.01263033039867878, + 0.007011855021119118, + 0.004128560423851013, + 0.005986588541418314, + 0.0034063842613250017, + 0.006723941303789616, + 0.003245771396905184, + 0.010741443373262882, + 0.005252372939139605, + 0.004264548886567354, + 0.003904070006683469, + 0.004644301254302263, + 0.0031297847162932158, + 0.004216470289975405, + 0.004413245711475611, + 0.005361149087548256, + 0.005678271874785423, + 0.004623198416084051, + 0.005548002663999796, + 0.002663389779627323, + 0.0033774622716009617, + 0.0032339885365217924, + 0.00365602970123291, + 0.002824920229613781, + 0.00577571103349328, + 0.0052879005670547485, + 0.0012859762646257877, + 0.0011957510141655803, + 0.001344827120192349, + 0.0019147988641634583, + 0.0034490078687667847, + 0.0010870668338611722, + 0.001926339347846806, + 0.0022389674559235573, + 0.0021935668773949146, + 0.005817621015012264, + 0.0029526189900934696, + 0.0036244066432118416, + 0.0031472169794142246, + 0.002114407252520323, + 0.004198762588202953, + 0.00261673959903419 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 0, + 2 + ], + "selected_reward": -0.03372034430503845 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a complex robotic arm located on a futuristic automotive assembly line?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the robotic arm's surface show a blend of polished gunmetal and matte carbon fiber materials?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the video show supercooled liquid cascading down and crystalline ice blooming across the arm's joints?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is there a bright, chaotic yellow and white electrical spark arcing from a frozen joint?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera positioned in a static, slightly low-angled extreme close-up shot?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01607552357017994 + ], + "opposite_logprob_tokens": [ + -4.141075611114502 + ], + "logprob_reward": -0.01607552357017994, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.01607552357017994, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11791600286960602, + 0.04589118808507919, + 0.02696610800921917, + 0.5499241948127747, + 0.05549536645412445, + 0.0345027782022953, + 0.01710355468094349, + 0.014378197491168976, + 0.03514482453465462, + 0.10267779231071472 + ], + "credit_map": [ + 0.007507686037570238, + 0.02235231176018715, + 0.01633833535015583, + 0.013099960051476955, + 0.02169075980782509, + 0.01526421308517456, + 0.010846841149032116, + 0.010815899819135666, + 0.0033064621966332197, + 0.006610042415559292, + 0.006038500461727381, + 0.005532874260097742, + 0.005011684261262417, + 0.006829896476119757, + 0.005938387010246515, + 0.006623344961553812, + 0.004912649281322956, + 0.002759926486760378, + 0.0032767353113740683, + 0.003955377731472254, + 0.0026946801226586103, + 0.0028417010325938463, + 0.0038382632192224264, + 0.0026867776177823544, + 0.056685470044612885, + 0.04038895294070244, + 0.10804516822099686, + 0.09824813157320023, + 0.0607558935880661, + 0.06599236279726028, + 0.09181229770183563, + 0.027995964512228966, + 0.00625750282779336, + 0.0040452247485518456, + 0.005155780352652073, + 0.013787409290671349, + 0.0028227241709828377, + 0.003597918199375272, + 0.01030736044049263, + 0.009521454572677612, + 0.004081727005541325, + 0.00541703449562192, + 0.0030026081949472427, + 0.004107325337827206, + 0.0024760800879448652, + 0.006080744322389364, + 0.00439815828576684, + 0.004939103964716196, + 0.0021054979879409075, + 0.0021523258183151484, + 0.0014706664951518178, + 0.002185235731303692, + 0.0021367205772548914, + 0.002093842951580882, + 0.002487270860001445, + 0.0024719955399632454, + 0.001819874276407063, + 0.0009968947852030396, + 0.0018114817794412374, + 0.001313582411967218, + 0.0018882215954363346, + 0.002061407547444105, + 0.0021909894421696663, + 0.002295746700838208, + 0.0023985118605196476, + 0.0030828809831291437, + 0.003752322169020772, + 0.004671446047723293, + 0.00875471904873848, + 0.0033337136264890432, + 0.003957530949264765, + 0.005193701479583979, + 0.007739601191133261, + 0.024449290707707405, + 0.009927447885274887, + 0.010313332080841064, + 0.015429693274199963, + 0.008395835757255554, + 0.01664731837809086, + 0.00977527815848589 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 0, + 9 + ], + "selected_reward": -0.01607552357017994 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a complex robotic arm located on a futuristic automotive assembly line?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the robotic arm's surface show a blend of polished gunmetal and matte carbon fiber materials?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the video show supercooled liquid cascading down and crystalline ice blooming across the arm's joints?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is there a bright, chaotic yellow and white electrical spark arcing from a frozen joint?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera positioned in a static, slightly low-angled extreme close-up shot?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.002199494047090411 + ], + "opposite_logprob_tokens": [ + -6.127199649810791 + ], + "logprob_reward": -0.002199494047090411, + "yes_no_margin_reward": 6.125, + "token_credit_reward": -0.002199494047090411, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1107284352183342, + 0.29217568039894104, + 0.04133699834346771, + 0.041560783982276917, + 0.1452733725309372, + 0.06856051832437515, + 0.04804322123527527, + 0.06441400200128555, + 0.06428201496601105, + 0.1236250028014183 + ], + "credit_map": [ + 0.008092842996120453, + 0.014793706126511097, + 0.01606910116970539, + 0.012433418072760105, + 0.017306508496403694, + 0.016530517488718033, + 0.013830108568072319, + 0.011672237887978554, + 0.014363832771778107, + 0.06530030816793442, + 0.029261577874422073, + 0.028073424473404884, + 0.06305346637964249, + 0.03440721333026886, + 0.025170141831040382, + 0.03254569694399834, + 0.005460604093968868, + 0.005193657241761684, + 0.004025919362902641, + 0.006218896713107824, + 0.0033580963499844074, + 0.0040683988481760025, + 0.005678185727447271, + 0.0073332404717803, + 0.004465341102331877, + 0.003804761217907071, + 0.008796002715826035, + 0.006486766040325165, + 0.003719267901033163, + 0.004154222551733255, + 0.0068216933868825436, + 0.0033127267379313707, + 0.018660854548215866, + 0.011251818388700485, + 0.010740146972239017, + 0.03325542062520981, + 0.008739961311221123, + 0.018423335626721382, + 0.02552768960595131, + 0.018674148246645927, + 0.010233583860099316, + 0.008369361981749535, + 0.006255647633224726, + 0.008853228762745857, + 0.005471367388963699, + 0.011549167335033417, + 0.008127918466925621, + 0.009700237773358822, + 0.00517299072816968, + 0.006325242575258017, + 0.004706747829914093, + 0.006365655455738306, + 0.007068918086588383, + 0.004505295772105455, + 0.008048965595662594, + 0.005849405191838741, + 0.005808030720800161, + 0.005710632540285587, + 0.013505453243851662, + 0.006177231669425964, + 0.005316965281963348, + 0.003664901480078697, + 0.006106474902480841, + 0.01812431775033474, + 0.006891733501106501, + 0.007363076321780682, + 0.007364001125097275, + 0.00612594885751605, + 0.01160189788788557, + 0.007430198602378368, + 0.008170390501618385, + 0.009334766305983067, + 0.008456199429929256, + 0.021922098472714424, + 0.010943129658699036, + 0.012954398058354855, + 0.026872126385569572, + 0.010118306614458561, + 0.01956518366932869, + 0.012793559581041336 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 4, + 9 + ], + "selected_reward": -0.002199494047090411 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a complex robotic arm located on a futuristic automotive assembly line?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the robotic arm's surface show a blend of polished gunmetal and matte carbon fiber materials?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the video show supercooled liquid cascading down and crystalline ice blooming across the arm's joints?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is there a bright, chaotic yellow and white electrical spark arcing from a frozen joint?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera positioned in a static, slightly low-angled extreme close-up shot?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.033693257719278336 + ], + "opposite_logprob_tokens": [ + -3.408693313598633 + ], + "logprob_reward": -0.033693257719278336, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.033693257719278336, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10461663454771042, + 0.11277611553668976, + 0.06794123351573944, + 0.1739601343870163, + 0.09223239868879318, + 0.09624911844730377, + 0.18933512270450592, + 0.023174159228801727, + 0.05705934017896652, + 0.08265569061040878 + ], + "credit_map": [ + 0.007345621474087238, + 0.012149767018854618, + 0.01678672432899475, + 0.013574891723692417, + 0.013734615407884121, + 0.01622898131608963, + 0.012648732401430607, + 0.01214730553328991, + 0.006267281714826822, + 0.02033630758523941, + 0.014518226496875286, + 0.01252808328717947, + 0.0194363072514534, + 0.014372010715305805, + 0.010652163065969944, + 0.01466574240475893, + 0.012840993702411652, + 0.009266170673072338, + 0.007152218371629715, + 0.007439367473125458, + 0.005731591489166021, + 0.007104489021003246, + 0.011415822431445122, + 0.006990580353885889, + 0.018275465816259384, + 0.01099388301372528, + 0.03552745655179024, + 0.03161317855119705, + 0.02000153809785843, + 0.020723426714539528, + 0.027902962639927864, + 0.008922227658331394, + 0.011296662501990795, + 0.006979446392506361, + 0.00886860117316246, + 0.018048088997602463, + 0.00780571810901165, + 0.011918140575289726, + 0.017493339255452156, + 0.009822404943406582, + 0.01392141543328762, + 0.012086708098649979, + 0.007879482582211494, + 0.011841444298624992, + 0.005964023992419243, + 0.021662021055817604, + 0.011592799797654152, + 0.011301219463348389, + 0.019135242328047752, + 0.021419817581772804, + 0.019097521901130676, + 0.02831760048866272, + 0.026097852736711502, + 0.01788318157196045, + 0.032651517540216446, + 0.024732399731874466, + 0.0026026361156255007, + 0.0022793167736381292, + 0.003858117153868079, + 0.0025660786777734756, + 0.002297694096341729, + 0.0019848253577947617, + 0.002903378801420331, + 0.004682112485170364, + 0.006951912771910429, + 0.006007527932524681, + 0.006145203486084938, + 0.005908203776925802, + 0.009925534017384052, + 0.006073001306504011, + 0.009083952754735947, + 0.006964005529880524, + 0.004632452502846718, + 0.014935959130525589, + 0.006029522977769375, + 0.00925997644662857, + 0.02082081511616707, + 0.0061305793933570385, + 0.014040417969226837, + 0.006805965676903725 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 3, + 1 + ], + "selected_reward": -0.033693257719278336 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "In the style of a sleek cyberpunk animation, a static, slightly low-angled, extreme close-up meticulously depicts a complex robotic arm on a futuristic automotive assembly line. Its surface, a blend of polished gunmetal and matte carbon fiber, gleams under the chilly factory ambient lighting. Suddenly, as a malfunction triggers the fire suppression system, supercooled liquid cascades down, and shimmering droplets can be seen streaking across the frame. The moment freezing rain strikes the arm, a delicate and swift web of crystalline ice blooms across its intricate hydraulic joints, frost quickly enveloping the vibrant cyan glow emanating from the pressurized tubes. The arm's fluid, deliberate movements are interspersed with violent, shuddering tremors, its motion halting mid-task. Immediately, a bright, chaotic electrical spark erupts from the now-frozen joint, arcing erratically between the frozen metal plates. This dazzling, frenzied yellow and white light contrasts sharply with the scene's cool, dark color palette, creating a tense and dramatic atmosphere of catastrophic technological failure against the soft-focus backdrop of the vast, automated factory floor.", + "seed": 8277, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000095/group_10/index_914_seed_8277_dp_21_rank_42_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 10, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.03684935476630926, + "vlm_logprob": -0.03684935476630926, + "vlm_yes_no_margin": 3.4, + "vlm_token_credit": -0.03684935476630926, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a complex robotic arm located on a futuristic automotive assembly line?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the robotic arm's surface show a blend of polished gunmetal and matte carbon fiber materials?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the video show supercooled liquid cascading down and crystalline ice blooming across the arm's joints?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is there a bright, chaotic yellow and white electrical spark arcing from a frozen joint?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera positioned in a static, slightly low-angled extreme close-up shot?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.036202192306518555 + ], + "opposite_logprob_tokens": [ + -3.4112021923065186 + ], + "logprob_reward": -0.036202192306518555, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.036202192306518555, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14286597073078156, + 0.0963030532002449, + 0.05615120008587837, + 0.0627165213227272, + 0.04816405847668648, + 0.05749806761741638, + 0.15368129312992096, + 0.1891959011554718, + 0.14956402778625488, + 0.04385994002223015 + ], + "credit_map": [ + 0.008771355263888836, + 0.02796955034136772, + 0.011373722925782204, + 0.012612995691597462, + 0.028842873871326447, + 0.02965044230222702, + 0.01664431393146515, + 0.007000715937465429, + 0.005453397985547781, + 0.010302470996975899, + 0.006775557063519955, + 0.012099076062440872, + 0.016144974157214165, + 0.01113410759717226, + 0.023709191009402275, + 0.010684281587600708, + 0.003220796585083008, + 0.010191014036536217, + 0.0026840604841709137, + 0.003394558560103178, + 0.013217885047197342, + 0.009287716820836067, + 0.0090140076354146, + 0.005141162779182196, + 0.0036424556747078896, + 0.012658756226301193, + 0.007046994753181934, + 0.00832658912986517, + 0.006305500864982605, + 0.009618731215596199, + 0.010801504366099834, + 0.004315986298024654, + 0.0030448928009718657, + 0.0027632354758679867, + 0.0039000390097498894, + 0.0038013330195099115, + 0.004731465131044388, + 0.008456060662865639, + 0.013735114596784115, + 0.007731916382908821, + 0.005891901906579733, + 0.00447803782299161, + 0.005437158513814211, + 0.005551997106522322, + 0.006569949444383383, + 0.008450957015156746, + 0.013603230006992817, + 0.007514835335314274, + 0.018788473680615425, + 0.010482092387974262, + 0.018585748970508575, + 0.018385808914899826, + 0.015241125598549843, + 0.017692990601062775, + 0.023353135213255882, + 0.031151926144957542, + 0.020810000598430634, + 0.015005226247012615, + 0.022777218371629715, + 0.019582564011216164, + 0.022516293451189995, + 0.022688284516334534, + 0.03488514572381973, + 0.030931169167160988, + 0.017474958673119545, + 0.04904623702168465, + 0.011759079992771149, + 0.009495541453361511, + 0.011491212993860245, + 0.009242355823516846, + 0.007515431847423315, + 0.03353921324014664, + 0.005479537881910801, + 0.013959456235170364, + 0.0037500064354389906, + 0.002887783804908395, + 0.004763892386108637, + 0.0022567540872842073, + 0.0072610266506671906, + 0.003501483704894781 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 6, + 8 + ], + "selected_reward": -0.036202192306518555 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a complex robotic arm located on a futuristic automotive assembly line?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the robotic arm's surface show a blend of polished gunmetal and matte carbon fiber materials?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the video show supercooled liquid cascading down and crystalline ice blooming across the arm's joints?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is there a bright, chaotic yellow and white electrical spark arcing from a frozen joint?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera positioned in a static, slightly low-angled extreme close-up shot?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03811531513929367 + ], + "opposite_logprob_tokens": [ + -3.2881152629852295 + ], + "logprob_reward": -0.03811531513929367, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.03811531513929367, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.101681649684906, + 0.1449594646692276, + 0.10444769263267517, + 0.038276731967926025, + 0.07724741846323013, + 0.08496621996164322, + 0.10675262659788132, + 0.12166821211576462, + 0.09088841825723648, + 0.12911160290241241 + ], + "credit_map": [ + 0.007666703313589096, + 0.02461506612598896, + 0.005876118317246437, + 0.005383452866226435, + 0.025576094165444374, + 0.01803460530936718, + 0.009711464866995811, + 0.00481813307851553, + 0.006718778982758522, + 0.01967931166291237, + 0.009418408386409283, + 0.015676692128181458, + 0.029413610696792603, + 0.017470894381403923, + 0.03250624239444733, + 0.014075512066483498, + 0.004047450143843889, + 0.006762130651623011, + 0.0037842444144189358, + 0.005360611714422703, + 0.01092592068016529, + 0.031553782522678375, + 0.03541183099150658, + 0.00660170242190361, + 0.002658698707818985, + 0.009379659779369831, + 0.0038449387066066265, + 0.004660752601921558, + 0.0037188930436968803, + 0.005165474023669958, + 0.005565524101257324, + 0.0032827884424477816, + 0.005307513754814863, + 0.0032905270345509052, + 0.008193298242986202, + 0.006135579198598862, + 0.009404252283275127, + 0.018749136477708817, + 0.018711116164922714, + 0.0074559831991791725, + 0.007325677666813135, + 0.006610915530472994, + 0.008253510110080242, + 0.008206586353480816, + 0.008958316408097744, + 0.012038713321089745, + 0.020903371274471283, + 0.01266911905258894, + 0.012486843392252922, + 0.007184085436165333, + 0.011962264776229858, + 0.01435343362390995, + 0.009611748158931732, + 0.010148553177714348, + 0.018983040004968643, + 0.022022642195224762, + 0.013545372523367405, + 0.011549508199095726, + 0.01756918802857399, + 0.01744559220969677, + 0.013206996954977512, + 0.01033945195376873, + 0.020206572487950325, + 0.017805514857172966, + 0.008669503964483738, + 0.031102638691663742, + 0.007485340349376202, + 0.004718396812677383, + 0.005687699653208256, + 0.004859131295233965, + 0.004234538413584232, + 0.024131150916218758, + 0.013556379824876785, + 0.04608586058020592, + 0.007341880816966295, + 0.0053151994943618774, + 0.018058117479085922, + 0.007318213582038879, + 0.02236800454556942, + 0.009067938663065434 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 9, + 7 + ], + "selected_reward": -0.03811531513929367 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a complex robotic arm located on a futuristic automotive assembly line?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the robotic arm's surface show a blend of polished gunmetal and matte carbon fiber materials?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the video show supercooled liquid cascading down and crystalline ice blooming across the arm's joints?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is there a bright, chaotic yellow and white electrical spark arcing from a frozen joint?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera positioned in a static, slightly low-angled extreme close-up shot?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.062044139951467514 + ], + "opposite_logprob_tokens": [ + -2.812044143676758 + ], + "logprob_reward": -0.062044139951467514, + "yes_no_margin_reward": 2.75, + "token_credit_reward": -0.062044139951467514, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.054617688059806824, + 0.2898944020271301, + 0.422305166721344, + 0.028078030794858932, + 0.01677938550710678, + 0.022806307300925255, + 0.017501357942819595, + 0.016152840107679367, + 0.01634165644645691, + 0.11552312970161438 + ], + "credit_map": [ + 0.0035302063915878534, + 0.01149784680455923, + 0.004250268917530775, + 0.005092299077659845, + 0.011530705727636814, + 0.010703475214540958, + 0.005223426967859268, + 0.00278946734033525, + 0.013048316352069378, + 0.033083342015743256, + 0.014892308041453362, + 0.02724648453295231, + 0.059664398431777954, + 0.036874476820230484, + 0.07165475934743881, + 0.03343033045530319, + 0.01285556424409151, + 0.019288234412670135, + 0.012602410279214382, + 0.01953926496207714, + 0.029154155403375626, + 0.14889469742774963, + 0.15899036824703217, + 0.020980510860681534, + 0.0017468435689806938, + 0.006026704795658588, + 0.0025588409043848515, + 0.0035325109492987394, + 0.005222493316978216, + 0.0036980127915740013, + 0.003249584464356303, + 0.002043043030425906, + 0.001195664401166141, + 0.001054689520969987, + 0.0009994646534323692, + 0.0012815251247957349, + 0.0017547757597640157, + 0.005018807947635651, + 0.002323232125490904, + 0.003151229815557599, + 0.002523609669879079, + 0.002015170408412814, + 0.0021905892062932253, + 0.002018499420955777, + 0.0025767774786800146, + 0.0037937569431960583, + 0.0042000398971140385, + 0.0034878680016845465, + 0.0024692958686500788, + 0.001559130847454071, + 0.0028814836405217648, + 0.0020250387024134398, + 0.0016977343475446105, + 0.0013155603082850575, + 0.0025565598625689745, + 0.0029965564608573914, + 0.001820593373849988, + 0.0016838250448927283, + 0.0022463654167950153, + 0.002205783035606146, + 0.002857642015442252, + 0.0017739083850756288, + 0.0014762587379664183, + 0.0020884666591882706, + 0.0013155783526599407, + 0.004442023579031229, + 0.0013948483392596245, + 0.001099046552553773, + 0.001424069982022047, + 0.0013033057330176234, + 0.0012757395161315799, + 0.004087045323103666, + 0.011658882722258568, + 0.03462997078895569, + 0.00561189092695713, + 0.005336663220077753, + 0.011424538679420948, + 0.006769005209207535, + 0.02889833226799965, + 0.011193869635462761 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 1, + 9 + ], + "selected_reward": -0.062044139951467514 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a complex robotic arm located on a futuristic automotive assembly line?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the robotic arm's surface show a blend of polished gunmetal and matte carbon fiber materials?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the video show supercooled liquid cascading down and crystalline ice blooming across the arm's joints?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is there a bright, chaotic yellow and white electrical spark arcing from a frozen joint?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera positioned in a static, slightly low-angled extreme close-up shot?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03369452804327011 + ], + "opposite_logprob_tokens": [ + -3.4086945056915283 + ], + "logprob_reward": -0.03369452804327011, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.03369452804327011, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09713037312030792, + 0.25636759400367737, + 0.07494091987609863, + 0.062274955213069916, + 0.04748980328440666, + 0.04440053552389145, + 0.034590378403663635, + 0.05459962412714958, + 0.047919861972332, + 0.28028595447540283 + ], + "credit_map": [ + 0.008022820577025414, + 0.013062852434813976, + 0.005548624787479639, + 0.005343565251678228, + 0.0208758432418108, + 0.03289267048239708, + 0.00697315065190196, + 0.004410844761878252, + 0.011330779641866684, + 0.031390007585287094, + 0.015130058862268925, + 0.026509229093790054, + 0.05219501629471779, + 0.031881846487522125, + 0.060106366872787476, + 0.02782430127263069, + 0.002214806852862239, + 0.005982308182865381, + 0.003052694955840707, + 0.0031468202359974384, + 0.004867147654294968, + 0.029230214655399323, + 0.022491365671157837, + 0.003955557476729155, + 0.004204064607620239, + 0.012887701392173767, + 0.005450457334518433, + 0.008315080776810646, + 0.006904426030814648, + 0.008701260201632977, + 0.010832411237061024, + 0.004979549907147884, + 0.005037868861109018, + 0.003343218471854925, + 0.006815473549067974, + 0.004095415119081736, + 0.005823793821036816, + 0.009343397803604603, + 0.004569016396999359, + 0.008461622521281242, + 0.004296868573874235, + 0.003506978740915656, + 0.0037318491376936436, + 0.004603955894708633, + 0.0069654639810323715, + 0.007278321776539087, + 0.008414884097874165, + 0.005602211691439152, + 0.00342587660998106, + 0.005054180044680834, + 0.006469263695180416, + 0.0036364118568599224, + 0.004872682970017195, + 0.0026830898132175207, + 0.0024914504028856754, + 0.005957422312349081, + 0.004541868343949318, + 0.006663176231086254, + 0.0064430502243340015, + 0.0058416156098246574, + 0.007934825494885445, + 0.0037872877437621355, + 0.015250515192747116, + 0.0041372813284397125, + 0.004567031282931566, + 0.010265850462019444, + 0.004048625007271767, + 0.004541144706308842, + 0.00476972758769989, + 0.0023449279833585024, + 0.0029738459270447493, + 0.014408706687390804, + 0.027521289885044098, + 0.02145478129386902, + 0.027713770046830177, + 0.046650949865579605, + 0.05556448549032211, + 0.029358670115470886, + 0.03318975120782852, + 0.03883225470781326 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 1, + 0 + ], + "selected_reward": -0.03369452804327011 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a complex robotic arm located on a futuristic automotive assembly line?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the robotic arm's surface show a blend of polished gunmetal and matte carbon fiber materials?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the video show supercooled liquid cascading down and crystalline ice blooming across the arm's joints?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is there a bright, chaotic yellow and white electrical spark arcing from a frozen joint?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera positioned in a static, slightly low-angled extreme close-up shot?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.014190598390996456 + ], + "opposite_logprob_tokens": [ + -4.264190673828125 + ], + "logprob_reward": -0.014190598390996456, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.014190598390996456, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1209423840045929, + 0.24212683737277985, + 0.04675629734992981, + 0.037754856050014496, + 0.13132016360759735, + 0.03881271556019783, + 0.06543509662151337, + 0.04384729638695717, + 0.1031324565410614, + 0.16987183690071106 + ], + "credit_map": [ + 0.009066933766007423, + 0.030139153823256493, + 0.008226458914577961, + 0.00904423650354147, + 0.02387830801308155, + 0.026704072952270508, + 0.008751727640628815, + 0.005131489131599665, + 0.017536107450723648, + 0.03093588724732399, + 0.013522278517484665, + 0.018855463713407516, + 0.04262565076351166, + 0.030676424503326416, + 0.05693621188402176, + 0.03103882446885109, + 0.0027705510146915913, + 0.010825525037944317, + 0.0034227431751787663, + 0.00370863638818264, + 0.005496112164109945, + 0.012052810750901699, + 0.005634566769003868, + 0.0028453529812395573, + 0.0029832914005964994, + 0.006064999848604202, + 0.004010477568954229, + 0.004049890674650669, + 0.005475288722664118, + 0.0063974857330322266, + 0.005593529902398586, + 0.0031798947602510452, + 0.009196971543133259, + 0.005498907528817654, + 0.016854776069521904, + 0.008544080890715122, + 0.01097280066460371, + 0.014584898948669434, + 0.031250085681676865, + 0.03441765531897545, + 0.0034435121342539787, + 0.0030755256302654743, + 0.004841438494622707, + 0.004063224885612726, + 0.005652392283082008, + 0.005879147443920374, + 0.006992793641984463, + 0.004864685703068972, + 0.008602006360888481, + 0.004202554002404213, + 0.0070577384904026985, + 0.0075910366140306, + 0.006440773140639067, + 0.007684288080781698, + 0.010925070382654667, + 0.012931632809340954, + 0.004795955028384924, + 0.0035859383642673492, + 0.0056909359991550446, + 0.00547832204028964, + 0.005771439056843519, + 0.0045040990225970745, + 0.00709961773827672, + 0.006920985411852598, + 0.008694501593708992, + 0.0338766910135746, + 0.007552956230938435, + 0.004425737541168928, + 0.005699957720935345, + 0.004189800005406141, + 0.004897002596408129, + 0.033795811235904694, + 0.016864687204360962, + 0.0479145348072052, + 0.008277463726699352, + 0.009453807957470417, + 0.021643485873937607, + 0.009803150780498981, + 0.0419631153345108, + 0.013951586559414864 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 9, + 4 + ], + "selected_reward": -0.014190598390996456 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000095/group_11/index_tensor([4418, 4418])_group_11_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000095/group_11/index_tensor([4418, 4418])_group_11_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..fda3228e8838d0315dc86518a66e5f5b64df96c0 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000095/group_11/index_tensor([4418, 4418])_group_11_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "medium close-up shot from a level angle, the camera held still, captures with striking force a young man in a dark gray hooded sweatshirt standing in a dimly lit room. The sole cold light, coming from the side, casts deep shadows across his face, tightened with anger, illuminating the wall mirror before him, which reflects his own face, twisted and distorted with rage. The hood of his sweatshirt is pulled low, nearly obscuring his eyes, revealing only his furrowed brow and pursed lips. He jerks his right arm back, his clenched fist veins bulging and knuckles whitening from the exertion. Then, with a swift and decisive movement imbued with devastating force, he thrusts his fist forward, slamming it into the cold mirror. At the moment of impact, spiderweb-like cracks spread across the entire surface of the mirror before violently exploding. Countless shimmering glass shards fly explosively in all directions, tracing sharp paths through the air as if time has slowed down. His fist penetrates the center of the mirror and comes to rest in the jagged hole. The entire scene is filled with primal rage and a heart-pounding destructive catharsis, creating an atmosphere of extreme tension and despair.", + "seed": 11486, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000095/group_11/index_4418_seed_11486_dp_22_rank_44_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 11, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.032109671365469696, + "vlm_logprob": -0.032109671365469696, + "vlm_yes_no_margin": 4.45, + "vlm_token_credit": -0.032109671365469696, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a medium close-up of a young man wearing a dark gray hooded sweatshirt with the hood pulled low?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene lit by a single cold light source from the side, casting deep shadows on the man's face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the man perform the action of jerking his right arm back with a clenched fist and then thrusting it forward to slam into a wall mirror?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do spiderweb-like cracks spread across the mirror surface upon impact, followed by an explosion of glass shards flying in all directions?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera held still at a level angle throughout the scene?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.019734937697649002 + ], + "opposite_logprob_tokens": [ + -4.519734859466553 + ], + "logprob_reward": -0.019734937697649002, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.019734937697649002, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1214207112789154, + 0.02335251495242119, + 0.17404058575630188, + 0.15425577759742737, + 0.033467069268226624, + 0.12400206178426743, + 0.2686997354030609, + 0.020825518295168877, + 0.020503325387835503, + 0.059432778507471085 + ], + "credit_map": [ + 0.009316795505583286, + 0.007017230149358511, + 0.01154325157403946, + 0.009914366528391838, + 0.007411671336740255, + 0.033674731850624084, + 0.008712785318493843, + 0.03382987529039383, + 0.0029166266322135925, + 0.0017157631227746606, + 0.002688491018489003, + 0.001761185354553163, + 0.0032387082464993, + 0.005229013506323099, + 0.0024526044726371765, + 0.0033501225989311934, + 0.008040945045650005, + 0.0017615128308534622, + 0.0028205886483192444, + 0.00404124753549695, + 0.0025202243123203516, + 0.07669954746961594, + 0.0027445314917713404, + 0.07541198283433914, + 0.007219593971967697, + 0.0038191487547010183, + 0.003702439833432436, + 0.0031329100020229816, + 0.0029508969746530056, + 0.06598313897848129, + 0.0022961909417062998, + 0.06515144556760788, + 0.0026314870920032263, + 0.004265860188752413, + 0.002216548891738057, + 0.0035755466669797897, + 0.005704159382730722, + 0.006061893422156572, + 0.003399614943191409, + 0.00561195844784379, + 0.007790945470333099, + 0.009261311031877995, + 0.0068505858071148396, + 0.006902978289872408, + 0.009367766790091991, + 0.03744282200932503, + 0.007909566164016724, + 0.03847608342766762, + 0.08839092403650284, + 0.019886605441570282, + 0.11374863237142563, + 0.01085721980780363, + 0.006308484822511673, + 0.007632018532603979, + 0.004897750448435545, + 0.01697809249162674, + 0.0044692507945001125, + 0.0035378204192966223, + 0.0024079852737486362, + 0.001842509605921805, + 0.002006285823881626, + 0.0017543273279443383, + 0.002153311390429735, + 0.002654027659446001, + 0.003053167136386037, + 0.0027552114333957434, + 0.002220581052824855, + 0.0035635847598314285, + 0.0022048908285796642, + 0.0019123534439131618, + 0.002988425549119711, + 0.001805112580768764, + 0.0075733051635324955, + 0.00810123048722744, + 0.007019802462309599, + 0.012834821827709675, + 0.005468928255140781, + 0.005547572858631611, + 0.007242823019623756, + 0.005644294898957014 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 2, + 3 + ], + "selected_reward": -0.019734937697649002 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a medium close-up of a young man wearing a dark gray hooded sweatshirt with the hood pulled low?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene lit by a single cold light source from the side, casting deep shadows on the man's face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the man perform the action of jerking his right arm back with a clenched fist and then thrusting it forward to slam into a wall mirror?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do spiderweb-like cracks spread across the mirror surface upon impact, followed by an explosion of glass shards flying in all directions?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera held still at a level angle throughout the scene?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02060706540942192 + ], + "opposite_logprob_tokens": [ + -3.8956069946289062 + ], + "logprob_reward": -0.02060706540942192, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.02060706540942192, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08695589751005173, + 0.20042678713798523, + 0.20761220157146454, + 0.11305135488510132, + 0.0446263812482357, + 0.04883972555398941, + 0.2476370930671692, + 0.009672966785728931, + 0.022854045033454895, + 0.018323522061109543 + ], + "credit_map": [ + 0.00954427383840084, + 0.006345424335449934, + 0.01111409068107605, + 0.006830273196101189, + 0.00625159777700901, + 0.021297637373209, + 0.005803100764751434, + 0.019769497215747833, + 0.01127223577350378, + 0.005026679486036301, + 0.009224612265825272, + 0.007238414138555527, + 0.005978946574032307, + 0.0774439200758934, + 0.006758143659681082, + 0.07748383283615112, + 0.009618503041565418, + 0.0023534742649644613, + 0.003441161010414362, + 0.004074437078088522, + 0.0026591261848807335, + 0.09129046648740768, + 0.0032115760259330273, + 0.0909634679555893, + 0.005795284640043974, + 0.005210035480558872, + 0.002585980575531721, + 0.0032274131663143635, + 0.006909188348799944, + 0.04294031485915184, + 0.0032835304737091064, + 0.04309960827231407, + 0.004329164046794176, + 0.004846472293138504, + 0.0022526499815285206, + 0.004250273574143648, + 0.012385607697069645, + 0.0069725895300507545, + 0.004859886597841978, + 0.004729737062007189, + 0.0020445366390049458, + 0.001654889201745391, + 0.0027560789603739977, + 0.0013806133065372705, + 0.0016630179015919566, + 0.018580935895442963, + 0.0013901350321248174, + 0.019369518384337425, + 0.08061480522155762, + 0.01879839226603508, + 0.10462293773889542, + 0.010125198401510715, + 0.005996689200401306, + 0.007353123277425766, + 0.004599395673722029, + 0.015526551753282547, + 0.001435441430658102, + 0.0011572398943826556, + 0.001231217640452087, + 0.0010541334049776196, + 0.0008679446764290333, + 0.0013607675209641457, + 0.0012134475400671363, + 0.0013527738628908992, + 0.004295325372368097, + 0.0025406721979379654, + 0.0028728903271257877, + 0.004085779655724764, + 0.0017976920353248715, + 0.0019291479839012027, + 0.0032475844491273165, + 0.0020849546417593956, + 0.0029506944119930267, + 0.00230897543951869, + 0.0020237332209944725, + 0.003225018735975027, + 0.0016771173104643822, + 0.001909148064441979, + 0.0021607154048979282, + 0.0020681191235780716 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 2, + 1 + ], + "selected_reward": -0.02060706540942192 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a medium close-up of a young man wearing a dark gray hooded sweatshirt with the hood pulled low?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene lit by a single cold light source from the side, casting deep shadows on the man's face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the man perform the action of jerking his right arm back with a clenched fist and then thrusting it forward to slam into a wall mirror?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do spiderweb-like cracks spread across the mirror surface upon impact, followed by an explosion of glass shards flying in all directions?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera held still at a level angle throughout the scene?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005955095402896404 + ], + "opposite_logprob_tokens": [ + -5.130955219268799 + ], + "logprob_reward": -0.005955095402896404, + "yes_no_margin_reward": 5.125, + "token_credit_reward": -0.005955095402896404, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06441978365182877, + 0.027539122849702835, + 0.15105567872524261, + 0.04186563193798065, + 0.1146245226264, + 0.31647446751594543, + 0.10584450513124466, + 0.07288410514593124, + 0.03677307069301605, + 0.06851904839277267 + ], + "credit_map": [ + 0.005814925767481327, + 0.003126596799120307, + 0.005892459768801928, + 0.004059472121298313, + 0.004135449882596731, + 0.01916252262890339, + 0.0030727884732186794, + 0.01915556937456131, + 0.0017285540234297514, + 0.0012025241740047932, + 0.0015004572924226522, + 0.001428092597052455, + 0.001756519777700305, + 0.008949816226959229, + 0.0016330201178789139, + 0.009340140968561172, + 0.007872818037867546, + 0.0038440278731286526, + 0.009890677407383919, + 0.005093127954751253, + 0.004058313090354204, + 0.057911861687898636, + 0.004122384823858738, + 0.05826248601078987, + 0.0036104500759392977, + 0.0034306615125387907, + 0.005855014082044363, + 0.003339625895023346, + 0.003267009276896715, + 0.009414067491889, + 0.0033345920965075493, + 0.009614215232431889, + 0.005803187843412161, + 0.009866303764283657, + 0.006852542515844107, + 0.0060514784418046474, + 0.009625458158552647, + 0.03625958785414696, + 0.005469924304634333, + 0.03469603508710861, + 0.01627781055867672, + 0.0097952950745821, + 0.014070780016481876, + 0.010193879716098309, + 0.007323138881474733, + 0.12253157794475555, + 0.007970234379172325, + 0.12831178307533264, + 0.028414543718099594, + 0.009613996371626854, + 0.03718498349189758, + 0.006007310934364796, + 0.004557645879685879, + 0.008595782332122326, + 0.004509652033448219, + 0.00696059875190258, + 0.01864204742014408, + 0.015031544491648674, + 0.009197844192385674, + 0.005265326239168644, + 0.0059590935707092285, + 0.0045053670182824135, + 0.004547575954347849, + 0.009735306724905968, + 0.00717683183029294, + 0.003939748276025057, + 0.005065445322543383, + 0.0045458595268428326, + 0.003200015053153038, + 0.0029198986012488604, + 0.005754052195698023, + 0.00417122058570385, + 0.009198262356221676, + 0.008980809710919857, + 0.007488196715712547, + 0.013260926119983196, + 0.006704468280076981, + 0.007441976107656956, + 0.008328376337885857, + 0.007116039749234915 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 2, + 4 + ], + "selected_reward": -0.005955095402896404 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a medium close-up of a young man wearing a dark gray hooded sweatshirt with the hood pulled low?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene lit by a single cold light source from the side, casting deep shadows on the man's face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the man perform the action of jerking his right arm back with a clenched fist and then thrusting it forward to slam into a wall mirror?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do spiderweb-like cracks spread across the mirror surface upon impact, followed by an explosion of glass shards flying in all directions?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera held still at a level angle throughout the scene?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.001335205975919962 + ], + "opposite_logprob_tokens": [ + -6.626335144042969 + ], + "logprob_reward": -0.001335205975919962, + "yes_no_margin_reward": 6.625, + "token_credit_reward": -0.001335205975919962, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.051399145275354385, + 0.04373501241207123, + 0.05481450632214546, + 0.11430619657039642, + 0.052567340433597565, + 0.19725950062274933, + 0.18952107429504395, + 0.13617467880249023, + 0.11930398643016815, + 0.04091854766011238 + ], + "credit_map": [ + 0.004999874159693718, + 0.0030260912608355284, + 0.005351631436496973, + 0.0037093739956617355, + 0.0030026601161807775, + 0.013454978354275227, + 0.004559161141514778, + 0.013295380398631096, + 0.002038925653323531, + 0.0015522160101681948, + 0.0017053713090717793, + 0.00202663941308856, + 0.0013520343927666545, + 0.01691923290491104, + 0.0015701286029070616, + 0.01657046750187874, + 0.003810314927250147, + 0.0019615862984210253, + 0.005228704307228327, + 0.002451914129778743, + 0.0019207445438951254, + 0.019507646560668945, + 0.0014297596644610167, + 0.018503841012716293, + 0.005865727085620165, + 0.005599661730229855, + 0.0028768745250999928, + 0.00347762368619442, + 0.005493210628628731, + 0.04348351061344147, + 0.003910115920007229, + 0.04359947890043259, + 0.0030786478891968727, + 0.004924024920910597, + 0.0025894450955092907, + 0.0034221194218844175, + 0.008337668143212795, + 0.01295563019812107, + 0.004271029494702816, + 0.012988781556487083, + 0.01032282318919897, + 0.007392373401671648, + 0.007832275703549385, + 0.007305003236979246, + 0.005755260586738586, + 0.0749456062912941, + 0.005755993537604809, + 0.07795017957687378, + 0.05852022022008896, + 0.014525135979056358, + 0.07635211199522018, + 0.008819269947707653, + 0.005278930999338627, + 0.009385989978909492, + 0.0052519324235618114, + 0.01138751208782196, + 0.03291301429271698, + 0.016380183398723602, + 0.019501594826579094, + 0.011158921755850315, + 0.01253926195204258, + 0.015712639316916466, + 0.012828168459236622, + 0.015140900388360023, + 0.02956789918243885, + 0.013305488973855972, + 0.01509436871856451, + 0.010439111851155758, + 0.013038263656198978, + 0.010104975663125515, + 0.015909653156995773, + 0.011844230815768242, + 0.006679655518382788, + 0.004747329279780388, + 0.006647566799074411, + 0.007418390829116106, + 0.0033265671227127314, + 0.004128040745854378, + 0.004263443406671286, + 0.0037075579166412354 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 6, + 7 + ], + "selected_reward": -0.001335205975919962 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a medium close-up of a young man wearing a dark gray hooded sweatshirt with the hood pulled low?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene lit by a single cold light source from the side, casting deep shadows on the man's face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the man perform the action of jerking his right arm back with a clenched fist and then thrusting it forward to slam into a wall mirror?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do spiderweb-like cracks spread across the mirror surface upon impact, followed by an explosion of glass shards flying in all directions?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera held still at a level angle throughout the scene?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.11291605234146118 + ], + "opposite_logprob_tokens": [ + -2.2379159927368164 + ], + "logprob_reward": -0.11291605234146118, + "yes_no_margin_reward": 2.125, + "token_credit_reward": -0.11291605234146118, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.03212935850024223, + 0.011672173626720905, + 0.04886048287153244, + 0.17296655476093292, + 0.03962205722928047, + 0.05018755421042442, + 0.5957603454589844, + 0.014928021468222141, + 0.01828128844499588, + 0.015592163428664207 + ], + "credit_map": [ + 0.0029766142833977938, + 0.0014241118915379047, + 0.003226774511858821, + 0.0017840428045019507, + 0.0021513360552489758, + 0.008932509459555149, + 0.002080485923215747, + 0.009553485549986362, + 0.0012699473882094026, + 0.0008542761788703501, + 0.0012124011991545558, + 0.0009265100234188139, + 0.0008777300827205181, + 0.002531178295612335, + 0.0011089129839092493, + 0.002891217591241002, + 0.004379368387162685, + 0.0030346896965056658, + 0.007419595029205084, + 0.0035734064877033234, + 0.002778018359094858, + 0.012272676452994347, + 0.0023805969394743443, + 0.013022132217884064, + 0.0081362659111619, + 0.005044970661401749, + 0.0032300485763698816, + 0.003540078643709421, + 0.008049612864851952, + 0.07014156877994537, + 0.00465265242382884, + 0.07017135620117188, + 0.003479100065305829, + 0.005668444558978081, + 0.0024939649738371372, + 0.003671901999041438, + 0.009088842198252678, + 0.005890917964279652, + 0.004815399646759033, + 0.004513485357165337, + 0.004723530262708664, + 0.007308002561330795, + 0.006628853734582663, + 0.004646694287657738, + 0.006544496398419142, + 0.006832224782556295, + 0.005222278647124767, + 0.008281474933028221, + 0.19453564286231995, + 0.04443846270442009, + 0.252214640378952, + 0.02400798350572586, + 0.014564214274287224, + 0.017477963119745255, + 0.011098683811724186, + 0.037422772496938705, + 0.002177325077354908, + 0.0033465565647929907, + 0.0016157734207808971, + 0.001762444619089365, + 0.0013859538594260812, + 0.0017054032068699598, + 0.0014863734832033515, + 0.0014481908874586225, + 0.0019811480306088924, + 0.0021865295711904764, + 0.002566661685705185, + 0.004050566349178553, + 0.0013668148312717676, + 0.001567212981171906, + 0.002757979091256857, + 0.001804375322535634, + 0.003727114060893655, + 0.0020240475423634052, + 0.001614655600860715, + 0.002649800619110465, + 0.0010156233329325914, + 0.00151943473611027, + 0.0014767436077818274, + 0.001564744277857244 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 3, + 5 + ], + "selected_reward": -0.11291605234146118 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "medium close-up shot from a level angle, the camera held still, captures with striking force a young man in a dark gray hooded sweatshirt standing in a dimly lit room. The sole cold light, coming from the side, casts deep shadows across his face, tightened with anger, illuminating the wall mirror before him, which reflects his own face, twisted and distorted with rage. The hood of his sweatshirt is pulled low, nearly obscuring his eyes, revealing only his furrowed brow and pursed lips. He jerks his right arm back, his clenched fist veins bulging and knuckles whitening from the exertion. Then, with a swift and decisive movement imbued with devastating force, he thrusts his fist forward, slamming it into the cold mirror. At the moment of impact, spiderweb-like cracks spread across the entire surface of the mirror before violently exploding. Countless shimmering glass shards fly explosively in all directions, tracing sharp paths through the air as if time has slowed down. His fist penetrates the center of the mirror and comes to rest in the jagged hole. The entire scene is filled with primal rage and a heart-pounding destructive catharsis, creating an atmosphere of extreme tension and despair.", + "seed": 11487, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000095/group_11/index_4418_seed_11487_dp_22_rank_44_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 11, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.09629708314314485, + "vlm_logprob": -0.09629708314314485, + "vlm_yes_no_margin": 3.4, + "vlm_token_credit": -0.09629708314314485, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a medium close-up of a young man wearing a dark gray hooded sweatshirt with the hood pulled low?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene lit by a single cold light source from the side, casting deep shadows on the man's face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the man perform the action of jerking his right arm back with a clenched fist and then thrusting it forward to slam into a wall mirror?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do spiderweb-like cracks spread across the mirror surface upon impact, followed by an explosion of glass shards flying in all directions?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera held still at a level angle throughout the scene?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.019751884043216705 + ], + "opposite_logprob_tokens": [ + -4.519752025604248 + ], + "logprob_reward": -0.019751884043216705, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.019751884043216705, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2398645430803299, + 0.05498526245355606, + 0.037402305752038956, + 0.049738746136426926, + 0.04370476305484772, + 0.0789608582854271, + 0.1201784610748291, + 0.11613410711288452, + 0.10341529548168182, + 0.1556156426668167 + ], + "credit_map": [ + 0.035655755549669266, + 0.022708864882588387, + 0.04654235020279884, + 0.03137027844786644, + 0.010175137780606747, + 0.036149151623249054, + 0.021255813539028168, + 0.03600721433758736, + 0.007636862341314554, + 0.004864302463829517, + 0.009738120250403881, + 0.005435089115053415, + 0.00276015792042017, + 0.008477561175823212, + 0.004816495347768068, + 0.01125667430460453, + 0.004866963718086481, + 0.003740204032510519, + 0.0068957069888710976, + 0.004566223826259375, + 0.003523655002936721, + 0.0049280812963843346, + 0.0032908443827182055, + 0.005590628366917372, + 0.0056699421256780624, + 0.008028830401599407, + 0.006255876272916794, + 0.007434748578816652, + 0.003224713495001197, + 0.005910922307521105, + 0.0068693808279931545, + 0.006344331428408623, + 0.0054792482405900955, + 0.00860190112143755, + 0.007924908772110939, + 0.006506241858005524, + 0.002786470577120781, + 0.0038931372109800577, + 0.004088579211384058, + 0.004424275830388069, + 0.008127843029797077, + 0.009555405005812645, + 0.014071917161345482, + 0.012847717851400375, + 0.005406991578638554, + 0.00719232065603137, + 0.011845860630273819, + 0.009912810288369656, + 0.020724380388855934, + 0.020424116402864456, + 0.016666529700160027, + 0.018029535189270973, + 0.008276279084384441, + 0.010297060012817383, + 0.012497002258896828, + 0.013263564556837082, + 0.009679682552814484, + 0.016847800463438034, + 0.016093777492642403, + 0.02140158973634243, + 0.005499507300555706, + 0.015291031450033188, + 0.009173683822154999, + 0.0221470445394516, + 0.006477018818259239, + 0.02093782089650631, + 0.011539447121322155, + 0.014463059604167938, + 0.005523278843611479, + 0.01195551548153162, + 0.01674765720963478, + 0.01577150821685791, + 0.022835714742541313, + 0.03208552300930023, + 0.02641388028860092, + 0.03209887072443962, + 0.0112627437338233, + 0.0072239357978105545, + 0.013224974274635315, + 0.010470008477568626 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 6 + ], + "selected_reward": -0.019751884043216705 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a medium close-up of a young man wearing a dark gray hooded sweatshirt with the hood pulled low?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene lit by a single cold light source from the side, casting deep shadows on the man's face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the man perform the action of jerking his right arm back with a clenched fist and then thrusting it forward to slam into a wall mirror?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do spiderweb-like cracks spread across the mirror surface upon impact, followed by an explosion of glass shards flying in all directions?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera held still at a level angle throughout the scene?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03812770918011665 + ], + "opposite_logprob_tokens": [ + -3.2881276607513428 + ], + "logprob_reward": -0.03812770918011665, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.03812770918011665, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.16136333346366882, + 0.04906798154115677, + 0.06128218024969101, + 0.28201550245285034, + 0.065859854221344, + 0.09189984202384949, + 0.06305857002735138, + 0.08175146579742432, + 0.05163370445370674, + 0.09206750243902206 + ], + "credit_map": [ + 0.02224523015320301, + 0.01922614313662052, + 0.032290101051330566, + 0.02398916520178318, + 0.010155441239476204, + 0.020040391013026237, + 0.009079484269022942, + 0.024337371811270714, + 0.005634743254631758, + 0.006918145343661308, + 0.008463731966912746, + 0.006451304070651531, + 0.004422334022819996, + 0.005818559322506189, + 0.00491667352616787, + 0.006442489102482796, + 0.008471111766994, + 0.009306645020842552, + 0.011237265542149544, + 0.007459932938218117, + 0.005379070527851582, + 0.0056517235934734344, + 0.006223069503903389, + 0.007553362753242254, + 0.031739357858896255, + 0.04657163470983505, + 0.04488778114318848, + 0.07857366651296616, + 0.016802456229925156, + 0.019086213782429695, + 0.01732487604022026, + 0.02702951803803444, + 0.008486812934279442, + 0.014354660175740719, + 0.01686992309987545, + 0.01175876334309578, + 0.002225925214588642, + 0.003321577562019229, + 0.002960518002510071, + 0.005881673190742731, + 0.009930715896189213, + 0.011640995740890503, + 0.01774509623646736, + 0.014495338313281536, + 0.006359270308166742, + 0.009318822994828224, + 0.011642653495073318, + 0.010766947641968727, + 0.010127119719982147, + 0.010200876742601395, + 0.011096617206931114, + 0.010523084551095963, + 0.0038336573634296656, + 0.0046940753236413, + 0.005901042837649584, + 0.006682096514850855, + 0.007680041715502739, + 0.012277140282094479, + 0.01279543898999691, + 0.01377325039356947, + 0.005727918352931738, + 0.010444660671055317, + 0.006658036261796951, + 0.012394979596138, + 0.004252765793353319, + 0.009168395772576332, + 0.004950862377882004, + 0.008005930110812187, + 0.0025248343590646982, + 0.0075765042565763, + 0.007411818020045757, + 0.0077425893396139145, + 0.015234338119626045, + 0.019275102764368057, + 0.018490079790353775, + 0.016321837902069092, + 0.005412726663053036, + 0.005733646918088198, + 0.006439986638724804, + 0.00515978317707777 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 0, + 9 + ], + "selected_reward": -0.03812770918011665 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a medium close-up of a young man wearing a dark gray hooded sweatshirt with the hood pulled low?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene lit by a single cold light source from the side, casting deep shadows on the man's face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the man perform the action of jerking his right arm back with a clenched fist and then thrusting it forward to slam into a wall mirror?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do spiderweb-like cracks spread across the mirror surface upon impact, followed by an explosion of glass shards flying in all directions?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera held still at a level angle throughout the scene?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.029805496335029602 + ], + "opposite_logprob_tokens": [ + -3.5298054218292236 + ], + "logprob_reward": -0.029805496335029602, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.029805496335029602, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07607851922512054, + 0.02624966949224472, + 0.025901833549141884, + 0.038217730820178986, + 0.05738942325115204, + 0.10298264026641846, + 0.08613410592079163, + 0.14420254528522491, + 0.11623861640691757, + 0.3266049325466156 + ], + "credit_map": [ + 0.013044093735516071, + 0.007553112227469683, + 0.015145760029554367, + 0.009092365391552448, + 0.0034399358555674553, + 0.006934686563909054, + 0.006500693038105965, + 0.014367873780429363, + 0.0030377963557839394, + 0.004048830363899469, + 0.004119016695767641, + 0.0035916888155043125, + 0.0012350858887657523, + 0.004164741840213537, + 0.0019288583425804973, + 0.004123649559915066, + 0.003822142258286476, + 0.004805714823305607, + 0.005341785494238138, + 0.002985849976539612, + 0.0012954220874235034, + 0.0026758103631436825, + 0.0018917134730145335, + 0.003083393443375826, + 0.00453751627355814, + 0.0045148772187530994, + 0.00596590107306838, + 0.008083033375442028, + 0.002617356600239873, + 0.003858694341033697, + 0.004081517457962036, + 0.004558833781629801, + 0.006417432334274054, + 0.011460119858384132, + 0.010700976476073265, + 0.011132411658763885, + 0.0022147325798869133, + 0.005564429797232151, + 0.002423314843326807, + 0.00747600058093667, + 0.014600636437535286, + 0.012454970739781857, + 0.02365117520093918, + 0.017782626673579216, + 0.0041913301683962345, + 0.009103669784963131, + 0.011583458632230759, + 0.009614767506718636, + 0.013363247737288475, + 0.01585797779262066, + 0.014339274726808071, + 0.01236757356673479, + 0.006813852582126856, + 0.008105169981718063, + 0.007329157553613186, + 0.007957849651575089, + 0.010744264349341393, + 0.023979192599654198, + 0.02389952354133129, + 0.028930651023983955, + 0.006936078891158104, + 0.01781492680311203, + 0.01200727466493845, + 0.01989063434302807, + 0.008666622452437878, + 0.014055348932743073, + 0.01183945033699274, + 0.014942741952836514, + 0.006561005488038063, + 0.0203203447163105, + 0.016149913892149925, + 0.023703178390860558, + 0.03994419425725937, + 0.05788332223892212, + 0.04393455758690834, + 0.10265140980482101, + 0.020185649394989014, + 0.015478665940463543, + 0.03051169216632843, + 0.016015440225601196 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 7, + 8 + ], + "selected_reward": -0.029805496335029602 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a medium close-up of a young man wearing a dark gray hooded sweatshirt with the hood pulled low?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene lit by a single cold light source from the side, casting deep shadows on the man's face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the man perform the action of jerking his right arm back with a clenched fist and then thrusting it forward to slam into a wall mirror?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do spiderweb-like cracks spread across the mirror surface upon impact, followed by an explosion of glass shards flying in all directions?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera held still at a level angle throughout the scene?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.00673500495031476 + ], + "opposite_logprob_tokens": [ + -5.006734848022461 + ], + "logprob_reward": -0.00673500495031476, + "yes_no_margin_reward": 5.0, + "token_credit_reward": -0.00673500495031476, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.054693032056093216, + 0.01231912337243557, + 0.009344646707177162, + 0.018605824559926987, + 0.0264924094080925, + 0.05670002102851868, + 0.0588466078042984, + 0.0860847532749176, + 0.05750255286693573, + 0.619411051273346 + ], + "credit_map": [ + 0.00859370268881321, + 0.006209251936525106, + 0.009765507653355598, + 0.007382133975625038, + 0.002666552783921361, + 0.005929459352046251, + 0.004445547703653574, + 0.009700874797999859, + 0.0015213157748803496, + 0.0013258948456496, + 0.001878539682365954, + 0.0015919051365926862, + 0.0006730200257152319, + 0.002243567490950227, + 0.0010054464219138026, + 0.002079433761537075, + 0.001288643223233521, + 0.0014741080813109875, + 0.0015309423906728625, + 0.0009021410951390862, + 0.0005937748937867582, + 0.001319509232416749, + 0.0009174321894533932, + 0.0013180950190871954, + 0.00198350939899683, + 0.0028235625941306353, + 0.002446442376822233, + 0.003124390495941043, + 0.0013796553248539567, + 0.002051626332104206, + 0.0026013406459242105, + 0.002195296809077263, + 0.0029905391857028008, + 0.005229485686868429, + 0.005103975534439087, + 0.0062163216061890125, + 0.0009736706269904971, + 0.002419916447252035, + 0.0012742162216454744, + 0.0022842842154204845, + 0.008223318494856358, + 0.007190347649157047, + 0.012265948578715324, + 0.008731589652597904, + 0.0023511163890361786, + 0.006371023133397102, + 0.006105560343712568, + 0.005461119115352631, + 0.00981971062719822, + 0.011502410285174847, + 0.00992647185921669, + 0.008457744494080544, + 0.0037715507205575705, + 0.004749943036586046, + 0.004683852195739746, + 0.00593491829931736, + 0.0063233813270926476, + 0.014025446958839893, + 0.01442973967641592, + 0.01749993860721588, + 0.004700068384408951, + 0.010270224884152412, + 0.007023184560239315, + 0.011812766082584858, + 0.004149165470153093, + 0.006704248953610659, + 0.005159452557563782, + 0.007102786563336849, + 0.003566733095794916, + 0.011168035678565502, + 0.0077692922204732895, + 0.011882836930453777, + 0.07598531246185303, + 0.11358103901147842, + 0.08081676065921783, + 0.19506224989891052, + 0.03867374360561371, + 0.027217067778110504, + 0.05815482512116432, + 0.029920030385255814 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 7, + 6 + ], + "selected_reward": -0.00673500495031476 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a medium close-up of a young man wearing a dark gray hooded sweatshirt with the hood pulled low?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene lit by a single cold light source from the side, casting deep shadows on the man's face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the man perform the action of jerking his right arm back with a clenched fist and then thrusting it forward to slam into a wall mirror?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do spiderweb-like cracks spread across the mirror surface upon impact, followed by an explosion of glass shards flying in all directions?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera held still at a level angle throughout the scene?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.3870653212070465 + ], + "opposite_logprob_tokens": [ + -1.1370652914047241 + ], + "logprob_reward": -0.3870653212070465, + "yes_no_margin_reward": 0.75, + "token_credit_reward": -0.3870653212070465, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.044687122106552124, + 0.04615190252661705, + 0.028435617685317993, + 0.04183784872293472, + 0.11008918285369873, + 0.13333480060100555, + 0.27451300621032715, + 0.1131863221526146, + 0.10671446472406387, + 0.10104971379041672 + ], + "credit_map": [ + 0.005752696190029383, + 0.0074109844863414764, + 0.007378744427114725, + 0.006504509132355452, + 0.00296269659884274, + 0.005385343451052904, + 0.003556713229045272, + 0.005735435057431459, + 0.005963463801890612, + 0.009132741950452328, + 0.007695562671869993, + 0.005561329890042543, + 0.002890585223212838, + 0.00608741445466876, + 0.003242517588660121, + 0.005578288342803717, + 0.003801344195380807, + 0.004635107237845659, + 0.004988549277186394, + 0.0023894638288766146, + 0.0013041524216532707, + 0.004149036947637796, + 0.002539442153647542, + 0.004628520924597979, + 0.004755201283842325, + 0.006465948186814785, + 0.006167265120893717, + 0.008153359405696392, + 0.00301635661162436, + 0.005333699751645327, + 0.003941210452467203, + 0.004004806745797396, + 0.01209950353950262, + 0.022848526015877724, + 0.01873401179909706, + 0.025709697976708412, + 0.005497887264937162, + 0.00795654021203518, + 0.006146432366222143, + 0.011096580885350704, + 0.01801867038011551, + 0.014807439409196377, + 0.03404361382126808, + 0.01956694759428501, + 0.007125661708414555, + 0.009184122085571289, + 0.014152323827147484, + 0.016436012461781502, + 0.035360485315322876, + 0.047078512609004974, + 0.04528452828526497, + 0.04595000296831131, + 0.016009846702218056, + 0.02196952886879444, + 0.026514900848269463, + 0.03634520620107651, + 0.01062641479074955, + 0.018750794231891632, + 0.021213902160525322, + 0.025704270228743553, + 0.0068147084675729275, + 0.01073681190609932, + 0.00865876767784357, + 0.010680647566914558, + 0.009368080645799637, + 0.028614938259124756, + 0.01868680864572525, + 0.015759585425257683, + 0.0052802832797169685, + 0.008712765760719776, + 0.010701609775424004, + 0.009590396657586098, + 0.010908610187470913, + 0.016202403232455254, + 0.016773344948887825, + 0.02885100618004799, + 0.006021986715495586, + 0.00638981070369482, + 0.008911756798624992, + 0.006990795489400625 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 5, + 7 + ], + "selected_reward": -0.3870653212070465 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "medium close-up shot from a level angle, the camera held still, captures with striking force a young man in a dark gray hooded sweatshirt standing in a dimly lit room. The sole cold light, coming from the side, casts deep shadows across his face, tightened with anger, illuminating the wall mirror before him, which reflects his own face, twisted and distorted with rage. The hood of his sweatshirt is pulled low, nearly obscuring his eyes, revealing only his furrowed brow and pursed lips. He jerks his right arm back, his clenched fist veins bulging and knuckles whitening from the exertion. Then, with a swift and decisive movement imbued with devastating force, he thrusts his fist forward, slamming it into the cold mirror. At the moment of impact, spiderweb-like cracks spread across the entire surface of the mirror before violently exploding. Countless shimmering glass shards fly explosively in all directions, tracing sharp paths through the air as if time has slowed down. His fist penetrates the center of the mirror and comes to rest in the jagged hole. The entire scene is filled with primal rage and a heart-pounding destructive catharsis, creating an atmosphere of extreme tension and despair.", + "seed": 11488, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000095/group_11/index_4418_seed_11488_dp_23_rank_46_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 11, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.03060411164769903, + "vlm_logprob": -0.03060411164769903, + "vlm_yes_no_margin": 4.525, + "vlm_token_credit": -0.03060411164769903, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a medium close-up of a young man wearing a dark gray hooded sweatshirt with the hood pulled low?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene lit by a single cold light source from the side, casting deep shadows on the man's face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the man perform the action of jerking his right arm back with a clenched fist and then thrusting it forward to slam into a wall mirror?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do spiderweb-like cracks spread across the mirror surface upon impact, followed by an explosion of glass shards flying in all directions?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera held still at a level angle throughout the scene?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.023624368011951447 + ], + "opposite_logprob_tokens": [ + -4.398624420166016 + ], + "logprob_reward": -0.023624368011951447, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.023624368011951447, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.34127283096313477, + 0.0872279480099678, + 0.03106522001326084, + 0.048273637890815735, + 0.10342799872159958, + 0.07437024265527725, + 0.04789680615067482, + 0.06435936689376831, + 0.09553742408752441, + 0.10656850785017014 + ], + "credit_map": [ + 0.0364992581307888, + 0.03800927475094795, + 0.0762024000287056, + 0.05838565155863762, + 0.026242947205901146, + 0.03182726353406906, + 0.034120313823223114, + 0.03998573496937752, + 0.00939072947949171, + 0.011947175487875938, + 0.01719570718705654, + 0.012361310422420502, + 0.0048673879355192184, + 0.01233160775154829, + 0.008177375420928001, + 0.010956655256450176, + 0.003465494839474559, + 0.0037873494438827038, + 0.00462940102443099, + 0.002940376987680793, + 0.001923750969581306, + 0.0059309364296495914, + 0.004474862478673458, + 0.003913047723472118, + 0.004890794400125742, + 0.007457489147782326, + 0.006517770700156689, + 0.007232966832816601, + 0.0049011907540261745, + 0.005931297317147255, + 0.006004211027175188, + 0.005337920039892197, + 0.012178216129541397, + 0.013434109278023243, + 0.018983684480190277, + 0.014293557964265347, + 0.007825372740626335, + 0.016653556376695633, + 0.006463315803557634, + 0.013596179895102978, + 0.008250637911260128, + 0.008377429097890854, + 0.01040707714855671, + 0.009479478932917118, + 0.007899974472820759, + 0.009279416874051094, + 0.012126743793487549, + 0.008549490943551064, + 0.006199895869940519, + 0.008425362408161163, + 0.007800962310284376, + 0.004704697988927364, + 0.004383634775876999, + 0.005211056210100651, + 0.0043873912654817104, + 0.006783804856240749, + 0.007660289760679007, + 0.021196726709604263, + 0.010715476237237453, + 0.007194378413259983, + 0.0036179525777697563, + 0.003363993251696229, + 0.0055883340537548065, + 0.0050222184509038925, + 0.010767894797027111, + 0.01588473841547966, + 0.013231753371655941, + 0.013990792445838451, + 0.0072549679316580296, + 0.006073988042771816, + 0.017564494162797928, + 0.010768800973892212, + 0.022633414715528488, + 0.01263299211859703, + 0.015141909010708332, + 0.01083451323211193, + 0.004611470270901918, + 0.005696744658052921, + 0.012234179303050041, + 0.022783279418945312 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 4 + ], + "selected_reward": -0.023624368011951447 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a medium close-up of a young man wearing a dark gray hooded sweatshirt with the hood pulled low?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene lit by a single cold light source from the side, casting deep shadows on the man's face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the man perform the action of jerking his right arm back with a clenched fist and then thrusting it forward to slam into a wall mirror?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do spiderweb-like cracks spread across the mirror surface upon impact, followed by an explosion of glass shards flying in all directions?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera held still at a level angle throughout the scene?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.062079768627882004 + ], + "opposite_logprob_tokens": [ + -2.812079668045044 + ], + "logprob_reward": -0.062079768627882004, + "yes_no_margin_reward": 2.75, + "token_credit_reward": -0.062079768627882004, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2824835181236267, + 0.24562761187553406, + 0.08849643915891647, + 0.08793037384748459, + 0.12083301693201065, + 0.03925732523202896, + 0.030552536249160767, + 0.027408143505454063, + 0.04019201546907425, + 0.037218984216451645 + ], + "credit_map": [ + 0.030331948772072792, + 0.04960089921951294, + 0.043982479721307755, + 0.042691268026828766, + 0.016928287222981453, + 0.03953705355525017, + 0.03222241625189781, + 0.027189170941710472, + 0.03370610252022743, + 0.02286725677549839, + 0.0691559687256813, + 0.01868443377315998, + 0.016937509179115295, + 0.031095590442419052, + 0.02148052491247654, + 0.03170021250844002, + 0.009725073352456093, + 0.009667244739830494, + 0.009716566652059555, + 0.004893563687801361, + 0.008008765056729317, + 0.022117163985967636, + 0.008565752767026424, + 0.01580231264233589, + 0.010030831210315228, + 0.013588380068540573, + 0.01482241228222847, + 0.017941806465387344, + 0.006148789543658495, + 0.007003457751125097, + 0.00789901427924633, + 0.010495687834918499, + 0.013424331322312355, + 0.015107911080121994, + 0.021893924102187157, + 0.01598074659705162, + 0.009429682977497578, + 0.017182420939207077, + 0.008949066512286663, + 0.018864937126636505, + 0.004100144375115633, + 0.0030387744773179293, + 0.0055867345072329044, + 0.006385823246091604, + 0.005708151962608099, + 0.005349328741431236, + 0.005363004747778177, + 0.003725363640114665, + 0.00331349135376513, + 0.005736020393669605, + 0.003715284401550889, + 0.004516009707003832, + 0.0029476897325366735, + 0.002986677922308445, + 0.003495739307254553, + 0.003841622034087777, + 0.0032302900217473507, + 0.007832611910998821, + 0.0048730564303696156, + 0.002838075626641512, + 0.0015185410156846046, + 0.001438052044250071, + 0.002823872957378626, + 0.002853641752153635, + 0.0037308344617486, + 0.008625456131994724, + 0.005160412285476923, + 0.006538197863847017, + 0.0024620937183499336, + 0.003015655092895031, + 0.005467846058309078, + 0.005191521253436804, + 0.004740389063954353, + 0.003391772275790572, + 0.005738640669733286, + 0.0036560026928782463, + 0.002068189438432455, + 0.0019962445367127657, + 0.006589062977582216, + 0.009038681164383888 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 4 + ], + "selected_reward": -0.062079768627882004 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a medium close-up of a young man wearing a dark gray hooded sweatshirt with the hood pulled low?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene lit by a single cold light source from the side, casting deep shadows on the man's face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the man perform the action of jerking his right arm back with a clenched fist and then thrusting it forward to slam into a wall mirror?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do spiderweb-like cracks spread across the mirror surface upon impact, followed by an explosion of glass shards flying in all directions?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera held still at a level angle throughout the scene?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004646931774914265 + ], + "opposite_logprob_tokens": [ + -5.3796467781066895 + ], + "logprob_reward": -0.004646931774914265, + "yes_no_margin_reward": 5.375, + "token_credit_reward": -0.004646931774914265, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1308128833770752, + 0.05375426635146141, + 0.05875595286488533, + 0.034841638058423996, + 0.08389841020107269, + 0.17833732068538666, + 0.10303644835948944, + 0.06759442389011383, + 0.1385640650987625, + 0.15040460228919983 + ], + "credit_map": [ + 0.013530505821108818, + 0.01695040613412857, + 0.018781181424856186, + 0.024684758856892586, + 0.00835890881717205, + 0.02033074200153351, + 0.015231368131935596, + 0.01294501218944788, + 0.005614848807454109, + 0.007689992897212505, + 0.010103816166520119, + 0.005226893350481987, + 0.005184650421142578, + 0.008424991741776466, + 0.0059034256264567375, + 0.00560564873740077, + 0.007023310754448175, + 0.007757970597594976, + 0.0070998528972268105, + 0.004680843558162451, + 0.0035612701904028654, + 0.015222598798573017, + 0.0038005521055310965, + 0.009609553031623363, + 0.0034485182259231806, + 0.005242332816123962, + 0.004733553156256676, + 0.004332996904850006, + 0.0031680092215538025, + 0.004357153549790382, + 0.004593990743160248, + 0.004965084604918957, + 0.012414495460689068, + 0.0052287522703409195, + 0.020561272278428078, + 0.013502118177711964, + 0.004314393270760775, + 0.00999537669122219, + 0.006600670516490936, + 0.011281336657702923, + 0.01953171379864216, + 0.019276978448033333, + 0.0229876060038805, + 0.024491406977176666, + 0.019580505788326263, + 0.02049928531050682, + 0.034009579569101334, + 0.01796022802591324, + 0.014284092001616955, + 0.022737646475434303, + 0.01614576019346714, + 0.013224628753960133, + 0.009484210051596165, + 0.007809160277247429, + 0.009719310328364372, + 0.009631644934415817, + 0.008380071260035038, + 0.02029595524072647, + 0.01160947885364294, + 0.006726701278239489, + 0.004215893801301718, + 0.00510833878070116, + 0.005404236260801554, + 0.005853747483342886, + 0.015159453265368938, + 0.018373047932982445, + 0.019637437537312508, + 0.01458560861647129, + 0.013926477171480656, + 0.011237852275371552, + 0.023436712101101875, + 0.0222074743360281, + 0.022527853026986122, + 0.024337004870176315, + 0.017783096060156822, + 0.01912856660783291, + 0.010390766896307468, + 0.010840367525815964, + 0.022923219949007034, + 0.02247372828423977 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 9, + 8 + ], + "selected_reward": -0.004646931774914265 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a medium close-up of a young man wearing a dark gray hooded sweatshirt with the hood pulled low?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene lit by a single cold light source from the side, casting deep shadows on the man's face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the man perform the action of jerking his right arm back with a clenched fist and then thrusting it forward to slam into a wall mirror?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do spiderweb-like cracks spread across the mirror surface upon impact, followed by an explosion of glass shards flying in all directions?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera held still at a level angle throughout the scene?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0006337541271932423 + ], + "opposite_logprob_tokens": [ + -7.375633716583252 + ], + "logprob_reward": -0.0006337541271932423, + "yes_no_margin_reward": 7.375, + "token_credit_reward": -0.0006337541271932423, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07387931644916534, + 0.05132997781038284, + 0.02166491001844406, + 0.04676990956068039, + 0.0841570496559143, + 0.0911254957318306, + 0.0827159434556961, + 0.10419680178165436, + 0.19427244365215302, + 0.249888077378273 + ], + "credit_map": [ + 0.007710675243288279, + 0.010168255306780338, + 0.01073422096669674, + 0.01281064748764038, + 0.005505180452018976, + 0.012132566422224045, + 0.00801113247871399, + 0.006806642282754183, + 0.008550229482352734, + 0.005427271593362093, + 0.017561519518494606, + 0.00426991144195199, + 0.0023030596785247326, + 0.00436897249892354, + 0.00348205235786736, + 0.005366960074752569, + 0.002770182443782687, + 0.0028395189438015223, + 0.0022965276148170233, + 0.00234091910533607, + 0.0016081065405160189, + 0.0040663485415279865, + 0.0018146939110010862, + 0.003928614314645529, + 0.0049147536046803, + 0.006236433517187834, + 0.007977714762091637, + 0.008352415636181831, + 0.003365044016391039, + 0.004210417624562979, + 0.0053091333247721195, + 0.006403997540473938, + 0.008325574919581413, + 0.01171843335032463, + 0.013895212672650814, + 0.010789898224174976, + 0.007616381160914898, + 0.009619295597076416, + 0.008046580478549004, + 0.014145676046609879, + 0.00926855020225048, + 0.011420677416026592, + 0.010820120573043823, + 0.015464221127331257, + 0.009855885989964008, + 0.00811826717108488, + 0.016268188133835793, + 0.00990959070622921, + 0.01104121096432209, + 0.011908186599612236, + 0.015181172639131546, + 0.007357289548963308, + 0.005727534182369709, + 0.007330076768994331, + 0.007855290547013283, + 0.01631518453359604, + 0.012876307591795921, + 0.03611502796411514, + 0.018691638484597206, + 0.010913229547441006, + 0.005661054514348507, + 0.005520972423255444, + 0.007298918440937996, + 0.0071196542121469975, + 0.022071892395615578, + 0.0317574068903923, + 0.0312309842556715, + 0.01748722605407238, + 0.017366770654916763, + 0.015958290547132492, + 0.030070321634411812, + 0.028329551219940186, + 0.037628211081027985, + 0.038676679134368896, + 0.03354116901755333, + 0.02806900441646576, + 0.014667261391878128, + 0.013636281713843346, + 0.03264918178319931, + 0.051020294427871704 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 8, + 7 + ], + "selected_reward": -0.0006337541271932423 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a medium close-up of a young man wearing a dark gray hooded sweatshirt with the hood pulled low?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene lit by a single cold light source from the side, casting deep shadows on the man's face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the man perform the action of jerking his right arm back with a clenched fist and then thrusting it forward to slam into a wall mirror?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do spiderweb-like cracks spread across the mirror surface upon impact, followed by an explosion of glass shards flying in all directions?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera held still at a level angle throughout the scene?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.062035735696554184 + ], + "opposite_logprob_tokens": [ + -2.8120357990264893 + ], + "logprob_reward": -0.062035735696554184, + "yes_no_margin_reward": 2.75, + "token_credit_reward": -0.062035735696554184, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.26498478651046753, + 0.06612636148929596, + 0.03803964704275131, + 0.09205500781536102, + 0.11750391870737076, + 0.11829310655593872, + 0.10234727710485458, + 0.07971115410327911, + 0.05866840481758118, + 0.06227029860019684 + ], + "credit_map": [ + 0.03674923628568649, + 0.031017877161502838, + 0.07884431630373001, + 0.027235349640250206, + 0.010753723792731762, + 0.026543155312538147, + 0.018292924389243126, + 0.035548221319913864, + 0.0066188620403409, + 0.011138172820210457, + 0.009012081660330296, + 0.012905655428767204, + 0.0059984358958899975, + 0.005238342564553022, + 0.009566254913806915, + 0.005648557096719742, + 0.00363553361967206, + 0.005198776256293058, + 0.006595394108444452, + 0.004435628652572632, + 0.0024359195958822966, + 0.006557573098689318, + 0.0050255972892045975, + 0.0041552274487912655, + 0.010290250182151794, + 0.014294938184320927, + 0.01625973917543888, + 0.02001488022506237, + 0.005551861599087715, + 0.007312134839594364, + 0.008850878104567528, + 0.009480329230427742, + 0.012967726215720177, + 0.01425245963037014, + 0.020882025361061096, + 0.01434759795665741, + 0.010386723093688488, + 0.01776559092104435, + 0.00873635895550251, + 0.01816544681787491, + 0.013767987489700317, + 0.008315413258969784, + 0.016019973903894424, + 0.01930212415754795, + 0.01261824555695057, + 0.013972924090921879, + 0.01965022087097168, + 0.014646226540207863, + 0.01550683006644249, + 0.027980836108326912, + 0.016545630991458893, + 0.008837079629302025, + 0.008553776890039444, + 0.007062303833663464, + 0.008641260676085949, + 0.009219568222761154, + 0.009067947044968605, + 0.02141754887998104, + 0.011979592964053154, + 0.011091317981481552, + 0.0067331367172300816, + 0.0061344425193965435, + 0.007110337261110544, + 0.006176843773573637, + 0.007827279157936573, + 0.010234808549284935, + 0.008098533377051353, + 0.00520500261336565, + 0.004896343220025301, + 0.005072490777820349, + 0.009191310964524746, + 0.008142641745507717, + 0.009413737803697586, + 0.006465764716267586, + 0.00999792292714119, + 0.006175690796226263, + 0.0027301996015012264, + 0.002843438647687435, + 0.008652124553918839, + 0.015991419553756714 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 5, + 4 + ], + "selected_reward": -0.062035735696554184 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "medium close-up shot from a level angle, the camera held still, captures with striking force a young man in a dark gray hooded sweatshirt standing in a dimly lit room. The sole cold light, coming from the side, casts deep shadows across his face, tightened with anger, illuminating the wall mirror before him, which reflects his own face, twisted and distorted with rage. The hood of his sweatshirt is pulled low, nearly obscuring his eyes, revealing only his furrowed brow and pursed lips. He jerks his right arm back, his clenched fist veins bulging and knuckles whitening from the exertion. Then, with a swift and decisive movement imbued with devastating force, he thrusts his fist forward, slamming it into the cold mirror. At the moment of impact, spiderweb-like cracks spread across the entire surface of the mirror before violently exploding. Countless shimmering glass shards fly explosively in all directions, tracing sharp paths through the air as if time has slowed down. His fist penetrates the center of the mirror and comes to rest in the jagged hole. The entire scene is filled with primal rage and a heart-pounding destructive catharsis, creating an atmosphere of extreme tension and despair.", + "seed": 11489, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000095/group_11/index_4418_seed_11489_dp_23_rank_46_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 11, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.04327709618955851, + "vlm_logprob": -0.04327709618955851, + "vlm_yes_no_margin": 4.15, + "vlm_token_credit": -0.04327709618955851, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a medium close-up of a young man wearing a dark gray hooded sweatshirt with the hood pulled low?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene lit by a single cold light source from the side, casting deep shadows on the man's face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the man perform the action of jerking his right arm back with a clenched fist and then thrusting it forward to slam into a wall mirror?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do spiderweb-like cracks spread across the mirror surface upon impact, followed by an explosion of glass shards flying in all directions?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera held still at a level angle throughout the scene?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.023588858544826508 + ], + "opposite_logprob_tokens": [ + -4.39858865737915 + ], + "logprob_reward": -0.023588858544826508, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.023588858544826508, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1965508908033371, + 0.12162522971630096, + 0.0708380714058876, + 0.08294183760881424, + 0.09188642352819443, + 0.10554225742816925, + 0.0921991690993309, + 0.06266288459300995, + 0.07247921079397202, + 0.10327406227588654 + ], + "credit_map": [ + 0.022677751258015633, + 0.04436364024877548, + 0.018587881699204445, + 0.024030141532421112, + 0.016544518992304802, + 0.027724910527467728, + 0.020440151914954185, + 0.02218189649283886, + 0.0246137585490942, + 0.016094021499156952, + 0.020101038739085197, + 0.024227775633335114, + 0.008156916126608849, + 0.008449247106909752, + 0.009605657309293747, + 0.010376809164881706, + 0.009432652033865452, + 0.011650623753666878, + 0.010783123783767223, + 0.007431912701576948, + 0.012108738534152508, + 0.0054470812901854515, + 0.0073209661059081554, + 0.006662972271442413, + 0.008747213520109653, + 0.012021857313811779, + 0.009678501635789871, + 0.008858826942741871, + 0.010216481052339077, + 0.008128934539854527, + 0.016678933054208755, + 0.008611084893345833, + 0.010035843588411808, + 0.017719559371471405, + 0.01250063069164753, + 0.012450498528778553, + 0.008016731590032578, + 0.009488384239375591, + 0.013854793272912502, + 0.007819974794983864, + 0.011115564964711666, + 0.01374821551144123, + 0.01706872694194317, + 0.015746671706438065, + 0.013818824663758278, + 0.009797943755984306, + 0.016802538186311722, + 0.0074437642470002174, + 0.00858970358967781, + 0.013123087584972382, + 0.01114219892770052, + 0.014703859575092793, + 0.015364847145974636, + 0.006226647179573774, + 0.01233238447457552, + 0.010716438293457031, + 0.005713446531444788, + 0.008738992735743523, + 0.010064288973808289, + 0.012324447743594646, + 0.007102152798324823, + 0.0055611152201890945, + 0.008050370030105114, + 0.005108066834509373, + 0.007031853776425123, + 0.007405850570648909, + 0.011740485206246376, + 0.008213387802243233, + 0.00952997151762247, + 0.006066795438528061, + 0.011000258848071098, + 0.011490609496831894, + 0.007967817597091198, + 0.00817179400473833, + 0.014058053493499756, + 0.01598803699016571, + 0.014491084031760693, + 0.012861267663538456, + 0.015738315880298615, + 0.01399769727140665 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 5 + ], + "selected_reward": -0.023588858544826508 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a medium close-up of a young man wearing a dark gray hooded sweatshirt with the hood pulled low?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene lit by a single cold light source from the side, casting deep shadows on the man's face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the man perform the action of jerking his right arm back with a clenched fist and then thrusting it forward to slam into a wall mirror?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do spiderweb-like cracks spread across the mirror surface upon impact, followed by an explosion of glass shards flying in all directions?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera held still at a level angle throughout the scene?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0233097355812788 + ], + "opposite_logprob_tokens": [ + -3.7733097076416016 + ], + "logprob_reward": -0.0233097355812788, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.0233097355812788, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.23200514912605286, + 0.07473628222942352, + 0.18311019241809845, + 0.13678976893424988, + 0.1585562378168106, + 0.03996213525533676, + 0.07979162037372589, + 0.03665792569518089, + 0.02840287610888481, + 0.029987813904881477 + ], + "credit_map": [ + 0.028790265321731567, + 0.05754988640546799, + 0.023246807977557182, + 0.0307744350284338, + 0.022032983601093292, + 0.025996936485171318, + 0.020419737324118614, + 0.023194093257188797, + 0.00901663675904274, + 0.013064570724964142, + 0.011738152243196964, + 0.010441705584526062, + 0.005973780993372202, + 0.011111874133348465, + 0.006092243827879429, + 0.007297318894416094, + 0.021296752616763115, + 0.01862628199160099, + 0.020701223984360695, + 0.05321359261870384, + 0.03471260890364647, + 0.008059647865593433, + 0.01262944657355547, + 0.013870631344616413, + 0.016237575560808182, + 0.030421698465943336, + 0.017774518579244614, + 0.017543435096740723, + 0.016194136813282967, + 0.009680048562586308, + 0.018759191036224365, + 0.010179169476032257, + 0.018792545422911644, + 0.029892176389694214, + 0.023269832134246826, + 0.0220656581223011, + 0.013899330981075764, + 0.010571423918008804, + 0.025420835241675377, + 0.014644436538219452, + 0.004650990478694439, + 0.008845615200698376, + 0.005941952578723431, + 0.006662315223366022, + 0.0038174125365912914, + 0.001955844461917877, + 0.005523583386093378, + 0.002564425114542246, + 0.006472915876656771, + 0.013717029243707657, + 0.01149444654583931, + 0.013407133519649506, + 0.009369276463985443, + 0.005696283653378487, + 0.01379026472568512, + 0.005844272673130035, + 0.0037868069484829903, + 0.006787692662328482, + 0.0054178074933588505, + 0.00661889323964715, + 0.003490979550406337, + 0.0034104040823876858, + 0.004579572938382626, + 0.0025657678488641977, + 0.002754228888079524, + 0.003938487730920315, + 0.0034929451067000628, + 0.0036587051581591368, + 0.004182762000709772, + 0.0030418476089835167, + 0.004058589693158865, + 0.0032753103878349066, + 0.002634812379255891, + 0.003017251845449209, + 0.004084222484380007, + 0.004367554560303688, + 0.004046299960464239, + 0.002958965953439474, + 0.004971369169652462, + 0.003907337784767151 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 4 + ], + "selected_reward": -0.0233097355812788 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a medium close-up of a young man wearing a dark gray hooded sweatshirt with the hood pulled low?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene lit by a single cold light source from the side, casting deep shadows on the man's face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the man perform the action of jerking his right arm back with a clenched fist and then thrusting it forward to slam into a wall mirror?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do spiderweb-like cracks spread across the mirror surface upon impact, followed by an explosion of glass shards flying in all directions?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera held still at a level angle throughout the scene?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0059555694460868835 + ], + "opposite_logprob_tokens": [ + -5.130955696105957 + ], + "logprob_reward": -0.0059555694460868835, + "yes_no_margin_reward": 5.125, + "token_credit_reward": -0.0059555694460868835, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07569277286529541, + 0.06761210411787033, + 0.13095170259475708, + 0.1329287737607956, + 0.18631859123706818, + 0.05988706275820732, + 0.05294797569513321, + 0.07603283226490021, + 0.1284370869398117, + 0.08919113874435425 + ], + "credit_map": [ + 0.008713915944099426, + 0.011942924000322819, + 0.010147555731236935, + 0.011083959601819515, + 0.008220424875617027, + 0.008954541757702827, + 0.00841256882995367, + 0.008216878399252892, + 0.007524602115154266, + 0.013569386675953865, + 0.012988013215363026, + 0.00759829580783844, + 0.0065101138316094875, + 0.006896223872900009, + 0.005472052376717329, + 0.007053406909108162, + 0.015322301536798477, + 0.016198933124542236, + 0.01904110796749592, + 0.03352106735110283, + 0.02322182059288025, + 0.0061311726458370686, + 0.009368723258376122, + 0.008146568201482296, + 0.012434725649654865, + 0.02053230255842209, + 0.0174246896058321, + 0.016862450167536736, + 0.015585409477353096, + 0.010313887149095535, + 0.027790557593107224, + 0.011984741315245628, + 0.017153974622488022, + 0.02268049865961075, + 0.027148935943841934, + 0.03251713141798973, + 0.026759721338748932, + 0.017617210745811462, + 0.029644694179296494, + 0.012796417810022831, + 0.007034778129309416, + 0.010833147913217545, + 0.008604095317423344, + 0.006362432613968849, + 0.007206663955003023, + 0.00545082101598382, + 0.010501117445528507, + 0.003894002875313163, + 0.004729517735540867, + 0.006445346865803003, + 0.007399560417979956, + 0.00647470960393548, + 0.010930223390460014, + 0.003645431948825717, + 0.009101148694753647, + 0.004222031682729721, + 0.005613844376057386, + 0.007720704656094313, + 0.011180330999195576, + 0.01572238840162754, + 0.010854892432689667, + 0.009167944081127644, + 0.00869341567158699, + 0.00707930326461792, + 0.010739781893789768, + 0.019762670621275902, + 0.018281830474734306, + 0.015569907613098621, + 0.016838438808918, + 0.014374226331710815, + 0.017473308369517326, + 0.015396921895444393, + 0.0066214208491146564, + 0.012435836717486382, + 0.011583127081394196, + 0.01504268404096365, + 0.011368782259523869, + 0.010450560599565506, + 0.013380630873143673, + 0.008308089338243008 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 3, + 2 + ], + "selected_reward": -0.0059555694460868835 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a medium close-up of a young man wearing a dark gray hooded sweatshirt with the hood pulled low?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene lit by a single cold light source from the side, casting deep shadows on the man's face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the man perform the action of jerking his right arm back with a clenched fist and then thrusting it forward to slam into a wall mirror?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do spiderweb-like cracks spread across the mirror surface upon impact, followed by an explosion of glass shards flying in all directions?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera held still at a level angle throughout the scene?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0031911423429846764 + ], + "opposite_logprob_tokens": [ + -5.753190994262695 + ], + "logprob_reward": -0.0031911423429846764, + "yes_no_margin_reward": 5.75, + "token_credit_reward": -0.0031911423429846764, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06266696006059647, + 0.037828411906957626, + 0.09235119074583054, + 0.04968753084540367, + 0.07408122718334198, + 0.05375233665108681, + 0.07553397864103317, + 0.1532105952501297, + 0.22378139197826385, + 0.1771063506603241 + ], + "credit_map": [ + 0.005560368299484253, + 0.008983106352388859, + 0.006377348676323891, + 0.010003361850976944, + 0.006437457632273436, + 0.009733547456562519, + 0.006535413675010204, + 0.009036360308527946, + 0.006905704736709595, + 0.005990967620164156, + 0.007182580418884754, + 0.005233835428953171, + 0.002891644136980176, + 0.0030656929593533278, + 0.003338090144097805, + 0.0032198959961533546, + 0.009748106822371483, + 0.0097071947529912, + 0.011186914518475533, + 0.029291892424225807, + 0.019085295498371124, + 0.0037588721606880426, + 0.005435426253825426, + 0.004137478303164244, + 0.006404338404536247, + 0.007257435936480761, + 0.0056406911462545395, + 0.007197135593742132, + 0.004357376601547003, + 0.0046966117806732655, + 0.009576615877449512, + 0.004557330626994371, + 0.0074914307333528996, + 0.009419362992048264, + 0.009847461245954037, + 0.012220494449138641, + 0.007114754058420658, + 0.007694552186876535, + 0.014413198456168175, + 0.005879977252334356, + 0.005511170253157616, + 0.009854141622781754, + 0.007937227375805378, + 0.005656470078974962, + 0.006085618864744902, + 0.005086787510663271, + 0.010108324699103832, + 0.0035125992726534605, + 0.007433896418660879, + 0.013584733940660954, + 0.009255822747945786, + 0.010387486778199673, + 0.009727658703923225, + 0.006430085748434067, + 0.010927947238087654, + 0.0077863517217338085, + 0.011799768544733524, + 0.02934943325817585, + 0.023367157205939293, + 0.02548454888164997, + 0.014704185537993908, + 0.011929018422961235, + 0.025465818122029305, + 0.011110660620033741, + 0.01839112676680088, + 0.03268920257687569, + 0.02843008004128933, + 0.0264972522854805, + 0.03444128483533859, + 0.02621392346918583, + 0.028761930763721466, + 0.02835659682750702, + 0.013700693845748901, + 0.019946882501244545, + 0.021795181557536125, + 0.030153091996908188, + 0.022877316921949387, + 0.02205488830804825, + 0.025431634858250618, + 0.021146655082702637 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 9, + 7 + ], + "selected_reward": -0.0031911423429846764 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a medium close-up of a young man wearing a dark gray hooded sweatshirt with the hood pulled low?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene lit by a single cold light source from the side, casting deep shadows on the man's face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the man perform the action of jerking his right arm back with a clenched fist and then thrusting it forward to slam into a wall mirror?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do spiderweb-like cracks spread across the mirror surface upon impact, followed by an explosion of glass shards flying in all directions?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera held still at a level angle throughout the scene?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.16034017503261566 + ], + "opposite_logprob_tokens": [ + -1.9103401899337769 + ], + "logprob_reward": -0.16034017503261566, + "yes_no_margin_reward": 1.75, + "token_credit_reward": -0.16034017503261566, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08095842599868774, + 0.06937283277511597, + 0.16994476318359375, + 0.12582479417324066, + 0.07544391602277756, + 0.12174711376428604, + 0.1001027524471283, + 0.11531359702348709, + 0.06965582072734833, + 0.07163593173027039 + ], + "credit_map": [ + 0.007934856228530407, + 0.010344387963414192, + 0.007853739894926548, + 0.013445699587464333, + 0.007005906663835049, + 0.014482872560620308, + 0.009052783250808716, + 0.010838175192475319, + 0.008889621123671532, + 0.01323544979095459, + 0.012506040744483471, + 0.008775978349149227, + 0.005783585365861654, + 0.00704219751060009, + 0.0061843739822506905, + 0.006955590099096298, + 0.01540073100477457, + 0.019045459106564522, + 0.0161381084471941, + 0.05652802810072899, + 0.037678297609090805, + 0.007634398061782122, + 0.010265633463859558, + 0.007254111580550671, + 0.01042810920625925, + 0.020991723984479904, + 0.014792836271226406, + 0.01442486047744751, + 0.017762869596481323, + 0.010543103329837322, + 0.02419900894165039, + 0.012682282365858555, + 0.009477948769927025, + 0.008606178686022758, + 0.010428844951093197, + 0.012241696938872337, + 0.006893572863191366, + 0.006939748302102089, + 0.013405109755694866, + 0.0074508171528577805, + 0.015058235265314579, + 0.020334109663963318, + 0.01715194806456566, + 0.025755489245057106, + 0.010455786250531673, + 0.006507940590381622, + 0.017797816544771194, + 0.008685785345733166, + 0.009337526746094227, + 0.01549911592155695, + 0.012306408025324345, + 0.01863405853509903, + 0.01674719713628292, + 0.008214141242206097, + 0.011747694574296474, + 0.007616610266268253, + 0.01067604124546051, + 0.017529992386698723, + 0.01896819844841957, + 0.02640833705663681, + 0.011724314652383327, + 0.010033450089395046, + 0.012336352840065956, + 0.007636914029717445, + 0.007440168410539627, + 0.010003747418522835, + 0.009358581155538559, + 0.012384050525724888, + 0.009960361756384373, + 0.00548186618834734, + 0.009062297642230988, + 0.005964753683656454, + 0.006318503525108099, + 0.008099155500531197, + 0.009635879658162594, + 0.009518532082438469, + 0.009996670298278332, + 0.007767587900161743, + 0.01094570942223072, + 0.009353889152407646 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 3, + 5 + ], + "selected_reward": -0.16034017503261566 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000095/group_12/index_tensor([1273, 1273])_group_12_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000095/group_12/index_tensor([1273, 1273])_group_12_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..70c42930d5c6f5e6ebfe229ba0dded9b60966418 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000095/group_12/index_tensor([1273, 1273])_group_12_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "In a high-contrast black-and-white image, an extreme close-up slowly tilts downward, tracking a pearl that has just escaped from a broken necklace string. Illuminated by a single, intense sidelight, the pearl's surface reveals a rich gradation of colors, from bright silver to dark gray, shimmering with a warm sheen. It slowly tumbles down, finally landing on a dark, finely textured velvet fabric, its forward motion steady and graceful. The shot instantly shifts to an extreme macro perspective, the camera remaining still, focused on the resting pearl. Reflected on its mirror-like surface is a fragment of a woman's face—a wide-open eye, its outline strikingly sharp in the black and white, with a single, glistening tear seemingly lingering in its corner. The entire image, imbued with exquisite detail, exudes a tranquil atmosphere of loss and heartbreak.", + "seed": 12255, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000095/group_12/index_1273_seed_12255_dp_24_rank_48_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 12, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.027827679179608822, + "vlm_logprob": -0.027827679179608822, + "vlm_yes_no_margin": 4.025, + "vlm_token_credit": -0.027827679179608822, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video presented in a high-contrast black-and-white style?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the video begin with an extreme close-up of a pearl that has escaped from a broken necklace string?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the pearl slowly tumble down and land on a dark, finely textured velvet fabric?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the video show a reflection of a woman's face with a wide-open eye and a glistening tear on the pearl's surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start by tilting downward to track the pearl and then shift to a still extreme macro perspective?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01928052306175232 + ], + "opposite_logprob_tokens": [ + -5.019280433654785 + ], + "logprob_reward": -0.01928052306175232, + "yes_no_margin_reward": 5.0, + "token_credit_reward": -0.01928052306175232, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.27538248896598816, + 0.05035768821835518, + 0.065044105052948, + 0.05692233145236969, + 0.15287528932094574, + 0.024783911183476448, + 0.04772886633872986, + 0.1567649245262146, + 0.15309788286685944, + 0.01704249531030655 + ], + "credit_map": [ + 0.054348174482584, + 0.017680978402495384, + 0.028669869527220726, + 0.10336380451917648, + 0.01443711668252945, + 0.016230858862400055, + 0.014282271265983582, + 0.026369435712695122, + 0.010205105878412724, + 0.004748062696307898, + 0.0064403037540614605, + 0.0068267472088336945, + 0.006258601788431406, + 0.005195849575102329, + 0.005710503086447716, + 0.004972516093403101, + 0.0075716059654951096, + 0.007723226211965084, + 0.004984589293599129, + 0.007480834145098925, + 0.015773337334394455, + 0.007171256002038717, + 0.006734794937074184, + 0.007604467216879129, + 0.004959520883858204, + 0.004355477634817362, + 0.010165422223508358, + 0.006406813394278288, + 0.006546128541231155, + 0.009172389283776283, + 0.007513437420129776, + 0.007803148124366999, + 0.0165560245513916, + 0.006943398155272007, + 0.030421318486332893, + 0.01375085674226284, + 0.00904447864741087, + 0.023839274421334267, + 0.029095876961946487, + 0.023224065080285072, + 0.00367718655616045, + 0.0018805337604135275, + 0.0038250272627919912, + 0.0023012813180685043, + 0.002281289780512452, + 0.0025182992685586214, + 0.0036162706092000008, + 0.004684023093432188, + 0.005999016109853983, + 0.004003019537776709, + 0.0059915524907410145, + 0.006670279894024134, + 0.0030448792967945337, + 0.005936602130532265, + 0.008285676129162312, + 0.007797842845320702, + 0.021931052207946777, + 0.008933155797421932, + 0.03310374915599823, + 0.018417106941342354, + 0.012062042020261288, + 0.016985345631837845, + 0.020855238661170006, + 0.024477234110236168, + 0.021427873522043228, + 0.007988260127604008, + 0.028889009729027748, + 0.016832076013088226, + 0.01049593836069107, + 0.018131902441382408, + 0.02328539825975895, + 0.026047442108392715, + 0.0020023768302053213, + 0.0009545187349431217, + 0.002225888427346945, + 0.0016176074277609587, + 0.0015327280852943659, + 0.002494493965059519, + 0.002896317047998309, + 0.003318564733490348 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 7, + 8 + ], + "selected_reward": -0.01928052306175232 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video presented in a high-contrast black-and-white style?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the video begin with an extreme close-up of a pearl that has escaped from a broken necklace string?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the pearl slowly tumble down and land on a dark, finely textured velvet fabric?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the video show a reflection of a woman's face with a wide-open eye and a glistening tear on the pearl's surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start by tilting downward to track the pearl and then shift to a still extreme macro perspective?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.054965242743492126 + ], + "opposite_logprob_tokens": [ + -2.9299652576446533 + ], + "logprob_reward": -0.054965242743492126, + "yes_no_margin_reward": 2.875, + "token_credit_reward": -0.054965242743492126, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.22441509366035461, + 0.04730600118637085, + 0.1636986881494522, + 0.2755894958972931, + 0.10381651669740677, + 0.054590318351984024, + 0.05409702658653259, + 0.02446771040558815, + 0.042446643114089966, + 0.009572435170412064 + ], + "credit_map": [ + 0.040055058896541595, + 0.01775071956217289, + 0.01956041157245636, + 0.06212066486477852, + 0.01623186282813549, + 0.01885855384171009, + 0.02795601822435856, + 0.021881811320781708, + 0.007581575307995081, + 0.00559878908097744, + 0.005713639780879021, + 0.005751952063292265, + 0.005811725277453661, + 0.006100813392549753, + 0.0056615713983774185, + 0.005085932090878487, + 0.023720845580101013, + 0.03675833344459534, + 0.010024026967585087, + 0.01002313569188118, + 0.017419740557670593, + 0.02069089561700821, + 0.01134499255567789, + 0.03371672332286835, + 0.025304803624749184, + 0.021242426708340645, + 0.054791562259197235, + 0.020822692662477493, + 0.03878912702202797, + 0.04549446329474449, + 0.02820640429854393, + 0.04093803092837334, + 0.007999558001756668, + 0.00610584719106555, + 0.01420829351991415, + 0.011438548564910889, + 0.009113732725381851, + 0.023765716701745987, + 0.01816968061029911, + 0.013015133328735828, + 0.004583459813147783, + 0.0035906205885112286, + 0.007242973428219557, + 0.00853451993316412, + 0.006082974374294281, + 0.007033035159111023, + 0.008806915022432804, + 0.008715823292732239, + 0.00672245305031538, + 0.005881417542695999, + 0.00791983399540186, + 0.006156951654702425, + 0.0050742169842123985, + 0.0060368869453668594, + 0.007723116781562567, + 0.00858214870095253, + 0.0031927309464663267, + 0.0019357803976163268, + 0.004953416530042887, + 0.002624319866299629, + 0.0017358572222292423, + 0.0028609021101146936, + 0.003728665877133608, + 0.0034360375721007586, + 0.00559474341571331, + 0.0023590235505253077, + 0.0071724471636116505, + 0.0047363899648189545, + 0.0038270759396255016, + 0.005720924586057663, + 0.006309560965746641, + 0.006726477760821581, + 0.0014030573656782508, + 0.0005961815477348864, + 0.0014092035125941038, + 0.0010965438559651375, + 0.0008276873850263655, + 0.0010591555619612336, + 0.0017984065925702453, + 0.0013821993488818407 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 0, + 2 + ], + "selected_reward": -0.054965242743492126 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video presented in a high-contrast black-and-white style?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the video begin with an extreme close-up of a pearl that has escaped from a broken necklace string?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the pearl slowly tumble down and land on a dark, finely textured velvet fabric?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the video show a reflection of a woman's face with a wide-open eye and a glistening tear on the pearl's surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start by tilting downward to track the pearl and then shift to a still extreme macro perspective?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.007632262073457241 + ], + "opposite_logprob_tokens": [ + -4.882632255554199 + ], + "logprob_reward": -0.007632262073457241, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.007632262073457241, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11256846785545349, + 0.022899160161614418, + 0.12239716202020645, + 0.47563669085502625, + 0.05277015641331673, + 0.08668854087591171, + 0.055964142084121704, + 0.03092392347753048, + 0.02243843488395214, + 0.017713375389575958 + ], + "credit_map": [ + 0.022055769339203835, + 0.008553067222237587, + 0.010992755182087421, + 0.036013197153806686, + 0.006341846194118261, + 0.009511717595160007, + 0.008268746547400951, + 0.010831351391971111, + 0.002670956775546074, + 0.002339918166399002, + 0.0034233585465699434, + 0.003128637792542577, + 0.003300540382042527, + 0.0021378775127232075, + 0.003608734579756856, + 0.002289134543389082, + 0.018045080825686455, + 0.025702303275465965, + 0.007862389087677002, + 0.009235613979399204, + 0.013815225102007389, + 0.014839112758636475, + 0.010499898344278336, + 0.02239752747118473, + 0.037686675786972046, + 0.0245628971606493, + 0.09611418098211288, + 0.03541205823421478, + 0.04224998876452446, + 0.08090778440237045, + 0.058141760528087616, + 0.10056132078170776, + 0.005538935773074627, + 0.0028654986526817083, + 0.013037743978202343, + 0.005805261433124542, + 0.003684085560962558, + 0.008265819400548935, + 0.00633647944778204, + 0.007236327510327101, + 0.008558754809200764, + 0.00637780549004674, + 0.013353989459574223, + 0.011684684082865715, + 0.011454164050519466, + 0.010230239480733871, + 0.012803791090846062, + 0.012225111946463585, + 0.0062394822016358376, + 0.005991608370095491, + 0.0077559384517371655, + 0.006654549390077591, + 0.004753768444061279, + 0.006765271537005901, + 0.0085799191147089, + 0.00922360084950924, + 0.004470253828912973, + 0.0024962713941931725, + 0.004220670089125633, + 0.003432498313486576, + 0.0026018100325018167, + 0.003569110296666622, + 0.005070185288786888, + 0.005063122604042292, + 0.002409760607406497, + 0.0011035375064238906, + 0.0024077880661934614, + 0.0034282261040061712, + 0.0018338701920583844, + 0.002775624394416809, + 0.005529495421797037, + 0.0029501335229724646, + 0.002351186703890562, + 0.0007563078543171287, + 0.0030322798993438482, + 0.002259825821965933, + 0.0019322580192238092, + 0.0021313244942575693, + 0.0028679866809397936, + 0.0023822037037461996 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 2, + 0 + ], + "selected_reward": -0.007632262073457241 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video presented in a high-contrast black-and-white style?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the video begin with an extreme close-up of a pearl that has escaped from a broken necklace string?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the pearl slowly tumble down and land on a dark, finely textured velvet fabric?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the video show a reflection of a woman's face with a wide-open eye and a glistening tear on the pearl's surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start by tilting downward to track the pearl and then shift to a still extreme macro perspective?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.014201764017343521 + ], + "opposite_logprob_tokens": [ + -4.264201641082764 + ], + "logprob_reward": -0.014201764017343521, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.014201764017343521, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.05820437893271446, + 0.04532173275947571, + 0.04623675346374512, + 0.615587055683136, + 0.06005654111504555, + 0.045568935573101044, + 0.04234972596168518, + 0.02506186068058014, + 0.042338885366916656, + 0.01927408203482628 + ], + "credit_map": [ + 0.007011656649410725, + 0.0063750226981937885, + 0.008731530047953129, + 0.009645191952586174, + 0.00469384016469121, + 0.009247669950127602, + 0.00548521475866437, + 0.007014255505055189, + 0.006041429936885834, + 0.005146142560988665, + 0.005398284178227186, + 0.004941399674862623, + 0.00654159439727664, + 0.005292010027915239, + 0.004704013001173735, + 0.007256862707436085, + 0.005891139153391123, + 0.0070044551976025105, + 0.005360755603760481, + 0.005676668602973223, + 0.006534474901854992, + 0.005536537617444992, + 0.005456349812448025, + 0.00477637117728591, + 0.05574362352490425, + 0.04791778698563576, + 0.1230005994439125, + 0.047412995249032974, + 0.08249334990978241, + 0.10050574690103531, + 0.06335753202438354, + 0.09515547752380371, + 0.005581773351877928, + 0.005662983283400536, + 0.008872563019394875, + 0.009537489153444767, + 0.006083328742533922, + 0.0077894157730042934, + 0.007249704096466303, + 0.009279284626245499, + 0.004918953869491816, + 0.0036004893481731415, + 0.005719853099435568, + 0.006069113966077566, + 0.007037242408841848, + 0.005305887199938297, + 0.0072041102685034275, + 0.005713285878300667, + 0.005163911264389753, + 0.004308713600039482, + 0.006039861124008894, + 0.00480692507699132, + 0.003542394144460559, + 0.0050922054797410965, + 0.006427216809242964, + 0.006968499626964331, + 0.004208895843476057, + 0.0020336699672043324, + 0.003126448253169656, + 0.0022528888657689095, + 0.002054340671747923, + 0.002700410084798932, + 0.004251158330589533, + 0.004434049129486084, + 0.006442488636821508, + 0.002074637683108449, + 0.008353617042303085, + 0.004639005288481712, + 0.0038459962233901024, + 0.004730554297566414, + 0.005383852403610945, + 0.006868733558803797, + 0.0030542807653546333, + 0.0009254114702343941, + 0.003215635661035776, + 0.0021259086206555367, + 0.0023413740564137697, + 0.0020648781210184097, + 0.0027605206705629826, + 0.002786073600873351 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 4, + 0 + ], + "selected_reward": -0.014201764017343521 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video presented in a high-contrast black-and-white style?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the video begin with an extreme close-up of a pearl that has escaped from a broken necklace string?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the pearl slowly tumble down and land on a dark, finely textured velvet fabric?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the video show a reflection of a woman's face with a wide-open eye and a glistening tear on the pearl's surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start by tilting downward to track the pearl and then shift to a still extreme macro perspective?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0430586040019989 + ], + "opposite_logprob_tokens": [ + -3.1680586338043213 + ], + "logprob_reward": -0.0430586040019989, + "yes_no_margin_reward": 3.125, + "token_credit_reward": -0.0430586040019989, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08862430602312088, + 0.052024513483047485, + 0.12935250997543335, + 0.40102753043174744, + 0.0878957137465477, + 0.0340915322303772, + 0.06063562259078026, + 0.041583120822906494, + 0.07629067450761795, + 0.028474492952227592 + ], + "credit_map": [ + 0.01276727207005024, + 0.007027280051261187, + 0.009314950555562973, + 0.025269903242588043, + 0.006391322240233421, + 0.009407027624547482, + 0.010835755616426468, + 0.007610791362822056, + 0.006543219555169344, + 0.0051095252856612206, + 0.0062949163839221, + 0.005771127063781023, + 0.006377135403454304, + 0.00617982866242528, + 0.0069137560203671455, + 0.008835002779960632, + 0.01817813143134117, + 0.022668680176138878, + 0.014172986149787903, + 0.014250931330025196, + 0.01690838485956192, + 0.01646101102232933, + 0.010634049773216248, + 0.01607833057641983, + 0.03411875292658806, + 0.032240234315395355, + 0.08666577190160751, + 0.028957964852452278, + 0.05208442360162735, + 0.06857474148273468, + 0.0397222638130188, + 0.05866334214806557, + 0.008069328963756561, + 0.004636378027498722, + 0.016363702714443207, + 0.010625568218529224, + 0.00707557238638401, + 0.018128233030438423, + 0.012049466371536255, + 0.0109474565833807, + 0.005267361644655466, + 0.0033862944692373276, + 0.004153267480432987, + 0.0034274146892130375, + 0.0038660564459860325, + 0.0043904418125748634, + 0.005071501713246107, + 0.004529192112386227, + 0.007536378689110279, + 0.006645663175731897, + 0.007715095765888691, + 0.006918321829289198, + 0.0039043105207383633, + 0.008052179589867592, + 0.010201907716691494, + 0.009661763906478882, + 0.0065866210497915745, + 0.003560299053788185, + 0.004811359569430351, + 0.004229946527630091, + 0.003450850024819374, + 0.00458444794639945, + 0.006956304889172316, + 0.007403285708278418, + 0.010470808483660221, + 0.003769821021705866, + 0.013725657016038895, + 0.008109032176434994, + 0.006335429381579161, + 0.00963917002081871, + 0.010945119895040989, + 0.013295631855726242, + 0.004230997059494257, + 0.0012984378263354301, + 0.004364570137113333, + 0.0028435704298317432, + 0.0028004921041429043, + 0.0032238371204584837, + 0.005089126527309418, + 0.004623458255082369 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 2, + 0 + ], + "selected_reward": -0.0430586040019989 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "In a high-contrast black-and-white image, an extreme close-up slowly tilts downward, tracking a pearl that has just escaped from a broken necklace string. Illuminated by a single, intense sidelight, the pearl's surface reveals a rich gradation of colors, from bright silver to dark gray, shimmering with a warm sheen. It slowly tumbles down, finally landing on a dark, finely textured velvet fabric, its forward motion steady and graceful. The shot instantly shifts to an extreme macro perspective, the camera remaining still, focused on the resting pearl. Reflected on its mirror-like surface is a fragment of a woman's face—a wide-open eye, its outline strikingly sharp in the black and white, with a single, glistening tear seemingly lingering in its corner. The entire image, imbued with exquisite detail, exudes a tranquil atmosphere of loss and heartbreak.", + "seed": 12256, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000095/group_12/index_1273_seed_12256_dp_24_rank_48_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 12, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.02124470826238394, + "vlm_logprob": -0.02124470826238394, + "vlm_yes_no_margin": 4.275, + "vlm_token_credit": -0.02124470826238394, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video presented in a high-contrast black-and-white style?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the video begin with an extreme close-up of a pearl that has escaped from a broken necklace string?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the pearl slowly tumble down and land on a dark, finely textured velvet fabric?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the video show a reflection of a woman's face with a wide-open eye and a glistening tear on the pearl's surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start by tilting downward to track the pearl and then shift to a still extreme macro perspective?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01925468072295189 + ], + "opposite_logprob_tokens": [ + -5.019254684448242 + ], + "logprob_reward": -0.01925468072295189, + "yes_no_margin_reward": 5.0, + "token_credit_reward": -0.01925468072295189, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15641798079013824, + 0.06528152525424957, + 0.09411108493804932, + 0.06442995369434357, + 0.21319884061813354, + 0.056136030703783035, + 0.07620229572057724, + 0.20057450234889984, + 0.02618240751326084, + 0.04746532067656517 + ], + "credit_map": [ + 0.02148772031068802, + 0.027207812294363976, + 0.020670918747782707, + 0.018911685794591904, + 0.015905143693089485, + 0.018378453329205513, + 0.017074335366487503, + 0.016781920567154884, + 0.0056907981634140015, + 0.007406785152852535, + 0.009953715838491917, + 0.006503873039036989, + 0.012871586717665195, + 0.00952998362481594, + 0.004500660579651594, + 0.008824124000966549, + 0.006697442382574081, + 0.010376699268817902, + 0.013922572135925293, + 0.018492599949240685, + 0.007273445837199688, + 0.012997073121368885, + 0.013111944310367107, + 0.011239315383136272, + 0.005074577871710062, + 0.00473783491179347, + 0.009142165072262287, + 0.009136758744716644, + 0.00621612137183547, + 0.007843871600925922, + 0.012986878864467144, + 0.009291747584939003, + 0.009738794527947903, + 0.009418170899152756, + 0.014999819919466972, + 0.014287377707660198, + 0.013510060496628284, + 0.06707008183002472, + 0.017227863892912865, + 0.06694667041301727, + 0.004198731854557991, + 0.00433863652870059, + 0.0069160605780780315, + 0.004459411837160587, + 0.006501301657408476, + 0.012140185572206974, + 0.006435905117541552, + 0.011145802214741707, + 0.005759101826697588, + 0.004452617838978767, + 0.010736432857811451, + 0.008234561420977116, + 0.008558571338653564, + 0.009187531657516956, + 0.01919313706457615, + 0.010080353356897831, + 0.009013589471578598, + 0.007875747047364712, + 0.014345059171319008, + 0.015375367365777493, + 0.011520319618284702, + 0.05684155225753784, + 0.032949771732091904, + 0.05265311151742935, + 0.0027079612482339144, + 0.0015981538454070687, + 0.003892043372616172, + 0.002378220437094569, + 0.002478200476616621, + 0.004846958443522453, + 0.003860919736325741, + 0.004419952165335417, + 0.003500264836475253, + 0.0029785826336592436, + 0.007382170297205448, + 0.003682599635794759, + 0.0060936277732253075, + 0.009049798361957073, + 0.007121057715266943, + 0.00765722431242466 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 7, + 0 + ], + "selected_reward": -0.01925468072295189 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video presented in a high-contrast black-and-white style?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the video begin with an extreme close-up of a pearl that has escaped from a broken necklace string?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the pearl slowly tumble down and land on a dark, finely textured velvet fabric?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the video show a reflection of a woman's face with a wide-open eye and a glistening tear on the pearl's surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start by tilting downward to track the pearl and then shift to a still extreme macro perspective?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0337153896689415 + ], + "opposite_logprob_tokens": [ + -3.4087154865264893 + ], + "logprob_reward": -0.0337153896689415, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.0337153896689415, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3033972382545471, + 0.02165343053638935, + 0.03536722809076309, + 0.09330086410045624, + 0.11772345751523972, + 0.2813241183757782, + 0.05769886448979378, + 0.0663984939455986, + 0.011733144521713257, + 0.011403114534914494 + ], + "credit_map": [ + 0.03963194414973259, + 0.03225189074873924, + 0.044372253119945526, + 0.03571748360991478, + 0.03294777870178223, + 0.04030546173453331, + 0.03564294055104256, + 0.04252750426530838, + 0.0017792277503758669, + 0.001964415656402707, + 0.002465899335220456, + 0.0033053995575755835, + 0.0034165363758802414, + 0.0034437195863574743, + 0.002554166130721569, + 0.0027240689378231764, + 0.003209351561963558, + 0.004628823604434729, + 0.006027401890605688, + 0.004253325518220663, + 0.003919288981705904, + 0.005076861009001732, + 0.003744580550119281, + 0.004507598467171192, + 0.010331901721656322, + 0.006519156973809004, + 0.01832151971757412, + 0.01199326291680336, + 0.007975668646395206, + 0.01266427431255579, + 0.012473298236727715, + 0.013021792285144329, + 0.005881594493985176, + 0.004398548975586891, + 0.008613604120910168, + 0.01216137781739235, + 0.006466319318860769, + 0.03524880111217499, + 0.008003694005310535, + 0.036949533969163895, + 0.014540853910148144, + 0.009676704183220863, + 0.017655765637755394, + 0.016571978107094765, + 0.01349304337054491, + 0.0912361741065979, + 0.027362562716007233, + 0.09078707545995712, + 0.005621614400297403, + 0.005272070411592722, + 0.009778901934623718, + 0.006661114748567343, + 0.0064569623209536076, + 0.008520160801708698, + 0.008598641492426395, + 0.00678940350189805, + 0.0036594332195818424, + 0.0034040927421301603, + 0.006287125404924154, + 0.006881730165332556, + 0.005700964946299791, + 0.014098898507654667, + 0.014255429618060589, + 0.012110833078622818, + 0.0008265110664069653, + 0.0004741919692605734, + 0.001514630508609116, + 0.0006458757561631501, + 0.0006948542431928217, + 0.0032306998036801815, + 0.001276397262699902, + 0.0030699847266077995, + 0.0008283656788989902, + 0.000749842613004148, + 0.0016030088299885392, + 0.0008779683266766369, + 0.0011861507082358003, + 0.0024347223807126284, + 0.0015986459329724312, + 0.0021244122181087732 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 5, + 4 + ], + "selected_reward": -0.0337153896689415 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video presented in a high-contrast black-and-white style?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the video begin with an extreme close-up of a pearl that has escaped from a broken necklace string?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the pearl slowly tumble down and land on a dark, finely textured velvet fabric?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the video show a reflection of a woman's face with a wide-open eye and a glistening tear on the pearl's surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start by tilting downward to track the pearl and then shift to a still extreme macro perspective?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0036187181249260902 + ], + "opposite_logprob_tokens": [ + -5.6286187171936035 + ], + "logprob_reward": -0.0036187181249260902, + "yes_no_margin_reward": 5.625, + "token_credit_reward": -0.0036187181249260902, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08180776983499527, + 0.02915601246058941, + 0.057345692068338394, + 0.09648153185844421, + 0.13036783039569855, + 0.19578269124031067, + 0.09069773554801941, + 0.26021456718444824, + 0.03147132322192192, + 0.02667488157749176 + ], + "credit_map": [ + 0.012315746396780014, + 0.010547345504164696, + 0.01039961539208889, + 0.010240829549729824, + 0.00894985068589449, + 0.012416421435773373, + 0.005949043668806553, + 0.010988910682499409, + 0.003154134377837181, + 0.002953205956146121, + 0.004440017510205507, + 0.0032636327669024467, + 0.004797305911779404, + 0.0036756612826138735, + 0.0027180302422493696, + 0.004154023714363575, + 0.003945069387555122, + 0.007023575250059366, + 0.00689840828999877, + 0.010475109331309795, + 0.006334679201245308, + 0.005949078593403101, + 0.009210594929754734, + 0.007509179878979921, + 0.008874330669641495, + 0.008102066814899445, + 0.023437129333615303, + 0.011176719330251217, + 0.008494390174746513, + 0.011976810172200203, + 0.011885242536664009, + 0.01253484282642603, + 0.006756698712706566, + 0.006969100795686245, + 0.012653435580432415, + 0.013657700270414352, + 0.011139300651848316, + 0.03247059881687164, + 0.016466762870550156, + 0.030254242941737175, + 0.007938570342957973, + 0.004777016118168831, + 0.009479278698563576, + 0.007126715965569019, + 0.008825834840536118, + 0.0719204694032669, + 0.015456066466867924, + 0.0702587366104126, + 0.005954505875706673, + 0.005213250406086445, + 0.014059129171073437, + 0.008838001638650894, + 0.011153044179081917, + 0.01392699871212244, + 0.02080458030104637, + 0.010748221538960934, + 0.011740236543118954, + 0.0067681120708584785, + 0.02104102075099945, + 0.015471662394702435, + 0.009922908619046211, + 0.08557750284671783, + 0.02992606721818447, + 0.07976703345775604, + 0.001814508344978094, + 0.0013803367037326097, + 0.003964491654187441, + 0.001710101612843573, + 0.0023303122725337744, + 0.00865721795707941, + 0.0038837732281535864, + 0.007730584125965834, + 0.002022524131461978, + 0.0015053239185363054, + 0.005178773310035467, + 0.0021127245854586363, + 0.0027599437162280083, + 0.004927966743707657, + 0.003794832155108452, + 0.004372792784124613 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 5, + 4 + ], + "selected_reward": -0.0036187181249260902 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video presented in a high-contrast black-and-white style?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the video begin with an extreme close-up of a pearl that has escaped from a broken necklace string?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the pearl slowly tumble down and land on a dark, finely textured velvet fabric?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the video show a reflection of a woman's face with a wide-open eye and a glistening tear on the pearl's surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start by tilting downward to track the pearl and then shift to a still extreme macro perspective?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02329447865486145 + ], + "opposite_logprob_tokens": [ + -3.773294448852539 + ], + "logprob_reward": -0.02329447865486145, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.02329447865486145, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07014582306146622, + 0.010221223346889019, + 0.026625435799360275, + 0.024977631866931915, + 0.2320169359445572, + 0.5090565085411072, + 0.03839980810880661, + 0.07079771161079407, + 0.009657740592956543, + 0.008101162500679493 + ], + "credit_map": [ + 0.010317172855138779, + 0.009572722017765045, + 0.00883723609149456, + 0.010325581766664982, + 0.01101979985833168, + 0.007906200364232063, + 0.005503547843545675, + 0.006663569249212742, + 0.0013012003619223833, + 0.0007976041524671018, + 0.001961505739018321, + 0.0013072255533188581, + 0.0010458760662004352, + 0.001168634626083076, + 0.0014315600274130702, + 0.001207617111504078, + 0.001958723645657301, + 0.0028449278324842453, + 0.004248180892318487, + 0.0038318848237395287, + 0.0027191555127501488, + 0.004341600462794304, + 0.0028692935593426228, + 0.0038116713985800743, + 0.002049189293757081, + 0.0027206013910472393, + 0.004665802698582411, + 0.004225117154419422, + 0.002325757872313261, + 0.002724101534113288, + 0.003814415307715535, + 0.0024526461493223906, + 0.012054678983986378, + 0.011515473015606403, + 0.02019292116165161, + 0.02531706541776657, + 0.01704287901520729, + 0.06091767176985741, + 0.026796486228704453, + 0.058179762214422226, + 0.020274361595511436, + 0.011831440031528473, + 0.02279520593583584, + 0.017624709755182266, + 0.02207096479833126, + 0.19176407158374786, + 0.0366959311068058, + 0.18599987030029297, + 0.002750860992819071, + 0.002573553705587983, + 0.0059099034406244755, + 0.003974401392042637, + 0.004835519008338451, + 0.0054589929059147835, + 0.008611532859504223, + 0.004285044968128204, + 0.0031709771137684584, + 0.0027795112691819668, + 0.005638981703668833, + 0.00540211983025074, + 0.004238411318510771, + 0.019758127629756927, + 0.011482068337500095, + 0.018327519297599792, + 0.0006719453958794475, + 0.0004918455961160362, + 0.0009046848281286657, + 0.0006049676449038088, + 0.0007101060473360121, + 0.002637724159285426, + 0.0011808195849880576, + 0.0024556478019803762, + 0.0007225749432109296, + 0.0005909225437790155, + 0.001225392334163189, + 0.0005302033387124538, + 0.0009663323289714754, + 0.001648125471547246, + 0.0008298912434838712, + 0.0015877202386036515 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 4, + 7 + ], + "selected_reward": -0.02329447865486145 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video presented in a high-contrast black-and-white style?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the video begin with an extreme close-up of a pearl that has escaped from a broken necklace string?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the pearl slowly tumble down and land on a dark, finely textured velvet fabric?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the video show a reflection of a woman's face with a wide-open eye and a glistening tear on the pearl's surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start by tilting downward to track the pearl and then shift to a still extreme macro perspective?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.026340274140238762 + ], + "opposite_logprob_tokens": [ + -3.6513402462005615 + ], + "logprob_reward": -0.026340274140238762, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.026340274140238762, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.4093296527862549, + 0.014114897698163986, + 0.052627891302108765, + 0.05962898209691048, + 0.05559657886624336, + 0.2437901347875595, + 0.09717961400747299, + 0.023562200367450714, + 0.034560222178697586, + 0.009609862230718136 + ], + "credit_map": [ + 0.0832485482096672, + 0.05687045305967331, + 0.06057426333427429, + 0.06197876110672951, + 0.03502560779452324, + 0.0494677759706974, + 0.02200445719063282, + 0.04015977680683136, + 0.0017893449403345585, + 0.0013737697154283524, + 0.0017552579520270228, + 0.0019034503493458033, + 0.002706743311136961, + 0.001710365992039442, + 0.0012549730017781258, + 0.0016209930181503296, + 0.0037794492673128843, + 0.005549388471990824, + 0.006172631401568651, + 0.01151186041533947, + 0.005376960150897503, + 0.006132592912763357, + 0.0073814173229038715, + 0.0067235869355499744, + 0.005574627313762903, + 0.005291793495416641, + 0.011958476155996323, + 0.007108862977474928, + 0.005624971818178892, + 0.007845192216336727, + 0.008790182881057262, + 0.007434871047735214, + 0.003568023443222046, + 0.0035653042141348124, + 0.0051596881821751595, + 0.005192560609430075, + 0.006185776088386774, + 0.013111138716340065, + 0.007929608225822449, + 0.010884481482207775, + 0.009351043961942196, + 0.006266377400606871, + 0.009497870691120625, + 0.008538377471268177, + 0.01148828025907278, + 0.09520711749792099, + 0.013179921545088291, + 0.09026113152503967, + 0.006160567048937082, + 0.005397907458245754, + 0.013544363901019096, + 0.009551282972097397, + 0.01282602921128273, + 0.015207900665700436, + 0.022050511091947556, + 0.012441054917871952, + 0.0019039816688746214, + 0.0015036864206194878, + 0.003169846488162875, + 0.0026860523503273726, + 0.001591821201145649, + 0.0048817358911037445, + 0.004125415347516537, + 0.0036996619310230017, + 0.0022849540691822767, + 0.0017239043954759836, + 0.002750844694674015, + 0.0033409607131034136, + 0.0022540518548339605, + 0.008140696212649345, + 0.0062591638416051865, + 0.007805643603205681, + 0.0010345036862418056, + 0.0009518416482023895, + 0.0015745601849630475, + 0.0011381879448890686, + 0.0009345783037133515, + 0.0012994790449738503, + 0.001660569105297327, + 0.001016142312437296 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 5, + 6 + ], + "selected_reward": -0.026340274140238762 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "In a high-contrast black-and-white image, an extreme close-up slowly tilts downward, tracking a pearl that has just escaped from a broken necklace string. Illuminated by a single, intense sidelight, the pearl's surface reveals a rich gradation of colors, from bright silver to dark gray, shimmering with a warm sheen. It slowly tumbles down, finally landing on a dark, finely textured velvet fabric, its forward motion steady and graceful. The shot instantly shifts to an extreme macro perspective, the camera remaining still, focused on the resting pearl. Reflected on its mirror-like surface is a fragment of a woman's face—a wide-open eye, its outline strikingly sharp in the black and white, with a single, glistening tear seemingly lingering in its corner. The entire image, imbued with exquisite detail, exudes a tranquil atmosphere of loss and heartbreak.", + "seed": 12257, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000095/group_12/index_1273_seed_12257_dp_25_rank_50_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 12, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.042784165032207964, + "vlm_logprob": -0.042784165032207964, + "vlm_yes_no_margin": 3.475, + "vlm_token_credit": -0.042784165032207964, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video presented in a high-contrast black-and-white style?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the video begin with an extreme close-up of a pearl that has escaped from a broken necklace string?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the pearl slowly tumble down and land on a dark, finely textured velvet fabric?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the video show a reflection of a woman's face with a wide-open eye and a glistening tear on the pearl's surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start by tilting downward to track the pearl and then shift to a still extreme macro perspective?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.030697904527187347 + ], + "opposite_logprob_tokens": [ + -4.030697822570801 + ], + "logprob_reward": -0.030697904527187347, + "yes_no_margin_reward": 4.0, + "token_credit_reward": -0.030697904527187347, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.16307751834392548, + 0.07066784799098969, + 0.08222497254610062, + 0.03307605907320976, + 0.27940645813941956, + 0.08503369987010956, + 0.2236947864294052, + 0.024820487946271896, + 0.023616604506969452, + 0.014381539076566696 + ], + "credit_map": [ + 0.012245066463947296, + 0.013471033424139023, + 0.023312097415328026, + 0.026999153196811676, + 0.014300473965704441, + 0.015420284122228622, + 0.02632705494761467, + 0.03100235015153885, + 0.010070967487990856, + 0.0068977405317127705, + 0.011926812119781971, + 0.010424749925732613, + 0.006405662279576063, + 0.008313961327075958, + 0.00956813059747219, + 0.007059828843921423, + 0.006374184507876635, + 0.008402486331760883, + 0.011033697053790092, + 0.017662225291132927, + 0.011333172209560871, + 0.011917226016521454, + 0.008259892463684082, + 0.007242091465741396, + 0.0037455337587743998, + 0.0024188051465898752, + 0.005073366686701775, + 0.004364939872175455, + 0.002254243940114975, + 0.005292712245136499, + 0.0024584014900028706, + 0.007468059193342924, + 0.028601517900824547, + 0.012357191182672977, + 0.06136097386479378, + 0.028977496549487114, + 0.011942400597035885, + 0.04283146932721138, + 0.025036606937646866, + 0.0682988166809082, + 0.012328186072409153, + 0.004710094537585974, + 0.01367488969117403, + 0.004054939839988947, + 0.006681583356112242, + 0.032333895564079285, + 0.00329772662371397, + 0.007952382788062096, + 0.030188191682100296, + 0.014406931586563587, + 0.05767050012946129, + 0.028553565964102745, + 0.016184069216251373, + 0.030689621344208717, + 0.01918892003595829, + 0.026812972500920296, + 0.0035389033146202564, + 0.0018045944161713123, + 0.00397254666313529, + 0.0026773749850690365, + 0.0018304332625120878, + 0.005501433275640011, + 0.0019342914456501603, + 0.003560911398380995, + 0.004355162847787142, + 0.0016049244441092014, + 0.004753895103931427, + 0.0027390685863792896, + 0.0012576633598655462, + 0.0039015833754092455, + 0.0015616415766999125, + 0.0034426662605255842, + 0.003182128770276904, + 0.0007989134173840284, + 0.0023533410858362913, + 0.001257179887033999, + 0.0008688293164595962, + 0.0016938275657594204, + 0.0013438666937872767, + 0.002883452922105789 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 6, + 0 + ], + "selected_reward": -0.030697904527187347 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video presented in a high-contrast black-and-white style?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the video begin with an extreme close-up of a pearl that has escaped from a broken necklace string?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the pearl slowly tumble down and land on a dark, finely textured velvet fabric?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the video show a reflection of a woman's face with a wide-open eye and a glistening tear on the pearl's surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start by tilting downward to track the pearl and then shift to a still extreme macro perspective?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03810820356011391 + ], + "opposite_logprob_tokens": [ + -3.2881081104278564 + ], + "logprob_reward": -0.03810820356011391, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.03810820356011391, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.32743528485298157, + 0.12810491025447845, + 0.08033054322004318, + 0.09036844968795776, + 0.1069900169968605, + 0.10816120356321335, + 0.036099549382925034, + 0.05097091943025589, + 0.0592685304582119, + 0.012270626612007618 + ], + "credit_map": [ + 0.02435516007244587, + 0.03569796308875084, + 0.047568436712026596, + 0.039329249411821365, + 0.04559161514043808, + 0.03137075528502464, + 0.0436861626803875, + 0.059835925698280334, + 0.014531835913658142, + 0.020626535639166832, + 0.016695253551006317, + 0.013333024457097054, + 0.023037003353238106, + 0.014249889180064201, + 0.015583653934299946, + 0.010047707706689835, + 0.0057789599522948265, + 0.009184883907437325, + 0.010709013789892197, + 0.014657586812973022, + 0.014095520600676537, + 0.011697350069880486, + 0.008171848952770233, + 0.00603537866845727, + 0.009514006786048412, + 0.008474183268845081, + 0.01406126283109188, + 0.013338943012058735, + 0.00582946976646781, + 0.011038193479180336, + 0.006320933811366558, + 0.021791448816657066, + 0.01239857915788889, + 0.006791081745177507, + 0.023905009031295776, + 0.011960292235016823, + 0.006078911479562521, + 0.013147206045687199, + 0.010953962802886963, + 0.02175496704876423, + 0.011971445754170418, + 0.006930946838110685, + 0.017440736293792725, + 0.011578085832297802, + 0.00668526254594326, + 0.033401958644390106, + 0.006572894286364317, + 0.01357986219227314, + 0.0047193714417517185, + 0.002790614264085889, + 0.006945612374693155, + 0.006680075544863939, + 0.0025297377724200487, + 0.00424761651083827, + 0.003542737103998661, + 0.0046437825076282024, + 0.009380823001265526, + 0.0023829347919672728, + 0.009153714403510094, + 0.005170614458620548, + 0.003092372789978981, + 0.009566160850226879, + 0.003945834469050169, + 0.008278464898467064, + 0.007601809687912464, + 0.004970201291143894, + 0.012275033630430698, + 0.006687430199235678, + 0.0062288991175591946, + 0.00907997228205204, + 0.005053486209362745, + 0.007371694315224886, + 0.0031002680771052837, + 0.0006473002140410244, + 0.0017054688651114702, + 0.0010685281595215201, + 0.0006695669726468623, + 0.0012460638536140323, + 0.0008067878661677241, + 0.003026642370969057 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 5 + ], + "selected_reward": -0.03810820356011391 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video presented in a high-contrast black-and-white style?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the video begin with an extreme close-up of a pearl that has escaped from a broken necklace string?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the pearl slowly tumble down and land on a dark, finely textured velvet fabric?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the video show a reflection of a woman's face with a wide-open eye and a glistening tear on the pearl's surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start by tilting downward to track the pearl and then shift to a still extreme macro perspective?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.011083466000854969 + ], + "opposite_logprob_tokens": [ + -4.511083602905273 + ], + "logprob_reward": -0.011083466000854969, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.011083466000854969, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09167446196079254, + 0.06412378698587418, + 0.14654923975467682, + 0.05036436393857002, + 0.06356044113636017, + 0.28110697865486145, + 0.09526507556438446, + 0.12326209247112274, + 0.04652929678559303, + 0.0375642254948616 + ], + "credit_map": [ + 0.009039528667926788, + 0.009556117467582226, + 0.01457421574741602, + 0.013175777159631252, + 0.008400981314480305, + 0.010554890148341656, + 0.014082206413149834, + 0.012290744110941887, + 0.010119509883224964, + 0.010096141137182713, + 0.007745982147753239, + 0.007085290737450123, + 0.01035443414002657, + 0.006008568219840527, + 0.008020003326237202, + 0.004693853203207254, + 0.012000606395304203, + 0.017750289291143417, + 0.022024746984243393, + 0.018133867532014847, + 0.025839632377028465, + 0.02393878810107708, + 0.013494467362761497, + 0.013366853818297386, + 0.005660979077219963, + 0.006475738249719143, + 0.006838514003902674, + 0.007143966853618622, + 0.004539531655609608, + 0.008730767294764519, + 0.0036621922627091408, + 0.007312668487429619, + 0.006864383351057768, + 0.00308788288384676, + 0.01377085130661726, + 0.006826329976320267, + 0.0030782867688685656, + 0.00838944036513567, + 0.0055041927844285965, + 0.01603907160460949, + 0.038213226944208145, + 0.01990911364555359, + 0.053706277161836624, + 0.03973221406340599, + 0.01785345934331417, + 0.06640271097421646, + 0.019663481041789055, + 0.02562650665640831, + 0.01651977375149727, + 0.006193092558532953, + 0.018746186047792435, + 0.008217980153858662, + 0.0067226821556687355, + 0.019082391634583473, + 0.006418109871447086, + 0.013364852406084538, + 0.021621087566018105, + 0.00611575273796916, + 0.028074147179722786, + 0.008687752299010754, + 0.006867145653814077, + 0.02371915429830551, + 0.007640792988240719, + 0.020536266267299652, + 0.008030365221202374, + 0.00272575207054615, + 0.009942656382918358, + 0.004785355646163225, + 0.004204103723168373, + 0.007204381283372641, + 0.0029342330526560545, + 0.006702449172735214, + 0.007509132381528616, + 0.0021052320953458548, + 0.008571812883019447, + 0.0032494887709617615, + 0.0030395977664738894, + 0.005495219957083464, + 0.002555276034399867, + 0.005038466304540634 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 2, + 7 + ], + "selected_reward": -0.011083466000854969 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video presented in a high-contrast black-and-white style?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the video begin with an extreme close-up of a pearl that has escaped from a broken necklace string?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the pearl slowly tumble down and land on a dark, finely textured velvet fabric?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the video show a reflection of a woman's face with a wide-open eye and a glistening tear on the pearl's surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start by tilting downward to track the pearl and then shift to a still extreme macro perspective?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.10033027082681656 + ], + "opposite_logprob_tokens": [ + -2.350330352783203 + ], + "logprob_reward": -0.10033027082681656, + "yes_no_margin_reward": 2.25, + "token_credit_reward": -0.10033027082681656, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11238034814596176, + 0.10306733101606369, + 0.09977039694786072, + 0.038529008626937866, + 0.24120870232582092, + 0.14474450051784515, + 0.05782821774482727, + 0.03052002564072609, + 0.1608295440673828, + 0.011121906340122223 + ], + "credit_map": [ + 0.008904137648642063, + 0.0090798893943429, + 0.019141193479299545, + 0.01577594503760338, + 0.007937820628285408, + 0.009952093474566936, + 0.014685774222016335, + 0.02690349519252777, + 0.014222721569240093, + 0.015201381407678127, + 0.012628220021724701, + 0.011911150068044662, + 0.016341321170330048, + 0.01335891056805849, + 0.011010781861841679, + 0.00839284434914589, + 0.008320290595293045, + 0.012163413688540459, + 0.011343790218234062, + 0.012131542898714542, + 0.019796492531895638, + 0.017465034499764442, + 0.01006192434579134, + 0.008487912826240063, + 0.004051588475704193, + 0.004039218183606863, + 0.004776804707944393, + 0.007796311751008034, + 0.0026677357964217663, + 0.006670484319329262, + 0.0029853428713977337, + 0.005541522055864334, + 0.023279624059796333, + 0.013884393498301506, + 0.05154145136475563, + 0.029636569321155548, + 0.010220980271697044, + 0.03614594042301178, + 0.02238306775689125, + 0.05411667376756668, + 0.020488286390900612, + 0.008695224300026894, + 0.022143997251987457, + 0.011853726580739021, + 0.008701466023921967, + 0.05434398353099823, + 0.005669676698744297, + 0.01284813042730093, + 0.010304571129381657, + 0.004120281897485256, + 0.0131109319627285, + 0.009327110834419727, + 0.0034988040570169687, + 0.00750928046181798, + 0.0035885516554117203, + 0.0063686855137348175, + 0.003774090204387903, + 0.0021955117117613554, + 0.006453110370784998, + 0.0035825681407004595, + 0.0017330722184851766, + 0.006189704407006502, + 0.0022869950626045465, + 0.004304973874241114, + 0.024053949862718582, + 0.00932102557271719, + 0.03657057508826256, + 0.01890501007437706, + 0.012604499235749245, + 0.028180403634905815, + 0.010604188777506351, + 0.02058989368379116, + 0.002009655348956585, + 0.0008299312903545797, + 0.0024223916698247194, + 0.0013088671257719398, + 0.0005552521324716508, + 0.0011553902877494693, + 0.0010955814504995942, + 0.0017448366852477193 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 8, + 5 + ], + "selected_reward": -0.10033027082681656 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video presented in a high-contrast black-and-white style?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the video begin with an extreme close-up of a pearl that has escaped from a broken necklace string?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the pearl slowly tumble down and land on a dark, finely textured velvet fabric?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the video show a reflection of a woman's face with a wide-open eye and a glistening tear on the pearl's surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start by tilting downward to track the pearl and then shift to a still extreme macro perspective?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03370098024606705 + ], + "opposite_logprob_tokens": [ + -3.408700942993164 + ], + "logprob_reward": -0.03370098024606705, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.03370098024606705, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09028361737728119, + 0.07499172538518906, + 0.3104649782180786, + 0.06393566727638245, + 0.17778509855270386, + 0.06626202911138535, + 0.14310115575790405, + 0.04727035015821457, + 0.01715080253779888, + 0.008754627779126167 + ], + "credit_map": [ + 0.00747687229886651, + 0.01199027057737112, + 0.012967190705239773, + 0.009786593727767467, + 0.010538320057094097, + 0.009480412118136883, + 0.010576928965747356, + 0.01746702566742897, + 0.0099129443988204, + 0.012534006498754025, + 0.01034563034772873, + 0.008693237788975239, + 0.01047601830214262, + 0.008997763507068157, + 0.008174232207238674, + 0.0058578914031386375, + 0.02551797404885292, + 0.037173815071582794, + 0.043126147240400314, + 0.04012112319469452, + 0.05639791116118431, + 0.05346411466598511, + 0.028550824150443077, + 0.026113051921129227, + 0.007809395436197519, + 0.007147178519517183, + 0.014753206633031368, + 0.008697903715074062, + 0.004181562922894955, + 0.008629410527646542, + 0.005027612205594778, + 0.007689397782087326, + 0.01759207621216774, + 0.009008487686514854, + 0.037497010082006454, + 0.020891031250357628, + 0.008499584160745144, + 0.02407272905111313, + 0.02017359994351864, + 0.040050577372312546, + 0.00879956316202879, + 0.005228889174759388, + 0.012087116949260235, + 0.009526245296001434, + 0.005647761281579733, + 0.00840422511100769, + 0.006765196565538645, + 0.009803030639886856, + 0.018633397296071053, + 0.009659930132329464, + 0.03180669620633125, + 0.02490597777068615, + 0.009386330842971802, + 0.01856653392314911, + 0.012706255540251732, + 0.01743602380156517, + 0.009881311096251011, + 0.0026520946994423866, + 0.009290876798331738, + 0.004717167001217604, + 0.0019705360755324364, + 0.005925523117184639, + 0.003970390651375055, + 0.008862452581524849, + 0.002627583919093013, + 0.0012067159404978156, + 0.004055545665323734, + 0.0018732433672994375, + 0.0011260377941653132, + 0.0025692118797451258, + 0.001235295902006328, + 0.0024571684189140797, + 0.0012582674389705062, + 0.00075141666457057, + 0.0018506740452721715, + 0.0011750819394364953, + 0.0004785957280546427, + 0.0011904728598892689, + 0.0009237218182533979, + 0.0011263976339250803 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 4, + 6 + ], + "selected_reward": -0.03370098024606705 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "In a high-contrast black-and-white image, an extreme close-up slowly tilts downward, tracking a pearl that has just escaped from a broken necklace string. Illuminated by a single, intense sidelight, the pearl's surface reveals a rich gradation of colors, from bright silver to dark gray, shimmering with a warm sheen. It slowly tumbles down, finally landing on a dark, finely textured velvet fabric, its forward motion steady and graceful. The shot instantly shifts to an extreme macro perspective, the camera remaining still, focused on the resting pearl. Reflected on its mirror-like surface is a fragment of a woman's face—a wide-open eye, its outline strikingly sharp in the black and white, with a single, glistening tear seemingly lingering in its corner. The entire image, imbued with exquisite detail, exudes a tranquil atmosphere of loss and heartbreak.", + "seed": 12258, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000095/group_12/index_1273_seed_12258_dp_25_rank_50_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 12, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.046959365159273146, + "vlm_logprob": -0.046959365159273146, + "vlm_yes_no_margin": 3.3, + "vlm_token_credit": -0.046959365159273146, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video presented in a high-contrast black-and-white style?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the video begin with an extreme close-up of a pearl that has escaped from a broken necklace string?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the pearl slowly tumble down and land on a dark, finely textured velvet fabric?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the video show a reflection of a woman's face with a wide-open eye and a glistening tear on the pearl's surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start by tilting downward to track the pearl and then shift to a still extreme macro perspective?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.049061309546232224 + ], + "opposite_logprob_tokens": [ + -3.2990612983703613 + ], + "logprob_reward": -0.049061309546232224, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.049061309546232224, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11798807978630066, + 0.0363236665725708, + 0.08221177011728287, + 0.4243531823158264, + 0.04103745520114899, + 0.06911029666662216, + 0.03973562270402908, + 0.030373623594641685, + 0.07376410067081451, + 0.08510220795869827 + ], + "credit_map": [ + 0.011540698818862438, + 0.010677356272935867, + 0.012767860665917397, + 0.01690301112830639, + 0.010700567625463009, + 0.022439541295170784, + 0.009945476427674294, + 0.023013565689325333, + 0.002848047763109207, + 0.0029385227244347334, + 0.005672476254403591, + 0.0034753410145640373, + 0.0028837916906923056, + 0.006355310324579477, + 0.004210315644741058, + 0.00793986115604639, + 0.006841346155852079, + 0.005630312487483025, + 0.00908909272402525, + 0.014484451152384281, + 0.007684269454330206, + 0.01115828100591898, + 0.016305776312947273, + 0.011018245480954647, + 0.011689016595482826, + 0.0082018431276083, + 0.031651753932237625, + 0.01528958324342966, + 0.010865321382880211, + 0.15918764472007751, + 0.018293233588337898, + 0.16917477548122406, + 0.002276061801239848, + 0.0027444937732070684, + 0.007041303440928459, + 0.005708713084459305, + 0.00305912084877491, + 0.005121827125549316, + 0.008819709531962872, + 0.006266226526349783, + 0.003865589154884219, + 0.00591304711997509, + 0.015147130005061626, + 0.006830771919339895, + 0.006299459375441074, + 0.00830222200602293, + 0.011864365078508854, + 0.010887720622122288, + 0.0019671800546348095, + 0.0024606329388916492, + 0.012072499841451645, + 0.004737017210572958, + 0.0016480913618579507, + 0.005162429064512253, + 0.005538441706448793, + 0.006149331573396921, + 0.0016286720056086779, + 0.0019291453063488007, + 0.006694059353321791, + 0.003732290118932724, + 0.0026067327708005905, + 0.0029992354102432728, + 0.006524288561195135, + 0.004259196575731039, + 0.0040740990079939365, + 0.00479369331151247, + 0.017936240881681442, + 0.008710896596312523, + 0.0035936213098466396, + 0.010221488773822784, + 0.011900853365659714, + 0.012533212080597878, + 0.005348841194063425, + 0.005239560268819332, + 0.017433125525712967, + 0.010105046443641186, + 0.006508625112473965, + 0.010722354054450989, + 0.016125686466693878, + 0.013618973083794117 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 0, + 9 + ], + "selected_reward": -0.049061309546232224 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video presented in a high-contrast black-and-white style?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the video begin with an extreme close-up of a pearl that has escaped from a broken necklace string?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the pearl slowly tumble down and land on a dark, finely textured velvet fabric?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the video show a reflection of a woman's face with a wide-open eye and a glistening tear on the pearl's surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start by tilting downward to track the pearl and then shift to a still extreme macro perspective?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.10031053423881531 + ], + "opposite_logprob_tokens": [ + -2.3503105640411377 + ], + "logprob_reward": -0.10031053423881531, + "yes_no_margin_reward": 2.25, + "token_credit_reward": -0.10031053423881531, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.25374558568000793, + 0.11413523554801941, + 0.08146123588085175, + 0.2712697386741638, + 0.07837159931659698, + 0.09797439724206924, + 0.024122202768921852, + 0.031093785539269447, + 0.023578369989991188, + 0.02424783632159233 + ], + "credit_map": [ + 0.024077659472823143, + 0.014568047598004341, + 0.02761606127023697, + 0.018289314582943916, + 0.04059353470802307, + 0.05451587587594986, + 0.016173239797353745, + 0.05791187658905983, + 0.009697752073407173, + 0.011061631143093109, + 0.019188953563570976, + 0.011362678371369839, + 0.010164070874452591, + 0.016360506415367126, + 0.017066240310668945, + 0.019233394414186478, + 0.005230860784649849, + 0.005229840986430645, + 0.010122069157660007, + 0.014110899530351162, + 0.011148130521178246, + 0.011960333213210106, + 0.013298013247549534, + 0.010361088439822197, + 0.006855190731585026, + 0.00767749547958374, + 0.03245823457837105, + 0.015299485065042973, + 0.008760702796280384, + 0.08578740060329437, + 0.024787375703454018, + 0.08964387327432632, + 0.0046691750176250935, + 0.006314161233603954, + 0.0156758613884449, + 0.009522969834506512, + 0.004584471695125103, + 0.009213818237185478, + 0.017572473734617233, + 0.010818669572472572, + 0.004680989310145378, + 0.0062708063051104546, + 0.016589388251304626, + 0.0133744478225708, + 0.012039314024150372, + 0.010522010736167431, + 0.01994708925485611, + 0.0145503468811512, + 0.0012739922385662794, + 0.001728947157971561, + 0.005031669978052378, + 0.0031910219695419073, + 0.00232737697660923, + 0.003267105668783188, + 0.0037862451281398535, + 0.0035158423706889153, + 0.001723260385915637, + 0.0018445112509652972, + 0.009655148722231388, + 0.00331908673979342, + 0.0011312906863167882, + 0.0050402418710291386, + 0.003275712253525853, + 0.005104532930999994, + 0.0013857707381248474, + 0.0014541009441018105, + 0.004618794657289982, + 0.0029436370823532343, + 0.0016239910619333386, + 0.0032296283170580864, + 0.004668435547500849, + 0.0036540126893669367, + 0.0015598142053931952, + 0.001307091093622148, + 0.006337637081742287, + 0.002592352218925953, + 0.0016278971452265978, + 0.0032681410666555166, + 0.003643860574811697, + 0.003911043982952833 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 0, + 1 + ], + "selected_reward": -0.10031053423881531 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video presented in a high-contrast black-and-white style?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the video begin with an extreme close-up of a pearl that has escaped from a broken necklace string?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the pearl slowly tumble down and land on a dark, finely textured velvet fabric?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the video show a reflection of a woman's face with a wide-open eye and a glistening tear on the pearl's surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start by tilting downward to track the pearl and then shift to a still extreme macro perspective?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.012544183060526848 + ], + "opposite_logprob_tokens": [ + -4.38754415512085 + ], + "logprob_reward": -0.012544183060526848, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.012544183060526848, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08257158100605011, + 0.07249712944030762, + 0.06348008662462234, + 0.30738821625709534, + 0.23827190697193146, + 0.06019662320613861, + 0.036379799246788025, + 0.0590384416282177, + 0.03119686245918274, + 0.04897940531373024 + ], + "credit_map": [ + 0.00463377358391881, + 0.005127936601638794, + 0.005968875251710415, + 0.005298397038131952, + 0.024253008887171745, + 0.01411357056349516, + 0.008647745475172997, + 0.014528273604810238, + 0.005714732687920332, + 0.0052158525213599205, + 0.00935579277575016, + 0.004908536095172167, + 0.00505285058170557, + 0.015172230079770088, + 0.00947519950568676, + 0.017601938918232918, + 0.007353648543357849, + 0.0037173728924244642, + 0.00843289215117693, + 0.016262007877230644, + 0.005001853685826063, + 0.011004719883203506, + 0.004661655519157648, + 0.007045936770737171, + 0.007283646613359451, + 0.005831720307469368, + 0.016238190233707428, + 0.008849354460835457, + 0.006453955080360174, + 0.1231684610247612, + 0.008033283986151218, + 0.13152961432933807, + 0.013630381785333157, + 0.015595409087836742, + 0.05330363288521767, + 0.03079107217490673, + 0.016571898013353348, + 0.0284454133361578, + 0.04526250436902046, + 0.034671600908041, + 0.003417937085032463, + 0.003928724676370621, + 0.013114800676703453, + 0.00931707676500082, + 0.004607467912137508, + 0.009973219595849514, + 0.007512364070862532, + 0.008325031958520412, + 0.0020142316352576017, + 0.002517523942515254, + 0.007519111968576908, + 0.004776115529239178, + 0.002499260939657688, + 0.005545228719711304, + 0.005966398864984512, + 0.005541927646845579, + 0.0031071705743670464, + 0.004470100626349449, + 0.01565585285425186, + 0.008371091447770596, + 0.0021088963840156794, + 0.00912509299814701, + 0.005792244803160429, + 0.010407987982034683, + 0.0019254052313044667, + 0.0022131234873086214, + 0.008801787160336971, + 0.002991275629028678, + 0.0014249839587137103, + 0.004635998979210854, + 0.003872861387208104, + 0.0053314268589019775, + 0.0023480982054024935, + 0.0027933926321566105, + 0.015128720551729202, + 0.005848598666489124, + 0.0017505147261545062, + 0.007781583350151777, + 0.005069163627922535, + 0.008259331807494164 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 4, + 0 + ], + "selected_reward": -0.012544183060526848 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video presented in a high-contrast black-and-white style?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the video begin with an extreme close-up of a pearl that has escaped from a broken necklace string?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the pearl slowly tumble down and land on a dark, finely textured velvet fabric?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the video show a reflection of a woman's face with a wide-open eye and a glistening tear on the pearl's surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start by tilting downward to track the pearl and then shift to a still extreme macro perspective?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.043082013726234436 + ], + "opposite_logprob_tokens": [ + -3.1680819988250732 + ], + "logprob_reward": -0.043082013726234436, + "yes_no_margin_reward": 3.125, + "token_credit_reward": -0.043082013726234436, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1684236228466034, + 0.0322970412671566, + 0.03582245111465454, + 0.6042209267616272, + 0.04907144233584404, + 0.046627841889858246, + 0.015261887572705746, + 0.01592126302421093, + 0.017497921362519264, + 0.01485559530556202 + ], + "credit_map": [ + 0.010090415365993977, + 0.013501434586942196, + 0.012376880273222923, + 0.02771144174039364, + 0.03205307573080063, + 0.029305292293429375, + 0.012458998709917068, + 0.03092607669532299, + 0.0025737183168530464, + 0.0033857503440231085, + 0.0032636902760714293, + 0.0029782033525407314, + 0.0033350451849400997, + 0.004680004436522722, + 0.007085178047418594, + 0.004995449911803007, + 0.0038608508184552193, + 0.0017473094630986452, + 0.006175493821501732, + 0.005424485541880131, + 0.002594203455373645, + 0.0046231187880039215, + 0.006741821300238371, + 0.00465516559779644, + 0.01698932982981205, + 0.012432688847184181, + 0.04802646487951279, + 0.02221522480249405, + 0.01959441415965557, + 0.20703209936618805, + 0.05147433653473854, + 0.2264563888311386, + 0.0028081187047064304, + 0.0033848288003355265, + 0.007709084078669548, + 0.007234364282339811, + 0.0026040312368422747, + 0.0055773756466805935, + 0.01290732529014349, + 0.006846309639513493, + 0.002038373611867428, + 0.0031736716628074646, + 0.011083395220339298, + 0.004598371218889952, + 0.003966342657804489, + 0.00480135390534997, + 0.010117708705365658, + 0.006848624907433987, + 0.0009762020781636238, + 0.0010238519171252847, + 0.003320845076814294, + 0.0020698190201073885, + 0.001189389149658382, + 0.0021452130749821663, + 0.0022817852441221476, + 0.002254781313240528, + 0.0006453073001466691, + 0.0010363627225160599, + 0.00373500376008451, + 0.0018118016887456179, + 0.0007103965617716312, + 0.001657669898122549, + 0.004276723135262728, + 0.0020479969680309296, + 0.0009364346624352038, + 0.0010757826967164874, + 0.004900330677628517, + 0.0019075284944847226, + 0.0008879138040356338, + 0.002613319316878915, + 0.0024058001581579447, + 0.002770810155197978, + 0.0008284660871140659, + 0.0009573661955073476, + 0.003282370511442423, + 0.0016918857581913471, + 0.0008664922206662595, + 0.0019467601086944342, + 0.0028246704023331404, + 0.002457583090290427 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 0, + 4 + ], + "selected_reward": -0.043082013726234436 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video presented in a high-contrast black-and-white style?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the video begin with an extreme close-up of a pearl that has escaped from a broken necklace string?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the pearl slowly tumble down and land on a dark, finely textured velvet fabric?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the video show a reflection of a woman's face with a wide-open eye and a glistening tear on the pearl's surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start by tilting downward to track the pearl and then shift to a still extreme macro perspective?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.029798785224556923 + ], + "opposite_logprob_tokens": [ + -3.529798746109009 + ], + "logprob_reward": -0.029798785224556923, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.029798785224556923, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.16011101007461548, + 0.10692045837640762, + 0.06462474167346954, + 0.2865063548088074, + 0.053841352462768555, + 0.2319173514842987, + 0.03258611634373665, + 0.014850451610982418, + 0.02587508223950863, + 0.02276700921356678 + ], + "credit_map": [ + 0.012343993410468102, + 0.013853223994374275, + 0.019189292564988136, + 0.028388619422912598, + 0.027242373675107956, + 0.021738102659583092, + 0.013634529896080494, + 0.02372087724506855, + 0.00805939082056284, + 0.012394564226269722, + 0.012585241347551346, + 0.011246493086218834, + 0.01124016847461462, + 0.018266312777996063, + 0.01145678386092186, + 0.021671516820788383, + 0.005907854996621609, + 0.003821286605671048, + 0.008657420054078102, + 0.011697525158524513, + 0.005540703888982534, + 0.00893468875437975, + 0.012069689109921455, + 0.007995571941137314, + 0.010390116833150387, + 0.010260449722409248, + 0.028056876733899117, + 0.019227957352995872, + 0.01230911910533905, + 0.08121494203805923, + 0.03669533506035805, + 0.08835157006978989, + 0.0026808716356754303, + 0.004408629145473242, + 0.009637458249926567, + 0.008165550418198109, + 0.004551438149064779, + 0.006786791607737541, + 0.009942903183400631, + 0.007667715195566416, + 0.013141402043402195, + 0.0207207091152668, + 0.0525834895670414, + 0.021620186045765877, + 0.015623681247234344, + 0.03112092986702919, + 0.039965689182281494, + 0.037141282111406326, + 0.0016964953392744064, + 0.002081911778077483, + 0.007840681821107864, + 0.0032521754037588835, + 0.00267945509403944, + 0.004344084765762091, + 0.005694824270904064, + 0.0049964929930865765, + 0.0009700662922114134, + 0.0010675265220925212, + 0.003355893772095442, + 0.0016407882794737816, + 0.0009752482874318957, + 0.002163155237212777, + 0.0023506218567490578, + 0.0023271518293768167, + 0.0015695245238021016, + 0.0016161046223714948, + 0.0063459486700594425, + 0.0035231253132224083, + 0.0013207998126745224, + 0.003203438827767968, + 0.004141127225011587, + 0.004155016969889402, + 0.0012074918486177921, + 0.0018676795298233628, + 0.005913266446441412, + 0.002853875979781151, + 0.000959812430664897, + 0.0032026555854827166, + 0.0029903813265264034, + 0.0037718461826443672 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 5, + 0 + ], + "selected_reward": -0.029798785224556923 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000095/group_13/index_tensor([8849, 8849])_group_13_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000095/group_13/index_tensor([8849, 8849])_group_13_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..e9d2ec198109130208d40dcb4b9a7615300ae62b --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000095/group_13/index_tensor([8849, 8849])_group_13_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "An extreme close-up static shot, filmed from a slightly elevated angle, focuses on a polished stainless steel dessert spoon as it steadily and deliberately moves toward a vibrant ruby-red gelatin dessert. Made of delicate grooves, the dessert rests on a pristine white ceramic plate, its surface shimmering and shimmering against a softly blurred background, almost liquid-like. As the spoon's tip gently nudged the sides of the teetering structure, mesmerizing ripples course through its entire translucent body, causing the dessert to subtly wriggle and sway. Soft, directed light from an unseen source filters through the scene, casting gentle, shifting shadows that cause the gelatin to glow from within, its shimmering surface shimmering with subtle movement, creating a tranquil and captivating atmosphere focused on texture and light.", + "seed": 24972, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000095/group_13/index_8849_seed_24972_dp_26_rank_52_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 13, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.012661361170466989, + "vlm_logprob": -0.012661361170466989, + "vlm_yes_no_margin": 5.4, + "vlm_token_credit": -0.012661361170466989, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of a polished stainless steel dessert spoon?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera angle slightly elevated and static?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the spoon move steadily and deliberately toward a vibrant ruby-red gelatin dessert?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the gelatin dessert rest on a pristine white ceramic plate with a softly blurred background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the spoon's tip gently nudge the dessert, causing ripples and a swaying motion?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02119416743516922 + ], + "opposite_logprob_tokens": [ + -4.396193981170654 + ], + "logprob_reward": -0.02119416743516922, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.02119416743516922, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.28874191641807556, + 0.08684982359409332, + 0.03460055962204933, + 0.08893687278032303, + 0.07560333609580994, + 0.056510988622903824, + 0.08285994082689285, + 0.05965343490242958, + 0.06819155812263489, + 0.15805162489414215 + ], + "credit_map": [ + 0.03694139048457146, + 0.01632034406065941, + 0.028550101444125175, + 0.057677511125802994, + 0.01889495924115181, + 0.08502443134784698, + 0.029251156374812126, + 0.01608200930058956, + 0.006300703156739473, + 0.0032529004383832216, + 0.008176672272384167, + 0.03232094645500183, + 0.003061912953853607, + 0.003869081847369671, + 0.025947032496333122, + 0.0039205788634717464, + 0.004348961636424065, + 0.0031553523149341345, + 0.004133927170187235, + 0.006827671546489, + 0.0019373245304450393, + 0.003120796289294958, + 0.006907178089022636, + 0.004169347696006298, + 0.007775556296110153, + 0.0063848416320979595, + 0.013606708496809006, + 0.011546891182661057, + 0.007386324927210808, + 0.012440617196261883, + 0.017106283456087112, + 0.012689655646681786, + 0.008578945882618427, + 0.0068808081559836864, + 0.005935061257332563, + 0.012130660004913807, + 0.0049237944185733795, + 0.009100275114178658, + 0.019781874492764473, + 0.008271920494735241, + 0.007527329958975315, + 0.006839832756668329, + 0.0050134202465415, + 0.011411081999540329, + 0.0046028695069253445, + 0.004973144270479679, + 0.010206455364823341, + 0.005936849862337112, + 0.011794034391641617, + 0.007311879191547632, + 0.007981982082128525, + 0.009495619684457779, + 0.007250422611832619, + 0.01665567420423031, + 0.01106049120426178, + 0.011309836991131306, + 0.009620224125683308, + 0.003336942056193948, + 0.003989906050264835, + 0.005914106033742428, + 0.005632558837532997, + 0.012878750450909138, + 0.01086003053933382, + 0.007420915178954601, + 0.00706643657758832, + 0.010878217406570911, + 0.005622916389256716, + 0.009177012369036674, + 0.008356569334864616, + 0.009367681108415127, + 0.009528269059956074, + 0.008194457739591599, + 0.014180782251060009, + 0.011602051556110382, + 0.010164999403059483, + 0.011837519705295563, + 0.007586672902107239, + 0.05109132081270218, + 0.013720277696847916, + 0.037867993116378784 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 3 + ], + "selected_reward": -0.02119416743516922 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of a polished stainless steel dessert spoon?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera angle slightly elevated and static?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the spoon move steadily and deliberately toward a vibrant ruby-red gelatin dessert?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the gelatin dessert rest on a pristine white ceramic plate with a softly blurred background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the spoon's tip gently nudge the dessert, causing ripples and a swaying motion?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03373025730252266 + ], + "opposite_logprob_tokens": [ + -3.4087302684783936 + ], + "logprob_reward": -0.03373025730252266, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.03373025730252266, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15291287004947662, + 0.058760978281497955, + 0.07368643581867218, + 0.11836007982492447, + 0.07192270457744598, + 0.06352563202381134, + 0.1626940369606018, + 0.08074144273996353, + 0.09793856739997864, + 0.11945714056491852 + ], + "credit_map": [ + 0.021389195695519447, + 0.009449994191527367, + 0.012745141983032227, + 0.02178521268069744, + 0.015986144542694092, + 0.043758708983659744, + 0.014654718339443207, + 0.013143773190677166, + 0.005745692178606987, + 0.004482773598283529, + 0.0055082314647734165, + 0.00966548826545477, + 0.009445949457585812, + 0.00745893269777298, + 0.010061525739729404, + 0.006392383016645908, + 0.006377054378390312, + 0.0049152672290802, + 0.012246835976839066, + 0.017704349011182785, + 0.0031893071718513966, + 0.008371912874281406, + 0.015402454882860184, + 0.005479250103235245, + 0.0134523781016469, + 0.009174769744277, + 0.008531591854989529, + 0.015091248787939548, + 0.010353248566389084, + 0.02129240892827511, + 0.022121958434581757, + 0.01834247075021267, + 0.008004488423466682, + 0.00802752934396267, + 0.005333248991519213, + 0.011703242547810078, + 0.0071273138746619225, + 0.008767446503043175, + 0.01535993441939354, + 0.00759950652718544, + 0.005539239384233952, + 0.00544769037514925, + 0.006030999589711428, + 0.011368218809366226, + 0.004486109130084515, + 0.007496651727706194, + 0.01331519614905119, + 0.009841535240411758, + 0.023663880303502083, + 0.019417032599449158, + 0.014402490109205246, + 0.023762095719575882, + 0.011073202826082706, + 0.0256706140935421, + 0.023102639243006706, + 0.021602092310786247, + 0.010381803847849369, + 0.006492340471595526, + 0.004983835853636265, + 0.010828837752342224, + 0.007202694192528725, + 0.009359818883240223, + 0.021430641412734985, + 0.010061475448310375, + 0.009893115609884262, + 0.01308854203671217, + 0.017079655081033707, + 0.011629877611994743, + 0.01739669218659401, + 0.007927740924060345, + 0.010361284017562866, + 0.010561666451394558, + 0.007691516540944576, + 0.00977330468595028, + 0.006771564949303865, + 0.009284894913434982, + 0.008967429399490356, + 0.030276745557785034, + 0.011641887947916985, + 0.035049788653850555 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 0, + 9 + ], + "selected_reward": -0.03373025730252266 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of a polished stainless steel dessert spoon?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera angle slightly elevated and static?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the spoon move steadily and deliberately toward a vibrant ruby-red gelatin dessert?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the gelatin dessert rest on a pristine white ceramic plate with a softly blurred background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the spoon's tip gently nudge the dessert, causing ripples and a swaying motion?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0007254829397425056 + ], + "opposite_logprob_tokens": [ + -7.250725269317627 + ], + "logprob_reward": -0.0007254829397425056, + "yes_no_margin_reward": 7.25, + "token_credit_reward": -0.0007254829397425056, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3223278820514679, + 0.053831059485673904, + 0.06103299930691719, + 0.05071534588932991, + 0.16756953299045563, + 0.030094806104898453, + 0.045383673161268234, + 0.11917959153652191, + 0.07423887401819229, + 0.0756261944770813 + ], + "credit_map": [ + 0.043796606361866, + 0.01820415072143078, + 0.026062767952680588, + 0.05294370651245117, + 0.026608426123857498, + 0.09848619252443314, + 0.032069992274045944, + 0.024156033992767334, + 0.004037960432469845, + 0.0037763449363410473, + 0.0026034563779830933, + 0.015354371629655361, + 0.007084521930664778, + 0.0037044407799839973, + 0.014130725525319576, + 0.003139240900054574, + 0.004700751975178719, + 0.0030185445211827755, + 0.007912524044513702, + 0.01468604151159525, + 0.0036222035996615887, + 0.00825754925608635, + 0.0134002435952425, + 0.005435141269117594, + 0.004661158192902803, + 0.003277468727901578, + 0.004878111183643341, + 0.01051408052444458, + 0.004652036819607019, + 0.004567704629153013, + 0.013257869519293308, + 0.004906915128231049, + 0.01633046567440033, + 0.010367607697844505, + 0.0096982941031456, + 0.02605055272579193, + 0.015844641253352165, + 0.02425631508231163, + 0.04352426901459694, + 0.021497396752238274, + 0.003681511851027608, + 0.003453124314546585, + 0.0020554717630147934, + 0.004894724115729332, + 0.0023507571313530207, + 0.0036189304664731026, + 0.00683561060577631, + 0.0032046774867922068, + 0.006419586017727852, + 0.003643650561571121, + 0.002682733815163374, + 0.003937963861972094, + 0.002892322139814496, + 0.009365994483232498, + 0.009696473367512226, + 0.006744950544089079, + 0.013462735339999199, + 0.007352042477577925, + 0.006995063740760088, + 0.01602165959775448, + 0.013567611575126648, + 0.015386239625513554, + 0.0334181971848011, + 0.012976040132343769, + 0.009253495372831821, + 0.011216189712285995, + 0.005308348219841719, + 0.005769710056483746, + 0.019197365269064903, + 0.007360001094639301, + 0.00679814675822854, + 0.009335624054074287, + 0.00600742781534791, + 0.009261378087103367, + 0.007864594459533691, + 0.010940099135041237, + 0.005796421784907579, + 0.013206761330366135, + 0.011368364095687866, + 0.011181145906448364 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 7 + ], + "selected_reward": -0.0007254829397425056 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of a polished stainless steel dessert spoon?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera angle slightly elevated and static?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the spoon move steadily and deliberately toward a vibrant ruby-red gelatin dessert?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the gelatin dessert rest on a pristine white ceramic plate with a softly blurred background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the spoon's tip gently nudge the dessert, causing ripples and a swaying motion?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0009211826254613698 + ], + "opposite_logprob_tokens": [ + -7.000921249389648 + ], + "logprob_reward": -0.0009211826254613698, + "yes_no_margin_reward": 7.0, + "token_credit_reward": -0.0009211826254613698, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.6767610311508179, + 0.022300466895103455, + 0.037106953561306, + 0.013573265634477139, + 0.045041218400001526, + 0.04106820002198219, + 0.06300174444913864, + 0.013461057096719742, + 0.023479122668504715, + 0.06420684605836868 + ], + "credit_map": [ + 0.09572236239910126, + 0.03355690836906433, + 0.05433907359838486, + 0.12815794348716736, + 0.043569307774305344, + 0.2224404513835907, + 0.06304754316806793, + 0.03592752292752266, + 0.0025278935208916664, + 0.002051592105999589, + 0.0020633682142943144, + 0.002827724441885948, + 0.003131770296022296, + 0.0027299413923174143, + 0.00431663915514946, + 0.002651538932695985, + 0.004124884959310293, + 0.0037178359925746918, + 0.006798955611884594, + 0.0054595861583948135, + 0.0025695262011140585, + 0.002809070749208331, + 0.0068259816616773605, + 0.00480111176148057, + 0.0016375269042328, + 0.0010491134598851204, + 0.0022771551739424467, + 0.0016579782823100686, + 0.0012461121659725904, + 0.0020761790219694376, + 0.002339412458240986, + 0.0012897881679236889, + 0.00503352889791131, + 0.0054308027029037476, + 0.004458954557776451, + 0.009785194881260395, + 0.003745177760720253, + 0.004551887046545744, + 0.0071786134503781796, + 0.004857055377215147, + 0.004564269911497831, + 0.004260081797838211, + 0.004342145752161741, + 0.005799694452434778, + 0.00436309864744544, + 0.004366821609437466, + 0.007799174636602402, + 0.005572912748903036, + 0.009146261028945446, + 0.00401691347360611, + 0.0049285260029137135, + 0.0057683284394443035, + 0.003705400973558426, + 0.019994661211967468, + 0.008429066278040409, + 0.007012586109340191, + 0.0018842422869056463, + 0.0008348054252564907, + 0.0011937113013118505, + 0.0013364548794925213, + 0.0013344617327675223, + 0.002363463630899787, + 0.0021609750110656023, + 0.002352942945435643, + 0.002957988064736128, + 0.004403888713568449, + 0.0018197867320850492, + 0.0020857101771980524, + 0.004660815466195345, + 0.0020699622109532356, + 0.002710612490773201, + 0.0027703591622412205, + 0.005059939343482256, + 0.007585626095533371, + 0.00797005370259285, + 0.010920912958681583, + 0.0056263720616698265, + 0.007045619189739227, + 0.010614504106342793, + 0.009383819997310638 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 6 + ], + "selected_reward": -0.0009211826254613698 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of a polished stainless steel dessert spoon?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera angle slightly elevated and static?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the spoon move steadily and deliberately toward a vibrant ruby-red gelatin dessert?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the gelatin dessert rest on a pristine white ceramic plate with a softly blurred background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the spoon's tip gently nudge the dessert, causing ripples and a swaying motion?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.006735715549439192 + ], + "opposite_logprob_tokens": [ + -5.006735801696777 + ], + "logprob_reward": -0.006735715549439192, + "yes_no_margin_reward": 5.0, + "token_credit_reward": -0.006735715549439192, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2211550772190094, + 0.07779161632061005, + 0.07637597620487213, + 0.06829704344272614, + 0.05717429891228676, + 0.040455546230077744, + 0.08494693040847778, + 0.20534278452396393, + 0.11559893190860748, + 0.05286180600523949 + ], + "credit_map": [ + 0.028363345190882683, + 0.013913789764046669, + 0.020845884457230568, + 0.039430201053619385, + 0.015343178063631058, + 0.06504527479410172, + 0.02266845293343067, + 0.015544945374131203, + 0.003870188957080245, + 0.003583677811548114, + 0.004591304808855057, + 0.028900541365146637, + 0.004016637336462736, + 0.0036189169622957706, + 0.024241993203759193, + 0.004968360066413879, + 0.0076202573254704475, + 0.0043642097152769566, + 0.005223201122134924, + 0.01589455083012581, + 0.006343287881463766, + 0.011234170757234097, + 0.016063660383224487, + 0.009632637724280357, + 0.006889647338539362, + 0.004612493794411421, + 0.005032102577388287, + 0.01293755043298006, + 0.005216330289840698, + 0.010871265083551407, + 0.013153699226677418, + 0.009583956561982632, + 0.005643934477120638, + 0.0061083268374204636, + 0.005205999594181776, + 0.013590187765657902, + 0.0039214519783854485, + 0.00650325883179903, + 0.010743692517280579, + 0.005457447376102209, + 0.004417368210852146, + 0.002721503609791398, + 0.0031105868984013796, + 0.005658818408846855, + 0.003148183459416032, + 0.005145259201526642, + 0.009607751853764057, + 0.006646075285971165, + 0.012604170478880405, + 0.009121323004364967, + 0.005366528406739235, + 0.006284312345087528, + 0.0050066011026501656, + 0.02337038703262806, + 0.014443691819906235, + 0.008749907836318016, + 0.01947995461523533, + 0.011236140504479408, + 0.011381341144442558, + 0.026573898270726204, + 0.029651949182152748, + 0.0267062745988369, + 0.05712719261646271, + 0.02318604476749897, + 0.014530584216117859, + 0.015786299481987953, + 0.007500163745135069, + 0.009342915378510952, + 0.028170254081487656, + 0.011880315840244293, + 0.012281564995646477, + 0.016106829047203064, + 0.0034870661329478025, + 0.005285943858325481, + 0.004441999364644289, + 0.005936201196163893, + 0.007635702379047871, + 0.008727867156267166, + 0.008629508316516876, + 0.008717520162463188 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 7, + 8 + ], + "selected_reward": -0.006735715549439192 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "An extreme close-up static shot, filmed from a slightly elevated angle, focuses on a polished stainless steel dessert spoon as it steadily and deliberately moves toward a vibrant ruby-red gelatin dessert. Made of delicate grooves, the dessert rests on a pristine white ceramic plate, its surface shimmering and shimmering against a softly blurred background, almost liquid-like. As the spoon's tip gently nudged the sides of the teetering structure, mesmerizing ripples course through its entire translucent body, causing the dessert to subtly wriggle and sway. Soft, directed light from an unseen source filters through the scene, casting gentle, shifting shadows that cause the gelatin to glow from within, its shimmering surface shimmering with subtle movement, creating a tranquil and captivating atmosphere focused on texture and light.", + "seed": 24973, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000095/group_13/index_8849_seed_24973_dp_26_rank_52_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 13, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.03301622485741973, + "vlm_logprob": -0.03301622485741973, + "vlm_yes_no_margin": 3.775, + "vlm_token_credit": -0.03301622485741973, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of a polished stainless steel dessert spoon?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera angle slightly elevated and static?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the spoon move steadily and deliberately toward a vibrant ruby-red gelatin dessert?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the gelatin dessert rest on a pristine white ceramic plate with a softly blurred background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the spoon's tip gently nudge the dessert, causing ripples and a swaying motion?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03953928127884865 + ], + "opposite_logprob_tokens": [ + -3.539539337158203 + ], + "logprob_reward": -0.03953928127884865, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.03953928127884865, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.6161975264549255, + 0.059751369059085846, + 0.031032199040055275, + 0.028006471693515778, + 0.0800454393029213, + 0.02617414854466915, + 0.03561779856681824, + 0.03293544426560402, + 0.027223525568842888, + 0.06301604211330414 + ], + "credit_map": [ + 0.033186521381139755, + 0.025146575644612312, + 0.22654736042022705, + 0.23620319366455078, + 0.021795207634568214, + 0.028167016804218292, + 0.026132024824619293, + 0.01901964098215103, + 0.011089907959103584, + 0.005420998204499483, + 0.00736686447635293, + 0.006213695276528597, + 0.007467913441359997, + 0.01086223591119051, + 0.004627344664186239, + 0.006702408194541931, + 0.006431848742067814, + 0.0025089269038289785, + 0.004720721859484911, + 0.002081067766994238, + 0.003060244955122471, + 0.00575217604637146, + 0.002440559444949031, + 0.004036652855575085, + 0.0053396583534777164, + 0.0024345475248992443, + 0.00263874395750463, + 0.002554219448938966, + 0.003872615983709693, + 0.003524905303493142, + 0.0025836490094661713, + 0.0050581335090100765, + 0.00994872022420168, + 0.006566802505403757, + 0.013642985373735428, + 0.01483048964291811, + 0.011483394540846348, + 0.010664300993084908, + 0.007029789965599775, + 0.005878959782421589, + 0.005276415962725878, + 0.002642750507220626, + 0.0028349256608635187, + 0.004868694115430117, + 0.0023722408805042505, + 0.0034035793505609035, + 0.002158698160201311, + 0.002616842510178685, + 0.004949104506522417, + 0.002287308918312192, + 0.0038420541677623987, + 0.00534094637259841, + 0.0041067772544920444, + 0.007040096912533045, + 0.003663310082629323, + 0.004388203378766775, + 0.004224055912345648, + 0.002178249415010214, + 0.004010695498436689, + 0.0032478475477546453, + 0.0038770860992372036, + 0.006742165423929691, + 0.004197483416646719, + 0.00445786165073514, + 0.005375907756388187, + 0.0035223013255745173, + 0.0034970201086252928, + 0.0030001900158822536, + 0.002960993442684412, + 0.00415424769744277, + 0.00261482666246593, + 0.002098038326948881, + 0.012342222966253757, + 0.004935660399496555, + 0.00893409550189972, + 0.009405847638845444, + 0.005302187521010637, + 0.006328303832560778, + 0.00930890254676342, + 0.006458825431764126 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 9 + ], + "selected_reward": -0.03953928127884865 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of a polished stainless steel dessert spoon?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera angle slightly elevated and static?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the spoon move steadily and deliberately toward a vibrant ruby-red gelatin dessert?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the gelatin dessert rest on a pristine white ceramic plate with a softly blurred background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the spoon's tip gently nudge the dessert, causing ripples and a swaying motion?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0700334906578064 + ], + "opposite_logprob_tokens": [ + -2.695033550262451 + ], + "logprob_reward": -0.0700334906578064, + "yes_no_margin_reward": 2.625, + "token_credit_reward": -0.0700334906578064, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.8171083927154541, + 0.01133419293910265, + 0.020355621352791786, + 0.049148671329021454, + 0.014090102165937424, + 0.023199576884508133, + 0.00975580420345068, + 0.02054010145366192, + 0.010488716885447502, + 0.023978762328624725 + ], + "credit_map": [ + 0.03832758218050003, + 0.026192286983132362, + 0.2992042899131775, + 0.3324771821498871, + 0.016896778717637062, + 0.05432126298546791, + 0.0315709114074707, + 0.018118133768439293, + 0.002923083258792758, + 0.0009266819688491523, + 0.001676670042797923, + 0.0016453092684969306, + 0.0009421341237612069, + 0.0015399132389575243, + 0.0009423584560863674, + 0.0007380417664535344, + 0.004198308102786541, + 0.0021437339019030333, + 0.004557279869914055, + 0.0014670732198283076, + 0.001557309296913445, + 0.0035758097656071186, + 0.001193349133245647, + 0.0016627582954242826, + 0.006357946433126926, + 0.0028903509955853224, + 0.008892318233847618, + 0.0040440126322209835, + 0.005286176688969135, + 0.009617868810892105, + 0.005287802778184414, + 0.006772194057703018, + 0.001649096142500639, + 0.0015157173620536923, + 0.0018244043458253145, + 0.003515911288559437, + 0.0015466021141037345, + 0.0017331260023638606, + 0.001552049070596695, + 0.0007531954324804246, + 0.0030513096135109663, + 0.002593864221125841, + 0.002177216811105609, + 0.0029389604460448027, + 0.0024049633648246527, + 0.004765980411320925, + 0.0023490809835493565, + 0.0029182019643485546, + 0.0016081734793260694, + 0.0008926833979785442, + 0.0013462776551023126, + 0.0014131967909634113, + 0.0011606451589614153, + 0.0012409542687237263, + 0.0009653217275626957, + 0.0011285524815320969, + 0.003828017972409725, + 0.0020028918515890837, + 0.002108584623783827, + 0.0023541173432022333, + 0.0016714950324967504, + 0.0029568865429610014, + 0.002303654793649912, + 0.003314453177154064, + 0.0021044507157057524, + 0.0010223125573247671, + 0.0011424889089539647, + 0.0011920810211449862, + 0.0012527578510344028, + 0.0014033638872206211, + 0.0012728145811706781, + 0.0010984472464770079, + 0.0044508930295705795, + 0.0018377761589363217, + 0.0031818803399801254, + 0.003419112414121628, + 0.0022687178570777178, + 0.002723229583352804, + 0.002798015484586358, + 0.003299137344583869 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 9 + ], + "selected_reward": -0.0700334906578064 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of a polished stainless steel dessert spoon?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera angle slightly elevated and static?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the spoon move steadily and deliberately toward a vibrant ruby-red gelatin dessert?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the gelatin dessert rest on a pristine white ceramic plate with a softly blurred background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the spoon's tip gently nudge the dessert, causing ripples and a swaying motion?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.00763498293235898 + ], + "opposite_logprob_tokens": [ + -4.882635116577148 + ], + "logprob_reward": -0.00763498293235898, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.00763498293235898, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.29596951603889465, + 0.05098896846175194, + 0.06389853358268738, + 0.20489101111888885, + 0.06397898495197296, + 0.0695478692650795, + 0.02640412002801895, + 0.08688661456108093, + 0.04745045304298401, + 0.08998393267393112 + ], + "credit_map": [ + 0.037347421050071716, + 0.020428989082574844, + 0.07043197751045227, + 0.03120848909020424, + 0.026643753051757812, + 0.07885867357254028, + 0.016817890107631683, + 0.014232316054403782, + 0.009352083317935467, + 0.00508840661495924, + 0.007073813118040562, + 0.005705600138753653, + 0.005612098146229982, + 0.006427013780921698, + 0.006124675739556551, + 0.005605274811387062, + 0.017198840156197548, + 0.008535911329090595, + 0.011387555859982967, + 0.006445792969316244, + 0.005837514530867338, + 0.006726742722094059, + 0.004272110294550657, + 0.0034940680488944054, + 0.025527583435177803, + 0.009824219159781933, + 0.04701001197099686, + 0.0137777179479599, + 0.01617804728448391, + 0.059177298098802567, + 0.015678534284234047, + 0.017717601731419563, + 0.0075324298813939095, + 0.006160717457532883, + 0.006870293524116278, + 0.008488047868013382, + 0.00966760702431202, + 0.012757883407175541, + 0.007354781497269869, + 0.00514722615480423, + 0.010912829078733921, + 0.007892339490354061, + 0.007043639197945595, + 0.01007431373000145, + 0.0069303689524531364, + 0.009406028315424919, + 0.0058943135663867, + 0.011394032277166843, + 0.004402792546898127, + 0.002556663239374757, + 0.0024798258673399687, + 0.003611263120546937, + 0.003108101198449731, + 0.003801880404353142, + 0.002895718440413475, + 0.0035478759091347456, + 0.014618739485740662, + 0.0088151004165411, + 0.007779373321682215, + 0.011107823811471462, + 0.00833039078861475, + 0.016371959820389748, + 0.010408461093902588, + 0.009454768151044846, + 0.006552053149789572, + 0.004585240501910448, + 0.005522421561181545, + 0.00474409619346261, + 0.0061017777770757675, + 0.009248974733054638, + 0.0059412685222923756, + 0.004754619672894478, + 0.011372615583240986, + 0.0053988550789654255, + 0.013199575245380402, + 0.010491807013750076, + 0.009295285679399967, + 0.014378372579813004, + 0.010727769695222378, + 0.015119659714400768 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 9 + ], + "selected_reward": -0.00763498293235898 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of a polished stainless steel dessert spoon?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera angle slightly elevated and static?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the spoon move steadily and deliberately toward a vibrant ruby-red gelatin dessert?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the gelatin dessert rest on a pristine white ceramic plate with a softly blurred background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the spoon's tip gently nudge the dessert, causing ripples and a swaying motion?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.009783296845853329 + ], + "opposite_logprob_tokens": [ + -4.6347832679748535 + ], + "logprob_reward": -0.009783296845853329, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.009783296845853329, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.24553713202476501, + 0.18567223846912384, + 0.22276443243026733, + 0.039084360003471375, + 0.13390791416168213, + 0.025140749290585518, + 0.019490636885166168, + 0.02639702707529068, + 0.051025208085775375, + 0.05098031461238861 + ], + "credit_map": [ + 0.02330450341105461, + 0.012282502837479115, + 0.04178113862872124, + 0.07797732949256897, + 0.014334194362163544, + 0.050524190068244934, + 0.01551961712539196, + 0.009813648648560047, + 0.02027740702033043, + 0.011413461528718472, + 0.04327841475605965, + 0.02554270066320896, + 0.02377568744122982, + 0.030817417427897453, + 0.015562114305794239, + 0.015005029737949371, + 0.04939267784357071, + 0.025341680273413658, + 0.026753125712275505, + 0.024069180712103844, + 0.024679480120539665, + 0.029063811525702477, + 0.018849559128284454, + 0.024614905938506126, + 0.008152198046445847, + 0.004234131891280413, + 0.003849007887765765, + 0.003812301903963089, + 0.0052377600222826, + 0.004361752886325121, + 0.004447106271982193, + 0.004990097600966692, + 0.0183053407818079, + 0.012723232619464397, + 0.012083950452506542, + 0.013765957206487656, + 0.022781746461987495, + 0.02894660271704197, + 0.01115971989929676, + 0.014141359366476536, + 0.0033426100853830576, + 0.0020496128126978874, + 0.0025520811323076487, + 0.002571099204942584, + 0.0031885886564850807, + 0.006523572374135256, + 0.0020204365719109774, + 0.0028927463572472334, + 0.00252318661659956, + 0.0019549699500203133, + 0.0021686607506126165, + 0.0031281220726668835, + 0.0022397555876523256, + 0.0025209705345332623, + 0.002818431705236435, + 0.002136539202183485, + 0.0027351013850420713, + 0.0017476407811045647, + 0.0029974060598760843, + 0.0026290861424058676, + 0.0034104958176612854, + 0.006238773465156555, + 0.0030138178262859583, + 0.0036247025709599257, + 0.010359234176576138, + 0.005580548662692308, + 0.007185176946222782, + 0.0059587894938886166, + 0.004690338857471943, + 0.006257868371903896, + 0.007038362789899111, + 0.003954889252781868, + 0.006739545613527298, + 0.0037041418254375458, + 0.007076772395521402, + 0.007949022576212883, + 0.005046469159424305, + 0.005169625859707594, + 0.008964288048446178, + 0.00633044121786952 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 1 + ], + "selected_reward": -0.009783296845853329 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of a polished stainless steel dessert spoon?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera angle slightly elevated and static?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the spoon move steadily and deliberately toward a vibrant ruby-red gelatin dessert?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the gelatin dessert rest on a pristine white ceramic plate with a softly blurred background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the spoon's tip gently nudge the dessert, causing ripples and a swaying motion?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03809007257223129 + ], + "opposite_logprob_tokens": [ + -3.2880899906158447 + ], + "logprob_reward": -0.03809007257223129, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.03809007257223129, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.20009726285934448, + 0.08251653611660004, + 0.045774634927511215, + 0.2628857493400574, + 0.09620486199855804, + 0.06845378875732422, + 0.0466848760843277, + 0.06144435331225395, + 0.047127172350883484, + 0.0888107419013977 + ], + "credit_map": [ + 0.02537578158080578, + 0.010628893971443176, + 0.05905844271183014, + 0.025809887796640396, + 0.012655499391257763, + 0.041692160069942474, + 0.013697708956897259, + 0.011178893968462944, + 0.010686627589166164, + 0.007253825664520264, + 0.022104348987340927, + 0.009735100902616978, + 0.00811962503939867, + 0.012672226876020432, + 0.006778406910598278, + 0.005166375543922186, + 0.011689106933772564, + 0.007868667133152485, + 0.005669310223311186, + 0.003309352323412895, + 0.004856838379055262, + 0.0049532325938344, + 0.0036833875346928835, + 0.0037447400391101837, + 0.038398586213588715, + 0.013936127535998821, + 0.059502895921468735, + 0.016930408775806427, + 0.019632849842309952, + 0.07142578810453415, + 0.022068534046411514, + 0.020990559831261635, + 0.01010977290570736, + 0.010583064518868923, + 0.008926532231271267, + 0.01046079397201538, + 0.021941954270005226, + 0.012887119315564632, + 0.008739586919546127, + 0.012556033208966255, + 0.01596834696829319, + 0.0076613835990428925, + 0.006925420369952917, + 0.008692033588886261, + 0.007207993883639574, + 0.008684806525707245, + 0.006235900800675154, + 0.007077902555465698, + 0.006724529433995485, + 0.0036584045737981796, + 0.004875516053289175, + 0.006292462348937988, + 0.005569220520555973, + 0.007194748148322105, + 0.005611847620457411, + 0.006758145522326231, + 0.010931123979389668, + 0.006559126079082489, + 0.0062786852940917015, + 0.008791779167950153, + 0.004789278842508793, + 0.010509411804378033, + 0.008140918798744678, + 0.005444028880447149, + 0.006918768398463726, + 0.004377539735287428, + 0.00501097971573472, + 0.005472663789987564, + 0.005955622531473637, + 0.009336495772004128, + 0.004859954118728638, + 0.005195146892219782, + 0.014018384739756584, + 0.005605440586805344, + 0.010741865262389183, + 0.014719584956765175, + 0.008080310188233852, + 0.010516985319554806, + 0.010431474074721336, + 0.014696696773171425 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 0, + 4 + ], + "selected_reward": -0.03809007257223129 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "An extreme close-up static shot, filmed from a slightly elevated angle, focuses on a polished stainless steel dessert spoon as it steadily and deliberately moves toward a vibrant ruby-red gelatin dessert. Made of delicate grooves, the dessert rests on a pristine white ceramic plate, its surface shimmering and shimmering against a softly blurred background, almost liquid-like. As the spoon's tip gently nudged the sides of the teetering structure, mesmerizing ripples course through its entire translucent body, causing the dessert to subtly wriggle and sway. Soft, directed light from an unseen source filters through the scene, casting gentle, shifting shadows that cause the gelatin to glow from within, its shimmering surface shimmering with subtle movement, creating a tranquil and captivating atmosphere focused on texture and light.", + "seed": 24974, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000095/group_13/index_8849_seed_24974_dp_27_rank_54_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 13, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.020919598289765418, + "vlm_logprob": -0.020919598289765418, + "vlm_yes_no_margin": 4.625, + "vlm_token_credit": -0.020919598289765418, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of a polished stainless steel dessert spoon?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera angle slightly elevated and static?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the spoon move steadily and deliberately toward a vibrant ruby-red gelatin dessert?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the gelatin dessert rest on a pristine white ceramic plate with a softly blurred background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the spoon's tip gently nudge the dessert, causing ripples and a swaying motion?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.027364538982510567 + ], + "opposite_logprob_tokens": [ + -3.902364492416382 + ], + "logprob_reward": -0.027364538982510567, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.027364538982510567, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3150860667228699, + 0.08792603760957718, + 0.05452234670519829, + 0.07914316654205322, + 0.06053849682211876, + 0.06197704002261162, + 0.07110694795846939, + 0.06171658635139465, + 0.0464077889919281, + 0.16157563030719757 + ], + "credit_map": [ + 0.047532983124256134, + 0.028343722224235535, + 0.12075114250183105, + 0.028248678892850876, + 0.018218759447336197, + 0.02572641521692276, + 0.012269778177142143, + 0.03399454057216644, + 0.013324752449989319, + 0.016662200912833214, + 0.013994169421494007, + 0.009592561982572079, + 0.010785036720335484, + 0.0065325479954481125, + 0.0062919617630541325, + 0.010742790065705776, + 0.007646526210010052, + 0.006483806297183037, + 0.005773700773715973, + 0.007775189820677042, + 0.0080952737480402, + 0.007835441268980503, + 0.005545302294194698, + 0.005367100704461336, + 0.01617448776960373, + 0.009523428045213223, + 0.013310824520885944, + 0.008118541911244392, + 0.009882571175694466, + 0.0086970878764987, + 0.005510796792805195, + 0.007925419136881828, + 0.0068900384940207005, + 0.008136248216032982, + 0.007964723743498325, + 0.00795978493988514, + 0.007719056215137243, + 0.007680934853851795, + 0.007474342826753855, + 0.006713360548019409, + 0.009452098980545998, + 0.010681468062102795, + 0.0077714379876852036, + 0.007429497316479683, + 0.007325601298362017, + 0.006265298463404179, + 0.006369459442794323, + 0.006682168692350388, + 0.012119557708501816, + 0.009495855309069157, + 0.008882937021553516, + 0.007025392726063728, + 0.007815340533852577, + 0.010069862008094788, + 0.006432890426367521, + 0.009265096858143806, + 0.007690940983593464, + 0.010970469564199448, + 0.005523870699107647, + 0.00867359060794115, + 0.008828624151647091, + 0.007551528979092836, + 0.006578010506927967, + 0.005899540614336729, + 0.005398190580308437, + 0.004983284045010805, + 0.004734399728477001, + 0.0063021741807460785, + 0.005497496575117111, + 0.005901852156966925, + 0.007208412513136864, + 0.006381967570632696, + 0.02038714475929737, + 0.015324599109590054, + 0.014792934991419315, + 0.02115960791707039, + 0.03186318278312683, + 0.0203773844987154, + 0.02188222110271454, + 0.015788521617650986 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 1 + ], + "selected_reward": -0.027364538982510567 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of a polished stainless steel dessert spoon?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera angle slightly elevated and static?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the spoon move steadily and deliberately toward a vibrant ruby-red gelatin dessert?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the gelatin dessert rest on a pristine white ceramic plate with a softly blurred background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the spoon's tip gently nudge the dessert, causing ripples and a swaying motion?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.054980359971523285 + ], + "opposite_logprob_tokens": [ + -2.9299802780151367 + ], + "logprob_reward": -0.054980359971523285, + "yes_no_margin_reward": 2.875, + "token_credit_reward": -0.054980359971523285, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12853550910949707, + 0.07531090080738068, + 0.09101879596710205, + 0.21586377918720245, + 0.2580691874027252, + 0.1609220802783966, + 0.0205741748213768, + 0.012695999816060066, + 0.012770449742674828, + 0.024239132180809975 + ], + "credit_map": [ + 0.019513892009854317, + 0.012857338413596153, + 0.04549084231257439, + 0.011564268730580807, + 0.007922266609966755, + 0.00925302505493164, + 0.007802189793437719, + 0.014131693169474602, + 0.014476336538791656, + 0.009584098123013973, + 0.01332981139421463, + 0.007226645015180111, + 0.008971311151981354, + 0.008589932695031166, + 0.005631734151393175, + 0.0075010317377746105, + 0.010874326340854168, + 0.010590099729597569, + 0.007807029411196709, + 0.009949935600161552, + 0.01177284587174654, + 0.014976094476878643, + 0.014991884119808674, + 0.010056579485535622, + 0.037109024822711945, + 0.02681598998606205, + 0.022809423506259918, + 0.027181724086403847, + 0.03287932649254799, + 0.027315499261021614, + 0.018742285668849945, + 0.0230104923248291, + 0.04072653129696846, + 0.0291612446308136, + 0.025941846892237663, + 0.03602654114365578, + 0.035674188286066055, + 0.033658936619758606, + 0.02664126642048359, + 0.030238626524806023, + 0.021658768877387047, + 0.023025061935186386, + 0.016108393669128418, + 0.02076198346912861, + 0.020141804590821266, + 0.01832951046526432, + 0.01890942081809044, + 0.021987149491906166, + 0.0025623226538300514, + 0.004093739669770002, + 0.0020810465794056654, + 0.0023728504311293364, + 0.002552082994952798, + 0.0027701312210410833, + 0.0018698228523135185, + 0.002272176556289196, + 0.0012441858416423202, + 0.0017613996751606464, + 0.00137090974021703, + 0.00133405695669353, + 0.0027807343285530806, + 0.001576912123709917, + 0.0013456690357998013, + 0.0012821315322071314, + 0.0016701252898201346, + 0.0013731308281421661, + 0.0014400167856365442, + 0.001621027709916234, + 0.0017273696139454842, + 0.0015986114740371704, + 0.002000019419938326, + 0.0013401482719928026, + 0.002881776774302125, + 0.0025914094876497984, + 0.002457992872223258, + 0.0027469582855701447, + 0.004937221296131611, + 0.0033827912993729115, + 0.002181322081014514, + 0.0030596607830375433 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 3, + 5 + ], + "selected_reward": -0.054980359971523285 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of a polished stainless steel dessert spoon?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera angle slightly elevated and static?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the spoon move steadily and deliberately toward a vibrant ruby-red gelatin dessert?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the gelatin dessert rest on a pristine white ceramic plate with a softly blurred background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the spoon's tip gently nudge the dessert, causing ripples and a swaying motion?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005263281520456076 + ], + "opposite_logprob_tokens": [ + -5.255263328552246 + ], + "logprob_reward": -0.005263281520456076, + "yes_no_margin_reward": 5.25, + "token_credit_reward": -0.005263281520456076, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0858018696308136, + 0.08081807941198349, + 0.10808821022510529, + 0.1544337421655655, + 0.11717428267002106, + 0.06095804274082184, + 0.06219673901796341, + 0.05889101326465607, + 0.04706072807312012, + 0.22457727789878845 + ], + "credit_map": [ + 0.00983337126672268, + 0.011554131284356117, + 0.008159201592206955, + 0.015375426039099693, + 0.008362175896763802, + 0.009212124161422253, + 0.014696530066430569, + 0.008608913980424404, + 0.012607328593730927, + 0.012686188332736492, + 0.012885704636573792, + 0.007762143854051828, + 0.008856689557433128, + 0.00969726126641035, + 0.006749155931174755, + 0.009573614224791527, + 0.011109155602753162, + 0.011285594664514065, + 0.009361285716295242, + 0.011200536042451859, + 0.01400942075997591, + 0.020978325977921486, + 0.01278906874358654, + 0.017354819923639297, + 0.020826518535614014, + 0.015408454462885857, + 0.01742856577038765, + 0.013409885577857494, + 0.014791850931942463, + 0.02739734575152397, + 0.013872326351702213, + 0.031298793852329254, + 0.018886776641011238, + 0.010665039531886578, + 0.012276653200387955, + 0.013939608819782734, + 0.01578488014638424, + 0.016998929902911186, + 0.013425511308014393, + 0.01519688405096531, + 0.007005524355918169, + 0.010234633460640907, + 0.008134676143527031, + 0.00793485902249813, + 0.00562676228582859, + 0.006843738257884979, + 0.005352147854864597, + 0.009825700893998146, + 0.008621050976216793, + 0.013270825147628784, + 0.006584586575627327, + 0.006369899958372116, + 0.0080500403419137, + 0.007673950400203466, + 0.006259745918214321, + 0.005366639234125614, + 0.0085257263854146, + 0.0074525694362819195, + 0.0075469533912837505, + 0.009061113931238651, + 0.0056264689192175865, + 0.006521374452859163, + 0.007364495191723108, + 0.006792313884943724, + 0.005813029129058123, + 0.00814736820757389, + 0.0038871329743415117, + 0.006535080727189779, + 0.003950818441808224, + 0.007740773726254702, + 0.005492500029504299, + 0.005494024604558945, + 0.025700286030769348, + 0.026578374207019806, + 0.013936124742031097, + 0.02435939572751522, + 0.04411611706018448, + 0.03729076310992241, + 0.022489896044135094, + 0.030106306076049805 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 3, + 4 + ], + "selected_reward": -0.005263281520456076 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of a polished stainless steel dessert spoon?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera angle slightly elevated and static?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the spoon move steadily and deliberately toward a vibrant ruby-red gelatin dessert?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the gelatin dessert rest on a pristine white ceramic plate with a softly blurred background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the spoon's tip gently nudge the dessert, causing ripples and a swaying motion?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.000922849983908236 + ], + "opposite_logprob_tokens": [ + -7.000922679901123 + ], + "logprob_reward": -0.000922849983908236, + "yes_no_margin_reward": 7.0, + "token_credit_reward": -0.000922849983908236, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.31982800364494324, + 0.0784553810954094, + 0.0489601232111454, + 0.10844983160495758, + 0.06035749241709709, + 0.06530824303627014, + 0.04562176391482353, + 0.04581219702959061, + 0.14147362112998962, + 0.08573338389396667 + ], + "credit_map": [ + 0.0505138635635376, + 0.030845018103718758, + 0.12537828087806702, + 0.026782091706991196, + 0.01719287969172001, + 0.028100604191422462, + 0.011979293078184128, + 0.02903595194220543, + 0.011791114695370197, + 0.013592686504125595, + 0.012809430249035358, + 0.0070566218346357346, + 0.008999637328088284, + 0.007603297010064125, + 0.005682243034243584, + 0.010920347645878792, + 0.006006631534546614, + 0.005247075576335192, + 0.004102795850485563, + 0.0056835683062672615, + 0.007618261966854334, + 0.00827160757035017, + 0.007099517155438662, + 0.0049306610599160194, + 0.012862919829785824, + 0.010713836178183556, + 0.011854679323732853, + 0.00945247057825327, + 0.010207407176494598, + 0.019840234890580177, + 0.009862839244306087, + 0.02365543693304062, + 0.007119740825146437, + 0.010673054493963718, + 0.006301605608314276, + 0.007549570873379707, + 0.008870403282344341, + 0.007303342688828707, + 0.0075019896030426025, + 0.005037777591496706, + 0.009630431421101093, + 0.0077127316035330296, + 0.011606940999627113, + 0.007685073651373386, + 0.005179081577807665, + 0.007645169273018837, + 0.007125424221158028, + 0.008723383769392967, + 0.0064643495716154575, + 0.005343281663954258, + 0.006647217087447643, + 0.006178542971611023, + 0.005989911500364542, + 0.0057060085237026215, + 0.004212727304548025, + 0.005079721566289663, + 0.005906970240175724, + 0.006189335603266954, + 0.005279930774122477, + 0.005305491853505373, + 0.006669509690254927, + 0.0071904598735272884, + 0.0048600658774375916, + 0.004410428460687399, + 0.017634013667702675, + 0.013407178223133087, + 0.020307207480072975, + 0.019369976595044136, + 0.01920703426003456, + 0.021987270563840866, + 0.014893104322254658, + 0.014667816460132599, + 0.009890018031001091, + 0.010668973438441753, + 0.007024213671684265, + 0.013224979862570763, + 0.01531931385397911, + 0.01154698058962822, + 0.008567958138883114, + 0.009490934200584888 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 3 + ], + "selected_reward": -0.000922849983908236 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of a polished stainless steel dessert spoon?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera angle slightly elevated and static?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the spoon move steadily and deliberately toward a vibrant ruby-red gelatin dessert?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the gelatin dessert rest on a pristine white ceramic plate with a softly blurred background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the spoon's tip gently nudge the dessert, causing ripples and a swaying motion?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.016066960990428925 + ], + "opposite_logprob_tokens": [ + -4.141067028045654 + ], + "logprob_reward": -0.016066960990428925, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.016066960990428925, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07328587770462036, + 0.09013523906469345, + 0.04940139129757881, + 0.12870746850967407, + 0.11185470223426819, + 0.07000480592250824, + 0.13836519420146942, + 0.027029870077967644, + 0.03684923052787781, + 0.274366170167923 + ], + "credit_map": [ + 0.008901023305952549, + 0.012279024347662926, + 0.019343608990311623, + 0.007985863834619522, + 0.006528711412101984, + 0.006415214855223894, + 0.005083312280476093, + 0.00674911355599761, + 0.013079501688480377, + 0.015547733753919601, + 0.012570660561323166, + 0.00822543166577816, + 0.010438951663672924, + 0.013505147770047188, + 0.007849439978599548, + 0.00891837663948536, + 0.007286507170647383, + 0.005278746131807566, + 0.004978879354894161, + 0.006953872740268707, + 0.006057021673768759, + 0.008289809338748455, + 0.006169624626636505, + 0.004386932123452425, + 0.015843046829104424, + 0.011919248849153519, + 0.014675089158117771, + 0.011497038416564465, + 0.012284738942980766, + 0.02287362329661846, + 0.01175952609628439, + 0.027855155989527702, + 0.014904175885021687, + 0.016923971474170685, + 0.011239576153457165, + 0.012977124191820621, + 0.01548248715698719, + 0.014127443544566631, + 0.016553055495023727, + 0.00964687392115593, + 0.009384230710566044, + 0.011927975341677666, + 0.008857832290232182, + 0.008837566711008549, + 0.007920178584754467, + 0.008469019085168839, + 0.006915888283401728, + 0.007692116778343916, + 0.01659470424056053, + 0.02391049824655056, + 0.015190819278359413, + 0.015437383204698563, + 0.016095733270049095, + 0.02051028236746788, + 0.014342869631946087, + 0.016282906755805016, + 0.003402385162189603, + 0.004201800096780062, + 0.0029227205086499453, + 0.004226688761264086, + 0.00266843824647367, + 0.0033053329680114985, + 0.003569461405277252, + 0.0027330429293215275, + 0.004814500920474529, + 0.00499089527875185, + 0.003256311407312751, + 0.005179206840693951, + 0.0037314901128411293, + 0.006310679949820042, + 0.004016443621367216, + 0.004549701698124409, + 0.02921266295015812, + 0.03057890385389328, + 0.02478252723813057, + 0.03838776424527168, + 0.04942427575588226, + 0.03544206917285919, + 0.035362813621759415, + 0.031175164505839348 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 6, + 3 + ], + "selected_reward": -0.016066960990428925 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "An extreme close-up static shot, filmed from a slightly elevated angle, focuses on a polished stainless steel dessert spoon as it steadily and deliberately moves toward a vibrant ruby-red gelatin dessert. Made of delicate grooves, the dessert rests on a pristine white ceramic plate, its surface shimmering and shimmering against a softly blurred background, almost liquid-like. As the spoon's tip gently nudged the sides of the teetering structure, mesmerizing ripples course through its entire translucent body, causing the dessert to subtly wriggle and sway. Soft, directed light from an unseen source filters through the scene, casting gentle, shifting shadows that cause the gelatin to glow from within, its shimmering surface shimmering with subtle movement, creating a tranquil and captivating atmosphere focused on texture and light.", + "seed": 24975, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000095/group_13/index_8849_seed_24975_dp_27_rank_54_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 13, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.011083666700869799, + "vlm_logprob": -0.011083666700869799, + "vlm_yes_no_margin": 5.275, + "vlm_token_credit": -0.011083666700869799, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of a polished stainless steel dessert spoon?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera angle slightly elevated and static?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the spoon move steadily and deliberately toward a vibrant ruby-red gelatin dessert?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the gelatin dessert rest on a pristine white ceramic plate with a softly blurred background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the spoon's tip gently nudge the dessert, causing ripples and a swaying motion?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02802095003426075 + ], + "opposite_logprob_tokens": [ + -4.028020858764648 + ], + "logprob_reward": -0.02802095003426075, + "yes_no_margin_reward": 4.0, + "token_credit_reward": -0.02802095003426075, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2493443489074707, + 0.13281743228435516, + 0.0678035318851471, + 0.08418860286474228, + 0.04531199857592583, + 0.0967109203338623, + 0.08735648542642593, + 0.06509113311767578, + 0.04256058856844902, + 0.1288149207830429 + ], + "credit_map": [ + 0.03511768952012062, + 0.03892049193382263, + 0.034084729850292206, + 0.022579126060009003, + 0.053664106875658035, + 0.02682725340127945, + 0.02073056995868683, + 0.017420385032892227, + 0.020472191274166107, + 0.01637241803109646, + 0.021115010604262352, + 0.013744551688432693, + 0.020576028153300285, + 0.014182745479047298, + 0.015685128048062325, + 0.010669351555407047, + 0.010646660812199116, + 0.008866796270012856, + 0.008970960043370724, + 0.007039754185825586, + 0.013884454034268856, + 0.006337879225611687, + 0.006965055130422115, + 0.005091970786452293, + 0.01303484570235014, + 0.01104627177119255, + 0.014231065288186073, + 0.006785776000469923, + 0.01613897830247879, + 0.007894325070083141, + 0.008037573657929897, + 0.0070197670720517635, + 0.007935741916298866, + 0.0043707313016057014, + 0.004826087970286608, + 0.006074733566492796, + 0.008158659562468529, + 0.004978928714990616, + 0.005525951739400625, + 0.003441165667027235, + 0.014596915803849697, + 0.011581799015402794, + 0.015766995027661324, + 0.010808216407895088, + 0.01825648359954357, + 0.008794758468866348, + 0.008663929998874664, + 0.00824181828647852, + 0.015517499297857285, + 0.01056694146245718, + 0.014495342038571835, + 0.010531419888138771, + 0.010082611814141273, + 0.00864963885396719, + 0.010956515558063984, + 0.0065565151162445545, + 0.009496111422777176, + 0.007810528855770826, + 0.009477624669671059, + 0.006172871217131615, + 0.008392748422920704, + 0.011518889106810093, + 0.007737823761999607, + 0.0044845351949334145, + 0.006869465112686157, + 0.004990112502127886, + 0.005137534346431494, + 0.005661700386554003, + 0.006250966340303421, + 0.004554239101707935, + 0.005306600593030453, + 0.00378997391089797, + 0.019730696454644203, + 0.023645147681236267, + 0.018421631306409836, + 0.012098871171474457, + 0.017185289412736893, + 0.01702452450990677, + 0.011606180109083652, + 0.009102587588131428 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 9 + ], + "selected_reward": -0.02802095003426075 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of a polished stainless steel dessert spoon?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera angle slightly elevated and static?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the spoon move steadily and deliberately toward a vibrant ruby-red gelatin dessert?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the gelatin dessert rest on a pristine white ceramic plate with a softly blurred background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the spoon's tip gently nudge the dessert, causing ripples and a swaying motion?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01609347201883793 + ], + "opposite_logprob_tokens": [ + -4.1410932540893555 + ], + "logprob_reward": -0.01609347201883793, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.01609347201883793, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.18743328750133514, + 0.09121837466955185, + 0.04603632539510727, + 0.047090791165828705, + 0.15271002054214478, + 0.15545853972434998, + 0.10449422150850296, + 0.06407102197408676, + 0.0297759510576725, + 0.12171151489019394 + ], + "credit_map": [ + 0.02427372708916664, + 0.025997838005423546, + 0.03424713760614395, + 0.017177941277623177, + 0.03837452828884125, + 0.02085455320775509, + 0.012256821617484093, + 0.014250732026994228, + 0.013653049245476723, + 0.013598544523119926, + 0.011969420127570629, + 0.005836728028953075, + 0.018376559019088745, + 0.01131484191864729, + 0.008850621059536934, + 0.0076186093501746655, + 0.007975437678396702, + 0.005640971940010786, + 0.007629209663718939, + 0.005405651871114969, + 0.006295977160334587, + 0.003580978838726878, + 0.005577402655035257, + 0.003930692095309496, + 0.0074043418280780315, + 0.00546437781304121, + 0.004635647404938936, + 0.005278154741972685, + 0.008563132956624031, + 0.006589096039533615, + 0.004301323089748621, + 0.0048547121696174145, + 0.028599422425031662, + 0.01398127805441618, + 0.02167723886668682, + 0.016203323379158974, + 0.025314759463071823, + 0.015805257484316826, + 0.01752721704542637, + 0.013601509854197502, + 0.0328495092689991, + 0.01813860610127449, + 0.020252099260687828, + 0.013645481318235397, + 0.026109440252184868, + 0.021839940920472145, + 0.011255229823291302, + 0.011368206702172756, + 0.015898624435067177, + 0.014255644753575325, + 0.021916475147008896, + 0.012088533490896225, + 0.010017025284469128, + 0.009904958307743073, + 0.010327872820198536, + 0.01008507702499628, + 0.012032192200422287, + 0.006429533939808607, + 0.006764580029994249, + 0.004232875537127256, + 0.006543150171637535, + 0.015473561361432076, + 0.005559502635151148, + 0.007035618647933006, + 0.005487383343279362, + 0.003518186043947935, + 0.002854709280654788, + 0.003979084547609091, + 0.007213184144347906, + 0.002147865714505315, + 0.0026173428632318974, + 0.001958192093297839, + 0.02459803596138954, + 0.020059583708643913, + 0.012821217067539692, + 0.010660416446626186, + 0.01984226703643799, + 0.013902154751121998, + 0.011473555117845535, + 0.008354280143976212 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 5, + 4 + ], + "selected_reward": -0.01609347201883793 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of a polished stainless steel dessert spoon?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera angle slightly elevated and static?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the spoon move steadily and deliberately toward a vibrant ruby-red gelatin dessert?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the gelatin dessert rest on a pristine white ceramic plate with a softly blurred background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the spoon's tip gently nudge the dessert, causing ripples and a swaying motion?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0019471748964861035 + ], + "opposite_logprob_tokens": [ + -6.251947402954102 + ], + "logprob_reward": -0.0019471748964861035, + "yes_no_margin_reward": 6.25, + "token_credit_reward": -0.0019471748964861035, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10890014469623566, + 0.04647896811366081, + 0.027015769854187965, + 0.11795857548713684, + 0.13874056935310364, + 0.07765085995197296, + 0.029429366812109947, + 0.4043230712413788, + 0.024280140176415443, + 0.025222521275281906 + ], + "credit_map": [ + 0.017955578863620758, + 0.01366913877427578, + 0.011028408072888851, + 0.011453187093138695, + 0.024118663743138313, + 0.014290865510702133, + 0.009294281713664532, + 0.0070900204591453075, + 0.007941222749650478, + 0.004986580461263657, + 0.006835591048002243, + 0.004670273512601852, + 0.008749688975512981, + 0.004923336207866669, + 0.004372530151158571, + 0.003999745473265648, + 0.004543035291135311, + 0.0035303235054016113, + 0.0033085737377405167, + 0.002655192045494914, + 0.003318371484056115, + 0.002937859622761607, + 0.004207644611597061, + 0.0025147716514766216, + 0.02236693538725376, + 0.015850285068154335, + 0.012922123074531555, + 0.008934189565479755, + 0.016553353518247604, + 0.020339947193861008, + 0.010125253349542618, + 0.010866495780646801, + 0.024314384907484055, + 0.014412539079785347, + 0.017227018252015114, + 0.012436719611287117, + 0.026399483904242516, + 0.016456253826618195, + 0.014592467807233334, + 0.012901704758405685, + 0.011897610500454903, + 0.008828527294099331, + 0.010656848549842834, + 0.007409277837723494, + 0.013764808885753155, + 0.009850675240159035, + 0.007299917750060558, + 0.00794319435954094, + 0.005361582152545452, + 0.0033785554114729166, + 0.003617117879912257, + 0.002869704505428672, + 0.003623587777838111, + 0.004973118659108877, + 0.003192489268258214, + 0.0024132111575454473, + 0.07632479071617126, + 0.04435493052005768, + 0.033729176968336105, + 0.05120423063635826, + 0.048326246440410614, + 0.08232918381690979, + 0.04085470736026764, + 0.027199819684028625, + 0.003999880515038967, + 0.0028921838384121656, + 0.0023616303224116564, + 0.003516722936183214, + 0.00466506602242589, + 0.002871404169127345, + 0.002390336012467742, + 0.0015829167095944285, + 0.003729849588125944, + 0.003555126953870058, + 0.004030883777886629, + 0.0024923495948314667, + 0.0028519933111965656, + 0.003981131128966808, + 0.002565779723227024, + 0.0020154074300080538 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 4, + 3 + ], + "selected_reward": -0.0019471748964861035 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of a polished stainless steel dessert spoon?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera angle slightly elevated and static?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the spoon move steadily and deliberately toward a vibrant ruby-red gelatin dessert?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the gelatin dessert rest on a pristine white ceramic plate with a softly blurred background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the spoon's tip gently nudge the dessert, causing ripples and a swaying motion?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0007190502947196364 + ], + "opposite_logprob_tokens": [ + -7.25071907043457 + ], + "logprob_reward": -0.0007190502947196364, + "yes_no_margin_reward": 7.25, + "token_credit_reward": -0.0007190502947196364, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1759556531906128, + 0.11135485768318176, + 0.06371194869279861, + 0.040787339210510254, + 0.05542077496647835, + 0.29176250100135803, + 0.035993561148643494, + 0.07279647886753082, + 0.07861489057540894, + 0.07360193133354187 + ], + "credit_map": [ + 0.03650909662246704, + 0.01899026334285736, + 0.01980799064040184, + 0.012568501755595207, + 0.03444620594382286, + 0.026199592277407646, + 0.013677548617124557, + 0.013756460510194302, + 0.01771640032529831, + 0.013279377482831478, + 0.01312408596277237, + 0.012262294068932533, + 0.019903305917978287, + 0.013692761771380901, + 0.012643424794077873, + 0.008733208291232586, + 0.011450117453932762, + 0.009673911146819592, + 0.00967040378600359, + 0.0060182055458426476, + 0.00882495753467083, + 0.006470999214798212, + 0.006077936384826899, + 0.005525416694581509, + 0.007057979237288237, + 0.004715243354439735, + 0.005169320851564407, + 0.003095049411058426, + 0.005902696866542101, + 0.007109700236469507, + 0.00411343015730381, + 0.0036239176988601685, + 0.007869967259466648, + 0.0055948044173419476, + 0.00699868518859148, + 0.005149121396243572, + 0.011128917336463928, + 0.008300630375742912, + 0.005733662750571966, + 0.0046449871733784676, + 0.07021549344062805, + 0.02415979467332363, + 0.026355886831879616, + 0.020735397934913635, + 0.056632332503795624, + 0.03588247671723366, + 0.02623468078672886, + 0.031546451151371, + 0.0055899852886796, + 0.005566688720136881, + 0.005298980511724949, + 0.004030443262308836, + 0.0041219196282327175, + 0.00395501172170043, + 0.0041298093274235725, + 0.00330072408542037, + 0.011500954627990723, + 0.010558051988482475, + 0.007896943017840385, + 0.008579731918871403, + 0.008987398818135262, + 0.0127621004357934, + 0.007776986341923475, + 0.004734313115477562, + 0.01109097246080637, + 0.009665205143392086, + 0.013578685000538826, + 0.010283473879098892, + 0.011297035962343216, + 0.009340024553239346, + 0.0070411646738648415, + 0.006318327970802784, + 0.01690749078989029, + 0.006485847290605307, + 0.008104516193270683, + 0.0059565408155322075, + 0.013747899793088436, + 0.011825655587017536, + 0.005430418532341719, + 0.005143560003489256 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 0, + 1 + ], + "selected_reward": -0.0007190502947196364 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of a polished stainless steel dessert spoon?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera angle slightly elevated and static?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the spoon move steadily and deliberately toward a vibrant ruby-red gelatin dessert?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the gelatin dessert rest on a pristine white ceramic plate with a softly blurred background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the spoon's tip gently nudge the dessert, causing ripples and a swaying motion?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.008637686260044575 + ], + "opposite_logprob_tokens": [ + -4.75863790512085 + ], + "logprob_reward": -0.008637686260044575, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.008637686260044575, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2125086933374405, + 0.16951420903205872, + 0.05685761198401451, + 0.09683609753847122, + 0.128506600856781, + 0.10220953822135925, + 0.027609946206212044, + 0.13663209974765778, + 0.026271561160683632, + 0.04305368289351463 + ], + "credit_map": [ + 0.04156327620148659, + 0.02404804155230522, + 0.01895880699157715, + 0.019846953451633453, + 0.038111452013254166, + 0.033689212054014206, + 0.01821402832865715, + 0.018076909705996513, + 0.023870689794421196, + 0.020747531205415726, + 0.023175058886408806, + 0.020797839388251305, + 0.02848302572965622, + 0.015312076546251774, + 0.024181297048926353, + 0.012946683913469315, + 0.010775046423077583, + 0.008871704339981079, + 0.008106040768325329, + 0.0052891867235302925, + 0.006559466477483511, + 0.005671687889844179, + 0.004996549338102341, + 0.006587929558008909, + 0.01835770346224308, + 0.011576337739825249, + 0.011941771022975445, + 0.006851098500192165, + 0.014505223371088505, + 0.01785293035209179, + 0.006769458297640085, + 0.008981565944850445, + 0.02105993777513504, + 0.012444264255464077, + 0.017333727329969406, + 0.012894462794065475, + 0.023111052811145782, + 0.015258654952049255, + 0.014576208777725697, + 0.01182828564196825, + 0.016966482624411583, + 0.0136524997651577, + 0.011649870313704014, + 0.01100799348205328, + 0.01415327563881874, + 0.01504143700003624, + 0.008276315405964851, + 0.011461654677987099, + 0.004644392989575863, + 0.003339927177876234, + 0.0031140786595642567, + 0.0027956953272223473, + 0.004238078836351633, + 0.0036515758838504553, + 0.003213595598936081, + 0.002612599404528737, + 0.027851110324263573, + 0.014080188237130642, + 0.011793470941483974, + 0.016631361097097397, + 0.015894263982772827, + 0.02659955434501171, + 0.013721623457968235, + 0.01006051991134882, + 0.004178086761385202, + 0.0031712071504443884, + 0.003476654179394245, + 0.0034343162551522255, + 0.0034566230606287718, + 0.003188271541148424, + 0.003013056004419923, + 0.0023533450439572334, + 0.007770632393658161, + 0.004052479285746813, + 0.004995982628315687, + 0.0037778110709041357, + 0.006147031206637621, + 0.006870754063129425, + 0.004536178428679705, + 0.004902814049273729 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 7 + ], + "selected_reward": -0.008637686260044575 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000095/group_14/index_tensor([4472, 4472])_group_14_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000095/group_14/index_tensor([4472, 4472])_group_14_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..633b8934bc7bf1e44a6af718ea4129a2cf90de4d --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000095/group_14/index_tensor([4472, 4472])_group_14_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "An ultra-detailed, high-speed phantom camera-style photograph tracks smoothly with a massive, ghostly white jellyfish as it navigates the abyssal darkness of the deep sea. Maintaining an eye-level perspective with the creature, the camera moves perfectly parallel to its tranquil forward momentum. Its massive, translucent bell contracts in mesmerizing ultra-slow motion, a powerful yet graceful pulse rippling across its gelatinous surface and displacing water in visible crystalline vortices. The only light comes from the creature's own soft, internal bioluminescence, casting a milky white glow that illuminates the inky blue waters and reveals an intricate network of canals within its body. Following this beating heart, countless trailing tentacles, a mixture of delicate, ethereal strands and thicker, ribbon-like oral arms, drift behind in a complex, fluid ballet, hypnotically undulating and spiraling. The ultra-slow motion captures every tiny detail with crystal clarity – the subtle texture of the chimes, the delicate spiral dance of each tentacle, and the microscopic grains of marine snow drifting lazily by, creating a hypnotic and breathtaking visual spectacle of otherworldly beauty and profound silence.", + "seed": 6513, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000095/group_14/index_4472_seed_6513_dp_28_rank_56_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 14, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.02475091451779008, + "vlm_logprob": -0.02475091451779008, + "vlm_yes_no_margin": 4.275, + "vlm_token_credit": -0.02475091451779008, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a massive, ghostly white jellyfish?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the setting the abyssal darkness of the deep sea with inky blue water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the jellyfish's bell contract in ultra-slow motion with visible crystalline vortices displacing water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are countless trailing tentacles and oral arms visible drifting behind the jellyfish in an undulating motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera moving perfectly parallel to the jellyfish at an eye-level perspective?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.006590535398572683 + ], + "opposite_logprob_tokens": [ + -5.506590366363525 + ], + "logprob_reward": -0.006590535398572683, + "yes_no_margin_reward": 5.5, + "token_credit_reward": -0.006590535398572683, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.18931181728839874, + 0.09607645869255066, + 0.07448049634695053, + 0.09917504340410233, + 0.06600525230169296, + 0.11698741465806961, + 0.08398395776748657, + 0.1522749364376068, + 0.04164453223347664, + 0.08006008714437485 + ], + "credit_map": [ + 0.024655429646372795, + 0.015312792733311653, + 0.024662980809807777, + 0.020303474739193916, + 0.024456923827528954, + 0.01698184199631214, + 0.03271229565143585, + 0.030226081609725952, + 0.01543192844837904, + 0.008903986774384975, + 0.013256249949336052, + 0.010506723076105118, + 0.014287426136434078, + 0.011033478192985058, + 0.012087006121873856, + 0.010569656267762184, + 0.010013191029429436, + 0.00819984171539545, + 0.009176387451589108, + 0.010748563334345818, + 0.007176773156970739, + 0.008913070894777775, + 0.010792200453579426, + 0.00946046318858862, + 0.013818877749145031, + 0.0075473240576684475, + 0.01269451528787613, + 0.01876317709684372, + 0.011346228420734406, + 0.01152452640235424, + 0.014182881452143192, + 0.009297506883740425, + 0.00809570588171482, + 0.00794044230133295, + 0.008780228905379772, + 0.008554655127227306, + 0.007559322752058506, + 0.006790605373680592, + 0.011096913367509842, + 0.007187375798821449, + 0.016237054020166397, + 0.013906462118029594, + 0.01595151796936989, + 0.01550219301134348, + 0.011486723087728024, + 0.011040466837584972, + 0.015668991953134537, + 0.017193995416164398, + 0.00958501547574997, + 0.01138488668948412, + 0.010692143812775612, + 0.011350590735673904, + 0.008054327219724655, + 0.013131335377693176, + 0.008927189745008945, + 0.010858462192118168, + 0.01738767698407173, + 0.01808024011552334, + 0.02001497521996498, + 0.018236637115478516, + 0.01899723708629608, + 0.018053734675049782, + 0.028210779651999474, + 0.013293652795255184, + 0.005170745309442282, + 0.005550631787627935, + 0.00557233951985836, + 0.005616120994091034, + 0.007069988641887903, + 0.0031035318970680237, + 0.006426357664167881, + 0.0031348152551800013, + 0.013308084569871426, + 0.007144639268517494, + 0.012839622795581818, + 0.00741227250546217, + 0.006447880994528532, + 0.009986034594476223, + 0.007249905727803707, + 0.01567164622247219 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 7, + 5 + ], + "selected_reward": -0.006590535398572683 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a massive, ghostly white jellyfish?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the setting the abyssal darkness of the deep sea with inky blue water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the jellyfish's bell contract in ultra-slow motion with visible crystalline vortices displacing water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are countless trailing tentacles and oral arms visible drifting behind the jellyfish in an undulating motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera moving perfectly parallel to the jellyfish at an eye-level perspective?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.007632853463292122 + ], + "opposite_logprob_tokens": [ + -4.882632732391357 + ], + "logprob_reward": -0.007632853463292122, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.007632853463292122, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15707361698150635, + 0.11581111699342728, + 0.1077534481883049, + 0.05192139744758606, + 0.08529072999954224, + 0.07308419793844223, + 0.10041704773902893, + 0.10004667192697525, + 0.08151140064001083, + 0.12709036469459534 + ], + "credit_map": [ + 0.01754571497440338, + 0.012531059794127941, + 0.019288823008537292, + 0.014953085221350193, + 0.017958324402570724, + 0.014459837228059769, + 0.025924548506736755, + 0.03441222757101059, + 0.019683577120304108, + 0.009984503500163555, + 0.01761600375175476, + 0.01604091189801693, + 0.01060572825372219, + 0.015919074416160583, + 0.01153081189841032, + 0.014430511742830276, + 0.012817067094147205, + 0.009355741553008556, + 0.012992923147976398, + 0.019898058846592903, + 0.006419436540454626, + 0.014310822822153568, + 0.010338340885937214, + 0.02162105403840542, + 0.0064112781547009945, + 0.0043128477409482, + 0.007502289023250341, + 0.011533218435943127, + 0.003338393522426486, + 0.006526332814246416, + 0.006258130539208651, + 0.0060389102436602116, + 0.009966859593987465, + 0.008635192178189754, + 0.010905621573328972, + 0.010288510471582413, + 0.009366367943584919, + 0.010827168822288513, + 0.014246579259634018, + 0.011054430156946182, + 0.009249845519661903, + 0.008344762027263641, + 0.009235143661499023, + 0.011583219282329082, + 0.008396362885832787, + 0.009338739328086376, + 0.009151708334684372, + 0.007784419227391481, + 0.012203053571283817, + 0.010578014887869358, + 0.0130724823102355, + 0.016415633261203766, + 0.007644937373697758, + 0.017157230526208878, + 0.011543560773134232, + 0.011802131310105324, + 0.012726441025733948, + 0.0076219262555241585, + 0.01338232308626175, + 0.013983059674501419, + 0.0068532200530171394, + 0.013043250888586044, + 0.017129523679614067, + 0.015306920744478703, + 0.014174515381455421, + 0.008507649414241314, + 0.011562892235815525, + 0.009528196416795254, + 0.008734038099646568, + 0.007485960144549608, + 0.010978488251566887, + 0.010539662092924118, + 0.02277124486863613, + 0.009897775948047638, + 0.021068301051855087, + 0.015688201412558556, + 0.011608723551034927, + 0.01332330983132124, + 0.012762955389916897, + 0.019969835877418518 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 1 + ], + "selected_reward": -0.007632853463292122 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a massive, ghostly white jellyfish?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the setting the abyssal darkness of the deep sea with inky blue water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the jellyfish's bell contract in ultra-slow motion with visible crystalline vortices displacing water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are countless trailing tentacles and oral arms visible drifting behind the jellyfish in an undulating motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera moving perfectly parallel to the jellyfish at an eye-level perspective?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04865988716483116 + ], + "opposite_logprob_tokens": [ + -3.0486598014831543 + ], + "logprob_reward": -0.04865988716483116, + "yes_no_margin_reward": 3.0, + "token_credit_reward": -0.04865988716483116, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.20065978169441223, + 0.07070880383253098, + 0.08825553953647614, + 0.1121772900223732, + 0.06169629842042923, + 0.0833156481385231, + 0.04496629536151886, + 0.2527318000793457, + 0.04491806402802467, + 0.040570490062236786 + ], + "credit_map": [ + 0.022113492712378502, + 0.01743108220398426, + 0.024726223200559616, + 0.017664235085248947, + 0.02573121152818203, + 0.0197673961520195, + 0.03650083765387535, + 0.036725323647260666, + 0.011138376779854298, + 0.006417537573724985, + 0.011471518315374851, + 0.006819440051913261, + 0.006715697236359119, + 0.008172554895281792, + 0.008846249431371689, + 0.011127429082989693, + 0.014782874844968319, + 0.010260563343763351, + 0.013604291714727879, + 0.012113725766539574, + 0.00670933723449707, + 0.011454379186034203, + 0.010132723487913609, + 0.009197639301419258, + 0.013528678566217422, + 0.007201234344393015, + 0.013003362342715263, + 0.020793763920664787, + 0.011863844469189644, + 0.01462278701364994, + 0.019587958231568336, + 0.011575654149055481, + 0.007950864732265472, + 0.007610270753502846, + 0.007808510214090347, + 0.007668382953852415, + 0.006119974888861179, + 0.006845674943178892, + 0.009255971759557724, + 0.008436649106442928, + 0.008061232045292854, + 0.012989264912903309, + 0.010097325779497623, + 0.009506411850452423, + 0.011078578419983387, + 0.00868939608335495, + 0.013763251714408398, + 0.009130188263952732, + 0.005380549468100071, + 0.004251768346875906, + 0.005870523862540722, + 0.004304796457290649, + 0.0034212497994303703, + 0.009203863330185413, + 0.007151576690375805, + 0.005381968338042498, + 0.028691913932561874, + 0.02873888611793518, + 0.034333065152168274, + 0.03138105571269989, + 0.03223276138305664, + 0.03207153081893921, + 0.044714026153087616, + 0.02056855708360672, + 0.0058259074576199055, + 0.004855982959270477, + 0.005577378906309605, + 0.006225199904292822, + 0.005823587998747826, + 0.005112688988447189, + 0.006151264533400536, + 0.005346049088984728, + 0.0074002486653625965, + 0.003911763895303011, + 0.005644084420055151, + 0.004026134964078665, + 0.003659873967990279, + 0.0049599893391132355, + 0.004571895580738783, + 0.006396500859409571 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 0, + 3 + ], + "selected_reward": -0.04865988716483116 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a massive, ghostly white jellyfish?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the setting the abyssal darkness of the deep sea with inky blue water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the jellyfish's bell contract in ultra-slow motion with visible crystalline vortices displacing water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are countless trailing tentacles and oral arms visible drifting behind the jellyfish in an undulating motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera moving perfectly parallel to the jellyfish at an eye-level perspective?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0059441933408379555 + ], + "opposite_logprob_tokens": [ + -5.13094425201416 + ], + "logprob_reward": -0.0059441933408379555, + "yes_no_margin_reward": 5.125, + "token_credit_reward": -0.0059441933408379555, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09524990618228912, + 0.061205796897411346, + 0.08888517320156097, + 0.04599802568554878, + 0.08649653941392899, + 0.2429818958044052, + 0.03183705732226372, + 0.04595889896154404, + 0.033100757747888565, + 0.2682859003543854 + ], + "credit_map": [ + 0.013490901328623295, + 0.0071106660179793835, + 0.012184870429337025, + 0.009602692909538746, + 0.009025361388921738, + 0.010946152731776237, + 0.01542341336607933, + 0.01746586337685585, + 0.01067422330379486, + 0.005682681687176228, + 0.009054767899215221, + 0.0048746755346655846, + 0.008275209926068783, + 0.007754607126116753, + 0.007532052230089903, + 0.007357582915574312, + 0.008666355162858963, + 0.006955196615308523, + 0.012352894060313702, + 0.014172137714922428, + 0.005851391702890396, + 0.009351587854325771, + 0.012154270894825459, + 0.019381340593099594, + 0.005381667520850897, + 0.0040496764704585075, + 0.005587344523519278, + 0.005503850989043713, + 0.005793689284473658, + 0.008526504971086979, + 0.006902544759213924, + 0.0042527527548372746, + 0.011411563493311405, + 0.009002338163554668, + 0.011777893640100956, + 0.007378122769296169, + 0.007572632748633623, + 0.01477828435599804, + 0.015191514976322651, + 0.009384198114275932, + 0.02221779339015484, + 0.03564278036355972, + 0.02994689717888832, + 0.03873071074485779, + 0.031007425859570503, + 0.020459625869989395, + 0.04608204960823059, + 0.018894637003540993, + 0.0035736379213631153, + 0.0027974427212029696, + 0.004109653178602457, + 0.0033615927677601576, + 0.003927240148186684, + 0.00435481546446681, + 0.006334247067570686, + 0.0033784329425543547, + 0.005394378211349249, + 0.005074089393019676, + 0.006072232499718666, + 0.006053325720131397, + 0.005638123489916325, + 0.004889053758233786, + 0.008600243367254734, + 0.004237455781549215, + 0.0037728555034846067, + 0.0036082095466554165, + 0.00446110125631094, + 0.005495831836014986, + 0.0036712137516587973, + 0.0028001528698951006, + 0.005530533380806446, + 0.0037608614657074213, + 0.023624679073691368, + 0.02439224161207676, + 0.03753981739282608, + 0.028394926339387894, + 0.0317121259868145, + 0.0345747135579586, + 0.039426084607839584, + 0.04862132668495178 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 5, + 0 + ], + "selected_reward": -0.0059441933408379555 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a massive, ghostly white jellyfish?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the setting the abyssal darkness of the deep sea with inky blue water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the jellyfish's bell contract in ultra-slow motion with visible crystalline vortices displacing water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are countless trailing tentacles and oral arms visible drifting behind the jellyfish in an undulating motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera moving perfectly parallel to the jellyfish at an eye-level perspective?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.05492710322141647 + ], + "opposite_logprob_tokens": [ + -2.929927110671997 + ], + "logprob_reward": -0.05492710322141647, + "yes_no_margin_reward": 2.875, + "token_credit_reward": -0.05492710322141647, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14763043820858002, + 0.13289734721183777, + 0.05893927812576294, + 0.09138381481170654, + 0.11724663525819778, + 0.15186968445777893, + 0.0714799016714096, + 0.075736865401268, + 0.11438404023647308, + 0.038431987166404724 + ], + "credit_map": [ + 0.023944251239299774, + 0.014168907888233662, + 0.02062159776687622, + 0.014637462794780731, + 0.017537394538521767, + 0.015745053067803383, + 0.018849674612283707, + 0.0221260953694582, + 0.024445340037345886, + 0.010891037061810493, + 0.02070506475865841, + 0.010246523655951023, + 0.020148612558841705, + 0.014854860492050648, + 0.01861947402358055, + 0.012986447662115097, + 0.009913703426718712, + 0.006779039744287729, + 0.008504576981067657, + 0.007541432045400143, + 0.004590687341988087, + 0.006839026231318712, + 0.007333675865083933, + 0.007437133230268955, + 0.011579478159546852, + 0.011166306212544441, + 0.011567622423171997, + 0.014897850342094898, + 0.009327593259513378, + 0.01211618073284626, + 0.010534816421568394, + 0.010193966329097748, + 0.01596180908381939, + 0.019292980432510376, + 0.01686899922788143, + 0.008777967654168606, + 0.00791037455201149, + 0.01772323064506054, + 0.015289236791431904, + 0.015422025695443153, + 0.02057819627225399, + 0.01998361572623253, + 0.019696956500411034, + 0.017870023846626282, + 0.01982361450791359, + 0.014132355339825153, + 0.02375430054962635, + 0.016030622646212578, + 0.01008189283311367, + 0.005382598377764225, + 0.01046901848167181, + 0.007599262520670891, + 0.008571654558181763, + 0.007167623843997717, + 0.014336539432406425, + 0.007871309295296669, + 0.00951212365180254, + 0.008682889863848686, + 0.009938115254044533, + 0.009640094824135303, + 0.008953617885708809, + 0.009209777228534222, + 0.01381014846265316, + 0.005990094039589167, + 0.014233351685106754, + 0.009990943595767021, + 0.015029079280793667, + 0.01869012601673603, + 0.009040077216923237, + 0.012420164421200752, + 0.018980083987116814, + 0.016000214964151382, + 0.005646490957587957, + 0.0038451082073152065, + 0.005331581924110651, + 0.0048924102447927, + 0.0030557371210306883, + 0.0043275970965623856, + 0.0042863707058131695, + 0.00704669114202261 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 0, + 1 + ], + "selected_reward": -0.05492710322141647 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "An ultra-detailed, high-speed phantom camera-style photograph tracks smoothly with a massive, ghostly white jellyfish as it navigates the abyssal darkness of the deep sea. Maintaining an eye-level perspective with the creature, the camera moves perfectly parallel to its tranquil forward momentum. Its massive, translucent bell contracts in mesmerizing ultra-slow motion, a powerful yet graceful pulse rippling across its gelatinous surface and displacing water in visible crystalline vortices. The only light comes from the creature's own soft, internal bioluminescence, casting a milky white glow that illuminates the inky blue waters and reveals an intricate network of canals within its body. Following this beating heart, countless trailing tentacles, a mixture of delicate, ethereal strands and thicker, ribbon-like oral arms, drift behind in a complex, fluid ballet, hypnotically undulating and spiraling. The ultra-slow motion captures every tiny detail with crystal clarity – the subtle texture of the chimes, the delicate spiral dance of each tentacle, and the microscopic grains of marine snow drifting lazily by, creating a hypnotic and breathtaking visual spectacle of otherworldly beauty and profound silence.", + "seed": 6514, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000095/group_14/index_4472_seed_6514_dp_28_rank_56_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 14, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.015158358123153448, + "vlm_logprob": -0.015158358123153448, + "vlm_yes_no_margin": 5.0, + "vlm_token_credit": -0.015158358123153448, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a massive, ghostly white jellyfish?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the setting the abyssal darkness of the deep sea with inky blue water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the jellyfish's bell contract in ultra-slow motion with visible crystalline vortices displacing water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are countless trailing tentacles and oral arms visible drifting behind the jellyfish in an undulating motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera moving perfectly parallel to the jellyfish at an eye-level perspective?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004145719110965729 + ], + "opposite_logprob_tokens": [ + -6.254145622253418 + ], + "logprob_reward": -0.004145719110965729, + "yes_no_margin_reward": 6.25, + "token_credit_reward": -0.004145719110965729, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.19696387648582458, + 0.22645817697048187, + 0.06839914619922638, + 0.0868222787976265, + 0.13916486501693726, + 0.036481648683547974, + 0.04814491048455238, + 0.06265036016702652, + 0.07758796960115433, + 0.057326789945364 + ], + "credit_map": [ + 0.03703230991959572, + 0.028434108942747116, + 0.025290759280323982, + 0.021908288821578026, + 0.019727597013115883, + 0.014898648485541344, + 0.03450228273868561, + 0.015169882215559483, + 0.035664185881614685, + 0.029462771490216255, + 0.028279496356844902, + 0.03345285728573799, + 0.022326625883579254, + 0.029432866722345352, + 0.027004890143871307, + 0.020834481343626976, + 0.009509886614978313, + 0.00895985122770071, + 0.011164204217493534, + 0.0064347172155976295, + 0.00730121647939086, + 0.006886348593980074, + 0.010306804440915585, + 0.0078361164778471, + 0.011685582809150219, + 0.008785809390246868, + 0.009536328725516796, + 0.00691733555868268, + 0.00697357440367341, + 0.011644238606095314, + 0.01728382706642151, + 0.01399558037519455, + 0.015333523042500019, + 0.012186220847070217, + 0.012265404686331749, + 0.013270946219563484, + 0.019817804917693138, + 0.026839058846235275, + 0.0201182272285223, + 0.019333671778440475, + 0.0056283543817698956, + 0.002960412995889783, + 0.005322710145264864, + 0.0046124146319925785, + 0.0039384993724524975, + 0.0039613377302885056, + 0.004467780236154795, + 0.005590135231614113, + 0.00636146729812026, + 0.004565409384667873, + 0.005624391604214907, + 0.005116886459290981, + 0.005866513587534428, + 0.005061659030616283, + 0.0068566943518817425, + 0.00869189016520977, + 0.006290842313319445, + 0.007654127664864063, + 0.006771547719836235, + 0.009672210551798344, + 0.006062282714992762, + 0.010891668498516083, + 0.00888154562562704, + 0.006426132749766111, + 0.008147938176989555, + 0.007131625432521105, + 0.010294313542544842, + 0.01201819721609354, + 0.008268256671726704, + 0.008334568701684475, + 0.011217380873858929, + 0.012175685726106167, + 0.007093613967299461, + 0.0063160983845591545, + 0.005406422074884176, + 0.004269562195986509, + 0.0074500879272818565, + 0.008848345838487148, + 0.008742411620914936, + 0.009200247935950756 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 4 + ], + "selected_reward": -0.004145719110965729 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a massive, ghostly white jellyfish?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the setting the abyssal darkness of the deep sea with inky blue water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the jellyfish's bell contract in ultra-slow motion with visible crystalline vortices displacing water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are countless trailing tentacles and oral arms visible drifting behind the jellyfish in an undulating motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera moving perfectly parallel to the jellyfish at an eye-level perspective?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0028248426970094442 + ], + "opposite_logprob_tokens": [ + -5.877824783325195 + ], + "logprob_reward": -0.0028248426970094442, + "yes_no_margin_reward": 5.875, + "token_credit_reward": -0.0028248426970094442, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.25784584879875183, + 0.10943052917718887, + 0.07707042247056961, + 0.1424172818660736, + 0.17871896922588348, + 0.062047380954027176, + 0.04278057441115379, + 0.03009852021932602, + 0.06068753823637962, + 0.038902852684259415 + ], + "credit_map": [ + 0.038911547511816025, + 0.02813728339970112, + 0.024884942919015884, + 0.032376501709222794, + 0.02472430095076561, + 0.017266007140278816, + 0.062153082340955734, + 0.02939220517873764, + 0.01574948988854885, + 0.012827777303755283, + 0.010665984824299812, + 0.0145109286531806, + 0.011804708279669285, + 0.019541360437870026, + 0.013724098913371563, + 0.010606193915009499, + 0.011264999397099018, + 0.007729690987616777, + 0.012608741410076618, + 0.00823276862502098, + 0.008620476350188255, + 0.009129778482019901, + 0.009074010886251926, + 0.010409967973828316, + 0.029200060293078423, + 0.012120709754526615, + 0.018505612388253212, + 0.013590755872428417, + 0.01884131133556366, + 0.012030481360852718, + 0.019606279209256172, + 0.018522078171372414, + 0.02510129287838936, + 0.012974615208804607, + 0.02258000522851944, + 0.027582796290516853, + 0.021092835813760757, + 0.01683816872537136, + 0.028739485889673233, + 0.02380979433655739, + 0.008818291127681732, + 0.005246724467724562, + 0.008347016759216785, + 0.009897003881633282, + 0.006799114868044853, + 0.006205490790307522, + 0.007598764728754759, + 0.009134983643889427, + 0.006927963346242905, + 0.004281897097826004, + 0.0060121905989944935, + 0.005020948592573404, + 0.003640698501840234, + 0.0032696258276700974, + 0.006195216439664364, + 0.007432037964463234, + 0.004519924521446228, + 0.0021224436350166798, + 0.002281115623190999, + 0.002479387680068612, + 0.0027916859835386276, + 0.0029125537257641554, + 0.006353361066430807, + 0.006638048682361841, + 0.005502745509147644, + 0.00583341671153903, + 0.0068625835701823235, + 0.009482854045927525, + 0.007026351522654295, + 0.0071820602752268314, + 0.008734045550227165, + 0.010063487105071545, + 0.0046884347684681416, + 0.0030382568947970867, + 0.004084298387169838, + 0.004768253304064274, + 0.005581843201071024, + 0.005729498341679573, + 0.003804905107244849, + 0.007207367103546858 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 3 + ], + "selected_reward": -0.0028248426970094442 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a massive, ghostly white jellyfish?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the setting the abyssal darkness of the deep sea with inky blue water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the jellyfish's bell contract in ultra-slow motion with visible crystalline vortices displacing water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are countless trailing tentacles and oral arms visible drifting behind the jellyfish in an undulating motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera moving perfectly parallel to the jellyfish at an eye-level perspective?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02329576015472412 + ], + "opposite_logprob_tokens": [ + -3.7732958793640137 + ], + "logprob_reward": -0.02329576015472412, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.02329576015472412, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.31227415800094604, + 0.12303928285837173, + 0.051638081669807434, + 0.06352715939283371, + 0.0818648636341095, + 0.0882759615778923, + 0.03528236970305443, + 0.17976878583431244, + 0.0428607277572155, + 0.021468602120876312 + ], + "credit_map": [ + 0.05124160274863243, + 0.0416250079870224, + 0.034737199544906616, + 0.03769158199429512, + 0.0294667799025774, + 0.02280130796134472, + 0.062053799629211426, + 0.032656874507665634, + 0.017608726397156715, + 0.01516043208539486, + 0.012377277947962284, + 0.015483501367270947, + 0.013876124285161495, + 0.02026410959661007, + 0.01595662534236908, + 0.012312483973801136, + 0.00693057756870985, + 0.005560848396271467, + 0.004496308043599129, + 0.00456456420943141, + 0.008275950327515602, + 0.007521076127886772, + 0.008525227196514606, + 0.005763529799878597, + 0.00856813695281744, + 0.004557406064122915, + 0.006733315531164408, + 0.005944730713963509, + 0.005628669168800116, + 0.008498254232108593, + 0.01302008330821991, + 0.010576561093330383, + 0.011055465787649155, + 0.007790523115545511, + 0.009143185801804066, + 0.01027307752519846, + 0.011538952589035034, + 0.010498665273189545, + 0.009700404480099678, + 0.011864587664604187, + 0.011431553401052952, + 0.008391547948122025, + 0.012617039494216442, + 0.012821749784052372, + 0.010038251988589764, + 0.007856295444071293, + 0.011440824717283249, + 0.013678697869181633, + 0.005050382576882839, + 0.003016218077391386, + 0.0030470711644738913, + 0.003745083464309573, + 0.006560497917234898, + 0.0038528344593942165, + 0.004660783335566521, + 0.005349500570446253, + 0.016834108158946037, + 0.014047681353986263, + 0.013538437895476818, + 0.01739388145506382, + 0.01845509186387062, + 0.05666591599583626, + 0.01672641932964325, + 0.026107262820005417, + 0.005275025963783264, + 0.003938851412385702, + 0.006179602351039648, + 0.005067032761871815, + 0.004779641050845385, + 0.0055168489925563335, + 0.005219023674726486, + 0.006884703878313303, + 0.002140024211257696, + 0.002524759853258729, + 0.001893637585453689, + 0.00304116471670568, + 0.0034037744626402855, + 0.003152579767629504, + 0.002778684487566352, + 0.002533977385610342 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 7, + 1 + ], + "selected_reward": -0.02329576015472412 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a massive, ghostly white jellyfish?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the setting the abyssal darkness of the deep sea with inky blue water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the jellyfish's bell contract in ultra-slow motion with visible crystalline vortices displacing water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are countless trailing tentacles and oral arms visible drifting behind the jellyfish in an undulating motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera moving perfectly parallel to the jellyfish at an eye-level perspective?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0024873053189367056 + ], + "opposite_logprob_tokens": [ + -6.0024871826171875 + ], + "logprob_reward": -0.0024873053189367056, + "yes_no_margin_reward": 6.0, + "token_credit_reward": -0.0024873053189367056, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2820046842098236, + 0.24539467692375183, + 0.07767646759748459, + 0.057681676000356674, + 0.08184202015399933, + 0.03801004961133003, + 0.027286915108561516, + 0.07539784163236618, + 0.06795769184827805, + 0.046747900545597076 + ], + "credit_map": [ + 0.0411209799349308, + 0.0336289256811142, + 0.0337526798248291, + 0.030205314978957176, + 0.026584140956401825, + 0.021411417052149773, + 0.06567666679620743, + 0.02962459623813629, + 0.03623584657907486, + 0.024826176464557648, + 0.03410002961754799, + 0.04292389377951622, + 0.026819437742233276, + 0.028912033885717392, + 0.02966434508562088, + 0.021912934258580208, + 0.010897241532802582, + 0.009451743215322495, + 0.009790186770260334, + 0.008201572112739086, + 0.012103430926799774, + 0.008772696368396282, + 0.010515064001083374, + 0.00794453639537096, + 0.009626456536352634, + 0.006534900516271591, + 0.006763805169612169, + 0.0060604410246014595, + 0.005104350857436657, + 0.005165455862879753, + 0.008799257688224316, + 0.009627014398574829, + 0.01116473600268364, + 0.005973642226308584, + 0.011253277771174908, + 0.009988791309297085, + 0.00958941224962473, + 0.009807360358536243, + 0.013266602531075478, + 0.010798207484185696, + 0.005838514305651188, + 0.0033575394190847874, + 0.007260056212544441, + 0.00452269334346056, + 0.0040820580907166, + 0.00317060062661767, + 0.004406623542308807, + 0.005371965933591127, + 0.003884713165462017, + 0.0025338404811918736, + 0.0026733882259577513, + 0.0028361957520246506, + 0.0038814288564026356, + 0.0026559499092400074, + 0.003797843586653471, + 0.00502355769276619, + 0.007868911139667034, + 0.006282806396484375, + 0.007144262548536062, + 0.010582265444099903, + 0.007261018268764019, + 0.01816975325345993, + 0.009032201021909714, + 0.009056631475687027, + 0.008422612212598324, + 0.007668760605156422, + 0.013092584908008575, + 0.009849658235907555, + 0.004876862280070782, + 0.0069159287959337234, + 0.005751636810600758, + 0.011379655450582504, + 0.005867293104529381, + 0.004887554328888655, + 0.0052026621997356415, + 0.005297302268445492, + 0.00626863818615675, + 0.006296621635556221, + 0.005112489685416222, + 0.007815346121788025 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 4 + ], + "selected_reward": -0.0024873053189367056 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a massive, ghostly white jellyfish?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the setting the abyssal darkness of the deep sea with inky blue water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the jellyfish's bell contract in ultra-slow motion with visible crystalline vortices displacing water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are countless trailing tentacles and oral arms visible drifting behind the jellyfish in an undulating motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera moving perfectly parallel to the jellyfish at an eye-level perspective?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04303816333413124 + ], + "opposite_logprob_tokens": [ + -3.1680381298065186 + ], + "logprob_reward": -0.04303816333413124, + "yes_no_margin_reward": 3.125, + "token_credit_reward": -0.04303816333413124, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2885749340057373, + 0.04690823704004288, + 0.055838096886873245, + 0.11907289922237396, + 0.06116149574518204, + 0.0680590271949768, + 0.08960691094398499, + 0.21653755009174347, + 0.02332041598856449, + 0.030920421704649925 + ], + "credit_map": [ + 0.057878248393535614, + 0.046533215790987015, + 0.035733528435230255, + 0.03334624320268631, + 0.031177429482340813, + 0.019423840567469597, + 0.04168381169438362, + 0.022798635065555573, + 0.006967855617403984, + 0.007297799922525883, + 0.006648238282650709, + 0.004276569467037916, + 0.005312399007380009, + 0.00603231368586421, + 0.005734974052757025, + 0.004638089332729578, + 0.008479633368551731, + 0.004588893614709377, + 0.008557184599339962, + 0.007218310609459877, + 0.005756695289164782, + 0.005899437703192234, + 0.006802218034863472, + 0.00853572878986597, + 0.018602848052978516, + 0.015696287155151367, + 0.015778079628944397, + 0.00901646539568901, + 0.015165263786911964, + 0.014645013026893139, + 0.015579147264361382, + 0.014589820057153702, + 0.007768144365400076, + 0.005284792743623257, + 0.005849367938935757, + 0.00530086038634181, + 0.008069027215242386, + 0.010555393062531948, + 0.009173346683382988, + 0.009160569868981838, + 0.008654534816741943, + 0.005904817022383213, + 0.009460865519940853, + 0.008092879317700863, + 0.008883101865649223, + 0.006284563336521387, + 0.012765444815158844, + 0.008012826554477215, + 0.010862749069929123, + 0.007275547832250595, + 0.007219888269901276, + 0.00943300873041153, + 0.01618545688688755, + 0.011320564895868301, + 0.014409391209483147, + 0.012900318950414658, + 0.019199658185243607, + 0.01794602908194065, + 0.017802676185965538, + 0.024624455720186234, + 0.022007256746292114, + 0.06621664017438889, + 0.029986202716827393, + 0.01875465363264084, + 0.0031742239370942116, + 0.002311425283551216, + 0.00370065332390368, + 0.002028390299528837, + 0.002140989527106285, + 0.0026770157273858786, + 0.0038915721233934164, + 0.0033961469307541847, + 0.0034255865029990673, + 0.0027809804305434227, + 0.0037168937269598246, + 0.0034323905128985643, + 0.005044631659984589, + 0.004614303354173899, + 0.0035725256893783808, + 0.004333114717155695 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 7, + 3 + ], + "selected_reward": -0.04303816333413124 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "An ultra-detailed, high-speed phantom camera-style photograph tracks smoothly with a massive, ghostly white jellyfish as it navigates the abyssal darkness of the deep sea. Maintaining an eye-level perspective with the creature, the camera moves perfectly parallel to its tranquil forward momentum. Its massive, translucent bell contracts in mesmerizing ultra-slow motion, a powerful yet graceful pulse rippling across its gelatinous surface and displacing water in visible crystalline vortices. The only light comes from the creature's own soft, internal bioluminescence, casting a milky white glow that illuminates the inky blue waters and reveals an intricate network of canals within its body. Following this beating heart, countless trailing tentacles, a mixture of delicate, ethereal strands and thicker, ribbon-like oral arms, drift behind in a complex, fluid ballet, hypnotically undulating and spiraling. The ultra-slow motion captures every tiny detail with crystal clarity – the subtle texture of the chimes, the delicate spiral dance of each tentacle, and the microscopic grains of marine snow drifting lazily by, creating a hypnotic and breathtaking visual spectacle of otherworldly beauty and profound silence.", + "seed": 6515, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000095/group_14/index_4472_seed_6515_dp_29_rank_58_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 14, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.014214873989112675, + "vlm_logprob": -0.014214873989112675, + "vlm_yes_no_margin": 5.425, + "vlm_token_credit": -0.014214873989112675, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a massive, ghostly white jellyfish?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the setting the abyssal darkness of the deep sea with inky blue water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the jellyfish's bell contract in ultra-slow motion with visible crystalline vortices displacing water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are countless trailing tentacles and oral arms visible drifting behind the jellyfish in an undulating motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera moving perfectly parallel to the jellyfish at an eye-level perspective?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0031264969147741795 + ], + "opposite_logprob_tokens": [ + -6.753126621246338 + ], + "logprob_reward": -0.0031264969147741795, + "yes_no_margin_reward": 6.75, + "token_credit_reward": -0.0031264969147741795, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0931306779384613, + 0.0858478769659996, + 0.12859396636486053, + 0.12380249798297882, + 0.06874491274356842, + 0.0462467186152935, + 0.18796341121196747, + 0.1354375183582306, + 0.07153497636318207, + 0.0586974136531353 + ], + "credit_map": [ + 0.00823137816041708, + 0.0062508052214980125, + 0.016101252287626266, + 0.006126755848526955, + 0.01650274731218815, + 0.006507692392915487, + 0.014853165484964848, + 0.018556887283921242, + 0.013014827854931355, + 0.010424643754959106, + 0.013677858747541904, + 0.00660511152818799, + 0.011827192269265652, + 0.005405877251178026, + 0.012866654433310032, + 0.012025723233819008, + 0.01033089216798544, + 0.016173891723155975, + 0.011438616551458836, + 0.014837795868515968, + 0.015435989946126938, + 0.02085713855922222, + 0.023154661059379578, + 0.01636497862637043, + 0.011217956431210041, + 0.01965184696018696, + 0.012144196778535843, + 0.017844881862401962, + 0.013092677108943462, + 0.018416941165924072, + 0.017077475786209106, + 0.014356533996760845, + 0.005222022067755461, + 0.006866406183689833, + 0.007002087309956551, + 0.012627050280570984, + 0.007835198193788528, + 0.009634419344365597, + 0.011284241452813148, + 0.008273487910628319, + 0.004751747008413076, + 0.004071963019669056, + 0.006216647103428841, + 0.005356867332011461, + 0.004103099461644888, + 0.006389325950294733, + 0.007821213454008102, + 0.0075358618050813675, + 0.013190053403377533, + 0.0116103645414114, + 0.02986118569970131, + 0.019668441265821457, + 0.047141462564468384, + 0.035090718418359756, + 0.019300851970911026, + 0.012100339867174625, + 0.01458844542503357, + 0.013275813311338425, + 0.017873618751764297, + 0.020179834216833115, + 0.0157127995043993, + 0.021668341010808945, + 0.012540359981358051, + 0.01959831453859806, + 0.005125494208186865, + 0.0045718722976744175, + 0.010164524428546429, + 0.009235554374754429, + 0.008458508178591728, + 0.008953049778938293, + 0.011899652890861034, + 0.013126319274306297, + 0.0055148969404399395, + 0.004597908351570368, + 0.007588877342641354, + 0.00670615304261446, + 0.007673916406929493, + 0.007214255165308714, + 0.008071965537965298, + 0.01132944691926241 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 7, + 2 + ], + "selected_reward": -0.0031264969147741795 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a massive, ghostly white jellyfish?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the setting the abyssal darkness of the deep sea with inky blue water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the jellyfish's bell contract in ultra-slow motion with visible crystalline vortices displacing water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are countless trailing tentacles and oral arms visible drifting behind the jellyfish in an undulating motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera moving perfectly parallel to the jellyfish at an eye-level perspective?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.001718117156997323 + ], + "opposite_logprob_tokens": [ + -6.376718044281006 + ], + "logprob_reward": -0.001718117156997323, + "yes_no_margin_reward": 6.375, + "token_credit_reward": -0.001718117156997323, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11524748802185059, + 0.07047401368618011, + 0.08367457240819931, + 0.029185699298977852, + 0.14623308181762695, + 0.1467382162809372, + 0.14591383934020996, + 0.055159032344818115, + 0.14643338322639465, + 0.06094067916274071 + ], + "credit_map": [ + 0.012935408391058445, + 0.006748782470822334, + 0.03764314949512482, + 0.007906030863523483, + 0.008307789452373981, + 0.005166210234165192, + 0.009065928868949413, + 0.027474191039800644, + 0.008497641421854496, + 0.004653402604162693, + 0.014580443501472473, + 0.003461558138951659, + 0.008829861879348755, + 0.0066818189807236195, + 0.008549853228032589, + 0.015219435095787048, + 0.014217334799468517, + 0.007553085684776306, + 0.009946752339601517, + 0.006411075592041016, + 0.013668213039636612, + 0.014540315605700016, + 0.009392108768224716, + 0.007945682853460312, + 0.004375345539301634, + 0.005491663236171007, + 0.00393571937456727, + 0.0028754454106092453, + 0.003239343874156475, + 0.0033201149199157953, + 0.0031649419106543064, + 0.0027831243351101875, + 0.016394011676311493, + 0.01654411293566227, + 0.016388773918151855, + 0.023554688319563866, + 0.013566777110099792, + 0.021716482937335968, + 0.02128162793815136, + 0.016786618158221245, + 0.016339782625436783, + 0.01695641316473484, + 0.02153118886053562, + 0.01832794025540352, + 0.013495744206011295, + 0.022254055365920067, + 0.02058872953057289, + 0.017244359478354454, + 0.011833288706839085, + 0.009745372459292412, + 0.026495182886719704, + 0.01779911108314991, + 0.03271836042404175, + 0.02450677938759327, + 0.013514156453311443, + 0.009301576763391495, + 0.008914101868867874, + 0.006101708393543959, + 0.006947156507521868, + 0.006196911446750164, + 0.0064427824690938, + 0.009723479859530926, + 0.003974409773945808, + 0.006858483422547579, + 0.012763489969074726, + 0.009888295084238052, + 0.01908279024064541, + 0.016597270965576172, + 0.022503476589918137, + 0.019977496936917305, + 0.02078741230070591, + 0.024833155795931816, + 0.00561478128656745, + 0.0058536045253276825, + 0.009143436327576637, + 0.008152099326252937, + 0.006665358319878578, + 0.006229657679796219, + 0.009642005898058414, + 0.009639738127589226 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 8, + 4 + ], + "selected_reward": -0.001718117156997323 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a massive, ghostly white jellyfish?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the setting the abyssal darkness of the deep sea with inky blue water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the jellyfish's bell contract in ultra-slow motion with visible crystalline vortices displacing water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are countless trailing tentacles and oral arms visible drifting behind the jellyfish in an undulating motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera moving perfectly parallel to the jellyfish at an eye-level perspective?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.009791087359189987 + ], + "opposite_logprob_tokens": [ + -4.634790897369385 + ], + "logprob_reward": -0.009791087359189987, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.009791087359189987, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08180736005306244, + 0.21408553421497345, + 0.08822104334831238, + 0.06942901760339737, + 0.0937499850988388, + 0.04874667525291443, + 0.18718241155147552, + 0.12269686162471771, + 0.046769753098487854, + 0.04731135815382004 + ], + "credit_map": [ + 0.009774573147296906, + 0.007582998368889093, + 0.016489749774336815, + 0.006635497324168682, + 0.018993442878127098, + 0.006166321225464344, + 0.010500152595341206, + 0.005664620082825422, + 0.02530737593770027, + 0.015316102653741837, + 0.02623579278588295, + 0.01168063748627901, + 0.03963728994131088, + 0.03558032587170601, + 0.02502262219786644, + 0.03530540317296982, + 0.009196189232170582, + 0.010819380171597004, + 0.008857020176947117, + 0.009674442932009697, + 0.009268836118280888, + 0.01249756757169962, + 0.015835266560316086, + 0.012072338722646236, + 0.0065597123466432095, + 0.01061961054801941, + 0.006717110518366098, + 0.010109513998031616, + 0.0072780791670084, + 0.010638018138706684, + 0.008695874363183975, + 0.008811100386083126, + 0.007499952334910631, + 0.010458680801093578, + 0.00956309400498867, + 0.013238622806966305, + 0.01058819517493248, + 0.016298241913318634, + 0.015859095379710197, + 0.010244105011224747, + 0.005402692593634129, + 0.00415525957942009, + 0.006644498091191053, + 0.006076634395867586, + 0.0048028710298240185, + 0.0064055160619318485, + 0.009415696375072002, + 0.005843510385602713, + 0.014870699495077133, + 0.009321507066488266, + 0.03275708854198456, + 0.020539451390504837, + 0.051134612411260605, + 0.03516889736056328, + 0.014418493025004864, + 0.00897165946662426, + 0.014843817800283432, + 0.017647363245487213, + 0.014683220535516739, + 0.013737271539866924, + 0.018409647047519684, + 0.021618137136101723, + 0.010323039256036282, + 0.011434362269937992, + 0.003725506365299225, + 0.002989106811583042, + 0.006400437094271183, + 0.005064597353339195, + 0.006242774892598391, + 0.005619111005216837, + 0.007375339977443218, + 0.009352879598736763, + 0.004485283978283405, + 0.004161280579864979, + 0.006657822988927364, + 0.00699173379689455, + 0.0050343615002930164, + 0.005285667255520821, + 0.008963675238192081, + 0.005731532350182533 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 6, + 7 + ], + "selected_reward": -0.009791087359189987 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a massive, ghostly white jellyfish?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the setting the abyssal darkness of the deep sea with inky blue water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the jellyfish's bell contract in ultra-slow motion with visible crystalline vortices displacing water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are countless trailing tentacles and oral arms visible drifting behind the jellyfish in an undulating motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera moving perfectly parallel to the jellyfish at an eye-level perspective?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0015126940561458468 + ], + "opposite_logprob_tokens": [ + -6.50151252746582 + ], + "logprob_reward": -0.0015126940561458468, + "yes_no_margin_reward": 6.5, + "token_credit_reward": -0.0015126940561458468, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07801201939582825, + 0.1333295702934265, + 0.039707574993371964, + 0.13163109123706818, + 0.20344141125679016, + 0.052744653075933456, + 0.12180086970329285, + 0.03691249340772629, + 0.07009200006723404, + 0.13232827186584473 + ], + "credit_map": [ + 0.009382106363773346, + 0.004593616351485252, + 0.023643914610147476, + 0.00502643920481205, + 0.010068078525364399, + 0.004195180721580982, + 0.007735903840512037, + 0.01336678210645914, + 0.011463209986686707, + 0.010083403438329697, + 0.014199720695614815, + 0.006392580922693014, + 0.025891868397593498, + 0.023012859746813774, + 0.018161123618483543, + 0.024124812334775925, + 0.005783236585557461, + 0.004030598793178797, + 0.004738322459161282, + 0.003260124009102583, + 0.0071334680542349815, + 0.005587202496826649, + 0.005342845804989338, + 0.0038317767903208733, + 0.016411270946264267, + 0.022371536120772362, + 0.015073085203766823, + 0.016879592090845108, + 0.013469257391989231, + 0.017245616763830185, + 0.015123523771762848, + 0.015057208016514778, + 0.014047804288566113, + 0.022348571568727493, + 0.0212493147701025, + 0.04105042293667793, + 0.018891777843236923, + 0.025853710249066353, + 0.029881004244089127, + 0.030118810012936592, + 0.005872133653610945, + 0.004868169315159321, + 0.007276125717908144, + 0.007018503732979298, + 0.004522873088717461, + 0.006214732304215431, + 0.01021582167595625, + 0.006756294053047895, + 0.007940125651657581, + 0.00791892223060131, + 0.018446611240506172, + 0.01354646123945713, + 0.03245914727449417, + 0.023941610008478165, + 0.01042341161519289, + 0.007124579511582851, + 0.0049958969466388226, + 0.00444596353918314, + 0.004956118296831846, + 0.005679112393409014, + 0.004097320139408112, + 0.005324308294802904, + 0.003549592336639762, + 0.0038641802966594696, + 0.007303171791136265, + 0.0057289074175059795, + 0.010108459740877151, + 0.012570487335324287, + 0.0065187797881662846, + 0.004670276772230864, + 0.011935576796531677, + 0.011256336234509945, + 0.013724271208047867, + 0.010566020384430885, + 0.033342983573675156, + 0.015260463580489159, + 0.011613021604716778, + 0.011139404028654099, + 0.0258144810795784, + 0.010867621749639511 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 1, + 9 + ], + "selected_reward": -0.0015126940561458468 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a massive, ghostly white jellyfish?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the setting the abyssal darkness of the deep sea with inky blue water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the jellyfish's bell contract in ultra-slow motion with visible crystalline vortices displacing water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are countless trailing tentacles and oral arms visible drifting behind the jellyfish in an undulating motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera moving perfectly parallel to the jellyfish at an eye-level perspective?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.05492597445845604 + ], + "opposite_logprob_tokens": [ + -2.9299259185791016 + ], + "logprob_reward": -0.05492597445845604, + "yes_no_margin_reward": 2.875, + "token_credit_reward": -0.05492597445845604, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07263350486755371, + 0.15885299444198608, + 0.09183302521705627, + 0.077341228723526, + 0.05003416910767555, + 0.08409270644187927, + 0.15538856387138367, + 0.14125774800777435, + 0.09269151836633682, + 0.07587455213069916 + ], + "credit_map": [ + 0.009080307558178902, + 0.008917170576751232, + 0.010973719879984856, + 0.007712428458034992, + 0.010829232633113861, + 0.004960412159562111, + 0.014120873063802719, + 0.006039352156221867, + 0.02656744420528412, + 0.011739854700863361, + 0.03912393003702164, + 0.011484937742352486, + 0.017367469146847725, + 0.01112920418381691, + 0.012933352962136269, + 0.028506793081760406, + 0.008337102830410004, + 0.010855862870812416, + 0.008868188597261906, + 0.010597790591418743, + 0.010959255509078503, + 0.015851840376853943, + 0.014705680310726166, + 0.011657298542559147, + 0.006468089297413826, + 0.01355954073369503, + 0.008169585838913918, + 0.012314234860241413, + 0.007653879467397928, + 0.010940645821392536, + 0.0097667770460248, + 0.00846846867352724, + 0.006641912739723921, + 0.005344665143638849, + 0.0062401858158409595, + 0.007843920029699802, + 0.004447670187801123, + 0.006813523359596729, + 0.0067152404226362705, + 0.005987048149108887, + 0.009427640587091446, + 0.008365970104932785, + 0.011598157696425915, + 0.010481612756848335, + 0.007342123426496983, + 0.011722630821168423, + 0.013593530282378197, + 0.011561025865375996, + 0.011910862289369106, + 0.006330435164272785, + 0.02482927031815052, + 0.016294915229082108, + 0.0411313958466053, + 0.030581852421164513, + 0.015649249777197838, + 0.008660568855702877, + 0.015361456200480461, + 0.013744623400270939, + 0.017597559839487076, + 0.01666358672082424, + 0.021036984398961067, + 0.02557671070098877, + 0.011513986624777317, + 0.019762830808758736, + 0.007154391147196293, + 0.005387128796428442, + 0.011783824302256107, + 0.013916853815317154, + 0.013177626766264439, + 0.009885339997708797, + 0.015732649713754654, + 0.015653694048523903, + 0.00580673897638917, + 0.004950417205691338, + 0.010799759067595005, + 0.008614842779934406, + 0.013744496740400791, + 0.011148474179208279, + 0.010495486669242382, + 0.010314324870705605 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 6, + 7 + ], + "selected_reward": -0.05492597445845604 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "An ultra-detailed, high-speed phantom camera-style photograph tracks smoothly with a massive, ghostly white jellyfish as it navigates the abyssal darkness of the deep sea. Maintaining an eye-level perspective with the creature, the camera moves perfectly parallel to its tranquil forward momentum. Its massive, translucent bell contracts in mesmerizing ultra-slow motion, a powerful yet graceful pulse rippling across its gelatinous surface and displacing water in visible crystalline vortices. The only light comes from the creature's own soft, internal bioluminescence, casting a milky white glow that illuminates the inky blue waters and reveals an intricate network of canals within its body. Following this beating heart, countless trailing tentacles, a mixture of delicate, ethereal strands and thicker, ribbon-like oral arms, drift behind in a complex, fluid ballet, hypnotically undulating and spiraling. The ultra-slow motion captures every tiny detail with crystal clarity – the subtle texture of the chimes, the delicate spiral dance of each tentacle, and the microscopic grains of marine snow drifting lazily by, creating a hypnotic and breathtaking visual spectacle of otherworldly beauty and profound silence.", + "seed": 6516, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000095/group_14/index_4472_seed_6516_dp_29_rank_58_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 14, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.02028389382176101, + "vlm_logprob": -0.02028389382176101, + "vlm_yes_no_margin": 4.825, + "vlm_token_credit": -0.02028389382176101, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a massive, ghostly white jellyfish?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the setting the abyssal darkness of the deep sea with inky blue water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the jellyfish's bell contract in ultra-slow motion with visible crystalline vortices displacing water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are countless trailing tentacles and oral arms visible drifting behind the jellyfish in an undulating motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera moving perfectly parallel to the jellyfish at an eye-level perspective?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005279645789414644 + ], + "opposite_logprob_tokens": [ + -5.505279541015625 + ], + "logprob_reward": -0.005279645789414644, + "yes_no_margin_reward": 5.5, + "token_credit_reward": -0.005279645789414644, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07686436176300049, + 0.08985713124275208, + 0.08326214551925659, + 0.1408986896276474, + 0.16482894122600555, + 0.2274816930294037, + 0.02951044961810112, + 0.07320773601531982, + 0.06472566723823547, + 0.04936324432492256 + ], + "credit_map": [ + 0.005236106459051371, + 0.008722390979528427, + 0.007994757033884525, + 0.009219233877956867, + 0.009696216322481632, + 0.015684694051742554, + 0.009751470759510994, + 0.010559484362602234, + 0.0073685189709067345, + 0.017597505822777748, + 0.01001970935612917, + 0.009900848381221294, + 0.014446038752794266, + 0.008078204467892647, + 0.012075337581336498, + 0.010370959527790546, + 0.006148562766611576, + 0.011058629490435123, + 0.012990389950573444, + 0.013012702576816082, + 0.009237018413841724, + 0.006510229781270027, + 0.012439150363206863, + 0.011865448206663132, + 0.007413029670715332, + 0.014524764381349087, + 0.025584954768419266, + 0.02998805232346058, + 0.012477426789700985, + 0.010695703327655792, + 0.025019090622663498, + 0.015195662155747414, + 0.017245030030608177, + 0.009308036416769028, + 0.038536109030246735, + 0.01590251922607422, + 0.010600064881145954, + 0.029230210930109024, + 0.020708506926894188, + 0.02329844981431961, + 0.002738912357017398, + 0.005449100863188505, + 0.019967807456851006, + 0.08198688924312592, + 0.006343943066895008, + 0.002704982180148363, + 0.09957535564899445, + 0.008714686147868633, + 0.0036675671581178904, + 0.00374835217371583, + 0.004587091505527496, + 0.004352291580289602, + 0.0034326938912272453, + 0.003514036303386092, + 0.0027703193482011557, + 0.003438093699514866, + 0.009917886927723885, + 0.009211630560457706, + 0.01050088182091713, + 0.01022602804005146, + 0.00816592201590538, + 0.007884681224822998, + 0.008321301080286503, + 0.00897939596325159, + 0.006804876960813999, + 0.005549887660890818, + 0.008004806004464626, + 0.006996769458055496, + 0.007377756293863058, + 0.007307542487978935, + 0.016791705042123795, + 0.005892311222851276, + 0.005623285658657551, + 0.005048887338489294, + 0.005635258741676807, + 0.006994838826358318, + 0.008284389972686768, + 0.006375215481966734, + 0.00770528893917799, + 0.0036960712168365717 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 4, + 3 + ], + "selected_reward": -0.005279645789414644 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a massive, ghostly white jellyfish?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the setting the abyssal darkness of the deep sea with inky blue water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the jellyfish's bell contract in ultra-slow motion with visible crystalline vortices displacing water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are countless trailing tentacles and oral arms visible drifting behind the jellyfish in an undulating motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera moving perfectly parallel to the jellyfish at an eye-level perspective?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0036220441106706858 + ], + "opposite_logprob_tokens": [ + -5.628622055053711 + ], + "logprob_reward": -0.0036220441106706858, + "yes_no_margin_reward": 5.625, + "token_credit_reward": -0.0036220441106706858, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09453080594539642, + 0.08226564526557922, + 0.09187720715999603, + 0.19393883645534515, + 0.10267766565084457, + 0.051036588847637177, + 0.1052764505147934, + 0.06451123207807541, + 0.08059840649366379, + 0.13328717648983002 + ], + "credit_map": [ + 0.0102731604129076, + 0.010717728175222874, + 0.011917099356651306, + 0.012052422389388084, + 0.015020580030977726, + 0.01015346497297287, + 0.015246446244418621, + 0.009149907156825066, + 0.0063974009826779366, + 0.01768045499920845, + 0.010865313932299614, + 0.00988965854048729, + 0.012164893560111523, + 0.005965097341686487, + 0.010038371197879314, + 0.009264450520277023, + 0.0076453592628240585, + 0.014177780598402023, + 0.012073115445673466, + 0.01570386253297329, + 0.011621546931564808, + 0.007866611704230309, + 0.0120750917121768, + 0.010713832452893257, + 0.012318101711571217, + 0.015532135032117367, + 0.04802878201007843, + 0.02927221916615963, + 0.017390646040439606, + 0.013655321672558784, + 0.031560759991407394, + 0.026180848479270935, + 0.010439656674861908, + 0.007727168500423431, + 0.020349448546767235, + 0.011339686810970306, + 0.007252065930515528, + 0.017280515283346176, + 0.011546510271728039, + 0.016742607578635216, + 0.0036511458456516266, + 0.004406776744872332, + 0.004315790254622698, + 0.01225884910672903, + 0.003766776528209448, + 0.003690431360155344, + 0.014822804369032383, + 0.004124009050428867, + 0.008264890871942043, + 0.012551981024444103, + 0.014797793701291084, + 0.022556122392416, + 0.01090563926845789, + 0.01300751231610775, + 0.010305678471922874, + 0.012886825948953629, + 0.01155002135783434, + 0.006858960725367069, + 0.008268031291663647, + 0.00997812207788229, + 0.005103467497974634, + 0.004750339780002832, + 0.00799631979316473, + 0.010005966760218143, + 0.016587024554610252, + 0.00897467415779829, + 0.010156639851629734, + 0.00600016675889492, + 0.006250347010791302, + 0.011267314665019512, + 0.013841863721609116, + 0.007520374376326799, + 0.03318515792489052, + 0.011350487358868122, + 0.018717151135206223, + 0.013701708987355232, + 0.010523847304284573, + 0.011275813914835453, + 0.01683455891907215, + 0.01769845187664032 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 9, + 6 + ], + "selected_reward": -0.0036220441106706858 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a massive, ghostly white jellyfish?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the setting the abyssal darkness of the deep sea with inky blue water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the jellyfish's bell contract in ultra-slow motion with visible crystalline vortices displacing water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are countless trailing tentacles and oral arms visible drifting behind the jellyfish in an undulating motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera moving perfectly parallel to the jellyfish at an eye-level perspective?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.011083937250077724 + ], + "opposite_logprob_tokens": [ + -4.511084079742432 + ], + "logprob_reward": -0.011083937250077724, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.011083937250077724, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11456739157438278, + 0.0715397447347641, + 0.15922218561172485, + 0.13270457088947296, + 0.0904579907655716, + 0.06815061718225479, + 0.03361058980226517, + 0.12828923761844635, + 0.084723562002182, + 0.11673417687416077 + ], + "credit_map": [ + 0.008371102623641491, + 0.01324673742055893, + 0.01073413249105215, + 0.016160832718014717, + 0.015216103754937649, + 0.01974979229271412, + 0.017225855961441994, + 0.013862831518054008, + 0.00423200661316514, + 0.0065401094034314156, + 0.012126905843615532, + 0.011012164875864983, + 0.006535897497087717, + 0.007445182651281357, + 0.014478159137070179, + 0.009169320575892925, + 0.009551577270030975, + 0.01587802916765213, + 0.035413987934589386, + 0.02021767944097519, + 0.013794844038784504, + 0.011023697443306446, + 0.03158165514469147, + 0.02176070585846901, + 0.008716587908565998, + 0.01732211746275425, + 0.017900846898555756, + 0.02388509176671505, + 0.01368861086666584, + 0.015124409459531307, + 0.021866006776690483, + 0.014200895093381405, + 0.011728165671229362, + 0.008761939592659473, + 0.01493328157812357, + 0.012848996557295322, + 0.0073294672183692455, + 0.008018549531698227, + 0.013977807946503162, + 0.012859778478741646, + 0.0024396951775997877, + 0.0029491845052689314, + 0.00534712104126811, + 0.022796323522925377, + 0.0034146280959248543, + 0.002222585491836071, + 0.02534165233373642, + 0.0036394232884049416, + 0.0047380318865180016, + 0.0037434149999171495, + 0.0061573744751513, + 0.0056654405780136585, + 0.0035734977573156357, + 0.0035375538282096386, + 0.003119287546724081, + 0.003075985936447978, + 0.013935352675616741, + 0.015607024542987347, + 0.0189967043697834, + 0.022062523290514946, + 0.009307792410254478, + 0.013810968957841396, + 0.017835604026913643, + 0.01673327013850212, + 0.011478241533041, + 0.005208865739405155, + 0.010219832882285118, + 0.011136113665997982, + 0.0057235099375247955, + 0.00729350233450532, + 0.02418467588722706, + 0.00947881955653429, + 0.0138253690674901, + 0.011141425929963589, + 0.012241408228874207, + 0.010669470764696598, + 0.01463229674845934, + 0.019524095579981804, + 0.016963526606559753, + 0.017736585810780525 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 3, + 7 + ], + "selected_reward": -0.011083937250077724 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a massive, ghostly white jellyfish?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the setting the abyssal darkness of the deep sea with inky blue water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the jellyfish's bell contract in ultra-slow motion with visible crystalline vortices displacing water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are countless trailing tentacles and oral arms visible drifting behind the jellyfish in an undulating motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera moving perfectly parallel to the jellyfish at an eye-level perspective?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0024880189448595047 + ], + "opposite_logprob_tokens": [ + -6.002488136291504 + ], + "logprob_reward": -0.0024880189448595047, + "yes_no_margin_reward": 6.0, + "token_credit_reward": -0.0024880189448595047, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06314214318990707, + 0.034800928086042404, + 0.06474260985851288, + 0.3714883029460907, + 0.043006040155887604, + 0.15437039732933044, + 0.008443137630820274, + 0.10870945453643799, + 0.1041589304804802, + 0.0471380241215229 + ], + "credit_map": [ + 0.004427505191415548, + 0.007374098524451256, + 0.00510813295841217, + 0.008592875674366951, + 0.009397276677191257, + 0.011863645166158676, + 0.007532633375376463, + 0.008845976553857327, + 0.002308826195076108, + 0.005384159740060568, + 0.004229885060340166, + 0.004747430328279734, + 0.00529737863689661, + 0.0035867481492459774, + 0.0055524930357933044, + 0.0036940101999789476, + 0.003346638521179557, + 0.007943100295960903, + 0.015326264314353466, + 0.009656202979385853, + 0.004988525994122028, + 0.0041498527862131596, + 0.01141269225627184, + 0.007919331081211567, + 0.020831746980547905, + 0.018768837675452232, + 0.1419016420841217, + 0.0483752004802227, + 0.01554764062166214, + 0.015867343172430992, + 0.04292016103863716, + 0.0672757476568222, + 0.0047587133012712, + 0.004946656059473753, + 0.007337454240769148, + 0.004650475457310677, + 0.0037404592148959637, + 0.005294478498399258, + 0.0069463602267205715, + 0.005331444088369608, + 0.001936615677550435, + 0.003917047753930092, + 0.013444127514958382, + 0.05520923435688019, + 0.003976784646511078, + 0.0025364996399730444, + 0.06708415597677231, + 0.006265917327255011, + 0.000816310232039541, + 0.000729123770724982, + 0.001271191518753767, + 0.0019404649501666427, + 0.0007541038212366402, + 0.0009242334635928273, + 0.000980027369223535, + 0.0010276827961206436, + 0.010713649913668633, + 0.014443077147006989, + 0.017694562673568726, + 0.0181883592158556, + 0.008063467219471931, + 0.011780275031924248, + 0.0146460747346282, + 0.013179999776184559, + 0.014398905448615551, + 0.008202948607504368, + 0.014187360182404518, + 0.01273956336081028, + 0.007547630928456783, + 0.0131430234760046, + 0.02055721916258335, + 0.013382287696003914, + 0.006129892077296972, + 0.0030583273619413376, + 0.005782902706414461, + 0.004486439749598503, + 0.0044715516269207, + 0.007092325948178768, + 0.009659412316977978, + 0.006457173265516758 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 5, + 7 + ], + "selected_reward": -0.0024880189448595047 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a massive, ghostly white jellyfish?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the setting the abyssal darkness of the deep sea with inky blue water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the jellyfish's bell contract in ultra-slow motion with visible crystalline vortices displacing water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are countless trailing tentacles and oral arms visible drifting behind the jellyfish in an undulating motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera moving perfectly parallel to the jellyfish at an eye-level perspective?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0789458230137825 + ], + "opposite_logprob_tokens": [ + -2.5789458751678467 + ], + "logprob_reward": -0.0789458230137825, + "yes_no_margin_reward": 2.5, + "token_credit_reward": -0.0789458230137825, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08579010516405106, + 0.07996384054422379, + 0.06278014183044434, + 0.43522754311561584, + 0.07363413274288177, + 0.04350518807768822, + 0.02990741655230522, + 0.06782599538564682, + 0.08684694021940231, + 0.03451869264245033 + ], + "credit_map": [ + 0.007190872449427843, + 0.009511122480034828, + 0.009796545840799809, + 0.012007376179099083, + 0.008848070167005062, + 0.01410285010933876, + 0.014941866509616375, + 0.009391405619680882, + 0.0060921646654605865, + 0.0122373066842556, + 0.011915522627532482, + 0.012264389544725418, + 0.011840161867439747, + 0.007069883402436972, + 0.010362786240875721, + 0.008181633427739143, + 0.0037327155005186796, + 0.0043777902610599995, + 0.020017283037304878, + 0.00856543704867363, + 0.0034622452221810818, + 0.0036749127320945263, + 0.008649314753711224, + 0.010300448164343834, + 0.024667393416166306, + 0.024152545258402824, + 0.1567479968070984, + 0.05454437807202339, + 0.024705760180950165, + 0.01853744313120842, + 0.051918886601924896, + 0.07995317131280899, + 0.008640429936349392, + 0.006639651022851467, + 0.013415888883173466, + 0.010221787728369236, + 0.0063374401070177555, + 0.007872465066611767, + 0.010518843308091164, + 0.009987633675336838, + 0.00317318900488317, + 0.003947922494262457, + 0.005435742903500795, + 0.008777657523751259, + 0.005426494404673576, + 0.0037895527202636003, + 0.007493144366890192, + 0.005461487453430891, + 0.0025178284849971533, + 0.004110874608159065, + 0.00390020664781332, + 0.005597099661827087, + 0.0050432439893484116, + 0.002659114543348551, + 0.0026660412549972534, + 0.0034130094572901726, + 0.01034367922693491, + 0.008252428844571114, + 0.01013348251581192, + 0.011639206670224667, + 0.0046192496083676815, + 0.006814287975430489, + 0.007882973179221153, + 0.00814069528132677, + 0.013358863070607185, + 0.007136589381843805, + 0.013270635157823563, + 0.010981268249452114, + 0.006473053712397814, + 0.00677600409835577, + 0.01787467487156391, + 0.010975859127938747, + 0.006103730294853449, + 0.002914999146014452, + 0.004742585122585297, + 0.0033934235107153654, + 0.003235885174944997, + 0.0038101838435977697, + 0.006813693325966597, + 0.0035041961818933487 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 8, + 0 + ], + "selected_reward": -0.0789458230137825 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000095/group_15/index_tensor([5442, 5442])_group_15_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000095/group_15/index_tensor([5442, 5442])_group_15_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..1d482573ecb1db430c7f79737c748b891cb53e65 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000095/group_15/index_tensor([5442, 5442])_group_15_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "An extreme close-up, static shot depicts a human eye, filling the screen with unsettling detail. The pupil is a vast, pitch-black void, dilated to its absolute limit, devouring the intricate webbed pattern of the pale, ice-blue iris. A sharp, cold-white point of light reflects off the glassy surface of the cornea, suggesting a harsh, unseen light source. The scrotum, far from pure white, depicts a delicate network of crimson, engorged veins. The skin of the upper eyelid, displaying a faint, damp sheen, is caught in an uncontrollable, agitated tremor; it intermittently flutters and twitches, rapid, involuntary micro-movements that betray a state of extreme duress. The eye itself remains fixed, staring directly forward with unwavering intensity, appearing both hollow and piercing. The harsh, direct light sculpts the subtle texture of the surrounding skin and casts deep shadows, creating a cool, high-contrast palette and an irresistible atmosphere of raw tension and psychological distress.", + "seed": 489, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000095/group_15/index_5442_seed_489_dp_30_rank_60_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 15, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.10482951095327735, + "vlm_logprob": -0.10482951095327735, + "vlm_yes_no_margin": 3.675, + "vlm_token_credit": -0.10482951095327735, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of a human eye that fills the screen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the pupil depicted as a vast, pitch-black void dilated to its absolute limit?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the iris pale and ice-blue with an intricate webbed pattern?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the upper eyelid skin show a faint, damp sheen and exhibit an agitated tremor with rapid micro-movements?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the eye fixed staring directly forward with unwavering intensity?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0036427113227546215 + ], + "opposite_logprob_tokens": [ + -7.128642559051514 + ], + "logprob_reward": -0.0036427113227546215, + "yes_no_margin_reward": 7.125, + "token_credit_reward": -0.0036427113227546215, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1321643739938736, + 0.03941858932375908, + 0.06208880990743637, + 0.056343212723731995, + 0.03647357225418091, + 0.35881906747817993, + 0.1394982635974884, + 0.02126622386276722, + 0.0988863930106163, + 0.05504154413938522 + ], + "credit_map": [ + 0.025111235678195953, + 0.00834599044173956, + 0.01874527707695961, + 0.009072779677808285, + 0.010158714838325977, + 0.015952203422784805, + 0.01060135941952467, + 0.034176815301179886, + 0.005847749765962362, + 0.003934105392545462, + 0.0037044433411210775, + 0.004534692503511906, + 0.004559461493045092, + 0.003283575177192688, + 0.004536821506917477, + 0.009017734788358212, + 0.006555566098541021, + 0.002869413699954748, + 0.015871958807110786, + 0.005642592906951904, + 0.0033947844058275223, + 0.007887644693255424, + 0.004653528332710266, + 0.01521331537514925, + 0.006341155152767897, + 0.003173806704580784, + 0.00856686569750309, + 0.00755712017416954, + 0.005523202940821648, + 0.006611240096390247, + 0.005468389485031366, + 0.013101429678499699, + 0.004536919295787811, + 0.0020606520120054483, + 0.006189906504005194, + 0.00534699996933341, + 0.0039015922229737043, + 0.003472947282716632, + 0.004804895259439945, + 0.006159655284136534, + 0.023497777059674263, + 0.007326201535761356, + 0.1200845018029213, + 0.01963721215724945, + 0.009718498215079308, + 0.0351896658539772, + 0.010096105746924877, + 0.13326908648014069, + 0.006798335816711187, + 0.003456149250268936, + 0.033824313431978226, + 0.005453235004097223, + 0.007688180543482304, + 0.05395393446087837, + 0.005545121151953936, + 0.022778991609811783, + 0.001921677845530212, + 0.0012823543511331081, + 0.002968477550894022, + 0.002017589285969734, + 0.0047963568940758705, + 0.0029806627426296473, + 0.002169906860217452, + 0.0031291956547647715, + 0.007420709822326899, + 0.0025742067955434322, + 0.025857217609882355, + 0.011327738873660564, + 0.005453112069517374, + 0.009209033101797104, + 0.005644362419843674, + 0.03140001371502876, + 0.00742938369512558, + 0.0027971460949629545, + 0.0078571867197752, + 0.010218233801424503, + 0.004362855106592178, + 0.006769981700927019, + 0.004519644193351269, + 0.01108711026608944 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 6, + 0 + ], + "selected_reward": -0.0036427113227546215 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of a human eye that fills the screen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the pupil depicted as a vast, pitch-black void dilated to its absolute limit?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the iris pale and ice-blue with an intricate webbed pattern?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the upper eyelid skin show a faint, damp sheen and exhibit an agitated tremor with rapid micro-movements?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the eye fixed staring directly forward with unwavering intensity?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.38706645369529724 + ], + "opposite_logprob_tokens": [ + -1.1370664834976196 + ], + "logprob_reward": -0.38706645369529724, + "yes_no_margin_reward": 0.75, + "token_credit_reward": -0.38706645369529724, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09782208502292633, + 0.04076502472162247, + 0.06377516686916351, + 0.037758130580186844, + 0.04779134690761566, + 0.35653451085090637, + 0.1380613148212433, + 0.13896971940994263, + 0.031324584037065506, + 0.04719811677932739 + ], + "credit_map": [ + 0.01882142387330532, + 0.005073282402008772, + 0.0235489159822464, + 0.005608665756881237, + 0.011778149753808975, + 0.013136218301951885, + 0.007746197283267975, + 0.012109236791729927, + 0.006746278144419193, + 0.0025327822659164667, + 0.010676666162908077, + 0.003385403426364064, + 0.003077807603403926, + 0.005509704817086458, + 0.0033889536280184984, + 0.005447429604828358, + 0.006472055334597826, + 0.0032789581455290318, + 0.009299425408244133, + 0.0049562579952180386, + 0.0038749128580093384, + 0.007258469704538584, + 0.004324664361774921, + 0.024310419335961342, + 0.0034455303102731705, + 0.003053561318665743, + 0.005966936703771353, + 0.003719041822478175, + 0.0051439907401800156, + 0.0037498415913432837, + 0.005098546855151653, + 0.0075806803070008755, + 0.0036871209740638733, + 0.0018800303805619478, + 0.010625744238495827, + 0.004043892025947571, + 0.00664268946275115, + 0.004330498166382313, + 0.003687110263854265, + 0.012894254177808762, + 0.020793573930859566, + 0.007512291427701712, + 0.11422504484653473, + 0.01859663985669613, + 0.020571738481521606, + 0.0317339226603508, + 0.014736098237335682, + 0.1283651739358902, + 0.006369660142809153, + 0.0031793457455933094, + 0.033034540712833405, + 0.004648824688047171, + 0.012685198336839676, + 0.05135694518685341, + 0.007582744583487511, + 0.019204044714570045, + 0.011090421117842197, + 0.0040197838097810745, + 0.04025379940867424, + 0.008342898450791836, + 0.01551925577223301, + 0.02030659094452858, + 0.0062004560604691505, + 0.03323649987578392, + 0.002584352856501937, + 0.0008289816905744374, + 0.009291546419262886, + 0.0023839366622269154, + 0.0018181715859100223, + 0.004414434544742107, + 0.0017538319807499647, + 0.00824932660907507, + 0.004433873109519482, + 0.0028481981717050076, + 0.009021955542266369, + 0.005542312748730183, + 0.006092768628150225, + 0.004291983786970377, + 0.004455048590898514, + 0.010511975735425949 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 7, + 6 + ], + "selected_reward": -0.38706645369529724 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of a human eye that fills the screen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the pupil depicted as a vast, pitch-black void dilated to its absolute limit?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the iris pale and ice-blue with an intricate webbed pattern?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the upper eyelid skin show a faint, damp sheen and exhibit an agitated tremor with rapid micro-movements?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the eye fixed staring directly forward with unwavering intensity?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01254853792488575 + ], + "opposite_logprob_tokens": [ + -4.387548446655273 + ], + "logprob_reward": -0.01254853792488575, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.01254853792488575, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.29148003458976746, + 0.11582314223051071, + 0.04861397296190262, + 0.05413614958524704, + 0.10407125949859619, + 0.1820252388715744, + 0.04108069837093353, + 0.08594765514135361, + 0.04009921848773956, + 0.03672254458069801 + ], + "credit_map": [ + 0.051907237619161606, + 0.013103321194648743, + 0.08786831796169281, + 0.0149075360968709, + 0.030905170366168022, + 0.03988040238618851, + 0.026675162836909294, + 0.026232918724417686, + 0.015438736416399479, + 0.0056385379284620285, + 0.02790050208568573, + 0.007891703397035599, + 0.01593044213950634, + 0.017442885786294937, + 0.01404022891074419, + 0.011540113016963005, + 0.00685676746070385, + 0.0030606479849666357, + 0.005683240480720997, + 0.003130607306957245, + 0.006399293430149555, + 0.005074057728052139, + 0.005182432942092419, + 0.013226929120719433, + 0.005525045096874237, + 0.0026931483298540115, + 0.007642351556569338, + 0.007525182794779539, + 0.006092171650379896, + 0.005715141072869301, + 0.005910612642765045, + 0.013032503426074982, + 0.006860360503196716, + 0.0036629335954785347, + 0.028411008417606354, + 0.005996102467179298, + 0.008527877740561962, + 0.00885368138551712, + 0.0064813909120857716, + 0.03527791053056717, + 0.010548028163611889, + 0.0038284054026007652, + 0.058164142072200775, + 0.01037926971912384, + 0.011721591465175152, + 0.016642125323414803, + 0.007217672653496265, + 0.0635240226984024, + 0.003129436867311597, + 0.0012042071903124452, + 0.008735700510442257, + 0.0019315739627927542, + 0.0034815471153706312, + 0.007128724828362465, + 0.0028075852897018194, + 0.012661928310990334, + 0.005822249688208103, + 0.0022779011633247137, + 0.024779153987765312, + 0.005360719747841358, + 0.009733714163303375, + 0.010334166698157787, + 0.00462216604501009, + 0.02301759086549282, + 0.0034704930149018764, + 0.0015984131023287773, + 0.00835068803280592, + 0.004983756225556135, + 0.0020066455472260714, + 0.00409120786935091, + 0.0027697859331965446, + 0.01282823272049427, + 0.0042278519831597805, + 0.002271089470013976, + 0.005002566147595644, + 0.0044985017739236355, + 0.005452233832329512, + 0.0055679683573544025, + 0.003722552442923188, + 0.005979785695672035 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 5, + 1 + ], + "selected_reward": -0.01254853792488575 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of a human eye that fills the screen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the pupil depicted as a vast, pitch-black void dilated to its absolute limit?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the iris pale and ice-blue with an intricate webbed pattern?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the upper eyelid skin show a faint, damp sheen and exhibit an agitated tremor with rapid micro-movements?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the eye fixed staring directly forward with unwavering intensity?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.10030999779701233 + ], + "opposite_logprob_tokens": [ + -2.3503100872039795 + ], + "logprob_reward": -0.10030999779701233, + "yes_no_margin_reward": 2.25, + "token_credit_reward": -0.10030999779701233, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2734726369380951, + 0.03800099715590477, + 0.08426336199045181, + 0.0803099125623703, + 0.04940439760684967, + 0.3210183084011078, + 0.06894811987876892, + 0.031330280005931854, + 0.018335962668061256, + 0.034916121512651443 + ], + "credit_map": [ + 0.049305979162454605, + 0.01394957397133112, + 0.0873408168554306, + 0.015801232308149338, + 0.020960230380296707, + 0.04419873282313347, + 0.015502668917179108, + 0.026413381099700928, + 0.0038646783214062452, + 0.002265162067487836, + 0.008627711795270443, + 0.0032435611356049776, + 0.004998726770281792, + 0.004713087808340788, + 0.004115551710128784, + 0.006172516383230686, + 0.012358914129436016, + 0.004468545783311129, + 0.00944398995488882, + 0.005656412802636623, + 0.005103468894958496, + 0.009336602874100208, + 0.006336827762424946, + 0.031558599323034286, + 0.009081264957785606, + 0.0038282740861177444, + 0.017452821135520935, + 0.007031826302409172, + 0.005137819331139326, + 0.008651770651340485, + 0.006142334546893835, + 0.02298380434513092, + 0.004566205199807882, + 0.002067180583253503, + 0.012397512793540955, + 0.005280720070004463, + 0.0026344414800405502, + 0.004211805760860443, + 0.0034560009371489286, + 0.014790530316531658, + 0.020343860611319542, + 0.00677642272785306, + 0.10735711455345154, + 0.0193592831492424, + 0.009414168074727058, + 0.031738344579935074, + 0.008927435614168644, + 0.11710168421268463, + 0.005662816110998392, + 0.0017969337059184909, + 0.010462646372616291, + 0.0029827847611159086, + 0.0033404442947357893, + 0.01686599850654602, + 0.00423752935603261, + 0.02359897457063198, + 0.004009983967989683, + 0.0015095897251740098, + 0.00586533360183239, + 0.002295190002769232, + 0.006440300960093737, + 0.0033620106987655163, + 0.003289557062089443, + 0.0045583127066493034, + 0.0016613203333690763, + 0.0008276649750769138, + 0.003843659069389105, + 0.0019284600857645273, + 0.0020529436878859997, + 0.002089991932734847, + 0.0015226949471980333, + 0.004409227054566145, + 0.004388534463942051, + 0.0014154501259326935, + 0.008073266595602036, + 0.00352148013189435, + 0.0020398974884301424, + 0.003236884716898203, + 0.0016925858799368143, + 0.010548023506999016 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 0, + 2 + ], + "selected_reward": -0.10030999779701233 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of a human eye that fills the screen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the pupil depicted as a vast, pitch-black void dilated to its absolute limit?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the iris pale and ice-blue with an intricate webbed pattern?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the upper eyelid skin show a faint, damp sheen and exhibit an agitated tremor with rapid micro-movements?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the eye fixed staring directly forward with unwavering intensity?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.020579854026436806 + ], + "opposite_logprob_tokens": [ + -3.8955798149108887 + ], + "logprob_reward": -0.020579854026436806, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.020579854026436806, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.277230829000473, + 0.06999379396438599, + 0.05870296433568001, + 0.044371169060468674, + 0.028377925977110863, + 0.16197684407234192, + 0.16672326624393463, + 0.0728103518486023, + 0.0879802331328392, + 0.031832605600357056 + ], + "credit_map": [ + 0.03147868439555168, + 0.019521309062838554, + 0.07091029733419418, + 0.024265095591545105, + 0.028049293905496597, + 0.03630787134170532, + 0.027533091604709625, + 0.03916520997881889, + 0.00998107623308897, + 0.0066840010695159435, + 0.005936856847256422, + 0.008013954386115074, + 0.007746277842670679, + 0.008412975817918777, + 0.009379573166370392, + 0.013839086517691612, + 0.008988148532807827, + 0.004093730356544256, + 0.005746298935264349, + 0.005289253778755665, + 0.005658479407429695, + 0.0053650313057005405, + 0.006407622247934341, + 0.017154397442936897, + 0.004660222679376602, + 0.00222569378092885, + 0.00952157098799944, + 0.004159692209213972, + 0.004115904215723276, + 0.0052021867595613, + 0.004339268431067467, + 0.010146630927920341, + 0.0026546157896518707, + 0.002356792101636529, + 0.004647824913263321, + 0.0031997968908399343, + 0.004188247956335545, + 0.0029111176263540983, + 0.003146508475765586, + 0.005273025017231703, + 0.010974536649882793, + 0.005382824223488569, + 0.04349609836935997, + 0.008900236338376999, + 0.019491827115416527, + 0.017481457442045212, + 0.012321528978645802, + 0.043928321450948715, + 0.0067564938217401505, + 0.0031193529721349478, + 0.04993511736392975, + 0.005384413059800863, + 0.009541806764900684, + 0.06936348974704742, + 0.006055179983377457, + 0.016567397862672806, + 0.006015298888087273, + 0.003141094697639346, + 0.012946510687470436, + 0.0052248756401240826, + 0.011083915829658508, + 0.00986701250076294, + 0.006671425420790911, + 0.01786021701991558, + 0.007713729981333017, + 0.0023626924958080053, + 0.022964147850871086, + 0.007582650054246187, + 0.0072504132986068726, + 0.009072259068489075, + 0.004734759218990803, + 0.02629958465695381, + 0.0047125015407800674, + 0.0020890652667731047, + 0.0038866596296429634, + 0.003614149522036314, + 0.0049895294941961765, + 0.0031584561802446842, + 0.003317493014037609, + 0.006064748857170343 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 6, + 5 + ], + "selected_reward": -0.020579854026436806 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "An extreme close-up, static shot depicts a human eye, filling the screen with unsettling detail. The pupil is a vast, pitch-black void, dilated to its absolute limit, devouring the intricate webbed pattern of the pale, ice-blue iris. A sharp, cold-white point of light reflects off the glassy surface of the cornea, suggesting a harsh, unseen light source. The scrotum, far from pure white, depicts a delicate network of crimson, engorged veins. The skin of the upper eyelid, displaying a faint, damp sheen, is caught in an uncontrollable, agitated tremor; it intermittently flutters and twitches, rapid, involuntary micro-movements that betray a state of extreme duress. The eye itself remains fixed, staring directly forward with unwavering intensity, appearing both hollow and piercing. The harsh, direct light sculpts the subtle texture of the surrounding skin and casts deep shadows, creating a cool, high-contrast palette and an irresistible atmosphere of raw tension and psychological distress.", + "seed": 490, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000095/group_15/index_5442_seed_490_dp_30_rank_60_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 15, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.08641250999644398, + "vlm_logprob": -0.08641250999644398, + "vlm_yes_no_margin": 3.7, + "vlm_token_credit": -0.08641250999644398, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of a human eye that fills the screen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the pupil depicted as a vast, pitch-black void dilated to its absolute limit?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the iris pale and ice-blue with an intricate webbed pattern?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the upper eyelid skin show a faint, damp sheen and exhibit an agitated tremor with rapid micro-movements?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the eye fixed staring directly forward with unwavering intensity?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.006314919795840979 + ], + "opposite_logprob_tokens": [ + -6.881314754486084 + ], + "logprob_reward": -0.006314919795840979, + "yes_no_margin_reward": 6.875, + "token_credit_reward": -0.006314919795840979, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13226555287837982, + 0.05053020268678665, + 0.133402481675148, + 0.03361254557967186, + 0.026196245104074478, + 0.2902633547782898, + 0.05804784968495369, + 0.22398939728736877, + 0.02847030758857727, + 0.02322203479707241 + ], + "credit_map": [ + 0.011702684685587883, + 0.01204606145620346, + 0.03583773225545883, + 0.010769722051918507, + 0.0107130641117692, + 0.014627316035330296, + 0.012794731184840202, + 0.023774243891239166, + 0.004318421706557274, + 0.004735591355711222, + 0.012668844312429428, + 0.004146381746977568, + 0.005904372315853834, + 0.005698905326426029, + 0.004361504688858986, + 0.008696184493601322, + 0.017918240278959274, + 0.039486113935709, + 0.00805725622922182, + 0.014385014772415161, + 0.010710068047046661, + 0.007673556450754404, + 0.010265563614666462, + 0.02490665204823017, + 0.004746237304061651, + 0.004604424349963665, + 0.0038131435867398977, + 0.0041804322972893715, + 0.00416757445782423, + 0.003090818412601948, + 0.003462636610493064, + 0.0055472771637141705, + 0.0034712222404778004, + 0.003438245505094528, + 0.004014473874121904, + 0.0029043874237686396, + 0.004822797141969204, + 0.0023383076768368483, + 0.003515661461278796, + 0.0016911483835428953, + 0.043192096054553986, + 0.0905165821313858, + 0.012097532860934734, + 0.026905672624707222, + 0.021385937929153442, + 0.019228411838412285, + 0.01693938672542572, + 0.05999773368239403, + 0.006794771645218134, + 0.01365021150559187, + 0.0037084787618368864, + 0.00430048443377018, + 0.004850093275308609, + 0.0070786625146865845, + 0.0038900598883628845, + 0.013775086030364037, + 0.024789830669760704, + 0.05592608451843262, + 0.01151714101433754, + 0.013816912658512592, + 0.01517467387020588, + 0.027280565351247787, + 0.015151193365454674, + 0.060333002358675, + 0.0034625716507434845, + 0.00648887362331152, + 0.0024798482190817595, + 0.0030253115110099316, + 0.003842732636258006, + 0.0026500162202864885, + 0.0020919176749885082, + 0.004429035820066929, + 0.002303696470335126, + 0.0027766262646764517, + 0.0024705170653760433, + 0.004247256089001894, + 0.002322417451068759, + 0.003953201230615377, + 0.0025495262816548347, + 0.002598793711513281 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 7, + 2 + ], + "selected_reward": -0.006314919795840979 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of a human eye that fills the screen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the pupil depicted as a vast, pitch-black void dilated to its absolute limit?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the iris pale and ice-blue with an intricate webbed pattern?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the upper eyelid skin show a faint, damp sheen and exhibit an agitated tremor with rapid micro-movements?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the eye fixed staring directly forward with unwavering intensity?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.3134738504886627 + ], + "opposite_logprob_tokens": [ + -1.3134738206863403 + ], + "logprob_reward": -0.3134738504886627, + "yes_no_margin_reward": 1.0, + "token_credit_reward": -0.3134738504886627, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10989105701446533, + 0.1321963667869568, + 0.12322744727134705, + 0.034517619758844376, + 0.019418124109506607, + 0.06152242049574852, + 0.08123184740543365, + 0.396097332239151, + 0.013009052723646164, + 0.028888754546642303 + ], + "credit_map": [ + 0.013429808430373669, + 0.01249171607196331, + 0.023476485162973404, + 0.011286708526313305, + 0.009793691337108612, + 0.011463502421975136, + 0.009810138493776321, + 0.018139002844691277, + 0.010898374021053314, + 0.010843059048056602, + 0.037528373301029205, + 0.009859573096036911, + 0.009588674642145634, + 0.0095424959436059, + 0.011768913827836514, + 0.03216690570116043, + 0.01966850832104683, + 0.037789735943078995, + 0.007104502525180578, + 0.011089389212429523, + 0.009203584864735603, + 0.0074990736320614815, + 0.008112763054668903, + 0.02275989018380642, + 0.0038624736480414867, + 0.004573906306177378, + 0.004110776819288731, + 0.004570887424051762, + 0.004692177288234234, + 0.0043890937231481075, + 0.004157917574048042, + 0.0041603888384997845, + 0.002138374838978052, + 0.004206537734717131, + 0.0023471678141504526, + 0.002061185659840703, + 0.0031975030433386564, + 0.0013171345926821232, + 0.0020231399685144424, + 0.002127080224454403, + 0.00876154750585556, + 0.01839035004377365, + 0.0028993149753659964, + 0.005609596148133278, + 0.004460377153009176, + 0.004672989249229431, + 0.004111638758331537, + 0.012616603635251522, + 0.008716410957276821, + 0.019818272441625595, + 0.00399353913962841, + 0.009154928848147392, + 0.007421036250889301, + 0.007992060855031013, + 0.00632550148293376, + 0.017810093238949776, + 0.04471122846007347, + 0.10038652271032333, + 0.0203092098236084, + 0.025210311636328697, + 0.027336331084370613, + 0.0462973415851593, + 0.026569392532110214, + 0.10527696460485458, + 0.0009895509574562311, + 0.0021992071997374296, + 0.001278419396840036, + 0.0031828894279897213, + 0.0014691432006657124, + 0.001213763258419931, + 0.0015161166666075587, + 0.001159963314421475, + 0.0028461231850087643, + 0.0041366685181856155, + 0.0024165755603462458, + 0.005354834254831076, + 0.0029673429671674967, + 0.005281136836856604, + 0.002608836628496647, + 0.003277235897257924 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 1, + 2 + ], + "selected_reward": -0.3134738504886627 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of a human eye that fills the screen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the pupil depicted as a vast, pitch-black void dilated to its absolute limit?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the iris pale and ice-blue with an intricate webbed pattern?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the upper eyelid skin show a faint, damp sheen and exhibit an agitated tremor with rapid micro-movements?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the eye fixed staring directly forward with unwavering intensity?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.008653049357235432 + ], + "opposite_logprob_tokens": [ + -4.758653163909912 + ], + "logprob_reward": -0.008653049357235432, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.008653049357235432, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.27481508255004883, + 0.09216974675655365, + 0.07783018052577972, + 0.08405973762273788, + 0.06288127601146698, + 0.06588194519281387, + 0.08126484602689743, + 0.14460329711437225, + 0.08142959326505661, + 0.03506428375840187 + ], + "credit_map": [ + 0.03189102187752724, + 0.031786151230335236, + 0.055035434663295746, + 0.026617540046572685, + 0.03083023987710476, + 0.03305456042289734, + 0.033666569739580154, + 0.03193357214331627, + 0.008891837671399117, + 0.010567380115389824, + 0.015765678137540817, + 0.010571393184363842, + 0.011294371448457241, + 0.012661593034863472, + 0.012741999700665474, + 0.009675491601228714, + 0.00578316068276763, + 0.00924640242010355, + 0.012415711767971516, + 0.013697807677090168, + 0.0069713471457362175, + 0.006813488435000181, + 0.012302646413445473, + 0.01059961412101984, + 0.0071821087040007114, + 0.01726461760699749, + 0.009277205914258957, + 0.009013311937451363, + 0.012718932703137398, + 0.006124210078269243, + 0.006006935145705938, + 0.01647242344915867, + 0.0051934863440692425, + 0.007658584974706173, + 0.011815481819212437, + 0.008343673311173916, + 0.009270614944398403, + 0.004907937720417976, + 0.006694801617413759, + 0.008996698074042797, + 0.00879763439297676, + 0.018530594184994698, + 0.0032238257117569447, + 0.006339425686746836, + 0.005451444070786238, + 0.005427684634923935, + 0.004772634711116552, + 0.013338704593479633, + 0.008865357376635075, + 0.016290877014398575, + 0.005977765191346407, + 0.009169459342956543, + 0.013056947849690914, + 0.007989608682692051, + 0.00902415718883276, + 0.010890670120716095, + 0.016021106392145157, + 0.03650342673063278, + 0.007156075444072485, + 0.009918485768139362, + 0.011936542578041553, + 0.01648346334695816, + 0.01004369929432869, + 0.03654049336910248, + 0.011347994208335876, + 0.02208634652197361, + 0.004576194565743208, + 0.007017459720373154, + 0.009506480768322945, + 0.006386890541762114, + 0.005100451409816742, + 0.015407776460051537, + 0.0031182183884084225, + 0.004566146992146969, + 0.004553207661956549, + 0.005048755090683699, + 0.006328212562948465, + 0.003478477243334055, + 0.004924300592392683, + 0.003046964993700385 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 7, + 1 + ], + "selected_reward": -0.008653049357235432 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of a human eye that fills the screen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the pupil depicted as a vast, pitch-black void dilated to its absolute limit?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the iris pale and ice-blue with an intricate webbed pattern?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the upper eyelid skin show a faint, damp sheen and exhibit an agitated tremor with rapid micro-movements?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the eye fixed staring directly forward with unwavering intensity?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.048660002648830414 + ], + "opposite_logprob_tokens": [ + -3.0486600399017334 + ], + "logprob_reward": -0.048660002648830414, + "yes_no_margin_reward": 3.0, + "token_credit_reward": -0.048660002648830414, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1255873739719391, + 0.06500466167926788, + 0.10871841758489609, + 0.2603645920753479, + 0.09288520365953445, + 0.1047128438949585, + 0.07164930552244186, + 0.10750214010477066, + 0.04348622262477875, + 0.02008925937116146 + ], + "credit_map": [ + 0.014478398486971855, + 0.012492538429796696, + 0.03005256876349449, + 0.011475345119833946, + 0.01490066945552826, + 0.010124257765710354, + 0.013884361833333969, + 0.018179237842559814, + 0.004885032773017883, + 0.008380644023418427, + 0.006625846493989229, + 0.007376209832727909, + 0.012685263529419899, + 0.007997211068868637, + 0.0077461968176066875, + 0.00930824689567089, + 0.014530419372022152, + 0.03219667077064514, + 0.008043544366955757, + 0.011335611343383789, + 0.007896999828517437, + 0.0068215616047382355, + 0.009723793715238571, + 0.01816980540752411, + 0.025388598442077637, + 0.06305032968521118, + 0.025592094287276268, + 0.027915384620428085, + 0.022112535312771797, + 0.019717739894986153, + 0.015491131693124771, + 0.06109677255153656, + 0.008695884607732296, + 0.021111376583576202, + 0.013323329389095306, + 0.010426110588014126, + 0.013830626383423805, + 0.007926549762487411, + 0.010384825989603996, + 0.0071864924393594265, + 0.014623802155256271, + 0.03045913577079773, + 0.005295184440910816, + 0.009415381588041782, + 0.00801072921603918, + 0.008157928474247456, + 0.007075107656419277, + 0.02167557179927826, + 0.00989056657999754, + 0.020354466512799263, + 0.0038875897880643606, + 0.005868126638233662, + 0.004474144894629717, + 0.0057326857931911945, + 0.005896897986531258, + 0.015544818714261055, + 0.012162123806774616, + 0.027644211426377296, + 0.005474245175719261, + 0.007215503137558699, + 0.00748538039624691, + 0.012137752957642078, + 0.00726200919598341, + 0.028120914474129677, + 0.003927502781152725, + 0.007704024203121662, + 0.004811638500541449, + 0.01105960551649332, + 0.005071874242275953, + 0.0027653519064188004, + 0.005050944164395332, + 0.0030952796805649996, + 0.0018808814929798245, + 0.002584659494459629, + 0.0033441344276070595, + 0.003608024213463068, + 0.0016598086804151535, + 0.002308367984369397, + 0.0022637376096099615, + 0.002439644420519471 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 0, + 2 + ], + "selected_reward": -0.048660002648830414 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of a human eye that fills the screen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the pupil depicted as a vast, pitch-black void dilated to its absolute limit?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the iris pale and ice-blue with an intricate webbed pattern?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the upper eyelid skin show a faint, damp sheen and exhibit an agitated tremor with rapid micro-movements?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the eye fixed staring directly forward with unwavering intensity?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.05496072769165039 + ], + "opposite_logprob_tokens": [ + -2.9299607276916504 + ], + "logprob_reward": -0.05496072769165039, + "yes_no_margin_reward": 2.875, + "token_credit_reward": -0.05496072769165039, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.16073517501354218, + 0.07439951598644257, + 0.0686054453253746, + 0.2717619240283966, + 0.07361719757318497, + 0.020365437492728233, + 0.05396473407745361, + 0.09291326254606247, + 0.1636434644460678, + 0.019993845373392105 + ], + "credit_map": [ + 0.015069437213242054, + 0.013762278482317924, + 0.04430258274078369, + 0.013127891346812248, + 0.011118238791823387, + 0.01434818934649229, + 0.014424123801290989, + 0.03458244353532791, + 0.006006385199725628, + 0.007045440375804901, + 0.01656332053244114, + 0.0057261292822659016, + 0.006985635031014681, + 0.00778501620516181, + 0.007851873524487019, + 0.016435712575912476, + 0.008937702514231205, + 0.02085944451391697, + 0.004403247032314539, + 0.006159428972750902, + 0.006708874367177486, + 0.0051919445395469666, + 0.006339835934340954, + 0.010004963725805283, + 0.030318690463900566, + 0.057795796543359756, + 0.02416849322617054, + 0.028783340007066727, + 0.028764771297574043, + 0.023109955713152885, + 0.02320619486272335, + 0.055614687502384186, + 0.00926218368113041, + 0.020914092659950256, + 0.005054024048149586, + 0.007042254786938429, + 0.005178044084459543, + 0.00539078563451767, + 0.0048121060244739056, + 0.015963705256581306, + 0.002270464086905122, + 0.0035967822186648846, + 0.002595246536657214, + 0.002036941470578313, + 0.004057397600263357, + 0.0022883524652570486, + 0.0016271349741145968, + 0.0018931204685941339, + 0.0073832240886986256, + 0.015810370445251465, + 0.0033669231925159693, + 0.004911401309072971, + 0.003810285357758403, + 0.003758392296731472, + 0.004199284128844738, + 0.010724854655563831, + 0.01011912152171135, + 0.023406723514199257, + 0.005072983913123608, + 0.006027244962751865, + 0.006757131312042475, + 0.010759057477116585, + 0.006236866116523743, + 0.024534139782190323, + 0.02340407483279705, + 0.04655611887574196, + 0.010717807337641716, + 0.01630658470094204, + 0.015377810224890709, + 0.011521623469889164, + 0.011469115503132343, + 0.028290333226323128, + 0.0023044792469590902, + 0.002630764152854681, + 0.0018184182699769735, + 0.0025129758287221193, + 0.0020832472946494818, + 0.0040782964788377285, + 0.0019012391567230225, + 0.002664424479007721 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 8, + 0 + ], + "selected_reward": -0.05496072769165039 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "An extreme close-up, static shot depicts a human eye, filling the screen with unsettling detail. The pupil is a vast, pitch-black void, dilated to its absolute limit, devouring the intricate webbed pattern of the pale, ice-blue iris. A sharp, cold-white point of light reflects off the glassy surface of the cornea, suggesting a harsh, unseen light source. The scrotum, far from pure white, depicts a delicate network of crimson, engorged veins. The skin of the upper eyelid, displaying a faint, damp sheen, is caught in an uncontrollable, agitated tremor; it intermittently flutters and twitches, rapid, involuntary micro-movements that betray a state of extreme duress. The eye itself remains fixed, staring directly forward with unwavering intensity, appearing both hollow and piercing. The harsh, direct light sculpts the subtle texture of the surrounding skin and casts deep shadows, creating a cool, high-contrast palette and an irresistible atmosphere of raw tension and psychological distress.", + "seed": 491, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000095/group_15/index_5442_seed_491_dp_31_rank_62_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 15, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.04898848906159401, + "vlm_logprob": -0.04898848906159401, + "vlm_yes_no_margin": 4.1, + "vlm_token_credit": -0.04898848906159401, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of a human eye that fills the screen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the pupil depicted as a vast, pitch-black void dilated to its absolute limit?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the iris pale and ice-blue with an intricate webbed pattern?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the upper eyelid skin show a faint, damp sheen and exhibit an agitated tremor with rapid micro-movements?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the eye fixed staring directly forward with unwavering intensity?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004833679646253586 + ], + "opposite_logprob_tokens": [ + -7.254833698272705 + ], + "logprob_reward": -0.004833679646253586, + "yes_no_margin_reward": 7.25, + "token_credit_reward": -0.004833679646253586, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1214756965637207, + 0.15815113484859467, + 0.035552915185689926, + 0.03573434054851532, + 0.04964499548077583, + 0.19997839629650116, + 0.13620512187480927, + 0.025532521307468414, + 0.17742277681827545, + 0.060302115976810455 + ], + "credit_map": [ + 0.01885075308382511, + 0.00864314567297697, + 0.012732090428471565, + 0.014010048471391201, + 0.011880747973918915, + 0.01844530738890171, + 0.015859607607126236, + 0.021053997799754143, + 0.032607030123472214, + 0.011869163252413273, + 0.015395916067063808, + 0.011067955754697323, + 0.013404949568212032, + 0.023797286674380302, + 0.01665141060948372, + 0.03335743024945259, + 0.004341959021985531, + 0.003943106159567833, + 0.005084234289824963, + 0.002935070311650634, + 0.003985189832746983, + 0.004407825879752636, + 0.003909698221832514, + 0.006945829838514328, + 0.004331671167165041, + 0.002799296984449029, + 0.004460398573428392, + 0.0043128118850290775, + 0.0042055225931108, + 0.0038188959006220102, + 0.0042775277979671955, + 0.007528215181082487, + 0.005299135576933622, + 0.00414053350687027, + 0.004350021947175264, + 0.00871965941041708, + 0.006378678604960442, + 0.005554113537073135, + 0.007218599785119295, + 0.007984252646565437, + 0.022954285144805908, + 0.013849216513335705, + 0.02490793727338314, + 0.015831951051950455, + 0.02041693963110447, + 0.03032822720706463, + 0.024206917732954025, + 0.04748290777206421, + 0.02193167246878147, + 0.009341512806713581, + 0.016072122380137444, + 0.013115627691149712, + 0.013015804812312126, + 0.01898779347538948, + 0.017654381692409515, + 0.026086213067173958, + 0.002663571620360017, + 0.0020662543829530478, + 0.003006754443049431, + 0.002049544360488653, + 0.002350930357351899, + 0.003850682405754924, + 0.0031705962028354406, + 0.006374188233166933, + 0.016808949410915375, + 0.014127216301858425, + 0.03217767924070358, + 0.01924341917037964, + 0.015138749964535236, + 0.029044250026345253, + 0.026931457221508026, + 0.02395106852054596, + 0.004583058878779411, + 0.004413051065057516, + 0.009353653527796268, + 0.004307011142373085, + 0.00820912141352892, + 0.013567578978836536, + 0.0069816866889595985, + 0.008886953815817833 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 8, + 1 + ], + "selected_reward": -0.004833679646253586 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of a human eye that fills the screen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the pupil depicted as a vast, pitch-black void dilated to its absolute limit?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the iris pale and ice-blue with an intricate webbed pattern?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the upper eyelid skin show a faint, damp sheen and exhibit an agitated tremor with rapid micro-movements?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the eye fixed staring directly forward with unwavering intensity?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.12706495821475983 + ], + "opposite_logprob_tokens": [ + -2.1270649433135986 + ], + "logprob_reward": -0.12706495821475983, + "yes_no_margin_reward": 2.0, + "token_credit_reward": -0.12706495821475983, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15564395487308502, + 0.1331416666507721, + 0.04942052811384201, + 0.04442784562706947, + 0.05074680596590042, + 0.0377449132502079, + 0.05342644825577736, + 0.06945206224918365, + 0.3491673469543457, + 0.05682840570807457 + ], + "credit_map": [ + 0.019194653257727623, + 0.012990646995604038, + 0.012822928838431835, + 0.023341162130236626, + 0.016508713364601135, + 0.024920526891946793, + 0.02137334644794464, + 0.024491973221302032, + 0.022176282480359077, + 0.012094137258827686, + 0.014913489110767841, + 0.014829421415925026, + 0.012775097042322159, + 0.02418944053351879, + 0.016189606860280037, + 0.01597418636083603, + 0.007002950645983219, + 0.004143135622143745, + 0.005357988178730011, + 0.007625173777341843, + 0.005173371639102697, + 0.006853051949292421, + 0.00609672861173749, + 0.007168126292526722, + 0.004726682789623737, + 0.003039764706045389, + 0.005484068766236305, + 0.006209108047187328, + 0.005374435801059008, + 0.007779324892908335, + 0.005279598291963339, + 0.006534860469400883, + 0.005119283217936754, + 0.004904478322714567, + 0.0055964915081858635, + 0.006010001990944147, + 0.006312610115855932, + 0.0053949193097651005, + 0.008613047190010548, + 0.008795971050858498, + 0.003038605209439993, + 0.002870237221941352, + 0.005969631485641003, + 0.0033839335665106773, + 0.004261532332748175, + 0.00478066923096776, + 0.005838034674525261, + 0.007602270692586899, + 0.006283862050622702, + 0.004386052489280701, + 0.006884448230266571, + 0.0062044113874435425, + 0.005786676425486803, + 0.007004106882959604, + 0.006965951528400183, + 0.009910938329994678, + 0.007627655752003193, + 0.006047039292752743, + 0.006552067119628191, + 0.006445172242820263, + 0.008919211104512215, + 0.009678315371274948, + 0.009342062287032604, + 0.014840538613498211, + 0.03675244003534317, + 0.02693670615553856, + 0.04904730245471001, + 0.04002201929688454, + 0.02818191982805729, + 0.04639995098114014, + 0.06289459764957428, + 0.05893241614103317, + 0.004581022541970015, + 0.0029733639676123857, + 0.012261095456779003, + 0.005364381242543459, + 0.00441753538325429, + 0.01450107991695404, + 0.007285723928362131, + 0.0054442002438008785 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 0, + 1 + ], + "selected_reward": -0.12706495821475983 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of a human eye that fills the screen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the pupil depicted as a vast, pitch-black void dilated to its absolute limit?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the iris pale and ice-blue with an intricate webbed pattern?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the upper eyelid skin show a faint, damp sheen and exhibit an agitated tremor with rapid micro-movements?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the eye fixed staring directly forward with unwavering intensity?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.009795455262064934 + ], + "opposite_logprob_tokens": [ + -4.634795665740967 + ], + "logprob_reward": -0.009795455262064934, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.009795455262064934, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2548135221004486, + 0.1635512411594391, + 0.05307694897055626, + 0.09321696311235428, + 0.119051493704319, + 0.06955389678478241, + 0.034909941256046295, + 0.04949651286005974, + 0.05121331661939621, + 0.1111161857843399 + ], + "credit_map": [ + 0.05161416158080101, + 0.01868462562561035, + 0.029309751465916634, + 0.028627002611756325, + 0.01627628691494465, + 0.034306205809116364, + 0.03315917029976845, + 0.04283631592988968, + 0.03773977980017662, + 0.01092631183564663, + 0.015584946610033512, + 0.00991154182702303, + 0.013234943151473999, + 0.026918239891529083, + 0.018463972955942154, + 0.030771488323807716, + 0.009333542548120022, + 0.004077686462551355, + 0.0070951818488538265, + 0.00393487187102437, + 0.00457622017711401, + 0.007357738446444273, + 0.005488152615725994, + 0.011213548481464386, + 0.01027478463947773, + 0.006358148530125618, + 0.011886774562299252, + 0.01521286927163601, + 0.01100121345371008, + 0.010387428104877472, + 0.012183108367025852, + 0.015912627801299095, + 0.013770048506557941, + 0.007893231697380543, + 0.012961680069565773, + 0.02570621483027935, + 0.010364963673055172, + 0.011080282740294933, + 0.01875559613108635, + 0.01851946860551834, + 0.006708524655550718, + 0.004329003393650055, + 0.00943389069288969, + 0.014844128862023354, + 0.005062697920948267, + 0.008697940967977047, + 0.009673877619206905, + 0.010803819634020329, + 0.004137853160500526, + 0.002667112974449992, + 0.003948446363210678, + 0.004012607503682375, + 0.004441157914698124, + 0.004332057200372219, + 0.004376223310828209, + 0.0069944788701832294, + 0.006562161725014448, + 0.003281818702816963, + 0.005629081279039383, + 0.009569725953042507, + 0.006048231385648251, + 0.00619888212531805, + 0.005698562134057283, + 0.0065080453641712666, + 0.0068414839915931225, + 0.004281365312635899, + 0.005952712148427963, + 0.011468887329101562, + 0.004944699816405773, + 0.004608949646353722, + 0.007286774925887585, + 0.005828435532748699, + 0.011972090229392052, + 0.007712002843618393, + 0.00877461303025484, + 0.022590726613998413, + 0.010165643878281116, + 0.012433591298758984, + 0.021473929286003113, + 0.015993576496839523 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 4 + ], + "selected_reward": -0.009795455262064934 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of a human eye that fills the screen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the pupil depicted as a vast, pitch-black void dilated to its absolute limit?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the iris pale and ice-blue with an intricate webbed pattern?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the upper eyelid skin show a faint, damp sheen and exhibit an agitated tremor with rapid micro-movements?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the eye fixed staring directly forward with unwavering intensity?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.08904905617237091 + ], + "opposite_logprob_tokens": [ + -2.4640491008758545 + ], + "logprob_reward": -0.08904905617237091, + "yes_no_margin_reward": 2.375, + "token_credit_reward": -0.08904905617237091, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.16582249104976654, + 0.30196666717529297, + 0.055462006479501724, + 0.021934278309345245, + 0.02844812534749508, + 0.02281448245048523, + 0.05324184522032738, + 0.03941410407423973, + 0.29007524251937866, + 0.020820774137973785 + ], + "credit_map": [ + 0.03178834170103073, + 0.01523878239095211, + 0.017707427963614464, + 0.014926492236554623, + 0.014935662969946861, + 0.02713109739124775, + 0.01794011890888214, + 0.02615458145737648, + 0.07073565572500229, + 0.0228602085262537, + 0.028363870456814766, + 0.01691063493490219, + 0.02316666767001152, + 0.04630901291966438, + 0.027398165315389633, + 0.06622245907783508, + 0.008023388683795929, + 0.0056139882653951645, + 0.0066015031188726425, + 0.005617107730358839, + 0.008145864121615887, + 0.006936779711395502, + 0.00615676399320364, + 0.00836661271750927, + 0.0018997450824826956, + 0.0015852130018174648, + 0.0026560253463685513, + 0.0036860215477645397, + 0.002853656653314829, + 0.00272241304628551, + 0.002722969511523843, + 0.0038082352839410305, + 0.002882275264710188, + 0.0025250560138374567, + 0.0032790745608508587, + 0.005702989641577005, + 0.003679669927805662, + 0.0030616377480328083, + 0.0035509730223566294, + 0.003766449633985758, + 0.0020605954341590405, + 0.0014789216220378876, + 0.002215207787230611, + 0.0029838494956493378, + 0.0021378877572715282, + 0.0036176282446831465, + 0.003921429626643658, + 0.004398961551487446, + 0.0056015620939433575, + 0.0036668151151388884, + 0.006164578255265951, + 0.007793810218572617, + 0.00586449122056365, + 0.006320910528302193, + 0.005140373483300209, + 0.01268930546939373, + 0.00515980739146471, + 0.0024725291877985, + 0.004303986206650734, + 0.007703954819589853, + 0.0037273885682225227, + 0.005272075068205595, + 0.004728818777948618, + 0.006045540329068899, + 0.023395486176013947, + 0.01674031838774681, + 0.04857414960861206, + 0.05013984441757202, + 0.026207026094198227, + 0.05784710869193077, + 0.03847215697169304, + 0.028699150308966637, + 0.001400974695570767, + 0.0021091834641993046, + 0.0020274671260267496, + 0.0028525020461529493, + 0.002860240638256073, + 0.0036526520270854235, + 0.0031711466144770384, + 0.00274660624563694 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 8, + 0 + ], + "selected_reward": -0.08904905617237091 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of a human eye that fills the screen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the pupil depicted as a vast, pitch-black void dilated to its absolute limit?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the iris pale and ice-blue with an intricate webbed pattern?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the upper eyelid skin show a faint, damp sheen and exhibit an agitated tremor with rapid micro-movements?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the eye fixed staring directly forward with unwavering intensity?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01419929601252079 + ], + "opposite_logprob_tokens": [ + -4.264199256896973 + ], + "logprob_reward": -0.01419929601252079, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.01419929601252079, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2249501496553421, + 0.12163548171520233, + 0.07615863531827927, + 0.054212361574172974, + 0.03593875840306282, + 0.06799980252981186, + 0.06935879588127136, + 0.0743575319647789, + 0.24593789875507355, + 0.029450641945004463 + ], + "credit_map": [ + 0.03799399733543396, + 0.018190480768680573, + 0.019621077924966812, + 0.018091648817062378, + 0.02126712165772915, + 0.03652484342455864, + 0.024899035692214966, + 0.04836193099617958, + 0.020564723759889603, + 0.008103249594569206, + 0.011810860596597195, + 0.00716201588511467, + 0.010563341900706291, + 0.02063274383544922, + 0.012484084814786911, + 0.03031446598470211, + 0.011774662882089615, + 0.005845147650688887, + 0.007915959693491459, + 0.006171148736029863, + 0.006966613233089447, + 0.011181924492120743, + 0.007173560559749603, + 0.019129613414406776, + 0.00716808857396245, + 0.00349641521461308, + 0.005933014210313559, + 0.005963355768471956, + 0.007442712318152189, + 0.006809229496866465, + 0.006289750337600708, + 0.011109796352684498, + 0.004998135380446911, + 0.0031190274748951197, + 0.003842118661850691, + 0.003115658648312092, + 0.005221689585596323, + 0.004762326832860708, + 0.00341968541033566, + 0.007460114546120167, + 0.006494138389825821, + 0.005735147278755903, + 0.00754998903721571, + 0.006452514789998531, + 0.009159090928733349, + 0.009424686431884766, + 0.00713458564132452, + 0.016049647703766823, + 0.007368891499936581, + 0.005808096379041672, + 0.00808835867792368, + 0.00701866764575243, + 0.007408798672258854, + 0.010609375312924385, + 0.009421098046004772, + 0.01363550778478384, + 0.007693388964980841, + 0.005880529060959816, + 0.007384483236819506, + 0.007969528436660767, + 0.008786534890532494, + 0.009929228574037552, + 0.00962857250124216, + 0.017085261642932892, + 0.024681348353624344, + 0.018824830651283264, + 0.026331761851906776, + 0.0385492742061615, + 0.023490682244300842, + 0.029170513153076172, + 0.042903851717710495, + 0.041985638439655304, + 0.003919146489351988, + 0.002284499816596508, + 0.0031651181634515524, + 0.002850225893780589, + 0.0036700840573757887, + 0.005056979134678841, + 0.0037132310681045055, + 0.004791355691850185 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 0, + 1 + ], + "selected_reward": -0.01419929601252079 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "An extreme close-up, static shot depicts a human eye, filling the screen with unsettling detail. The pupil is a vast, pitch-black void, dilated to its absolute limit, devouring the intricate webbed pattern of the pale, ice-blue iris. A sharp, cold-white point of light reflects off the glassy surface of the cornea, suggesting a harsh, unseen light source. The scrotum, far from pure white, depicts a delicate network of crimson, engorged veins. The skin of the upper eyelid, displaying a faint, damp sheen, is caught in an uncontrollable, agitated tremor; it intermittently flutters and twitches, rapid, involuntary micro-movements that betray a state of extreme duress. The eye itself remains fixed, staring directly forward with unwavering intensity, appearing both hollow and piercing. The harsh, direct light sculpts the subtle texture of the surrounding skin and casts deep shadows, creating a cool, high-contrast palette and an irresistible atmosphere of raw tension and psychological distress.", + "seed": 492, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000095/group_15/index_5442_seed_492_dp_31_rank_62_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 15, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.3268126188777387, + "vlm_logprob": -0.3268126188777387, + "vlm_yes_no_margin": 2.975, + "vlm_token_credit": -0.3268126188777387, + "vlm_yes_ratio": 0.8, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of a human eye that fills the screen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the pupil depicted as a vast, pitch-black void dilated to its absolute limit?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the iris pale and ice-blue with an intricate webbed pattern?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the upper eyelid skin show a faint, damp sheen and exhibit an agitated tremor with rapid micro-movements?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the eye fixed staring directly forward with unwavering intensity?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.006694271694868803 + ], + "opposite_logprob_tokens": [ + -7.256694316864014 + ], + "logprob_reward": -0.006694271694868803, + "yes_no_margin_reward": 7.25, + "token_credit_reward": -0.006694271694868803, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2632037401199341, + 0.09161422401666641, + 0.08038484305143356, + 0.10186061263084412, + 0.10221866518259048, + 0.12427009642124176, + 0.07497325539588928, + 0.045830197632312775, + 0.07344730943441391, + 0.042197056114673615 + ], + "credit_map": [ + 0.05884857475757599, + 0.01865457557141781, + 0.035824231803417206, + 0.0280679352581501, + 0.02338947355747223, + 0.03662232682108879, + 0.025352558121085167, + 0.03644406422972679, + 0.020980725064873695, + 0.006729335058480501, + 0.010165204294025898, + 0.009175083599984646, + 0.008318175561726093, + 0.01026594266295433, + 0.01150149293243885, + 0.014478272758424282, + 0.01102893240749836, + 0.007992506958544254, + 0.01347361970692873, + 0.010561628267168999, + 0.010584719479084015, + 0.009749257005751133, + 0.009494330734014511, + 0.007499853614717722, + 0.013387959450483322, + 0.0066355327144265175, + 0.00888556707650423, + 0.015628201887011528, + 0.015938468277454376, + 0.010158638469874859, + 0.018020905554294586, + 0.013205337338149548, + 0.0132595244795084, + 0.010733969509601593, + 0.014798175543546677, + 0.00936366431415081, + 0.01451635081321001, + 0.009575826115906239, + 0.013742232695221901, + 0.016228919848799706, + 0.01648746244609356, + 0.012315782718360424, + 0.016514852643013, + 0.015094400383532047, + 0.018938006833195686, + 0.011488511227071285, + 0.01633133739233017, + 0.017099756747484207, + 0.009602226316928864, + 0.00681949220597744, + 0.010068737901747227, + 0.009754583239555359, + 0.006516924127936363, + 0.010368756018579006, + 0.00980099942535162, + 0.012041530571877956, + 0.0055836644023656845, + 0.004064497537910938, + 0.010683693923056126, + 0.005507851019501686, + 0.006149166263639927, + 0.003823828184977174, + 0.005511481314897537, + 0.004506017081439495, + 0.012123791500926018, + 0.004891790449619293, + 0.016984205693006516, + 0.008614535443484783, + 0.0065453085117042065, + 0.00785914994776249, + 0.00827267300337553, + 0.00815584696829319, + 0.007800829131156206, + 0.0036582753527909517, + 0.00589063810184598, + 0.00458156131207943, + 0.004414641298353672, + 0.004724449478089809, + 0.00494961766526103, + 0.006177046336233616 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 5, + 4 + ], + "selected_reward": -0.006694271694868803 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of a human eye that fills the screen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the pupil depicted as a vast, pitch-black void dilated to its absolute limit?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the iris pale and ice-blue with an intricate webbed pattern?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the upper eyelid skin show a faint, damp sheen and exhibit an agitated tremor with rapid micro-movements?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the eye fixed staring directly forward with unwavering intensity?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -1.4063396453857422 + ], + "opposite_logprob_tokens": [ + -0.2813396155834198 + ], + "logprob_reward": -1.4063396453857422, + "yes_no_margin_reward": -1.125, + "token_credit_reward": -1.4063396453857422, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.15595833957195282, + 0.18612965941429138, + 0.13156220316886902, + 0.07050741463899612, + 0.12368223071098328, + 0.10785216093063354, + 0.058935754001140594, + 0.04964646324515343, + 0.0639510303735733, + 0.05177473649382591 + ], + "credit_map": [ + 0.02365727908909321, + 0.01205002423375845, + 0.021235210821032524, + 0.018974915146827698, + 0.017776528373360634, + 0.019748009741306305, + 0.021676059812307358, + 0.020840324461460114, + 0.025159146636724472, + 0.014007884077727795, + 0.03349669277667999, + 0.023651763796806335, + 0.026505736634135246, + 0.01739988662302494, + 0.027407102286815643, + 0.018501436337828636, + 0.017800452187657356, + 0.011888443492352962, + 0.020549416542053223, + 0.016795407980680466, + 0.015735315158963203, + 0.012316388078033924, + 0.02000824734568596, + 0.016468530520796776, + 0.009202060289680958, + 0.004018515348434448, + 0.008119513280689716, + 0.012722200714051723, + 0.006326370406895876, + 0.004963293205946684, + 0.014134797267615795, + 0.011020664125680923, + 0.01544156577438116, + 0.012292837724089622, + 0.02184203453361988, + 0.013907048851251602, + 0.016407839953899384, + 0.012005934491753578, + 0.016586875542998314, + 0.015198081731796265, + 0.014187753200531006, + 0.011597405187785625, + 0.011878419667482376, + 0.013170929625630379, + 0.016756294295191765, + 0.010786352679133415, + 0.014800844714045525, + 0.014674164354801178, + 0.007131601683795452, + 0.005696065723896027, + 0.009078172035515308, + 0.005357956048101187, + 0.0070565566420555115, + 0.007031992543488741, + 0.007335794623941183, + 0.010247611440718174, + 0.006270428188145161, + 0.005411176010966301, + 0.009337115101516247, + 0.0068747419863939285, + 0.005429904907941818, + 0.005048830062150955, + 0.005774423014372587, + 0.005499841179698706, + 0.0070869880728423595, + 0.007586358115077019, + 0.0077753812074661255, + 0.011667393147945404, + 0.0059445565566420555, + 0.004621052648872137, + 0.011778014712035656, + 0.007491291034966707, + 0.00612209178507328, + 0.0052104820497334, + 0.007655927911400795, + 0.007343487348407507, + 0.006839320994913578, + 0.004554195795208216, + 0.008041313849389553, + 0.006007911171764135 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 2 + ], + "selected_reward": -1.4063396453857422 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of a human eye that fills the screen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the pupil depicted as a vast, pitch-black void dilated to its absolute limit?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the iris pale and ice-blue with an intricate webbed pattern?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the upper eyelid skin show a faint, damp sheen and exhibit an agitated tremor with rapid micro-movements?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the eye fixed staring directly forward with unwavering intensity?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.016102269291877747 + ], + "opposite_logprob_tokens": [ + -4.141102313995361 + ], + "logprob_reward": -0.016102269291877747, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.016102269291877747, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.16158145666122437, + 0.10201313346624374, + 0.11454231292009354, + 0.17504483461380005, + 0.09345298260450363, + 0.13296088576316833, + 0.06796290725469589, + 0.046589694917201996, + 0.046379078179597855, + 0.05947272852063179 + ], + "credit_map": [ + 0.018936719745397568, + 0.018813643604516983, + 0.02171626128256321, + 0.01658204011619091, + 0.027588117867708206, + 0.02013223059475422, + 0.018097924068570137, + 0.019714513793587685, + 0.01355573907494545, + 0.010529295541346073, + 0.015011858195066452, + 0.012764321640133858, + 0.01792445033788681, + 0.009891452267765999, + 0.0123805096372962, + 0.009955507703125477, + 0.010366037487983704, + 0.008929967880249023, + 0.01166638731956482, + 0.023022089153528214, + 0.010769483633339405, + 0.008994417265057564, + 0.02750810608267784, + 0.013285820372402668, + 0.020305031910538673, + 0.01608744077384472, + 0.0421762578189373, + 0.021033549681305885, + 0.014903102070093155, + 0.017404407262802124, + 0.023664074018597603, + 0.019470974802970886, + 0.012289443053305149, + 0.010325470939278603, + 0.008551295846700668, + 0.009568910114467144, + 0.011883944272994995, + 0.010759078897535801, + 0.013099882751703262, + 0.016974953934550285, + 0.0164426751434803, + 0.013499893248081207, + 0.0160762220621109, + 0.0181009192019701, + 0.02327647991478443, + 0.011089076288044453, + 0.01688249409198761, + 0.01759311929345131, + 0.006926618982106447, + 0.00749402167275548, + 0.009410242550075054, + 0.006607759278267622, + 0.009442680515348911, + 0.008546232245862484, + 0.009181882254779339, + 0.010353472083806992, + 0.005302885547280312, + 0.004345796070992947, + 0.010091325268149376, + 0.006384752225130796, + 0.004548341501504183, + 0.004479401744902134, + 0.005722204688936472, + 0.005714986007660627, + 0.006257458124309778, + 0.005685738287866116, + 0.008075615391135216, + 0.006003024056553841, + 0.004144805483520031, + 0.00361566711217165, + 0.00772286020219326, + 0.004873908590525389, + 0.008257195353507996, + 0.006304608657956123, + 0.008079694584012032, + 0.006206115707755089, + 0.008946367539465427, + 0.005438445135951042, + 0.007451141718775034, + 0.00878915935754776 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 0, + 5 + ], + "selected_reward": -0.016102269291877747 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of a human eye that fills the screen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the pupil depicted as a vast, pitch-black void dilated to its absolute limit?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the iris pale and ice-blue with an intricate webbed pattern?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the upper eyelid skin show a faint, damp sheen and exhibit an agitated tremor with rapid micro-movements?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the eye fixed staring directly forward with unwavering intensity?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.14285677671432495 + ], + "opposite_logprob_tokens": [ + -2.0178568363189697 + ], + "logprob_reward": -0.14285677671432495, + "yes_no_margin_reward": 1.875, + "token_credit_reward": -0.14285677671432495, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12273307889699936, + 0.08421815186738968, + 0.14623412489891052, + 0.30484139919281006, + 0.11168109625577927, + 0.06402111053466797, + 0.04052860289812088, + 0.041043128818273544, + 0.0566130094230175, + 0.028086360543966293 + ], + "credit_map": [ + 0.017197290435433388, + 0.009583096951246262, + 0.015820756554603577, + 0.015332844108343124, + 0.012792537920176983, + 0.014423152431845665, + 0.01889605075120926, + 0.018687346950173378, + 0.01291534025222063, + 0.007063672412186861, + 0.012322429567575455, + 0.010011566802859306, + 0.012472185306251049, + 0.008776841685175896, + 0.010055787861347198, + 0.010600321926176548, + 0.021435251459479332, + 0.009965788573026657, + 0.019426915794610977, + 0.01855745166540146, + 0.017166096717119217, + 0.014185375533998013, + 0.02229345217347145, + 0.0232037752866745, + 0.04220874607563019, + 0.02244461514055729, + 0.0537649542093277, + 0.04061590135097504, + 0.027325715869665146, + 0.024932196363806725, + 0.05378074571490288, + 0.03976847603917122, + 0.013503512367606163, + 0.010111303068697453, + 0.009450654499232769, + 0.011488275602459908, + 0.008967521600425243, + 0.015222187153995037, + 0.015925666317343712, + 0.027011962607502937, + 0.007570204790681601, + 0.005206063389778137, + 0.009629102423787117, + 0.008102728053927422, + 0.009049884043633938, + 0.006950976327061653, + 0.008909179829061031, + 0.008602965623140335, + 0.005326706450432539, + 0.004434193018823862, + 0.007517943624407053, + 0.003912432584911585, + 0.004523023031651974, + 0.0047082444652915, + 0.004849947988986969, + 0.005256106611341238, + 0.004682376980781555, + 0.003915413282811642, + 0.0073699429631233215, + 0.005696197971701622, + 0.004946515429764986, + 0.002832225523889065, + 0.0064341905526816845, + 0.00516626238822937, + 0.008670901879668236, + 0.0063950675539672375, + 0.007475615479052067, + 0.007237919140607119, + 0.006685737054795027, + 0.005422507878392935, + 0.008049150928854942, + 0.006676105782389641, + 0.003751432988792658, + 0.003066721372306347, + 0.0031965849921107292, + 0.004981253296136856, + 0.0025704733561724424, + 0.002333216369152069, + 0.0035535984206944704, + 0.004633075091987848 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 2, + 0 + ], + "selected_reward": -0.14285677671432495 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of a human eye that fills the screen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the pupil depicted as a vast, pitch-black void dilated to its absolute limit?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the iris pale and ice-blue with an intricate webbed pattern?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the upper eyelid skin show a faint, damp sheen and exhibit an agitated tremor with rapid micro-movements?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the eye fixed staring directly forward with unwavering intensity?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.06207013130187988 + ], + "opposite_logprob_tokens": [ + -2.81207013130188 + ], + "logprob_reward": -0.06207013130187988, + "yes_no_margin_reward": 2.75, + "token_credit_reward": -0.06207013130187988, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.20961549878120422, + 0.11683530360460281, + 0.1188952848315239, + 0.09527605772018433, + 0.10495617985725403, + 0.153903067111969, + 0.056486401706933975, + 0.03926372900605202, + 0.05814787372946739, + 0.04662060737609863 + ], + "credit_map": [ + 0.02556687593460083, + 0.018284965306520462, + 0.02354479767382145, + 0.021133841946721077, + 0.030697830021381378, + 0.02515515126287937, + 0.02959807775914669, + 0.03563394770026207, + 0.016877584159374237, + 0.009316464886069298, + 0.023210294544696808, + 0.012366021052002907, + 0.015016848221421242, + 0.010767383500933647, + 0.01487188134342432, + 0.014408816583454609, + 0.014051958918571472, + 0.0092612961307168, + 0.015299146063625813, + 0.01724931411445141, + 0.016009537503123283, + 0.010374289937317371, + 0.023569099605083466, + 0.013080629520118237, + 0.012287887744605541, + 0.007393051404505968, + 0.02083730325102806, + 0.011711213737726212, + 0.010088074952363968, + 0.008760006166994572, + 0.011984755285084248, + 0.012213760986924171, + 0.013355834409594536, + 0.011602664366364479, + 0.010874263010919094, + 0.01576172187924385, + 0.012389836832880974, + 0.008320331573486328, + 0.01697206124663353, + 0.01567946746945381, + 0.02641352079808712, + 0.014435612596571445, + 0.024160323664546013, + 0.01428950484842062, + 0.018118541687726974, + 0.018991809338331223, + 0.018190089613199234, + 0.01930365338921547, + 0.006666183937340975, + 0.005834120325744152, + 0.010147880762815475, + 0.005748555529862642, + 0.007392336148768663, + 0.00692413654178381, + 0.007390634156763554, + 0.0063825491815805435, + 0.00447109853848815, + 0.0036825172137469053, + 0.005793822929263115, + 0.008233188651502132, + 0.0037407875061035156, + 0.003190668299794197, + 0.006241591181606054, + 0.0039100502617657185, + 0.006231543608009815, + 0.0054989224299788475, + 0.006952040828764439, + 0.006540011614561081, + 0.006751269567757845, + 0.004387723747640848, + 0.011609508655965328, + 0.01017685141414404, + 0.005176271311938763, + 0.004372702911496162, + 0.006062853615731001, + 0.008749970234930515, + 0.006122800055891275, + 0.00430616969242692, + 0.006684683728963137, + 0.005145154893398285 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 5, + 2 + ], + "selected_reward": -0.06207013130187988 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "No", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000095/group_2/index_tensor([8782, 8782])_group_2_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000095/group_2/index_tensor([8782, 8782])_group_2_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..6f19c9d2ba5ea9a6c61620376565f9c3550535b7 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000095/group_2/index_tensor([8782, 8782])_group_2_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "high-angle drone camera slowly zooms forward and tilts slightly downward, creating a fluid, ink-and-wash painting effect, its edges adorned with a natural, smudged effect. Illuminated by the cool moonlight, a derelict ancient temple nestled deep in a mountain valley emerges. Its crumbling eaves and moss-covered dark tiles appear as varying shades of ink against the night sky. A woman in white, the sole bright spot in this dynamic painting, clad in a flowing white silk dress. Her voluminous sleeves and cascading jet-black hair unfurl in the night breeze, like nimble brushstrokes gliding across rice paper. Silently, almost parallel to the ground, she drifts forward, her movements light and steady. Like a wisp of smoke, she crosses the overgrown courtyard, passes through the open, stained vermilion gate, and descends into the darkness of the temple's main hall. The moonlight outlines the scene with clear contours and large, thick shadows, forming a strong black and white contrast. All dynamic elements have a softness and fluidity like ink dissolving in water, creating an extremely quiet, mysterious and poignant oriental fantasy atmosphere.", + "seed": 7214, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000095/group_2/index_8782_seed_7214_dp_4_rank_8_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 2, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.1055773664265871, + "vlm_logprob": -0.1055773664265871, + "vlm_yes_no_margin": 2.875, + "vlm_token_credit": -0.1055773664265871, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a derelict ancient temple located in a mountain valley?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a woman wearing a flowing white silk dress visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the woman drift forward silently while crossing an overgrown courtyard and passing through a stained vermilion gate?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the video exhibit an ink-and-wash painting style with smudged edges and fluid, dissolving ink effects?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot a high-angle view that slowly zooms forward and tilts slightly downward?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.32182231545448303 + ], + "opposite_logprob_tokens": [ + -1.3218222856521606 + ], + "logprob_reward": -0.32182231545448303, + "yes_no_margin_reward": 1.0, + "token_credit_reward": -0.32182231545448303, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.24715274572372437, + 0.1446259468793869, + 0.07894644886255264, + 0.0930028185248375, + 0.04387354478240013, + 0.10817144066095352, + 0.0635235533118248, + 0.08211014419794083, + 0.0819394439458847, + 0.056653887033462524 + ], + "credit_map": [ + 0.023022722452878952, + 0.05356970056891441, + 0.026391535997390747, + 0.03380060940980911, + 0.030485650524497032, + 0.02447551302611828, + 0.029327478259801865, + 0.026079513132572174, + 0.013249550946056843, + 0.02264924719929695, + 0.011430367827415466, + 0.0196941290050745, + 0.019876619800925255, + 0.023352809250354767, + 0.020543839782476425, + 0.013829383999109268, + 0.007623085752129555, + 0.01165597140789032, + 0.01081248838454485, + 0.0073158107697963715, + 0.008623918518424034, + 0.0140910679474473, + 0.010967308655381203, + 0.007856792770326138, + 0.006493531633168459, + 0.007635890040546656, + 0.009779063053429127, + 0.016622690483927727, + 0.00839410163462162, + 0.015806037932634354, + 0.016480054706335068, + 0.011791451834142208, + 0.00561146205291152, + 0.00516437878832221, + 0.0056112296879291534, + 0.004419037140905857, + 0.007590347435325384, + 0.007925852201879025, + 0.0030499412678182125, + 0.004501297138631344, + 0.011493785306811333, + 0.010055171325802803, + 0.024725642055273056, + 0.013316906057298183, + 0.017269039526581764, + 0.01474686898291111, + 0.007443106733262539, + 0.009120920673012733, + 0.005931324325501919, + 0.007502357941120863, + 0.012358355335891247, + 0.014262762852013111, + 0.007098457310348749, + 0.005062977783381939, + 0.005665384698659182, + 0.00564193120226264, + 0.007562538143247366, + 0.008524838835000992, + 0.0099491560831666, + 0.01484985463321209, + 0.012385547161102295, + 0.009252697229385376, + 0.008600668981671333, + 0.010984838008880615, + 0.009475317783653736, + 0.01163958478718996, + 0.009288112632930279, + 0.011434393934905529, + 0.014058923348784447, + 0.007957647554576397, + 0.006946605164557695, + 0.011138860136270523, + 0.007468260824680328, + 0.004629942588508129, + 0.004796822089701891, + 0.006107569206506014, + 0.008843972347676754, + 0.009033339098095894, + 0.00759598845615983, + 0.008177992887794971 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 5 + ], + "selected_reward": -0.32182231545448303 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a derelict ancient temple located in a mountain valley?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a woman wearing a flowing white silk dress visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the woman drift forward silently while crossing an overgrown courtyard and passing through a stained vermilion gate?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the video exhibit an ink-and-wash painting style with smudged edges and fluid, dissolving ink effects?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot a high-angle view that slowly zooms forward and tilts slightly downward?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07898855954408646 + ], + "opposite_logprob_tokens": [ + -2.578988552093506 + ], + "logprob_reward": -0.07898855954408646, + "yes_no_margin_reward": 2.5, + "token_credit_reward": -0.07898855954408646, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1453108787536621, + 0.2320820689201355, + 0.053481001406908035, + 0.051030032336711884, + 0.07782414555549622, + 0.07495680451393127, + 0.19909489154815674, + 0.05437823757529259, + 0.026726627722382545, + 0.08511537313461304 + ], + "credit_map": [ + 0.011506558395922184, + 0.015888294205069542, + 0.014863857999444008, + 0.017390593886375427, + 0.026170488446950912, + 0.011982171796262264, + 0.03273206576704979, + 0.014776843599975109, + 0.0149199478328228, + 0.025772621855139732, + 0.05959334224462509, + 0.04112647473812103, + 0.02576172724366188, + 0.026027709245681763, + 0.026407916098833084, + 0.012472287751734257, + 0.00588134815916419, + 0.006957645062357187, + 0.007507104426622391, + 0.006356668658554554, + 0.005281840916723013, + 0.008113189600408077, + 0.008078417740762234, + 0.0053047859109938145, + 0.004710353910923004, + 0.006339794024825096, + 0.005960611626505852, + 0.007168369833379984, + 0.005604301113635302, + 0.01110673975199461, + 0.0053047663532197475, + 0.004835094790905714, + 0.009622704237699509, + 0.006710094399750233, + 0.012043888680636883, + 0.008584495633840561, + 0.011133883148431778, + 0.013556089252233505, + 0.007888556458055973, + 0.008284423500299454, + 0.00778057100251317, + 0.005688554607331753, + 0.012035969644784927, + 0.00913697574287653, + 0.008738734759390354, + 0.012094435282051563, + 0.008237797766923904, + 0.011243761517107487, + 0.01686844229698181, + 0.0237737987190485, + 0.03922607749700546, + 0.055591292679309845, + 0.015993565320968628, + 0.014971240423619747, + 0.0190413985401392, + 0.013629055581986904, + 0.006755494512617588, + 0.01068428810685873, + 0.005347266793251038, + 0.007713247090578079, + 0.008245410397648811, + 0.005619429051876068, + 0.004548175260424614, + 0.005464921705424786, + 0.003284828970208764, + 0.004686160013079643, + 0.002577183535322547, + 0.0029202389996498823, + 0.003941569477319717, + 0.002630729228258133, + 0.0038770076353102922, + 0.0028089075349271297, + 0.015374083071947098, + 0.0064032962545752525, + 0.012507052160799503, + 0.007379170972853899, + 0.016634346917271614, + 0.010427653789520264, + 0.007787014823406935, + 0.008602743968367577 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 6, + 0 + ], + "selected_reward": -0.07898855954408646 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a derelict ancient temple located in a mountain valley?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a woman wearing a flowing white silk dress visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the woman drift forward silently while crossing an overgrown courtyard and passing through a stained vermilion gate?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the video exhibit an ink-and-wash painting style with smudged edges and fluid, dissolving ink effects?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot a high-angle view that slowly zooms forward and tilts slightly downward?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.012556660920381546 + ], + "opposite_logprob_tokens": [ + -4.387556552886963 + ], + "logprob_reward": -0.012556660920381546, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.012556660920381546, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.17003558576107025, + 0.09469231963157654, + 0.09861096739768982, + 0.03888631612062454, + 0.07276001572608948, + 0.12916283309459686, + 0.13352236151695251, + 0.11758879572153091, + 0.08334742486476898, + 0.06139335781335831 + ], + "credit_map": [ + 0.018416771665215492, + 0.026789631694555283, + 0.018201632425189018, + 0.024798544123768806, + 0.01679731346666813, + 0.018589088693261147, + 0.024286314845085144, + 0.022156275808811188, + 0.008348512463271618, + 0.014447936788201332, + 0.006237671244889498, + 0.009709656238555908, + 0.012582925148308277, + 0.016721025109291077, + 0.017794359475374222, + 0.008850233629345894, + 0.009305478073656559, + 0.014005200937390327, + 0.009773526340723038, + 0.010266553610563278, + 0.01302543468773365, + 0.013411792926490307, + 0.016778577119112015, + 0.012044407427310944, + 0.004822885617613792, + 0.003934978507459164, + 0.003600803902372718, + 0.0035946224816143513, + 0.007700872141867876, + 0.007228184025734663, + 0.005232496652752161, + 0.0027714704629033804, + 0.008256927132606506, + 0.0069990153424441814, + 0.011805874295532703, + 0.008169382810592651, + 0.01013820432126522, + 0.011778678745031357, + 0.006891342345625162, + 0.008720594458281994, + 0.018014676868915558, + 0.012857583351433277, + 0.016536105424165726, + 0.014895397238433361, + 0.021787501871585846, + 0.02172597497701645, + 0.008753446862101555, + 0.014592143706977367, + 0.010921255685389042, + 0.01660219207406044, + 0.023607559502124786, + 0.03713936731219292, + 0.010943211615085602, + 0.010885639116168022, + 0.011765905655920506, + 0.011657224968075752, + 0.010358095169067383, + 0.013938593678176403, + 0.017576122656464577, + 0.020387860015034676, + 0.01538710668683052, + 0.013170100748538971, + 0.013057339936494827, + 0.01371357124298811, + 0.010806413367390633, + 0.012316307984292507, + 0.00880552176386118, + 0.01084250956773758, + 0.013128383085131645, + 0.00882401317358017, + 0.007504963781684637, + 0.01111931074410677, + 0.010912037454545498, + 0.004918893799185753, + 0.01046304777264595, + 0.005339754745364189, + 0.01096782460808754, + 0.0072480058297514915, + 0.0049512023106217384, + 0.006592592690140009 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 6, + 5 + ], + "selected_reward": -0.012556660920381546 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a derelict ancient temple located in a mountain valley?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a woman wearing a flowing white silk dress visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the woman drift forward silently while crossing an overgrown courtyard and passing through a stained vermilion gate?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the video exhibit an ink-and-wash painting style with smudged edges and fluid, dissolving ink effects?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot a high-angle view that slowly zooms forward and tilts slightly downward?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.014206817373633385 + ], + "opposite_logprob_tokens": [ + -4.264206886291504 + ], + "logprob_reward": -0.014206817373633385, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.014206817373633385, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.17437799274921417, + 0.056963108479976654, + 0.07097818702459335, + 0.08253876119852066, + 0.11091188341379166, + 0.07180984318256378, + 0.060610488057136536, + 0.1494453400373459, + 0.06922636181116104, + 0.15313802659511566 + ], + "credit_map": [ + 0.020958205685019493, + 0.03352455794811249, + 0.015169484540820122, + 0.021904004737734795, + 0.024225125089287758, + 0.01953822188079357, + 0.02088022790849209, + 0.01817817986011505, + 0.00623844750225544, + 0.009760669432580471, + 0.004099537618458271, + 0.009166818112134933, + 0.006239649839699268, + 0.009595136158168316, + 0.007472031284123659, + 0.0043908171355724335, + 0.00885485764592886, + 0.012966277077794075, + 0.005761131178587675, + 0.00743140559643507, + 0.00874065887182951, + 0.007597016170620918, + 0.012764941900968552, + 0.006861899513751268, + 0.006435786373913288, + 0.009190838783979416, + 0.008322696201503277, + 0.01148068904876709, + 0.008155785501003265, + 0.01563076674938202, + 0.012451698072254658, + 0.010870502330362797, + 0.013077711686491966, + 0.014732283540070057, + 0.013565308414399624, + 0.01558218989521265, + 0.013014188036322594, + 0.019507965072989464, + 0.010196954011917114, + 0.011235283687710762, + 0.009807285852730274, + 0.006193557754158974, + 0.011907889507710934, + 0.00947378296405077, + 0.010627498850226402, + 0.010541241616010666, + 0.005977415945380926, + 0.00728117348626256, + 0.006955614313483238, + 0.006810045335441828, + 0.008980028331279755, + 0.007445037364959717, + 0.008919057436287403, + 0.007567658554762602, + 0.005854662042111158, + 0.008078386075794697, + 0.012682028114795685, + 0.017699016258120537, + 0.013476108200848103, + 0.02114345319569111, + 0.021260689944028854, + 0.021711943671107292, + 0.018108803778886795, + 0.023363299667835236, + 0.008329051546752453, + 0.012751673348248005, + 0.007732097990810871, + 0.009639499709010124, + 0.011302096769213676, + 0.006111486814916134, + 0.0065438891761004925, + 0.006816565059125423, + 0.027345089241862297, + 0.015629304572939873, + 0.02293320931494236, + 0.018063921481370926, + 0.027011649683117867, + 0.01151095051318407, + 0.014099868945777416, + 0.016544034704566002 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 7 + ], + "selected_reward": -0.014206817373633385 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a derelict ancient temple located in a mountain valley?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a woman wearing a flowing white silk dress visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the woman drift forward silently while crossing an overgrown courtyard and passing through a stained vermilion gate?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the video exhibit an ink-and-wash painting style with smudged edges and fluid, dissolving ink effects?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot a high-angle view that slowly zooms forward and tilts slightly downward?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.1003124788403511 + ], + "opposite_logprob_tokens": [ + -2.3503124713897705 + ], + "logprob_reward": -0.1003124788403511, + "yes_no_margin_reward": 2.25, + "token_credit_reward": -0.1003124788403511, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1505778431892395, + 0.0828663781285286, + 0.08835015445947647, + 0.08749765902757645, + 0.07243598997592926, + 0.07980259507894516, + 0.2970183789730072, + 0.05338495597243309, + 0.0356430746614933, + 0.052422937005758286 + ], + "credit_map": [ + 0.014782112091779709, + 0.021690549328923225, + 0.023877540603280067, + 0.017786279320716858, + 0.014984543435275555, + 0.013407736085355282, + 0.020437277853488922, + 0.02361181750893593, + 0.007589843589812517, + 0.012633315287530422, + 0.006810074672102928, + 0.008362563326954842, + 0.014087694697082043, + 0.012471780180931091, + 0.013727123849093914, + 0.007183977402746677, + 0.009006629697978497, + 0.014693396165966988, + 0.006512327119708061, + 0.012898984365165234, + 0.011496036313474178, + 0.013833443634212017, + 0.01231439784169197, + 0.007594938389956951, + 0.007825052365660667, + 0.008539359085261822, + 0.007332806941121817, + 0.012855949811637402, + 0.011537973769009113, + 0.016647007316350937, + 0.014209138229489326, + 0.00855037197470665, + 0.007952367886900902, + 0.008011204190552235, + 0.008421868085861206, + 0.008850169368088245, + 0.010574393905699253, + 0.013944189995527267, + 0.006894818507134914, + 0.007786983624100685, + 0.009157948195934296, + 0.009080268442630768, + 0.007779367733746767, + 0.010956982150673866, + 0.013633830472826958, + 0.011046702042222023, + 0.00888485461473465, + 0.009262640960514545, + 0.038008444011211395, + 0.04133559390902519, + 0.04091503471136093, + 0.049447860568761826, + 0.03856314718723297, + 0.04001861810684204, + 0.026275930926203728, + 0.022453757002949715, + 0.005791232921183109, + 0.006487205624580383, + 0.006462689954787493, + 0.007783870678395033, + 0.009918776340782642, + 0.004512438550591469, + 0.0053132567554712296, + 0.007115482352674007, + 0.004289377946406603, + 0.004527088720351458, + 0.004916233941912651, + 0.004001512657850981, + 0.004706535488367081, + 0.004985612351447344, + 0.003987110685557127, + 0.004229601006954908, + 0.0057713668793439865, + 0.003720315406098962, + 0.007746042218059301, + 0.005218160804361105, + 0.009747871197760105, + 0.007998866960406303, + 0.004892039578408003, + 0.007328275125473738 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 0, + 2 + ], + "selected_reward": -0.1003124788403511 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "high-angle drone camera slowly zooms forward and tilts slightly downward, creating a fluid, ink-and-wash painting effect, its edges adorned with a natural, smudged effect. Illuminated by the cool moonlight, a derelict ancient temple nestled deep in a mountain valley emerges. Its crumbling eaves and moss-covered dark tiles appear as varying shades of ink against the night sky. A woman in white, the sole bright spot in this dynamic painting, clad in a flowing white silk dress. Her voluminous sleeves and cascading jet-black hair unfurl in the night breeze, like nimble brushstrokes gliding across rice paper. Silently, almost parallel to the ground, she drifts forward, her movements light and steady. Like a wisp of smoke, she crosses the overgrown courtyard, passes through the open, stained vermilion gate, and descends into the darkness of the temple's main hall. The moonlight outlines the scene with clear contours and large, thick shadows, forming a strong black and white contrast. All dynamic elements have a softness and fluidity like ink dissolving in water, creating an extremely quiet, mysterious and poignant oriental fantasy atmosphere.", + "seed": 7215, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000095/group_2/index_8782_seed_7215_dp_4_rank_8_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 2, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.06892634676769376, + "vlm_logprob": -0.06892634676769376, + "vlm_yes_no_margin": 3.25, + "vlm_token_credit": -0.06892634676769376, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a derelict ancient temple located in a mountain valley?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a woman wearing a flowing white silk dress visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the woman drift forward silently while crossing an overgrown courtyard and passing through a stained vermilion gate?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the video exhibit an ink-and-wash painting style with smudged edges and fluid, dissolving ink effects?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot a high-angle view that slowly zooms forward and tilts slightly downward?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.17545126378536224 + ], + "opposite_logprob_tokens": [ + -1.9254512786865234 + ], + "logprob_reward": -0.17545126378536224, + "yes_no_margin_reward": 1.75, + "token_credit_reward": -0.17545126378536224, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12327969074249268, + 0.05162494629621506, + 0.0753980204463005, + 0.06469769775867462, + 0.07962342351675034, + 0.1784171164035797, + 0.056898947805166245, + 0.2672143578529358, + 0.07653511315584183, + 0.02631072700023651 + ], + "credit_map": [ + 0.011509515345096588, + 0.015010851435363293, + 0.017400648444890976, + 0.01295000221580267, + 0.01238884124904871, + 0.017807047814130783, + 0.01970686949789524, + 0.01650591567158699, + 0.004911752417683601, + 0.005465119611471891, + 0.010963806882500648, + 0.005761454813182354, + 0.006126913242042065, + 0.003971730824559927, + 0.009070378728210926, + 0.005353791639208794, + 0.007340768352150917, + 0.008300221525132656, + 0.008565487340092659, + 0.008917195722460747, + 0.010743897408246994, + 0.008331368677318096, + 0.01364340540021658, + 0.009555669501423836, + 0.00790448673069477, + 0.008266928605735302, + 0.007530327420681715, + 0.006333223078399897, + 0.00864267349243164, + 0.0076238117180764675, + 0.009013500064611435, + 0.009382744319736958, + 0.012552188709378242, + 0.010129168629646301, + 0.007174245081841946, + 0.0050280108116567135, + 0.012778080999851227, + 0.011017589829862118, + 0.005519072525203228, + 0.015425066463649273, + 0.020346995443105698, + 0.01853875257074833, + 0.01932046376168728, + 0.027556806802749634, + 0.020915133878588676, + 0.02942587062716484, + 0.021653130650520325, + 0.020659958943724632, + 0.005813561379909515, + 0.005775496829301119, + 0.008286054246127605, + 0.00426913145929575, + 0.01123837660998106, + 0.011346197687089443, + 0.005087412428110838, + 0.005082717631012201, + 0.030888769775629044, + 0.03011150285601616, + 0.07365341484546661, + 0.01803794503211975, + 0.022187182679772377, + 0.024499017745256424, + 0.013196407817304134, + 0.05464010685682297, + 0.008355710655450821, + 0.01185637153685093, + 0.00595046766102314, + 0.006665358319878578, + 0.01503897923976183, + 0.009600907564163208, + 0.011349924840033054, + 0.00771739287301898, + 0.0020713070407509804, + 0.002928651636466384, + 0.0033111560624092817, + 0.0033717670012265444, + 0.0031683386769145727, + 0.003978549502789974, + 0.0031188796274363995, + 0.004362077452242374 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 5, + 0 + ], + "selected_reward": -0.17545126378536224 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a derelict ancient temple located in a mountain valley?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a woman wearing a flowing white silk dress visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the woman drift forward silently while crossing an overgrown courtyard and passing through a stained vermilion gate?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the video exhibit an ink-and-wash painting style with smudged edges and fluid, dissolving ink effects?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot a high-angle view that slowly zooms forward and tilts slightly downward?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04866965487599373 + ], + "opposite_logprob_tokens": [ + -3.0486695766448975 + ], + "logprob_reward": -0.04866965487599373, + "yes_no_margin_reward": 3.0, + "token_credit_reward": -0.04866965487599373, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.04547707736492157, + 0.050855014473199844, + 0.038756731897592545, + 0.04988870769739151, + 0.029236743226647377, + 0.08502164483070374, + 0.07326939702033997, + 0.11293605715036392, + 0.11479394137859344, + 0.39976468682289124 + ], + "credit_map": [ + 0.003869615728035569, + 0.004573972430080175, + 0.005864501930773258, + 0.005172520875930786, + 0.0047680409625172615, + 0.005649164319038391, + 0.010817554779350758, + 0.004761706106364727, + 0.004546007607132196, + 0.00478809280321002, + 0.007385860662907362, + 0.005464212503284216, + 0.007057329174131155, + 0.006365156266838312, + 0.008704097010195255, + 0.006544260308146477, + 0.003426057519391179, + 0.005842764861881733, + 0.003919755574315786, + 0.003287942847236991, + 0.006086127366870642, + 0.005716617219150066, + 0.004092070274055004, + 0.006385398097336292, + 0.006829180289059877, + 0.005941587500274181, + 0.005963532254099846, + 0.005236527882516384, + 0.006466193590313196, + 0.007902390323579311, + 0.005337165202945471, + 0.006212128791958094, + 0.006447004619985819, + 0.002830424113199115, + 0.002688732696697116, + 0.003410353558138013, + 0.003553941147401929, + 0.0024484649766236544, + 0.0033356444910168648, + 0.004522177390754223, + 0.009365309029817581, + 0.008476005867123604, + 0.009698044508695602, + 0.011933115310966969, + 0.010339999571442604, + 0.015000898391008377, + 0.009618229232728481, + 0.010590044781565666, + 0.006187876220792532, + 0.006824176758527756, + 0.007189588155597448, + 0.004166845232248306, + 0.015547946095466614, + 0.01968015916645527, + 0.006755264941602945, + 0.006917538587003946, + 0.01543601043522358, + 0.011763661168515682, + 0.03328898921608925, + 0.00790910329669714, + 0.011041761375963688, + 0.008811503648757935, + 0.004977057687938213, + 0.019707970321178436, + 0.01372439507395029, + 0.015339921228587627, + 0.011666637845337391, + 0.012515807524323463, + 0.017172062769532204, + 0.014096132479608059, + 0.01452387310564518, + 0.01575511135160923, + 0.027526114135980606, + 0.04651172459125519, + 0.041638463735580444, + 0.04674795642495155, + 0.04538072645664215, + 0.06608246266841888, + 0.0398532934486866, + 0.0860239714384079 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 8, + 7 + ], + "selected_reward": -0.04866965487599373 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a derelict ancient temple located in a mountain valley?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a woman wearing a flowing white silk dress visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the woman drift forward silently while crossing an overgrown courtyard and passing through a stained vermilion gate?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the video exhibit an ink-and-wash painting style with smudged edges and fluid, dissolving ink effects?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot a high-angle view that slowly zooms forward and tilts slightly downward?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.007643855642527342 + ], + "opposite_logprob_tokens": [ + -4.882643699645996 + ], + "logprob_reward": -0.007643855642527342, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.007643855642527342, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13877612352371216, + 0.07805185765028, + 0.10523176193237305, + 0.13068236410617828, + 0.0639079213142395, + 0.0589117631316185, + 0.059445034712553024, + 0.13174325227737427, + 0.07599804550409317, + 0.15725186467170715 + ], + "credit_map": [ + 0.013550269417464733, + 0.019679272547364235, + 0.01669241115450859, + 0.014817400835454464, + 0.014627241529524326, + 0.017535675317049026, + 0.025120705366134644, + 0.016753146424889565, + 0.007896668277680874, + 0.008323756977915764, + 0.016893282532691956, + 0.008687619119882584, + 0.008790362626314163, + 0.007696181535720825, + 0.011287909001111984, + 0.008476076647639275, + 0.010309531353414059, + 0.016159774735569954, + 0.010111097246408463, + 0.010398422367870808, + 0.011620339006185532, + 0.014165151864290237, + 0.017157457768917084, + 0.015309996902942657, + 0.013224348425865173, + 0.020244963467121124, + 0.014190459623932838, + 0.012836918234825134, + 0.016886640340089798, + 0.014973071403801441, + 0.016935279592871666, + 0.021390685811638832, + 0.00774062005802989, + 0.011462146416306496, + 0.004747434984892607, + 0.004829497542232275, + 0.008782914839684963, + 0.00880646612495184, + 0.0046005877666175365, + 0.012938256375491619, + 0.006340670399367809, + 0.005731461104005575, + 0.005058919079601765, + 0.007497511804103851, + 0.008567743003368378, + 0.013160486705601215, + 0.00569377513602376, + 0.006861198227852583, + 0.004726596176624298, + 0.006254239939153194, + 0.005939666647464037, + 0.005776763428002596, + 0.009345200844109058, + 0.012818834744393826, + 0.00809466652572155, + 0.006489066407084465, + 0.021472230553627014, + 0.026226188987493515, + 0.0178200863301754, + 0.009472916834056377, + 0.01527891680598259, + 0.011766275390982628, + 0.010077380575239658, + 0.019629258662462234, + 0.00823698379099369, + 0.011275318451225758, + 0.0076319193467497826, + 0.008981508202850819, + 0.010568776167929173, + 0.011041677556931973, + 0.008305570110678673, + 0.009956291876733303, + 0.01166731771081686, + 0.016549598425626755, + 0.019233617931604385, + 0.01694369688630104, + 0.02033747360110283, + 0.02664782479405403, + 0.016459181904792786, + 0.029413148760795593 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 0, + 7 + ], + "selected_reward": -0.007643855642527342 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a derelict ancient temple located in a mountain valley?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a woman wearing a flowing white silk dress visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the woman drift forward silently while crossing an overgrown courtyard and passing through a stained vermilion gate?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the video exhibit an ink-and-wash painting style with smudged edges and fluid, dissolving ink effects?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot a high-angle view that slowly zooms forward and tilts slightly downward?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.012549950741231441 + ], + "opposite_logprob_tokens": [ + -4.387549877166748 + ], + "logprob_reward": -0.012549950741231441, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.012549950741231441, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1680809110403061, + 0.11453726887702942, + 0.06806281208992004, + 0.10419304668903351, + 0.08006282895803452, + 0.05563115328550339, + 0.08193322271108627, + 0.1502368301153183, + 0.03480783849954605, + 0.1424541175365448 + ], + "credit_map": [ + 0.018361615017056465, + 0.02075457014143467, + 0.02639727108180523, + 0.02188500575721264, + 0.01882508024573326, + 0.017239458858966827, + 0.02428528107702732, + 0.02033262699842453, + 0.010729062370955944, + 0.013747715391218662, + 0.014980610460042953, + 0.01582157053053379, + 0.018380196765065193, + 0.007355549838393927, + 0.012469279579818249, + 0.02105328068137169, + 0.009942205622792244, + 0.008116334676742554, + 0.012004626914858818, + 0.007294118404388428, + 0.008553426712751389, + 0.006438171956688166, + 0.0077374824322760105, + 0.00797644630074501, + 0.01325929630547762, + 0.014710983261466026, + 0.01277326699346304, + 0.011426853947341442, + 0.012560715898871422, + 0.009439656510949135, + 0.01382946316152811, + 0.016192810609936714, + 0.016870252788066864, + 0.008012188598513603, + 0.003884972771629691, + 0.005853475537151098, + 0.013750557787716389, + 0.006752602756023407, + 0.004490481223911047, + 0.02044830285012722, + 0.006425759755074978, + 0.006978464312851429, + 0.005896785296499729, + 0.006093475501984358, + 0.007812644354999065, + 0.009888711385428905, + 0.0062086088582873344, + 0.006326709408313036, + 0.007058792281895876, + 0.0077124228700995445, + 0.006746344733983278, + 0.0047399140894412994, + 0.015344316139817238, + 0.02395661175251007, + 0.006373304408043623, + 0.010001517832279205, + 0.02369735948741436, + 0.024256359785795212, + 0.03463646024465561, + 0.010663861408829689, + 0.0168691985309124, + 0.011664025485515594, + 0.009423411451280117, + 0.019026145339012146, + 0.0027297064661979675, + 0.00313068856485188, + 0.0038935381453484297, + 0.003024446312338114, + 0.0043754735961556435, + 0.00686521315947175, + 0.00610815454274416, + 0.0046806191094219685, + 0.012993100099265575, + 0.01401455793529749, + 0.019138691946864128, + 0.017898358404636383, + 0.023082973435521126, + 0.020407022908329964, + 0.016977261751890182, + 0.017942147329449654 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 7, + 9 + ], + "selected_reward": -0.012549950741231441 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a derelict ancient temple located in a mountain valley?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a woman wearing a flowing white silk dress visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the woman drift forward silently while crossing an overgrown courtyard and passing through a stained vermilion gate?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the video exhibit an ink-and-wash painting style with smudged edges and fluid, dissolving ink effects?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot a high-angle view that slowly zooms forward and tilts slightly downward?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.10031700879335403 + ], + "opposite_logprob_tokens": [ + -2.3503170013427734 + ], + "logprob_reward": -0.10031700879335403, + "yes_no_margin_reward": 2.25, + "token_credit_reward": -0.10031700879335403, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06073756888508797, + 0.04259075969457626, + 0.04900912195444107, + 0.12313064187765121, + 0.07389101386070251, + 0.08502037823200226, + 0.12352515012025833, + 0.19694790244102478, + 0.050821028649806976, + 0.19432643055915833 + ], + "credit_map": [ + 0.00715478602796793, + 0.01121615618467331, + 0.006482250522822142, + 0.006221339106559753, + 0.0072770374827086926, + 0.005289475433528423, + 0.00953192450106144, + 0.007564603816717863, + 0.004091324284672737, + 0.0038566996809095144, + 0.010996603406965733, + 0.0032171078491955996, + 0.006502335425466299, + 0.003535994328558445, + 0.003661906812340021, + 0.006728791166096926, + 0.005668682046234608, + 0.00444764643907547, + 0.006298836786299944, + 0.006539064459502697, + 0.006748372223228216, + 0.004299656022340059, + 0.009079952724277973, + 0.00592691358178854, + 0.017151441425085068, + 0.02019442431628704, + 0.014513520523905754, + 0.011928356252610683, + 0.014708399772644043, + 0.011038350872695446, + 0.011873255483806133, + 0.02172289416193962, + 0.01391302328556776, + 0.0077415588311851025, + 0.008958295919001102, + 0.005005254875868559, + 0.00878899171948433, + 0.008811969310045242, + 0.009176185354590416, + 0.011495734564960003, + 0.00945389736443758, + 0.009592747315764427, + 0.009451049380004406, + 0.00986968632787466, + 0.011222627013921738, + 0.016033073887228966, + 0.009930901229381561, + 0.00946639385074377, + 0.010857893154025078, + 0.012341616675257683, + 0.017059095203876495, + 0.010892177931964397, + 0.026639314368367195, + 0.022815750911831856, + 0.011227229610085487, + 0.011692078784108162, + 0.02626360021531582, + 0.01785189099609852, + 0.06032118201255798, + 0.014484049752354622, + 0.017912840470671654, + 0.016303284093737602, + 0.009060008451342583, + 0.0347510389983654, + 0.005320581141859293, + 0.0065918355248868465, + 0.00500252703204751, + 0.0068057687021791935, + 0.005010807886719704, + 0.010230316780507565, + 0.005254280753433704, + 0.006604907568544149, + 0.013207255862653255, + 0.022589867934584618, + 0.021691307425498962, + 0.018095646053552628, + 0.02432374097406864, + 0.03298034146428108, + 0.020933164283633232, + 0.040505118668079376 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 9, + 6 + ], + "selected_reward": -0.10031700879335403 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "high-angle drone camera slowly zooms forward and tilts slightly downward, creating a fluid, ink-and-wash painting effect, its edges adorned with a natural, smudged effect. Illuminated by the cool moonlight, a derelict ancient temple nestled deep in a mountain valley emerges. Its crumbling eaves and moss-covered dark tiles appear as varying shades of ink against the night sky. A woman in white, the sole bright spot in this dynamic painting, clad in a flowing white silk dress. Her voluminous sleeves and cascading jet-black hair unfurl in the night breeze, like nimble brushstrokes gliding across rice paper. Silently, almost parallel to the ground, she drifts forward, her movements light and steady. Like a wisp of smoke, she crosses the overgrown courtyard, passes through the open, stained vermilion gate, and descends into the darkness of the temple's main hall. The moonlight outlines the scene with clear contours and large, thick shadows, forming a strong black and white contrast. All dynamic elements have a softness and fluidity like ink dissolving in water, creating an extremely quiet, mysterious and poignant oriental fantasy atmosphere.", + "seed": 7216, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000095/group_2/index_8782_seed_7216_dp_5_rank_10_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 2, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.05892499256879091, + "vlm_logprob": -0.05892499256879091, + "vlm_yes_no_margin": 3.324999976158142, + "vlm_token_credit": -0.05892499256879091, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a derelict ancient temple located in a mountain valley?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a woman wearing a flowing white silk dress visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the woman drift forward silently while crossing an overgrown courtyard and passing through a stained vermilion gate?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the video exhibit an ink-and-wash painting style with smudged edges and fluid, dissolving ink effects?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot a high-angle view that slowly zooms forward and tilts slightly downward?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.13427622616291046 + ], + "opposite_logprob_tokens": [ + -2.1342761516571045 + ], + "logprob_reward": -0.13427622616291046, + "yes_no_margin_reward": 1.9999998807907104, + "token_credit_reward": -0.13427622616291046, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13800206780433655, + 0.04572824761271477, + 0.24299593269824982, + 0.03606797009706497, + 0.2360229641199112, + 0.08971591293811798, + 0.0482306107878685, + 0.06913867592811584, + 0.041924893856048584, + 0.05217277631163597 + ], + "credit_map": [ + 0.013802812434732914, + 0.01537474524229765, + 0.018987784162163734, + 0.017071006819605827, + 0.027515551075339317, + 0.013062797486782074, + 0.014443966560065746, + 0.01774340681731701, + 0.004408382344990969, + 0.007714167702943087, + 0.004146512597799301, + 0.005371804814785719, + 0.005868831183761358, + 0.006373545154929161, + 0.008081663399934769, + 0.003763342509046197, + 0.033795081079006195, + 0.027219215407967567, + 0.027645723894238472, + 0.025405893102288246, + 0.044801972806453705, + 0.03818625211715698, + 0.02292005345225334, + 0.02302173525094986, + 0.003557416144758463, + 0.003319192212074995, + 0.004215017892420292, + 0.003248636145144701, + 0.005682844668626785, + 0.004832420032471418, + 0.0056170145981013775, + 0.005595428869128227, + 0.02107611671090126, + 0.02164858765900135, + 0.02568593993782997, + 0.020545758306980133, + 0.04545211419463158, + 0.032750651240348816, + 0.028812440112233162, + 0.04005136340856552, + 0.007951529696583748, + 0.00709772203117609, + 0.03600170090794563, + 0.0035685698967427015, + 0.006445595528930426, + 0.022213386371731758, + 0.002916396129876375, + 0.0035210158675909042, + 0.01282837986946106, + 0.004106808919459581, + 0.004865395370870829, + 0.003518033539876342, + 0.0027796593494713306, + 0.004881617147475481, + 0.002826665062457323, + 0.012424050830304623, + 0.008375019766390324, + 0.003969196230173111, + 0.021311083808541298, + 0.005253649782389402, + 0.007003219798207283, + 0.006114578805863857, + 0.005403756629675627, + 0.011708169244229794, + 0.004064398352056742, + 0.007645898498594761, + 0.004918545484542847, + 0.003945227712392807, + 0.006660806480795145, + 0.006227818783372641, + 0.004302553832530975, + 0.004159645177423954, + 0.006319972686469555, + 0.007190565578639507, + 0.006089489907026291, + 0.004272617399692535, + 0.007568547036498785, + 0.007876130752265453, + 0.006675671320408583, + 0.006179782096296549 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 4, + 0 + ], + "selected_reward": -0.13427622616291046 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a derelict ancient temple located in a mountain valley?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a woman wearing a flowing white silk dress visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the woman drift forward silently while crossing an overgrown courtyard and passing through a stained vermilion gate?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the video exhibit an ink-and-wash painting style with smudged edges and fluid, dissolving ink effects?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot a high-angle view that slowly zooms forward and tilts slightly downward?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.06206699460744858 + ], + "opposite_logprob_tokens": [ + -2.8120670318603516 + ], + "logprob_reward": -0.06206699460744858, + "yes_no_margin_reward": 2.75, + "token_credit_reward": -0.06206699460744858, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.22902911901474, + 0.14391760528087616, + 0.1707925796508789, + 0.03330635279417038, + 0.03311494365334511, + 0.0341409407556057, + 0.14650659263134003, + 0.09414754807949066, + 0.08259190618991852, + 0.03245248273015022 + ], + "credit_map": [ + 0.021243296563625336, + 0.030874717980623245, + 0.03591018170118332, + 0.03394056111574173, + 0.027103351429104805, + 0.022569701075553894, + 0.031267281621694565, + 0.02612001821398735, + 0.015327917411923409, + 0.027039607986807823, + 0.01724550873041153, + 0.015912601724267006, + 0.022702250629663467, + 0.021138189360499382, + 0.013571579940617085, + 0.010979960672557354, + 0.026568910107016563, + 0.018923016265034676, + 0.01736481301486492, + 0.01612711139023304, + 0.031216289848089218, + 0.026143696159124374, + 0.014599308371543884, + 0.019849436357617378, + 0.002797300461679697, + 0.0042509352788329124, + 0.0034925583750009537, + 0.004986855201423168, + 0.006964844185858965, + 0.0054594953544437885, + 0.002899725688621402, + 0.002454637084156275, + 0.005055766552686691, + 0.00413904432207346, + 0.0045814719051122665, + 0.003001031232997775, + 0.0046841963194310665, + 0.0046036941930651665, + 0.0040688482113182545, + 0.002980890916660428, + 0.0037798869889229536, + 0.0034857201389968395, + 0.0031670951284468174, + 0.0023443589452654123, + 0.005304102785885334, + 0.009428807534277439, + 0.0033916933462023735, + 0.0032392770517617464, + 0.04676350578665733, + 0.009933040477335453, + 0.03880666196346283, + 0.010634164325892925, + 0.0053870766423642635, + 0.011060711927711964, + 0.0064132604748010635, + 0.017508164048194885, + 0.011341540142893791, + 0.007002767641097307, + 0.025416474789381027, + 0.006577591877430677, + 0.008930784650146961, + 0.009656828828155994, + 0.00735172163695097, + 0.01786983385682106, + 0.0072166756726801395, + 0.016252370551228523, + 0.009638645686209202, + 0.0083385668694973, + 0.013040222227573395, + 0.013810585252940655, + 0.007901259697973728, + 0.006393582094460726, + 0.003376139560714364, + 0.005016302224248648, + 0.0038726148195564747, + 0.0034428846556693316, + 0.004867233335971832, + 0.005861876532435417, + 0.0029516052454710007, + 0.003063824260607362 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 6 + ], + "selected_reward": -0.06206699460744858 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a derelict ancient temple located in a mountain valley?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a woman wearing a flowing white silk dress visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the woman drift forward silently while crossing an overgrown courtyard and passing through a stained vermilion gate?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the video exhibit an ink-and-wash painting style with smudged edges and fluid, dissolving ink effects?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot a high-angle view that slowly zooms forward and tilts slightly downward?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.012557013891637325 + ], + "opposite_logprob_tokens": [ + -4.387557029724121 + ], + "logprob_reward": -0.012557013891637325, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.012557013891637325, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10510476678609848, + 0.06996333599090576, + 0.077525295317173, + 0.026378527283668518, + 0.41386184096336365, + 0.07524056732654572, + 0.04068220406770706, + 0.08971317112445831, + 0.07025078684091568, + 0.03127950802445412 + ], + "credit_map": [ + 0.009824514389038086, + 0.012581931427121162, + 0.02043929137289524, + 0.009579498320817947, + 0.0173932034522295, + 0.009914267808198929, + 0.01673712581396103, + 0.008634933270514011, + 0.006566085387021303, + 0.007127776741981506, + 0.007707657758146524, + 0.006506762932986021, + 0.019354121759533882, + 0.010165962390601635, + 0.007825823500752449, + 0.004709139000624418, + 0.015281563624739647, + 0.006058021914213896, + 0.0063240258023142815, + 0.006071721203625202, + 0.011778413318097591, + 0.012254409492015839, + 0.008149492554366589, + 0.01160765066742897, + 0.002965095918625593, + 0.0030129796359688044, + 0.0021646542008966208, + 0.0032329850364476442, + 0.005138974171131849, + 0.0039293197914958, + 0.0031495094299316406, + 0.0027850100304931402, + 0.03884465992450714, + 0.0420740582048893, + 0.051747843623161316, + 0.038911204785108566, + 0.08075161278247833, + 0.05430730804800987, + 0.046875275671482086, + 0.06034988537430763, + 0.007581030484288931, + 0.004679459612816572, + 0.023727118968963623, + 0.003339918330311775, + 0.010374725796282291, + 0.015181592665612698, + 0.005064185243099928, + 0.005292536225169897, + 0.00438835471868515, + 0.003361061215400696, + 0.011530308052897453, + 0.002680715639144182, + 0.003501594066619873, + 0.004531270824372768, + 0.002735625486820936, + 0.007953273132443428, + 0.010064280591905117, + 0.01251738891005516, + 0.0133881326764822, + 0.012726368382573128, + 0.009141220711171627, + 0.012249344028532505, + 0.00911281630396843, + 0.01051362045109272, + 0.006789754144847393, + 0.012941434048116207, + 0.00854523852467537, + 0.006338205654174089, + 0.011229735799133778, + 0.010432499460875988, + 0.005785326007753611, + 0.008188593201339245, + 0.00386333791539073, + 0.0035911481827497482, + 0.004100276622921228, + 0.0035650134086608887, + 0.004196306690573692, + 0.00453784316778183, + 0.0031283183488994837, + 0.004297263920307159 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 0, + 7 + ], + "selected_reward": -0.012557013891637325 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a derelict ancient temple located in a mountain valley?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a woman wearing a flowing white silk dress visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the woman drift forward silently while crossing an overgrown courtyard and passing through a stained vermilion gate?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the video exhibit an ink-and-wash painting style with smudged edges and fluid, dissolving ink effects?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot a high-angle view that slowly zooms forward and tilts slightly downward?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.006746964529156685 + ], + "opposite_logprob_tokens": [ + -5.006746768951416 + ], + "logprob_reward": -0.006746964529156685, + "yes_no_margin_reward": 5.0, + "token_credit_reward": -0.006746964529156685, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10093891620635986, + 0.07275210320949554, + 0.12352696806192398, + 0.06603789329528809, + 0.08804154396057129, + 0.15237495303153992, + 0.06046845391392708, + 0.04170790687203407, + 0.1581556648015976, + 0.13599556684494019 + ], + "credit_map": [ + 0.00775857362896204, + 0.009885587729513645, + 0.027095116674900055, + 0.009662175551056862, + 0.014916443265974522, + 0.01091246772557497, + 0.009805731475353241, + 0.010902820155024529, + 0.007529479451477528, + 0.009364569559693336, + 0.008890878409147263, + 0.009626642800867558, + 0.009195840917527676, + 0.011634583584964275, + 0.009814408607780933, + 0.006695697084069252, + 0.01801067776978016, + 0.015595226548612118, + 0.01233470905572176, + 0.012319684028625488, + 0.02353840135037899, + 0.019053256139159203, + 0.010036862455308437, + 0.01263813953846693, + 0.007599623408168554, + 0.007255383301526308, + 0.00787582527846098, + 0.0047514368779957294, + 0.016002167016267776, + 0.009663674049079418, + 0.007091647479683161, + 0.005798137281090021, + 0.00860659684985876, + 0.013519282452762127, + 0.01500720251351595, + 0.00829631183296442, + 0.012256595306098461, + 0.011200477369129658, + 0.008886539377272129, + 0.010268528014421463, + 0.01334468089044094, + 0.018344178795814514, + 0.05830777436494827, + 0.006456386297941208, + 0.014354938641190529, + 0.02494053915143013, + 0.008127727545797825, + 0.008498704060912132, + 0.015265491791069508, + 0.004857492633163929, + 0.01975078508257866, + 0.004218103364109993, + 0.004115019924938679, + 0.005981709808111191, + 0.003204798325896263, + 0.0030750499572604895, + 0.004922972992062569, + 0.0059980065561831, + 0.009180099703371525, + 0.004756022244691849, + 0.004259887617081404, + 0.004084423184394836, + 0.005170710384845734, + 0.0033357797656208277, + 0.0181326474994421, + 0.027729474008083344, + 0.020234936848282814, + 0.014016973786056042, + 0.027232395485043526, + 0.019457431510090828, + 0.012193269096314907, + 0.019158538430929184, + 0.02210322953760624, + 0.019706541672348976, + 0.02063458226621151, + 0.0111700976267457, + 0.02068045549094677, + 0.016310082748532295, + 0.014928906224668026, + 0.01046166755259037 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 5, + 9 + ], + "selected_reward": -0.006746964529156685 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a derelict ancient temple located in a mountain valley?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a woman wearing a flowing white silk dress visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the woman drift forward silently while crossing an overgrown courtyard and passing through a stained vermilion gate?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the video exhibit an ink-and-wash painting style with smudged edges and fluid, dissolving ink effects?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot a high-angle view that slowly zooms forward and tilts slightly downward?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07897776365280151 + ], + "opposite_logprob_tokens": [ + -2.5789778232574463 + ], + "logprob_reward": -0.07897776365280151, + "yes_no_margin_reward": 2.5, + "token_credit_reward": -0.07897776365280151, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11885309219360352, + 0.07150525599718094, + 0.12066280096769333, + 0.06967058777809143, + 0.10892107337713242, + 0.2878410816192627, + 0.04529104381799698, + 0.07850481569766998, + 0.07304418832063675, + 0.025706037878990173 + ], + "credit_map": [ + 0.012350218370556831, + 0.012987056747078896, + 0.015317023731768131, + 0.012546907179057598, + 0.021610232070088387, + 0.013609576970338821, + 0.018225006759166718, + 0.012207073159515858, + 0.008020211942493916, + 0.01056650746613741, + 0.006837120745331049, + 0.006833280436694622, + 0.010598084889352322, + 0.012497692368924618, + 0.010333189740777016, + 0.005819168407469988, + 0.019119789823889732, + 0.013440520502626896, + 0.010491494089365005, + 0.012306678108870983, + 0.02502742037177086, + 0.018004368990659714, + 0.008279329165816307, + 0.013993209227919579, + 0.004956965334713459, + 0.010655536316335201, + 0.006728367879986763, + 0.005416669882833958, + 0.01631307043135166, + 0.009602520614862442, + 0.007084795273840427, + 0.008912666700780392, + 0.01087900996208191, + 0.017038865014910698, + 0.013281428255140781, + 0.01055027637630701, + 0.019148236140608788, + 0.014923689886927605, + 0.008991778828203678, + 0.014107788912951946, + 0.023402346298098564, + 0.028452500700950623, + 0.10711796581745148, + 0.013334103859961033, + 0.029899094253778458, + 0.046062711626291275, + 0.021430443972349167, + 0.01814192347228527, + 0.011805388145148754, + 0.004059218801558018, + 0.009722292423248291, + 0.0030166145879775286, + 0.004947962239384651, + 0.004223068710416555, + 0.0022414610721170902, + 0.005275040864944458, + 0.008641126565635204, + 0.009410135447978973, + 0.012227194383740425, + 0.014122886583209038, + 0.00667667668312788, + 0.009744251146912575, + 0.009004118852317333, + 0.008678424172103405, + 0.006647941190749407, + 0.01577490195631981, + 0.007967759855091572, + 0.006671049166470766, + 0.012417598627507687, + 0.011521677486598492, + 0.006046435330063105, + 0.005996823776513338, + 0.00285058724693954, + 0.002437600400298834, + 0.002515319734811783, + 0.002871364587917924, + 0.0038931400049477816, + 0.00453021889552474, + 0.0022702321875840425, + 0.00433757621794939 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 2, + 0 + ], + "selected_reward": -0.07897776365280151 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "high-angle drone camera slowly zooms forward and tilts slightly downward, creating a fluid, ink-and-wash painting effect, its edges adorned with a natural, smudged effect. Illuminated by the cool moonlight, a derelict ancient temple nestled deep in a mountain valley emerges. Its crumbling eaves and moss-covered dark tiles appear as varying shades of ink against the night sky. A woman in white, the sole bright spot in this dynamic painting, clad in a flowing white silk dress. Her voluminous sleeves and cascading jet-black hair unfurl in the night breeze, like nimble brushstrokes gliding across rice paper. Silently, almost parallel to the ground, she drifts forward, her movements light and steady. Like a wisp of smoke, she crosses the overgrown courtyard, passes through the open, stained vermilion gate, and descends into the darkness of the temple's main hall. The moonlight outlines the scene with clear contours and large, thick shadows, forming a strong black and white contrast. All dynamic elements have a softness and fluidity like ink dissolving in water, creating an extremely quiet, mysterious and poignant oriental fantasy atmosphere.", + "seed": 7217, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000095/group_2/index_8782_seed_7217_dp_5_rank_10_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 2, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.0376403066329658, + "vlm_logprob": -0.0376403066329658, + "vlm_yes_no_margin": 3.949999976158142, + "vlm_token_credit": -0.0376403066329658, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a derelict ancient temple located in a mountain valley?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a woman wearing a flowing white silk dress visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the woman drift forward silently while crossing an overgrown courtyard and passing through a stained vermilion gate?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the video exhibit an ink-and-wash painting style with smudged edges and fluid, dissolving ink effects?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot a high-angle view that slowly zooms forward and tilts slightly downward?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.028511036187410355 + ], + "opposite_logprob_tokens": [ + -3.7785110473632812 + ], + "logprob_reward": -0.028511036187410355, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.028511036187410355, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0874248519539833, + 0.05724981799721718, + 0.09580814838409424, + 0.023179305717349052, + 0.09183571487665176, + 0.05432302877306938, + 0.38067615032196045, + 0.036385584622621536, + 0.1431858241558075, + 0.029931599274277687 + ], + "credit_map": [ + 0.016606174409389496, + 0.014624616131186485, + 0.009314129129052162, + 0.009923294186592102, + 0.0067731463350355625, + 0.006529968697577715, + 0.01755300536751747, + 0.0061005232855677605, + 0.007283476646989584, + 0.005643701180815697, + 0.009360873140394688, + 0.011008420959115028, + 0.0032570238690823317, + 0.006296392530202866, + 0.006985881831496954, + 0.0074140517972409725, + 0.013422541320323944, + 0.013525285758078098, + 0.00958287250250578, + 0.01354124117642641, + 0.007961159572005272, + 0.012911089695990086, + 0.006631540600210428, + 0.018232429400086403, + 0.0024875265080481768, + 0.0023429831489920616, + 0.004529832396656275, + 0.0032103811390697956, + 0.0015676779439672828, + 0.0019373054383322597, + 0.0024983882904052734, + 0.004605212714523077, + 0.004912546370178461, + 0.023595057427883148, + 0.0032960649114102125, + 0.005145474337041378, + 0.032660409808158875, + 0.005550255533307791, + 0.0048546502366662025, + 0.011821269057691097, + 0.006712848786264658, + 0.006885802838951349, + 0.006520987953990698, + 0.00527278333902359, + 0.005680692847818136, + 0.007145852316170931, + 0.006292831152677536, + 0.00981123372912407, + 0.03122478537261486, + 0.028096137568354607, + 0.04088875651359558, + 0.022875288501381874, + 0.09538182616233826, + 0.09852582961320877, + 0.026979386806488037, + 0.03670414537191391, + 0.004457033239305019, + 0.004227780271321535, + 0.0071673085913062096, + 0.0031868815422058105, + 0.00524930190294981, + 0.004952411167323589, + 0.0034309630282223225, + 0.003713910235092044, + 0.011367416009306908, + 0.0060525937005877495, + 0.009049014188349247, + 0.004641009494662285, + 0.055133190006017685, + 0.04459084942936897, + 0.004514703527092934, + 0.00783704686909914, + 0.003051101230084896, + 0.004662021063268185, + 0.004917954094707966, + 0.0031207865104079247, + 0.0027753079775720835, + 0.004487860016524792, + 0.0022041487973183393, + 0.0047124214470386505 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 8, + 2 + ], + "selected_reward": -0.028511036187410355 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a derelict ancient temple located in a mountain valley?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a woman wearing a flowing white silk dress visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the woman drift forward silently while crossing an overgrown courtyard and passing through a stained vermilion gate?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the video exhibit an ink-and-wash painting style with smudged edges and fluid, dissolving ink effects?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot a high-angle view that slowly zooms forward and tilts slightly downward?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004655949305742979 + ], + "opposite_logprob_tokens": [ + -5.379655838012695 + ], + "logprob_reward": -0.004655949305742979, + "yes_no_margin_reward": 5.375, + "token_credit_reward": -0.004655949305742979, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.044197928160429, + 0.03173365816473961, + 0.04918207600712776, + 0.04453492909669876, + 0.17925406992435455, + 0.10117600113153458, + 0.04278922453522682, + 0.0507955439388752, + 0.24806807935237885, + 0.2082684487104416 + ], + "credit_map": [ + 0.007753023412078619, + 0.0079380813986063, + 0.008174898102879524, + 0.004064707085490227, + 0.0035707505885511637, + 0.0051233526319265366, + 0.004185774363577366, + 0.003387338714674115, + 0.005232034716755152, + 0.003762512933462858, + 0.0036640267353504896, + 0.005018656142055988, + 0.0016564176185056567, + 0.003706936025992036, + 0.004719013348221779, + 0.003974060993641615, + 0.006114894989877939, + 0.006284680683165789, + 0.003638762515038252, + 0.007875635288655758, + 0.0032573852222412825, + 0.006730655673891306, + 0.004195342771708965, + 0.011084718629717827, + 0.005733610596507788, + 0.0048088133335113525, + 0.007903915829956532, + 0.005981676280498505, + 0.0036912227515131235, + 0.003040003590285778, + 0.006537753157317638, + 0.006837934721261263, + 0.009356833063066006, + 0.06287215650081635, + 0.006321772467344999, + 0.00938924215734005, + 0.03743188828229904, + 0.010262830182909966, + 0.011174339801073074, + 0.032444994896650314, + 0.00941788125783205, + 0.016075465828180313, + 0.00967487320303917, + 0.010936666280031204, + 0.014268288388848305, + 0.014621126465499401, + 0.013952690176665783, + 0.012229014188051224, + 0.0038315404672175646, + 0.004604813177138567, + 0.004431514535099268, + 0.005517204292118549, + 0.006056188140064478, + 0.0076959095895290375, + 0.005204036366194487, + 0.005448020063340664, + 0.005991815123707056, + 0.0054272874258458614, + 0.008704327046871185, + 0.004142817575484514, + 0.009535725228488445, + 0.007800988852977753, + 0.004750054329633713, + 0.0044425297528505325, + 0.01685367338359356, + 0.01435113325715065, + 0.013164587318897247, + 0.011498740874230862, + 0.0810806080698967, + 0.0721864327788353, + 0.01104727666825056, + 0.027885638177394867, + 0.01582571677863598, + 0.021379379555583, + 0.018728138878941536, + 0.018798919394612312, + 0.029365431517362595, + 0.036953993141651154, + 0.020598163828253746, + 0.04661870375275612 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 9, + 4 + ], + "selected_reward": -0.004655949305742979 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a derelict ancient temple located in a mountain valley?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a woman wearing a flowing white silk dress visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the woman drift forward silently while crossing an overgrown courtyard and passing through a stained vermilion gate?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the video exhibit an ink-and-wash painting style with smudged edges and fluid, dissolving ink effects?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot a high-angle view that slowly zooms forward and tilts slightly downward?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.009794510900974274 + ], + "opposite_logprob_tokens": [ + -4.63479471206665 + ], + "logprob_reward": -0.009794510900974274, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.009794510900974274, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.04315749555826187, + 0.04731640964746475, + 0.05148807913064957, + 0.039011042565107346, + 0.29247021675109863, + 0.05035122111439705, + 0.18652565777301788, + 0.06323417276144028, + 0.18612723052501678, + 0.04031839966773987 + ], + "credit_map": [ + 0.006850409787148237, + 0.004663445521146059, + 0.006757075432687998, + 0.007742272689938545, + 0.0030787240248173475, + 0.004570177756249905, + 0.005598037037998438, + 0.0038973549380898476, + 0.008453300222754478, + 0.005351055413484573, + 0.0062548089772462845, + 0.007118352688848972, + 0.003890658263117075, + 0.0051115937530994415, + 0.0038174944929778576, + 0.007319154217839241, + 0.005975265521556139, + 0.004644243512302637, + 0.004707104526460171, + 0.01109990756958723, + 0.0025031326804310083, + 0.005195217672735453, + 0.00615630391985178, + 0.011206908151507378, + 0.0037256639916449785, + 0.0035379640758037567, + 0.004075923003256321, + 0.005415667779743671, + 0.0033257873728871346, + 0.004504702053964138, + 0.008490909822285175, + 0.005934427957981825, + 0.014639737084507942, + 0.11046396940946579, + 0.00917850248515606, + 0.011121455579996109, + 0.09182699769735336, + 0.013352455571293831, + 0.012646310031414032, + 0.029240818694233894, + 0.006974596530199051, + 0.005629675462841988, + 0.00555846095085144, + 0.005355348810553551, + 0.00417108042165637, + 0.006489641033113003, + 0.007713300175964832, + 0.008459120988845825, + 0.01183866336941719, + 0.018081633374094963, + 0.023686951026320457, + 0.022775081917643547, + 0.029857778921723366, + 0.04850520193576813, + 0.013032044284045696, + 0.018748316913843155, + 0.007859165780246258, + 0.0069866483099758625, + 0.010375797748565674, + 0.006726306863129139, + 0.008273900486528873, + 0.007841365411877632, + 0.006836413871496916, + 0.008334577083587646, + 0.018242545425891876, + 0.00843017641454935, + 0.01412292756140232, + 0.006420494988560677, + 0.07050292938947678, + 0.0493871308863163, + 0.005917941685765982, + 0.013103104196488857, + 0.0037131633143872023, + 0.006008706521242857, + 0.0057439301162958145, + 0.004128788132220507, + 0.006243179552257061, + 0.006272173020988703, + 0.002806117059662938, + 0.005402346607297659 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 6, + 8 + ], + "selected_reward": -0.009794510900974274 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a derelict ancient temple located in a mountain valley?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a woman wearing a flowing white silk dress visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the woman drift forward silently while crossing an overgrown courtyard and passing through a stained vermilion gate?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the video exhibit an ink-and-wash painting style with smudged edges and fluid, dissolving ink effects?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot a high-angle view that slowly zooms forward and tilts slightly downward?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.018195031210780144 + ], + "opposite_logprob_tokens": [ + -4.018195152282715 + ], + "logprob_reward": -0.018195031210780144, + "yes_no_margin_reward": 4.0, + "token_credit_reward": -0.018195031210780144, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06319668889045715, + 0.09705711156129837, + 0.19652707874774933, + 0.018376421183347702, + 0.12512370944023132, + 0.2055344581604004, + 0.12461903691291809, + 0.07649695128202438, + 0.041719138622283936, + 0.051349420100450516 + ], + "credit_map": [ + 0.010743211023509502, + 0.01160331442952156, + 0.009937475435435772, + 0.004578168038278818, + 0.005334915593266487, + 0.006589442491531372, + 0.009097514674067497, + 0.0053126453422009945, + 0.013576135039329529, + 0.008850918151438236, + 0.012825654819607735, + 0.017178786918520927, + 0.0071518379263579845, + 0.011273486539721489, + 0.012547848746180534, + 0.013652448542416096, + 0.02764734998345375, + 0.029649630188941956, + 0.025771230459213257, + 0.02073611691594124, + 0.0165182463824749, + 0.02888675034046173, + 0.013539864681661129, + 0.03377788886427879, + 0.0022577790077775717, + 0.002539937850087881, + 0.002887427108362317, + 0.002482954878360033, + 0.0012445724569261074, + 0.0024339184165000916, + 0.002077268436551094, + 0.002452561631798744, + 0.008105709217488766, + 0.03778572008013725, + 0.006646991241723299, + 0.009617036208510399, + 0.010606322437524796, + 0.00811127107590437, + 0.008259007707238197, + 0.03599165380001068, + 0.02598683163523674, + 0.03593845292925835, + 0.026474980637431145, + 0.020099736750125885, + 0.020157011225819588, + 0.025297418236732483, + 0.019568804651498795, + 0.032011229544878006, + 0.009066220372915268, + 0.01020044181495905, + 0.013176391832530499, + 0.010361219756305218, + 0.02950464002788067, + 0.031367380172014236, + 0.00923323817551136, + 0.011709500104188919, + 0.009841798804700375, + 0.00898030772805214, + 0.00841696560382843, + 0.0077470033429563046, + 0.012932553887367249, + 0.008629477582871914, + 0.009931321255862713, + 0.01001752633601427, + 0.004084235522896051, + 0.003245777217671275, + 0.005256446544080973, + 0.003215544857084751, + 0.009287072345614433, + 0.007892509922385216, + 0.002278924221172929, + 0.006458628922700882, + 0.005125454626977444, + 0.00823249202221632, + 0.006461888551712036, + 0.006153618451207876, + 0.006726916879415512, + 0.006996911950409412, + 0.004011338111013174, + 0.007640797179192305 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 2, + 4 + ], + "selected_reward": -0.018195031210780144 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a derelict ancient temple located in a mountain valley?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a woman wearing a flowing white silk dress visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the woman drift forward silently while crossing an overgrown courtyard and passing through a stained vermilion gate?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the video exhibit an ink-and-wash painting style with smudged edges and fluid, dissolving ink effects?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot a high-angle view that slowly zooms forward and tilts slightly downward?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.12704500555992126 + ], + "opposite_logprob_tokens": [ + -2.127044916152954 + ], + "logprob_reward": -0.12704500555992126, + "yes_no_margin_reward": 1.9999998807907104, + "token_credit_reward": -0.12704500555992126, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.04916653037071228, + 0.027620214968919754, + 0.05331750959157944, + 0.05601325258612633, + 0.36632806062698364, + 0.09243705868721008, + 0.20286822319030762, + 0.03975236415863037, + 0.05784089118242264, + 0.05465587228536606 + ], + "credit_map": [ + 0.007382733281701803, + 0.009631005115807056, + 0.006520650815218687, + 0.00666944170370698, + 0.003945447038859129, + 0.00659593241289258, + 0.004644846078008413, + 0.0037764699663966894, + 0.0032766745425760746, + 0.002418144838884473, + 0.004460569005459547, + 0.005174766760319471, + 0.0015734549378976226, + 0.0027967779897153378, + 0.004284409340471029, + 0.0036354181356728077, + 0.005798265803605318, + 0.006425405386835337, + 0.008550268597900867, + 0.008558196946978569, + 0.0038945006672292948, + 0.0069262804463505745, + 0.004858155269175768, + 0.008306438103318214, + 0.0053465100936591625, + 0.005515166558325291, + 0.007841638289391994, + 0.00937614869326353, + 0.004937794525176287, + 0.004529798403382301, + 0.009036924690008163, + 0.009429270401597023, + 0.01992006227374077, + 0.12941022217273712, + 0.011248632334172726, + 0.01930942013859749, + 0.07852885127067566, + 0.021390682086348534, + 0.01916353777050972, + 0.0673566684126854, + 0.008069667033851147, + 0.013208133168518543, + 0.009134870953857899, + 0.010334158316254616, + 0.01166443433612585, + 0.012424344196915627, + 0.013560634106397629, + 0.014040812849998474, + 0.0149531289935112, + 0.01266536209732294, + 0.02523801289498806, + 0.01180035900324583, + 0.044256553053855896, + 0.062248360365629196, + 0.012931144796311855, + 0.018775297328829765, + 0.005525250919163227, + 0.0038750916719436646, + 0.006228122394531965, + 0.0038830540142953396, + 0.005768640898168087, + 0.005077088251709938, + 0.0035681233275681734, + 0.00582699291408062, + 0.002882472239434719, + 0.0031772921793162823, + 0.0036045187152922153, + 0.002962730126455426, + 0.019671302288770676, + 0.018152419477701187, + 0.0022835549898445606, + 0.005106599070131779, + 0.0068956962786614895, + 0.0108376145362854, + 0.009574866853654385, + 0.005983309354633093, + 0.005584544502198696, + 0.007402434479445219, + 0.0030090934596955776, + 0.005368316546082497 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 6, + 5 + ], + "selected_reward": -0.12704500555992126 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000095/group_3/index_tensor([6580, 6580])_group_3_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000095/group_3/index_tensor([6580, 6580])_group_3_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..90d34aa06aaf0451128912367e84849e7b4d90a3 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000095/group_3/index_tensor([6580, 6580])_group_3_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "static, eye-level medium shot, framed with documentary-style realism, captures a weary champion sitting perfectly centered on a worn wooden bench in a desolate, silent locker room. His sweat-dampened black hair clings to his forehead, his muscular shoulders slump with exhaustion, and a simple white towel hangs from his neck. He wears only a pair of dark gym shorts, and the white bandages on his hands are partially untied, dangling loosely. His face, etched with the physical toll of his victory, is a mask of exhaustion; his gaze is unfocused and glassy, staring past the camera into an unseen distance. Beside him, a massive, gleaming golden trophy sits on the bench, close yet pointedly overlooked, its polished surface catching the cold, sterile light from a single overhead fluorescent bulb, casting a bright, almost mocking reflection against the drab, gray concrete walls and the recessed metal lockers that line the room. He took a slow, deep breath, barely lifting his chest, and then let out a deliberately heavy, audible sigh, as if draining away the last of his energy, his body slumping further. The cheers of victory were a forgotten echo, replaced by the deep, hollow emptiness of the silent room and the grim reality of victory's consequences.", + "seed": 12487, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000095/group_3/index_6580_seed_12487_dp_6_rank_12_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 3, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.21379857864230872, + "vlm_logprob": -0.21379857864230872, + "vlm_yes_no_margin": 2.775, + "vlm_token_credit": -0.21379857864230872, + "vlm_yes_ratio": 0.8, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot framed as a static, eye-level medium shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a weary champion sitting perfectly centered on a worn wooden bench?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the character wearing dark gym shorts with white bandages on his hands that are partially untied and dangling?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a massive, gleaming golden trophy visible on the bench beside the character?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the environment a desolate, silent locker room with drab gray concrete walls and recessed metal lockers?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.08084350824356079 + ], + "opposite_logprob_tokens": [ + -3.205843448638916 + ], + "logprob_reward": -0.08084350824356079, + "yes_no_margin_reward": 3.125, + "token_credit_reward": -0.08084350824356079, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2577243447303772, + 0.03824672847986221, + 0.04753606393933296, + 0.0650988519191742, + 0.04145198315382004, + 0.0648711621761322, + 0.10309577733278275, + 0.12902523577213287, + 0.10845868289470673, + 0.14449121057987213 + ], + "credit_map": [ + 0.02542690560221672, + 0.012890743091702461, + 0.08206124603748322, + 0.017584027722477913, + 0.015185965225100517, + 0.07850649207830429, + 0.008028564974665642, + 0.018040405586361885, + 0.0031919232569634914, + 0.003428500844165683, + 0.009930955246090889, + 0.003277404233813286, + 0.002952928887680173, + 0.010094881057739258, + 0.002245948649942875, + 0.0031241904944181442, + 0.004775844980031252, + 0.0028705978766083717, + 0.012763709761202335, + 0.003601386211812496, + 0.002404331462457776, + 0.01541439350694418, + 0.0026536991354078054, + 0.0030521033331751823, + 0.008870062418282032, + 0.0057535795494914055, + 0.011475668288767338, + 0.005346357356756926, + 0.004825163166970015, + 0.01696513406932354, + 0.004362792707979679, + 0.007500097621232271, + 0.0044479588977992535, + 0.004300189670175314, + 0.006713469047099352, + 0.003878864226862788, + 0.006226262077689171, + 0.008662712760269642, + 0.0033744536340236664, + 0.003848074236884713, + 0.005933312699198723, + 0.006320206914097071, + 0.016420084983110428, + 0.005862125661224127, + 0.006336848251521587, + 0.01677747070789337, + 0.002592876786366105, + 0.00462823873385787, + 0.014069427736103535, + 0.0063819498755037785, + 0.02222180739045143, + 0.006833448074758053, + 0.00946143176406622, + 0.02927372418344021, + 0.00489323865622282, + 0.009960745461285114, + 0.01125567127019167, + 0.007145140320062637, + 0.03257283940911293, + 0.008532938547432423, + 0.007185013499110937, + 0.042223602533340454, + 0.008606855757534504, + 0.011503173038363457, + 0.013810898177325726, + 0.0060025653801858425, + 0.02611999586224556, + 0.008026537485420704, + 0.01012794952839613, + 0.02870655059814453, + 0.005734421778470278, + 0.009929763153195381, + 0.012042446061968803, + 0.012438586913049221, + 0.03780673071742058, + 0.01163495797663927, + 0.008089302107691765, + 0.042134083807468414, + 0.009768396615982056, + 0.010576705448329449 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 7 + ], + "selected_reward": -0.08084350824356079 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot framed as a static, eye-level medium shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a weary champion sitting perfectly centered on a worn wooden bench?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the character wearing dark gym shorts with white bandages on his hands that are partially untied and dangling?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a massive, gleaming golden trophy visible on the bench beside the character?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the environment a desolate, silent locker room with drab gray concrete walls and recessed metal lockers?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03812633082270622 + ], + "opposite_logprob_tokens": [ + -3.288126230239868 + ], + "logprob_reward": -0.03812633082270622, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.03812633082270622, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2261408120393753, + 0.0605161152780056, + 0.0276353619992733, + 0.36063846945762634, + 0.020240766927599907, + 0.034051328897476196, + 0.0528775155544281, + 0.08386635780334473, + 0.02543698251247406, + 0.10859625041484833 + ], + "credit_map": [ + 0.01857762783765793, + 0.008090917021036148, + 0.07808983325958252, + 0.014692535623908043, + 0.007938975468277931, + 0.07703433185815811, + 0.005622690077871084, + 0.016093911603093147, + 0.005656562279909849, + 0.008228282444179058, + 0.008128305897116661, + 0.007289614994078875, + 0.0049849869683384895, + 0.011520820669829845, + 0.0032768764067441225, + 0.01143066305667162, + 0.0031227844301611185, + 0.0022044992074370384, + 0.005877547897398472, + 0.0019838132429867983, + 0.0024912317749112844, + 0.007104521617293358, + 0.0016671321354806423, + 0.0031838316936045885, + 0.02205531857907772, + 0.008869593963027, + 0.11940935999155045, + 0.030763696879148483, + 0.014688801020383835, + 0.13855773210525513, + 0.008325631730258465, + 0.017968343570828438, + 0.002607540460303426, + 0.0017779979389160872, + 0.0032567342277616262, + 0.002337896730750799, + 0.0021564755588769913, + 0.004008254036307335, + 0.002352391602471471, + 0.0017434773035347462, + 0.0030310405418276787, + 0.0021577002480626106, + 0.007602164521813393, + 0.00229736533947289, + 0.004618334583938122, + 0.00836510956287384, + 0.0030853438656777143, + 0.002894270932301879, + 0.007246449589729309, + 0.0032240753062069416, + 0.010400854982435703, + 0.0035674183163791895, + 0.0038162481505423784, + 0.016386594623327255, + 0.0033006269950419664, + 0.004935251083225012, + 0.005675658583641052, + 0.001905712066218257, + 0.028582390397787094, + 0.007418391294777393, + 0.0019923290237784386, + 0.03125546872615814, + 0.002621531719341874, + 0.00441486993804574, + 0.0026078778319060802, + 0.0018711101729422808, + 0.0051652672700583935, + 0.0020922122057527304, + 0.0023767377715557814, + 0.005553854163736105, + 0.002827337011694908, + 0.0029425867833197117, + 0.00965237244963646, + 0.010763286612927914, + 0.01620238833129406, + 0.015548540279269218, + 0.012480650097131729, + 0.013118199072778225, + 0.017089787870645523, + 0.013741027563810349 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 0, + 9 + ], + "selected_reward": -0.03812633082270622 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot framed as a static, eye-level medium shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a weary champion sitting perfectly centered on a worn wooden bench?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the character wearing dark gym shorts with white bandages on his hands that are partially untied and dangling?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a massive, gleaming golden trophy visible on the bench beside the character?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the environment a desolate, silent locker room with drab gray concrete walls and recessed metal lockers?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04307105019688606 + ], + "opposite_logprob_tokens": [ + -3.1680710315704346 + ], + "logprob_reward": -0.04307105019688606, + "yes_no_margin_reward": 3.125, + "token_credit_reward": -0.04307105019688606, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2274261713027954, + 0.09245456755161285, + 0.025705918669700623, + 0.0391552671790123, + 0.0518639050424099, + 0.07652482390403748, + 0.0354592464864254, + 0.2595234215259552, + 0.03082626312971115, + 0.1610604226589203 + ], + "credit_map": [ + 0.017996061593294144, + 0.0067203897051513195, + 0.07494635879993439, + 0.014832695946097374, + 0.0142945796251297, + 0.07612094283103943, + 0.01047146413475275, + 0.012043680995702744, + 0.006171705666929483, + 0.0036090780049562454, + 0.030851416289806366, + 0.0065513900481164455, + 0.0030166751239448786, + 0.03438011184334755, + 0.0025170648004859686, + 0.005357122048735619, + 0.0018194473814219236, + 0.0022442894987761974, + 0.0068497550673782825, + 0.0016331676160916686, + 0.002112109214067459, + 0.007490516174584627, + 0.0015656384639441967, + 0.0019909958355128765, + 0.005193392746150494, + 0.0027549222577363253, + 0.010657926090061665, + 0.002970433095470071, + 0.0028276697266846895, + 0.009013420902192593, + 0.0024215010926127434, + 0.0033160015009343624, + 0.0047925603576004505, + 0.004824057687073946, + 0.009367646649479866, + 0.004440240561962128, + 0.008758910931646824, + 0.00809240248054266, + 0.006814776919782162, + 0.004773308057337999, + 0.007303683087229729, + 0.00592083903029561, + 0.01645648293197155, + 0.005347241181880236, + 0.008687648922204971, + 0.01721840724349022, + 0.00813824962824583, + 0.007452270947396755, + 0.002740828087553382, + 0.0022633697371929884, + 0.009523678570985794, + 0.0026962291449308395, + 0.0027341789100319147, + 0.010806182399392128, + 0.0019810576923191547, + 0.002713722875341773, + 0.017437303438782692, + 0.003998544532805681, + 0.09575670212507248, + 0.020334014669060707, + 0.004647134803235531, + 0.10273083299398422, + 0.0029919592197984457, + 0.01162694115191698, + 0.0040290020406246185, + 0.001831425353884697, + 0.006412748247385025, + 0.002277134219184518, + 0.0030308887362480164, + 0.007493420969694853, + 0.002505467738956213, + 0.0032461760565638542, + 0.011573124676942825, + 0.013352758251130581, + 0.03932074084877968, + 0.012791736982762814, + 0.012648684903979301, + 0.04609384015202522, + 0.012509236112236977, + 0.012770305387675762 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 0, + 9 + ], + "selected_reward": -0.04307105019688606 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot framed as a static, eye-level medium shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a weary champion sitting perfectly centered on a worn wooden bench?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the character wearing dark gym shorts with white bandages on his hands that are partially untied and dangling?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a massive, gleaming golden trophy visible on the bench beside the character?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the environment a desolate, silent locker room with drab gray concrete walls and recessed metal lockers?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0086353225633502 + ], + "opposite_logprob_tokens": [ + -4.758635520935059 + ], + "logprob_reward": -0.0086353225633502, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.0086353225633502, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08730567991733551, + 0.10665558278560638, + 0.09423887729644775, + 0.16138094663619995, + 0.029156846925616264, + 0.015387116931378841, + 0.01824009232223034, + 0.36204713582992554, + 0.02578878216445446, + 0.09979898482561111 + ], + "credit_map": [ + 0.009543030522763729, + 0.015055320225656033, + 0.011957798153162003, + 0.013728789985179901, + 0.007713379338383675, + 0.015141677111387253, + 0.005097130313515663, + 0.009068558923900127, + 0.01221142616122961, + 0.02000860497355461, + 0.015552552416920662, + 0.015352734364569187, + 0.00818728655576706, + 0.01934325136244297, + 0.004913267679512501, + 0.01108645647764206, + 0.0074072894640266895, + 0.009191850200295448, + 0.025369655340909958, + 0.007923915982246399, + 0.004231378901749849, + 0.03066132217645645, + 0.0029782955534756184, + 0.006475167348980904, + 0.009682182222604752, + 0.003234215546399355, + 0.05376918241381645, + 0.013597757555544376, + 0.005980811547487974, + 0.06404013186693192, + 0.002349932910874486, + 0.008726725354790688, + 0.003098198212683201, + 0.0029443984385579824, + 0.006835819687694311, + 0.0031093598809093237, + 0.001989465905353427, + 0.007695689797401428, + 0.001162018976174295, + 0.00232189754024148, + 0.002022928325459361, + 0.0012326180003583431, + 0.0033053660299628973, + 0.0009461002191528678, + 0.0012451348593458533, + 0.004296372178941965, + 0.0009578976314514875, + 0.0013806989882141352, + 0.0023626978509128094, + 0.0011109611950814724, + 0.003342387732118368, + 0.0014283440541476011, + 0.0016226430889219046, + 0.0049808393232524395, + 0.0015555336140096188, + 0.0018366845324635506, + 0.024030903354287148, + 0.006422055885195732, + 0.13126464188098907, + 0.02831149473786354, + 0.0069111124612390995, + 0.14474907517433167, + 0.0035501173697412014, + 0.016807766631245613, + 0.0031639710068702698, + 0.0018637441098690033, + 0.005478882696479559, + 0.0016527329571545124, + 0.0032106407452374697, + 0.005829279311001301, + 0.001966916723176837, + 0.0026226150803267956, + 0.00966197531670332, + 0.009219055995345116, + 0.020214183256030083, + 0.012409243732690811, + 0.006306578870862722, + 0.02245580591261387, + 0.008757399395108223, + 0.010774742811918259 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 3, + 1 + ], + "selected_reward": -0.0086353225633502 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot framed as a static, eye-level medium shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a weary champion sitting perfectly centered on a worn wooden bench?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the character wearing dark gym shorts with white bandages on his hands that are partially untied and dangling?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a massive, gleaming golden trophy visible on the bench beside the character?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the environment a desolate, silent locker room with drab gray concrete walls and recessed metal lockers?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -0.8983166813850403 + ], + "opposite_logprob_tokens": [ + -0.5233166813850403 + ], + "logprob_reward": -0.8983166813850403, + "yes_no_margin_reward": -0.375, + "token_credit_reward": -0.8983166813850403, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.08488733321428299, + 0.07445301860570908, + 0.024214956909418106, + 0.0650596097111702, + 0.02050134912133217, + 0.041474636644124985, + 0.017944755032658577, + 0.4274090528488159, + 0.04666847363114357, + 0.19738677144050598 + ], + "credit_map": [ + 0.00885178055614233, + 0.009726081974804401, + 0.016934465616941452, + 0.008899630978703499, + 0.008143075741827488, + 0.019752927124500275, + 0.005525460001081228, + 0.007053921464830637, + 0.0056440504267811775, + 0.004512574523687363, + 0.022917017340660095, + 0.0056936293840408325, + 0.0032051235903054476, + 0.025190221145749092, + 0.0025748885236680508, + 0.004715523682534695, + 0.002107545966282487, + 0.0025729872286319733, + 0.005214464385062456, + 0.0022093222942203283, + 0.0018506882479414344, + 0.006100719794631004, + 0.001328995800577104, + 0.0028302359860390425, + 0.007858810015022755, + 0.007366306148469448, + 0.010439742356538773, + 0.006816806271672249, + 0.007520734332501888, + 0.013861875049769878, + 0.0034306510351598263, + 0.007764684967696667, + 0.002479547169059515, + 0.0020783026702702045, + 0.0034696797374635935, + 0.0019897371530532837, + 0.0018594235880300403, + 0.0051908125169575214, + 0.0010693068616092205, + 0.0023645416367799044, + 0.004062316846102476, + 0.0041312603279948235, + 0.009499437175691128, + 0.003723332891240716, + 0.002889239229261875, + 0.011790418066084385, + 0.0016194394556805491, + 0.003759196028113365, + 0.0022187104914337397, + 0.0016933983424678445, + 0.002699965378269553, + 0.002150034299120307, + 0.001158396014943719, + 0.00475412467494607, + 0.0008859601221047342, + 0.0023841678630560637, + 0.030123790726065636, + 0.0062017375603318214, + 0.16055196523666382, + 0.03335552290081978, + 0.007347406353801489, + 0.1678144633769989, + 0.003562286728993058, + 0.018451934680342674, + 0.006488691549748182, + 0.003957206383347511, + 0.009684406220912933, + 0.004607269540429115, + 0.0031603712122887373, + 0.011333925649523735, + 0.0021303780376911163, + 0.005306231323629618, + 0.01599798910319805, + 0.017069442197680473, + 0.054903484880924225, + 0.01454130094498396, + 0.014649558812379837, + 0.06234455853700638, + 0.005876455456018448, + 0.012003988027572632 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 9, + 0 + ], + "selected_reward": -0.8983166813850403 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "No" + ] + }, + { + "prompt": "static, eye-level medium shot, framed with documentary-style realism, captures a weary champion sitting perfectly centered on a worn wooden bench in a desolate, silent locker room. His sweat-dampened black hair clings to his forehead, his muscular shoulders slump with exhaustion, and a simple white towel hangs from his neck. He wears only a pair of dark gym shorts, and the white bandages on his hands are partially untied, dangling loosely. His face, etched with the physical toll of his victory, is a mask of exhaustion; his gaze is unfocused and glassy, staring past the camera into an unseen distance. Beside him, a massive, gleaming golden trophy sits on the bench, close yet pointedly overlooked, its polished surface catching the cold, sterile light from a single overhead fluorescent bulb, casting a bright, almost mocking reflection against the drab, gray concrete walls and the recessed metal lockers that line the room. He took a slow, deep breath, barely lifting his chest, and then let out a deliberately heavy, audible sigh, as if draining away the last of his energy, his body slumping further. The cheers of victory were a forgotten echo, replaced by the deep, hollow emptiness of the silent room and the grim reality of victory's consequences.", + "seed": 12488, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000095/group_3/index_6580_seed_12488_dp_6_rank_12_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 3, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.3676397386938334, + "vlm_logprob": -0.3676397386938334, + "vlm_yes_no_margin": 2.35, + "vlm_token_credit": -0.3676397386938334, + "vlm_yes_ratio": 0.8, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot framed as a static, eye-level medium shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a weary champion sitting perfectly centered on a worn wooden bench?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the character wearing dark gym shorts with white bandages on his hands that are partially untied and dangling?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a massive, gleaming golden trophy visible on the bench beside the character?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the environment a desolate, silent locker room with drab gray concrete walls and recessed metal lockers?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.13613387942314148 + ], + "opposite_logprob_tokens": [ + -2.511133909225464 + ], + "logprob_reward": -0.13613387942314148, + "yes_no_margin_reward": 2.375, + "token_credit_reward": -0.13613387942314148, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.23359720408916473, + 0.06674481928348541, + 0.051076166331768036, + 0.06215761601924896, + 0.06690146028995514, + 0.11932214349508286, + 0.06818781793117523, + 0.17119024693965912, + 0.04859200119972229, + 0.11223048716783524 + ], + "credit_map": [ + 0.03560352697968483, + 0.034357041120529175, + 0.032345570623874664, + 0.014961917884647846, + 0.019790319725871086, + 0.03481974080204964, + 0.023373179137706757, + 0.03834589943289757, + 0.008595912717282772, + 0.01230715960264206, + 0.007149344310164452, + 0.0079665407538414, + 0.006479996722191572, + 0.00951759796589613, + 0.006523247342556715, + 0.008205023594200611, + 0.007288865279406309, + 0.0059158531948924065, + 0.007183312438428402, + 0.004224737640470266, + 0.006035788916051388, + 0.005721549969166517, + 0.005165250040590763, + 0.009540810249745846, + 0.00986719410866499, + 0.0093015655875206, + 0.007713891100138426, + 0.005447618197649717, + 0.00547297578305006, + 0.006827357225120068, + 0.0060160476714372635, + 0.011510966345667839, + 0.008350787684321404, + 0.008066998794674873, + 0.011470519006252289, + 0.005914150737226009, + 0.006069133523851633, + 0.00856467243283987, + 0.005668850615620613, + 0.01279634702950716, + 0.015002942644059658, + 0.010786676779389381, + 0.015458973124623299, + 0.009420911781489849, + 0.019095811992883682, + 0.011339519172906876, + 0.01254963781684637, + 0.025667667388916016, + 0.008779952302575111, + 0.005514000076800585, + 0.011182502843439579, + 0.005450024735182524, + 0.007217164151370525, + 0.0068652736954391, + 0.010032153688371181, + 0.01314675621688366, + 0.02176038920879364, + 0.017696959897875786, + 0.025909176096320152, + 0.012110847048461437, + 0.020631028339266777, + 0.02184319868683815, + 0.02672516368329525, + 0.024513481184840202, + 0.005534375552088022, + 0.007258973550051451, + 0.007810409180819988, + 0.003386755706742406, + 0.004214127082377672, + 0.005179921165108681, + 0.007761782500892878, + 0.007445657160133123, + 0.014773672446608543, + 0.010849556885659695, + 0.017091698944568634, + 0.01490017306059599, + 0.01555157545953989, + 0.008614230901002884, + 0.015533522702753544, + 0.014916053973138332 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 7, + 5 + ], + "selected_reward": -0.13613387942314148 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot framed as a static, eye-level medium shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a weary champion sitting perfectly centered on a worn wooden bench?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the character wearing dark gym shorts with white bandages on his hands that are partially untied and dangling?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a massive, gleaming golden trophy visible on the bench beside the character?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the environment a desolate, silent locker room with drab gray concrete walls and recessed metal lockers?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.048677489161491394 + ], + "opposite_logprob_tokens": [ + -3.048677444458008 + ], + "logprob_reward": -0.048677489161491394, + "yes_no_margin_reward": 3.0, + "token_credit_reward": -0.048677489161491394, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1543053537607193, + 0.04852459207177162, + 0.04668392613530159, + 0.07663983851671219, + 0.06290324032306671, + 0.17520727217197418, + 0.057093456387519836, + 0.20291319489479065, + 0.07542485743761063, + 0.10030431300401688 + ], + "credit_map": [ + 0.0261750016361475, + 0.02005929686129093, + 0.020270604640245438, + 0.012431182898581028, + 0.009199229069054127, + 0.023544704541563988, + 0.01315027941018343, + 0.029475051909685135, + 0.00804633554071188, + 0.006053008604794741, + 0.006212144624441862, + 0.00474293390288949, + 0.004450913984328508, + 0.005429620388895273, + 0.0039957561530172825, + 0.009593877010047436, + 0.006706685293465853, + 0.005380264949053526, + 0.007932834327220917, + 0.004233336541801691, + 0.0032050551380962133, + 0.004948382265865803, + 0.004071069415658712, + 0.010206297039985657, + 0.011598446406424046, + 0.008006185293197632, + 0.010558608919382095, + 0.005432890262454748, + 0.005872673820704222, + 0.014410431496798992, + 0.009101198986172676, + 0.011659405194222927, + 0.01018332876265049, + 0.005103711038827896, + 0.016460927203297615, + 0.005417764652520418, + 0.003971527796238661, + 0.006692223716527224, + 0.005128053016960621, + 0.009945710189640522, + 0.018556371331214905, + 0.014782105572521687, + 0.023691391572356224, + 0.023573346436023712, + 0.02740357257425785, + 0.01667313277721405, + 0.01957247406244278, + 0.03095487505197525, + 0.00864844024181366, + 0.0051636663265526295, + 0.009749911725521088, + 0.005016137845814228, + 0.0068224151618778706, + 0.00680437171831727, + 0.006376488599926233, + 0.008512025699019432, + 0.02693178504705429, + 0.03008028119802475, + 0.024680353701114655, + 0.012865161523222923, + 0.017927974462509155, + 0.028460178524255753, + 0.03405698016285896, + 0.02791047841310501, + 0.009739778935909271, + 0.00549167487770319, + 0.012752842158079147, + 0.0042305514216423035, + 0.011333474889397621, + 0.00893581472337246, + 0.009390306659042835, + 0.013550410978496075, + 0.013610376976430416, + 0.00782341044396162, + 0.019636603072285652, + 0.009164148941636086, + 0.011611773632466793, + 0.007823968306183815, + 0.010178388096392155, + 0.02045564167201519 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 5, + 0 + ], + "selected_reward": -0.048677489161491394 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot framed as a static, eye-level medium shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a weary champion sitting perfectly centered on a worn wooden bench?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the character wearing dark gym shorts with white bandages on his hands that are partially untied and dangling?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a massive, gleaming golden trophy visible on the bench beside the character?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the environment a desolate, silent locker room with drab gray concrete walls and recessed metal lockers?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.043065112084150314 + ], + "opposite_logprob_tokens": [ + -3.168065071105957 + ], + "logprob_reward": -0.043065112084150314, + "yes_no_margin_reward": 3.125, + "token_credit_reward": -0.043065112084150314, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2106146514415741, + 0.06747881323099136, + 0.05661895498633385, + 0.04211407154798508, + 0.07226347923278809, + 0.12448883801698685, + 0.07403691112995148, + 0.16254137456417084, + 0.042618006467819214, + 0.14722485840320587 + ], + "credit_map": [ + 0.029307648539543152, + 0.018617235124111176, + 0.035210058093070984, + 0.010523752309381962, + 0.03109527938067913, + 0.03924839571118355, + 0.02284676767885685, + 0.023765528574585915, + 0.009516301564872265, + 0.008459036238491535, + 0.008994978852570057, + 0.0072962697595357895, + 0.006043142639100552, + 0.008689970709383488, + 0.010019068606197834, + 0.008460055105388165, + 0.009572002105414867, + 0.0055866665206849575, + 0.0059459032490849495, + 0.003155139507725835, + 0.007329413667321205, + 0.00775515241548419, + 0.009266121312975883, + 0.008008559234440327, + 0.005824968684464693, + 0.005728079471737146, + 0.0053013949654996395, + 0.002621436258777976, + 0.006352551747113466, + 0.007050229236483574, + 0.004905278328806162, + 0.004330138675868511, + 0.011135033331811428, + 0.006620684172958136, + 0.009856609627604485, + 0.005799062550067902, + 0.01016642339527607, + 0.01081040408462286, + 0.006983030587434769, + 0.01089224312454462, + 0.016280610114336014, + 0.013973613269627094, + 0.023218240588903427, + 0.010667137801647186, + 0.012171766720712185, + 0.013949110172688961, + 0.015595781616866589, + 0.018632594496011734, + 0.01310886163264513, + 0.009680655784904957, + 0.011596543714404106, + 0.007025511469691992, + 0.006469466723501682, + 0.007088571321219206, + 0.006953586358577013, + 0.012113723903894424, + 0.019010327756404877, + 0.021038280799984932, + 0.02139737457036972, + 0.012974397279322147, + 0.018239157274365425, + 0.017430610954761505, + 0.029687795788049698, + 0.022763429209589958, + 0.0062570408917963505, + 0.005260088015347719, + 0.005474422127008438, + 0.0033361155074089766, + 0.004679435398429632, + 0.004362302832305431, + 0.00686038751155138, + 0.006388220936059952, + 0.01794595830142498, + 0.015008795075118542, + 0.019227318465709686, + 0.019179847091436386, + 0.01319199614226818, + 0.01927204616367817, + 0.018022995442152023, + 0.025375917553901672 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 7, + 9 + ], + "selected_reward": -0.043065112084150314 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot framed as a static, eye-level medium shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a weary champion sitting perfectly centered on a worn wooden bench?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the character wearing dark gym shorts with white bandages on his hands that are partially untied and dangling?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a massive, gleaming golden trophy visible on the bench beside the character?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the environment a desolate, silent locker room with drab gray concrete walls and recessed metal lockers?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.009776568040251732 + ], + "opposite_logprob_tokens": [ + -4.634776592254639 + ], + "logprob_reward": -0.009776568040251732, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.009776568040251732, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12015172094106674, + 0.03257592022418976, + 0.018395820632576942, + 0.04208575189113617, + 0.03369961678981781, + 0.055966880172491074, + 0.05044177174568176, + 0.6131908297538757, + 0.017964890226721764, + 0.01552677433937788 + ], + "credit_map": [ + 0.016277972608804703, + 0.009949902072548866, + 0.016173124313354492, + 0.0068509322591125965, + 0.012175729498267174, + 0.016943974420428276, + 0.017459774389863014, + 0.02432030811905861, + 0.004282687790691853, + 0.0031173499301075935, + 0.004697072319686413, + 0.002406618557870388, + 0.0031436753924936056, + 0.0035083747934550047, + 0.005885865073651075, + 0.005534274969249964, + 0.0029699034057557583, + 0.0019252232741564512, + 0.0030881327111274004, + 0.001397543353959918, + 0.0010282385628670454, + 0.0022579198703169823, + 0.0026545452419668436, + 0.0030743132811039686, + 0.0058824592269957066, + 0.002841629320755601, + 0.007037408649921417, + 0.0020479289814829826, + 0.008492602035403252, + 0.006185523699969053, + 0.005316879600286484, + 0.004281322471797466, + 0.0052971490658819675, + 0.003556728595867753, + 0.0050523290410637856, + 0.004261293448507786, + 0.003344559110701084, + 0.004363789223134518, + 0.003217728342860937, + 0.004606040194630623, + 0.006370581686496735, + 0.005947600118815899, + 0.008161570876836777, + 0.00596016738563776, + 0.005687841679900885, + 0.007361964788287878, + 0.008150321431457996, + 0.008326832205057144, + 0.005946354009211063, + 0.005637698341161013, + 0.006910942029207945, + 0.004197011701762676, + 0.006155700422823429, + 0.005866285413503647, + 0.006619209423661232, + 0.00910856667906046, + 0.07936150580644608, + 0.07447183132171631, + 0.079126738011837, + 0.04308372735977173, + 0.07304754853248596, + 0.07999533414840698, + 0.10441720485687256, + 0.07968690991401672, + 0.0021915107499808073, + 0.001697020954452455, + 0.0035238536074757576, + 0.0011377009795978665, + 0.0032350076362490654, + 0.0017654352122917771, + 0.0021708887070417404, + 0.0022434722632169724, + 0.0018716870108619332, + 0.0019490518607199192, + 0.0021072791423648596, + 0.0010483056539669633, + 0.0029219789430499077, + 0.0016031828708946705, + 0.002110376488417387, + 0.0019149115541949868 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 0, + 5 + ], + "selected_reward": -0.009776568040251732 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot framed as a static, eye-level medium shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a weary champion sitting perfectly centered on a worn wooden bench?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the character wearing dark gym shorts with white bandages on his hands that are partially untied and dangling?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a massive, gleaming golden trophy visible on the bench beside the character?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the environment a desolate, silent locker room with drab gray concrete walls and recessed metal lockers?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -1.6005456447601318 + ], + "opposite_logprob_tokens": [ + -0.22554564476013184 + ], + "logprob_reward": -1.6005456447601318, + "yes_no_margin_reward": -1.375, + "token_credit_reward": -1.6005456447601318, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.3656034469604492, + 0.11045720428228378, + 0.04623671621084213, + 0.0360758975148201, + 0.12395621836185455, + 0.09695896506309509, + 0.057349927723407745, + 0.05461489409208298, + 0.02523690089583397, + 0.08350983262062073 + ], + "credit_map": [ + 0.0538303405046463, + 0.022634560242295265, + 0.05703438073396683, + 0.020129773765802383, + 0.030149664729833603, + 0.057395003736019135, + 0.045553695410490036, + 0.07887604832649231, + 0.016263049095869064, + 0.009388206526637077, + 0.016970431432127953, + 0.006084006745368242, + 0.009263008832931519, + 0.014378073625266552, + 0.013096767477691174, + 0.02501365728676319, + 0.007245982065796852, + 0.0037452559918165207, + 0.006807250436395407, + 0.002303243149071932, + 0.004362206440418959, + 0.00600061472505331, + 0.005392902530729771, + 0.010379258543252945, + 0.005236011929810047, + 0.005095217376947403, + 0.005598882678896189, + 0.003525893436744809, + 0.002910075942054391, + 0.0050677708350121975, + 0.0030553224496543407, + 0.005586725194007158, + 0.026370801031589508, + 0.009678001515567303, + 0.017110975459218025, + 0.00803778413683176, + 0.00907123927026987, + 0.02559911645948887, + 0.008458815515041351, + 0.01962948776781559, + 0.0121445432305336, + 0.005662929732352495, + 0.014992994256317616, + 0.009077752009034157, + 0.011506706476211548, + 0.01700124889612198, + 0.012355849146842957, + 0.014216938056051731, + 0.0075303553603589535, + 0.0033592041581869125, + 0.009168121963739395, + 0.0032838373444974422, + 0.007656984031200409, + 0.008304337970912457, + 0.009013189934194088, + 0.009033901616930962, + 0.006947083864361048, + 0.004309586714953184, + 0.0073837959207594395, + 0.003996000625193119, + 0.007145354989916086, + 0.007794357370585203, + 0.010283281095325947, + 0.006755433510988951, + 0.003333677537739277, + 0.002345347311347723, + 0.004654126241803169, + 0.0017719391034916043, + 0.0025421208702027798, + 0.003873175010085106, + 0.0030237503815442324, + 0.0036927652545273304, + 0.013716328889131546, + 0.0075388685800135136, + 0.011750804260373116, + 0.0068433103151619434, + 0.008912531659007072, + 0.013183807954192162, + 0.01098859403282404, + 0.010575585067272186 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 1 + ], + "selected_reward": -1.6005456447601318 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "No" + ] + }, + { + "prompt": "static, eye-level medium shot, framed with documentary-style realism, captures a weary champion sitting perfectly centered on a worn wooden bench in a desolate, silent locker room. His sweat-dampened black hair clings to his forehead, his muscular shoulders slump with exhaustion, and a simple white towel hangs from his neck. He wears only a pair of dark gym shorts, and the white bandages on his hands are partially untied, dangling loosely. His face, etched with the physical toll of his victory, is a mask of exhaustion; his gaze is unfocused and glassy, staring past the camera into an unseen distance. Beside him, a massive, gleaming golden trophy sits on the bench, close yet pointedly overlooked, its polished surface catching the cold, sterile light from a single overhead fluorescent bulb, casting a bright, almost mocking reflection against the drab, gray concrete walls and the recessed metal lockers that line the room. He took a slow, deep breath, barely lifting his chest, and then let out a deliberately heavy, audible sigh, as if draining away the last of his energy, his body slumping further. The cheers of victory were a forgotten echo, replaced by the deep, hollow emptiness of the silent room and the grim reality of victory's consequences.", + "seed": 12489, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000095/group_3/index_6580_seed_12489_dp_7_rank_14_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 3, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.04206390641629696, + "vlm_logprob": -0.04206390641629696, + "vlm_yes_no_margin": 3.55, + "vlm_token_credit": -0.04206390641629696, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot framed as a static, eye-level medium shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a weary champion sitting perfectly centered on a worn wooden bench?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the character wearing dark gym shorts with white bandages on his hands that are partially untied and dangling?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a massive, gleaming golden trophy visible on the bench beside the character?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the environment a desolate, silent locker room with drab gray concrete walls and recessed metal lockers?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0611189529299736 + ], + "opposite_logprob_tokens": [ + -4.061119079589844 + ], + "logprob_reward": -0.0611189529299736, + "yes_no_margin_reward": 4.0, + "token_credit_reward": -0.0611189529299736, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.19576352834701538, + 0.15369662642478943, + 0.03750017657876015, + 0.03983241319656372, + 0.1433548480272293, + 0.03720870241522789, + 0.11658817529678345, + 0.016131944954395294, + 0.025015907362103462, + 0.23490770161151886 + ], + "credit_map": [ + 0.017211757600307465, + 0.016829704865813255, + 0.02576594240963459, + 0.014016957953572273, + 0.012961351312696934, + 0.04439258202910423, + 0.009171267040073872, + 0.05541396141052246, + 0.009326054714620113, + 0.005046391859650612, + 0.014089657925069332, + 0.004187385551631451, + 0.004072878044098616, + 0.05661925673484802, + 0.003787185065448284, + 0.056567806750535965, + 0.002882156055420637, + 0.003154950449243188, + 0.0038357314188033342, + 0.00337195280008018, + 0.00272081489674747, + 0.01040931511670351, + 0.0022644815035164356, + 0.008860772475600243, + 0.0022032649721950293, + 0.00307100941427052, + 0.004421495832502842, + 0.00466552609577775, + 0.0028590650763362646, + 0.007175501901656389, + 0.004442522767931223, + 0.010994026437401772, + 0.007661559619009495, + 0.003114585066214204, + 0.011362072080373764, + 0.0036036258097738028, + 0.002762575400993228, + 0.05657424405217171, + 0.003822092432528734, + 0.054454099386930466, + 0.0036661126650869846, + 0.001728564384393394, + 0.004579547327011824, + 0.0024064795579761267, + 0.002404636237770319, + 0.010717158205807209, + 0.0032254077959805727, + 0.008480795659124851, + 0.0061386325396597385, + 0.003260341938585043, + 0.00990303885191679, + 0.003819152247160673, + 0.0031572156585752964, + 0.0427115373313427, + 0.004367311019450426, + 0.043230947107076645, + 0.0017676141578704119, + 0.001344449003227055, + 0.002642218489199877, + 0.0016005303477868438, + 0.001385737443342805, + 0.0023443177342414856, + 0.0017610405338928103, + 0.003286037128418684, + 0.0031413265969604254, + 0.0026279950980097055, + 0.0036052309442311525, + 0.0022611687891185284, + 0.0023567022290080786, + 0.005692367907613516, + 0.0017799061024561524, + 0.0035512119065970182, + 0.032627250999212265, + 0.024650119245052338, + 0.03424069285392761, + 0.028532128781080246, + 0.018312636762857437, + 0.025833938270807266, + 0.035349853336811066, + 0.03536108508706093 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 0, + 1 + ], + "selected_reward": -0.0611189529299736 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot framed as a static, eye-level medium shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a weary champion sitting perfectly centered on a worn wooden bench?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the character wearing dark gym shorts with white bandages on his hands that are partially untied and dangling?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a massive, gleaming golden trophy visible on the bench beside the character?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the environment a desolate, silent locker room with drab gray concrete walls and recessed metal lockers?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.020608466118574142 + ], + "opposite_logprob_tokens": [ + -3.895608425140381 + ], + "logprob_reward": -0.020608466118574142, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.020608466118574142, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.23371905088424683, + 0.15812677145004272, + 0.0668259710073471, + 0.035244159400463104, + 0.10248582810163498, + 0.12330823391675949, + 0.08465331047773361, + 0.029270144179463387, + 0.03549521043896675, + 0.1308712214231491 + ], + "credit_map": [ + 0.02182907424867153, + 0.01370302028954029, + 0.024384798482060432, + 0.011777978390455246, + 0.01359819620847702, + 0.06861423701047897, + 0.00982943456619978, + 0.06998235732316971, + 0.009138542227447033, + 0.004400607664138079, + 0.012745695188641548, + 0.005656106863170862, + 0.004531464073807001, + 0.05991799011826515, + 0.004789301194250584, + 0.05694708228111267, + 0.0037858036812394857, + 0.003009631996974349, + 0.005405225791037083, + 0.005001053214073181, + 0.004395969212055206, + 0.020480040460824966, + 0.004767720587551594, + 0.0199805349111557, + 0.002016938989982009, + 0.0027098453138023615, + 0.0038069842848926783, + 0.005692136008292437, + 0.004132199566811323, + 0.004526829347014427, + 0.005460822023451328, + 0.006898411083966494, + 0.005086902529001236, + 0.004584429785609245, + 0.0082853389903903, + 0.006383996456861496, + 0.0033631352707743645, + 0.03462667763233185, + 0.004829096142202616, + 0.0353262722492218, + 0.005925252102315426, + 0.0036862429697066545, + 0.009919174015522003, + 0.0044044991955161095, + 0.003128963988274336, + 0.0464688315987587, + 0.003713817335665226, + 0.04606147110462189, + 0.004950141068547964, + 0.0018061459995806217, + 0.006598083768039942, + 0.0024713154416531324, + 0.001944420626387, + 0.03292256221175194, + 0.0026116734370589256, + 0.03134898468852043, + 0.0021043219603598118, + 0.0025181113742291927, + 0.0029935059137642384, + 0.0023561851121485233, + 0.001733179553411901, + 0.0069694179110229015, + 0.0023397800978273153, + 0.008255650289356709, + 0.002776677953079343, + 0.004739499185234308, + 0.004663646221160889, + 0.004945457447320223, + 0.0035885151010006666, + 0.003821759484708309, + 0.005180734675377607, + 0.005778925027698278, + 0.015454494394361973, + 0.014793230220675468, + 0.018204407766461372, + 0.01971432939171791, + 0.010537991300225258, + 0.013505319133400917, + 0.01961553655564785, + 0.01904592663049698 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 9 + ], + "selected_reward": -0.020608466118574142 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot framed as a static, eye-level medium shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a weary champion sitting perfectly centered on a worn wooden bench?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the character wearing dark gym shorts with white bandages on his hands that are partially untied and dangling?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a massive, gleaming golden trophy visible on the bench beside the character?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the environment a desolate, silent locker room with drab gray concrete walls and recessed metal lockers?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.026349563151597977 + ], + "opposite_logprob_tokens": [ + -3.6513495445251465 + ], + "logprob_reward": -0.026349563151597977, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.026349563151597977, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.29040440917015076, + 0.07934920489788055, + 0.0393812358379364, + 0.03440120816230774, + 0.2469000518321991, + 0.07275276631116867, + 0.08509413152933121, + 0.03385418653488159, + 0.04440711438655853, + 0.07345565408468246 + ], + "credit_map": [ + 0.02422630973160267, + 0.015372013673186302, + 0.03375738859176636, + 0.0172251146286726, + 0.02023214101791382, + 0.08175107836723328, + 0.012843784876167774, + 0.08499658107757568, + 0.007112327963113785, + 0.005041474476456642, + 0.007440801244229078, + 0.00771314138546586, + 0.005573981907218695, + 0.022129422053694725, + 0.005329967010766268, + 0.019008083269000053, + 0.003374628722667694, + 0.0028004078194499016, + 0.0034639746882021427, + 0.004790705628693104, + 0.004380817990750074, + 0.00785617996007204, + 0.006585509981960058, + 0.006129014305770397, + 0.0029653741512447596, + 0.0035254815593361855, + 0.0039553274400532246, + 0.003486725501716137, + 0.0035465024411678314, + 0.00533969234675169, + 0.005199214443564415, + 0.0063828881829977036, + 0.01395565364509821, + 0.006168078631162643, + 0.01935156248509884, + 0.00948327872902155, + 0.0054353331215679646, + 0.0947422981262207, + 0.007663004565984011, + 0.09010081738233566, + 0.003750487696379423, + 0.0024293598253279924, + 0.005865122191607952, + 0.002163226017728448, + 0.0023266233038157225, + 0.027081245556473732, + 0.0026839906349778175, + 0.026452716439962387, + 0.005039758048951626, + 0.0023053414188325405, + 0.006437474396079779, + 0.002538624219596386, + 0.0020834116730839014, + 0.032884132117033005, + 0.002864025766029954, + 0.03094136156141758, + 0.003234825562685728, + 0.004538138397037983, + 0.004742296412587166, + 0.0032585717272013426, + 0.002275758422911167, + 0.005658895708620548, + 0.003470949362963438, + 0.006674750708043575, + 0.006105128210037947, + 0.006953790783882141, + 0.008125968277454376, + 0.003972651436924934, + 0.004118261393159628, + 0.005775582976639271, + 0.0030079209245741367, + 0.00634781364351511, + 0.010571734979748726, + 0.006849588360637426, + 0.011240019463002682, + 0.010171286761760712, + 0.007342743221670389, + 0.009282595477998257, + 0.0069467537105083466, + 0.011050935834646225 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 6 + ], + "selected_reward": -0.026349563151597977 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot framed as a static, eye-level medium shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a weary champion sitting perfectly centered on a worn wooden bench?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the character wearing dark gym shorts with white bandages on his hands that are partially untied and dangling?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a massive, gleaming golden trophy visible on the bench beside the character?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the environment a desolate, silent locker room with drab gray concrete walls and recessed metal lockers?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07896344363689423 + ], + "opposite_logprob_tokens": [ + -2.5789635181427 + ], + "logprob_reward": -0.07896344363689423, + "yes_no_margin_reward": 2.5, + "token_credit_reward": -0.07896344363689423, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08879707008600235, + 0.12564921379089355, + 0.13749384880065918, + 0.03118235059082508, + 0.39485085010528564, + 0.052634820342063904, + 0.04718397557735443, + 0.016344353556632996, + 0.03134433925151825, + 0.07451916486024857 + ], + "credit_map": [ + 0.010998040437698364, + 0.010059562511742115, + 0.01374376192688942, + 0.009691743180155754, + 0.010224764235317707, + 0.01291098166257143, + 0.010404634289443493, + 0.01076358463615179, + 0.00807119719684124, + 0.004682225175201893, + 0.010544605553150177, + 0.005635251756757498, + 0.005288525950163603, + 0.045431241393089294, + 0.004624263383448124, + 0.041371915489435196, + 0.007392726372927427, + 0.0059717800468206406, + 0.011391758918762207, + 0.008196234703063965, + 0.006893841549754143, + 0.047026973217725754, + 0.006098109297454357, + 0.044522419571876526, + 0.002515592146664858, + 0.003387355711311102, + 0.003031060565263033, + 0.005005020648241043, + 0.005185552407056093, + 0.0036376218777149916, + 0.004337349906563759, + 0.004082798026502132, + 0.02123088389635086, + 0.009383464232087135, + 0.0314483679831028, + 0.00933046080172062, + 0.007846429012715816, + 0.15935403108596802, + 0.00779402581974864, + 0.1484631896018982, + 0.003808147506788373, + 0.003710881108418107, + 0.005366200115531683, + 0.004065846558660269, + 0.0029854336753487587, + 0.015658805146813393, + 0.0018824934959411621, + 0.015157015062868595, + 0.00262029143050313, + 0.001899617025628686, + 0.0036453153006732464, + 0.002282798755913973, + 0.0016510937130078673, + 0.01666872575879097, + 0.002883568173274398, + 0.015532570891082287, + 0.001780036254785955, + 0.0018127282382920384, + 0.003158775856718421, + 0.0020814097952097654, + 0.0017614245880395174, + 0.0017637900309637189, + 0.0015375767834484577, + 0.002448612591251731, + 0.00425996258854866, + 0.004784177523106337, + 0.006450226064771414, + 0.0024326774291694164, + 0.0029426750261336565, + 0.004478059709072113, + 0.002154272049665451, + 0.0038422923535108566, + 0.00933101400732994, + 0.00839949119836092, + 0.00985020212829113, + 0.008377361111342907, + 0.008549004793167114, + 0.008007160387933254, + 0.010993517935276031, + 0.011011416092514992 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 2, + 1 + ], + "selected_reward": -0.07896344363689423 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot framed as a static, eye-level medium shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a weary champion sitting perfectly centered on a worn wooden bench?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the character wearing dark gym shorts with white bandages on his hands that are partially untied and dangling?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a massive, gleaming golden trophy visible on the bench beside the character?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the environment a desolate, silent locker room with drab gray concrete walls and recessed metal lockers?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.023279106244444847 + ], + "opposite_logprob_tokens": [ + -3.7732791900634766 + ], + "logprob_reward": -0.023279106244444847, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.023279106244444847, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2552046477794647, + 0.21804460883140564, + 0.04866576939821243, + 0.041844893246889114, + 0.1987999528646469, + 0.08687203377485275, + 0.016450660303235054, + 0.02931491658091545, + 0.014466673135757446, + 0.09033582359552383 + ], + "credit_map": [ + 0.02055414766073227, + 0.021086789667606354, + 0.030839309096336365, + 0.013307786546647549, + 0.011654924601316452, + 0.0774741992354393, + 0.009090145118534565, + 0.07119735330343246, + 0.011951319873332977, + 0.006205547600984573, + 0.017886830493807793, + 0.0049360585398972034, + 0.004124040715396404, + 0.08685742318630219, + 0.004029459785670042, + 0.08205391466617584, + 0.0027141545433551073, + 0.0024656844325363636, + 0.0036866681184619665, + 0.0018901106668636203, + 0.0014841846423223615, + 0.018021544441580772, + 0.001981703331694007, + 0.016421718522906303, + 0.0036606667563319206, + 0.005303259938955307, + 0.006042973604053259, + 0.002797611290588975, + 0.002311521442607045, + 0.009516860358417034, + 0.002620236249640584, + 0.009591765701770782, + 0.011202758178114891, + 0.004552270285785198, + 0.015468456782400608, + 0.004095601849257946, + 0.0035025491379201412, + 0.08037576824426651, + 0.00441261800006032, + 0.07518992573022842, + 0.0060968357138335705, + 0.003392336657270789, + 0.008082653395831585, + 0.002905820496380329, + 0.001729333307594061, + 0.03195531666278839, + 0.002184703480452299, + 0.030525028705596924, + 0.002292047254741192, + 0.002001643180847168, + 0.002609104383736849, + 0.0018981730099767447, + 0.0015717995120212436, + 0.0025865647476166487, + 0.00116624403744936, + 0.002325082430616021, + 0.0030014035291969776, + 0.004475264810025692, + 0.0029507402796298265, + 0.002226191107183695, + 0.0015848545590415597, + 0.0065175630152225494, + 0.0018220401834696531, + 0.0067368606105446815, + 0.0017359431367367506, + 0.0017037284560501575, + 0.002299885032698512, + 0.001395014813169837, + 0.0016154780751094222, + 0.0018421690911054611, + 0.0013609271263703704, + 0.0025135278701782227, + 0.01247852947562933, + 0.00965434405952692, + 0.01241099089384079, + 0.010488037019968033, + 0.0064949821680784225, + 0.011665408499538898, + 0.011662717908620834, + 0.015480812638998032 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 4 + ], + "selected_reward": -0.023279106244444847 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "static, eye-level medium shot, framed with documentary-style realism, captures a weary champion sitting perfectly centered on a worn wooden bench in a desolate, silent locker room. His sweat-dampened black hair clings to his forehead, his muscular shoulders slump with exhaustion, and a simple white towel hangs from his neck. He wears only a pair of dark gym shorts, and the white bandages on his hands are partially untied, dangling loosely. His face, etched with the physical toll of his victory, is a mask of exhaustion; his gaze is unfocused and glassy, staring past the camera into an unseen distance. Beside him, a massive, gleaming golden trophy sits on the bench, close yet pointedly overlooked, its polished surface catching the cold, sterile light from a single overhead fluorescent bulb, casting a bright, almost mocking reflection against the drab, gray concrete walls and the recessed metal lockers that line the room. He took a slow, deep breath, barely lifting his chest, and then let out a deliberately heavy, audible sigh, as if draining away the last of his energy, his body slumping further. The cheers of victory were a forgotten echo, replaced by the deep, hollow emptiness of the silent room and the grim reality of victory's consequences.", + "seed": 12490, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000095/group_3/index_6580_seed_12490_dp_7_rank_14_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 3, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.03058710414916277, + "vlm_logprob": -0.03058710414916277, + "vlm_yes_no_margin": 4.0, + "vlm_token_credit": -0.03058710414916277, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot framed as a static, eye-level medium shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a weary champion sitting perfectly centered on a worn wooden bench?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the character wearing dark gym shorts with white bandages on his hands that are partially untied and dangling?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a massive, gleaming golden trophy visible on the bench beside the character?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the environment a desolate, silent locker room with drab gray concrete walls and recessed metal lockers?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.06124386936426163 + ], + "opposite_logprob_tokens": [ + -3.811243772506714 + ], + "logprob_reward": -0.06124386936426163, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.06124386936426163, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.18832768499851227, + 0.07111268490552902, + 0.21155057847499847, + 0.0651237964630127, + 0.020154403522610664, + 0.07464046776294708, + 0.096769779920578, + 0.03830041363835335, + 0.10333894938230515, + 0.13068129122257233 + ], + "credit_map": [ + 0.022000392898917198, + 0.018733937293291092, + 0.030493618920445442, + 0.027381254360079765, + 0.01713932305574417, + 0.03048061951994896, + 0.02223358489573002, + 0.019864948466420174, + 0.008011342957615852, + 0.010494689457118511, + 0.010072192177176476, + 0.006672328338027, + 0.007971996441483498, + 0.010345849208533764, + 0.0061397976242005825, + 0.01140447799116373, + 0.02591096982359886, + 0.025550905615091324, + 0.030410435050725937, + 0.037588510662317276, + 0.013435613363981247, + 0.033100467175245285, + 0.012338954024016857, + 0.033214692026376724, + 0.00569102494046092, + 0.007763361092656851, + 0.010077623650431633, + 0.008696086704730988, + 0.0042559914290905, + 0.012434984557330608, + 0.0035433322191238403, + 0.012661389075219631, + 0.0024402206763625145, + 0.0025030472315847874, + 0.0037330258637666702, + 0.002241274807602167, + 0.0017413736786693335, + 0.0027751715388149023, + 0.00142998609226197, + 0.003290304448455572, + 0.008302144706249237, + 0.012962245382368565, + 0.009581961669027805, + 0.014145786873996258, + 0.0057588182389736176, + 0.009693549014627934, + 0.004410034045577049, + 0.009785926900804043, + 0.009709259495139122, + 0.015747467055916786, + 0.011915051378309727, + 0.020327232778072357, + 0.005661974661052227, + 0.013723409734666348, + 0.007093631662428379, + 0.012591740116477013, + 0.00416194973513484, + 0.0027345179114490747, + 0.004959553480148315, + 0.003764472668990493, + 0.003310779109597206, + 0.008725201711058617, + 0.0022255140356719494, + 0.008418423123657703, + 0.019834140315651894, + 0.010888176038861275, + 0.015825198963284492, + 0.011217471212148666, + 0.007855742238461971, + 0.020047735422849655, + 0.007366744335740805, + 0.010303733870387077, + 0.011033536866307259, + 0.009057128801941872, + 0.015076261945068836, + 0.016826167702674866, + 0.010996022261679173, + 0.036177244037389755, + 0.007493348326534033, + 0.024021577090024948 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 0, + 9 + ], + "selected_reward": -0.06124386936426163 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot framed as a static, eye-level medium shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a weary champion sitting perfectly centered on a worn wooden bench?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the character wearing dark gym shorts with white bandages on his hands that are partially untied and dangling?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a massive, gleaming golden trophy visible on the bench beside the character?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the environment a desolate, silent locker room with drab gray concrete walls and recessed metal lockers?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.014214457012712955 + ], + "opposite_logprob_tokens": [ + -4.264214515686035 + ], + "logprob_reward": -0.014214457012712955, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.014214457012712955, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3329044282436371, + 0.08980939537286758, + 0.14287139475345612, + 0.030167974531650543, + 0.03415948525071144, + 0.02563849464058876, + 0.06834477931261063, + 0.05351565033197403, + 0.10747160017490387, + 0.11511684954166412 + ], + "credit_map": [ + 0.023843497037887573, + 0.02639906294643879, + 0.050014298409223557, + 0.03984716907143593, + 0.022915860638022423, + 0.04307319596409798, + 0.029121357947587967, + 0.09768997877836227, + 0.009632408618927002, + 0.012036769650876522, + 0.012066073715686798, + 0.007491240277886391, + 0.01399227511137724, + 0.011317126452922821, + 0.008086350746452808, + 0.015187140554189682, + 0.01541542261838913, + 0.015765463933348656, + 0.016384702175855637, + 0.02146073244512081, + 0.0205556508153677, + 0.022152774035930634, + 0.012745638377964497, + 0.018390996381640434, + 0.003356361761689186, + 0.0021746798884123564, + 0.005051514599472284, + 0.004210385028272867, + 0.002474436769261956, + 0.004842066206037998, + 0.0025332390796393156, + 0.005525288172066212, + 0.0036583219189196825, + 0.005162590648978949, + 0.0046392641961574554, + 0.005634792614728212, + 0.0022173887118697166, + 0.005894398782402277, + 0.0024207148235291243, + 0.0045320079661905766, + 0.0023308780509978533, + 0.0037024624180048704, + 0.0034993914887309074, + 0.0022326672915369272, + 0.0034114622976630926, + 0.0054628243669867516, + 0.0016536241164430976, + 0.003345184726640582, + 0.006773428060114384, + 0.010007050819694996, + 0.008346175774931908, + 0.01229850947856903, + 0.004045852925628424, + 0.011765111237764359, + 0.004313229117542505, + 0.01079540979117155, + 0.004248268436640501, + 0.003921785857528448, + 0.006506525445729494, + 0.0041620382107794285, + 0.003854291746392846, + 0.01614486612379551, + 0.0026072529144585133, + 0.012070623226463795, + 0.01867160201072693, + 0.011159182526171207, + 0.01493238378316164, + 0.01666409708559513, + 0.008479470387101173, + 0.01328811515122652, + 0.007940851151943207, + 0.01633589342236519, + 0.008159181103110313, + 0.005205079913139343, + 0.014904201030731201, + 0.004146168474107981, + 0.007470432668924332, + 0.04524671658873558, + 0.005848242435604334, + 0.024136817082762718 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 9 + ], + "selected_reward": -0.014214457012712955 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot framed as a static, eye-level medium shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a weary champion sitting perfectly centered on a worn wooden bench?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the character wearing dark gym shorts with white bandages on his hands that are partially untied and dangling?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a massive, gleaming golden trophy visible on the bench beside the character?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the environment a desolate, silent locker room with drab gray concrete walls and recessed metal lockers?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.029799941927194595 + ], + "opposite_logprob_tokens": [ + -3.5297999382019043 + ], + "logprob_reward": -0.029799941927194595, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.029799941927194595, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2478356808423996, + 0.13621704280376434, + 0.10734602808952332, + 0.061674512922763824, + 0.05739269405603409, + 0.01772226020693779, + 0.0350237712264061, + 0.029199589043855667, + 0.1993376612663269, + 0.10825075209140778 + ], + "credit_map": [ + 0.02383445017039776, + 0.015292029827833176, + 0.0369088388979435, + 0.026187637820839882, + 0.016738153994083405, + 0.032849084585905075, + 0.023221097886562347, + 0.07280442863702774, + 0.01552026066929102, + 0.025935376062989235, + 0.017159119248390198, + 0.02354617230594158, + 0.012707564979791641, + 0.017029447481036186, + 0.007844969630241394, + 0.01647414080798626, + 0.012998771853744984, + 0.009913215413689613, + 0.012395822443068027, + 0.013799689710140228, + 0.01752016507089138, + 0.01511434093117714, + 0.011573259718716145, + 0.014030775055289268, + 0.007028622552752495, + 0.010821078903973103, + 0.008880079723894596, + 0.010528172366321087, + 0.005431579891592264, + 0.007888847962021828, + 0.0037553401198238134, + 0.007340799551457167, + 0.006382407620549202, + 0.00911672506481409, + 0.007172533310949802, + 0.007372171152383089, + 0.008768963627517223, + 0.005852172616869211, + 0.0057500568218529224, + 0.006977668032050133, + 0.0019589432049542665, + 0.0021719029173254967, + 0.0022870602551847696, + 0.0025479125324636698, + 0.0022564518731087446, + 0.002029760042205453, + 0.0016238122479990125, + 0.002846420044079423, + 0.003517010249197483, + 0.004438110161572695, + 0.003937034867703915, + 0.006371175870299339, + 0.0024837872479110956, + 0.006740738172084093, + 0.00210001808591187, + 0.005435900297015905, + 0.00300157954916358, + 0.0024989338126033545, + 0.004310315474867821, + 0.0035631589125841856, + 0.0023310473188757896, + 0.006629165727645159, + 0.0014254495035856962, + 0.005439943168312311, + 0.031322646886110306, + 0.01502138376235962, + 0.022984135895967484, + 0.024756070226430893, + 0.011371348984539509, + 0.04397939145565033, + 0.010148877277970314, + 0.03975382819771767, + 0.007648097351193428, + 0.008016719482839108, + 0.012886219657957554, + 0.008028118871152401, + 0.007953902706503868, + 0.03825035318732262, + 0.0057394797913730145, + 0.01972787268459797 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 1 + ], + "selected_reward": -0.029799941927194595 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot framed as a static, eye-level medium shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a weary champion sitting perfectly centered on a worn wooden bench?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the character wearing dark gym shorts with white bandages on his hands that are partially untied and dangling?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a massive, gleaming golden trophy visible on the bench beside the character?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the environment a desolate, silent locker room with drab gray concrete walls and recessed metal lockers?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004632692784070969 + ], + "opposite_logprob_tokens": [ + -5.379632472991943 + ], + "logprob_reward": -0.004632692784070969, + "yes_no_margin_reward": 5.375, + "token_credit_reward": -0.004632692784070969, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06134946644306183, + 0.1617693305015564, + 0.20120278000831604, + 0.14680935442447662, + 0.13209286332130432, + 0.04639497771859169, + 0.03909587487578392, + 0.028746232390403748, + 0.09562498331069946, + 0.08691418170928955 + ], + "credit_map": [ + 0.009179819375276566, + 0.005560245364904404, + 0.008552201092243195, + 0.010971986688673496, + 0.005353373009711504, + 0.006587428506463766, + 0.006960057653486729, + 0.008184350095689297, + 0.018712054938077927, + 0.03475329652428627, + 0.01973257213830948, + 0.0192888081073761, + 0.008967104367911816, + 0.027552489191293716, + 0.007842138409614563, + 0.024920864030718803, + 0.020109133794903755, + 0.03968501091003418, + 0.02539404109120369, + 0.032020699232816696, + 0.010464989580214024, + 0.030461713671684265, + 0.009784272871911526, + 0.033282894641160965, + 0.013491819612681866, + 0.025458242744207382, + 0.019121699035167694, + 0.02646666020154953, + 0.008877359330654144, + 0.023167241364717484, + 0.00942233856767416, + 0.02080398239195347, + 0.013144210912287235, + 0.023153314366936684, + 0.01661568135023117, + 0.02170908823609352, + 0.0070569757372140884, + 0.023132527247071266, + 0.007098380476236343, + 0.02018267661333084, + 0.004571502096951008, + 0.008449492044746876, + 0.005365108605474234, + 0.007331231143325567, + 0.003207354573532939, + 0.008913214318454266, + 0.0023616482503712177, + 0.00619542459025979, + 0.004180882126092911, + 0.00598974758759141, + 0.004865714348852634, + 0.007570915389806032, + 0.002156161703169346, + 0.00665272818878293, + 0.0021904606837779284, + 0.005489263217896223, + 0.003247298067435622, + 0.00399353913962841, + 0.004089301452040672, + 0.0048888144083321095, + 0.001999896252527833, + 0.0041427006945014, + 0.0016328422352671623, + 0.004751836881041527, + 0.015358499251306057, + 0.00890237744897604, + 0.011327502317726612, + 0.014402887783944607, + 0.005637435708194971, + 0.019942034035921097, + 0.005305899772793055, + 0.014748343266546726, + 0.006620313040912151, + 0.00951335858553648, + 0.010616138577461243, + 0.013935216702520847, + 0.005629895254969597, + 0.02304938994348049, + 0.004951999988406897, + 0.012597862631082535 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 1, + 3 + ], + "selected_reward": -0.004632692784070969 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot framed as a static, eye-level medium shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a weary champion sitting perfectly centered on a worn wooden bench?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the character wearing dark gym shorts with white bandages on his hands that are partially untied and dangling?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a massive, gleaming golden trophy visible on the bench beside the character?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the environment a desolate, silent locker room with drab gray concrete walls and recessed metal lockers?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0430445596575737 + ], + "opposite_logprob_tokens": [ + -3.1680445671081543 + ], + "logprob_reward": -0.0430445596575737, + "yes_no_margin_reward": 3.125, + "token_credit_reward": -0.0430445596575737, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14426001906394958, + 0.13311652839183807, + 0.06485146284103394, + 0.0471380390226841, + 0.023590298369526863, + 0.029458776116371155, + 0.1437341570854187, + 0.052394215017557144, + 0.29777345061302185, + 0.06368298083543777 + ], + "credit_map": [ + 0.014444885775446892, + 0.010946329683065414, + 0.021336600184440613, + 0.015087291598320007, + 0.015492227859795094, + 0.015629254281520844, + 0.015601839870214462, + 0.03572160378098488, + 0.015374402515590191, + 0.0280084777623415, + 0.016757439821958542, + 0.01707744039595127, + 0.008458749391138554, + 0.02149355411529541, + 0.006378757767379284, + 0.01956772804260254, + 0.00849762000143528, + 0.013870729133486748, + 0.007725530304014683, + 0.006960064172744751, + 0.0046753054484725, + 0.009160713292658329, + 0.002684863982722163, + 0.011276647448539734, + 0.0045015872456133366, + 0.009009821340441704, + 0.006695540156215429, + 0.007230781018733978, + 0.0024534682743251324, + 0.008029364049434662, + 0.002753615379333496, + 0.006463867612183094, + 0.0026639362331479788, + 0.003882360877469182, + 0.0035549167077988386, + 0.0028940525371581316, + 0.001673223334364593, + 0.003815568285062909, + 0.0018285319674760103, + 0.0032777125015854836, + 0.0028572380542755127, + 0.005193673074245453, + 0.003609986510127783, + 0.0038903618697077036, + 0.0020666210912168026, + 0.006042586173862219, + 0.0014406733680516481, + 0.00435763830319047, + 0.01369831059128046, + 0.022062109783291817, + 0.017446929588913918, + 0.023957455530762672, + 0.00787980668246746, + 0.02755737118422985, + 0.007286980282515287, + 0.023845214396715164, + 0.005852542817592621, + 0.007703025359660387, + 0.006816906854510307, + 0.009199881926178932, + 0.0028829225338995457, + 0.009616031311452389, + 0.002830652054399252, + 0.007492257282137871, + 0.05076175928115845, + 0.021197644993662834, + 0.037898845970630646, + 0.035342615097761154, + 0.016598209738731384, + 0.06004386022686958, + 0.013145063072443008, + 0.06278549134731293, + 0.0044077131897211075, + 0.003530453657731414, + 0.008368594571948051, + 0.003921421244740486, + 0.004646113608032465, + 0.024045757949352264, + 0.003322548931464553, + 0.011440386064350605 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 0, + 6 + ], + "selected_reward": -0.0430445596575737 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000095/group_4/index_tensor([7168, 7168])_group_4_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000095/group_4/index_tensor([7168, 7168])_group_4_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..971a7d1117ef73174bfeb244494712940f4a2c58 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000095/group_4/index_tensor([7168, 7168])_group_4_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "static, eye-level medium shot captures a scene rendered entirely in the style of traditional Chinese ink painting. A lone, elegant shovel, its long handle depicted in a confident stroke of dark ink, stands amidst dark, fertile soil, layered with rich, moist grays and blacks. Nearby, a handheld sprayer, rendered with minimalist, economical brushstrokes, emits a fine mist of fertilizer. The atomized particles are captured as soft, diffuse clouds of light gray ink, gently drifting downward. These particles settle onto the dark ground like fine dust, subtly punctuated by rich washes of ink. The entire monochrome composition, executed with bold, expressive brushstrokes on a surface with the visible fibrous texture of rice paper, evokes a tranquil and poetic atmosphere, before the entire image pauses for a moment and smoothly fades to black.", + "seed": 12397, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000095/group_4/index_7168_seed_12397_dp_8_rank_16_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 4, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.431781655177474, + "vlm_logprob": -0.431781655177474, + "vlm_yes_no_margin": 1.575, + "vlm_token_credit": -0.431781655177474, + "vlm_yes_ratio": 0.8, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the visual style of the video a traditional Chinese ink painting?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a lone shovel with a long handle visible standing in dark, fertile soil?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a handheld sprayer emitting a fine mist of fertilizer visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the video display the visible fibrous texture of rice paper as the surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video smoothly fade to black at the end?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.027157703414559364 + ], + "opposite_logprob_tokens": [ + -3.652157783508301 + ], + "logprob_reward": -0.027157703414559364, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.027157703414559364, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.04378725588321686, + 0.016403747722506523, + 0.02598731964826584, + 0.01606127619743347, + 0.036122675985097885, + 0.0620785616338253, + 0.03455819562077522, + 0.05970897525548935, + 0.0850682258605957, + 0.6202237010002136 + ], + "credit_map": [ + 0.008519522845745087, + 0.004856074694544077, + 0.004949152935296297, + 0.0034648324362933636, + 0.005506305489689112, + 0.0053403740748763084, + 0.005882977042347193, + 0.0052680219523608685, + 0.002633910160511732, + 0.002185243181884289, + 0.002142446581274271, + 0.001444864203222096, + 0.0016915054293349385, + 0.0017137936083599925, + 0.0022002137266099453, + 0.0023917732760310173, + 0.006254319101572037, + 0.0029845028184354305, + 0.0023837201297283173, + 0.0019554400350898504, + 0.005330664571374655, + 0.0027898179832845926, + 0.002414795570075512, + 0.0018740625819191337, + 0.0037255045026540756, + 0.002845459384843707, + 0.0015446534380316734, + 0.0011654305271804333, + 0.0019052480347454548, + 0.0017626459011808038, + 0.0017015434568747878, + 0.0014107926981523633, + 0.005009263288229704, + 0.00741773983463645, + 0.003803143510594964, + 0.00314669543877244, + 0.005119813606142998, + 0.0043728528544306755, + 0.0040934826247394085, + 0.0031596876215189695, + 0.009735764935612679, + 0.00573647441342473, + 0.009523145854473114, + 0.005087897647172213, + 0.007894397713243961, + 0.012844408862292767, + 0.004172792658209801, + 0.007083685602992773, + 0.0026886772830039263, + 0.001605618977919221, + 0.0019045034423470497, + 0.0021900718566030264, + 0.005940453615039587, + 0.012317424640059471, + 0.0017843479290604591, + 0.006127099506556988, + 0.0036972907837480307, + 0.002075752941891551, + 0.002853711601346731, + 0.0032307254150509834, + 0.010769887827336788, + 0.02260270155966282, + 0.0029955869540572166, + 0.011483321897685528, + 0.005941222887486219, + 0.005173644982278347, + 0.005814831238240004, + 0.006416319403797388, + 0.01227646041661501, + 0.028591671958565712, + 0.005649844650179148, + 0.015204239636659622, + 0.02873127907514572, + 0.014555552043020725, + 0.010775720700621605, + 0.023481162264943123, + 0.11589270830154419, + 0.25815117359161377, + 0.029902640730142593, + 0.13873355090618134 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 8, + 5 + ], + "selected_reward": -0.027157703414559364 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the visual style of the video a traditional Chinese ink painting?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a lone shovel with a long handle visible standing in dark, fertile soil?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a handheld sprayer emitting a fine mist of fertilizer visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the video display the visible fibrous texture of rice paper as the surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video smoothly fade to black at the end?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.42893674969673157 + ], + "opposite_logprob_tokens": [ + -1.0539367198944092 + ], + "logprob_reward": -0.42893674969673157, + "yes_no_margin_reward": 0.625, + "token_credit_reward": -0.42893674969673157, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.016936175525188446, + 0.022479187697172165, + 0.01851520501077175, + 0.013514572754502296, + 0.008181886747479439, + 0.02065272442996502, + 0.013132325373589993, + 0.011239847168326378, + 0.0975915938615799, + 0.7777565717697144 + ], + "credit_map": [ + 0.0030635667499154806, + 0.001741061802022159, + 0.0016931951977312565, + 0.002283907262608409, + 0.002111215377226472, + 0.0017222381429746747, + 0.0023303108755499125, + 0.0019906796514987946, + 0.0035159196704626083, + 0.0029945173300802708, + 0.0026907813735306263, + 0.0022006016224622726, + 0.0026268279179930687, + 0.002359955571591854, + 0.00365883088670671, + 0.002431750763207674, + 0.0035820791963487864, + 0.0023686436470597982, + 0.0019653874915093184, + 0.00160756625700742, + 0.003055982291698456, + 0.001899340539239347, + 0.0023364245425909758, + 0.0016997806960716844, + 0.0019749721977859735, + 0.0018081868765875697, + 0.0017801446374505758, + 0.0010478603653609753, + 0.0020664671901613474, + 0.0012954457197338343, + 0.0020997615065425634, + 0.0014417337952181697, + 0.0014326469972729683, + 0.0015374940121546388, + 0.0010106121189892292, + 0.0009500118903815746, + 0.0008889273158274591, + 0.0009115344728343189, + 0.000745729892514646, + 0.0007049298146739602, + 0.0024927412159740925, + 0.0015344449784606695, + 0.0023046154528856277, + 0.0011429772712290287, + 0.0030171910766512156, + 0.005730161909013987, + 0.0017849985742941499, + 0.0026455933693796396, + 0.0014757090248167515, + 0.0011920492397621274, + 0.0007648394675925374, + 0.0008048902382142842, + 0.0019004008499905467, + 0.0038378688041120768, + 0.0010076616890728474, + 0.0021489060018211603, + 0.0008871659520082176, + 0.0007610234897583723, + 0.0008210329688154161, + 0.0006732226465828717, + 0.001875076675787568, + 0.0037486967630684376, + 0.0005900801043026149, + 0.0018835477530956268, + 0.004387244116514921, + 0.003840019693598151, + 0.001840517041273415, + 0.003812333568930626, + 0.018317362293601036, + 0.040772538632154465, + 0.004696293734014034, + 0.019925285130739212, + 0.035242579877376556, + 0.016124218702316284, + 0.012691101059317589, + 0.028964325785636902, + 0.14699944853782654, + 0.3248107135295868, + 0.03762543573975563, + 0.17529870569705963 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 8, + 1 + ], + "selected_reward": -0.42893674969673157 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the visual style of the video a traditional Chinese ink painting?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a lone shovel with a long handle visible standing in dark, fertile soil?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a handheld sprayer emitting a fine mist of fertilizer visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the video display the visible fibrous texture of rice paper as the surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video smoothly fade to black at the end?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.6328355073928833 + ], + "opposite_logprob_tokens": [ + -0.7578355073928833 + ], + "logprob_reward": -0.6328355073928833, + "yes_no_margin_reward": 0.125, + "token_credit_reward": -0.6328355073928833, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.040982943028211594, + 0.023046066984534264, + 0.04126344993710518, + 0.008094321005046368, + 0.011064660735428333, + 0.02214665710926056, + 0.024288125336170197, + 0.01334321964532137, + 0.07789851725101471, + 0.7378720045089722 + ], + "credit_map": [ + 0.006171666085720062, + 0.0041028037667274475, + 0.005572804715484381, + 0.0031128849368542433, + 0.005042008589953184, + 0.004774306435137987, + 0.0067229545675218105, + 0.005483516491949558, + 0.004201303236186504, + 0.0027874656952917576, + 0.0026649879291653633, + 0.002947060624137521, + 0.0027858922258019447, + 0.0024404090363532305, + 0.002906965324655175, + 0.0023119840770959854, + 0.008590102195739746, + 0.005584524013102055, + 0.003759479383006692, + 0.003104840638116002, + 0.008253715932369232, + 0.005128540564328432, + 0.0037672079633921385, + 0.0030750425066798925, + 0.000931781716644764, + 0.0008620872395113111, + 0.0011746196541935205, + 0.0008700860780663788, + 0.0009398082038387656, + 0.0008375914767384529, + 0.0011977007379755378, + 0.0012806468876078725, + 0.0016756233526393771, + 0.0016960111679509282, + 0.00123180344235152, + 0.0010277662659063935, + 0.001321738469414413, + 0.0013478752225637436, + 0.0017401180230081081, + 0.0010237247915938497, + 0.004127588123083115, + 0.004099570214748383, + 0.0020477960351854563, + 0.001862160163000226, + 0.0031812640372663736, + 0.0028106567915529013, + 0.0017975986702367663, + 0.002220024587586522, + 0.0022132338490337133, + 0.0018125750357285142, + 0.00213347258977592, + 0.0013898659963160753, + 0.003570695873349905, + 0.007343629840761423, + 0.0016298431437462568, + 0.004194809123873711, + 0.0012654298916459084, + 0.0015983612975105643, + 0.001403374713845551, + 0.0021986551582813263, + 0.001891792519018054, + 0.0020899369847029448, + 0.0012803500285372138, + 0.001615319517441094, + 0.004990552086383104, + 0.0031764961313456297, + 0.004363123327493668, + 0.00472998945042491, + 0.013566301204264164, + 0.029218358919024467, + 0.0038996629882603884, + 0.013954036869108677, + 0.033614177256822586, + 0.015084288083016872, + 0.014909274876117706, + 0.02793491631746292, + 0.13924039900302887, + 0.30616870522499084, + 0.035782404243946075, + 0.1651378870010376 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 8, + 2 + ], + "selected_reward": -0.6328355073928833 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the visual style of the video a traditional Chinese ink painting?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a lone shovel with a long handle visible standing in dark, fertile soil?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a handheld sprayer emitting a fine mist of fertilizer visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the video display the visible fibrous texture of rice paper as the surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video smoothly fade to black at the end?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.016087371855974197 + ], + "opposite_logprob_tokens": [ + -4.141087532043457 + ], + "logprob_reward": -0.016087371855974197, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.016087371855974197, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.05580919608473778, + 0.013730470091104507, + 0.019875280559062958, + 0.017930276691913605, + 0.040787141770124435, + 0.03823136165738106, + 0.053610850125551224, + 0.05137234926223755, + 0.1501637101173401, + 0.5584893226623535 + ], + "credit_map": [ + 0.007688979618251324, + 0.006457096431404352, + 0.00881946086883545, + 0.004606274887919426, + 0.0076225255616009235, + 0.006354567129164934, + 0.007494472432881594, + 0.006765818689018488, + 0.00196562591008842, + 0.001663232920691371, + 0.002222208073362708, + 0.001135833328589797, + 0.0024653247091919184, + 0.0015508937649428844, + 0.0012840876588597894, + 0.0014432641910389066, + 0.0050779664888978004, + 0.002502109855413437, + 0.0017485799035057425, + 0.0014621438458561897, + 0.003962964750826359, + 0.0019511629361659288, + 0.0015864509623497725, + 0.0015839015832170844, + 0.0026841608341783285, + 0.002449468243867159, + 0.0019072035793215036, + 0.00203215004876256, + 0.002119221491739154, + 0.0017170771025121212, + 0.0024257872719317675, + 0.002595209050923586, + 0.0072968490421772, + 0.007250273134559393, + 0.0038827573880553246, + 0.004315060563385487, + 0.0052782283164560795, + 0.004579366184771061, + 0.0041879890486598015, + 0.003996618092060089, + 0.005384874064475298, + 0.0029397679027169943, + 0.006100369617342949, + 0.0022564639803022146, + 0.005053279455751181, + 0.008368450216948986, + 0.0029067457653582096, + 0.0052214134484529495, + 0.004378429614007473, + 0.00317259319126606, + 0.005166173446923494, + 0.004046900197863579, + 0.00788438506424427, + 0.016605421900749207, + 0.0032902925740927458, + 0.009066655300557613, + 0.004003375303000212, + 0.002415119204670191, + 0.004573040176182985, + 0.003954860381782055, + 0.00794101133942604, + 0.016756273806095123, + 0.0029375513549894094, + 0.008791117928922176, + 0.0069564757868647575, + 0.005896926391869783, + 0.0031782181467860937, + 0.00595482112839818, + 0.027823740616440773, + 0.06236489489674568, + 0.00783409085124731, + 0.03015454113483429, + 0.02568475343286991, + 0.01315320935100317, + 0.010325855575501919, + 0.020893214270472527, + 0.1044999212026596, + 0.23251041769981384, + 0.026665328070521355, + 0.12475664913654327 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 8, + 0 + ], + "selected_reward": -0.016087371855974197 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the visual style of the video a traditional Chinese ink painting?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a lone shovel with a long handle visible standing in dark, fertile soil?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a handheld sprayer emitting a fine mist of fertilizer visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the video display the visible fibrous texture of rice paper as the surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video smoothly fade to black at the end?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -1.0538909435272217 + ], + "opposite_logprob_tokens": [ + -0.4288909435272217 + ], + "logprob_reward": -1.0538909435272217, + "yes_no_margin_reward": -0.625, + "token_credit_reward": -1.0538909435272217, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.026631705462932587, + 0.008666476234793663, + 0.012744011357426643, + 0.017072005197405815, + 0.04224167391657829, + 0.1569960117340088, + 0.029253734275698662, + 0.10713427513837814, + 0.3705109655857086, + 0.228749081492424 + ], + "credit_map": [ + 0.004322898108512163, + 0.0021075529512017965, + 0.004456567112356424, + 0.001967106480151415, + 0.0023120585829019547, + 0.003078087465837598, + 0.005577738396823406, + 0.0028096972964704037, + 0.0013184379786252975, + 0.0009128791280090809, + 0.0009604254155419767, + 0.001370005775243044, + 0.0009077220456674695, + 0.000907067849766463, + 0.0011532586067914963, + 0.0011366800172254443, + 0.002197983441874385, + 0.0011159186251461506, + 0.0017556804232299328, + 0.0010170835303142667, + 0.0020878666546195745, + 0.0013340231962502003, + 0.0017564588924869895, + 0.0014789981069043279, + 0.004412518348544836, + 0.001795235788449645, + 0.0015702091623097658, + 0.0014445842243731022, + 0.002046407200396061, + 0.0022929462138563395, + 0.0014822608791291714, + 0.002027844777330756, + 0.005877021700143814, + 0.005292294546961784, + 0.00464932806789875, + 0.002360529964789748, + 0.007002962753176689, + 0.009106222540140152, + 0.0037629075814038515, + 0.00419040909036994, + 0.012954559177160263, + 0.01561396662145853, + 0.023326439782977104, + 0.011994268745183945, + 0.022657349705696106, + 0.034664351493120193, + 0.014563788659870625, + 0.021221298724412918, + 0.0025960055645555258, + 0.006760902237147093, + 0.0026887203566730022, + 0.0037425460759550333, + 0.004177021328359842, + 0.0029514373745769262, + 0.003020609961822629, + 0.0033164953347295523, + 0.0049943216145038605, + 0.00286219734698534, + 0.002394525334239006, + 0.004167197737842798, + 0.020673885941505432, + 0.043778739869594574, + 0.00513156782835722, + 0.0231318436563015, + 0.016622833907604218, + 0.015820063650608063, + 0.007492119446396828, + 0.014417947269976139, + 0.06952860206365585, + 0.15357902646064758, + 0.017732633277773857, + 0.07531779259443283, + 0.010292631573975086, + 0.0049357302486896515, + 0.003818884026259184, + 0.008556822314858437, + 0.04330794885754585, + 0.09542656689882278, + 0.01126533467322588, + 0.051145195960998535 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 9, + 5 + ], + "selected_reward": -1.0538909435272217 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "No" + ] + }, + { + "prompt": "static, eye-level medium shot captures a scene rendered entirely in the style of traditional Chinese ink painting. A lone, elegant shovel, its long handle depicted in a confident stroke of dark ink, stands amidst dark, fertile soil, layered with rich, moist grays and blacks. Nearby, a handheld sprayer, rendered with minimalist, economical brushstrokes, emits a fine mist of fertilizer. The atomized particles are captured as soft, diffuse clouds of light gray ink, gently drifting downward. These particles settle onto the dark ground like fine dust, subtly punctuated by rich washes of ink. The entire monochrome composition, executed with bold, expressive brushstrokes on a surface with the visible fibrous texture of rice paper, evokes a tranquil and poetic atmosphere, before the entire image pauses for a moment and smoothly fades to black.", + "seed": 12398, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000095/group_4/index_7168_seed_12398_dp_8_rank_16_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 4, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.4369648888707161, + "vlm_logprob": -0.4369648888707161, + "vlm_yes_no_margin": 1.775, + "vlm_token_credit": -0.4369648888707161, + "vlm_yes_ratio": 0.6, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the visual style of the video a traditional Chinese ink painting?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a lone shovel with a long handle visible standing in dark, fertile soil?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a handheld sprayer emitting a fine mist of fertilizer visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the video display the visible fibrous texture of rice paper as the surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video smoothly fade to black at the end?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.027687406167387962 + ], + "opposite_logprob_tokens": [ + -3.6526873111724854 + ], + "logprob_reward": -0.027687406167387962, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.027687406167387962, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12553608417510986, + 0.09892810881137848, + 0.056159608066082, + 0.038534168154001236, + 0.1377938985824585, + 0.31627652049064636, + 0.11972052603960037, + 0.03441033139824867, + 0.04121094197034836, + 0.031429722905159 + ], + "credit_map": [ + 0.016799088567495346, + 0.017719736322760582, + 0.01631885953247547, + 0.012678459286689758, + 0.01666240207850933, + 0.019354436546564102, + 0.011498337611556053, + 0.014504766091704369, + 0.017390325665473938, + 0.017436223104596138, + 0.012586052529513836, + 0.011696078814566135, + 0.01069278921931982, + 0.010598397813737392, + 0.007971163839101791, + 0.010557078756392002, + 0.007609752472490072, + 0.0075815836898982525, + 0.008462829515337944, + 0.009042641147971153, + 0.005220657214522362, + 0.0062401858158409595, + 0.006077710073441267, + 0.0059242467395961285, + 0.004762065596878529, + 0.006073599215596914, + 0.004310724325478077, + 0.002733580768108368, + 0.005706499796360731, + 0.00801004283130169, + 0.0030346803832799196, + 0.0039029745385050774, + 0.019347496330738068, + 0.03430274501442909, + 0.022737959399819374, + 0.007222163025289774, + 0.013405687175691128, + 0.02147611603140831, + 0.006228756159543991, + 0.013072975911200047, + 0.02824314683675766, + 0.08981312811374664, + 0.042489249259233475, + 0.009339793585240841, + 0.022544141858816147, + 0.09981026500463486, + 0.008944815024733543, + 0.015091988258063793, + 0.021733013913035393, + 0.01784263551235199, + 0.02080089971423149, + 0.006547222379595041, + 0.013961653225123882, + 0.019101455807685852, + 0.006797081790864468, + 0.012936561368405819, + 0.004890910349786282, + 0.0052959369495511055, + 0.004481709096580744, + 0.0018931733211502433, + 0.007347129750996828, + 0.005029709078371525, + 0.0022568320855498314, + 0.0032149283215403557, + 0.006791457999497652, + 0.0051362840458750725, + 0.006499935872852802, + 0.0025178012438118458, + 0.008157136850059032, + 0.006114193703979254, + 0.002475515240803361, + 0.0035186177119612694, + 0.004928345791995525, + 0.004562928806990385, + 0.003833947004750371, + 0.0019989165011793375, + 0.005831389222294092, + 0.0044960444793105125, + 0.0024341417010873556, + 0.0033440133556723595 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 4, + 0 + ], + "selected_reward": -0.027687406167387962 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the visual style of the video a traditional Chinese ink painting?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a lone shovel with a long handle visible standing in dark, fertile soil?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a handheld sprayer emitting a fine mist of fertilizer visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the video display the visible fibrous texture of rice paper as the surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video smoothly fade to black at the end?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.08910063654184341 + ], + "opposite_logprob_tokens": [ + -2.4641005992889404 + ], + "logprob_reward": -0.08910063654184341, + "yes_no_margin_reward": 2.375, + "token_credit_reward": -0.08910063654184341, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.16177937388420105, + 0.060151729732751846, + 0.06334652751684189, + 0.12488330155611038, + 0.05660667642951012, + 0.13020379841327667, + 0.07758037000894547, + 0.18005001544952393, + 0.0741531252861023, + 0.07124502211809158 + ], + "credit_map": [ + 0.022015666589140892, + 0.02091783657670021, + 0.03116449899971485, + 0.020158523693680763, + 0.021267088130116463, + 0.01484741736203432, + 0.016898369416594505, + 0.014509988017380238, + 0.008668534457683563, + 0.006682050414383411, + 0.008098073303699493, + 0.008517337031662464, + 0.007137712556868792, + 0.009499983862042427, + 0.004751979373395443, + 0.0067960647866129875, + 0.007841050624847412, + 0.007064278703182936, + 0.012862159870564938, + 0.012212073430418968, + 0.0056315078400075436, + 0.0057298364117741585, + 0.006051939446479082, + 0.005953687708824873, + 0.016724932938814163, + 0.017002075910568237, + 0.017068346962332726, + 0.010860488750040531, + 0.016950713470578194, + 0.01917243003845215, + 0.013247771188616753, + 0.013856564648449421, + 0.00864606723189354, + 0.013641786761581898, + 0.009240045212209225, + 0.003164438297972083, + 0.005331189371645451, + 0.008273919112980366, + 0.00334665784612298, + 0.004962580278515816, + 0.013472722843289375, + 0.01819942332804203, + 0.03331461548805237, + 0.005597975570708513, + 0.006308316718786955, + 0.04103806987404823, + 0.006178510375320911, + 0.00609417911618948, + 0.009686479344964027, + 0.01462370902299881, + 0.011135393753647804, + 0.004800789523869753, + 0.009678985923528671, + 0.013990255072712898, + 0.004869509022682905, + 0.008795257657766342, + 0.027630213648080826, + 0.02531270496547222, + 0.022405080497264862, + 0.01254256907850504, + 0.04053717851638794, + 0.027474399656057358, + 0.01056467741727829, + 0.013583211228251457, + 0.010504678823053837, + 0.010598056949675083, + 0.010180406272411346, + 0.0059409793466329575, + 0.00889620278030634, + 0.013909194618463516, + 0.005902581848204136, + 0.008221028372645378, + 0.010011845268309116, + 0.008863602764904499, + 0.008365132845938206, + 0.006354726850986481, + 0.01490310300141573, + 0.01139457430690527, + 0.0046378448605537415, + 0.006714198738336563 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 0, + 5 + ], + "selected_reward": -0.08910063654184341 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the visual style of the video a traditional Chinese ink painting?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a lone shovel with a long handle visible standing in dark, fertile soil?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a handheld sprayer emitting a fine mist of fertilizer visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the video display the visible fibrous texture of rice paper as the surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video smoothly fade to black at the end?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -0.8261886835098267 + ], + "opposite_logprob_tokens": [ + -0.5761886835098267 + ], + "logprob_reward": -0.8261886835098267, + "yes_no_margin_reward": -0.25, + "token_credit_reward": -0.8261886835098267, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.05214366689324379, + 0.028950262814760208, + 0.1002851203083992, + 0.04985431581735611, + 0.14966613054275513, + 0.3375006914138794, + 0.04849107190966606, + 0.1422615498304367, + 0.04731728881597519, + 0.04352997615933418 + ], + "credit_map": [ + 0.0067696524783968925, + 0.005350014194846153, + 0.005490423180162907, + 0.007491284981369972, + 0.005694341380149126, + 0.007320537697523832, + 0.0063000754453241825, + 0.007727336138486862, + 0.004160064272582531, + 0.004300820641219616, + 0.004458966199308634, + 0.0036934069357812405, + 0.0035151508636772633, + 0.0030315397307276726, + 0.002899324055761099, + 0.0028909891843795776, + 0.012237239629030228, + 0.011649377644062042, + 0.01862528547644615, + 0.021268868818879128, + 0.008920274674892426, + 0.00972207635641098, + 0.010263605043292046, + 0.0075983889400959015, + 0.006168301682919264, + 0.007917230017483234, + 0.006258857902139425, + 0.004820073023438454, + 0.007480244617909193, + 0.005847425665706396, + 0.005967407487332821, + 0.00539478100836277, + 0.02131667733192444, + 0.03532608225941658, + 0.022947372868657112, + 0.00929811131209135, + 0.014085054397583008, + 0.022223901003599167, + 0.007927894592285156, + 0.01654103398323059, + 0.02352004498243332, + 0.060609497129917145, + 0.08339622616767883, + 0.010027767159044743, + 0.014093425124883652, + 0.12067553400993347, + 0.010834887623786926, + 0.014343300834298134, + 0.010299429297447205, + 0.011868865229189396, + 0.00463221687823534, + 0.00224842457100749, + 0.004234209191054106, + 0.00949676800519228, + 0.002137070521712303, + 0.003574085421860218, + 0.02150980569422245, + 0.02323194220662117, + 0.017527064308524132, + 0.008083527907729149, + 0.03259554132819176, + 0.021895961835980415, + 0.00736983260139823, + 0.01004787441343069, + 0.006216993555426598, + 0.007463544607162476, + 0.005785370711237192, + 0.0027163575869053602, + 0.009599524550139904, + 0.0077089122496545315, + 0.0028945747762918472, + 0.004932011477649212, + 0.007192191667854786, + 0.006524813361465931, + 0.004852783400565386, + 0.0019222405971959233, + 0.008193383924663067, + 0.007910570129752159, + 0.00218825601041317, + 0.004745734855532646 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 4, + 7 + ], + "selected_reward": -0.8261886835098267 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the visual style of the video a traditional Chinese ink painting?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a lone shovel with a long handle visible standing in dark, fertile soil?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a handheld sprayer emitting a fine mist of fertilizer visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the video display the visible fibrous texture of rice paper as the surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video smoothly fade to black at the end?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.018209779635071754 + ], + "opposite_logprob_tokens": [ + -4.018209934234619 + ], + "logprob_reward": -0.018209779635071754, + "yes_no_margin_reward": 4.0, + "token_credit_reward": -0.018209779635071754, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08234056085348129, + 0.13306207954883575, + 0.10006192326545715, + 0.0912509560585022, + 0.1495913714170456, + 0.14715240895748138, + 0.07422466576099396, + 0.05605482682585716, + 0.03544772043824196, + 0.13081353902816772 + ], + "credit_map": [ + 0.01326119713485241, + 0.01581871323287487, + 0.007695209234952927, + 0.006625431589782238, + 0.009973092935979366, + 0.009921093471348286, + 0.006191342603415251, + 0.01285447645932436, + 0.024541256949305534, + 0.022743888199329376, + 0.015484774485230446, + 0.012064272537827492, + 0.015615236945450306, + 0.0164046511054039, + 0.009343000128865242, + 0.016864996403455734, + 0.013985577039420605, + 0.014231356792151928, + 0.01165476068854332, + 0.011575204320251942, + 0.011543642729520798, + 0.013471164740622044, + 0.012509915977716446, + 0.011090302839875221, + 0.011343455873429775, + 0.015669452026486397, + 0.008769718930125237, + 0.008991813287138939, + 0.013991955667734146, + 0.013495437800884247, + 0.009678307920694351, + 0.00931081734597683, + 0.019332759082317352, + 0.043290168046951294, + 0.015623299404978752, + 0.008652758784592152, + 0.011443492025136948, + 0.02629215642809868, + 0.008800276555120945, + 0.016156470403075218, + 0.02180924080312252, + 0.03995221108198166, + 0.02033688873052597, + 0.006556229665875435, + 0.008710653521120548, + 0.035774655640125275, + 0.0058252098970115185, + 0.008187326602637768, + 0.012859196402132511, + 0.018332527950406075, + 0.008501754142343998, + 0.0037241103127598763, + 0.006620300002396107, + 0.013133685104548931, + 0.0034084301441907883, + 0.007644659839570522, + 0.006103829946368933, + 0.01116703450679779, + 0.006036838516592979, + 0.003817894496023655, + 0.011427555233240128, + 0.01022811233997345, + 0.003028151346370578, + 0.004245410207659006, + 0.005936556030064821, + 0.00472206762060523, + 0.004443780053406954, + 0.0027175003197044134, + 0.004078854341059923, + 0.006725417450070381, + 0.0030758713837713003, + 0.003747672075405717, + 0.02060793898999691, + 0.018027719110250473, + 0.018952716141939163, + 0.008528831414878368, + 0.022107044234871864, + 0.02265782281756401, + 0.00847004447132349, + 0.011461411602795124 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 5, + 1 + ], + "selected_reward": -0.018209779635071754 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the visual style of the video a traditional Chinese ink painting?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a lone shovel with a long handle visible standing in dark, fertile soil?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a handheld sprayer emitting a fine mist of fertilizer visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the video display the visible fibrous texture of rice paper as the surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video smoothly fade to black at the end?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -1.2236379384994507 + ], + "opposite_logprob_tokens": [ + -0.3486379086971283 + ], + "logprob_reward": -1.2236379384994507, + "yes_no_margin_reward": -0.875, + "token_credit_reward": -1.2236379384994507, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.06496305018663406, + 0.10507375746965408, + 0.100057452917099, + 0.038562048226594925, + 0.14054957032203674, + 0.187516450881958, + 0.1170119121670723, + 0.09194198995828629, + 0.04793769121170044, + 0.10638607293367386 + ], + "credit_map": [ + 0.009668562561273575, + 0.009417259134352207, + 0.008871856145560741, + 0.007654407527297735, + 0.005643599666655064, + 0.007566910237073898, + 0.007596638519316912, + 0.008543813601136208, + 0.017424695193767548, + 0.014714861288666725, + 0.012707456015050411, + 0.01309024915099144, + 0.011894263327121735, + 0.013419155962765217, + 0.009382986463606358, + 0.012440088205039501, + 0.013508394360542297, + 0.011971306055784225, + 0.01710134744644165, + 0.014287449419498444, + 0.011891046538949013, + 0.012976208701729774, + 0.009246332570910454, + 0.009075363166630268, + 0.00587459234520793, + 0.0067634969018399715, + 0.005239428952336311, + 0.002863420872017741, + 0.004861929919570684, + 0.005723530426621437, + 0.0030500078573822975, + 0.004185641650110483, + 0.021979717537760735, + 0.03204315900802612, + 0.019155452027916908, + 0.006999759469181299, + 0.01318952813744545, + 0.02384399063885212, + 0.006555404048413038, + 0.01678255759179592, + 0.021451452746987343, + 0.042672477662563324, + 0.03169158846139908, + 0.0064820111729204655, + 0.010696100071072578, + 0.057586975395679474, + 0.0070847016759216785, + 0.009851138107478619, + 0.017610911279916763, + 0.026754729449748993, + 0.01950845681130886, + 0.006242844276130199, + 0.00987001322209835, + 0.019967705011367798, + 0.0062017980962991714, + 0.010855454951524734, + 0.013432659208774567, + 0.011268378235399723, + 0.011843974702060223, + 0.006870438810437918, + 0.021521354094147682, + 0.01367172971367836, + 0.00538327032700181, + 0.007950184866786003, + 0.007817890495061874, + 0.0064832717180252075, + 0.005545439664274454, + 0.0029451800510287285, + 0.009363142773509026, + 0.007949477061629295, + 0.0032291612587869167, + 0.0046041314490139484, + 0.01600264012813568, + 0.015480979345738888, + 0.01478835754096508, + 0.007144798059016466, + 0.0211589727550745, + 0.01862255483865738, + 0.006099313963204622, + 0.007088459562510252 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 4, + 6 + ], + "selected_reward": -1.2236379384994507 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "No", + "Yes", + "No" + ] + }, + { + "prompt": "static, eye-level medium shot captures a scene rendered entirely in the style of traditional Chinese ink painting. A lone, elegant shovel, its long handle depicted in a confident stroke of dark ink, stands amidst dark, fertile soil, layered with rich, moist grays and blacks. Nearby, a handheld sprayer, rendered with minimalist, economical brushstrokes, emits a fine mist of fertilizer. The atomized particles are captured as soft, diffuse clouds of light gray ink, gently drifting downward. These particles settle onto the dark ground like fine dust, subtly punctuated by rich washes of ink. The entire monochrome composition, executed with bold, expressive brushstrokes on a surface with the visible fibrous texture of rice paper, evokes a tranquil and poetic atmosphere, before the entire image pauses for a moment and smoothly fades to black.", + "seed": 12399, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000095/group_4/index_7168_seed_12399_dp_9_rank_18_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 4, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.20822973258327693, + "vlm_logprob": -0.20822973258327693, + "vlm_yes_no_margin": 3.525, + "vlm_token_credit": -0.20822973258327693, + "vlm_yes_ratio": 0.8, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the visual style of the video a traditional Chinese ink painting?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a lone shovel with a long handle visible standing in dark, fertile soil?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a handheld sprayer emitting a fine mist of fertilizer visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the video display the visible fibrous texture of rice paper as the surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video smoothly fade to black at the end?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02129313349723816 + ], + "opposite_logprob_tokens": [ + -3.8962931632995605 + ], + "logprob_reward": -0.02129313349723816, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.02129313349723816, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15573875606060028, + 0.0838623046875, + 0.06175630912184715, + 0.059007927775382996, + 0.07113688439130783, + 0.13542377948760986, + 0.12888313829898834, + 0.061869535595178604, + 0.12099310010671616, + 0.12132828682661057 + ], + "credit_map": [ + 0.025661824271082878, + 0.024163680151104927, + 0.015642834827303886, + 0.01704183779656887, + 0.031941007822752, + 0.015955936163663864, + 0.014284957200288773, + 0.011046679690480232, + 0.013248571194708347, + 0.01214316114783287, + 0.009583169594407082, + 0.007304736878722906, + 0.019303342327475548, + 0.00797201320528984, + 0.00848944392055273, + 0.005817861761897802, + 0.010019142180681229, + 0.008801550604403019, + 0.006189186125993729, + 0.005841911304742098, + 0.01553078182041645, + 0.004866303876042366, + 0.006461988668888807, + 0.004045446403324604, + 0.008680672384798527, + 0.009920877404510975, + 0.0064857117831707, + 0.004887084010988474, + 0.009528311900794506, + 0.00548299727961421, + 0.007580508477985859, + 0.006441765930503607, + 0.011247592978179455, + 0.010640502907335758, + 0.006759449373930693, + 0.006301057059317827, + 0.010365832597017288, + 0.007896835915744305, + 0.008857766166329384, + 0.009067850187420845, + 0.025110401213169098, + 0.03519595041871071, + 0.011585609056055546, + 0.013293774798512459, + 0.01364039909094572, + 0.013624152168631554, + 0.011034009978175163, + 0.01193948183208704, + 0.035889770835638046, + 0.03184850886464119, + 0.013355073519051075, + 0.007658594753593206, + 0.0050959805957973, + 0.007825558073818684, + 0.007506413385272026, + 0.019703233614563942, + 0.00962739810347557, + 0.007687784731388092, + 0.009883496910333633, + 0.006356470286846161, + 0.00876681413501501, + 0.0057167476043105125, + 0.005613595712929964, + 0.008217228576540947, + 0.015889659523963928, + 0.013476714491844177, + 0.01921815238893032, + 0.015643948689103127, + 0.015336835756897926, + 0.013310159556567669, + 0.012182433158159256, + 0.015935199335217476, + 0.014920858666300774, + 0.013217851519584656, + 0.017642633989453316, + 0.018276195973157883, + 0.012116663157939911, + 0.016520217061042786, + 0.014035363681614399, + 0.014598499052226543 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 5, + 6 + ], + "selected_reward": -0.02129313349723816 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the visual style of the video a traditional Chinese ink painting?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a lone shovel with a long handle visible standing in dark, fertile soil?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a handheld sprayer emitting a fine mist of fertilizer visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the video display the visible fibrous texture of rice paper as the surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video smoothly fade to black at the end?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.029826786369085312 + ], + "opposite_logprob_tokens": [ + -3.5298268795013428 + ], + "logprob_reward": -0.029826786369085312, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.029826786369085312, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.19469907879829407, + 0.051060404628515244, + 0.04596719145774841, + 0.04175960272550583, + 0.060289401561021805, + 0.10945408791303635, + 0.19540554285049438, + 0.08047613501548767, + 0.06947991997003555, + 0.1514086127281189 + ], + "credit_map": [ + 0.033374492079019547, + 0.0274666715413332, + 0.019175143912434578, + 0.01808934658765793, + 0.05039234831929207, + 0.018018029630184174, + 0.015578709542751312, + 0.01260434091091156, + 0.008024533279240131, + 0.006350363604724407, + 0.0068376148119568825, + 0.005799753125756979, + 0.010756508447229862, + 0.004740052856504917, + 0.0047966488637030125, + 0.003754928708076477, + 0.007135982625186443, + 0.005772074684500694, + 0.006104905158281326, + 0.004592796787619591, + 0.007931698113679886, + 0.004802042618393898, + 0.004881082568317652, + 0.004746612627059221, + 0.006477830000221729, + 0.007029653061181307, + 0.0043919943273067474, + 0.00452728196978569, + 0.008979297243058681, + 0.0031064122449606657, + 0.005030927713960409, + 0.0022162008099257946, + 0.009319273754954338, + 0.0064647747203707695, + 0.007236868608742952, + 0.004625383764505386, + 0.014224899001419544, + 0.006728422362357378, + 0.005846682004630566, + 0.0058430940844118595, + 0.023008020594716072, + 0.032413333654403687, + 0.006556730251759291, + 0.010260580107569695, + 0.013950558379292488, + 0.007529176771640778, + 0.009212462231516838, + 0.006523232441395521, + 0.033632006496191025, + 0.03396061435341835, + 0.0196552537381649, + 0.015440043061971664, + 0.034888941794633865, + 0.01603686809539795, + 0.02033253014087677, + 0.021459294483065605, + 0.007893913425505161, + 0.00965537503361702, + 0.018602412194013596, + 0.007860795594751835, + 0.006937618833035231, + 0.006994361523538828, + 0.006733054295182228, + 0.015798596665263176, + 0.00931271817535162, + 0.008761148899793625, + 0.011307177133858204, + 0.009865492582321167, + 0.008703124709427357, + 0.005033818073570728, + 0.007925244979560375, + 0.008571190759539604, + 0.012852887623012066, + 0.017713801935315132, + 0.025455275550484657, + 0.0196516290307045, + 0.0182055477052927, + 0.017484473064541817, + 0.023548396304249763, + 0.016496609896421432 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 0, + 9 + ], + "selected_reward": -0.029826786369085312 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the visual style of the video a traditional Chinese ink painting?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a lone shovel with a long handle visible standing in dark, fertile soil?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a handheld sprayer emitting a fine mist of fertilizer visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the video display the visible fibrous texture of rice paper as the surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video smoothly fade to black at the end?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -0.9743054509162903 + ], + "opposite_logprob_tokens": [ + -0.4743054509162903 + ], + "logprob_reward": -0.9743054509162903, + "yes_no_margin_reward": -0.5, + "token_credit_reward": -0.9743054509162903, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.09295794367790222, + 0.05066218972206116, + 0.044803787022829056, + 0.03894878178834915, + 0.04860826954245567, + 0.12979288399219513, + 0.1284254640340805, + 0.115898497402668, + 0.12839242815971375, + 0.22150981426239014 + ], + "credit_map": [ + 0.01383237261325121, + 0.012640749104321003, + 0.010211781598627567, + 0.006710987538099289, + 0.01704883947968483, + 0.010915801860392094, + 0.007962465286254883, + 0.013634943403303623, + 0.008505206555128098, + 0.010163228958845139, + 0.005480853375047445, + 0.005158364772796631, + 0.005883474834263325, + 0.0055168005637824535, + 0.005229521542787552, + 0.004724740516394377, + 0.008311392739415169, + 0.010136507451534271, + 0.003963451832532883, + 0.003115240950137377, + 0.009531634859740734, + 0.003617594949901104, + 0.003461790969595313, + 0.0026661751326173544, + 0.0063864379189908504, + 0.005123182665556669, + 0.0037048293743282557, + 0.003513260977342725, + 0.008479430340230465, + 0.0041299741715192795, + 0.004435733892023563, + 0.0031759333796799183, + 0.006383994594216347, + 0.004580280743539333, + 0.0050395457074046135, + 0.004331708885729313, + 0.01078746560961008, + 0.003911959007382393, + 0.008563979528844357, + 0.005009339656680822, + 0.021967627108097076, + 0.026079511269927025, + 0.014945013448596, + 0.013712761923670769, + 0.01289069838821888, + 0.014340724796056747, + 0.014262028969824314, + 0.011594507843255997, + 0.03334642946720123, + 0.03493417054414749, + 0.01176372915506363, + 0.01019353698939085, + 0.006130197551101446, + 0.00920045841485262, + 0.009855955839157104, + 0.013000966049730778, + 0.01233674306422472, + 0.01346237026154995, + 0.019292762503027916, + 0.011848855763673782, + 0.014929059892892838, + 0.010281586088240147, + 0.00872456282377243, + 0.025022557005286217, + 0.01457760389894247, + 0.012413433752954006, + 0.02647729218006134, + 0.012173259630799294, + 0.02053331770002842, + 0.014651988632977009, + 0.010858248919248581, + 0.016707269474864006, + 0.016476087272167206, + 0.015144869685173035, + 0.03551934286952019, + 0.033717043697834015, + 0.031373582780361176, + 0.028542442247271538, + 0.02693561092019081, + 0.03380083665251732 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 5, + 6 + ], + "selected_reward": -0.9743054509162903 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the visual style of the video a traditional Chinese ink painting?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a lone shovel with a long handle visible standing in dark, fertile soil?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a handheld sprayer emitting a fine mist of fertilizer visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the video display the visible fibrous texture of rice paper as the surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video smoothly fade to black at the end?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.014209050685167313 + ], + "opposite_logprob_tokens": [ + -4.264209270477295 + ], + "logprob_reward": -0.014209050685167313, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.014209050685167313, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10837279260158539, + 0.05589083954691887, + 0.04271963983774185, + 0.03575433790683746, + 0.0662897601723671, + 0.06839081645011902, + 0.125056654214859, + 0.08515827357769012, + 0.15421496331691742, + 0.25815197825431824 + ], + "credit_map": [ + 0.016048284247517586, + 0.014605477452278137, + 0.015998639166355133, + 0.009757613763213158, + 0.015812071040272713, + 0.013812797144055367, + 0.008526849560439587, + 0.013811057433485985, + 0.00966516975313425, + 0.009822293184697628, + 0.004126924555748701, + 0.00425034761428833, + 0.015004044398665428, + 0.005396616645157337, + 0.00394098786637187, + 0.003684450639411807, + 0.007966168224811554, + 0.008956988342106342, + 0.002772645791992545, + 0.0031236354261636734, + 0.01031768973916769, + 0.004178052302449942, + 0.002665257779881358, + 0.0027391999028623104, + 0.005939377471804619, + 0.007108768448233604, + 0.003556401701644063, + 0.003819432808086276, + 0.005004335660487413, + 0.0041190548799932, + 0.0038216121029108763, + 0.002385352738201618, + 0.011085794307291508, + 0.010712547227740288, + 0.0058581712655723095, + 0.005404053255915642, + 0.012076097540557384, + 0.006026783958077431, + 0.00710209971293807, + 0.008024207316339016, + 0.008524315431714058, + 0.008669575676321983, + 0.008570080623030663, + 0.008062474429607391, + 0.007071690168231726, + 0.008434677496552467, + 0.0066948095336556435, + 0.012363185174763203, + 0.017082789912819862, + 0.022631315514445305, + 0.01224455051124096, + 0.010459288023412228, + 0.023552829399704933, + 0.012192197144031525, + 0.014955605380237103, + 0.011938062496483326, + 0.008762278594076633, + 0.01291353814303875, + 0.016780974343419075, + 0.008088956587016582, + 0.01064321305602789, + 0.00564409326761961, + 0.0075521632097661495, + 0.0147730503231287, + 0.016932852566242218, + 0.024550627917051315, + 0.02279779687523842, + 0.026159200817346573, + 0.014439807273447514, + 0.015531692653894424, + 0.01211329735815525, + 0.021689681336283684, + 0.020949650555849075, + 0.026752479374408722, + 0.044463396072387695, + 0.042295895516872406, + 0.027877818793058395, + 0.026651673018932343, + 0.046131327748298645, + 0.02302970178425312 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 8, + 6 + ], + "selected_reward": -0.014209050685167313 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the visual style of the video a traditional Chinese ink painting?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a lone shovel with a long handle visible standing in dark, fertile soil?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a handheld sprayer emitting a fine mist of fertilizer visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the video display the visible fibrous texture of rice paper as the surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video smoothly fade to black at the end?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0015142414486035705 + ], + "opposite_logprob_tokens": [ + -6.501514434814453 + ], + "logprob_reward": -0.0015142414486035705, + "yes_no_margin_reward": 6.5, + "token_credit_reward": -0.0015142414486035705, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06791676580905914, + 0.0333685427904129, + 0.02200433239340782, + 0.01715788245201111, + 0.02129119634628296, + 0.036218296736478806, + 0.13053977489471436, + 0.12200625240802765, + 0.18480722606182098, + 0.36468976736068726 + ], + "credit_map": [ + 0.010403633117675781, + 0.00992238987237215, + 0.008082158863544464, + 0.006137744057923555, + 0.011882685124874115, + 0.007707558106631041, + 0.00796227902173996, + 0.005818312056362629, + 0.004307342227548361, + 0.004080852027982473, + 0.005405813921242952, + 0.0029961352702230215, + 0.00552768399938941, + 0.004519877955317497, + 0.0034809859935194254, + 0.003049852792173624, + 0.003135437611490488, + 0.002384352730587125, + 0.003954605665057898, + 0.002089301124215126, + 0.0031247755978256464, + 0.00200794474221766, + 0.002427543979138136, + 0.0028803704772144556, + 0.0026301105972379446, + 0.001781127299182117, + 0.0021434032823890448, + 0.001542402314953506, + 0.003531858790665865, + 0.0015837959945201874, + 0.0018935005646198988, + 0.0020516845397651196, + 0.00290275109000504, + 0.0028372034430503845, + 0.003085087053477764, + 0.002309520961716771, + 0.0043732416816055775, + 0.0022675003856420517, + 0.0018147595692425966, + 0.001701133674941957, + 0.005171450786292553, + 0.00445183739066124, + 0.005024312064051628, + 0.0038691649679094553, + 0.004409471061080694, + 0.004216786008328199, + 0.002951150294393301, + 0.006124123930931091, + 0.024742908775806427, + 0.025080222636461258, + 0.014994398690760136, + 0.011332839727401733, + 0.016060905531048775, + 0.009917364455759525, + 0.013975508511066437, + 0.014435624703764915, + 0.015439500100910664, + 0.014870494604110718, + 0.018354330211877823, + 0.016639990732073784, + 0.018471935763955116, + 0.010725628584623337, + 0.012435349635779858, + 0.015069027431309223, + 0.019231269136071205, + 0.027320055291056633, + 0.028843680396676064, + 0.03337496519088745, + 0.016545677557587624, + 0.017538808286190033, + 0.018072811886668205, + 0.02387995272874832, + 0.03794717416167259, + 0.031042106449604034, + 0.0633818581700325, + 0.06011961027979851, + 0.04380635917186737, + 0.040226180106401443, + 0.04681418836116791, + 0.041352275758981705 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 8, + 6 + ], + "selected_reward": -0.0015142414486035705 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "No", + "Yes", + "Yes" + ] + }, + { + "prompt": "static, eye-level medium shot captures a scene rendered entirely in the style of traditional Chinese ink painting. A lone, elegant shovel, its long handle depicted in a confident stroke of dark ink, stands amidst dark, fertile soil, layered with rich, moist grays and blacks. Nearby, a handheld sprayer, rendered with minimalist, economical brushstrokes, emits a fine mist of fertilizer. The atomized particles are captured as soft, diffuse clouds of light gray ink, gently drifting downward. These particles settle onto the dark ground like fine dust, subtly punctuated by rich washes of ink. The entire monochrome composition, executed with bold, expressive brushstrokes on a surface with the visible fibrous texture of rice paper, evokes a tranquil and poetic atmosphere, before the entire image pauses for a moment and smoothly fades to black.", + "seed": 12400, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000095/group_4/index_7168_seed_12400_dp_9_rank_18_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 4, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.03481618389487266, + "vlm_logprob": -0.03481618389487266, + "vlm_yes_no_margin": 3.575, + "vlm_token_credit": -0.03481618389487266, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the visual style of the video a traditional Chinese ink painting?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a lone shovel with a long handle visible standing in dark, fertile soil?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a handheld sprayer emitting a fine mist of fertilizer visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the video display the visible fibrous texture of rice paper as the surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video smoothly fade to black at the end?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.022081349045038223 + ], + "opposite_logprob_tokens": [ + -3.8970813751220703 + ], + "logprob_reward": -0.022081349045038223, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.022081349045038223, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.04529478773474693, + 0.039896875619888306, + 0.028811015188694, + 0.054501187056303024, + 0.034906864166259766, + 0.02399091236293316, + 0.16616863012313843, + 0.080841064453125, + 0.1906975954771042, + 0.33489102125167847 + ], + "credit_map": [ + 0.005975091364234686, + 0.007512978743761778, + 0.005004511680454016, + 0.004262495785951614, + 0.004262313712388277, + 0.007668949663639069, + 0.004651503637433052, + 0.005956942681223154, + 0.008134616538882256, + 0.006230674684047699, + 0.0034929097164422274, + 0.003119130851700902, + 0.003485705005005002, + 0.007848779670894146, + 0.004409099463373423, + 0.003175957128405571, + 0.004907716531306505, + 0.004212572239339352, + 0.0033421246334910393, + 0.0023331893607974052, + 0.002280804794281721, + 0.004916694480925798, + 0.002982364036142826, + 0.003835547249764204, + 0.0074996636249125, + 0.008400004357099533, + 0.004824200179427862, + 0.0039743417873978615, + 0.005967226345092058, + 0.012834000401198864, + 0.004603459034115076, + 0.006398291327059269, + 0.006962759420275688, + 0.004409254062920809, + 0.004829226527363062, + 0.0033453553915023804, + 0.003563967300578952, + 0.0030976743437349796, + 0.0038218232803046703, + 0.004876803606748581, + 0.004599353764206171, + 0.0021966062486171722, + 0.0036867940798401833, + 0.0031448728404939175, + 0.002225020434707403, + 0.003528426168486476, + 0.002218599896878004, + 0.002391238696873188, + 0.017087962478399277, + 0.05445416271686554, + 0.008182604797184467, + 0.007256101351231337, + 0.011752127669751644, + 0.013729267753660679, + 0.007990164682269096, + 0.04571623355150223, + 0.017981667071580887, + 0.009457321837544441, + 0.010300425812602043, + 0.005348312668502331, + 0.004667623899877071, + 0.016779763624072075, + 0.0047274380922317505, + 0.011578512378036976, + 0.01897883415222168, + 0.06058583781123161, + 0.014399216510355473, + 0.02046707086265087, + 0.013659154996275902, + 0.021332571282982826, + 0.0117063969373703, + 0.029568525031208992, + 0.02815351076424122, + 0.04788611829280853, + 0.06047497317194939, + 0.022914916276931763, + 0.03842991217970848, + 0.056885212659835815, + 0.03602223098278046, + 0.04412414878606796 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 8, + 6 + ], + "selected_reward": -0.022081349045038223 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the visual style of the video a traditional Chinese ink painting?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a lone shovel with a long handle visible standing in dark, fertile soil?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a handheld sprayer emitting a fine mist of fertilizer visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the video display the visible fibrous texture of rice paper as the surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video smoothly fade to black at the end?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.008651750162243843 + ], + "opposite_logprob_tokens": [ + -4.7586517333984375 + ], + "logprob_reward": -0.008651750162243843, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.008651750162243843, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06723581999540329, + 0.08552230149507523, + 0.016872717067599297, + 0.017075203359127045, + 0.03428291156888008, + 0.018817273899912834, + 0.4529648721218109, + 0.06049555167555809, + 0.07166268676519394, + 0.17507067322731018 + ], + "credit_map": [ + 0.007584216073155403, + 0.015046242624521255, + 0.005113291088491678, + 0.005666406825184822, + 0.005262333899736404, + 0.014121024869382381, + 0.006333698518574238, + 0.008108600974082947, + 0.009956554509699345, + 0.011204416863620281, + 0.010441801510751247, + 0.008138397708535194, + 0.00749989366158843, + 0.013379059731960297, + 0.00865049846470356, + 0.016251685097813606, + 0.002484254539012909, + 0.002516767242923379, + 0.001612241379916668, + 0.001371580408886075, + 0.001579173607751727, + 0.003279621945694089, + 0.0018937578424811363, + 0.0021353205665946007, + 0.0029126747976988554, + 0.002488864352926612, + 0.0017307172529399395, + 0.001994699239730835, + 0.001422551111318171, + 0.0019214327912777662, + 0.0022190366871654987, + 0.002385226543992758, + 0.005810547154396772, + 0.004224208649247885, + 0.003527804045006633, + 0.0033020656555891037, + 0.0034410639200359583, + 0.007656008470803499, + 0.0026061481330543756, + 0.0037150674033910036, + 0.0024357212241739035, + 0.0022858374286442995, + 0.0019161173840984702, + 0.0015108586521819234, + 0.0022293587680906057, + 0.003766066860407591, + 0.0025197414215654135, + 0.0021535721607506275, + 0.04652220383286476, + 0.15474459528923035, + 0.01548724714666605, + 0.01855272240936756, + 0.0329352542757988, + 0.03657787665724754, + 0.02021954394876957, + 0.12792539596557617, + 0.012852045707404613, + 0.0069941082037985325, + 0.008452962152659893, + 0.005560777615755796, + 0.004921627230942249, + 0.009795481339097023, + 0.0037193798925727606, + 0.008199172094464302, + 0.011050330474972725, + 0.011309880763292313, + 0.008805656805634499, + 0.010231971740722656, + 0.009043309837579727, + 0.007343309931457043, + 0.008261197246611118, + 0.00561703322455287, + 0.017032044008374214, + 0.02325080893933773, + 0.029039178043603897, + 0.014181303791701794, + 0.022486548870801926, + 0.027009692043066025, + 0.01992458663880825, + 0.022146502509713173 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 9, + 1 + ], + "selected_reward": -0.008651750162243843 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the visual style of the video a traditional Chinese ink painting?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a lone shovel with a long handle visible standing in dark, fertile soil?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a handheld sprayer emitting a fine mist of fertilizer visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the video display the visible fibrous texture of rice paper as the surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video smoothly fade to black at the end?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.05496196821331978 + ], + "opposite_logprob_tokens": [ + -2.929961919784546 + ], + "logprob_reward": -0.05496196821331978, + "yes_no_margin_reward": 2.875, + "token_credit_reward": -0.05496196821331978, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12459755688905716, + 0.061477985233068466, + 0.02983075939118862, + 0.02960309386253357, + 0.08648776262998581, + 0.06657123565673828, + 0.2224561870098114, + 0.09910077601671219, + 0.12502041459083557, + 0.15485423803329468 + ], + "credit_map": [ + 0.021319307386875153, + 0.025354713201522827, + 0.009906170889735222, + 0.010160849429666996, + 0.010444379411637783, + 0.021419482305645943, + 0.012143081054091454, + 0.013849573209881783, + 0.010629192925989628, + 0.009716011583805084, + 0.006874504033476114, + 0.004624571185559034, + 0.004251772537827492, + 0.01335739903151989, + 0.006013114005327225, + 0.00601141806691885, + 0.00614446634426713, + 0.004088626243174076, + 0.0030707179103046656, + 0.0031602121889591217, + 0.0021663911174982786, + 0.0031807732302695513, + 0.003106391755864024, + 0.004913181532174349, + 0.003943097777664661, + 0.0037601811345666647, + 0.0027957444544881582, + 0.0029489980079233646, + 0.002230473794043064, + 0.0033773782197386026, + 0.0037085823714733124, + 0.00683863740414381, + 0.019594313576817513, + 0.011202234774827957, + 0.011088051833212376, + 0.0078109316527843475, + 0.00809100829064846, + 0.010223821736872196, + 0.007772419136017561, + 0.010704983957111835, + 0.014691159129142761, + 0.006977429613471031, + 0.01055411621928215, + 0.006332159973680973, + 0.004569946322590113, + 0.00943246204406023, + 0.007775263860821724, + 0.0062387012876570225, + 0.02551965042948723, + 0.07205968350172043, + 0.01063913106918335, + 0.009615747258067131, + 0.015992416068911552, + 0.018272211775183678, + 0.009712910279631615, + 0.060644425451755524, + 0.02397686056792736, + 0.01484877709299326, + 0.012816532514989376, + 0.0105799101293087, + 0.007076195906847715, + 0.008466202765703201, + 0.0076263598166406155, + 0.013709936290979385, + 0.013886888511478901, + 0.023577511310577393, + 0.018423305824398994, + 0.013013049028813839, + 0.014905870892107487, + 0.013575003482401371, + 0.015898218378424644, + 0.01174056064337492, + 0.014941681176424026, + 0.01837448962032795, + 0.032656677067279816, + 0.02222394198179245, + 0.012912224046885967, + 0.01818040758371353, + 0.02110949158668518, + 0.01445532776415348 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 9, + 8 + ], + "selected_reward": -0.05496196821331978 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the visual style of the video a traditional Chinese ink painting?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a lone shovel with a long handle visible standing in dark, fertile soil?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a handheld sprayer emitting a fine mist of fertilizer visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the video display the visible fibrous texture of rice paper as the surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video smoothly fade to black at the end?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.026358967646956444 + ], + "opposite_logprob_tokens": [ + -3.6513590812683105 + ], + "logprob_reward": -0.026358967646956444, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.026358967646956444, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07902669906616211, + 0.042700398713350296, + 0.018891170620918274, + 0.05162803828716278, + 0.06679044663906097, + 0.028449425473809242, + 0.15374307334423065, + 0.06568465381860733, + 0.064978688955307, + 0.428107351064682 + ], + "credit_map": [ + 0.009902586229145527, + 0.01788540743291378, + 0.007707125972956419, + 0.006088527385145426, + 0.007217356003820896, + 0.016384901478886604, + 0.00582736125215888, + 0.00801344309002161, + 0.006423341576009989, + 0.005967131350189447, + 0.005231698974967003, + 0.003417631611227989, + 0.004242326132953167, + 0.007464280817657709, + 0.004220548551529646, + 0.00573344063013792, + 0.003242655424401164, + 0.0028079955372959375, + 0.0020926690194755793, + 0.0014158481499180198, + 0.002105818362906575, + 0.0033286192920058966, + 0.0021029228810220957, + 0.0017946426523849368, + 0.005970175378024578, + 0.006852869410067797, + 0.004634552169591188, + 0.0037880856543779373, + 0.00394954951480031, + 0.008344135247170925, + 0.004443841055035591, + 0.013644829392433167, + 0.013109478168189526, + 0.00839855708181858, + 0.007671774365007877, + 0.00676519051194191, + 0.007250196300446987, + 0.010558896698057652, + 0.0056926412507891655, + 0.007343719247728586, + 0.004068347159773111, + 0.0028351368382573128, + 0.003369694110006094, + 0.003524145809933543, + 0.0027503047604113817, + 0.006037827115505934, + 0.002576958853751421, + 0.003287013154476881, + 0.016851164400577545, + 0.050342053174972534, + 0.007931455038487911, + 0.007258167024701834, + 0.010713648051023483, + 0.012641454115509987, + 0.0077258567325770855, + 0.040279291570186615, + 0.010929306969046593, + 0.00941524002701044, + 0.009292214177548885, + 0.00691553670912981, + 0.005194708704948425, + 0.010669199749827385, + 0.005265974439680576, + 0.008002473041415215, + 0.010358802042901516, + 0.019174177199602127, + 0.004791046492755413, + 0.004304639529436827, + 0.005570349749177694, + 0.0068456209264695644, + 0.005343062337487936, + 0.008590995334088802, + 0.03878406807780266, + 0.055626191198825836, + 0.07348988205194473, + 0.03659242391586304, + 0.05403328314423561, + 0.06674924492835999, + 0.05051540955901146, + 0.05231689661741257 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 6, + 0 + ], + "selected_reward": -0.026358967646956444 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the visual style of the video a traditional Chinese ink painting?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a lone shovel with a long handle visible standing in dark, fertile soil?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a handheld sprayer emitting a fine mist of fertilizer visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the video display the visible fibrous texture of rice paper as the surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video smoothly fade to black at the end?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.06202688440680504 + ], + "opposite_logprob_tokens": [ + -2.8120269775390625 + ], + "logprob_reward": -0.06202688440680504, + "yes_no_margin_reward": 2.75, + "token_credit_reward": -0.06202688440680504, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.014504040591418743, + 0.009448878467082977, + 0.011124584823846817, + 0.011094833724200726, + 0.020296843722462654, + 0.024684399366378784, + 0.044029347598552704, + 0.0985160619020462, + 0.17990203201770782, + 0.5863989591598511 + ], + "credit_map": [ + 0.0021841342095285654, + 0.0026501521933823824, + 0.0014304009964689612, + 0.0010578467044979334, + 0.0014071306213736534, + 0.002403444377705455, + 0.001142917200922966, + 0.002228014636784792, + 0.0012322718976065516, + 0.0014379643835127354, + 0.0010896347230300307, + 0.0007096778135746717, + 0.0009960511233657598, + 0.0015760083915665746, + 0.000988895888440311, + 0.001418374595232308, + 0.0018846462480723858, + 0.0014732042327523232, + 0.0011314754374325275, + 0.0009605423547327518, + 0.001097124069929123, + 0.0017574538942426443, + 0.0008879022789187729, + 0.001932236016727984, + 0.002081760670989752, + 0.0015100146410986781, + 0.0010631862096488476, + 0.001072879764251411, + 0.001394982566125691, + 0.0017380823846906424, + 0.0009562660125084221, + 0.001277661183848977, + 0.00300804921425879, + 0.0025019447784870863, + 0.0022364920005202293, + 0.0023041190579533577, + 0.001845210324972868, + 0.004003637470304966, + 0.0015246066031977534, + 0.0028727843891829252, + 0.0031650483142584562, + 0.0024242799263447523, + 0.00260558002628386, + 0.0020164602901786566, + 0.0028429203666746616, + 0.005772758740931749, + 0.0029799588955938816, + 0.0028773932717740536, + 0.006676303222775459, + 0.010505770333111286, + 0.005234249401837587, + 0.0036592544056475163, + 0.0035269742365926504, + 0.003996817860752344, + 0.003458430990576744, + 0.006971547845751047, + 0.015473322942852974, + 0.013900421559810638, + 0.009781072847545147, + 0.01280331239104271, + 0.011830136179924011, + 0.013863304629921913, + 0.009143921546638012, + 0.0117205660790205, + 0.01821807026863098, + 0.05300857126712799, + 0.01433572918176651, + 0.014263564720749855, + 0.016154007986187935, + 0.019546521827578545, + 0.015589185990393162, + 0.02878638729453087, + 0.049972053617239, + 0.08655062317848206, + 0.10003584623336792, + 0.0423152931034565, + 0.06398908793926239, + 0.101364366710186, + 0.06277278810739517, + 0.07939891517162323 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 8, + 7 + ], + "selected_reward": -0.06202688440680504 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000095/group_5/index_tensor([4168, 4168])_group_5_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000095/group_5/index_tensor([4168, 4168])_group_5_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..376d055c53a60b382194d1e5d60752cd6472b4ed --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000095/group_5/index_tensor([4168, 4168])_group_5_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "An immersive first-person perspective, the camera pans and pans smoothly with the body's movements. Within the frame, a pair of slender, powerful hands tightly grasp an ancient longsword. The dark leather wrapped around the hilt has a warm patina from ageless wear. The slender blade gleams with a cold, restrained metallic luster under the soft, diffuse light filtering through the mist. You are perched atop a precipitous mountain, the jagged rock beneath your feet, surrounded by thick, flowing clouds, which gather and dissipate, occasionally revealing the indigo-hued outlines of the distant mountains, like an ink painting. With a deep breath, you raise the sword with steady composure, launching into a flowing sword dance. Each swing is precise and powerful, the tip of the sword tracing graceful, sharp silver arcs through the air, emitting a soft, sharp whistling sound as it slices through the mist. Your body lunges, spins, and stretches, its wide white linen sleeves fluttering in the air, seemingly merging with the flowing mist. The entire movement is seamless and rhythmic, each stretch awakening dormant ancient veins, imbued with a fusion of focus, tranquility, and strength. The surrounding landscape, the twisting ancient pines, the shifting mist, and the trajectory of your body and blade are all composed of elegant, flowing lines. Against a predominant palette of cool gray and indigo, they create a dynamic, solitary, and profound artistic conception.", + "seed": 26689, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000095/group_5/index_4168_seed_26689_dp_10_rank_20_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 5, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.08184535540640354, + "vlm_logprob": -0.08184535540640354, + "vlm_yes_no_margin": 3.525, + "vlm_token_credit": -0.08184535540640354, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video shot from a first-person perspective showing the character's hands?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the character's hands grasping an ancient longsword with dark leather wrapped around the hilt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the character perform a flowing sword dance involving lunges, spins, and stretches?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the setting a precipitous mountain with jagged rocks, thick flowing clouds, and distant indigo-hued mountains?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera pan smoothly in sync with the body's movements?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.3197437524795532 + ], + "opposite_logprob_tokens": [ + -1.3197437524795532 + ], + "logprob_reward": -0.3197437524795532, + "yes_no_margin_reward": 1.0, + "token_credit_reward": -0.3197437524795532, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0879240334033966, + 0.03049330599606037, + 0.1713193953037262, + 0.030705777928233147, + 0.1433010697364807, + 0.1838381588459015, + 0.15321306884288788, + 0.09073176234960556, + 0.03180355578660965, + 0.07666989415884018 + ], + "credit_map": [ + 0.007275980897247791, + 0.01808340847492218, + 0.012918385677039623, + 0.0150306923314929, + 0.009639283642172813, + 0.005804918240755796, + 0.011404033750295639, + 0.007767326198518276, + 0.00331634352914989, + 0.0045752148143947124, + 0.002990313805639744, + 0.004177815746515989, + 0.004791758488863707, + 0.004422788042575121, + 0.0031287020538002253, + 0.003090365557000041, + 0.01894034631550312, + 0.033097315579652786, + 0.022557904943823814, + 0.023334339261054993, + 0.017055882140994072, + 0.031143683940172195, + 0.012614593841135502, + 0.012575321830809116, + 0.002871939679607749, + 0.004290489479899406, + 0.0034602724481374025, + 0.0035439948551356792, + 0.0032556415535509586, + 0.004335128236562014, + 0.0037602134980261326, + 0.005188096780329943, + 0.013793864287436008, + 0.025562098249793053, + 0.018482552841305733, + 0.015266467817127705, + 0.013996453024446964, + 0.018962770700454712, + 0.01692051813006401, + 0.020316338166594505, + 0.027561014518141747, + 0.03650753200054169, + 0.02451445907354355, + 0.01470660138875246, + 0.014269066974520683, + 0.03431086242198944, + 0.015226384624838829, + 0.016742227599024773, + 0.014298264868557453, + 0.021004751324653625, + 0.024981731548905373, + 0.011963802389800549, + 0.03112531267106533, + 0.016972383484244347, + 0.016444111242890358, + 0.01642271690070629, + 0.0059282006695866585, + 0.010173571296036243, + 0.007994619198143482, + 0.008593462407588959, + 0.01056516170501709, + 0.02118758112192154, + 0.009520704858005047, + 0.016768459230661392, + 0.002060430124402046, + 0.002763083903118968, + 0.0030003341380506754, + 0.0046514286659657955, + 0.005367806646972895, + 0.005071842577308416, + 0.004756920505315065, + 0.004131706897169352, + 0.0063080252148211, + 0.008853449486196041, + 0.008396215736865997, + 0.007593062240630388, + 0.011000515893101692, + 0.01403260137885809, + 0.012053582817316055, + 0.008432434871792793 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 2, + 6 + ], + "selected_reward": -0.3197437524795532 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video shot from a first-person perspective showing the character's hands?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the character's hands grasping an ancient longsword with dark leather wrapped around the hilt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the character perform a flowing sword dance involving lunges, spins, and stretches?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the setting a precipitous mountain with jagged rocks, thick flowing clouds, and distant indigo-hued mountains?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera pan smoothly in sync with the body's movements?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.054969754070043564 + ], + "opposite_logprob_tokens": [ + -2.9299697875976562 + ], + "logprob_reward": -0.054969754070043564, + "yes_no_margin_reward": 2.875, + "token_credit_reward": -0.054969754070043564, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07404332607984543, + 0.03172548487782478, + 0.08998940885066986, + 0.08345606923103333, + 0.13466355204582214, + 0.0720541775226593, + 0.059341028332710266, + 0.29627230763435364, + 0.05303199216723442, + 0.10542263835668564 + ], + "credit_map": [ + 0.005706848111003637, + 0.018943073228001595, + 0.012431665323674679, + 0.012708972208201885, + 0.005721500143408775, + 0.005992904771119356, + 0.006346368696540594, + 0.0061919959262013435, + 0.0036373352631926537, + 0.006539956666529179, + 0.0038319046143442392, + 0.004320518579334021, + 0.0041855573654174805, + 0.0032275717239826918, + 0.003666781820356846, + 0.0023158600088208914, + 0.009006068110466003, + 0.018991604447364807, + 0.01243639923632145, + 0.015733718872070312, + 0.00982308853417635, + 0.010150197893381119, + 0.007342046592384577, + 0.006506290286779404, + 0.008138138800859451, + 0.010950871743261814, + 0.010305891744792461, + 0.00943987350910902, + 0.008284841664135456, + 0.017972510308027267, + 0.009786564856767654, + 0.008577373810112476, + 0.02029067650437355, + 0.019822150468826294, + 0.023980431258678436, + 0.013394559733569622, + 0.013280319049954414, + 0.015972241759300232, + 0.009022236801683903, + 0.018900930881500244, + 0.008383366279304028, + 0.01322952751070261, + 0.006824480835348368, + 0.007322785910218954, + 0.008041715249419212, + 0.01027834601700306, + 0.011102583259344101, + 0.006871371995657682, + 0.005595432128757238, + 0.008410110138356686, + 0.010947288945317268, + 0.008881519548594952, + 0.008048963733017445, + 0.006957977544516325, + 0.006900502368807793, + 0.00359923648647964, + 0.009060017764568329, + 0.008649815805256367, + 0.00843022484332323, + 0.007786849047988653, + 0.017506366595625877, + 0.12224122136831284, + 0.011903170496225357, + 0.11069464683532715, + 0.004956674296408892, + 0.005067782010883093, + 0.004132539965212345, + 0.005389530211687088, + 0.006310985889285803, + 0.010016258805990219, + 0.008384100161492825, + 0.008774123154580593, + 0.008571363985538483, + 0.013393261469900608, + 0.011188389733433723, + 0.012023082934319973, + 0.014501637779176235, + 0.016041766852140427, + 0.01821245066821575, + 0.011490686796605587 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 4, + 9 + ], + "selected_reward": -0.054969754070043564 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video shot from a first-person perspective showing the character's hands?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the character's hands grasping an ancient longsword with dark leather wrapped around the hilt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the character perform a flowing sword dance involving lunges, spins, and stretches?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the setting a precipitous mountain with jagged rocks, thick flowing clouds, and distant indigo-hued mountains?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera pan smoothly in sync with the body's movements?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.011080636642873287 + ], + "opposite_logprob_tokens": [ + -4.511080741882324 + ], + "logprob_reward": -0.011080636642873287, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.011080636642873287, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12429188191890717, + 0.05639954283833504, + 0.09555570036172867, + 0.08347605913877487, + 0.0832207053899765, + 0.18464171886444092, + 0.08966893702745438, + 0.042511794716119766, + 0.08817604184150696, + 0.15205764770507812 + ], + "credit_map": [ + 0.010453701950609684, + 0.03305644541978836, + 0.020131351426243782, + 0.020684896036982536, + 0.009447675198316574, + 0.008747153915464878, + 0.011265797540545464, + 0.010504858568310738, + 0.005054336041212082, + 0.012324721552431583, + 0.0051189144141972065, + 0.008803543634712696, + 0.008383982814848423, + 0.007465817965567112, + 0.0051909019239246845, + 0.004057320766150951, + 0.00732226250693202, + 0.015804922208189964, + 0.00976918637752533, + 0.016256537288427353, + 0.011249407194554806, + 0.016954610124230385, + 0.0070871938951313496, + 0.011111583560705185, + 0.006538899149745703, + 0.013048729859292507, + 0.009017042815685272, + 0.008595761843025684, + 0.008683428168296814, + 0.019817523658275604, + 0.008650357834994793, + 0.009124312549829483, + 0.011219202540814877, + 0.010517231188714504, + 0.017867999151349068, + 0.011241452768445015, + 0.006715746130794287, + 0.00965788122266531, + 0.0073645762167871, + 0.008636613376438618, + 0.022672325372695923, + 0.043606385588645935, + 0.01659785769879818, + 0.016243018209934235, + 0.013150012120604515, + 0.03486018627882004, + 0.022485291585326195, + 0.015026634559035301, + 0.006270920392125845, + 0.01158931665122509, + 0.01300467737019062, + 0.011725876480340958, + 0.014964144676923752, + 0.011120228096842766, + 0.014183346182107925, + 0.006810424383729696, + 0.0028092542197555304, + 0.0028501441702246666, + 0.003802751423791051, + 0.003677363507449627, + 0.0059664021246135235, + 0.01120289508253336, + 0.0031961118802428246, + 0.009006872773170471, + 0.008674189448356628, + 0.007677207235246897, + 0.008627263829112053, + 0.010937746614217758, + 0.009818051941692829, + 0.014739204198122025, + 0.01357074175029993, + 0.0141316382214427, + 0.012035354971885681, + 0.018548816442489624, + 0.020217014476656914, + 0.019538339227437973, + 0.022047823294997215, + 0.02034987136721611, + 0.024219589307904243, + 0.015100828371942043 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 9, + 0 + ], + "selected_reward": -0.011080636642873287 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video shot from a first-person perspective showing the character's hands?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the character's hands grasping an ancient longsword with dark leather wrapped around the hilt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the character perform a flowing sword dance involving lunges, spins, and stretches?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the setting a precipitous mountain with jagged rocks, thick flowing clouds, and distant indigo-hued mountains?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera pan smoothly in sync with the body's movements?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.00525094848126173 + ], + "opposite_logprob_tokens": [ + -5.255250930786133 + ], + "logprob_reward": -0.00525094848126173, + "yes_no_margin_reward": 5.25, + "token_credit_reward": -0.00525094848126173, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15858851373195648, + 0.1793311983346939, + 0.10057488828897476, + 0.03117220662534237, + 0.07026369869709015, + 0.04496774822473526, + 0.18691198527812958, + 0.11348119378089905, + 0.048479679971933365, + 0.06622884422540665 + ], + "credit_map": [ + 0.012203127145767212, + 0.038893572986125946, + 0.02332061342895031, + 0.029191531240940094, + 0.014517804607748985, + 0.012134876102209091, + 0.013622372411191463, + 0.014704600907862186, + 0.017869578674435616, + 0.03249712660908699, + 0.017142435535788536, + 0.027246877551078796, + 0.021387580782175064, + 0.020071182399988174, + 0.026237986981868744, + 0.01687842607498169, + 0.009601256810128689, + 0.022825807332992554, + 0.01244030799716711, + 0.012653384357690811, + 0.0102795185521245, + 0.019760562106966972, + 0.005980617832392454, + 0.007033428642898798, + 0.0027752637397497892, + 0.00383106735534966, + 0.003588954219594598, + 0.004050806164741516, + 0.003900183830410242, + 0.005807679612189531, + 0.0031746195163577795, + 0.00404363265261054, + 0.011215885169804096, + 0.010755669325590134, + 0.012592539191246033, + 0.007521255407482386, + 0.004793515894562006, + 0.009917532093822956, + 0.005037286318838596, + 0.008430018089711666, + 0.00885213166475296, + 0.006193188950419426, + 0.0092318095266819, + 0.0038074820768088102, + 0.0034741198178380728, + 0.004787822254002094, + 0.0040443893522024155, + 0.004576807841658592, + 0.014019593596458435, + 0.02683398500084877, + 0.017628930509090424, + 0.02836017869412899, + 0.02845355123281479, + 0.02282680571079254, + 0.02766750566661358, + 0.021121442317962646, + 0.004133009817451239, + 0.0051323166117072105, + 0.006534365937113762, + 0.004822919145226479, + 0.008631987497210503, + 0.04123092442750931, + 0.006242887582629919, + 0.03675277531147003, + 0.003994925878942013, + 0.0042598797008395195, + 0.0034560784697532654, + 0.006600130815058947, + 0.006752862129360437, + 0.010316002182662487, + 0.007111864630132914, + 0.005987935699522495, + 0.004748818464577198, + 0.009109089151024818, + 0.005476922262459993, + 0.008211168460547924, + 0.009276085533201694, + 0.010842522606253624, + 0.010808971710503101, + 0.007755267899483442 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 1, + 0 + ], + "selected_reward": -0.00525094848126173 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video shot from a first-person perspective showing the character's hands?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the character's hands grasping an ancient longsword with dark leather wrapped around the hilt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the character perform a flowing sword dance involving lunges, spins, and stretches?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the setting a precipitous mountain with jagged rocks, thick flowing clouds, and distant indigo-hued mountains?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera pan smoothly in sync with the body's movements?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.018181685358285904 + ], + "opposite_logprob_tokens": [ + -4.018181800842285 + ], + "logprob_reward": -0.018181685358285904, + "yes_no_margin_reward": 4.0, + "token_credit_reward": -0.018181685358285904, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08620817214250565, + 0.05770306661725044, + 0.10627877712249756, + 0.04551906883716583, + 0.09533517807722092, + 0.29232358932495117, + 0.0727580115199089, + 0.1534256488084793, + 0.03363875672221184, + 0.05680973082780838 + ], + "credit_map": [ + 0.009025518782436848, + 0.013872895389795303, + 0.012787935324013233, + 0.012933391146361828, + 0.009823267348110676, + 0.007671434432268143, + 0.009725082665681839, + 0.01036864798516035, + 0.005281379446387291, + 0.012005741707980633, + 0.006948066409677267, + 0.008393727242946625, + 0.008089859038591385, + 0.006174901034682989, + 0.005925368517637253, + 0.004884024616330862, + 0.010419633239507675, + 0.0196272861212492, + 0.01446348987519741, + 0.01829744130373001, + 0.010859435424208641, + 0.016990210860967636, + 0.009511349722743034, + 0.006109934765845537, + 0.005193484015762806, + 0.005668510217219591, + 0.007169602904468775, + 0.005924283992499113, + 0.004441550932824612, + 0.005824409890919924, + 0.0060620857402682304, + 0.005235141143202782, + 0.017285319045186043, + 0.01563776284456253, + 0.015641270205378532, + 0.011665573343634605, + 0.005770327989012003, + 0.012807980179786682, + 0.006200707517564297, + 0.010326236486434937, + 0.04272116720676422, + 0.07854198664426804, + 0.040180329233407974, + 0.014920152723789215, + 0.017414020374417305, + 0.06303326785564423, + 0.014609363861382008, + 0.02090328373014927, + 0.006320421118289232, + 0.008239565417170525, + 0.009835774078965187, + 0.005962328054010868, + 0.01133777666836977, + 0.00918168667703867, + 0.011041417717933655, + 0.010839033871889114, + 0.007065578829497099, + 0.011104507371783257, + 0.009615251794457436, + 0.00967128574848175, + 0.01844204030930996, + 0.04634435102343559, + 0.008748565800487995, + 0.042434077709913254, + 0.003422257024794817, + 0.002765267388895154, + 0.004302522167563438, + 0.004785791039466858, + 0.004188129212707281, + 0.004647697787731886, + 0.0053977956995368, + 0.004129292909055948, + 0.00458170659840107, + 0.006630764342844486, + 0.0057726153172552586, + 0.0058812955394387245, + 0.007990412414073944, + 0.009316008538007736, + 0.010757886804640293, + 0.0058790394105017185 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 7, + 2 + ], + "selected_reward": -0.018181685358285904 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "An immersive first-person perspective, the camera pans and pans smoothly with the body's movements. Within the frame, a pair of slender, powerful hands tightly grasp an ancient longsword. The dark leather wrapped around the hilt has a warm patina from ageless wear. The slender blade gleams with a cold, restrained metallic luster under the soft, diffuse light filtering through the mist. You are perched atop a precipitous mountain, the jagged rock beneath your feet, surrounded by thick, flowing clouds, which gather and dissipate, occasionally revealing the indigo-hued outlines of the distant mountains, like an ink painting. With a deep breath, you raise the sword with steady composure, launching into a flowing sword dance. Each swing is precise and powerful, the tip of the sword tracing graceful, sharp silver arcs through the air, emitting a soft, sharp whistling sound as it slices through the mist. Your body lunges, spins, and stretches, its wide white linen sleeves fluttering in the air, seemingly merging with the flowing mist. The entire movement is seamless and rhythmic, each stretch awakening dormant ancient veins, imbued with a fusion of focus, tranquility, and strength. The surrounding landscape, the twisting ancient pines, the shifting mist, and the trajectory of your body and blade are all composed of elegant, flowing lines. Against a predominant palette of cool gray and indigo, they create a dynamic, solitary, and profound artistic conception.", + "seed": 26690, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000095/group_5/index_4168_seed_26690_dp_10_rank_20_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 5, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.026695015281438826, + "vlm_logprob": -0.026695015281438826, + "vlm_yes_no_margin": 4.05, + "vlm_token_credit": -0.026695015281438826, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video shot from a first-person perspective showing the character's hands?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the character's hands grasping an ancient longsword with dark leather wrapped around the hilt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the character perform a flowing sword dance involving lunges, spins, and stretches?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the setting a precipitous mountain with jagged rocks, thick flowing clouds, and distant indigo-hued mountains?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera pan smoothly in sync with the body's movements?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.021819883957505226 + ], + "opposite_logprob_tokens": [ + -4.021820068359375 + ], + "logprob_reward": -0.021819883957505226, + "yes_no_margin_reward": 4.0, + "token_credit_reward": -0.021819883957505226, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.061531294137239456, + 0.6327612996101379, + 0.1429154872894287, + 0.012053325772285461, + 0.024995997548103333, + 0.020905813202261925, + 0.008004453033208847, + 0.012414340861141682, + 0.022685810923576355, + 0.06173219904303551 + ], + "credit_map": [ + 0.009096362628042698, + 0.008172698318958282, + 0.006460475269705057, + 0.005653106141835451, + 0.009105868637561798, + 0.007535786367952824, + 0.008901722729206085, + 0.006605273578315973, + 0.016012929379940033, + 0.024914858862757683, + 0.010531165637075901, + 0.018715841695666313, + 0.2730165421962738, + 0.25743043422698975, + 0.018123365938663483, + 0.01401615422219038, + 0.020024020224809647, + 0.0232058335095644, + 0.015729563310742378, + 0.013682984746992588, + 0.018479442223906517, + 0.02235373482108116, + 0.012552028521895409, + 0.016887875273823738, + 0.0013565697008743882, + 0.0012575637083500624, + 0.0016052562277764082, + 0.00183065515011549, + 0.001636415719985962, + 0.0015690861037001014, + 0.001639967318624258, + 0.0011578104458749294, + 0.0018960002344101667, + 0.0033071045763790607, + 0.002403068821877241, + 0.0017670351080596447, + 0.005060107447206974, + 0.0053272005170583725, + 0.002599321072921157, + 0.0026361614000052214, + 0.0019699118565768003, + 0.001078034401871264, + 0.0013887063832953572, + 0.0012304766569286585, + 0.001914417021907866, + 0.001989461947232485, + 0.006563315633684397, + 0.004771490581333637, + 0.0007975144544616342, + 0.0008630374795757234, + 0.0009408031473867595, + 0.0016143115935847163, + 0.0011834593024104834, + 0.0007935152971185744, + 0.00083212397294119, + 0.0009796874364838004, + 0.0010045618982985616, + 0.0010220075491815805, + 0.0011955287773162127, + 0.0017192906234413385, + 0.0015014726668596268, + 0.0015806442825123668, + 0.0023316482547670603, + 0.0020591863431036472, + 0.001495135365985334, + 0.00332773313857615, + 0.0015660799108445644, + 0.0035907267592847347, + 0.0027754520997405052, + 0.002720048651099205, + 0.004572256933897734, + 0.002638377482071519, + 0.005345609970390797, + 0.009417709894478321, + 0.0062937322072684765, + 0.011749560944736004, + 0.006652673706412315, + 0.005999787710607052, + 0.010809843428432941, + 0.00546328304335475 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 2, + 9 + ], + "selected_reward": -0.021819883957505226 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video shot from a first-person perspective showing the character's hands?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the character's hands grasping an ancient longsword with dark leather wrapped around the hilt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the character perform a flowing sword dance involving lunges, spins, and stretches?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the setting a precipitous mountain with jagged rocks, thick flowing clouds, and distant indigo-hued mountains?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera pan smoothly in sync with the body's movements?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07897501438856125 + ], + "opposite_logprob_tokens": [ + -2.578974962234497 + ], + "logprob_reward": -0.07897501438856125, + "yes_no_margin_reward": 2.5, + "token_credit_reward": -0.07897501438856125, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06445848196744919, + 0.46295294165611267, + 0.07379617542028427, + 0.0988740622997284, + 0.040944598615169525, + 0.08335456252098083, + 0.01859578676521778, + 0.018658151850104332, + 0.05186882242560387, + 0.08649636805057526 + ], + "credit_map": [ + 0.007523823529481888, + 0.006243910640478134, + 0.006239080335944891, + 0.007343066390603781, + 0.009432043880224228, + 0.009981608018279076, + 0.009789534844458103, + 0.007905417121946812, + 0.015001187101006508, + 0.019917823374271393, + 0.0074677783995866776, + 0.01390357967466116, + 0.19206953048706055, + 0.1771814525127411, + 0.01910325698554516, + 0.018308358266949654, + 0.006908036768436432, + 0.009835930541157722, + 0.006306227762252092, + 0.009058759547770023, + 0.007629270665347576, + 0.012318741530179977, + 0.008749939501285553, + 0.012989276088774204, + 0.007542290724813938, + 0.01018900889903307, + 0.014423615299165249, + 0.01807270012795925, + 0.013136358000338078, + 0.013346453197300434, + 0.01331310998648405, + 0.008850525133311749, + 0.0030235317535698414, + 0.006582675501704216, + 0.004031041171401739, + 0.003225120482966304, + 0.008761073462665081, + 0.006962181068956852, + 0.004121568985283375, + 0.004237405024468899, + 0.00805300660431385, + 0.0053092618472874165, + 0.006850965786725283, + 0.00771822826936841, + 0.006252465303987265, + 0.006952905096113682, + 0.024032017216086388, + 0.01818571612238884, + 0.0020051056053489447, + 0.0023637227714061737, + 0.002914927201345563, + 0.0026327832601964474, + 0.0022375506814569235, + 0.002250904683023691, + 0.0017818126361817122, + 0.00240897829644382, + 0.0012584225041791797, + 0.0019704385194927454, + 0.001953147118911147, + 0.0029936260543763638, + 0.0016195724019780755, + 0.0020745890215039253, + 0.003559485310688615, + 0.0032288695219904184, + 0.003664022544398904, + 0.009827641770243645, + 0.003524437081068754, + 0.005398513749241829, + 0.008903268724679947, + 0.0051527912728488445, + 0.01001761294901371, + 0.005380532704293728, + 0.0068415128625929356, + 0.014028121717274189, + 0.007751731667667627, + 0.013559558428823948, + 0.00872205663472414, + 0.01245129108428955, + 0.013241434469819069, + 0.009900664910674095 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 3, + 9 + ], + "selected_reward": -0.07897501438856125 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video shot from a first-person perspective showing the character's hands?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the character's hands grasping an ancient longsword with dark leather wrapped around the hilt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the character perform a flowing sword dance involving lunges, spins, and stretches?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the setting a precipitous mountain with jagged rocks, thick flowing clouds, and distant indigo-hued mountains?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera pan smoothly in sync with the body's movements?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.012543476186692715 + ], + "opposite_logprob_tokens": [ + -4.387543678283691 + ], + "logprob_reward": -0.012543476186692715, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.012543476186692715, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08282548189163208, + 0.39870062470436096, + 0.13136924803256989, + 0.08207301795482635, + 0.044529832899570465, + 0.10527176409959793, + 0.023990757763385773, + 0.021636251360177994, + 0.06025008484721184, + 0.04935295507311821 + ], + "credit_map": [ + 0.009962573647499084, + 0.007958409376442432, + 0.008558373898267746, + 0.009318331256508827, + 0.010537467896938324, + 0.009343297220766544, + 0.014171448536217213, + 0.012975575402379036, + 0.009547783061861992, + 0.016354383900761604, + 0.005367278587073088, + 0.008686295710504055, + 0.17640016973018646, + 0.16202032566070557, + 0.01236532349139452, + 0.007959061302244663, + 0.01604527048766613, + 0.023588508367538452, + 0.01227637194097042, + 0.011512861587107182, + 0.02048296108841896, + 0.025352967903017998, + 0.010008614510297775, + 0.012101695872843266, + 0.0075296200811862946, + 0.007967553101480007, + 0.01040128618478775, + 0.016992272809147835, + 0.0099311713129282, + 0.01129155419766903, + 0.010156906209886074, + 0.007802651729434729, + 0.00353582133539021, + 0.006417825352400541, + 0.0044489032588899136, + 0.003814799478277564, + 0.010427252389490604, + 0.007419238798320293, + 0.004933806601911783, + 0.003532186383381486, + 0.00870021153241396, + 0.006367824040353298, + 0.007666429970413446, + 0.007581029552966356, + 0.007623381447046995, + 0.009058925323188305, + 0.03361279517412186, + 0.024661168456077576, + 0.0019059526966884732, + 0.003012267407029867, + 0.002360628917813301, + 0.002737100701779127, + 0.002688537584617734, + 0.003972447477281094, + 0.0027101344894617796, + 0.004603687673807144, + 0.001678399508818984, + 0.002378797624260187, + 0.0021039070561528206, + 0.0032364060170948505, + 0.0025105474051088095, + 0.0024721340741962194, + 0.004467618186026812, + 0.0027884424198418856, + 0.003396976040676236, + 0.008423407562077045, + 0.004275709856301546, + 0.010430216789245605, + 0.007301642559468746, + 0.005930842831730843, + 0.012457548640668392, + 0.008033739402890205, + 0.003569649066776037, + 0.0068977647460997105, + 0.004318784456700087, + 0.01017221063375473, + 0.004670904483646154, + 0.007247724570333958, + 0.006812693551182747, + 0.005663225892931223 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 2, + 5 + ], + "selected_reward": -0.012543476186692715 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video shot from a first-person perspective showing the character's hands?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the character's hands grasping an ancient longsword with dark leather wrapped around the hilt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the character perform a flowing sword dance involving lunges, spins, and stretches?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the setting a precipitous mountain with jagged rocks, thick flowing clouds, and distant indigo-hued mountains?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera pan smoothly in sync with the body's movements?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005947748199105263 + ], + "opposite_logprob_tokens": [ + -5.130947589874268 + ], + "logprob_reward": -0.005947748199105263, + "yes_no_margin_reward": 5.125, + "token_credit_reward": -0.005947748199105263, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14611665904521942, + 0.06418468803167343, + 0.09400109201669693, + 0.03476129099726677, + 0.2560616135597229, + 0.08800666779279709, + 0.15571632981300354, + 0.07213759422302246, + 0.027660077437758446, + 0.06135406717658043 + ], + "credit_map": [ + 0.02116658352315426, + 0.015895716845989227, + 0.017241664230823517, + 0.011405586265027523, + 0.024516329169273376, + 0.020391833037137985, + 0.021365273743867874, + 0.014133664779365063, + 0.0022073921281844378, + 0.0035246785264462233, + 0.0025323324371129274, + 0.00416979007422924, + 0.02380741946399212, + 0.02163563296198845, + 0.0038252139929682016, + 0.002482221694663167, + 0.013339104130864143, + 0.014578566886484623, + 0.00929478369653225, + 0.009464205242693424, + 0.011970305815339088, + 0.014703852124512196, + 0.010550613515079021, + 0.010099655948579311, + 0.003860845696181059, + 0.003845046041533351, + 0.004637250676751137, + 0.004585300572216511, + 0.0035814058501273394, + 0.005657348781824112, + 0.0033512860536575317, + 0.0052428049966692924, + 0.018703410401940346, + 0.03620346635580063, + 0.025591185316443443, + 0.023325955495238304, + 0.05178532376885414, + 0.04720636457204819, + 0.02779768407344818, + 0.02544819377362728, + 0.008028256706893444, + 0.006047659087926149, + 0.006584785878658295, + 0.0076455832459032536, + 0.009278491139411926, + 0.011829179711639881, + 0.024766718968749046, + 0.013825984671711922, + 0.012434786185622215, + 0.014019578695297241, + 0.02509094960987568, + 0.02706056833267212, + 0.011817198246717453, + 0.01659660041332245, + 0.01684875413775444, + 0.03184787184000015, + 0.00592097919434309, + 0.006025331094861031, + 0.0062632933259010315, + 0.009682867676019669, + 0.008296245709061623, + 0.011984775774180889, + 0.012931311503052711, + 0.011032775975763798, + 0.0020004648249596357, + 0.00611363910138607, + 0.0025025925133377314, + 0.002712814835831523, + 0.003198703983798623, + 0.002337393118068576, + 0.006486475467681885, + 0.0023079898674041033, + 0.004995883908122778, + 0.010604551993310452, + 0.005204849876463413, + 0.010205301456153393, + 0.00891620758920908, + 0.007708180230110884, + 0.007842844352126122, + 0.005876245442777872 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 6, + 0 + ], + "selected_reward": -0.005947748199105263 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video shot from a first-person perspective showing the character's hands?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the character's hands grasping an ancient longsword with dark leather wrapped around the hilt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the character perform a flowing sword dance involving lunges, spins, and stretches?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the setting a precipitous mountain with jagged rocks, thick flowing clouds, and distant indigo-hued mountains?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera pan smoothly in sync with the body's movements?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.014188953675329685 + ], + "opposite_logprob_tokens": [ + -4.264188766479492 + ], + "logprob_reward": -0.014188953675329685, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.014188953675329685, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.17849905788898468, + 0.3088649809360504, + 0.1441984623670578, + 0.05246760696172714, + 0.051666952669620514, + 0.0740480050444603, + 0.04035356268286705, + 0.050179373472929, + 0.039865899831056595, + 0.059856116771698 + ], + "credit_map": [ + 0.017901338636875153, + 0.015311026014387608, + 0.020754847675561905, + 0.01565961167216301, + 0.031069111078977585, + 0.028708714991807938, + 0.028168221935629845, + 0.020926179364323616, + 0.007831462658941746, + 0.013129112310707569, + 0.0037048112135380507, + 0.007085464894771576, + 0.13499847054481506, + 0.124953493475914, + 0.009659930132329464, + 0.007502247113734484, + 0.017710665240883827, + 0.024408873170614243, + 0.013197490945458412, + 0.013928466476500034, + 0.02060428448021412, + 0.027497945353388786, + 0.01316744089126587, + 0.013683292083442211, + 0.004240203648805618, + 0.0048116953112185, + 0.00712934834882617, + 0.010337294079363346, + 0.006327200680971146, + 0.008212115615606308, + 0.004668364301323891, + 0.006741383578628302, + 0.004052684176713228, + 0.006477602757513523, + 0.005017477087676525, + 0.00438203290104866, + 0.009603817947208881, + 0.009695400483906269, + 0.0051891617476940155, + 0.007248773239552975, + 0.006056313868612051, + 0.00533547718077898, + 0.005059479270130396, + 0.006173023022711277, + 0.007165123708546162, + 0.008004353381693363, + 0.020977580919861794, + 0.0152766527608037, + 0.004255840089172125, + 0.0037208390422165394, + 0.004880621563643217, + 0.007531596813350916, + 0.005393350962549448, + 0.005486847832798958, + 0.003953946754336357, + 0.005130520090460777, + 0.00408779876306653, + 0.004790560342371464, + 0.00533140916377306, + 0.008355685509741306, + 0.005871909204870462, + 0.005125791300088167, + 0.00889157596975565, + 0.0077246432192623615, + 0.0029868739657104015, + 0.005199823062866926, + 0.0038692678790539503, + 0.006479061208665371, + 0.004980084020644426, + 0.004531865008175373, + 0.006766606122255325, + 0.005052318796515465, + 0.0048967828042805195, + 0.010890483856201172, + 0.00647524232044816, + 0.007793002296239138, + 0.005901050288230181, + 0.008952287957072258, + 0.008926169946789742, + 0.006021098233759403 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 2 + ], + "selected_reward": -0.014188953675329685 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "An immersive first-person perspective, the camera pans and pans smoothly with the body's movements. Within the frame, a pair of slender, powerful hands tightly grasp an ancient longsword. The dark leather wrapped around the hilt has a warm patina from ageless wear. The slender blade gleams with a cold, restrained metallic luster under the soft, diffuse light filtering through the mist. You are perched atop a precipitous mountain, the jagged rock beneath your feet, surrounded by thick, flowing clouds, which gather and dissipate, occasionally revealing the indigo-hued outlines of the distant mountains, like an ink painting. With a deep breath, you raise the sword with steady composure, launching into a flowing sword dance. Each swing is precise and powerful, the tip of the sword tracing graceful, sharp silver arcs through the air, emitting a soft, sharp whistling sound as it slices through the mist. Your body lunges, spins, and stretches, its wide white linen sleeves fluttering in the air, seemingly merging with the flowing mist. The entire movement is seamless and rhythmic, each stretch awakening dormant ancient veins, imbued with a fusion of focus, tranquility, and strength. The surrounding landscape, the twisting ancient pines, the shifting mist, and the trajectory of your body and blade are all composed of elegant, flowing lines. Against a predominant palette of cool gray and indigo, they create a dynamic, solitary, and profound artistic conception.", + "seed": 26691, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000095/group_5/index_4168_seed_26691_dp_11_rank_22_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 5, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.06198788797482848, + "vlm_logprob": -0.06198788797482848, + "vlm_yes_no_margin": 3.95, + "vlm_token_credit": -0.06198788797482848, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video shot from a first-person perspective showing the character's hands?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the character's hands grasping an ancient longsword with dark leather wrapped around the hilt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the character perform a flowing sword dance involving lunges, spins, and stretches?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the setting a precipitous mountain with jagged rocks, thick flowing clouds, and distant indigo-hued mountains?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera pan smoothly in sync with the body's movements?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.23472601175308228 + ], + "opposite_logprob_tokens": [ + -1.6097259521484375 + ], + "logprob_reward": -0.23472601175308228, + "yes_no_margin_reward": 1.375, + "token_credit_reward": -0.23472601175308228, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07630840688943863, + 0.05878274142742157, + 0.05095489323139191, + 0.06724761426448822, + 0.033539313822984695, + 0.04770124331116676, + 0.12484712153673172, + 0.21518868207931519, + 0.12202072143554688, + 0.20340928435325623 + ], + "credit_map": [ + 0.00938237551599741, + 0.008883255533874035, + 0.008240270428359509, + 0.010875480249524117, + 0.011778157204389572, + 0.008718682453036308, + 0.010636845603585243, + 0.007793341763317585, + 0.010306098498404026, + 0.0073638129979372025, + 0.007763704750686884, + 0.006270426791161299, + 0.005070099141448736, + 0.007395890075713396, + 0.009085006080567837, + 0.005527707282453775, + 0.006223136559128761, + 0.00805873516947031, + 0.005130473058670759, + 0.008616961538791656, + 0.006454657297581434, + 0.005009892396628857, + 0.005289961118251085, + 0.006171075161546469, + 0.004981942940503359, + 0.006908987648785114, + 0.007556873373687267, + 0.007362655829638243, + 0.007352289743721485, + 0.009813982993364334, + 0.008689106442034245, + 0.014581768773496151, + 0.004560444969683886, + 0.004248170182108879, + 0.002769016195088625, + 0.00482635572552681, + 0.004630763083696365, + 0.003931122832000256, + 0.005208426155149937, + 0.003365012351423502, + 0.008054767735302448, + 0.00540398433804512, + 0.005279844161123037, + 0.005171995609998703, + 0.006860644090920687, + 0.004085049964487553, + 0.00831588450819254, + 0.004529072903096676, + 0.010241149924695492, + 0.011177035048604012, + 0.014731007628142834, + 0.013297454454004765, + 0.01310055609792471, + 0.01442672312259674, + 0.01998239755630493, + 0.027890799567103386, + 0.022847754880785942, + 0.04773593321442604, + 0.02892041578888893, + 0.022712789475917816, + 0.026742927730083466, + 0.02000306360423565, + 0.02362421527504921, + 0.022601578384637833, + 0.012915056198835373, + 0.010408011265099049, + 0.015549755655229092, + 0.00832430087029934, + 0.012276962399482727, + 0.016926081851124763, + 0.022120844572782516, + 0.02349971793591976, + 0.015952061861753464, + 0.028674770146608353, + 0.03718411922454834, + 0.023101888597011566, + 0.021770920604467392, + 0.01434272900223732, + 0.033151742070913315, + 0.029231049120426178 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 9, + 6 + ], + "selected_reward": -0.23472601175308228 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video shot from a first-person perspective showing the character's hands?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the character's hands grasping an ancient longsword with dark leather wrapped around the hilt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the character perform a flowing sword dance involving lunges, spins, and stretches?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the setting a precipitous mountain with jagged rocks, thick flowing clouds, and distant indigo-hued mountains?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera pan smoothly in sync with the body's movements?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.054967161267995834 + ], + "opposite_logprob_tokens": [ + -2.929967164993286 + ], + "logprob_reward": -0.054967161267995834, + "yes_no_margin_reward": 2.875, + "token_credit_reward": -0.054967161267995834, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07370463013648987, + 0.13892850279808044, + 0.06636344641447067, + 0.0549653023481369, + 0.050730347633361816, + 0.05489657074213028, + 0.15776380896568298, + 0.22913801670074463, + 0.06015244126319885, + 0.11335689574480057 + ], + "credit_map": [ + 0.010702712461352348, + 0.008479329757392406, + 0.007106855511665344, + 0.01089286245405674, + 0.008622781373560429, + 0.010456507094204426, + 0.009886538609862328, + 0.007557043340057135, + 0.010623502545058727, + 0.016747089102864265, + 0.009903615340590477, + 0.009317032061517239, + 0.008582980372011662, + 0.0388902984559536, + 0.008021126501262188, + 0.03684286028146744, + 0.00845112930983305, + 0.013374457135796547, + 0.006767562124878168, + 0.010987832210958004, + 0.008762902580201626, + 0.00614888733252883, + 0.005738471169024706, + 0.006132200360298157, + 0.005756814498454332, + 0.008534775115549564, + 0.0065468912944197655, + 0.005482810083776712, + 0.006611202843487263, + 0.009794264100492, + 0.0057653686963021755, + 0.006473171059042215, + 0.005774728953838348, + 0.004838664550334215, + 0.006332355085760355, + 0.006978047080338001, + 0.005973476450890303, + 0.006412525661289692, + 0.008564213290810585, + 0.005856337491422892, + 0.005619213916361332, + 0.007174297701567411, + 0.003681415691971779, + 0.003671053098514676, + 0.009044486097991467, + 0.005091623868793249, + 0.01263746339827776, + 0.00797701720148325, + 0.013058872893452644, + 0.019905339926481247, + 0.02301584556698799, + 0.029991254210472107, + 0.02360657788813114, + 0.015480607748031616, + 0.018876483663916588, + 0.013828824274241924, + 0.021641502156853676, + 0.04394731670618057, + 0.04204173386096954, + 0.02859223634004593, + 0.023589378222823143, + 0.02672683820128441, + 0.022301023826003075, + 0.02029799111187458, + 0.007502434309571981, + 0.00635394174605608, + 0.008234013803303242, + 0.0050634536892175674, + 0.006422887556254864, + 0.009060761891305447, + 0.008856783621013165, + 0.008658166974782944, + 0.009245792403817177, + 0.014474685303866863, + 0.025630297139286995, + 0.01303207129240036, + 0.013665618374943733, + 0.010740593075752258, + 0.013286950066685677, + 0.013280888088047504 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 6, + 1 + ], + "selected_reward": -0.054967161267995834 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video shot from a first-person perspective showing the character's hands?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the character's hands grasping an ancient longsword with dark leather wrapped around the hilt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the character perform a flowing sword dance involving lunges, spins, and stretches?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the setting a precipitous mountain with jagged rocks, thick flowing clouds, and distant indigo-hued mountains?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera pan smoothly in sync with the body's movements?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0036202622577548027 + ], + "opposite_logprob_tokens": [ + -5.628620147705078 + ], + "logprob_reward": -0.0036202622577548027, + "yes_no_margin_reward": 5.625, + "token_credit_reward": -0.0036202622577548027, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0796896442770958, + 0.13551492989063263, + 0.06391995400190353, + 0.12877656519412994, + 0.040907010436058044, + 0.03518021106719971, + 0.1180717870593071, + 0.2716521620750427, + 0.04988216608762741, + 0.07640548795461655 + ], + "credit_map": [ + 0.009597589261829853, + 0.010780278593301773, + 0.00823102705180645, + 0.012814034707844257, + 0.01292977575212717, + 0.008760049007833004, + 0.007774601224809885, + 0.008802298456430435, + 0.015844501554965973, + 0.012430577538907528, + 0.007971206679940224, + 0.012002372182905674, + 0.008603192865848541, + 0.03739612177014351, + 0.0102738868445158, + 0.030993077903985977, + 0.005242161452770233, + 0.008876999840140343, + 0.007844019681215286, + 0.009880470111966133, + 0.010405662469565868, + 0.007212326396256685, + 0.006371316034346819, + 0.008087006397545338, + 0.007296315394341946, + 0.01613215170800686, + 0.010309881530702114, + 0.008230726234614849, + 0.009691878221929073, + 0.037371017038822174, + 0.007620050106197596, + 0.032124556601047516, + 0.005009862594306469, + 0.004656479228287935, + 0.00328904134221375, + 0.008199017494916916, + 0.004928953014314175, + 0.00460842065513134, + 0.006130848545581102, + 0.004084393847733736, + 0.005086041521281004, + 0.0039299470372498035, + 0.004981535486876965, + 0.0034132676664739847, + 0.005164797883480787, + 0.0029585331212729216, + 0.005362389609217644, + 0.004283703397959471, + 0.009884790517389774, + 0.012086108326911926, + 0.02085150219500065, + 0.019063903018832207, + 0.01305328868329525, + 0.01301727257668972, + 0.010509259067475796, + 0.019605673849582672, + 0.028431255370378494, + 0.03739672154188156, + 0.03643570467829704, + 0.03759383782744408, + 0.03165009990334511, + 0.02867760695517063, + 0.03634893521666527, + 0.03511800616979599, + 0.0035633593797683716, + 0.0088801933452487, + 0.006630931049585342, + 0.00503102783113718, + 0.005239737685769796, + 0.004496540874242783, + 0.008934028446674347, + 0.007106352597475052, + 0.005252174101769924, + 0.00821740087121725, + 0.009729147888720036, + 0.011995057575404644, + 0.013250035233795643, + 0.007431702222675085, + 0.012512020766735077, + 0.00801796093583107 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 1, + 3 + ], + "selected_reward": -0.0036202622577548027 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video shot from a first-person perspective showing the character's hands?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the character's hands grasping an ancient longsword with dark leather wrapped around the hilt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the character perform a flowing sword dance involving lunges, spins, and stretches?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the setting a precipitous mountain with jagged rocks, thick flowing clouds, and distant indigo-hued mountains?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera pan smoothly in sync with the body's movements?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004092770162969828 + ], + "opposite_logprob_tokens": [ + -5.504092693328857 + ], + "logprob_reward": -0.004092770162969828, + "yes_no_margin_reward": 5.5, + "token_credit_reward": -0.004092770162969828, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14425042271614075, + 0.07938270270824432, + 0.05245043337345123, + 0.20639832317829132, + 0.0816858559846878, + 0.11810605227947235, + 0.06091631203889847, + 0.08704960346221924, + 0.0725933238863945, + 0.09716696292161942 + ], + "credit_map": [ + 0.014684105291962624, + 0.016852013766765594, + 0.013415484689176083, + 0.021425914019346237, + 0.02183934859931469, + 0.019550157710909843, + 0.022002171725034714, + 0.014481223188340664, + 0.008226001635193825, + 0.008116657845675945, + 0.006174890790134668, + 0.0048000747337937355, + 0.006148061715066433, + 0.02017376199364662, + 0.005914795212447643, + 0.01982845924794674, + 0.004162420518696308, + 0.004429689608514309, + 0.007957620546221733, + 0.0076807173900306225, + 0.008140922524034977, + 0.0063393027521669865, + 0.004855292849242687, + 0.008884468115866184, + 0.012744304724037647, + 0.019992729648947716, + 0.014192077331244946, + 0.011189062148332596, + 0.013890804722905159, + 0.06564100086688995, + 0.007368925027549267, + 0.06137942522764206, + 0.010750416666269302, + 0.01250498741865158, + 0.007811747957020998, + 0.01115446537733078, + 0.011301604099571705, + 0.009012769907712936, + 0.008554007858037949, + 0.010595852509140968, + 0.013192012906074524, + 0.01524763461202383, + 0.012567021884024143, + 0.008665931411087513, + 0.011462949216365814, + 0.010936001315712929, + 0.02227010577917099, + 0.023764392361044884, + 0.007145291659981012, + 0.009441775269806385, + 0.008446147665381432, + 0.007847611792385578, + 0.007913777604699135, + 0.006416781805455685, + 0.00626390939578414, + 0.007441013585776091, + 0.012566676363348961, + 0.010181890800595284, + 0.018468284979462624, + 0.009301264770328999, + 0.00823933631181717, + 0.011967196129262447, + 0.007947981357574463, + 0.00837698020040989, + 0.007723246701061726, + 0.0119525957852602, + 0.009539126418530941, + 0.006733213551342487, + 0.008436203002929688, + 0.009926758706569672, + 0.01018975768238306, + 0.008092415519058704, + 0.00763958552852273, + 0.011639820411801338, + 0.010508479550480843, + 0.01505747064948082, + 0.011345447041094303, + 0.009194142185151577, + 0.018441153690218925, + 0.013340858742594719 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 0, + 5 + ], + "selected_reward": -0.004092770162969828 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video shot from a first-person perspective showing the character's hands?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the character's hands grasping an ancient longsword with dark leather wrapped around the hilt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the character perform a flowing sword dance involving lunges, spins, and stretches?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the setting a precipitous mountain with jagged rocks, thick flowing clouds, and distant indigo-hued mountains?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera pan smoothly in sync with the body's movements?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.012533234432339668 + ], + "opposite_logprob_tokens": [ + -4.387533187866211 + ], + "logprob_reward": -0.012533234432339668, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.012533234432339668, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12525855004787445, + 0.17111527919769287, + 0.04605990648269653, + 0.08131688088178635, + 0.04969347268342972, + 0.0425662100315094, + 0.11347419768571854, + 0.1941695511341095, + 0.1078077107667923, + 0.06853823363780975 + ], + "credit_map": [ + 0.010822439566254616, + 0.014479574747383595, + 0.012151597999036312, + 0.012807031162083149, + 0.029280463233590126, + 0.017090678215026855, + 0.015529219061136246, + 0.013097543269395828, + 0.010417972691357136, + 0.012069922871887684, + 0.008887503296136856, + 0.010227429680526257, + 0.013693587854504585, + 0.05302790552377701, + 0.01384301483631134, + 0.048947952687740326, + 0.004291377030313015, + 0.006129344459623098, + 0.004721307195723057, + 0.009367816150188446, + 0.006129168905317783, + 0.004448302090167999, + 0.00591311464086175, + 0.005059477407485247, + 0.00615500146523118, + 0.013102036900818348, + 0.007971623912453651, + 0.00953192450106144, + 0.008402838371694088, + 0.015772560611367226, + 0.007961185649037361, + 0.01241970993578434, + 0.0046290685422718525, + 0.006422471720725298, + 0.005164028611034155, + 0.007563351187855005, + 0.006666027940809727, + 0.006321797147393227, + 0.007297329604625702, + 0.005629399791359901, + 0.005558765958994627, + 0.005888726562261581, + 0.004412297625094652, + 0.004963541403412819, + 0.007023877929896116, + 0.0037624661345034838, + 0.006506376899778843, + 0.0044501605443656445, + 0.011147737503051758, + 0.015599800273776054, + 0.013687452301383018, + 0.012075508013367653, + 0.013853145763278008, + 0.012559820897877216, + 0.010624737478792667, + 0.02392599731683731, + 0.017214255407452583, + 0.021732911467552185, + 0.04710479453206062, + 0.024054573848843575, + 0.01576562225818634, + 0.02785881794989109, + 0.024633504450321198, + 0.015805067494511604, + 0.011482157744467258, + 0.012738864868879318, + 0.015291562303900719, + 0.010006411001086235, + 0.011666545644402504, + 0.01885446533560753, + 0.015759989619255066, + 0.012007714249193668, + 0.00530205387622118, + 0.008187908679246902, + 0.009371797554194927, + 0.009242565371096134, + 0.0075062476098537445, + 0.006369347684085369, + 0.012804795056581497, + 0.009753521531820297 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 1, + 0 + ], + "selected_reward": -0.012533234432339668 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "An immersive first-person perspective, the camera pans and pans smoothly with the body's movements. Within the frame, a pair of slender, powerful hands tightly grasp an ancient longsword. The dark leather wrapped around the hilt has a warm patina from ageless wear. The slender blade gleams with a cold, restrained metallic luster under the soft, diffuse light filtering through the mist. You are perched atop a precipitous mountain, the jagged rock beneath your feet, surrounded by thick, flowing clouds, which gather and dissipate, occasionally revealing the indigo-hued outlines of the distant mountains, like an ink painting. With a deep breath, you raise the sword with steady composure, launching into a flowing sword dance. Each swing is precise and powerful, the tip of the sword tracing graceful, sharp silver arcs through the air, emitting a soft, sharp whistling sound as it slices through the mist. Your body lunges, spins, and stretches, its wide white linen sleeves fluttering in the air, seemingly merging with the flowing mist. The entire movement is seamless and rhythmic, each stretch awakening dormant ancient veins, imbued with a fusion of focus, tranquility, and strength. The surrounding landscape, the twisting ancient pines, the shifting mist, and the trajectory of your body and blade are all composed of elegant, flowing lines. Against a predominant palette of cool gray and indigo, they create a dynamic, solitary, and profound artistic conception.", + "seed": 26692, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000095/group_5/index_4168_seed_26692_dp_11_rank_22_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 5, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.02331348955631256, + "vlm_logprob": -0.02331348955631256, + "vlm_yes_no_margin": 3.975, + "vlm_token_credit": -0.02331348955631256, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video shot from a first-person perspective showing the character's hands?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the character's hands grasping an ancient longsword with dark leather wrapped around the hilt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the character perform a flowing sword dance involving lunges, spins, and stretches?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the setting a precipitous mountain with jagged rocks, thick flowing clouds, and distant indigo-hued mountains?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera pan smoothly in sync with the body's movements?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.028555523604154587 + ], + "opposite_logprob_tokens": [ + -3.7785556316375732 + ], + "logprob_reward": -0.028555523604154587, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.028555523604154587, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12727151811122894, + 0.06259410083293915, + 0.3998301327228546, + 0.1475282460451126, + 0.028756706044077873, + 0.03470302373170853, + 0.10344161093235016, + 0.030209098011255264, + 0.03677818924188614, + 0.02888736128807068 + ], + "credit_map": [ + 0.0077704559080302715, + 0.013759339228272438, + 0.04273522272706032, + 0.01157242152839899, + 0.007808572147041559, + 0.007803906686604023, + 0.00977553054690361, + 0.026046056300401688, + 0.00711197592318058, + 0.005844175349920988, + 0.007747681811451912, + 0.006345456000417471, + 0.012162556871771812, + 0.007940594106912613, + 0.006145558785647154, + 0.009296099655330181, + 0.02595490962266922, + 0.028079928830266, + 0.07810802757740021, + 0.026395170018076897, + 0.030998865142464638, + 0.04865190014243126, + 0.06443534791469574, + 0.0972059816122055, + 0.009049504064023495, + 0.010140842758119106, + 0.02100243791937828, + 0.013026897795498371, + 0.015554043464362621, + 0.04215458035469055, + 0.016622396185994148, + 0.019977552816271782, + 0.0030614358838647604, + 0.0034277066588401794, + 0.0036751774605363607, + 0.003331308253109455, + 0.0036701352801173925, + 0.003703444031998515, + 0.004637179430574179, + 0.0032503181137144566, + 0.004588810261338949, + 0.003115213941782713, + 0.004364542197436094, + 0.005046261008828878, + 0.003456865204498172, + 0.004605620168149471, + 0.004831152502447367, + 0.004694557748734951, + 0.006022699177265167, + 0.014504277147352695, + 0.015999766066670418, + 0.009173986501991749, + 0.0181837510317564, + 0.01326718833297491, + 0.009312104433774948, + 0.016977839171886444, + 0.0025685483124107122, + 0.006057355087250471, + 0.003993825986981392, + 0.0035513422917574644, + 0.003572281915694475, + 0.0033053504303097725, + 0.0028126405086368322, + 0.004347754642367363, + 0.0027548568323254585, + 0.004478190094232559, + 0.006071549840271473, + 0.004642280749976635, + 0.005933903157711029, + 0.004285835660994053, + 0.004454030655324459, + 0.004157541319727898, + 0.002074492396786809, + 0.0053342003375291824, + 0.0018730907468125224, + 0.0018502359744161367, + 0.009963974356651306, + 0.004072986077517271, + 0.0016968903364613652, + 0.0020214940886944532 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 3, + 0 + ], + "selected_reward": -0.028555523604154587 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video shot from a first-person perspective showing the character's hands?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the character's hands grasping an ancient longsword with dark leather wrapped around the hilt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the character perform a flowing sword dance involving lunges, spins, and stretches?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the setting a precipitous mountain with jagged rocks, thick flowing clouds, and distant indigo-hued mountains?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera pan smoothly in sync with the body's movements?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04865364357829094 + ], + "opposite_logprob_tokens": [ + -3.0486536026000977 + ], + "logprob_reward": -0.04865364357829094, + "yes_no_margin_reward": 3.0, + "token_credit_reward": -0.04865364357829094, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.16143175959587097, + 0.13460002839565277, + 0.20162753760814667, + 0.07194367051124573, + 0.04719033092260361, + 0.17822150886058807, + 0.08274802565574646, + 0.03534964844584465, + 0.04568418860435486, + 0.04120328277349472 + ], + "credit_map": [ + 0.013035264797508717, + 0.015024712309241295, + 0.0456140898168087, + 0.01269066147506237, + 0.015774833038449287, + 0.014448043890297413, + 0.012553883716464043, + 0.03229028359055519, + 0.01521120686084032, + 0.010976634919643402, + 0.011802947148680687, + 0.01102888397872448, + 0.01733933947980404, + 0.022058626636862755, + 0.018672222271561623, + 0.027510181069374084, + 0.016643859446048737, + 0.016592448577284813, + 0.04409056156873703, + 0.013116609305143356, + 0.01620449312031269, + 0.02612200379371643, + 0.02795940451323986, + 0.04089818894863129, + 0.00419188104569912, + 0.008019551634788513, + 0.01101001352071762, + 0.009200280532240868, + 0.0074387495405972, + 0.014260539785027504, + 0.0073055485263466835, + 0.010517114773392677, + 0.004789240658283234, + 0.006042527500540018, + 0.007906707003712654, + 0.004060348961502314, + 0.0049837613478302956, + 0.008322685956954956, + 0.005521385930478573, + 0.005563677754253149, + 0.010993470437824726, + 0.013689878396689892, + 0.022422606125473976, + 0.024359844624996185, + 0.015705155208706856, + 0.03423392400145531, + 0.019237609580159187, + 0.037579040974378586, + 0.005130646750330925, + 0.013046232052147388, + 0.010048888623714447, + 0.007994351908564568, + 0.01104590855538845, + 0.012302836403250694, + 0.008852689526975155, + 0.014326482079923153, + 0.0030297322664409876, + 0.0060520414263010025, + 0.006231282372027636, + 0.0044210427440702915, + 0.0031878340523689985, + 0.003786256304010749, + 0.0039562322199344635, + 0.004685230553150177, + 0.0036680561024695635, + 0.005506230052560568, + 0.007435344159603119, + 0.006048235110938549, + 0.007852785289287567, + 0.0049569010734558105, + 0.005355975590646267, + 0.004860665183514357, + 0.003068722551688552, + 0.004509966354817152, + 0.0033291254658252, + 0.005386877804994583, + 0.009930158965289593, + 0.004507084842771292, + 0.005963003262877464, + 0.004508348647505045 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 5, + 0 + ], + "selected_reward": -0.04865364357829094 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video shot from a first-person perspective showing the character's hands?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the character's hands grasping an ancient longsword with dark leather wrapped around the hilt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the character perform a flowing sword dance involving lunges, spins, and stretches?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the setting a precipitous mountain with jagged rocks, thick flowing clouds, and distant indigo-hued mountains?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera pan smoothly in sync with the body's movements?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.008641703985631466 + ], + "opposite_logprob_tokens": [ + -4.758641719818115 + ], + "logprob_reward": -0.008641703985631466, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.008641703985631466, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15564563870429993, + 0.13495293259620667, + 0.04491249471902847, + 0.08455505222082138, + 0.06384458392858505, + 0.06844675540924072, + 0.2641078233718872, + 0.041559550911188126, + 0.0552746057510376, + 0.08670055121183395 + ], + "credit_map": [ + 0.01042442861944437, + 0.020767930895090103, + 0.04683795943856239, + 0.011312207207083702, + 0.015899119898676872, + 0.009892269968986511, + 0.011923160403966904, + 0.028588570654392242, + 0.010079910978674889, + 0.007597756572067738, + 0.00620571942999959, + 0.007031939458101988, + 0.011434446088969707, + 0.029059113934636116, + 0.02412279136478901, + 0.03942125290632248, + 0.004729215055704117, + 0.004075869917869568, + 0.006753634661436081, + 0.005262054968625307, + 0.0057329763658344746, + 0.007228126283735037, + 0.005252285860478878, + 0.005878329277038574, + 0.00845594983547926, + 0.00882893055677414, + 0.013384658843278885, + 0.008872988633811474, + 0.009251491166651249, + 0.018367918208241463, + 0.009103967808187008, + 0.00828914251178503, + 0.005114854779094458, + 0.00923119392246008, + 0.008786377497017384, + 0.0051347073167562485, + 0.00833065714687109, + 0.010839750990271568, + 0.008581947535276413, + 0.007825091481208801, + 0.006003580521792173, + 0.006433404050767422, + 0.010772056877613068, + 0.009292960166931152, + 0.007459059823304415, + 0.007466527633368969, + 0.010946650989353657, + 0.010072512552142143, + 0.014741835184395313, + 0.04384427145123482, + 0.03626283258199692, + 0.02487538382411003, + 0.03753180056810379, + 0.031933922320604324, + 0.033109210431575775, + 0.04180857166647911, + 0.0028346586041152477, + 0.006622994318604469, + 0.006090732291340828, + 0.006833417806774378, + 0.004534550476819277, + 0.004464723635464907, + 0.0049259355291724205, + 0.005252536851912737, + 0.004615247715264559, + 0.006059147417545319, + 0.00928153283894062, + 0.0075113195925951, + 0.00932183489203453, + 0.006118326913565397, + 0.0065691121853888035, + 0.005798086058348417, + 0.00406330032274127, + 0.013513424433767796, + 0.004913484677672386, + 0.00411289231851697, + 0.03533449396491051, + 0.012059247121214867, + 0.0035063286777585745, + 0.00919738132506609 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 0, + 1 + ], + "selected_reward": -0.008641703985631466 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video shot from a first-person perspective showing the character's hands?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the character's hands grasping an ancient longsword with dark leather wrapped around the hilt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the character perform a flowing sword dance involving lunges, spins, and stretches?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the setting a precipitous mountain with jagged rocks, thick flowing clouds, and distant indigo-hued mountains?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera pan smoothly in sync with the body's movements?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.012536413036286831 + ], + "opposite_logprob_tokens": [ + -4.387536525726318 + ], + "logprob_reward": -0.012536413036286831, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.012536413036286831, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12520085275173187, + 0.0486762598156929, + 0.08920302242040634, + 0.0648328959941864, + 0.029280632734298706, + 0.17162366211414337, + 0.09493587166070938, + 0.034545280039310455, + 0.025778716430068016, + 0.3159227967262268 + ], + "credit_map": [ + 0.010335645638406277, + 0.012513712979853153, + 0.022667407989501953, + 0.016760172322392464, + 0.015145025216042995, + 0.016514698043465614, + 0.010845344513654709, + 0.02041884884238243, + 0.007901356555521488, + 0.006411608774214983, + 0.007355202455073595, + 0.004425249062478542, + 0.007733114063739777, + 0.005151779856532812, + 0.004961985629051924, + 0.004735961556434631, + 0.006883576977998018, + 0.004412287380546331, + 0.008901645429432392, + 0.006139849778264761, + 0.006848447490483522, + 0.01622394286096096, + 0.021484797820448875, + 0.018308468163013458, + 0.003956857603043318, + 0.0036276853643357754, + 0.0065968697890639305, + 0.004672274924814701, + 0.008893744088709354, + 0.0171255711466074, + 0.008282271213829517, + 0.011677620001137257, + 0.0023107423912733793, + 0.003846406936645508, + 0.003356519853696227, + 0.002702290192246437, + 0.003428826341405511, + 0.00574474735185504, + 0.004156728275120258, + 0.0037343695294111967, + 0.016738703474402428, + 0.012244459241628647, + 0.02451523020863533, + 0.01778414472937584, + 0.015010024420917034, + 0.02751772664487362, + 0.024557823315262794, + 0.033255547285079956, + 0.006908932235091925, + 0.011495962738990784, + 0.014926960691809654, + 0.008366923779249191, + 0.01919272169470787, + 0.013905180618166924, + 0.010170322842895985, + 0.00996887031942606, + 0.004218888934701681, + 0.0043847630731761456, + 0.004253605380654335, + 0.0033150522504001856, + 0.0037301976699382067, + 0.005534650292247534, + 0.004270592704415321, + 0.004837529733777046, + 0.0024344101548194885, + 0.00397551478818059, + 0.004435971844941378, + 0.0029292567633092403, + 0.0035110185854136944, + 0.003399754874408245, + 0.0027088935021311045, + 0.002383895218372345, + 0.015037829987704754, + 0.03927091881632805, + 0.02267477661371231, + 0.017275666818022728, + 0.11872474104166031, + 0.048844873905181885, + 0.023610606789588928, + 0.03048335202038288 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 5, + 0 + ], + "selected_reward": -0.012536413036286831 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video shot from a first-person perspective showing the character's hands?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the character's hands grasping an ancient longsword with dark leather wrapped around the hilt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the character perform a flowing sword dance involving lunges, spins, and stretches?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the setting a precipitous mountain with jagged rocks, thick flowing clouds, and distant indigo-hued mountains?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera pan smoothly in sync with the body's movements?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.018180163577198982 + ], + "opposite_logprob_tokens": [ + -4.0181803703308105 + ], + "logprob_reward": -0.018180163577198982, + "yes_no_margin_reward": 4.0, + "token_credit_reward": -0.018180163577198982, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1773064285516739, + 0.05441240966320038, + 0.08256760239601135, + 0.06917303055524826, + 0.11315495520830154, + 0.07720180600881577, + 0.1419205516576767, + 0.05768466740846634, + 0.06717315316200256, + 0.15940535068511963 + ], + "credit_map": [ + 0.013579687103629112, + 0.027239469811320305, + 0.04322214797139168, + 0.018065260723233223, + 0.02126677706837654, + 0.011046521365642548, + 0.01901577040553093, + 0.02387080155313015, + 0.005108949262648821, + 0.004177896771579981, + 0.004653130192309618, + 0.0033605587668716908, + 0.0037700363900512457, + 0.010541379451751709, + 0.00857516098767519, + 0.01422529760748148, + 0.007491097319871187, + 0.008191445842385292, + 0.01385204866528511, + 0.007191170006990433, + 0.007985877804458141, + 0.013364310376346111, + 0.010727116838097572, + 0.013764538802206516, + 0.004338950850069523, + 0.003907827660441399, + 0.006061742547899485, + 0.0041620065458118916, + 0.008958864957094193, + 0.020443584769964218, + 0.008950244635343552, + 0.012349809519946575, + 0.0101265087723732, + 0.016394460573792458, + 0.013411516323685646, + 0.013690118677914143, + 0.013648906722664833, + 0.016649970784783363, + 0.0154108302667737, + 0.013822644017636776, + 0.006912209093570709, + 0.005431769415736198, + 0.011895024217665195, + 0.012348642572760582, + 0.006002992391586304, + 0.008692031726241112, + 0.013322430662810802, + 0.012596704065799713, + 0.01042245328426361, + 0.0237769465893507, + 0.0169229693710804, + 0.01450879592448473, + 0.02532770484685898, + 0.020588671788573265, + 0.01673496514558792, + 0.013638041913509369, + 0.004647411406040192, + 0.0103368628770113, + 0.008622918277978897, + 0.00618564011529088, + 0.005740908440202475, + 0.00731493066996336, + 0.007092386018484831, + 0.00774361239746213, + 0.005991032347083092, + 0.008944040164351463, + 0.013060753233730793, + 0.009730818681418896, + 0.006108694244176149, + 0.008308960124850273, + 0.007216382306069136, + 0.007812468335032463, + 0.008505750447511673, + 0.018802255392074585, + 0.012220812030136585, + 0.017128026112914085, + 0.05354262888431549, + 0.023631077259778976, + 0.01047517266124487, + 0.015099630691111088 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 6 + ], + "selected_reward": -0.018180163577198982 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000095/group_6/index_tensor([3152, 3152])_group_6_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000095/group_6/index_tensor([3152, 3152])_group_6_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..ac840a9e9a04b2f61e481acc64d83d2729d2f32a --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000095/group_6/index_tensor([3152, 3152])_group_6_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "classic Hitchcock zoom lens, from a slightly bird's-eye view, captures an ancient wooden cargo ship struggling in the turbulent black sea. The inky black waters whip up mountainous waves, their crests ripped into pale foam by the wind. The sky is overcast, and occasional flashes of lightning provide the sole source of light, momentarily illuminating everything amidst the torrential downpour. The ship's hull is a weathered, dark timber, its towering masts groaning painfully in the storm. The decks are piled high with dark brown oak barrels secured with rusty iron hoops. These barrels roll and collide perilously, clattering with a dull roar as the ship pitches devastatingly between the crests and troughs of the waves. As the camera slowly zooms forward, it simultaneously pulls back, causing the ship's size to remain nearly constant within the frame while the raging sea and swirling clouds in the background appear dramatically stretched, as if time and space itself were warping and expanding. This perfectly recreates the hallmark, dizzying dreamlike quality of Hitchcock's films, creating an inescapable, surreal atmosphere of tension and dread.", + "seed": 15748, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000095/group_6/index_3152_seed_15748_dp_12_rank_24_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 6, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.1476696539670229, + "vlm_logprob": -0.1476696539670229, + "vlm_yes_no_margin": 3.025, + "vlm_token_credit": -0.1476696539670229, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an ancient wooden cargo ship with towering masts?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are dark brown oak barrels piled high on the ship's deck, secured with rusty iron hoops?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the environment a turbulent black sea with mountainous waves topped with pale foam under an overcast sky with lightning flashes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the barrels on the deck roll and collide while the ship pitches between wave crests and troughs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera perform a simultaneous slow zoom forward and pull back, keeping the ship's size constant while stretching the background?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.022009167820215225 + ], + "opposite_logprob_tokens": [ + -4.147009372711182 + ], + "logprob_reward": -0.022009167820215225, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.022009167820215225, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12555155158042908, + 0.10907649248838425, + 0.09080634266138077, + 0.11300990730524063, + 0.14146241545677185, + 0.05599614977836609, + 0.0889790877699852, + 0.06077397242188454, + 0.14688582718372345, + 0.06745824962854385 + ], + "credit_map": [ + 0.014705497771501541, + 0.01309152040630579, + 0.01678699068725109, + 0.02079533413052559, + 0.011110681109130383, + 0.018325747922062874, + 0.01461535319685936, + 0.01612042263150215, + 0.017980733886361122, + 0.015270219184458256, + 0.00919654406607151, + 0.013199021108448505, + 0.011189823038876057, + 0.013607295230031013, + 0.013156566768884659, + 0.015476285479962826, + 0.006673483178019524, + 0.007074876222759485, + 0.013659193180501461, + 0.010319104418158531, + 0.014500672928988934, + 0.01283353939652443, + 0.013338910415768623, + 0.012406565248966217, + 0.011975506320595741, + 0.009870577603578568, + 0.01160738430917263, + 0.01270761713385582, + 0.026091068983078003, + 0.015706518664956093, + 0.01119470689445734, + 0.013856529258191586, + 0.011084632948040962, + 0.01202890370041132, + 0.014209950342774391, + 0.01247639674693346, + 0.0254064854234457, + 0.02474752813577652, + 0.017574908211827278, + 0.02393360435962677, + 0.006390537600964308, + 0.005714226979762316, + 0.005137500818818808, + 0.007767688017338514, + 0.008494572713971138, + 0.008875030092895031, + 0.006222181022167206, + 0.007394408341497183, + 0.011299462988972664, + 0.015549367293715477, + 0.009886497631669044, + 0.007642622571438551, + 0.010608112439513206, + 0.010481567122042179, + 0.01348794437944889, + 0.010023515671491623, + 0.0066505675204098225, + 0.006201975047588348, + 0.006721052806824446, + 0.009307816624641418, + 0.006230424623936415, + 0.009910535998642445, + 0.008744458667933941, + 0.00700713787227869, + 0.013731790706515312, + 0.01445033773779869, + 0.02307797223329544, + 0.017866438254714012, + 0.013473672792315483, + 0.028779035434126854, + 0.018769042566418648, + 0.016737544909119606, + 0.005902272183448076, + 0.007952606305480003, + 0.007830239832401276, + 0.005519161466509104, + 0.005278901197016239, + 0.007532760500907898, + 0.010087276808917522, + 0.01735503226518631 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 4, + 0 + ], + "selected_reward": -0.022009167820215225 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an ancient wooden cargo ship with towering masts?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are dark brown oak barrels piled high on the ship's deck, secured with rusty iron hoops?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the environment a turbulent black sea with mountainous waves topped with pale foam under an overcast sky with lightning flashes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the barrels on the deck roll and collide while the ship pitches between wave crests and troughs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera perform a simultaneous slow zoom forward and pull back, keeping the ship's size constant while stretching the background?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03372852876782417 + ], + "opposite_logprob_tokens": [ + -3.40872859954834 + ], + "logprob_reward": -0.03372852876782417, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.03372852876782417, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.19313183426856995, + 0.13636957108974457, + 0.09530557692050934, + 0.03170088306069374, + 0.09226044267416, + 0.12561804056167603, + 0.07085122913122177, + 0.06120331212878227, + 0.08393355458974838, + 0.10962562263011932 + ], + "credit_map": [ + 0.025039996951818466, + 0.022156206890940666, + 0.02350776456296444, + 0.027253134176135063, + 0.014099903404712677, + 0.0321754552423954, + 0.02092277631163597, + 0.027976596727967262, + 0.029503583908081055, + 0.019560279324650764, + 0.012474007904529572, + 0.012348992750048637, + 0.008624338544905186, + 0.014755867421627045, + 0.019112300127744675, + 0.01999019645154476, + 0.012790100648999214, + 0.007756300736218691, + 0.00976311881095171, + 0.012647401541471481, + 0.014331960119307041, + 0.013091176748275757, + 0.012092439457774162, + 0.012833073735237122, + 0.004189538303762674, + 0.003043001051992178, + 0.0026671667583286762, + 0.0035241059958934784, + 0.004614756442606449, + 0.006585726048797369, + 0.0036092218942940235, + 0.003467369358986616, + 0.01314821932464838, + 0.012865904718637466, + 0.010400950908660889, + 0.010870001278817654, + 0.012402646243572235, + 0.010883400216698647, + 0.00742757273837924, + 0.014261743985116482, + 0.013974620960652828, + 0.02077229879796505, + 0.013239326886832714, + 0.01686258614063263, + 0.013941895216703415, + 0.018706651404500008, + 0.01186275016516447, + 0.01625790260732174, + 0.00870212446898222, + 0.014466434717178345, + 0.005428196396678686, + 0.007889286614954472, + 0.008429360575973988, + 0.009066922590136528, + 0.010152780450880527, + 0.006716116331517696, + 0.0071685235016047955, + 0.007322687190026045, + 0.007100862450897694, + 0.007005686871707439, + 0.006815337110310793, + 0.010150793008506298, + 0.008040099404752254, + 0.007599322125315666, + 0.008028638549149036, + 0.010887634009122849, + 0.012253683991730213, + 0.009777756407856941, + 0.008799075148999691, + 0.011713339947164059, + 0.008289807476103306, + 0.014183612540364265, + 0.012315734289586544, + 0.019412074238061905, + 0.013522271066904068, + 0.009251432493329048, + 0.01355317234992981, + 0.017428113147616386, + 0.009233024902641773, + 0.014909803867340088 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 5 + ], + "selected_reward": -0.03372852876782417 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an ancient wooden cargo ship with towering masts?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are dark brown oak barrels piled high on the ship's deck, secured with rusty iron hoops?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the environment a turbulent black sea with mountainous waves topped with pale foam under an overcast sky with lightning flashes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the barrels on the deck roll and collide while the ship pitches between wave crests and troughs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera perform a simultaneous slow zoom forward and pull back, keeping the ship's size constant while stretching the background?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01606954075396061 + ], + "opposite_logprob_tokens": [ + -4.141069412231445 + ], + "logprob_reward": -0.01606954075396061, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.01606954075396061, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08223112672567368, + 0.3773348331451416, + 0.0632505714893341, + 0.021302975714206696, + 0.0511101670563221, + 0.07049956172704697, + 0.09422314912080765, + 0.06480022519826889, + 0.09283193945884705, + 0.08241545408964157 + ], + "credit_map": [ + 0.010146480984985828, + 0.011494799517095089, + 0.008758378215134144, + 0.011349918320775032, + 0.008174464106559753, + 0.011885854415595531, + 0.008363252505660057, + 0.012057981453835964, + 0.0404982715845108, + 0.035364869982004166, + 0.04567871615290642, + 0.03853940963745117, + 0.024758297950029373, + 0.05829896032810211, + 0.07474702596664429, + 0.05944928154349327, + 0.012657692655920982, + 0.005015880335122347, + 0.005273275543004274, + 0.008357840590178967, + 0.005890914238989353, + 0.005366834811866283, + 0.010630969889461994, + 0.010057167150080204, + 0.002073123585432768, + 0.0017583450535312295, + 0.0018299755174666643, + 0.0021825775038450956, + 0.004587613977491856, + 0.0032886175904423, + 0.002787683391943574, + 0.0027950392104685307, + 0.0052782222628593445, + 0.004579950589686632, + 0.005675288382917643, + 0.00505989883095026, + 0.007596843875944614, + 0.008364762179553509, + 0.006326225586235523, + 0.008228973485529423, + 0.00858582928776741, + 0.008557818830013275, + 0.00665141548961401, + 0.007990053854882717, + 0.014145582914352417, + 0.01012258231639862, + 0.005448358599096537, + 0.008997920900583267, + 0.01505250483751297, + 0.016933875158429146, + 0.009747226722538471, + 0.012846570461988449, + 0.009526682086288929, + 0.012212362140417099, + 0.010264457203447819, + 0.007639470975846052, + 0.006132257170975208, + 0.008284899406135082, + 0.010405744425952435, + 0.007321655750274658, + 0.006845914293080568, + 0.010735640302300453, + 0.006502276752144098, + 0.008571837097406387, + 0.006952465046197176, + 0.005354180466383696, + 0.018571771681308746, + 0.00911160372197628, + 0.007888012565672398, + 0.024623485282063484, + 0.008355402387678623, + 0.011975012719631195, + 0.007535294629633427, + 0.009812435135245323, + 0.015250463970005512, + 0.008660823106765747, + 0.009063605219125748, + 0.010551688261330128, + 0.009543517604470253, + 0.011997617781162262 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 6, + 8 + ], + "selected_reward": -0.01606954075396061 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an ancient wooden cargo ship with towering masts?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are dark brown oak barrels piled high on the ship's deck, secured with rusty iron hoops?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the environment a turbulent black sea with mountainous waves topped with pale foam under an overcast sky with lightning flashes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the barrels on the deck roll and collide while the ship pitches between wave crests and troughs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera perform a simultaneous slow zoom forward and pull back, keeping the ship's size constant while stretching the background?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.033702366054058075 + ], + "opposite_logprob_tokens": [ + -3.4087023735046387 + ], + "logprob_reward": -0.033702366054058075, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.033702366054058075, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10830274969339371, + 0.14319781959056854, + 0.11214064806699753, + 0.08902308344841003, + 0.0746741071343422, + 0.09159422665834427, + 0.18541748821735382, + 0.0778517872095108, + 0.05139479413628578, + 0.06640329957008362 + ], + "credit_map": [ + 0.014383936300873756, + 0.012032159604132175, + 0.0151031743735075, + 0.013147421181201935, + 0.009795794263482094, + 0.01632777787744999, + 0.010972599498927593, + 0.016539884731173515, + 0.023129500448703766, + 0.014630570076406002, + 0.015662875026464462, + 0.016498832032084465, + 0.011313313618302345, + 0.019083859398961067, + 0.022882862016558647, + 0.01999600976705551, + 0.01329499576240778, + 0.00936485081911087, + 0.009441004134714603, + 0.015326007269322872, + 0.022689269855618477, + 0.010125982575118542, + 0.014253724366426468, + 0.017644813284277916, + 0.011574574746191502, + 0.010080250911414623, + 0.010465670377016068, + 0.013203646056354046, + 0.012047963216900826, + 0.014024293050169945, + 0.008292765356600285, + 0.009333920665085316, + 0.010906754992902279, + 0.009594731032848358, + 0.008692585863173008, + 0.00835468154400587, + 0.011314808391034603, + 0.007607004139572382, + 0.007709867786616087, + 0.010493678972125053, + 0.013566303066909313, + 0.01066668052226305, + 0.010053446516394615, + 0.011005394160747528, + 0.007756172213703394, + 0.011597944423556328, + 0.014135115779936314, + 0.012813164852559566, + 0.016609694808721542, + 0.02126871794462204, + 0.024217087775468826, + 0.015935642644762993, + 0.029650431126356125, + 0.02962971106171608, + 0.02357415109872818, + 0.024532048031687737, + 0.007712235674262047, + 0.00857495702803135, + 0.01337906252592802, + 0.008330978453159332, + 0.0115980738773942, + 0.011664371937513351, + 0.006777109112590551, + 0.00981499720364809, + 0.007218200247734785, + 0.007768853101879358, + 0.0066700042225420475, + 0.0053476872853934765, + 0.006123640108853579, + 0.00852455385029316, + 0.003695343853905797, + 0.00604651216417551, + 0.006360069382935762, + 0.007213786710053682, + 0.007034330163151026, + 0.006286424119025469, + 0.008293384686112404, + 0.012867889367043972, + 0.0084843710064888, + 0.009863047860562801 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 1, + 2 + ], + "selected_reward": -0.033702366054058075 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an ancient wooden cargo ship with towering masts?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are dark brown oak barrels piled high on the ship's deck, secured with rusty iron hoops?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the environment a turbulent black sea with mountainous waves topped with pale foam under an overcast sky with lightning flashes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the barrels on the deck roll and collide while the ship pitches between wave crests and troughs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera perform a simultaneous slow zoom forward and pull back, keeping the ship's size constant while stretching the background?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.6328386664390564 + ], + "opposite_logprob_tokens": [ + -0.7578386664390564 + ], + "logprob_reward": -0.6328386664390564, + "yes_no_margin_reward": 0.125, + "token_credit_reward": -0.6328386664390564, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.03941263630986214, + 0.050768833607435226, + 0.08224595338106155, + 0.06888118386268616, + 0.03910656273365021, + 0.1426258087158203, + 0.228955939412117, + 0.07238628715276718, + 0.20818787813186646, + 0.06742890179157257 + ], + "credit_map": [ + 0.005412023514509201, + 0.003802402876317501, + 0.004060475155711174, + 0.006065929774194956, + 0.0027443417347967625, + 0.0060464199632406235, + 0.0049666669219732285, + 0.006314374972134829, + 0.005520348437130451, + 0.006337946746498346, + 0.005198933184146881, + 0.006535060703754425, + 0.00287612178362906, + 0.00784845557063818, + 0.008315968327224255, + 0.008135996758937836, + 0.009157969616353512, + 0.007049908861517906, + 0.013040540739893913, + 0.011085410602390766, + 0.009668064303696156, + 0.009494230151176453, + 0.011360871605575085, + 0.011388963088393211, + 0.009299020282924175, + 0.008163860999047756, + 0.006260243710130453, + 0.009528543800115585, + 0.011570746079087257, + 0.00945512019097805, + 0.006979746278375387, + 0.007623902056366205, + 0.0054275523871183395, + 0.004560577683150768, + 0.004860685206949711, + 0.005678599234670401, + 0.0052560619078576565, + 0.003810200607404113, + 0.0041291676461696625, + 0.005383720155805349, + 0.01856227219104767, + 0.018997175619006157, + 0.014051997102797031, + 0.01901363395154476, + 0.020806338638067245, + 0.02112545818090439, + 0.012106449343264103, + 0.017962483689188957, + 0.02348647639155388, + 0.0365229956805706, + 0.029319049790501595, + 0.025686392560601234, + 0.03313685208559036, + 0.03324435278773308, + 0.021700698882341385, + 0.025859124958515167, + 0.008521339856088161, + 0.005172251258045435, + 0.016341911628842354, + 0.006164322607219219, + 0.010493754409253597, + 0.009887694381177425, + 0.005645328667014837, + 0.0101596899330616, + 0.014907028526067734, + 0.017285944893956184, + 0.032587867230176926, + 0.02438029833137989, + 0.020091334357857704, + 0.041380979120731354, + 0.019336745142936707, + 0.038217682391405106, + 0.008494912646710873, + 0.008653839118778706, + 0.013012127950787544, + 0.00754279550164938, + 0.005593816749751568, + 0.00687320064753294, + 0.006506755016744137, + 0.01075145322829485 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 8, + 5 + ], + "selected_reward": -0.6328386664390564 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "classic Hitchcock zoom lens, from a slightly bird's-eye view, captures an ancient wooden cargo ship struggling in the turbulent black sea. The inky black waters whip up mountainous waves, their crests ripped into pale foam by the wind. The sky is overcast, and occasional flashes of lightning provide the sole source of light, momentarily illuminating everything amidst the torrential downpour. The ship's hull is a weathered, dark timber, its towering masts groaning painfully in the storm. The decks are piled high with dark brown oak barrels secured with rusty iron hoops. These barrels roll and collide perilously, clattering with a dull roar as the ship pitches devastatingly between the crests and troughs of the waves. As the camera slowly zooms forward, it simultaneously pulls back, causing the ship's size to remain nearly constant within the frame while the raging sea and swirling clouds in the background appear dramatically stretched, as if time and space itself were warping and expanding. This perfectly recreates the hallmark, dizzying dreamlike quality of Hitchcock's films, creating an inescapable, surreal atmosphere of tension and dread.", + "seed": 15749, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000095/group_6/index_3152_seed_15749_dp_12_rank_24_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 6, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.19242739044129847, + "vlm_logprob": -0.19242739044129847, + "vlm_yes_no_margin": 2.35, + "vlm_token_credit": -0.19242739044129847, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an ancient wooden cargo ship with towering masts?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are dark brown oak barrels piled high on the ship's deck, secured with rusty iron hoops?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the environment a turbulent black sea with mountainous waves topped with pale foam under an overcast sky with lightning flashes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the barrels on the deck roll and collide while the ship pitches between wave crests and troughs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera perform a simultaneous slow zoom forward and pull back, keeping the ship's size constant while stretching the background?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.20945383608341217 + ], + "opposite_logprob_tokens": [ + -1.709453821182251 + ], + "logprob_reward": -0.20945383608341217, + "yes_no_margin_reward": 1.5, + "token_credit_reward": -0.20945383608341217, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.05971764400601387, + 0.07625068724155426, + 0.037897028028964996, + 0.1073402464389801, + 0.06527680158615112, + 0.048117496073246, + 0.07925783097743988, + 0.02253780886530876, + 0.058058179914951324, + 0.44554632902145386 + ], + "credit_map": [ + 0.0061904857866466045, + 0.005142843350768089, + 0.012259143404662609, + 0.007209504023194313, + 0.005440074950456619, + 0.007752931211143732, + 0.00640849070623517, + 0.009314167313277721, + 0.005355175118893385, + 0.009549953043460846, + 0.008522416464984417, + 0.011954979039728642, + 0.01214445661753416, + 0.008996916003525257, + 0.011050643399357796, + 0.008676145225763321, + 0.003330457489937544, + 0.003716286038979888, + 0.008363880217075348, + 0.005136515013873577, + 0.004530089441686869, + 0.003617113921791315, + 0.0047888318076729774, + 0.0044138552621006966, + 0.007370660547167063, + 0.004165915306657553, + 0.0063962130807340145, + 0.008374308235943317, + 0.004615401849150658, + 0.016491224989295006, + 0.034553855657577515, + 0.02537267655134201, + 0.006258394569158554, + 0.004058847203850746, + 0.00764461187645793, + 0.008213442750275135, + 0.0037799207493662834, + 0.005898977164179087, + 0.01328680757433176, + 0.016135800629854202, + 0.005291423294693232, + 0.002229398349300027, + 0.0044199442490935326, + 0.007550389040261507, + 0.002787601901218295, + 0.0043365792371332645, + 0.006475923117250204, + 0.01502623874694109, + 0.005423706490546465, + 0.005398341454565525, + 0.010342810302972794, + 0.019002430140972137, + 0.0040426356717944145, + 0.006751419045031071, + 0.01194807793945074, + 0.01634841039776802, + 0.002201739465817809, + 0.0027806286234408617, + 0.002831483958289027, + 0.002551501616835594, + 0.002103849546983838, + 0.0022952447179704905, + 0.004076626151800156, + 0.0036967352498322725, + 0.005192112177610397, + 0.004148810636252165, + 0.007251877803355455, + 0.012345097959041595, + 0.004582423251122236, + 0.006219498813152313, + 0.011505682952702045, + 0.006812677718698978, + 0.014521905221045017, + 0.008013663813471794, + 0.05409986898303032, + 0.0881858691573143, + 0.01427980326116085, + 0.014263734221458435, + 0.18383793532848358, + 0.06834351271390915 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 3, + 6 + ], + "selected_reward": -0.20945383608341217 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an ancient wooden cargo ship with towering masts?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are dark brown oak barrels piled high on the ship's deck, secured with rusty iron hoops?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the environment a turbulent black sea with mountainous waves topped with pale foam under an overcast sky with lightning flashes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the barrels on the deck roll and collide while the ship pitches between wave crests and troughs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera perform a simultaneous slow zoom forward and pull back, keeping the ship's size constant while stretching the background?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0700620487332344 + ], + "opposite_logprob_tokens": [ + -2.6950621604919434 + ], + "logprob_reward": -0.0700620487332344, + "yes_no_margin_reward": 2.625, + "token_credit_reward": -0.0700620487332344, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.049667999148368835, + 0.07063250988721848, + 0.06091177463531494, + 0.37817248702049255, + 0.10776214301586151, + 0.03653667867183685, + 0.04149530827999115, + 0.04444411024451256, + 0.12442021816968918, + 0.08595675975084305 + ], + "credit_map": [ + 0.0060164351016283035, + 0.005447796545922756, + 0.006164418533444405, + 0.004715106450021267, + 0.006664151791483164, + 0.006148204673081636, + 0.005237157456576824, + 0.009274729527533054, + 0.006165783852338791, + 0.007291202433407307, + 0.009645317681133747, + 0.009771464392542839, + 0.010121969506144524, + 0.008908894844353199, + 0.01032112818211317, + 0.008406746201217175, + 0.004241513088345528, + 0.003512172494083643, + 0.009419712238013744, + 0.01047671027481556, + 0.005897052586078644, + 0.008974692784249783, + 0.008717743679881096, + 0.009672180749475956, + 0.03032384254038334, + 0.0067601315677165985, + 0.012658191844820976, + 0.020971281453967094, + 0.014423253014683723, + 0.0708267018198967, + 0.13138790428638458, + 0.09082116186618805, + 0.008467767387628555, + 0.007630733773112297, + 0.010775522328913212, + 0.014365882612764835, + 0.012611676938831806, + 0.01569836027920246, + 0.019946200773119926, + 0.018265997990965843, + 0.0021279046777635813, + 0.0031601416412740946, + 0.004105780739337206, + 0.010582942515611649, + 0.0016022446798160672, + 0.0017122741555795074, + 0.0069239819422364235, + 0.00632140738889575, + 0.004023436922580004, + 0.003703213296830654, + 0.005259249359369278, + 0.006579261273145676, + 0.0024137927684932947, + 0.0063230725936591625, + 0.0072740213945508, + 0.005919260438531637, + 0.004381155129522085, + 0.004797584842890501, + 0.006178659852594137, + 0.007141701877117157, + 0.00343916192650795, + 0.004183315206319094, + 0.007694208063185215, + 0.006628324743360281, + 0.009964277036488056, + 0.005934556946158409, + 0.012686137109994888, + 0.04133526608347893, + 0.005752866622060537, + 0.008315914310514927, + 0.03187694773077965, + 0.008554257452487946, + 0.00604041013866663, + 0.004054634366184473, + 0.006681893486529589, + 0.010859171859920025, + 0.005310701671987772, + 0.004831886850297451, + 0.02682984247803688, + 0.021348217502236366 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 8, + 4 + ], + "selected_reward": -0.0700620487332344 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an ancient wooden cargo ship with towering masts?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are dark brown oak barrels piled high on the ship's deck, secured with rusty iron hoops?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the environment a turbulent black sea with mountainous waves topped with pale foam under an overcast sky with lightning flashes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the barrels on the deck roll and collide while the ship pitches between wave crests and troughs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera perform a simultaneous slow zoom forward and pull back, keeping the ship's size constant while stretching the background?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03370017558336258 + ], + "opposite_logprob_tokens": [ + -3.4087002277374268 + ], + "logprob_reward": -0.03370017558336258, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.03370017558336258, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08750862628221512, + 0.062171295285224915, + 0.03744750842452049, + 0.6896083950996399, + 0.022168247029185295, + 0.014485115185379982, + 0.01400159765034914, + 0.01516230870038271, + 0.03408636897802353, + 0.023360488936305046 + ], + "credit_map": [ + 0.012749443762004375, + 0.006595509126782417, + 0.011375272646546364, + 0.008830586448311806, + 0.006807691417634487, + 0.02006303332746029, + 0.009712529368698597, + 0.011374572291970253, + 0.007273069117218256, + 0.005981127731502056, + 0.009316960349678993, + 0.008154405280947685, + 0.006706618703901768, + 0.008745281957089901, + 0.007295555435121059, + 0.008698280900716782, + 0.004441343247890472, + 0.0018986038630828261, + 0.0061724488623440266, + 0.004325648304075003, + 0.0028183581307530403, + 0.006682372651994228, + 0.005949003156274557, + 0.005159737076610327, + 0.06541399657726288, + 0.012015370652079582, + 0.015825092792510986, + 0.03723372519016266, + 0.026059364899992943, + 0.15167084336280823, + 0.22992119193077087, + 0.151468887925148, + 0.002252641599625349, + 0.001206218614242971, + 0.0027482849545776844, + 0.003213155083358288, + 0.0017734646098688245, + 0.004408122040331364, + 0.003111712634563446, + 0.003454649355262518, + 0.0013543402310460806, + 0.0009660687646828592, + 0.0013117125490680337, + 0.001913408632390201, + 0.0013815081911161542, + 0.002466548001393676, + 0.002282814122736454, + 0.002808715682476759, + 0.0012097713770344853, + 0.0011999427806586027, + 0.0020189830102026463, + 0.003090022597461939, + 0.0009160903282463551, + 0.0013960391515865922, + 0.002462828066200018, + 0.0017079205717891455, + 0.001623087446205318, + 0.0014298239257186651, + 0.0018759658560156822, + 0.0017258762381970882, + 0.0011131549254059792, + 0.0022110643330961466, + 0.0028587467968463898, + 0.002324589528143406, + 0.002095339121297002, + 0.0013315165415406227, + 0.003046458587050438, + 0.013541849330067635, + 0.0011533164652064443, + 0.0012554399436339736, + 0.009507383219897747, + 0.0021550683304667473, + 0.002663076389580965, + 0.0021292075980454683, + 0.0028986025135964155, + 0.0029312558472156525, + 0.0018918417626991868, + 0.002384245628491044, + 0.0033884774893522263, + 0.005073783453553915 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 0, + 1 + ], + "selected_reward": -0.03370017558336258 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an ancient wooden cargo ship with towering masts?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are dark brown oak barrels piled high on the ship's deck, secured with rusty iron hoops?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the environment a turbulent black sea with mountainous waves topped with pale foam under an overcast sky with lightning flashes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the barrels on the deck roll and collide while the ship pitches between wave crests and troughs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera perform a simultaneous slow zoom forward and pull back, keeping the ship's size constant while stretching the background?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.016077635809779167 + ], + "opposite_logprob_tokens": [ + -4.141077518463135 + ], + "logprob_reward": -0.016077635809779167, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.016077635809779167, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0400548130273819, + 0.05849558115005493, + 0.050077956169843674, + 0.35680750012397766, + 0.03867104649543762, + 0.023049186915159225, + 0.026812082156538963, + 0.01603015512228012, + 0.07353575527667999, + 0.3164658546447754 + ], + "credit_map": [ + 0.0060218023136258125, + 0.0030951648950576782, + 0.005081729963421822, + 0.004371500574052334, + 0.0031139268539845943, + 0.007781031541526318, + 0.004355163779109716, + 0.006234494037926197, + 0.007605040445923805, + 0.006534117739647627, + 0.007889225147664547, + 0.006976528558880091, + 0.007621915079653263, + 0.00656232051551342, + 0.006669211201369762, + 0.008637221530079842, + 0.003182473126798868, + 0.0026427058037370443, + 0.006758744362741709, + 0.008069811388850212, + 0.0029468571301549673, + 0.007360011804848909, + 0.007244036998599768, + 0.011873318813741207, + 0.0321810357272625, + 0.006223008967936039, + 0.008071720600128174, + 0.019828373566269875, + 0.013863935135304928, + 0.07493168115615845, + 0.12110863626003265, + 0.0805991068482399, + 0.002536510117352009, + 0.002235793974250555, + 0.005638569127768278, + 0.00614754157140851, + 0.004526508506387472, + 0.005194724537432194, + 0.007327277213335037, + 0.005064122844487429, + 0.001655553001910448, + 0.0016419141320511699, + 0.002674599876627326, + 0.004752554930746555, + 0.0020705084316432476, + 0.0017828618874773383, + 0.003983718808740377, + 0.00448747631162405, + 0.001965886214748025, + 0.0021898827981203794, + 0.003251885762438178, + 0.00465809041634202, + 0.0014062972040846944, + 0.0034238530788570642, + 0.004783684853464365, + 0.00513250008225441, + 0.0016885704826563597, + 0.001812680740840733, + 0.0019992615561932325, + 0.001971134450286627, + 0.001789459609426558, + 0.002400001510977745, + 0.0023249925579875708, + 0.002044054213911295, + 0.004760200623422861, + 0.0027612242847681046, + 0.007127765100449324, + 0.028772877529263496, + 0.0020460907835513353, + 0.003130056895315647, + 0.020027004182338715, + 0.004910535644739866, + 0.00959679763764143, + 0.0058270604349672794, + 0.03400778770446777, + 0.059857845306396484, + 0.010621173307299614, + 0.010293712839484215, + 0.12857890129089355, + 0.05768256261944771 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 9, + 8 + ], + "selected_reward": -0.016077635809779167 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an ancient wooden cargo ship with towering masts?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are dark brown oak barrels piled high on the ship's deck, secured with rusty iron hoops?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the environment a turbulent black sea with mountainous waves topped with pale foam under an overcast sky with lightning flashes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the barrels on the deck roll and collide while the ship pitches between wave crests and troughs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera perform a simultaneous slow zoom forward and pull back, keeping the ship's size constant while stretching the background?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.6328432559967041 + ], + "opposite_logprob_tokens": [ + -0.7578432559967041 + ], + "logprob_reward": -0.6328432559967041, + "yes_no_margin_reward": 0.125, + "token_credit_reward": -0.6328432559967041, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07411316782236099, + 0.13876527547836304, + 0.0444524884223938, + 0.2051856368780136, + 0.07229219377040863, + 0.10565221309661865, + 0.07495066523551941, + 0.07763657718896866, + 0.08159127831459045, + 0.12536051869392395 + ], + "credit_map": [ + 0.00896041002124548, + 0.007340356707572937, + 0.01095270924270153, + 0.009120621718466282, + 0.007712072692811489, + 0.011051218956708908, + 0.009016045369207859, + 0.009959738701581955, + 0.01638760045170784, + 0.015808982774615288, + 0.018151141703128815, + 0.01854022592306137, + 0.018241239711642265, + 0.020013023167848587, + 0.014043180271983147, + 0.017579885199666023, + 0.006266670301556587, + 0.002689969725906849, + 0.007057732902467251, + 0.006838199216872454, + 0.00472458777949214, + 0.00523867504671216, + 0.005881038028746843, + 0.0057556177489459515, + 0.033238526433706284, + 0.012325531803071499, + 0.016691064462065697, + 0.02134031429886818, + 0.01164014171808958, + 0.06556790322065353, + 0.021806923672556877, + 0.022575251758098602, + 0.004370649345219135, + 0.004033051431179047, + 0.008507223799824715, + 0.013898986391723156, + 0.005608535371720791, + 0.010687071830034256, + 0.014110608957707882, + 0.01107607502490282, + 0.007397499401122332, + 0.006274483632296324, + 0.010259579867124557, + 0.022443871945142746, + 0.00731636630371213, + 0.011256690137088299, + 0.015574970282614231, + 0.025128761306405067, + 0.006225845776498318, + 0.009355460293591022, + 0.009968077763915062, + 0.01963810808956623, + 0.004451854154467583, + 0.007058662362396717, + 0.009385844692587852, + 0.00886682327836752, + 0.007587655447423458, + 0.011468863114714622, + 0.010294681414961815, + 0.00816313922405243, + 0.007514516357332468, + 0.008450400084257126, + 0.013729480095207691, + 0.010427850298583508, + 0.0058672004379332066, + 0.0032999676186591387, + 0.0068525285460054874, + 0.02777555212378502, + 0.0029520350508391857, + 0.005548869259655476, + 0.019987229257822037, + 0.009307902306318283, + 0.006023443304002285, + 0.0032707464415580034, + 0.009654896333813667, + 0.018430229276418686, + 0.00476515619084239, + 0.007367963902652264, + 0.04455280676484108, + 0.0312952920794487 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 1, + 9 + ], + "selected_reward": -0.6328432559967041 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "classic Hitchcock zoom lens, from a slightly bird's-eye view, captures an ancient wooden cargo ship struggling in the turbulent black sea. The inky black waters whip up mountainous waves, their crests ripped into pale foam by the wind. The sky is overcast, and occasional flashes of lightning provide the sole source of light, momentarily illuminating everything amidst the torrential downpour. The ship's hull is a weathered, dark timber, its towering masts groaning painfully in the storm. The decks are piled high with dark brown oak barrels secured with rusty iron hoops. These barrels roll and collide perilously, clattering with a dull roar as the ship pitches devastatingly between the crests and troughs of the waves. As the camera slowly zooms forward, it simultaneously pulls back, causing the ship's size to remain nearly constant within the frame while the raging sea and swirling clouds in the background appear dramatically stretched, as if time and space itself were warping and expanding. This perfectly recreates the hallmark, dizzying dreamlike quality of Hitchcock's films, creating an inescapable, surreal atmosphere of tension and dread.", + "seed": 15750, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000095/group_6/index_3152_seed_15750_dp_13_rank_26_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 6, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.16368794105947018, + "vlm_logprob": -0.16368794105947018, + "vlm_yes_no_margin": 2.95, + "vlm_token_credit": -0.16368794105947018, + "vlm_yes_ratio": 0.8, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an ancient wooden cargo ship with towering masts?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are dark brown oak barrels piled high on the ship's deck, secured with rusty iron hoops?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the environment a turbulent black sea with mountainous waves topped with pale foam under an overcast sky with lightning flashes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the barrels on the deck roll and collide while the ship pitches between wave crests and troughs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera perform a simultaneous slow zoom forward and pull back, keeping the ship's size constant while stretching the background?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.024101020768284798 + ], + "opposite_logprob_tokens": [ + -4.024101257324219 + ], + "logprob_reward": -0.024101020768284798, + "yes_no_margin_reward": 4.0, + "token_credit_reward": -0.024101020768284798, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07383305579423904, + 0.06717055290937424, + 0.09676865488290787, + 0.20829829573631287, + 0.15401509404182434, + 0.08207349479198456, + 0.09466943144798279, + 0.06434269994497299, + 0.10362949967384338, + 0.055199217051267624 + ], + "credit_map": [ + 0.006253574974834919, + 0.004757368471473455, + 0.011238643899559975, + 0.012183151207864285, + 0.007496314123272896, + 0.01440736185759306, + 0.009244822897017002, + 0.008251816034317017, + 0.005524087231606245, + 0.004565957933664322, + 0.01306470949202776, + 0.008021409623324871, + 0.007435687817633152, + 0.008005952462553978, + 0.01421021856367588, + 0.006342531647533178, + 0.008057069964706898, + 0.01638765260577202, + 0.018059054389595985, + 0.009578175842761993, + 0.011024557054042816, + 0.008047768846154213, + 0.016801878809928894, + 0.008812488988041878, + 0.018534062430262566, + 0.04352611303329468, + 0.020109843462705612, + 0.021526599302887917, + 0.04484755918383598, + 0.01996530219912529, + 0.020000837743282318, + 0.01978798396885395, + 0.013464015908539295, + 0.0292207021266222, + 0.021909214556217194, + 0.020475992932915688, + 0.02400357462465763, + 0.010515469126403332, + 0.014395755715668201, + 0.020030366256833076, + 0.0057490309700369835, + 0.005106293596327305, + 0.009603352285921574, + 0.00870219711214304, + 0.009986085817217827, + 0.009272647090256214, + 0.015818174928426743, + 0.017835713922977448, + 0.006246289238333702, + 0.0057058800011873245, + 0.015854110941290855, + 0.015453148633241653, + 0.011548304930329323, + 0.008165708743035793, + 0.01581372693181038, + 0.01588226668536663, + 0.0035695023834705353, + 0.008687261492013931, + 0.00940194446593523, + 0.008617972023785114, + 0.007802573498338461, + 0.006794769782572985, + 0.009971155785024166, + 0.009497520513832569, + 0.0075437449850142, + 0.008593066595494747, + 0.015401297248899937, + 0.016403691843152046, + 0.00998140312731266, + 0.01389151718467474, + 0.01588219404220581, + 0.015932584181427956, + 0.003995301201939583, + 0.005786808207631111, + 0.004190734587609768, + 0.009097807109355927, + 0.008003376424312592, + 0.006903065834194422, + 0.008293340913951397, + 0.008928780443966389 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 4, + 8 + ], + "selected_reward": -0.024101020768284798 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an ancient wooden cargo ship with towering masts?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are dark brown oak barrels piled high on the ship's deck, secured with rusty iron hoops?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the environment a turbulent black sea with mountainous waves topped with pale foam under an overcast sky with lightning flashes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the barrels on the deck roll and collide while the ship pitches between wave crests and troughs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera perform a simultaneous slow zoom forward and pull back, keeping the ship's size constant while stretching the background?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.062060609459877014 + ], + "opposite_logprob_tokens": [ + -2.812060594558716 + ], + "logprob_reward": -0.062060609459877014, + "yes_no_margin_reward": 2.75, + "token_credit_reward": -0.062060609459877014, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11350798606872559, + 0.16096217930316925, + 0.08188959211111069, + 0.12116781622171402, + 0.10309730470180511, + 0.10437041521072388, + 0.12008221447467804, + 0.08347334712743759, + 0.053492043167352676, + 0.057957109063863754 + ], + "credit_map": [ + 0.012894902378320694, + 0.007817089557647705, + 0.01586339622735977, + 0.02083241008222103, + 0.008081821724772453, + 0.0190641600638628, + 0.018527675420045853, + 0.010426520369946957, + 0.009731925092637539, + 0.006471207365393639, + 0.0490432046353817, + 0.017349304631352425, + 0.00860069040209055, + 0.01423014048486948, + 0.038893043994903564, + 0.01664266176521778, + 0.007245014887303114, + 0.010668493807315826, + 0.014425189234316349, + 0.012659450061619282, + 0.008279407396912575, + 0.007743946276605129, + 0.009658530354499817, + 0.011209560558199883, + 0.007163095287978649, + 0.017068779096007347, + 0.018204066902399063, + 0.010051957331597805, + 0.021006664261221886, + 0.015771912410855293, + 0.020083677023649216, + 0.011817664839327335, + 0.008872881531715393, + 0.016604693606495857, + 0.015645086765289307, + 0.017034487798810005, + 0.01306704431772232, + 0.007852574810385704, + 0.0100908437743783, + 0.013929695822298527, + 0.008790051564574242, + 0.009001304395496845, + 0.011829621158540249, + 0.01312904991209507, + 0.01282757893204689, + 0.014049048535525799, + 0.015507921576499939, + 0.019235841929912567, + 0.008747235871851444, + 0.008305283263325691, + 0.02051456645131111, + 0.027073664590716362, + 0.013830618001520634, + 0.010460064746439457, + 0.01775403879582882, + 0.01339674461632967, + 0.006194571498781443, + 0.007981532253324986, + 0.016925904899835587, + 0.01076391153037548, + 0.006526496727019548, + 0.008242789655923843, + 0.01590799354016781, + 0.010930145159363747, + 0.003241009311750531, + 0.003721974091604352, + 0.0073272003792226315, + 0.009169002063572407, + 0.0032542645931243896, + 0.004561812151223421, + 0.010539221577346325, + 0.011677558533847332, + 0.0047584231942892075, + 0.006078158039599657, + 0.0063068498857319355, + 0.008797344751656055, + 0.008004944771528244, + 0.0049552470445632935, + 0.009448709897696972, + 0.009607427753508091 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 3, + 6 + ], + "selected_reward": -0.062060609459877014 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an ancient wooden cargo ship with towering masts?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are dark brown oak barrels piled high on the ship's deck, secured with rusty iron hoops?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the environment a turbulent black sea with mountainous waves topped with pale foam under an overcast sky with lightning flashes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the barrels on the deck roll and collide while the ship pitches between wave crests and troughs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera perform a simultaneous slow zoom forward and pull back, keeping the ship's size constant while stretching the background?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02634398825466633 + ], + "opposite_logprob_tokens": [ + -3.651344060897827 + ], + "logprob_reward": -0.02634398825466633, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.02634398825466633, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.17761872708797455, + 0.08059521019458771, + 0.07437540590763092, + 0.07066256552934647, + 0.051038727164268494, + 0.07634633779525757, + 0.07631119340658188, + 0.14151690900325775, + 0.08667387068271637, + 0.1648610681295395 + ], + "credit_map": [ + 0.0324549600481987, + 0.013276726938784122, + 0.02260628528892994, + 0.024250376969575882, + 0.015597227960824966, + 0.033513739705085754, + 0.022369908168911934, + 0.013549488969147205, + 0.004700477235019207, + 0.004924001172184944, + 0.0250601377338171, + 0.007816043682396412, + 0.0070008039474487305, + 0.006505720317363739, + 0.01353195309638977, + 0.011056065559387207, + 0.007198295556008816, + 0.008585856296122074, + 0.012736085802316666, + 0.006763341836631298, + 0.005262367427349091, + 0.011599702760577202, + 0.013088369742035866, + 0.009141379967331886, + 0.006982066202908754, + 0.01108411606401205, + 0.007894618436694145, + 0.009210855700075626, + 0.011888273991644382, + 0.007414809428155422, + 0.00919505674391985, + 0.006992766167968512, + 0.006355101242661476, + 0.00822571199387312, + 0.005741105880588293, + 0.005372960586100817, + 0.009282595477998257, + 0.005652296356856823, + 0.005316853988915682, + 0.0050920965149998665, + 0.00505614560097456, + 0.005623073782771826, + 0.008291708305478096, + 0.007139200810343027, + 0.00948028452694416, + 0.007696132641285658, + 0.015333645977079868, + 0.017726141959428787, + 0.005303626414388418, + 0.004990902729332447, + 0.01274125836789608, + 0.012244674377143383, + 0.007820995524525642, + 0.008733895607292652, + 0.011529265902936459, + 0.012946565635502338, + 0.011982319876551628, + 0.011397093534469604, + 0.031442925333976746, + 0.01570211909711361, + 0.0108103072270751, + 0.014184641651809216, + 0.02459094114601612, + 0.021406540647149086, + 0.008113786578178406, + 0.006886816117912531, + 0.01527352724224329, + 0.014145508408546448, + 0.005978002212941647, + 0.010628762654960155, + 0.012530232779681683, + 0.01311721932142973, + 0.012872333638370037, + 0.011784705333411694, + 0.014590410515666008, + 0.03426975756883621, + 0.011439433321356773, + 0.01936843805015087, + 0.028374454006552696, + 0.03216153755784035 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 7 + ], + "selected_reward": -0.02634398825466633 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an ancient wooden cargo ship with towering masts?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are dark brown oak barrels piled high on the ship's deck, secured with rusty iron hoops?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the environment a turbulent black sea with mountainous waves topped with pale foam under an overcast sky with lightning flashes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the barrels on the deck roll and collide while the ship pitches between wave crests and troughs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera perform a simultaneous slow zoom forward and pull back, keeping the ship's size constant while stretching the background?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.012538531795144081 + ], + "opposite_logprob_tokens": [ + -4.387538433074951 + ], + "logprob_reward": -0.012538531795144081, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.012538531795144081, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1280064433813095, + 0.09460684657096863, + 0.15243598818778992, + 0.13079644739627838, + 0.047350093722343445, + 0.1090637594461441, + 0.1294327825307846, + 0.057621367275714874, + 0.09231741726398468, + 0.058368854224681854 + ], + "credit_map": [ + 0.016189372166991234, + 0.013013417832553387, + 0.016591966152191162, + 0.01803404651582241, + 0.00946001149713993, + 0.018528930842876434, + 0.023467976599931717, + 0.012720727361738682, + 0.009402824565768242, + 0.006950829643756151, + 0.021072931587696075, + 0.009028682485222816, + 0.008482635021209717, + 0.013541903346776962, + 0.014895815402269363, + 0.011231223121285439, + 0.017155099660158157, + 0.012578952126204967, + 0.03052017278969288, + 0.014657540246844292, + 0.011361238546669483, + 0.012519991025328636, + 0.016265815123915672, + 0.037377189844846725, + 0.013014818541705608, + 0.01024335715919733, + 0.02057557925581932, + 0.013432824052870274, + 0.01202476304024458, + 0.017391668632626534, + 0.027348756790161133, + 0.01676468551158905, + 0.007840799167752266, + 0.004266579635441303, + 0.006960626225918531, + 0.006774126086384058, + 0.004477341193705797, + 0.005258440505713224, + 0.005577773321419954, + 0.006194406654685736, + 0.006873483769595623, + 0.010974723845720291, + 0.015086905099451542, + 0.01572483964264393, + 0.012404235079884529, + 0.014326240867376328, + 0.01407527830451727, + 0.01959805190563202, + 0.007539512123912573, + 0.009832561947405338, + 0.020921597257256508, + 0.03272624686360359, + 0.012786482460796833, + 0.00830400176346302, + 0.02094416134059429, + 0.016378220170736313, + 0.004089744761586189, + 0.0043836296536028385, + 0.010985114611685276, + 0.007343733217567205, + 0.004604090005159378, + 0.006593984551727772, + 0.01246323250234127, + 0.007157835178077221, + 0.006564602255821228, + 0.009200753644108772, + 0.012737169861793518, + 0.015019034035503864, + 0.00943086389452219, + 0.013155372813344002, + 0.013543290086090565, + 0.012666333466768265, + 0.004878540523350239, + 0.004701311234384775, + 0.006494264118373394, + 0.011968401260674, + 0.00504343630746007, + 0.006742129102349281, + 0.00887933000922203, + 0.009661444462835789 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 3, + 6 + ], + "selected_reward": -0.012538531795144081 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an ancient wooden cargo ship with towering masts?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are dark brown oak barrels piled high on the ship's deck, secured with rusty iron hoops?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the environment a turbulent black sea with mountainous waves topped with pale foam under an overcast sky with lightning flashes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the barrels on the deck roll and collide while the ship pitches between wave crests and troughs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera perform a simultaneous slow zoom forward and pull back, keeping the ship's size constant while stretching the background?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -0.6933955550193787 + ], + "opposite_logprob_tokens": [ + -0.6933955550193787 + ], + "logprob_reward": -0.6933955550193787, + "yes_no_margin_reward": 0.0, + "token_credit_reward": -0.6933955550193787, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.18035003542900085, + 0.11729094386100769, + 0.07037532329559326, + 0.0615682452917099, + 0.06007178872823715, + 0.07923007756471634, + 0.16499483585357666, + 0.14197596907615662, + 0.10034340620040894, + 0.02379932627081871 + ], + "credit_map": [ + 0.019566841423511505, + 0.018773531541228294, + 0.02552379108965397, + 0.0324217788875103, + 0.013244573958218098, + 0.019212426617741585, + 0.03518550097942352, + 0.016421597450971603, + 0.005718855652958155, + 0.006110760383307934, + 0.03670137748122215, + 0.012422014959156513, + 0.006861594505608082, + 0.008113821968436241, + 0.03082999400794506, + 0.010532524436712265, + 0.007770310156047344, + 0.004335988778620958, + 0.015243222005665302, + 0.0046484158374369144, + 0.003810314694419503, + 0.005413626786321402, + 0.008963673375546932, + 0.020189780741930008, + 0.00706733763217926, + 0.008665106259286404, + 0.007549640256911516, + 0.006817758549004793, + 0.009582175873219967, + 0.005857477895915508, + 0.009105414152145386, + 0.0069233388639986515, + 0.0073606218211352825, + 0.007035022135823965, + 0.0081892479211092, + 0.007295429240912199, + 0.008887329138815403, + 0.007352516055107117, + 0.007078318856656551, + 0.006873309146612883, + 0.005744203459471464, + 0.008059489540755749, + 0.012217084877192974, + 0.011311248876154423, + 0.006767184007912874, + 0.011901301331818104, + 0.011523719877004623, + 0.011705859564244747, + 0.009405119344592094, + 0.010816053487360477, + 0.02954748272895813, + 0.0449068583548069, + 0.014134293422102928, + 0.00793700572103262, + 0.029050303623080254, + 0.019197730347514153, + 0.013336180709302425, + 0.00925314612686634, + 0.027200400829315186, + 0.019556986168026924, + 0.010586935095489025, + 0.017102649435400963, + 0.020874731242656708, + 0.024064956232905388, + 0.006829426623880863, + 0.00905933603644371, + 0.013143588788807392, + 0.014139109291136265, + 0.009838856756687164, + 0.01555776409804821, + 0.014993500895798206, + 0.01678183116018772, + 0.002860060427337885, + 0.002308549825102091, + 0.00294527062214911, + 0.003638147609308362, + 0.0025198361836373806, + 0.0025862683542072773, + 0.003169999225065112, + 0.0037711970508098602 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 6, + 7 + ], + "selected_reward": -0.6933955550193787 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "No" + ] + }, + { + "prompt": "classic Hitchcock zoom lens, from a slightly bird's-eye view, captures an ancient wooden cargo ship struggling in the turbulent black sea. The inky black waters whip up mountainous waves, their crests ripped into pale foam by the wind. The sky is overcast, and occasional flashes of lightning provide the sole source of light, momentarily illuminating everything amidst the torrential downpour. The ship's hull is a weathered, dark timber, its towering masts groaning painfully in the storm. The decks are piled high with dark brown oak barrels secured with rusty iron hoops. These barrels roll and collide perilously, clattering with a dull roar as the ship pitches devastatingly between the crests and troughs of the waves. As the camera slowly zooms forward, it simultaneously pulls back, causing the ship's size to remain nearly constant within the frame while the raging sea and swirling clouds in the background appear dramatically stretched, as if time and space itself were warping and expanding. This perfectly recreates the hallmark, dizzying dreamlike quality of Hitchcock's films, creating an inescapable, surreal atmosphere of tension and dread.", + "seed": 15751, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000095/group_6/index_3152_seed_15751_dp_13_rank_26_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 6, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.1659644627943635, + "vlm_logprob": -0.1659644627943635, + "vlm_yes_no_margin": 3.3, + "vlm_token_credit": -0.1659644627943635, + "vlm_yes_ratio": 0.8, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an ancient wooden cargo ship with towering masts?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are dark brown oak barrels piled high on the ship's deck, secured with rusty iron hoops?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the environment a turbulent black sea with mountainous waves topped with pale foam under an overcast sky with lightning flashes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the barrels on the deck roll and collide while the ship pitches between wave crests and troughs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera perform a simultaneous slow zoom forward and pull back, keeping the ship's size constant while stretching the background?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.013269447721540928 + ], + "opposite_logprob_tokens": [ + -4.763269424438477 + ], + "logprob_reward": -0.013269447721540928, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.013269447721540928, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.03960948809981346, + 0.016212325543165207, + 0.025849599391222, + 0.03325984999537468, + 0.5317067503929138, + 0.02666962891817093, + 0.015562553890049458, + 0.01874467544257641, + 0.23108039796352386, + 0.061304692178964615 + ], + "credit_map": [ + 0.003724656766280532, + 0.004470010753720999, + 0.008078651502728462, + 0.0031215010676532984, + 0.003682323032990098, + 0.004419813398271799, + 0.00549235288053751, + 0.006620178930461407, + 0.001678989501670003, + 0.002004132606089115, + 0.001902270712889731, + 0.00283678132109344, + 0.001960568595677614, + 0.001342466101050377, + 0.0018327611032873392, + 0.0026543543208390474, + 0.0034360974095761776, + 0.0031313207000494003, + 0.0036418940871953964, + 0.0035071398597210646, + 0.00557000283151865, + 0.0022038251627236605, + 0.00201226444914937, + 0.0023470541927963495, + 0.003469482297077775, + 0.0052631450816988945, + 0.003844128455966711, + 0.004932726267725229, + 0.0053641581907868385, + 0.004224652890115976, + 0.0030268304981291294, + 0.003134726779535413, + 0.023260483518242836, + 0.02038556896150112, + 0.19022853672504425, + 0.20467057824134827, + 0.010985859669744968, + 0.017217738553881645, + 0.037885189056396484, + 0.027072809636592865, + 0.002846652641892433, + 0.003208120819181204, + 0.003378658788278699, + 0.0035017523914575577, + 0.0028214033227413893, + 0.003650778904557228, + 0.004943068604916334, + 0.002319194143638015, + 0.001410777447745204, + 0.0022174417972564697, + 0.001507679233327508, + 0.0014634758699685335, + 0.001554590300656855, + 0.0031085594091564417, + 0.00228119152598083, + 0.002018839120864868, + 0.00214572437107563, + 0.0012090150266885757, + 0.0033604849595576525, + 0.0020721086766570807, + 0.00121786305680871, + 0.0026665739715099335, + 0.0032429611310362816, + 0.002829945180565119, + 0.0062682367861270905, + 0.006619012914597988, + 0.023353086784482002, + 0.016599440947175026, + 0.0078126210719347, + 0.006335894111543894, + 0.09913560748100281, + 0.06495649367570877, + 0.005065632052719593, + 0.0065687596797943115, + 0.011204692535102367, + 0.006226785015314817, + 0.006687050685286522, + 0.010798467323184013, + 0.006136984098702669, + 0.008616317994892597 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 8, + 9 + ], + "selected_reward": -0.013269447721540928 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an ancient wooden cargo ship with towering masts?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are dark brown oak barrels piled high on the ship's deck, secured with rusty iron hoops?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the environment a turbulent black sea with mountainous waves topped with pale foam under an overcast sky with lightning flashes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the barrels on the deck roll and collide while the ship pitches between wave crests and troughs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera perform a simultaneous slow zoom forward and pull back, keeping the ship's size constant while stretching the background?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.014203879982233047 + ], + "opposite_logprob_tokens": [ + -4.264204025268555 + ], + "logprob_reward": -0.014203879982233047, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.014203879982233047, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.056256379932165146, + 0.019739890471100807, + 0.02634546533226967, + 0.012746772728860378, + 0.10471507906913757, + 0.03215806186199188, + 0.07942003756761551, + 0.030545925721526146, + 0.48323023319244385, + 0.15484215319156647 + ], + "credit_map": [ + 0.005347935948520899, + 0.005112441256642342, + 0.007157338783144951, + 0.004642406012862921, + 0.005374366417527199, + 0.012933333404362202, + 0.006313933525234461, + 0.009374622255563736, + 0.001721267239190638, + 0.0016198026714846492, + 0.002285915892571211, + 0.0032436177134513855, + 0.0027362974360585213, + 0.0021127029322087765, + 0.003336554393172264, + 0.002683729864656925, + 0.0036070419009774923, + 0.002964309649541974, + 0.0032933198381215334, + 0.002021500375121832, + 0.004563652444630861, + 0.003290997352451086, + 0.002569166012108326, + 0.004035477060824633, + 0.0013171108439564705, + 0.0021296485792845488, + 0.001365411328151822, + 0.0011619019787758589, + 0.0026439055800437927, + 0.0018332557519897819, + 0.0010977425845339894, + 0.0011977952672168612, + 0.007786234375089407, + 0.004626289941370487, + 0.023066971451044083, + 0.032170649617910385, + 0.007368078455328941, + 0.005418586544692516, + 0.01719881407916546, + 0.007079453207552433, + 0.003684737253934145, + 0.004266528878360987, + 0.005706030409783125, + 0.004804814700037241, + 0.0026829277630895376, + 0.003436494152992964, + 0.004878811072558165, + 0.002697714138776064, + 0.00951859075576067, + 0.009166563861072063, + 0.009079168550670147, + 0.007585334125906229, + 0.0066824317909777164, + 0.013731990940868855, + 0.011646856553852558, + 0.012009101919829845, + 0.0031964280642569065, + 0.003451928962022066, + 0.0035681764129549265, + 0.00406116247177124, + 0.0037567424587905407, + 0.004704570863395929, + 0.0035825613886117935, + 0.004224355798214674, + 0.013346740044653416, + 0.014235656708478928, + 0.047007832676172256, + 0.03471332788467407, + 0.016707906499505043, + 0.015195004642009735, + 0.20506860315799713, + 0.136955127120018, + 0.009479939937591553, + 0.012069652788341045, + 0.03253977745771408, + 0.015504133887588978, + 0.020179670304059982, + 0.028911635279655457, + 0.016437841579318047, + 0.01971948705613613 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 9, + 4 + ], + "selected_reward": -0.014203879982233047 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an ancient wooden cargo ship with towering masts?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are dark brown oak barrels piled high on the ship's deck, secured with rusty iron hoops?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the environment a turbulent black sea with mountainous waves topped with pale foam under an overcast sky with lightning flashes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the barrels on the deck roll and collide while the ship pitches between wave crests and troughs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera perform a simultaneous slow zoom forward and pull back, keeping the ship's size constant while stretching the background?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.026340853422880173 + ], + "opposite_logprob_tokens": [ + -3.651340961456299 + ], + "logprob_reward": -0.026340853422880173, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.026340853422880173, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3168308436870575, + 0.05333711951971054, + 0.09311793744564056, + 0.03688615933060646, + 0.1535276174545288, + 0.02678021974861622, + 0.03622964024543762, + 0.029341991990804672, + 0.045390911400318146, + 0.20855751633644104 + ], + "credit_map": [ + 0.026571890339255333, + 0.02082356996834278, + 0.07704152911901474, + 0.020046332851052284, + 0.01904769241809845, + 0.06294421851634979, + 0.04222332686185837, + 0.048132311552762985, + 0.0046012927778065205, + 0.004332883283495903, + 0.0070017543621361256, + 0.009768851101398468, + 0.005494635086506605, + 0.004990690853446722, + 0.008690683171153069, + 0.008456334471702576, + 0.016799772158265114, + 0.010297169908881187, + 0.012910455465316772, + 0.008303548209369183, + 0.012442370876669884, + 0.011331716552376747, + 0.008386911824345589, + 0.012645996175706387, + 0.004039816092699766, + 0.004672545939683914, + 0.006367632187902927, + 0.004122514743357897, + 0.004021085798740387, + 0.0038150339387357235, + 0.004717463627457619, + 0.005130066070705652, + 0.00826962199062109, + 0.005990287754684687, + 0.0505349263548851, + 0.05687501281499863, + 0.004689532332122326, + 0.005549982655793428, + 0.01366878766566515, + 0.007949473336338997, + 0.004701460245996714, + 0.0023678315337747335, + 0.00398672791197896, + 0.004087772686034441, + 0.0023423884995281696, + 0.003019909607246518, + 0.0037691008765250444, + 0.002505031181499362, + 0.0031926792580634356, + 0.0058331177569925785, + 0.0037030596286058426, + 0.004483154974877834, + 0.004055507481098175, + 0.004531355109065771, + 0.005454240366816521, + 0.00497653242200613, + 0.0032944052945822477, + 0.002967714099213481, + 0.003767356975004077, + 0.0045064291916787624, + 0.002768912585452199, + 0.0040613203309476376, + 0.0041279261931777, + 0.0038479308132082224, + 0.004363775718957186, + 0.0038005977403372526, + 0.004878383595496416, + 0.0035717671271413565, + 0.004287088289856911, + 0.0049258447252213955, + 0.011133581399917603, + 0.0084298737347126, + 0.014837538823485374, + 0.015687093138694763, + 0.0470007099211216, + 0.01976335234940052, + 0.018452739343047142, + 0.04148781672120094, + 0.021933402866125107, + 0.029394876211881638 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 4 + ], + "selected_reward": -0.026340853422880173 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an ancient wooden cargo ship with towering masts?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are dark brown oak barrels piled high on the ship's deck, secured with rusty iron hoops?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the environment a turbulent black sea with mountainous waves topped with pale foam under an overcast sky with lightning flashes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the barrels on the deck roll and collide while the ship pitches between wave crests and troughs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera perform a simultaneous slow zoom forward and pull back, keeping the ship's size constant while stretching the background?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.018181685358285904 + ], + "opposite_logprob_tokens": [ + -4.018181800842285 + ], + "logprob_reward": -0.018181685358285904, + "yes_no_margin_reward": 4.0, + "token_credit_reward": -0.018181685358285904, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.029436033219099045, + 0.029348060488700867, + 0.04644106701016426, + 0.023500438779592514, + 0.6579645872116089, + 0.024620801210403442, + 0.04294353350996971, + 0.0226809773594141, + 0.07075071334838867, + 0.05231379345059395 + ], + "credit_map": [ + 0.00280366325750947, + 0.0024632522836327553, + 0.005362818483263254, + 0.0035871227737516165, + 0.0024259581696242094, + 0.004811897408217192, + 0.004170722793787718, + 0.0038105964194983244, + 0.003378717228770256, + 0.003397636814042926, + 0.003598510753363371, + 0.004230434074997902, + 0.004167095758020878, + 0.003437938168644905, + 0.0031660855747759342, + 0.003971643280237913, + 0.006926567293703556, + 0.006821079179644585, + 0.004981862846761942, + 0.0042250449769198895, + 0.0098177595064044, + 0.003883174853399396, + 0.0045048510655760765, + 0.005280728451907635, + 0.002725380938500166, + 0.0037996803876012564, + 0.002971887355670333, + 0.002500982955098152, + 0.0045432583428919315, + 0.002720350632444024, + 0.0022533887531608343, + 0.001985509181395173, + 0.030880242586135864, + 0.025046657770872116, + 0.23223893344402313, + 0.2514547109603882, + 0.014191346243023872, + 0.02130250446498394, + 0.051308728754520416, + 0.031541451811790466, + 0.004230405669659376, + 0.002194343600422144, + 0.0025978353805840015, + 0.004456769675016403, + 0.0018278839997947216, + 0.0025042700581252575, + 0.00414681900292635, + 0.0026624747551977634, + 0.005269801244139671, + 0.003877734299749136, + 0.005641072057187557, + 0.004770707804709673, + 0.003313386579975486, + 0.007035406306385994, + 0.006306607276201248, + 0.006728819105774164, + 0.0022302314173430204, + 0.0020130539778620005, + 0.003790083108469844, + 0.002617516787722707, + 0.0019578197970986366, + 0.00377851165831089, + 0.002835674211382866, + 0.0034580864012241364, + 0.002606774680316448, + 0.0027366664726287127, + 0.007849347777664661, + 0.005080414470285177, + 0.002863729139789939, + 0.0028175634797662497, + 0.028567545115947723, + 0.018228668719530106, + 0.005391506943851709, + 0.004525320138782263, + 0.011020726524293423, + 0.004971199203282595, + 0.0050401934422552586, + 0.008424333296716213, + 0.005898982286453247, + 0.007041529286652803 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 8, + 9 + ], + "selected_reward": -0.018181685358285904 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an ancient wooden cargo ship with towering masts?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are dark brown oak barrels piled high on the ship's deck, secured with rusty iron hoops?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the environment a turbulent black sea with mountainous waves topped with pale foam under an overcast sky with lightning flashes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the barrels on the deck roll and collide while the ship pitches between wave crests and troughs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera perform a simultaneous slow zoom forward and pull back, keeping the ship's size constant while stretching the background?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -0.7578264474868774 + ], + "opposite_logprob_tokens": [ + -0.6328264474868774 + ], + "logprob_reward": -0.7578264474868774, + "yes_no_margin_reward": -0.125, + "token_credit_reward": -0.7578264474868774, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.06682819873094559, + 0.017938608303666115, + 0.021572479978203773, + 0.02082996256649494, + 0.6323162913322449, + 0.02672462910413742, + 0.1063215434551239, + 0.047088369727134705, + 0.028124293312430382, + 0.0322556309401989 + ], + "credit_map": [ + 0.004129840061068535, + 0.0045005870051681995, + 0.01830342970788479, + 0.006596776656806469, + 0.004516094923019409, + 0.007351549342274666, + 0.011847443878650665, + 0.009582480415701866, + 0.0015386211453005672, + 0.0019089451525360346, + 0.0022625054698437452, + 0.002915550721809268, + 0.00220870622433722, + 0.0015579767059534788, + 0.0019849278032779694, + 0.003561377990990877, + 0.004270251840353012, + 0.00227547250688076, + 0.0022258670069277287, + 0.0021880713757127523, + 0.004261249676346779, + 0.0020231958478689194, + 0.0019387901993468404, + 0.0023895809426903725, + 0.002362461993470788, + 0.002244890434667468, + 0.0025444754865020514, + 0.0019911560229957104, + 0.0032336770091205835, + 0.0027579525485634804, + 0.00222327234223485, + 0.003472078824415803, + 0.02841765433549881, + 0.024069417268037796, + 0.22729584574699402, + 0.2432698905467987, + 0.011361324228346348, + 0.020457273349165916, + 0.04602757468819618, + 0.03141731396317482, + 0.0023391626309603453, + 0.002424049424007535, + 0.004902102518826723, + 0.004131088498979807, + 0.002063590567559004, + 0.00419620843604207, + 0.0044426280073821545, + 0.002225798787549138, + 0.011867126449942589, + 0.011990760453045368, + 0.012118995189666748, + 0.009655281901359558, + 0.009552913717925549, + 0.017421327531337738, + 0.015711355954408646, + 0.018003784120082855, + 0.007566755637526512, + 0.00412090914323926, + 0.00924910232424736, + 0.00510262418538332, + 0.0034523573704063892, + 0.007153308484703302, + 0.004015674814581871, + 0.006427639629691839, + 0.002298180013895035, + 0.0018580042524263263, + 0.004163665696978569, + 0.0025306374300271273, + 0.0025765348691493273, + 0.0036208860110491514, + 0.00736355921253562, + 0.0037128268741071224, + 0.00398072088137269, + 0.0031011018436402082, + 0.003648376790806651, + 0.003689425764605403, + 0.0041821058839559555, + 0.005751267075538635, + 0.0031355146784335375, + 0.004767121747136116 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 6, + 0 + ], + "selected_reward": -0.7578264474868774 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "No" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000095/group_7/index_tensor([6920, 6920])_group_7_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000095/group_7/index_tensor([6920, 6920])_group_7_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..2743744b5d45ffe4d5ace5414835485d64b99815 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000095/group_7/index_tensor([6920, 6920])_group_7_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "Shot from a slightly elevated angle, the extreme wide shot begins with a slow, majestic sideways trajectory across the deep, silent abyss. Ethereal pillars of sunlight penetrate the distant surface, creating a shifting divine light that illuminates countless particles of marine snow drifting in the current. The water itself is a breathtaking gradient, transitioning from a brilliant cyan near the top to a deep, mysterious indigo in the unseen depths below. At the center of this vast azure void, a lone sea turtle swims serenely, its form diminutive against the vast backdrop. It propels itself forward with the powerful yet graceful, slow, rhythmic sweeps of its enormous front flippers, each movement deliberate and effortless. The light captures the intricate mosaic of its brown and olive-green carapace and the leathery, ancient texture of its skin, creating a solitary, majestic figure on an endless journey through the tranquil grandeur of the underwater world, evoking an atmosphere of tranquil solitude.", + "seed": 20675, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000095/group_7/index_6920_seed_20675_dp_14_rank_28_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 7, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.07667647814378142, + "vlm_logprob": -0.07667647814378142, + "vlm_yes_no_margin": 3.625, + "vlm_token_credit": -0.07667647814378142, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set underwater with a gradient transitioning from cyan at the top to deep indigo below?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a lone sea turtle visible swimming at the center of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the turtle move using slow, rhythmic sweeps of its enormous front flippers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the turtle's carapace visible with a brown and olive-green mosaic pattern?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot an extreme wide shot with a slow sideways trajectory?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.010126405395567417 + ], + "opposite_logprob_tokens": [ + -4.88512659072876 + ], + "logprob_reward": -0.010126405395567417, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.010126405395567417, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.49421292543411255, + 0.06124042719602585, + 0.03816600888967514, + 0.06515819579362869, + 0.20503702759742737, + 0.03373127058148384, + 0.030779315158724785, + 0.025750352069735527, + 0.014163988642394543, + 0.031760457903146744 + ], + "credit_map": [ + 0.029613453894853592, + 0.025997865945100784, + 0.07418600469827652, + 0.019054129719734192, + 0.06354692578315735, + 0.21389199793338776, + 0.014229366555809975, + 0.05369323119521141, + 0.002877061255276203, + 0.0035521024838089943, + 0.010265626013278961, + 0.0037571536377072334, + 0.006668253801763058, + 0.023784339427947998, + 0.0038492644671350718, + 0.006486632861196995, + 0.002742806216701865, + 0.0028453113045543432, + 0.006845422554761171, + 0.0028059452306479216, + 0.004060103092342615, + 0.011635747738182545, + 0.0035392746794968843, + 0.0036914008669555187, + 0.0038737605791538954, + 0.00309216626919806, + 0.011632155627012253, + 0.0033020887058228254, + 0.007967018522322178, + 0.025469593703746796, + 0.004143689759075642, + 0.005677729845046997, + 0.006795130670070648, + 0.008282795548439026, + 0.02961849421262741, + 0.007932178676128387, + 0.02377139963209629, + 0.08611973375082016, + 0.007841766811907291, + 0.03467554971575737, + 0.0032089303713291883, + 0.0026609476190060377, + 0.005596695002168417, + 0.004517738241702318, + 0.004389578942209482, + 0.0059583294205367565, + 0.004060936626046896, + 0.003338115056976676, + 0.002017175778746605, + 0.0019430529791861773, + 0.0038527026772499084, + 0.0033930677454918623, + 0.0033307350240647793, + 0.009803161025047302, + 0.0026955457869917154, + 0.0037438776344060898, + 0.0014578336849808693, + 0.0021640148479491472, + 0.003103021765127778, + 0.004250771831721067, + 0.0029278479050844908, + 0.004302775487303734, + 0.005954352673143148, + 0.0015897363191470504, + 0.0008184399921447039, + 0.0011031203903257847, + 0.0015746467979624867, + 0.0021380700636655092, + 0.002039455808699131, + 0.003507116809487343, + 0.0018040540162473917, + 0.0011790855787694454, + 0.00246056099422276, + 0.002840040484443307, + 0.004151598550379276, + 0.00457588117569685, + 0.0038141191471368074, + 0.006161731667816639, + 0.004453942645341158, + 0.0033025885932147503 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 3 + ], + "selected_reward": -0.010126405395567417 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set underwater with a gradient transitioning from cyan at the top to deep indigo below?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a lone sea turtle visible swimming at the center of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the turtle move using slow, rhythmic sweeps of its enormous front flippers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the turtle's carapace visible with a brown and olive-green mosaic pattern?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot an extreme wide shot with a slow sideways trajectory?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.006745188031345606 + ], + "opposite_logprob_tokens": [ + -5.006745338439941 + ], + "logprob_reward": -0.006745188031345606, + "yes_no_margin_reward": 5.0, + "token_credit_reward": -0.006745188031345606, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.6096665859222412, + 0.08783628046512604, + 0.03032536618411541, + 0.03833066299557686, + 0.07543759047985077, + 0.028952457010746002, + 0.02277335338294506, + 0.05117417126893997, + 0.03817528858780861, + 0.017328297719359398 + ], + "credit_map": [ + 0.051616787910461426, + 0.0238984115421772, + 0.06873936951160431, + 0.027986301109194756, + 0.08203405141830444, + 0.2558042109012604, + 0.024704856798052788, + 0.07488252967596054, + 0.004566180985420942, + 0.00399541761726141, + 0.013758450746536255, + 0.007191894110292196, + 0.009478462859988213, + 0.031934015452861786, + 0.007721402682363987, + 0.009190445765852928, + 0.0019110451685264707, + 0.0024511057417839766, + 0.004888738505542278, + 0.0039022830314934254, + 0.0034246069844812155, + 0.00640122452750802, + 0.004570983350276947, + 0.002775374101474881, + 0.0017739234026521444, + 0.002659647725522518, + 0.0041592903435230255, + 0.005308825522661209, + 0.0034544854424893856, + 0.010267792269587517, + 0.006308732554316521, + 0.00439795758575201, + 0.0030667749233543873, + 0.007151324767619371, + 0.006999851204454899, + 0.003847001353278756, + 0.007520805113017559, + 0.02834738977253437, + 0.007130193989723921, + 0.011374244466423988, + 0.002379924524575472, + 0.0019919793121516705, + 0.0029495172202587128, + 0.0048193722032010555, + 0.0031395587138831615, + 0.0033867426682263613, + 0.007160274777561426, + 0.0031250847969204187, + 0.0013574576005339622, + 0.0014341421192511916, + 0.002381410915404558, + 0.0035926697310060263, + 0.002851314377039671, + 0.005828056950122118, + 0.0033284048549830914, + 0.001999894855543971, + 0.002037745201960206, + 0.0034646790008991957, + 0.005655425600707531, + 0.0078071230091154575, + 0.007148043718189001, + 0.010163537226617336, + 0.01173569168895483, + 0.0031619216315448284, + 0.0018631991697475314, + 0.0017946006264537573, + 0.005188873969018459, + 0.004678203258663416, + 0.003609637962654233, + 0.010499581694602966, + 0.007312984671443701, + 0.003228201298043132, + 0.001145500922575593, + 0.0014077721862122416, + 0.0021526296623051167, + 0.003087009070441127, + 0.0017040780512616038, + 0.0030527980998158455, + 0.003517028409987688, + 0.0012614800361916423 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 4 + ], + "selected_reward": -0.006745188031345606 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set underwater with a gradient transitioning from cyan at the top to deep indigo below?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a lone sea turtle visible swimming at the center of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the turtle move using slow, rhythmic sweeps of its enormous front flippers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the turtle's carapace visible with a brown and olive-green mosaic pattern?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot an extreme wide shot with a slow sideways trajectory?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.029806768521666527 + ], + "opposite_logprob_tokens": [ + -3.5298068523406982 + ], + "logprob_reward": -0.029806768521666527, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.029806768521666527, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.4677223861217499, + 0.09965113550424576, + 0.04794421046972275, + 0.10194038599729538, + 0.1072077751159668, + 0.024738838896155357, + 0.037341225892305374, + 0.03960898518562317, + 0.04264187812805176, + 0.031203139573335648 + ], + "credit_map": [ + 0.021887384355068207, + 0.024853115901350975, + 0.05660459026694298, + 0.01873939111828804, + 0.057423967868089676, + 0.19954976439476013, + 0.022132596001029015, + 0.06653158366680145, + 0.0060668159276247025, + 0.005965894088149071, + 0.010377268306910992, + 0.0124980378895998, + 0.012629318982362747, + 0.0187360942363739, + 0.022856000810861588, + 0.01052170805633068, + 0.002619873732328415, + 0.003596025751903653, + 0.0047781802713871, + 0.007628383114933968, + 0.004916754085570574, + 0.006889252457767725, + 0.012696114368736744, + 0.004819626919925213, + 0.005090747028589249, + 0.004408185835927725, + 0.015420114621520042, + 0.0058998349122703075, + 0.012207888066768646, + 0.038676295429468155, + 0.010026898235082626, + 0.010210423730313778, + 0.0037461607716977596, + 0.007499256171286106, + 0.013827881775796413, + 0.006748717278242111, + 0.010327585972845554, + 0.0412144772708416, + 0.009963450953364372, + 0.013880247250199318, + 0.0015109553933143616, + 0.0017738882452249527, + 0.003469125833362341, + 0.0030993714462965727, + 0.002516221720725298, + 0.005107212346047163, + 0.004411773290485144, + 0.0028502901550382376, + 0.00237415567971766, + 0.002453793305903673, + 0.006265901494771242, + 0.007090060971677303, + 0.004811741411685944, + 0.005937262438237667, + 0.005261395126581192, + 0.0031469156965613365, + 0.001992244506254792, + 0.004235436674207449, + 0.005244226660579443, + 0.005784736480563879, + 0.00431048683822155, + 0.006596124731004238, + 0.008433930575847626, + 0.003011799417436123, + 0.0023571737110614777, + 0.0032509632874280214, + 0.004202914424240589, + 0.008571387268602848, + 0.004527675919234753, + 0.005729636177420616, + 0.010612234473228455, + 0.003389897756278515, + 0.0019079853082075715, + 0.0017624108586460352, + 0.0053404844366014, + 0.0035713273100554943, + 0.00285144429653883, + 0.008214466273784637, + 0.005190971307456493, + 0.002364050131291151 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 3 + ], + "selected_reward": -0.029806768521666527 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set underwater with a gradient transitioning from cyan at the top to deep indigo below?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a lone sea turtle visible swimming at the center of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the turtle move using slow, rhythmic sweeps of its enormous front flippers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the turtle's carapace visible with a brown and olive-green mosaic pattern?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot an extreme wide shot with a slow sideways trajectory?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.023287607356905937 + ], + "opposite_logprob_tokens": [ + -3.773287534713745 + ], + "logprob_reward": -0.023287607356905937, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.023287607356905937, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.4851900339126587, + 0.0909043401479721, + 0.03367313742637634, + 0.05453724414110184, + 0.12933072447776794, + 0.07354684174060822, + 0.019860578700900078, + 0.060402389615774155, + 0.024089451879262924, + 0.028465263545513153 + ], + "credit_map": [ + 0.04235752671957016, + 0.01860031858086586, + 0.04209914058446884, + 0.02372020296752453, + 0.06435663998126984, + 0.19663263857364655, + 0.02718793973326683, + 0.07023563981056213, + 0.00528974924236536, + 0.004830596502870321, + 0.01085802260786295, + 0.010931327939033508, + 0.009075948968529701, + 0.022083215415477753, + 0.016859836876392365, + 0.010975644923746586, + 0.0027745047118514776, + 0.0027597367297858, + 0.00428122328594327, + 0.004354925360530615, + 0.00352561310864985, + 0.005988894030451775, + 0.006048665847629309, + 0.003939575981348753, + 0.0025260804686695337, + 0.0031360858120024204, + 0.006504278630018234, + 0.004572221543639898, + 0.0061261956579983234, + 0.018486706539988518, + 0.006111659575253725, + 0.007074020337313414, + 0.009003590792417526, + 0.013421879149973392, + 0.009960906580090523, + 0.01612754724919796, + 0.011682586744427681, + 0.03028026595711708, + 0.02322201617062092, + 0.01563192717730999, + 0.004912256728857756, + 0.00396141828969121, + 0.00826908741146326, + 0.007879511453211308, + 0.008113086223602295, + 0.020913342013955116, + 0.010580199770629406, + 0.00891793705523014, + 0.0016531208530068398, + 0.0014283304335549474, + 0.0018836937379091978, + 0.0030195193830877542, + 0.0025065152440220118, + 0.004800337366759777, + 0.0022587578278034925, + 0.0023103035055100918, + 0.0028286876622587442, + 0.0044351317919790745, + 0.006235156673938036, + 0.01090258453041315, + 0.005933103617280722, + 0.009038407355546951, + 0.01733933761715889, + 0.0036899810656905174, + 0.001511722570285201, + 0.0015425215242430568, + 0.0023578349500894547, + 0.004150426480919123, + 0.0031487008091062307, + 0.003342778654769063, + 0.005562749691307545, + 0.0024727170821279287, + 0.0016171680763363838, + 0.0023021912202239037, + 0.004253447987139225, + 0.003222559578716755, + 0.003343201009556651, + 0.007673701271414757, + 0.0035121708642691374, + 0.0025408233050256968 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 1 + ], + "selected_reward": -0.023287607356905937 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set underwater with a gradient transitioning from cyan at the top to deep indigo below?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a lone sea turtle visible swimming at the center of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the turtle move using slow, rhythmic sweeps of its enormous front flippers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the turtle's carapace visible with a brown and olive-green mosaic pattern?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot an extreme wide shot with a slow sideways trajectory?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.31341642141342163 + ], + "opposite_logprob_tokens": [ + -1.3134164810180664 + ], + "logprob_reward": -0.31341642141342163, + "yes_no_margin_reward": 1.0, + "token_credit_reward": -0.31341642141342163, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.6628779768943787, + 0.06871095299720764, + 0.023940542712807655, + 0.0821753665804863, + 0.023368170484900475, + 0.03148190677165985, + 0.028596501797437668, + 0.03374452143907547, + 0.02517786994576454, + 0.01992613822221756 + ], + "credit_map": [ + 0.039429545402526855, + 0.023679960519075394, + 0.11909264326095581, + 0.021876148879528046, + 0.08624011278152466, + 0.2896858751773834, + 0.019027968868613243, + 0.0638456866145134, + 0.006538716144859791, + 0.0049036783166229725, + 0.006382007151842117, + 0.0062158359214663506, + 0.010674286633729935, + 0.015614747069776058, + 0.009123342111706734, + 0.009258344769477844, + 0.001720037660561502, + 0.002808416960760951, + 0.00240290816873312, + 0.00241230521351099, + 0.004489256534725428, + 0.003219124162569642, + 0.004596675280481577, + 0.0022918193135410547, + 0.0032788419630378485, + 0.0028558329213410616, + 0.012378646992146969, + 0.004855492617934942, + 0.009868825785815716, + 0.03307012841105461, + 0.005446816328912973, + 0.010420787148177624, + 0.0022106426768004894, + 0.0017835680628195405, + 0.00368151324801147, + 0.003051133593544364, + 0.00218708417378366, + 0.004102534148842096, + 0.004104967229068279, + 0.002246727468445897, + 0.0019430008251219988, + 0.0018094658153131604, + 0.003511513350531459, + 0.005136388819664717, + 0.003910414408892393, + 0.005876030772924423, + 0.005736959632486105, + 0.003558133030310273, + 0.002177358139306307, + 0.002094345400109887, + 0.0034979013726115227, + 0.005053593777120113, + 0.003350483952090144, + 0.005293315276503563, + 0.004137732554227114, + 0.002991772722452879, + 0.0028717194218188524, + 0.0033340335357934237, + 0.0034719742834568024, + 0.004989063832908869, + 0.00262500811368227, + 0.004603805486112833, + 0.008611657656729221, + 0.003237255848944187, + 0.0014951126649975777, + 0.0022524830419570208, + 0.00279600010253489, + 0.004511202219873667, + 0.002635051729157567, + 0.0044192299246788025, + 0.0049496181309223175, + 0.0021191707346588373, + 0.0015764090931043029, + 0.0017456954810768366, + 0.0022776515688747168, + 0.0036291400901973248, + 0.0021170845720916986, + 0.002840356668457389, + 0.0038315749261528254, + 0.001908225822262466 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 1 + ], + "selected_reward": -0.31341642141342163 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "Shot from a slightly elevated angle, the extreme wide shot begins with a slow, majestic sideways trajectory across the deep, silent abyss. Ethereal pillars of sunlight penetrate the distant surface, creating a shifting divine light that illuminates countless particles of marine snow drifting in the current. The water itself is a breathtaking gradient, transitioning from a brilliant cyan near the top to a deep, mysterious indigo in the unseen depths below. At the center of this vast azure void, a lone sea turtle swims serenely, its form diminutive against the vast backdrop. It propels itself forward with the powerful yet graceful, slow, rhythmic sweeps of its enormous front flippers, each movement deliberate and effortless. The light captures the intricate mosaic of its brown and olive-green carapace and the leathery, ancient texture of its skin, creating a solitary, majestic figure on an endless journey through the tranquil grandeur of the underwater world, evoking an atmosphere of tranquil solitude.", + "seed": 20676, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000095/group_7/index_6920_seed_20676_dp_14_rank_28_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 7, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.12873884122818707, + "vlm_logprob": -0.12873884122818707, + "vlm_yes_no_margin": 3.425, + "vlm_token_credit": -0.12873884122818707, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set underwater with a gradient transitioning from cyan at the top to deep indigo below?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a lone sea turtle visible swimming at the center of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the turtle move using slow, rhythmic sweeps of its enormous front flippers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the turtle's carapace visible with a brown and olive-green mosaic pattern?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot an extreme wide shot with a slow sideways trajectory?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.010856373235583305 + ], + "opposite_logprob_tokens": [ + -4.7608561515808105 + ], + "logprob_reward": -0.010856373235583305, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.010856373235583305, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08880261331796646, + 0.06159340962767601, + 0.029502714052796364, + 0.02248464711010456, + 0.29743531346321106, + 0.04300841689109802, + 0.11687915027141571, + 0.08030923455953598, + 0.19813840091228485, + 0.061846088618040085 + ], + "credit_map": [ + 0.009472209960222244, + 0.011433340609073639, + 0.013472543098032475, + 0.008028334006667137, + 0.012395287863910198, + 0.01660418137907982, + 0.008177326060831547, + 0.009219391271471977, + 0.005057981237769127, + 0.009457520209252834, + 0.007971584796905518, + 0.006369565147906542, + 0.005483645014464855, + 0.014231659471988678, + 0.00550498953089118, + 0.007516464684158564, + 0.0019750494975596666, + 0.003058489179238677, + 0.0029211819637566805, + 0.0034439510200172663, + 0.0035300173331052065, + 0.0055131153203547, + 0.003165437374264002, + 0.005895471200346947, + 0.002162245335057378, + 0.002680964767932892, + 0.003039943752810359, + 0.0020689712837338448, + 0.002172528998926282, + 0.005334362853318453, + 0.0023505885619670153, + 0.0026750401593744755, + 0.02292291633784771, + 0.01559991855174303, + 0.11484667658805847, + 0.012548085302114487, + 0.016049817204475403, + 0.01855626329779625, + 0.008633626624941826, + 0.08827799558639526, + 0.0034996848553419113, + 0.008568002842366695, + 0.008956024423241615, + 0.002875338075682521, + 0.003843653481453657, + 0.009454846382141113, + 0.002098707016557455, + 0.0037121574860066175, + 0.00663714436814189, + 0.012357876636087894, + 0.0385126955807209, + 0.005434283055365086, + 0.005179752130061388, + 0.036599114537239075, + 0.004220219794660807, + 0.007938062772154808, + 0.007471274584531784, + 0.009385659359395504, + 0.009752079844474792, + 0.0061347936280071735, + 0.0066862632520496845, + 0.019199104979634285, + 0.004396385513246059, + 0.017283674329519272, + 0.01361666526645422, + 0.010757092386484146, + 0.08045458793640137, + 0.00959143415093422, + 0.00736918393522501, + 0.03281720355153084, + 0.005062556825578213, + 0.038469668477773666, + 0.004449468106031418, + 0.005440415348857641, + 0.011592496186494827, + 0.007038806565105915, + 0.005729727912694216, + 0.01531948521733284, + 0.006491788197308779, + 0.005783898290246725 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 8, + 6 + ], + "selected_reward": -0.010856373235583305 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set underwater with a gradient transitioning from cyan at the top to deep indigo below?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a lone sea turtle visible swimming at the center of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the turtle move using slow, rhythmic sweeps of its enormous front flippers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the turtle's carapace visible with a brown and olive-green mosaic pattern?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot an extreme wide shot with a slow sideways trajectory?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.009791205637156963 + ], + "opposite_logprob_tokens": [ + -4.634791374206543 + ], + "logprob_reward": -0.009791205637156963, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.009791205637156963, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12689583003520966, + 0.07603027671575546, + 0.04119356349110603, + 0.045746367424726486, + 0.10022326558828354, + 0.026098228991031647, + 0.1323319524526596, + 0.08246249705553055, + 0.35125985741615295, + 0.01775825396180153 + ], + "credit_map": [ + 0.011411734856665134, + 0.01333556417375803, + 0.014002700336277485, + 0.017416253685951233, + 0.019695356488227844, + 0.01643501967191696, + 0.024274416267871857, + 0.010324770584702492, + 0.005014501046389341, + 0.00787177961319685, + 0.011033532209694386, + 0.007974928244948387, + 0.006270312704145908, + 0.02095775119960308, + 0.009721535257995129, + 0.007185933645814657, + 0.0023194821551442146, + 0.004145705606788397, + 0.005379411857575178, + 0.003844501217827201, + 0.0029587210156023502, + 0.013403832912445068, + 0.003824587445706129, + 0.005317315459251404, + 0.0028615305200219154, + 0.0035657575353980064, + 0.006758155766874552, + 0.006783903576433659, + 0.005045299883931875, + 0.006648648530244827, + 0.008092785254120827, + 0.005990283098071814, + 0.005051645450294018, + 0.004803556017577648, + 0.02987564541399479, + 0.004238353110849857, + 0.005427513271570206, + 0.014135591685771942, + 0.0029969916213303804, + 0.033693961799144745, + 0.0017139154952019453, + 0.0023768513929098845, + 0.002424776554107666, + 0.002945752115920186, + 0.0032651573419570923, + 0.005070572253316641, + 0.0039484999142587185, + 0.004352701362222433, + 0.008989282883703709, + 0.0077775027602910995, + 0.05545467510819435, + 0.004914402961730957, + 0.004555556923151016, + 0.014503808692097664, + 0.003253831062465906, + 0.03288288041949272, + 0.0067664626985788345, + 0.006045067682862282, + 0.02939693257212639, + 0.004593990743160248, + 0.004908414091914892, + 0.010988839901983738, + 0.0036368148867040873, + 0.016125962138175964, + 0.022627603262662888, + 0.013652299530804157, + 0.14729954302310944, + 0.013142894953489304, + 0.009655619971454144, + 0.08236847817897797, + 0.0073797269724309444, + 0.055133670568466187, + 0.001418147119693458, + 0.001995197031646967, + 0.0025875610299408436, + 0.003280916251242161, + 0.0022142508532851934, + 0.0021002788562327623, + 0.00282649346627295, + 0.0013354076072573662 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 6, + 0 + ], + "selected_reward": -0.009791205637156963 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set underwater with a gradient transitioning from cyan at the top to deep indigo below?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a lone sea turtle visible swimming at the center of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the turtle move using slow, rhythmic sweeps of its enormous front flippers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the turtle's carapace visible with a brown and olive-green mosaic pattern?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot an extreme wide shot with a slow sideways trajectory?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02635025978088379 + ], + "opposite_logprob_tokens": [ + -3.651350259780884 + ], + "logprob_reward": -0.02635025978088379, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.02635025978088379, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10099917650222778, + 0.06547930836677551, + 0.047616735100746155, + 0.04520454257726669, + 0.07739054411649704, + 0.039676640182733536, + 0.19815313816070557, + 0.13527408242225647, + 0.25402170419692993, + 0.03618411347270012 + ], + "credit_map": [ + 0.010117142461240292, + 0.009987447410821915, + 0.010484026744961739, + 0.01379203237593174, + 0.015307871624827385, + 0.015252210199832916, + 0.017258619889616966, + 0.008799831382930279, + 0.004975719377398491, + 0.006446281913667917, + 0.009423482231795788, + 0.006484686397016048, + 0.005765206646174192, + 0.017846273258328438, + 0.008235612884163857, + 0.006302048452198505, + 0.003448573872447014, + 0.005340029485523701, + 0.006936306599527597, + 0.004754427820444107, + 0.003496810793876648, + 0.013555762358009815, + 0.004733147565275431, + 0.005351671949028969, + 0.0033658663742244244, + 0.0036766231060028076, + 0.00731290876865387, + 0.0044125900603830814, + 0.0045897625386714935, + 0.011416698805987835, + 0.005677169654518366, + 0.004752923734486103, + 0.009299980476498604, + 0.006016852799803019, + 0.024646420031785965, + 0.006531619932502508, + 0.0044829025864601135, + 0.014113607816398144, + 0.004241907969117165, + 0.008057255297899246, + 0.002891247160732746, + 0.004236955661326647, + 0.0058701676316559315, + 0.005141520872712135, + 0.0054843248799443245, + 0.0047614392824471, + 0.007385496515780687, + 0.003905486548319459, + 0.015863381326198578, + 0.01220270898193121, + 0.08236073702573776, + 0.007679170463234186, + 0.007793546188622713, + 0.016939839348196983, + 0.005072361323982477, + 0.05024139583110809, + 0.010241675190627575, + 0.013553889468312263, + 0.040112465620040894, + 0.012261711061000824, + 0.008573259226977825, + 0.011353664100170135, + 0.01297750324010849, + 0.026199907064437866, + 0.0168707724660635, + 0.01112850196659565, + 0.10690265148878098, + 0.009712941013276577, + 0.0074687534943223, + 0.04576389491558075, + 0.00604313425719738, + 0.050131041556596756, + 0.002281000604853034, + 0.0030301809310913086, + 0.005712852347642183, + 0.005933901760727167, + 0.0038205264136195183, + 0.007355149835348129, + 0.005548167508095503, + 0.002502336399629712 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 6, + 7 + ], + "selected_reward": -0.02635025978088379 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set underwater with a gradient transitioning from cyan at the top to deep indigo below?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a lone sea turtle visible swimming at the center of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the turtle move using slow, rhythmic sweeps of its enormous front flippers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the turtle's carapace visible with a brown and olive-green mosaic pattern?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot an extreme wide shot with a slow sideways trajectory?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.020579388365149498 + ], + "opposite_logprob_tokens": [ + -3.8955793380737305 + ], + "logprob_reward": -0.020579388365149498, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.020579388365149498, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06830569356679916, + 0.03795149177312851, + 0.03526296466588974, + 0.02795145846903324, + 0.13044473528862, + 0.09278243780136108, + 0.16847726702690125, + 0.11626861244440079, + 0.298778235912323, + 0.02377709187567234 + ], + "credit_map": [ + 0.006763156037777662, + 0.007671873085200787, + 0.010341683402657509, + 0.008377054706215858, + 0.009363759309053421, + 0.00942285731434822, + 0.01041414774954319, + 0.0059511661529541016, + 0.0030898733530193567, + 0.004168728366494179, + 0.00520337326452136, + 0.00434014480561018, + 0.00335747841745615, + 0.009096375666558743, + 0.004935069940984249, + 0.003760450752452016, + 0.001980322180315852, + 0.0030732518061995506, + 0.0059065609239041805, + 0.0027441473212093115, + 0.003494437551125884, + 0.011380933225154877, + 0.003230633679777384, + 0.003452679608017206, + 0.0017691190587356687, + 0.003134007565677166, + 0.003711757715791464, + 0.0036266613751649857, + 0.003157737897709012, + 0.003363378345966339, + 0.004919394385069609, + 0.004269405733793974, + 0.007030623499304056, + 0.006691797636449337, + 0.04477165639400482, + 0.007180738262832165, + 0.007961638271808624, + 0.04242362082004547, + 0.005215913988649845, + 0.009168758988380432, + 0.00695256981998682, + 0.009843229316174984, + 0.009640893898904324, + 0.013288140296936035, + 0.012061035260558128, + 0.011960748583078384, + 0.01781451515853405, + 0.011221317574381828, + 0.008028273470699787, + 0.006824240554124117, + 0.06187878176569939, + 0.008240089751780033, + 0.008542679250240326, + 0.05252451449632645, + 0.007360832300037146, + 0.015077865682542324, + 0.008333212696015835, + 0.008555256761610508, + 0.03847520425915718, + 0.008557860739529133, + 0.00787283107638359, + 0.008567730896174908, + 0.008824251592159271, + 0.027082286775112152, + 0.01973743364214897, + 0.012661062180995941, + 0.12290867418050766, + 0.011927718296647072, + 0.01045988779515028, + 0.05911713466048241, + 0.007675439119338989, + 0.05429092422127724, + 0.0017697137081995606, + 0.002192576415836811, + 0.0037669623270630836, + 0.004129709675908089, + 0.0031010080128908157, + 0.003214355790987611, + 0.0035612236242741346, + 0.0020415454637259245 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 6, + 4 + ], + "selected_reward": -0.020579388365149498 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set underwater with a gradient transitioning from cyan at the top to deep indigo below?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a lone sea turtle visible swimming at the center of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the turtle move using slow, rhythmic sweeps of its enormous front flippers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the turtle's carapace visible with a brown and olive-green mosaic pattern?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot an extreme wide shot with a slow sideways trajectory?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.5761169791221619 + ], + "opposite_logprob_tokens": [ + -0.8261169791221619 + ], + "logprob_reward": -0.5761169791221619, + "yes_no_margin_reward": 0.25, + "token_credit_reward": -0.5761169791221619, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.05404824763536453, + 0.03164168447256088, + 0.05516412481665611, + 0.06928286701440811, + 0.057626333087682724, + 0.049798425287008286, + 0.26500123739242554, + 0.16151063144207, + 0.23191727697849274, + 0.02400917001068592 + ], + "credit_map": [ + 0.006774033419787884, + 0.00614764541387558, + 0.006563116330653429, + 0.005976397544145584, + 0.008587418124079704, + 0.010000761598348618, + 0.005499646533280611, + 0.004499224480241537, + 0.002801731927320361, + 0.006043384317308664, + 0.003165072062984109, + 0.004045272246003151, + 0.0031432625837624073, + 0.005829555913805962, + 0.003134818747639656, + 0.003478587605059147, + 0.003712876932695508, + 0.005732798483222723, + 0.008857589215040207, + 0.006175702903419733, + 0.006239504087716341, + 0.011571952141821384, + 0.007851554080843925, + 0.005022147670388222, + 0.006002945359796286, + 0.007989873178303242, + 0.01117122732102871, + 0.005592609755694866, + 0.006106350105255842, + 0.018640240654349327, + 0.007383052725344896, + 0.00639656325802207, + 0.006675036624073982, + 0.007258361671119928, + 0.007348833605647087, + 0.009115702472627163, + 0.005892958492040634, + 0.0072688208892941475, + 0.0071897972375154495, + 0.006876825820654631, + 0.0035731045063585043, + 0.006129743531346321, + 0.005929984152317047, + 0.003384639509022236, + 0.0046434602700173855, + 0.016073474660515785, + 0.0031261425465345383, + 0.00693787494674325, + 0.023427654057741165, + 0.01594376005232334, + 0.11012224853038788, + 0.010219613090157509, + 0.010548965074121952, + 0.027021318674087524, + 0.006590650882571936, + 0.06112701818346977, + 0.014256469905376434, + 0.011014078743755817, + 0.06317228823900223, + 0.006768531631678343, + 0.007121644448488951, + 0.016710948199033737, + 0.006135409232228994, + 0.03633124753832817, + 0.014117072336375713, + 0.009709380567073822, + 0.09572102874517441, + 0.010212766006588936, + 0.006995575502514839, + 0.058704305440187454, + 0.004792911931872368, + 0.03166424483060837, + 0.0015716601628810167, + 0.002657797420397401, + 0.0035645500756800175, + 0.003993205726146698, + 0.003034972120076418, + 0.0038207711186259985, + 0.003280128352344036, + 0.0020860861986875534 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 8, + 7 + ], + "selected_reward": -0.5761169791221619 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "Shot from a slightly elevated angle, the extreme wide shot begins with a slow, majestic sideways trajectory across the deep, silent abyss. Ethereal pillars of sunlight penetrate the distant surface, creating a shifting divine light that illuminates countless particles of marine snow drifting in the current. The water itself is a breathtaking gradient, transitioning from a brilliant cyan near the top to a deep, mysterious indigo in the unseen depths below. At the center of this vast azure void, a lone sea turtle swims serenely, its form diminutive against the vast backdrop. It propels itself forward with the powerful yet graceful, slow, rhythmic sweeps of its enormous front flippers, each movement deliberate and effortless. The light captures the intricate mosaic of its brown and olive-green carapace and the leathery, ancient texture of its skin, creating a solitary, majestic figure on an endless journey through the tranquil grandeur of the underwater world, evoking an atmosphere of tranquil solitude.", + "seed": 20677, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000095/group_7/index_6920_seed_20677_dp_15_rank_30_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 7, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.09948454955592752, + "vlm_logprob": -0.09948454955592752, + "vlm_yes_no_margin": 3.6, + "vlm_token_credit": -0.09948454955592752, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set underwater with a gradient transitioning from cyan at the top to deep indigo below?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a lone sea turtle visible swimming at the center of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the turtle move using slow, rhythmic sweeps of its enormous front flippers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the turtle's carapace visible with a brown and olive-green mosaic pattern?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot an extreme wide shot with a slow sideways trajectory?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.011993792839348316 + ], + "opposite_logprob_tokens": [ + -4.636993885040283 + ], + "logprob_reward": -0.011993792839348316, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.011993792839348316, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11278087645769119, + 0.1390332728624344, + 0.058783672749996185, + 0.03419559448957443, + 0.10348641127347946, + 0.05231824517250061, + 0.09835682064294815, + 0.15851670503616333, + 0.04119974747300148, + 0.20132867991924286 + ], + "credit_map": [ + 0.012149368412792683, + 0.01607305370271206, + 0.011823738925158978, + 0.010968825779855251, + 0.010125232860445976, + 0.02768847718834877, + 0.01339606661349535, + 0.010556112974882126, + 0.01828704960644245, + 0.014574265107512474, + 0.014636249281466007, + 0.011978624388575554, + 0.01002518367022276, + 0.02852604165673256, + 0.013088804669678211, + 0.027917055413126945, + 0.006015649996697903, + 0.006769655272364616, + 0.007627519778907299, + 0.0050803376361727715, + 0.0050219325348734856, + 0.01802998036146164, + 0.005229833070188761, + 0.005008766893297434, + 0.005208971910178661, + 0.003525710664689541, + 0.004473006818443537, + 0.003368895500898361, + 0.0032031226437538862, + 0.005302187986671925, + 0.004192846827208996, + 0.004920853301882744, + 0.019610410556197166, + 0.009477159939706326, + 0.010028257966041565, + 0.009349397383630276, + 0.0068960776552557945, + 0.021382523700594902, + 0.00930103287100792, + 0.017441553995013237, + 0.00554968137294054, + 0.004902856424450874, + 0.005048384889960289, + 0.009227098897099495, + 0.006220012437552214, + 0.007462912704795599, + 0.008008237928152084, + 0.005899057257920504, + 0.007463050074875355, + 0.009866531938314438, + 0.012406467460095882, + 0.010675049386918545, + 0.010195398703217506, + 0.029024846851825714, + 0.01242723036557436, + 0.006298244930803776, + 0.013583711348474026, + 0.014252781867980957, + 0.017320789396762848, + 0.019412187859416008, + 0.016263579949736595, + 0.04144589230418205, + 0.022581517696380615, + 0.013656246475875378, + 0.002620014827698469, + 0.0037236616481095552, + 0.005718186963349581, + 0.007790667936205864, + 0.004941842053085566, + 0.005103712901473045, + 0.007064115256071091, + 0.004237547982484102, + 0.014013255015015602, + 0.02308381162583828, + 0.05870606750249863, + 0.014115681871771812, + 0.0122922258451581, + 0.052124518901109695, + 0.011991151608526707, + 0.015001969411969185 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 7, + 1 + ], + "selected_reward": -0.011993792839348316 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set underwater with a gradient transitioning from cyan at the top to deep indigo below?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a lone sea turtle visible swimming at the center of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the turtle move using slow, rhythmic sweeps of its enormous front flippers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the turtle's carapace visible with a brown and olive-green mosaic pattern?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot an extreme wide shot with a slow sideways trajectory?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004645626526325941 + ], + "opposite_logprob_tokens": [ + -5.379645824432373 + ], + "logprob_reward": -0.004645626526325941, + "yes_no_margin_reward": 5.375, + "token_credit_reward": -0.004645626526325941, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.26765066385269165, + 0.12821915745735168, + 0.1007043793797493, + 0.0652085691690445, + 0.06425803154706955, + 0.06078497692942619, + 0.10809160768985748, + 0.055370282381772995, + 0.07879045605659485, + 0.0709218829870224 + ], + "credit_map": [ + 0.024867819622159004, + 0.029708612710237503, + 0.02980063110589981, + 0.03640902042388916, + 0.03474503010511398, + 0.048035722225904465, + 0.0387418158352375, + 0.025342021137475967, + 0.014216849580407143, + 0.016495175659656525, + 0.011533860117197037, + 0.017487473785877228, + 0.016800805926322937, + 0.019989287480711937, + 0.019592292606830597, + 0.012103419750928879, + 0.011618531309068203, + 0.00992731936275959, + 0.010574541985988617, + 0.014519603922963142, + 0.01256297342479229, + 0.02201196365058422, + 0.011918491683900356, + 0.007570959627628326, + 0.006798876449465752, + 0.005259344819933176, + 0.006631017196923494, + 0.011483414098620415, + 0.008520743809640408, + 0.009758848696947098, + 0.010311057791113853, + 0.006445271894335747, + 0.008776070550084114, + 0.004409664776176214, + 0.006810817401856184, + 0.011920235119760036, + 0.008249464444816113, + 0.007318404968827963, + 0.011057627387344837, + 0.0057157473638653755, + 0.005860211327672005, + 0.005474284291267395, + 0.008010477758944035, + 0.009029141627252102, + 0.007710753940045834, + 0.009123124182224274, + 0.010729923844337463, + 0.004847063682973385, + 0.0062254648655653, + 0.008146095089614391, + 0.012080960907042027, + 0.014339889399707317, + 0.01988365687429905, + 0.01957816071808338, + 0.020413566380739212, + 0.007423821836709976, + 0.0037810143548995256, + 0.005219471175223589, + 0.006066803354769945, + 0.009467079304158688, + 0.007129834033548832, + 0.00782014336436987, + 0.008517381735146046, + 0.00736855436116457, + 0.004877710249274969, + 0.006206672638654709, + 0.016698315739631653, + 0.013167278841137886, + 0.009979397058486938, + 0.008651265874505043, + 0.010754511691629887, + 0.0084553062915802, + 0.007591377478092909, + 0.007107057608664036, + 0.010711876675486565, + 0.00840399507433176, + 0.008440354838967323, + 0.010808405466377735, + 0.008144090883433819, + 0.009714731015264988 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 6 + ], + "selected_reward": -0.004645626526325941 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set underwater with a gradient transitioning from cyan at the top to deep indigo below?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a lone sea turtle visible swimming at the center of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the turtle move using slow, rhythmic sweeps of its enormous front flippers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the turtle's carapace visible with a brown and olive-green mosaic pattern?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot an extreme wide shot with a slow sideways trajectory?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.033707667142152786 + ], + "opposite_logprob_tokens": [ + -3.408707618713379 + ], + "logprob_reward": -0.033707667142152786, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.033707667142152786, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1602528691291809, + 0.10502547025680542, + 0.07283283770084381, + 0.07031194120645523, + 0.04851701855659485, + 0.05182595178484917, + 0.04416162520647049, + 0.06773165613412857, + 0.0785694271326065, + 0.30077120661735535 + ], + "credit_map": [ + 0.019235434010624886, + 0.018203940242528915, + 0.016139693558216095, + 0.019040994346141815, + 0.022366423159837723, + 0.020712725818157196, + 0.024361638352274895, + 0.02019200101494789, + 0.010586196556687355, + 0.012057353742420673, + 0.01450774073600769, + 0.014363361522555351, + 0.013808539137244225, + 0.012590828351676464, + 0.018403738737106323, + 0.00870770588517189, + 0.007745534647256136, + 0.006436667405068874, + 0.0074881911277771, + 0.011961043812334538, + 0.009132971987128258, + 0.012238692492246628, + 0.01179223507642746, + 0.006037489976733923, + 0.007037639152258635, + 0.004419992677867413, + 0.006883732508867979, + 0.015376724302768707, + 0.007984301075339317, + 0.008485624566674232, + 0.013189094141125679, + 0.0069348253309726715, + 0.004485952202230692, + 0.003255437593907118, + 0.005936098750680685, + 0.010869014076888561, + 0.00578303262591362, + 0.004554655868560076, + 0.009095990099012852, + 0.004536834079772234, + 0.004769936669617891, + 0.004078834783285856, + 0.006233690772205591, + 0.007762707304209471, + 0.005770170595496893, + 0.009794501587748528, + 0.009229766204953194, + 0.00418633921071887, + 0.0026652785018086433, + 0.003360035130754113, + 0.006116077303886414, + 0.005599984899163246, + 0.006912798620760441, + 0.008284119889140129, + 0.008143117651343346, + 0.0030802052933722734, + 0.006278145592659712, + 0.006408588029444218, + 0.00590208126232028, + 0.009927412495017052, + 0.010078223422169685, + 0.011600254103541374, + 0.011004514060914516, + 0.006532432045787573, + 0.0044138082303106785, + 0.005813406780362129, + 0.00936772208660841, + 0.0149543983861804, + 0.011184684932231903, + 0.010454311966896057, + 0.013702591881155968, + 0.008678500540554523, + 0.010923540219664574, + 0.030343836173415184, + 0.08637239038944244, + 0.013412272557616234, + 0.01695271022617817, + 0.12222626060247421, + 0.007865529507398605, + 0.012674639001488686 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 0, + 1 + ], + "selected_reward": -0.033707667142152786 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set underwater with a gradient transitioning from cyan at the top to deep indigo below?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a lone sea turtle visible swimming at the center of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the turtle move using slow, rhythmic sweeps of its enormous front flippers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the turtle's carapace visible with a brown and olive-green mosaic pattern?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot an extreme wide shot with a slow sideways trajectory?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01818695291876793 + ], + "opposite_logprob_tokens": [ + -4.018187046051025 + ], + "logprob_reward": -0.01818695291876793, + "yes_no_margin_reward": 4.0, + "token_credit_reward": -0.01818695291876793, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1558825969696045, + 0.128848135471344, + 0.08439601212739944, + 0.041654616594314575, + 0.11352401226758957, + 0.04184287041425705, + 0.10243123024702072, + 0.08179353177547455, + 0.04264724627137184, + 0.20697976648807526 + ], + "credit_map": [ + 0.02068605087697506, + 0.01475891750305891, + 0.017347287386655807, + 0.01906595565378666, + 0.020198388025164604, + 0.02190135419368744, + 0.025551307946443558, + 0.016373317688703537, + 0.013237519189715385, + 0.018098268657922745, + 0.01102603878825903, + 0.013152397237718105, + 0.013064848259091377, + 0.027525901794433594, + 0.021383274346590042, + 0.011359890922904015, + 0.00798027217388153, + 0.008243578486144543, + 0.007343899458646774, + 0.012590325437486172, + 0.00974428839981556, + 0.014072995632886887, + 0.017304817214608192, + 0.0071158334612846375, + 0.004197923466563225, + 0.003734649857506156, + 0.0035210989881306887, + 0.00676315650343895, + 0.0067993467673659325, + 0.005119649693369865, + 0.008162285201251507, + 0.0033565014600753784, + 0.012042359448969364, + 0.007387995254248381, + 0.012158622033894062, + 0.0200552586466074, + 0.011984181590378284, + 0.016870440915226936, + 0.021470144391059875, + 0.011554998345673084, + 0.0036191341932862997, + 0.003741935361176729, + 0.0038913011085242033, + 0.006844054441899061, + 0.00786589551717043, + 0.005583326332271099, + 0.006856916937977076, + 0.003440301166847348, + 0.007604160346090794, + 0.010521708987653255, + 0.00930586364120245, + 0.01103780884295702, + 0.012252448126673698, + 0.029149221256375313, + 0.011817173101007938, + 0.010742845945060253, + 0.005592728964984417, + 0.008250745013356209, + 0.010324102826416492, + 0.010341613553464413, + 0.01144389808177948, + 0.019795186817646027, + 0.010565051808953285, + 0.005480204708874226, + 0.003190780756995082, + 0.003876214846968651, + 0.004460530821233988, + 0.006713917013257742, + 0.005777627695351839, + 0.007488141767680645, + 0.006688209716230631, + 0.0044518206268548965, + 0.008026285097002983, + 0.012448892928659916, + 0.06543829292058945, + 0.011438761837780476, + 0.012824428267776966, + 0.07841910421848297, + 0.00760151632130146, + 0.010782478377223015 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 0, + 1 + ], + "selected_reward": -0.01818695291876793 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set underwater with a gradient transitioning from cyan at the top to deep indigo below?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a lone sea turtle visible swimming at the center of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the turtle move using slow, rhythmic sweeps of its enormous front flippers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the turtle's carapace visible with a brown and olive-green mosaic pattern?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot an extreme wide shot with a slow sideways trajectory?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.4288887083530426 + ], + "opposite_logprob_tokens": [ + -1.0538886785507202 + ], + "logprob_reward": -0.4288887083530426, + "yes_no_margin_reward": 0.625, + "token_credit_reward": -0.4288887083530426, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11755140870809555, + 0.043571628630161285, + 0.028748469427227974, + 0.06599127501249313, + 0.08608125895261765, + 0.034213919192552567, + 0.09006182849407196, + 0.08134245872497559, + 0.10615599900484085, + 0.3462817668914795 + ], + "credit_map": [ + 0.008405284024775028, + 0.014773205853998661, + 0.010712075978517532, + 0.014987642876803875, + 0.014614449813961983, + 0.025947943329811096, + 0.01663728430867195, + 0.011473516002297401, + 0.00809543114155531, + 0.003837279276922345, + 0.004548798315227032, + 0.004643507767468691, + 0.003661779221147299, + 0.005203689448535442, + 0.005242256913334131, + 0.008338886313140392, + 0.003911061678081751, + 0.0024853991344571114, + 0.0032858646009117365, + 0.004400099162012339, + 0.0032361880876123905, + 0.003566491650417447, + 0.0049720932729542255, + 0.0028912732377648354, + 0.007079822942614555, + 0.006296715699136257, + 0.0072617316618561745, + 0.00801748875528574, + 0.00585718872025609, + 0.014470451511442661, + 0.011913455091416836, + 0.00509441876783967, + 0.01021330151706934, + 0.008194562047719955, + 0.008492078632116318, + 0.008760951459407806, + 0.007882430218160152, + 0.014371858909726143, + 0.014708167873322964, + 0.01345791108906269, + 0.0032565761357545853, + 0.002429915824905038, + 0.0028351249638944864, + 0.005271692294627428, + 0.004176673013716936, + 0.006371935363858938, + 0.006049793213605881, + 0.0038222100120037794, + 0.005755390506237745, + 0.009590224362909794, + 0.008116278797388077, + 0.010816050693392754, + 0.012172499671578407, + 0.022335181012749672, + 0.012193744070827961, + 0.009082453325390816, + 0.006734983995556831, + 0.005774450022727251, + 0.008569007739424706, + 0.013636060059070587, + 0.012618512846529484, + 0.010715038515627384, + 0.015820616856217384, + 0.007473791018128395, + 0.005236342083662748, + 0.008646364323794842, + 0.024912448599934578, + 0.009534320794045925, + 0.007523722946643829, + 0.03508061170578003, + 0.008561233058571815, + 0.0066609568893909454, + 0.016773894429206848, + 0.03798064962029457, + 0.08941303193569183, + 0.013400373980402946, + 0.01982351765036583, + 0.14494577050209045, + 0.0077045210637152195, + 0.016240019351243973 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 0, + 8 + ], + "selected_reward": -0.4288887083530426 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "Shot from a slightly elevated angle, the extreme wide shot begins with a slow, majestic sideways trajectory across the deep, silent abyss. Ethereal pillars of sunlight penetrate the distant surface, creating a shifting divine light that illuminates countless particles of marine snow drifting in the current. The water itself is a breathtaking gradient, transitioning from a brilliant cyan near the top to a deep, mysterious indigo in the unseen depths below. At the center of this vast azure void, a lone sea turtle swims serenely, its form diminutive against the vast backdrop. It propels itself forward with the powerful yet graceful, slow, rhythmic sweeps of its enormous front flippers, each movement deliberate and effortless. The light captures the intricate mosaic of its brown and olive-green carapace and the leathery, ancient texture of its skin, creating a solitary, majestic figure on an endless journey through the tranquil grandeur of the underwater world, evoking an atmosphere of tranquil solitude.", + "seed": 20678, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000095/group_7/index_6920_seed_20678_dp_15_rank_30_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 7, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.04721704591065645, + "vlm_logprob": -0.04721704591065645, + "vlm_yes_no_margin": 3.875, + "vlm_token_credit": -0.04721704591065645, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set underwater with a gradient transitioning from cyan at the top to deep indigo below?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a lone sea turtle visible swimming at the center of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the turtle move using slow, rhythmic sweeps of its enormous front flippers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the turtle's carapace visible with a brown and olive-green mosaic pattern?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot an extreme wide shot with a slow sideways trajectory?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.009238234721124172 + ], + "opposite_logprob_tokens": [ + -5.009238243103027 + ], + "logprob_reward": -0.009238234721124172, + "yes_no_margin_reward": 5.0, + "token_credit_reward": -0.009238234721124172, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.21031059324741364, + 0.14342100918293, + 0.12706154584884644, + 0.1944389045238495, + 0.06409047544002533, + 0.04458020254969597, + 0.03706006333231926, + 0.06312849372625351, + 0.03390892222523689, + 0.08199985325336456 + ], + "credit_map": [ + 0.02389518730342388, + 0.01614515110850334, + 0.023729106411337852, + 0.016180800274014473, + 0.026549512520432472, + 0.06798695772886276, + 0.014736548997461796, + 0.021087301895022392, + 0.02084164321422577, + 0.018636537715792656, + 0.01270844042301178, + 0.012325256131589413, + 0.01369519717991352, + 0.03161459416151047, + 0.018403565511107445, + 0.015195759013295174, + 0.013343004509806633, + 0.013513979502022266, + 0.025820359587669373, + 0.013552761636674404, + 0.008630407974123955, + 0.03141394257545471, + 0.009087768383324146, + 0.011699309572577477, + 0.030419237911701202, + 0.018210306763648987, + 0.04515552148222923, + 0.022702839225530624, + 0.013665886595845222, + 0.018144860863685608, + 0.02052609622478485, + 0.02561412751674652, + 0.008934954181313515, + 0.007583752274513245, + 0.008057906292378902, + 0.008056582883000374, + 0.009008921682834625, + 0.0070509170182049274, + 0.009134259074926376, + 0.0062631708569824696, + 0.005533276591449976, + 0.0038553718477487564, + 0.005105811636894941, + 0.005647962912917137, + 0.004565374460071325, + 0.01029854267835617, + 0.005114932078868151, + 0.0044589270837605, + 0.004572248551994562, + 0.007220612373203039, + 0.00727989012375474, + 0.0034853860270231962, + 0.0036969759967178106, + 0.005416094791144133, + 0.002570223528891802, + 0.0028186289127916098, + 0.008924990892410278, + 0.01646541617810726, + 0.004508648999035358, + 0.005237258039414883, + 0.005332658998668194, + 0.011838609352707863, + 0.0066341012716293335, + 0.004186798818409443, + 0.004019738174974918, + 0.005009028594940901, + 0.0028420807793736458, + 0.0043325661681592464, + 0.004126181360334158, + 0.006283588241785765, + 0.0036125662736594677, + 0.003683169838041067, + 0.009927879087626934, + 0.010582786053419113, + 0.012496968731284142, + 0.009255483746528625, + 0.009552989155054092, + 0.01475134864449501, + 0.00794680044054985, + 0.007485587615519762 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 1 + ], + "selected_reward": -0.009238234721124172 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set underwater with a gradient transitioning from cyan at the top to deep indigo below?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a lone sea turtle visible swimming at the center of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the turtle move using slow, rhythmic sweeps of its enormous front flippers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the turtle's carapace visible with a brown and olive-green mosaic pattern?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot an extreme wide shot with a slow sideways trajectory?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.007633563131093979 + ], + "opposite_logprob_tokens": [ + -4.882633686065674 + ], + "logprob_reward": -0.007633563131093979, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.007633563131093979, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.21920126676559448, + 0.20497392117977142, + 0.054364677518606186, + 0.14102350175380707, + 0.07082390785217285, + 0.13406552374362946, + 0.019406873732805252, + 0.03496401011943817, + 0.03316335007548332, + 0.08801291882991791 + ], + "credit_map": [ + 0.0152989337220788, + 0.015820128843188286, + 0.01952185668051243, + 0.034476473927497864, + 0.02961229905486107, + 0.04601864889264107, + 0.0346045158803463, + 0.02384842187166214, + 0.015459702350199223, + 0.017527218908071518, + 0.024447305127978325, + 0.028987767174839973, + 0.0206254031509161, + 0.03390220180153847, + 0.03588385134935379, + 0.0281404796987772, + 0.004519014619290829, + 0.005383480340242386, + 0.008260807953774929, + 0.005723393522202969, + 0.004575355444103479, + 0.013013217598199844, + 0.008410133421421051, + 0.004479279275983572, + 0.011387352831661701, + 0.011398376896977425, + 0.01571565680205822, + 0.024856802076101303, + 0.0199967622756958, + 0.022758306935429573, + 0.023377230390906334, + 0.011533025652170181, + 0.0050060986541211605, + 0.008252945728600025, + 0.007671087048947811, + 0.010304353199899197, + 0.011174547486007214, + 0.013203299604356289, + 0.01056884229183197, + 0.00464274175465107, + 0.010577350854873657, + 0.020331451669335365, + 0.019267132505774498, + 0.007492114789783955, + 0.010364091955125332, + 0.04852016642689705, + 0.009760389104485512, + 0.0077528320252895355, + 0.0016838718438521028, + 0.004410541150718927, + 0.0013697069371119142, + 0.0019021215848624706, + 0.002399592427536845, + 0.0045914542861282825, + 0.00199261330999434, + 0.001056972425431013, + 0.004039669409394264, + 0.008587706834077835, + 0.003301777644082904, + 0.003889436135068536, + 0.004658276215195656, + 0.005005673971027136, + 0.00335828959941864, + 0.0021231805440038443, + 0.0026316943112760782, + 0.002883905079215765, + 0.003582826815545559, + 0.007477239239960909, + 0.004191562067717314, + 0.004889695439487696, + 0.004127000458538532, + 0.0033794266637414694, + 0.008540144190192223, + 0.008110105991363525, + 0.010255886241793633, + 0.011430472135543823, + 0.014992759563028812, + 0.013417210429906845, + 0.014080124907195568, + 0.0071862200275063515 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 3 + ], + "selected_reward": -0.007633563131093979 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set underwater with a gradient transitioning from cyan at the top to deep indigo below?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a lone sea turtle visible swimming at the center of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the turtle move using slow, rhythmic sweeps of its enormous front flippers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the turtle's carapace visible with a brown and olive-green mosaic pattern?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot an extreme wide shot with a slow sideways trajectory?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02329133450984955 + ], + "opposite_logprob_tokens": [ + -3.7732913494110107 + ], + "logprob_reward": -0.02329133450984955, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.02329133450984955, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13854114711284637, + 0.10455215722322464, + 0.10214366763830185, + 0.13878493010997772, + 0.07854445278644562, + 0.12266676127910614, + 0.08550455421209335, + 0.0595882311463356, + 0.11452776938676834, + 0.05514628812670708 + ], + "credit_map": [ + 0.012467040680348873, + 0.009173907339572906, + 0.012497241608798504, + 0.016223175451159477, + 0.020977821201086044, + 0.038598351180553436, + 0.01789087802171707, + 0.010712740942835808, + 0.008432763628661633, + 0.010536368004977703, + 0.010184822604060173, + 0.011394967325031757, + 0.011723925359547138, + 0.023358868435025215, + 0.016723450273275375, + 0.012196985073387623, + 0.006737926509231329, + 0.010789194144308567, + 0.015567487105727196, + 0.012980058789253235, + 0.011482279747724533, + 0.022491246461868286, + 0.013230615295469761, + 0.008864859119057655, + 0.012804741971194744, + 0.013482161797583103, + 0.024948423728346825, + 0.020232245326042175, + 0.014433955773711205, + 0.0204282496124506, + 0.018830981105566025, + 0.013624172657728195, + 0.006436466239392757, + 0.010218135081231594, + 0.0135783851146698, + 0.008346443064510822, + 0.007979081943631172, + 0.019571904093027115, + 0.00709153339266777, + 0.0053225005976855755, + 0.008901908062398434, + 0.018270568922162056, + 0.017119817435741425, + 0.009817909449338913, + 0.010331213474273682, + 0.04102917015552521, + 0.010083192028105259, + 0.007112980354577303, + 0.00874011404812336, + 0.026682300493121147, + 0.0028661624528467655, + 0.005306555423885584, + 0.006496243644505739, + 0.025157637894153595, + 0.005281947087496519, + 0.004973600152879953, + 0.007321561221033335, + 0.01491444930434227, + 0.0045228200033307076, + 0.006008808501064777, + 0.00668808538466692, + 0.010010701604187489, + 0.004936417564749718, + 0.005185386165976524, + 0.010752510279417038, + 0.010221942327916622, + 0.01340495515614748, + 0.024342257529497147, + 0.015558069571852684, + 0.015152296982705593, + 0.013311713933944702, + 0.011784026399254799, + 0.0044107865542173386, + 0.00459087360650301, + 0.006324851419776678, + 0.007553485222160816, + 0.010666534304618835, + 0.008337455801665783, + 0.008535380475223064, + 0.004726921208202839 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 0, + 5 + ], + "selected_reward": -0.02329133450984955 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set underwater with a gradient transitioning from cyan at the top to deep indigo below?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a lone sea turtle visible swimming at the center of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the turtle move using slow, rhythmic sweeps of its enormous front flippers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the turtle's carapace visible with a brown and olive-green mosaic pattern?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot an extreme wide shot with a slow sideways trajectory?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.016067897900938988 + ], + "opposite_logprob_tokens": [ + -4.141067981719971 + ], + "logprob_reward": -0.016067897900938988, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.016067897900938988, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.24270512163639069, + 0.1096661165356636, + 0.13334175944328308, + 0.16343826055526733, + 0.14488469064235687, + 0.03999456390738487, + 0.039545949548482895, + 0.032708872109651566, + 0.04999103397130966, + 0.04372357949614525 + ], + "credit_map": [ + 0.014807509258389473, + 0.021652644500136375, + 0.01939994841814041, + 0.014129303395748138, + 0.04362523555755615, + 0.09237457811832428, + 0.014624246396124363, + 0.022091666236519814, + 0.013009698130190372, + 0.008545396849513054, + 0.01078573614358902, + 0.017466910183429718, + 0.01247202605009079, + 0.019588235765695572, + 0.017995337024331093, + 0.009802781045436859, + 0.011657070368528366, + 0.011095416732132435, + 0.021297458559274673, + 0.022931747138500214, + 0.013944759033620358, + 0.017307071015238762, + 0.022468052804470062, + 0.01264018751680851, + 0.024854430928826332, + 0.015215235762298107, + 0.0338844433426857, + 0.018353642895817757, + 0.011252432130277157, + 0.019013583660125732, + 0.020085111260414124, + 0.020779382437467575, + 0.012215943075716496, + 0.014433389529585838, + 0.017544785514473915, + 0.020989300683140755, + 0.026154469698667526, + 0.020064368844032288, + 0.02048186957836151, + 0.01300057116895914, + 0.003423879388719797, + 0.005216067191213369, + 0.005590851418673992, + 0.0033290402498096228, + 0.004358114209026098, + 0.011929752305150032, + 0.0034264212008565664, + 0.002720435382798314, + 0.004367271903902292, + 0.00647308025509119, + 0.004015433602035046, + 0.006991243455559015, + 0.004301051609218121, + 0.00433350820094347, + 0.005292033776640892, + 0.003772324649617076, + 0.0031169478315860033, + 0.004381761886179447, + 0.003325419733300805, + 0.0034005725756287575, + 0.003675981191918254, + 0.005444556940346956, + 0.0042990813963115215, + 0.005064550321549177, + 0.003887676866725087, + 0.004113127011805773, + 0.006289367564022541, + 0.012334644794464111, + 0.006456127390265465, + 0.006823836825788021, + 0.005322814919054508, + 0.0047634379006922245, + 0.0038350275717675686, + 0.0050883530639112, + 0.005191180855035782, + 0.005320950411260128, + 0.008368318900465965, + 0.005210372153669596, + 0.006102990824729204, + 0.004606388974934816 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 4 + ], + "selected_reward": -0.016067897900938988 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set underwater with a gradient transitioning from cyan at the top to deep indigo below?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a lone sea turtle visible swimming at the center of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the turtle move using slow, rhythmic sweeps of its enormous front flippers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the turtle's carapace visible with a brown and olive-green mosaic pattern?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot an extreme wide shot with a slow sideways trajectory?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.17985419929027557 + ], + "opposite_logprob_tokens": [ + -1.804854154586792 + ], + "logprob_reward": -0.17985419929027557, + "yes_no_margin_reward": 1.625, + "token_credit_reward": -0.17985419929027557, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0832814946770668, + 0.06793061643838882, + 0.10937881469726562, + 0.1627500206232071, + 0.09058588743209839, + 0.20366227626800537, + 0.1386570930480957, + 0.06687156111001968, + 0.05665435642004013, + 0.020227836444973946 + ], + "credit_map": [ + 0.006519276183098555, + 0.007467437535524368, + 0.01182385254651308, + 0.013816834427416325, + 0.008041608147323132, + 0.013726618140935898, + 0.012906143441796303, + 0.008979723788797855, + 0.005963355768471956, + 0.0053298077546060085, + 0.009693514555692673, + 0.00998678058385849, + 0.008493891917169094, + 0.008173475041985512, + 0.012866112403571606, + 0.007423677481710911, + 0.009412763640284538, + 0.011759959161281586, + 0.023356350138783455, + 0.013045523315668106, + 0.00799332931637764, + 0.023168256506323814, + 0.009445269592106342, + 0.011197361163794994, + 0.01626310683786869, + 0.018907641991972923, + 0.035487737506628036, + 0.01461406983435154, + 0.010669535957276821, + 0.038254264742136, + 0.01310123223811388, + 0.0154524315148592, + 0.012651008553802967, + 0.00791835505515337, + 0.01673879846930504, + 0.011467539705336094, + 0.0104877520352602, + 0.009518765844404697, + 0.011220510117709637, + 0.010583166033029556, + 0.018601149320602417, + 0.03395136073231697, + 0.02533990889787674, + 0.011172075755894184, + 0.01615552045404911, + 0.07263080030679703, + 0.012773123569786549, + 0.01303834281861782, + 0.009341984987258911, + 0.03305628150701523, + 0.015543154440820217, + 0.008809253573417664, + 0.010020715184509754, + 0.047485627233982086, + 0.006696159020066261, + 0.007703923154622316, + 0.007358736824244261, + 0.01500257570296526, + 0.005844274535775185, + 0.006854191888123751, + 0.007552369497716427, + 0.009376063011586666, + 0.008573147468268871, + 0.006310200784355402, + 0.0052566672675311565, + 0.00619284063577652, + 0.006104455795139074, + 0.01123143918812275, + 0.006999722216278315, + 0.008595511317253113, + 0.007164238952100277, + 0.005109479650855064, + 0.0018642009235918522, + 0.002170882886275649, + 0.0021462745498865843, + 0.0031192826572805643, + 0.003497286234050989, + 0.002542519010603428, + 0.00324835442006588, + 0.0016390361124649644 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 3, + 6 + ], + "selected_reward": -0.17985419929027557 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000095/group_8/index_tensor([6733, 6733])_group_8_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000095/group_8/index_tensor([6733, 6733])_group_8_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..4070aa9c31808574f6662abb4c9137eabb1277e5 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000095/group_8/index_tensor([6733, 6733])_group_8_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "smooth, slightly low-angle tracking shot follows a sleek, modern black sedan, its glossy paint reflecting the ambient light like liquid obsidian. Its bright white LED headlights slice sharply through the deepening twilight as it steadily glides from left to right across a wide, multi-lane asphalt road. The sidewalk is clearly marked with large, red and white directional arrows, their colors vibrant against the dark surface. In the background, a dense cityscape of towering glass and steel skyscrapers stands against a dramatic sky, ablaze with the warm golden glow of sunset, transitioning to a deep blue overhead. A prominent neon \"Dining Club\" sign hums, casting a vibrant pink and blue glow on the street and the sides of passing vehicles. The camera perfectly matches the car's constant speed, creating a fluid and captivating sense of movement, capturing the city's complex, electric atmosphere as it shifts from day to night.", + "seed": 1212, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000095/group_8/index_6733_seed_1212_dp_16_rank_32_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 8, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.03712666304782033, + "vlm_logprob": -0.03712666304782033, + "vlm_yes_no_margin": 3.75, + "vlm_token_credit": -0.03712666304782033, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a sleek, modern black sedan with glossy paint?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the car's headlights bright white LEDs that slice through the twilight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the car moving steadily from left to right across a wide, multi-lane asphalt road?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a dense cityscape of glass and steel skyscrapers with a sunset sky transitioning to deep blue?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera perform a smooth tracking shot matching the car's speed from a slightly low angle?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.012744174338877201 + ], + "opposite_logprob_tokens": [ + -4.762743949890137 + ], + "logprob_reward": -0.012744174338877201, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.012744174338877201, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.133308544754982, + 0.045823123306035995, + 0.06958659738302231, + 0.08533020317554474, + 0.09934989362955093, + 0.13215060532093048, + 0.07533859461545944, + 0.08269480615854263, + 0.13637840747833252, + 0.14003919064998627 + ], + "credit_map": [ + 0.007629651576280594, + 0.011101099662482738, + 0.010239492170512676, + 0.009832115843892097, + 0.032327860593795776, + 0.012468181550502777, + 0.03845623508095741, + 0.011253900825977325, + 0.006627778988331556, + 0.008077576756477356, + 0.004884680267423391, + 0.006038065999746323, + 0.004874797072261572, + 0.005685967393219471, + 0.004352465271949768, + 0.005281792022287846, + 0.0053250351920723915, + 0.010707869194447994, + 0.00572211155667901, + 0.008585651405155659, + 0.005556007381528616, + 0.005724402144551277, + 0.01023599598556757, + 0.01772952266037464, + 0.014673907309770584, + 0.009937592782080173, + 0.011260667815804482, + 0.012423710897564888, + 0.006876987870782614, + 0.007082620169967413, + 0.012905186042189598, + 0.010169528424739838, + 0.009876897558569908, + 0.008305074647068977, + 0.019936010241508484, + 0.015102483332157135, + 0.006469475105404854, + 0.009393543004989624, + 0.01568453200161457, + 0.01458188146352768, + 0.016489636152982712, + 0.009465992450714111, + 0.022358566522598267, + 0.03757382184267044, + 0.006079762242734432, + 0.00664988299831748, + 0.022396313026547432, + 0.011136634275317192, + 0.005600918550044298, + 0.0038650124333798885, + 0.020027486607432365, + 0.011395002715289593, + 0.005011528730392456, + 0.005247812252491713, + 0.012984857894480228, + 0.011205972172319889, + 0.0055648209527134895, + 0.006595018785446882, + 0.008574135601520538, + 0.023516932502388954, + 0.0037707812152802944, + 0.004514837637543678, + 0.007967877201735973, + 0.022190405055880547, + 0.00715330895036459, + 0.00704664783552289, + 0.030549053102731705, + 0.024748550727963448, + 0.007613727357238531, + 0.008973333053290844, + 0.03798997402191162, + 0.012303806841373444, + 0.01142487395554781, + 0.010102172382175922, + 0.023637864738702774, + 0.026887189596891403, + 0.012729329988360405, + 0.014307626523077488, + 0.024929989129304886, + 0.016020143404603004 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 8, + 0 + ], + "selected_reward": -0.012744174338877201 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a sleek, modern black sedan with glossy paint?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the car's headlights bright white LEDs that slice through the twilight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the car moving steadily from left to right across a wide, multi-lane asphalt road?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a dense cityscape of glass and steel skyscrapers with a sunset sky transitioning to deep blue?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera perform a smooth tracking shot matching the car's speed from a slightly low angle?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07900508493185043 + ], + "opposite_logprob_tokens": [ + -2.579005002975464 + ], + "logprob_reward": -0.07900508493185043, + "yes_no_margin_reward": 2.5, + "token_credit_reward": -0.07900508493185043, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.18050388991832733, + 0.04907853528857231, + 0.031020527705550194, + 0.051520321518182755, + 0.06210373342037201, + 0.16971449553966522, + 0.08905861526727676, + 0.06044696271419525, + 0.09048768132925034, + 0.2160651981830597 + ], + "credit_map": [ + 0.011396955698728561, + 0.010126577690243721, + 0.009815101511776447, + 0.009109968319535255, + 0.04898309335112572, + 0.016648653894662857, + 0.05848757550120354, + 0.01593596488237381, + 0.006194101180881262, + 0.009500207379460335, + 0.005539227742701769, + 0.005939282011240721, + 0.00530650420114398, + 0.0064834714867174625, + 0.004913515876978636, + 0.005202228669077158, + 0.003629507729783654, + 0.002168528037145734, + 0.0026735523715615273, + 0.004360034596174955, + 0.0035280354786664248, + 0.00410463847219944, + 0.002286772010847926, + 0.008269460871815681, + 0.0054310159757733345, + 0.0066234758123755455, + 0.007688259240239859, + 0.008341924287378788, + 0.003615745110437274, + 0.004045120906084776, + 0.008619396947324276, + 0.007155387196689844, + 0.008704576641321182, + 0.005429030861705542, + 0.010600299574434757, + 0.00990801490843296, + 0.00475449301302433, + 0.005099921021610498, + 0.008744338527321815, + 0.008863058872520924, + 0.02052430622279644, + 0.012281814590096474, + 0.029219960793852806, + 0.049851760268211365, + 0.008211960084736347, + 0.009031976573169231, + 0.026461327448487282, + 0.014131399802863598, + 0.005579234566539526, + 0.006305890157818794, + 0.020199207589030266, + 0.012215797789394855, + 0.009427078999578953, + 0.006089345086365938, + 0.015226058661937714, + 0.014016009867191315, + 0.004673556890338659, + 0.004113206639885902, + 0.008099342696368694, + 0.015638338401913643, + 0.003174313111230731, + 0.003511440707370639, + 0.007029077969491482, + 0.0142076900228858, + 0.005232128314673901, + 0.006617910228669643, + 0.02044663578271866, + 0.01191148441284895, + 0.004542967304587364, + 0.006982158403843641, + 0.02616065926849842, + 0.008593740873038769, + 0.01849398761987686, + 0.016385518014431, + 0.03504904732108116, + 0.041937652975320816, + 0.020359665155410767, + 0.02297382801771164, + 0.035975780338048935, + 0.02488972619175911 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 0, + 5 + ], + "selected_reward": -0.07900508493185043 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a sleek, modern black sedan with glossy paint?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the car's headlights bright white LEDs that slice through the twilight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the car moving steadily from left to right across a wide, multi-lane asphalt road?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a dense cityscape of glass and steel skyscrapers with a sunset sky transitioning to deep blue?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera perform a smooth tracking shot matching the car's speed from a slightly low angle?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03369291499257088 + ], + "opposite_logprob_tokens": [ + -3.4086928367614746 + ], + "logprob_reward": -0.03369291499257088, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.03369291499257088, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.21515049040317535, + 0.07362436503171921, + 0.07088673114776611, + 0.19239358603954315, + 0.05731262266635895, + 0.09514826536178589, + 0.07415500283241272, + 0.053441014140844345, + 0.06600801646709442, + 0.10187986493110657 + ], + "credit_map": [ + 0.013976558111608028, + 0.0108658317476511, + 0.018350567668676376, + 0.011902639642357826, + 0.05461332947015762, + 0.018212245777249336, + 0.06787917762994766, + 0.019350141286849976, + 0.008238792419433594, + 0.012154825031757355, + 0.009360593743622303, + 0.009488647803664207, + 0.008178756572306156, + 0.00700575252994895, + 0.008037164807319641, + 0.01115983072668314, + 0.006249929778277874, + 0.008408368565142155, + 0.0048135207034647465, + 0.012829065322875977, + 0.005870592314749956, + 0.006913766730576754, + 0.007303474936634302, + 0.01849800907075405, + 0.02342408336699009, + 0.024885285645723343, + 0.02428862266242504, + 0.029264939948916435, + 0.01600782573223114, + 0.01647118665277958, + 0.029607854783535004, + 0.028443794697523117, + 0.00617663050070405, + 0.0075833979062736034, + 0.006952788680791855, + 0.007703812792897224, + 0.0058899992145597935, + 0.00503716804087162, + 0.00959158968180418, + 0.00837723445147276, + 0.009214932098984718, + 0.0061805350705981255, + 0.017104653641581535, + 0.027380583807826042, + 0.004970173351466656, + 0.006244294345378876, + 0.014998296275734901, + 0.009054802358150482, + 0.003716853680089116, + 0.004486728459596634, + 0.01959886960685253, + 0.012058336287736893, + 0.005905694328248501, + 0.003968943376094103, + 0.012277049012482166, + 0.012142525054514408, + 0.003125852206721902, + 0.00289592775516212, + 0.005149552598595619, + 0.015835948288440704, + 0.0018346147844567895, + 0.0022456152364611626, + 0.005362512078136206, + 0.016990989446640015, + 0.003661690978333354, + 0.0030046699102967978, + 0.019813984632492065, + 0.01003320887684822, + 0.002218415727838874, + 0.00326317036524415, + 0.01852528005838394, + 0.0054875994101166725, + 0.009446600452065468, + 0.009435899555683136, + 0.013231677003204823, + 0.019458387047052383, + 0.0120294950902462, + 0.013689713552594185, + 0.014300623908638954, + 0.010287459939718246 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 9 + ], + "selected_reward": -0.03369291499257088 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a sleek, modern black sedan with glossy paint?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the car's headlights bright white LEDs that slice through the twilight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the car moving steadily from left to right across a wide, multi-lane asphalt road?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a dense cityscape of glass and steel skyscrapers with a sunset sky transitioning to deep blue?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera perform a smooth tracking shot matching the car's speed from a slightly low angle?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.05494481697678566 + ], + "opposite_logprob_tokens": [ + -2.9299447536468506 + ], + "logprob_reward": -0.05494481697678566, + "yes_no_margin_reward": 2.875, + "token_credit_reward": -0.05494481697678566, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2410098910331726, + 0.10926689207553864, + 0.2281516045331955, + 0.1376785933971405, + 0.03903491795063019, + 0.02884596399962902, + 0.08134456723928452, + 0.04297110438346863, + 0.04533974081277847, + 0.046356771141290665 + ], + "credit_map": [ + 0.014772764407098293, + 0.01132185384631157, + 0.011000530794262886, + 0.01193386409431696, + 0.0636744350194931, + 0.02781059220433235, + 0.07792787998914719, + 0.02256796509027481, + 0.020790012553334236, + 0.014716246165335178, + 0.012467063032090664, + 0.006605905015021563, + 0.011864803731441498, + 0.014670420438051224, + 0.01020780298858881, + 0.017944639548659325, + 0.022536501288414, + 0.02302904613316059, + 0.0192716084420681, + 0.03660452365875244, + 0.020025940611958504, + 0.021618427708745003, + 0.018759436905384064, + 0.06630612909793854, + 0.028185710310935974, + 0.015136688016355038, + 0.017672033980488777, + 0.0197096336632967, + 0.010516177862882614, + 0.014157894998788834, + 0.014538181014358997, + 0.017762266099452972, + 0.005195521283894777, + 0.0034070226829499006, + 0.007590575143694878, + 0.005664906930178404, + 0.0025715085212141275, + 0.0040796962566673756, + 0.004641386680305004, + 0.005884299520403147, + 0.0029477111529558897, + 0.0022735975217074156, + 0.005768436007201672, + 0.005935944151133299, + 0.0017319396138191223, + 0.0019714729860424995, + 0.005546058993786573, + 0.0026708042714744806, + 0.00741950049996376, + 0.006586071103811264, + 0.010761209763586521, + 0.009724842384457588, + 0.012108988128602505, + 0.010337739251554012, + 0.011960746720433235, + 0.012445463798940182, + 0.0037168683484196663, + 0.0034524418879300356, + 0.00501147098839283, + 0.008521066047251225, + 0.0022188688162714243, + 0.003337687347084284, + 0.00651064096018672, + 0.010202059522271156, + 0.0034414418041706085, + 0.004051772877573967, + 0.008404717780649662, + 0.007352829910814762, + 0.0028337365947663784, + 0.0038746821228414774, + 0.00943768210709095, + 0.005942881107330322, + 0.00385965290479362, + 0.004240124486386776, + 0.007055153604596853, + 0.00918459054082632, + 0.004187819082289934, + 0.005270084831863642, + 0.007489933632314205, + 0.005069411359727383 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 3 + ], + "selected_reward": -0.05494481697678566 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a sleek, modern black sedan with glossy paint?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the car's headlights bright white LEDs that slice through the twilight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the car moving steadily from left to right across a wide, multi-lane asphalt road?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a dense cityscape of glass and steel skyscrapers with a sunset sky transitioning to deep blue?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera perform a smooth tracking shot matching the car's speed from a slightly low angle?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005246323999017477 + ], + "opposite_logprob_tokens": [ + -5.255246162414551 + ], + "logprob_reward": -0.005246323999017477, + "yes_no_margin_reward": 5.25, + "token_credit_reward": -0.005246323999017477, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08654395490884781, + 0.029628105461597443, + 0.059530094265937805, + 0.08719731122255325, + 0.10176705569028854, + 0.12636041641235352, + 0.10218603163957596, + 0.08696126192808151, + 0.1233661025762558, + 0.19645963609218597 + ], + "credit_map": [ + 0.006024450063705444, + 0.005790530703961849, + 0.006597984582185745, + 0.007106157019734383, + 0.017364343628287315, + 0.014545204117894173, + 0.02126707136631012, + 0.007848212495446205, + 0.003929054364562035, + 0.003984467126429081, + 0.003526800312101841, + 0.0032510128803551197, + 0.0028881540056318045, + 0.0035491513554006815, + 0.002961773192510009, + 0.005537693854421377, + 0.004459381569176912, + 0.01256861537694931, + 0.004434456583112478, + 0.008056092076003551, + 0.006027825176715851, + 0.006424308754503727, + 0.009087670594453812, + 0.008471743203699589, + 0.014869445003569126, + 0.009140870533883572, + 0.012240308336913586, + 0.0143419299274683, + 0.006671614944934845, + 0.009328401647508144, + 0.008353770710527897, + 0.012250978499650955, + 0.013511604629456997, + 0.009472142904996872, + 0.020795907825231552, + 0.014333569444715977, + 0.0060755000449717045, + 0.009731650352478027, + 0.014492296613752842, + 0.013354386202991009, + 0.014482976868748665, + 0.009370344690978527, + 0.023499734699726105, + 0.03580606356263161, + 0.00733087258413434, + 0.010266212746500969, + 0.015281816013157368, + 0.010322394780814648, + 0.005283826030790806, + 0.006403907667845488, + 0.027454782277345657, + 0.016737516969442368, + 0.00932290405035019, + 0.006928571499884129, + 0.015478073619306087, + 0.01457645371556282, + 0.009751574136316776, + 0.01060186792165041, + 0.011245205998420715, + 0.018794335424900055, + 0.004820522386580706, + 0.005111453589051962, + 0.012365266680717468, + 0.014271036721765995, + 0.009503464214503765, + 0.012095419690012932, + 0.022166498005390167, + 0.018998589366674423, + 0.007439699489623308, + 0.011495987884700298, + 0.02692587859928608, + 0.014740565791726112, + 0.016439327970147133, + 0.013586812652647495, + 0.03341208025813103, + 0.04053214192390442, + 0.020680686458945274, + 0.019197452813386917, + 0.030242422595620155, + 0.022368714213371277 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 5, + 8 + ], + "selected_reward": -0.005246323999017477 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "smooth, slightly low-angle tracking shot follows a sleek, modern black sedan, its glossy paint reflecting the ambient light like liquid obsidian. Its bright white LED headlights slice sharply through the deepening twilight as it steadily glides from left to right across a wide, multi-lane asphalt road. The sidewalk is clearly marked with large, red and white directional arrows, their colors vibrant against the dark surface. In the background, a dense cityscape of towering glass and steel skyscrapers stands against a dramatic sky, ablaze with the warm golden glow of sunset, transitioning to a deep blue overhead. A prominent neon \"Dining Club\" sign hums, casting a vibrant pink and blue glow on the street and the sides of passing vehicles. The camera perfectly matches the car's constant speed, creating a fluid and captivating sense of movement, capturing the city's complex, electric atmosphere as it shifts from day to night.", + "seed": 1213, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000095/group_8/index_6733_seed_1213_dp_16_rank_32_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 8, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.032846881169825794, + "vlm_logprob": -0.032846881169825794, + "vlm_yes_no_margin": 3.925, + "vlm_token_credit": -0.032846881169825794, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a sleek, modern black sedan with glossy paint?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the car's headlights bright white LEDs that slice through the twilight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the car moving steadily from left to right across a wide, multi-lane asphalt road?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a dense cityscape of glass and steel skyscrapers with a sunset sky transitioning to deep blue?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera perform a smooth tracking shot matching the car's speed from a slightly low angle?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.012692502699792385 + ], + "opposite_logprob_tokens": [ + -5.012692451477051 + ], + "logprob_reward": -0.012692502699792385, + "yes_no_margin_reward": 5.0, + "token_credit_reward": -0.012692502699792385, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.128244087100029, + 0.06267746537923813, + 0.048451680690050125, + 0.10433004796504974, + 0.09158354252576828, + 0.05782829970121384, + 0.09081903100013733, + 0.07240669429302216, + 0.19693730771541595, + 0.14672181010246277 + ], + "credit_map": [ + 0.011427581310272217, + 0.012487531639635563, + 0.01673140376806259, + 0.015566634945571423, + 0.025009023025631905, + 0.02131248265504837, + 0.012757602147758007, + 0.012951835989952087, + 0.0059015946462750435, + 0.006440802942961454, + 0.0100718904286623, + 0.008599115535616875, + 0.007526258006691933, + 0.007755074184387922, + 0.00796427484601736, + 0.00841845478862524, + 0.005827402230352163, + 0.006195720285177231, + 0.004283785354346037, + 0.005406610667705536, + 0.009089507162570953, + 0.006668621674180031, + 0.006035034544765949, + 0.004945001099258661, + 0.01511397585272789, + 0.01989680528640747, + 0.010029208846390247, + 0.010285166092216969, + 0.015697767958045006, + 0.014249768108129501, + 0.010694793425500393, + 0.008362566120922565, + 0.0069455974735319614, + 0.007233538664877415, + 0.008759290911257267, + 0.01078537106513977, + 0.02205345220863819, + 0.015075759962201118, + 0.010575595311820507, + 0.010154936462640762, + 0.004925440531224012, + 0.0051175267435610294, + 0.005590211600065231, + 0.009543856605887413, + 0.008968021720647812, + 0.0083648432046175, + 0.009131078608334064, + 0.006187323480844498, + 0.008905521593987942, + 0.009452670812606812, + 0.009355049580335617, + 0.016489705070853233, + 0.014477666467428207, + 0.007269381545484066, + 0.0180059801787138, + 0.0068630618043243885, + 0.007500493433326483, + 0.009239871986210346, + 0.009630206041038036, + 0.011747804470360279, + 0.011562639847397804, + 0.004802036099135876, + 0.011400463059544563, + 0.006523186806589365, + 0.026243433356285095, + 0.017420461401343346, + 0.015174218453466892, + 0.041540663689374924, + 0.019960535690188408, + 0.018777893856167793, + 0.034258682280778885, + 0.023561429232358932, + 0.01180455181747675, + 0.0115897748619318, + 0.017439862713217735, + 0.025620657950639725, + 0.01972704939544201, + 0.014034200459718704, + 0.02387087605893612, + 0.0226348415017128 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 9, + 0 + ], + "selected_reward": -0.012692502699792385 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a sleek, modern black sedan with glossy paint?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the car's headlights bright white LEDs that slice through the twilight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the car moving steadily from left to right across a wide, multi-lane asphalt road?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a dense cityscape of glass and steel skyscrapers with a sunset sky transitioning to deep blue?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera perform a smooth tracking shot matching the car's speed from a slightly low angle?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07901841402053833 + ], + "opposite_logprob_tokens": [ + -2.5790183544158936 + ], + "logprob_reward": -0.07901841402053833, + "yes_no_margin_reward": 2.5, + "token_credit_reward": -0.07901841402053833, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10849077254533768, + 0.08008057624101639, + 0.0324033722281456, + 0.3383868634700775, + 0.03060929849743843, + 0.019079161807894707, + 0.05166909098625183, + 0.048038601875305176, + 0.22880864143371582, + 0.062433529645204544 + ], + "credit_map": [ + 0.013901659287512302, + 0.0107000432908535, + 0.019665934145450592, + 0.013247471302747726, + 0.01511024497449398, + 0.013742223381996155, + 0.0114010414108634, + 0.010722163133323193, + 0.00610911101102829, + 0.007664477452635765, + 0.01143572572618723, + 0.01220759842544794, + 0.012636201456189156, + 0.010216780006885529, + 0.011660574935376644, + 0.008150105364620686, + 0.004371765535324812, + 0.004459502175450325, + 0.0023204253520816565, + 0.003564199199900031, + 0.005953683052212, + 0.004083032254129648, + 0.004087202250957489, + 0.0035635626409202814, + 0.043975070118904114, + 0.03445228934288025, + 0.03092721663415432, + 0.041098542511463165, + 0.07757601141929626, + 0.04948723688721657, + 0.035619862377643585, + 0.025250662118196487, + 0.002097327494993806, + 0.0025894290301948786, + 0.003474618773907423, + 0.00477435952052474, + 0.004447387531399727, + 0.004321212414652109, + 0.005674370564520359, + 0.0032305954955518246, + 0.0015509235672652721, + 0.001748395268805325, + 0.002581211971119046, + 0.003173383418470621, + 0.0030246363021433353, + 0.0024786689318716526, + 0.0023417591582983732, + 0.0021801828406751156, + 0.0035776018630713224, + 0.004501829389482737, + 0.004248417913913727, + 0.008899363689124584, + 0.007020452059805393, + 0.0032142633572220802, + 0.013084094040095806, + 0.007123068906366825, + 0.00578707130625844, + 0.0036760116927325726, + 0.0052315182983875275, + 0.007640630938112736, + 0.0038270705845206976, + 0.004434774164110422, + 0.010379020124673843, + 0.0070625049993395805, + 0.04416833817958832, + 0.02200382761657238, + 0.022783447057008743, + 0.03629624471068382, + 0.02853064052760601, + 0.01554030366241932, + 0.0308841560035944, + 0.028601691126823425, + 0.009521734900772572, + 0.005436975508928299, + 0.0064014713279902935, + 0.007922629825770855, + 0.006538422778248787, + 0.007102862931787968, + 0.00840468518435955, + 0.011104748584330082 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 8, + 0 + ], + "selected_reward": -0.07901841402053833 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a sleek, modern black sedan with glossy paint?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the car's headlights bright white LEDs that slice through the twilight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the car moving steadily from left to right across a wide, multi-lane asphalt road?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a dense cityscape of glass and steel skyscrapers with a sunset sky transitioning to deep blue?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera perform a smooth tracking shot matching the car's speed from a slightly low angle?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.029795659705996513 + ], + "opposite_logprob_tokens": [ + -3.5297956466674805 + ], + "logprob_reward": -0.029795659705996513, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.029795659705996513, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09092701226472855, + 0.06846261769533157, + 0.038242146372795105, + 0.09967491775751114, + 0.4012793302536011, + 0.049294278025627136, + 0.06374712288379669, + 0.04268438369035721, + 0.08122256398200989, + 0.06446558237075806 + ], + "credit_map": [ + 0.006766259670257568, + 0.007718938402831554, + 0.02058914490044117, + 0.016883522272109985, + 0.008020369336009026, + 0.010601051151752472, + 0.009263875894248486, + 0.011083855293691158, + 0.0049042836762964725, + 0.006934660952538252, + 0.009314204566180706, + 0.012490160763263702, + 0.00888831913471222, + 0.009112446568906307, + 0.008955691009759903, + 0.007862849161028862, + 0.003447609953582287, + 0.004162238910794258, + 0.004255968611687422, + 0.005063348449766636, + 0.006934975739568472, + 0.0051534706726670265, + 0.005196282174438238, + 0.0040282499976456165, + 0.009396369569003582, + 0.013516624458134174, + 0.013929713517427444, + 0.009648196399211884, + 0.016907041892409325, + 0.015712136402726173, + 0.007205418776720762, + 0.013359415344893932, + 0.03131992742419243, + 0.030584637075662613, + 0.03963271155953407, + 0.055051982402801514, + 0.08436824381351471, + 0.06236137077212334, + 0.055316124111413956, + 0.04264433681964874, + 0.004596570041030645, + 0.00844701286405325, + 0.0069647980853915215, + 0.006107323803007603, + 0.007237253710627556, + 0.006042224355041981, + 0.005552275106310844, + 0.004346821457147598, + 0.0034626757260411978, + 0.004417336080223322, + 0.00723537802696228, + 0.01107576210051775, + 0.0058667478151619434, + 0.0037243436090648174, + 0.01658533699810505, + 0.0113795455545187, + 0.005235718097537756, + 0.005264756735414267, + 0.004855361767113209, + 0.007758149411529303, + 0.00600662874057889, + 0.0023193880915641785, + 0.0066413069143891335, + 0.00460307439789176, + 0.010040475986897945, + 0.008268564939498901, + 0.007506968453526497, + 0.017458410933613777, + 0.010034032166004181, + 0.0045437077060341835, + 0.014706900343298912, + 0.008663508109748363, + 0.007941631600260735, + 0.005671970080584288, + 0.005550611764192581, + 0.01116923987865448, + 0.007249516900628805, + 0.004602468106895685, + 0.01186826266348362, + 0.010411878116428852 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 3, + 0 + ], + "selected_reward": -0.029795659705996513 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a sleek, modern black sedan with glossy paint?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the car's headlights bright white LEDs that slice through the twilight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the car moving steadily from left to right across a wide, multi-lane asphalt road?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a dense cityscape of glass and steel skyscrapers with a sunset sky transitioning to deep blue?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera perform a smooth tracking shot matching the car's speed from a slightly low angle?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.038094662129879 + ], + "opposite_logprob_tokens": [ + -3.2880947589874268 + ], + "logprob_reward": -0.038094662129879, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.038094662129879, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12243472784757614, + 0.21774718165397644, + 0.050767604261636734, + 0.07211856544017792, + 0.13422776758670807, + 0.09820277243852615, + 0.06561049818992615, + 0.029013164341449738, + 0.15705786645412445, + 0.05281991884112358 + ], + "credit_map": [ + 0.01721319742500782, + 0.01734260469675064, + 0.014707068912684917, + 0.012990294955670834, + 0.021305063739418983, + 0.016036391258239746, + 0.010731951333582401, + 0.012108159251511097, + 0.02312033250927925, + 0.02190518006682396, + 0.038841310888528824, + 0.024257808923721313, + 0.02842091955244541, + 0.025049366056919098, + 0.02255588211119175, + 0.03359637036919594, + 0.008013961836695671, + 0.007468454539775848, + 0.00401066429913044, + 0.004301393870264292, + 0.011589487083256245, + 0.008198453113436699, + 0.0034917378798127174, + 0.0036934497766196728, + 0.009004072286188602, + 0.006971550174057484, + 0.00722063984721899, + 0.00859872717410326, + 0.016471464186906815, + 0.010823847725987434, + 0.007705532480031252, + 0.005322734359651804, + 0.008511118590831757, + 0.009599769487977028, + 0.015116224996745586, + 0.014829277992248535, + 0.03433676436543465, + 0.022671204060316086, + 0.017094967886805534, + 0.012068431824445724, + 0.008819903247058392, + 0.007670761551707983, + 0.010665269568562508, + 0.010901453904807568, + 0.016589457169175148, + 0.01394050382077694, + 0.020273052155971527, + 0.009342369623482227, + 0.0051162028685212135, + 0.006690599489957094, + 0.00503504229709506, + 0.011286109685897827, + 0.012545932084321976, + 0.004030751064419746, + 0.013715172186493874, + 0.007190692238509655, + 0.0030258020851761103, + 0.0050862873904407024, + 0.00485336733981967, + 0.0039027605671435595, + 0.00324690411798656, + 0.0021486268378794193, + 0.0038414252921938896, + 0.0029079921077936888, + 0.027182748541235924, + 0.013513215817511082, + 0.01486306544393301, + 0.031003795564174652, + 0.01719563826918602, + 0.010884741321206093, + 0.026683147996664047, + 0.015731513500213623, + 0.00872515793889761, + 0.003769629169255495, + 0.0041893464513123035, + 0.007785931695252657, + 0.005233152769505978, + 0.006715377327054739, + 0.007678085006773472, + 0.008723236620426178 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 8, + 4 + ], + "selected_reward": -0.038094662129879 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a sleek, modern black sedan with glossy paint?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the car's headlights bright white LEDs that slice through the twilight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the car moving steadily from left to right across a wide, multi-lane asphalt road?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a dense cityscape of glass and steel skyscrapers with a sunset sky transitioning to deep blue?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera perform a smooth tracking shot matching the car's speed from a slightly low angle?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004633167292922735 + ], + "opposite_logprob_tokens": [ + -5.379632949829102 + ], + "logprob_reward": -0.004633167292922735, + "yes_no_margin_reward": 5.375, + "token_credit_reward": -0.004633167292922735, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12980686128139496, + 0.06537404656410217, + 0.052334289997816086, + 0.07856802642345428, + 0.17062312364578247, + 0.08723611384630203, + 0.0575125552713871, + 0.05978008732199669, + 0.1697007715702057, + 0.12906411290168762 + ], + "credit_map": [ + 0.01421261765062809, + 0.01512096542865038, + 0.02254527062177658, + 0.016756100580096245, + 0.014072751626372337, + 0.02397402748465538, + 0.011495045386254787, + 0.011630075052380562, + 0.005285721737891436, + 0.00922302808612585, + 0.00797504186630249, + 0.009211327880620956, + 0.0099795488640666, + 0.008179686032235622, + 0.009464612230658531, + 0.006055074743926525, + 0.007314543705433607, + 0.004398492630571127, + 0.005484056659042835, + 0.005964362528175116, + 0.007355342619121075, + 0.0072410618886351585, + 0.00855220202356577, + 0.0060242293402552605, + 0.008590176701545715, + 0.012250883504748344, + 0.006430432666093111, + 0.007809078320860863, + 0.017736107110977173, + 0.011639414355158806, + 0.007949675433337688, + 0.006162257399410009, + 0.014058778062462807, + 0.013750109821557999, + 0.018755655735731125, + 0.024754753336310387, + 0.035363975912332535, + 0.02543933130800724, + 0.022991936653852463, + 0.015508584678173065, + 0.006450993474572897, + 0.008438308723270893, + 0.007079666946083307, + 0.013006663881242275, + 0.016758913174271584, + 0.014427966438233852, + 0.013975095003843307, + 0.00709850387647748, + 0.0045953006483614445, + 0.006606866605579853, + 0.00817903969436884, + 0.010431786999106407, + 0.007929627783596516, + 0.003960458096116781, + 0.008979874663054943, + 0.0068295979872345924, + 0.008921933360397816, + 0.008037927560508251, + 0.007576031144708395, + 0.009435438551008701, + 0.00711089139804244, + 0.0035276974085718393, + 0.009744641371071339, + 0.005425524897873402, + 0.031728584319353104, + 0.018749795854091644, + 0.017731018364429474, + 0.023694196715950966, + 0.022902216762304306, + 0.011398223228752613, + 0.021869482472538948, + 0.021627256646752357, + 0.02685500495135784, + 0.009185516275465488, + 0.012164242565631866, + 0.02278055064380169, + 0.010279080830514431, + 0.006858427543193102, + 0.01844894327223301, + 0.022492345422506332 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 8, + 0 + ], + "selected_reward": -0.004633167292922735 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "smooth, slightly low-angle tracking shot follows a sleek, modern black sedan, its glossy paint reflecting the ambient light like liquid obsidian. Its bright white LED headlights slice sharply through the deepening twilight as it steadily glides from left to right across a wide, multi-lane asphalt road. The sidewalk is clearly marked with large, red and white directional arrows, their colors vibrant against the dark surface. In the background, a dense cityscape of towering glass and steel skyscrapers stands against a dramatic sky, ablaze with the warm golden glow of sunset, transitioning to a deep blue overhead. A prominent neon \"Dining Club\" sign hums, casting a vibrant pink and blue glow on the street and the sides of passing vehicles. The camera perfectly matches the car's constant speed, creating a fluid and captivating sense of movement, capturing the city's complex, electric atmosphere as it shifts from day to night.", + "seed": 1214, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000095/group_8/index_6733_seed_1214_dp_17_rank_34_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 8, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.07011579321697355, + "vlm_logprob": -0.07011579321697355, + "vlm_yes_no_margin": 3.200000023841858, + "vlm_token_credit": -0.07011579321697355, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a sleek, modern black sedan with glossy paint?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the car's headlights bright white LEDs that slice through the twilight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the car moving steadily from left to right across a wide, multi-lane asphalt road?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a dense cityscape of glass and steel skyscrapers with a sunset sky transitioning to deep blue?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera perform a smooth tracking shot matching the car's speed from a slightly low angle?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0291694737970829 + ], + "opposite_logprob_tokens": [ + -3.6541695594787598 + ], + "logprob_reward": -0.0291694737970829, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.0291694737970829, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.027584640309214592, + 0.02465573512017727, + 0.03852834925055504, + 0.02407795377075672, + 0.48110106587409973, + 0.09338715672492981, + 0.06247527897357941, + 0.02684919722378254, + 0.08369528502225876, + 0.13764524459838867 + ], + "credit_map": [ + 0.0016440192703157663, + 0.0031187650747597218, + 0.002673744224011898, + 0.002948726760223508, + 0.0027390546165406704, + 0.0033190411049872637, + 0.005537934601306915, + 0.005603355821222067, + 0.0019165505655109882, + 0.003031963249668479, + 0.004313106648623943, + 0.0027039176784455776, + 0.0032498310320079327, + 0.0020561215933412313, + 0.004128259606659412, + 0.003255984978750348, + 0.003020097268745303, + 0.004624359775334597, + 0.0027298845816403627, + 0.004435977432876825, + 0.004470104817301035, + 0.007031621411442757, + 0.008111683651804924, + 0.0041046165861189365, + 0.0030502057634294033, + 0.002265067072585225, + 0.004316029604524374, + 0.003445055801421404, + 0.0018100187880918384, + 0.002633102238178253, + 0.002576657338067889, + 0.003981816582381725, + 0.02814682573080063, + 0.026408148929476738, + 0.056590043008327484, + 0.15378080308437347, + 0.02470347285270691, + 0.032317016273736954, + 0.09143111109733582, + 0.06772365421056747, + 0.0036948940251022577, + 0.005549325607717037, + 0.011645353399217129, + 0.03172047808766365, + 0.003056710120290518, + 0.0023841920774430037, + 0.02759811468422413, + 0.007738084997981787, + 0.006965549197047949, + 0.010404256172478199, + 0.012628704309463501, + 0.012529803439974785, + 0.003951867576688528, + 0.0025646144058555365, + 0.0071048312820494175, + 0.0063256509602069855, + 0.0018446041503921151, + 0.0015361037803813815, + 0.0036787965800613165, + 0.004429442808032036, + 0.0021712989546358585, + 0.0020228479988873005, + 0.00396808423101902, + 0.007198019418865442, + 0.0036853402853012085, + 0.00493788393214345, + 0.013326593674719334, + 0.012359164655208588, + 0.0045739044435322285, + 0.004104312043637037, + 0.011430165730416775, + 0.029277918860316277, + 0.01269204169511795, + 0.007900386117398739, + 0.03008400648832321, + 0.03901546448469162, + 0.00741557078436017, + 0.007842175662517548, + 0.016510507091879845, + 0.016185101121664047 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 9, + 5 + ], + "selected_reward": -0.0291694737970829 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a sleek, modern black sedan with glossy paint?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the car's headlights bright white LEDs that slice through the twilight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the car moving steadily from left to right across a wide, multi-lane asphalt road?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a dense cityscape of glass and steel skyscrapers with a sunset sky transitioning to deep blue?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera perform a smooth tracking shot matching the car's speed from a slightly low angle?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.14285874366760254 + ], + "opposite_logprob_tokens": [ + -2.0178587436676025 + ], + "logprob_reward": -0.14285874366760254, + "yes_no_margin_reward": 1.875, + "token_credit_reward": -0.14285874366760254, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07578665763139725, + 0.04538580775260925, + 0.03170250356197357, + 0.033733271062374115, + 0.20630882680416107, + 0.20091873407363892, + 0.10239534080028534, + 0.055961813777685165, + 0.09018073976039886, + 0.15762625634670258 + ], + "credit_map": [ + 0.005831572227180004, + 0.008639629930257797, + 0.0068815648555755615, + 0.009135316126048565, + 0.00789723638445139, + 0.009980528615415096, + 0.01201485563069582, + 0.015405950136482716, + 0.004151421599090099, + 0.004989061038941145, + 0.007140784524381161, + 0.005102210678160191, + 0.005867945030331612, + 0.0044083185493946075, + 0.007663828320801258, + 0.006062238942831755, + 0.0030863150022923946, + 0.004229051526635885, + 0.003007606603205204, + 0.004665832966566086, + 0.0033902477007359266, + 0.005066565237939358, + 0.005065467208623886, + 0.0031914152204990387, + 0.0033248828258365393, + 0.0032474978361278772, + 0.006844386458396912, + 0.0058680749498307705, + 0.0028257553931325674, + 0.003429292468354106, + 0.003769943956285715, + 0.004423434846103191, + 0.014602966606616974, + 0.012494941242039204, + 0.027514737099409103, + 0.05768626928329468, + 0.012463433668017387, + 0.01565883308649063, + 0.02925780601799488, + 0.036629848182201385, + 0.008257661014795303, + 0.009412873536348343, + 0.02402784675359726, + 0.06981338560581207, + 0.005537041462957859, + 0.004877767059952021, + 0.0632118210196495, + 0.01578034646809101, + 0.011863053776323795, + 0.01683252863585949, + 0.017597688362002373, + 0.01605445332825184, + 0.00656293798238039, + 0.004621184431016445, + 0.011922880075871944, + 0.01694061979651451, + 0.004113453906029463, + 0.0030006456654518843, + 0.007775094360113144, + 0.013320820406079292, + 0.003932320978492498, + 0.0037324291188269854, + 0.009218084625899792, + 0.010868960060179234, + 0.006112799514085054, + 0.006228909362107515, + 0.014548693783581257, + 0.014261412434279919, + 0.005516946315765381, + 0.004943336825817823, + 0.013864905573427677, + 0.02470373548567295, + 0.013727433979511261, + 0.013245702721178532, + 0.025780973955988884, + 0.02934741973876953, + 0.017111212015151978, + 0.01873561553657055, + 0.021779675036668777, + 0.017898214980959892 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 5, + 9 + ], + "selected_reward": -0.14285874366760254 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a sleek, modern black sedan with glossy paint?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the car's headlights bright white LEDs that slice through the twilight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the car moving steadily from left to right across a wide, multi-lane asphalt road?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a dense cityscape of glass and steel skyscrapers with a sunset sky transitioning to deep blue?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera perform a smooth tracking shot matching the car's speed from a slightly low angle?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02979809045791626 + ], + "opposite_logprob_tokens": [ + -3.5297980308532715 + ], + "logprob_reward": -0.02979809045791626, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.02979809045791626, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0620315782725811, + 0.04168045148253441, + 0.08132926374673843, + 0.03649180755019188, + 0.15033794939517975, + 0.11042589694261551, + 0.06871297955513, + 0.03960220515727997, + 0.32595208287239075, + 0.0834357962012291 + ], + "credit_map": [ + 0.0043379939161241055, + 0.006638678722083569, + 0.010864856652915478, + 0.00853115227073431, + 0.0065391953103244305, + 0.007585309911519289, + 0.008537521585822105, + 0.008996870368719101, + 0.004832342732697725, + 0.0035404704976826906, + 0.008973909541964531, + 0.005234508775174618, + 0.0030057544354349375, + 0.003795978846028447, + 0.00730872293934226, + 0.004988762084394693, + 0.006005220580846071, + 0.009562027640640736, + 0.007026306819170713, + 0.006349915638566017, + 0.010877660475671291, + 0.01186183001846075, + 0.0182439386844635, + 0.01140235923230648, + 0.002691141329705715, + 0.0045641884207725525, + 0.005703727714717388, + 0.00520631717517972, + 0.0033320337533950806, + 0.004600421991199255, + 0.004663113038986921, + 0.005730861332267523, + 0.009191833436489105, + 0.008569802157580853, + 0.015999913215637207, + 0.048488788306713104, + 0.008851531893014908, + 0.008518815971910954, + 0.03026406466960907, + 0.02045319229364395, + 0.008846063166856766, + 0.007107604760676622, + 0.018997153267264366, + 0.03191487863659859, + 0.004032931290566921, + 0.004354545380920172, + 0.026166200637817383, + 0.00900652538985014, + 0.005715337116271257, + 0.0083704125136137, + 0.012275642715394497, + 0.01381694059818983, + 0.004427631851285696, + 0.003657943569123745, + 0.01176523044705391, + 0.008683842606842518, + 0.0038787100929766893, + 0.0028182007372379303, + 0.004730083979666233, + 0.010132728144526482, + 0.002843706402927637, + 0.0024529118090867996, + 0.007049149833619595, + 0.005696715321391821, + 0.012656529434025288, + 0.013876122422516346, + 0.05303544923663139, + 0.050844836980104446, + 0.012516450136899948, + 0.01436649076640606, + 0.053841378539800644, + 0.11481481045484543, + 0.01190192997455597, + 0.006414173170924187, + 0.013766697607934475, + 0.014044254086911678, + 0.01045932900160551, + 0.006054073106497526, + 0.009020168334245682, + 0.011775173246860504 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 4, + 5 + ], + "selected_reward": -0.02979809045791626 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a sleek, modern black sedan with glossy paint?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the car's headlights bright white LEDs that slice through the twilight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the car moving steadily from left to right across a wide, multi-lane asphalt road?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a dense cityscape of glass and steel skyscrapers with a sunset sky transitioning to deep blue?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera perform a smooth tracking shot matching the car's speed from a slightly low angle?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.14280810952186584 + ], + "opposite_logprob_tokens": [ + -2.017808198928833 + ], + "logprob_reward": -0.14280810952186584, + "yes_no_margin_reward": 1.8750001192092896, + "token_credit_reward": -0.14280810952186584, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.05218925327062607, + 0.06712882965803146, + 0.13831719756126404, + 0.025198860093951225, + 0.10567010939121246, + 0.3181365430355072, + 0.13338930904865265, + 0.027930747717618942, + 0.08172260224819183, + 0.05031653493642807 + ], + "credit_map": [ + 0.003104504896327853, + 0.0041479808278381824, + 0.008819342590868473, + 0.005980064161121845, + 0.003744032233953476, + 0.007106317672878504, + 0.00786526221781969, + 0.011421745643019676, + 0.005315087735652924, + 0.008876057341694832, + 0.010989137925207615, + 0.008214323781430721, + 0.008426861837506294, + 0.005629492457956076, + 0.011280249804258347, + 0.008397621102631092, + 0.010062401182949543, + 0.016736812889575958, + 0.0099715581163764, + 0.017812006175518036, + 0.02024289406836033, + 0.02353447675704956, + 0.025763696059584618, + 0.014193366281688213, + 0.0025849826633930206, + 0.004100065212696791, + 0.002663022605702281, + 0.003539985278621316, + 0.0029851822182536125, + 0.003485097549855709, + 0.0034709256142377853, + 0.0023695991840213537, + 0.007544133812189102, + 0.005648006685078144, + 0.015007217414677143, + 0.028082117438316345, + 0.006672896444797516, + 0.009109076112508774, + 0.01471977774053812, + 0.018886877223849297, + 0.011515732854604721, + 0.017658928409218788, + 0.04493381083011627, + 0.11065591871738434, + 0.008704567328095436, + 0.0064230854623019695, + 0.09632137417793274, + 0.021923141553997993, + 0.006279791705310345, + 0.014201054349541664, + 0.022357158362865448, + 0.025260085240006447, + 0.007742223329842091, + 0.00724714295938611, + 0.028005270287394524, + 0.022296587005257607, + 0.002828923985362053, + 0.001689720549620688, + 0.003916488029062748, + 0.006907952483743429, + 0.0020602752920240164, + 0.002341232495382428, + 0.004753442946821451, + 0.0034327113535255194, + 0.0031064781360328197, + 0.003839082783088088, + 0.01376822404563427, + 0.014102146029472351, + 0.0034782395232468843, + 0.003963169641792774, + 0.01281746942549944, + 0.026647791266441345, + 0.005118436645716429, + 0.0022831852547824383, + 0.01185557246208191, + 0.010609905235469341, + 0.004401370882987976, + 0.006686332169920206, + 0.004993104841560125, + 0.0043686325661838055 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 2, + 6 + ], + "selected_reward": -0.14280810952186584 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a sleek, modern black sedan with glossy paint?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the car's headlights bright white LEDs that slice through the twilight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the car moving steadily from left to right across a wide, multi-lane asphalt road?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a dense cityscape of glass and steel skyscrapers with a sunset sky transitioning to deep blue?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera perform a smooth tracking shot matching the car's speed from a slightly low angle?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005944548640400171 + ], + "opposite_logprob_tokens": [ + -5.130944728851318 + ], + "logprob_reward": -0.005944548640400171, + "yes_no_margin_reward": 5.125, + "token_credit_reward": -0.005944548640400171, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06354460120201111, + 0.051694355905056, + 0.05371188744902611, + 0.04121475666761398, + 0.23616421222686768, + 0.18556024134159088, + 0.07629810273647308, + 0.05342137813568115, + 0.14399755001068115, + 0.09439292550086975 + ], + "credit_map": [ + 0.004292789380997419, + 0.005568527616560459, + 0.013685723766684532, + 0.01088776532560587, + 0.004265078343451023, + 0.004937450401484966, + 0.008218725211918354, + 0.011688536033034325, + 0.006450147833675146, + 0.004577521234750748, + 0.010716392658650875, + 0.006429325323551893, + 0.004957166500389576, + 0.003573135705664754, + 0.008647692389786243, + 0.006342971697449684, + 0.0036727937404066324, + 0.006329830270260572, + 0.004723367281258106, + 0.0054291971027851105, + 0.006397516932338476, + 0.00937613844871521, + 0.011124514043331146, + 0.006658524740487337, + 0.0037920658942312002, + 0.0039030276238918304, + 0.004844828508794308, + 0.009516352787613869, + 0.0027963612228631973, + 0.003997985273599625, + 0.008376233279705048, + 0.003987903241068125, + 0.012836416251957417, + 0.011478757485747337, + 0.028488313779234886, + 0.07910270988941193, + 0.010985889472067356, + 0.015259410254657269, + 0.04499557241797447, + 0.03301713615655899, + 0.006488466635346413, + 0.010843967087566853, + 0.021373184397816658, + 0.06542778760194778, + 0.005245673004537821, + 0.004388387780636549, + 0.057905104011297226, + 0.013887649402022362, + 0.007042437791824341, + 0.009875282645225525, + 0.013890636153519154, + 0.015187212266027927, + 0.004632569849491119, + 0.003249953966587782, + 0.010003047063946724, + 0.012416962534189224, + 0.004001703578978777, + 0.0022941348142921925, + 0.007556550204753876, + 0.008264233358204365, + 0.003961918875575066, + 0.004314329009503126, + 0.008024772629141808, + 0.015003729611635208, + 0.007831929251551628, + 0.01078938040882349, + 0.0222677830606699, + 0.018964232876896858, + 0.011483537964522839, + 0.010521108284592628, + 0.018120160326361656, + 0.044019412249326706, + 0.007751826662570238, + 0.005497232545167208, + 0.020029161125421524, + 0.027163077145814896, + 0.007381008937954903, + 0.0058693294413387775, + 0.012289131060242653, + 0.008412153460085392 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 5, + 8 + ], + "selected_reward": -0.005944548640400171 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "smooth, slightly low-angle tracking shot follows a sleek, modern black sedan, its glossy paint reflecting the ambient light like liquid obsidian. Its bright white LED headlights slice sharply through the deepening twilight as it steadily glides from left to right across a wide, multi-lane asphalt road. The sidewalk is clearly marked with large, red and white directional arrows, their colors vibrant against the dark surface. In the background, a dense cityscape of towering glass and steel skyscrapers stands against a dramatic sky, ablaze with the warm golden glow of sunset, transitioning to a deep blue overhead. A prominent neon \"Dining Club\" sign hums, casting a vibrant pink and blue glow on the street and the sides of passing vehicles. The camera perfectly matches the car's constant speed, creating a fluid and captivating sense of movement, capturing the city's complex, electric atmosphere as it shifts from day to night.", + "seed": 1215, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000095/group_8/index_6733_seed_1215_dp_17_rank_34_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 8, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.034256460051983594, + "vlm_logprob": -0.034256460051983594, + "vlm_yes_no_margin": 4.0, + "vlm_token_credit": -0.034256460051983594, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a sleek, modern black sedan with glossy paint?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the car's headlights bright white LEDs that slice through the twilight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the car moving steadily from left to right across a wide, multi-lane asphalt road?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a dense cityscape of glass and steel skyscrapers with a sunset sky transitioning to deep blue?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera perform a smooth tracking shot matching the car's speed from a slightly low angle?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.00983016099780798 + ], + "opposite_logprob_tokens": [ + -4.884829998016357 + ], + "logprob_reward": -0.00983016099780798, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.00983016099780798, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.04735855013132095, + 0.046747688204050064, + 0.021228449419140816, + 0.055962808430194855, + 0.4014744460582733, + 0.02221287414431572, + 0.045683566480875015, + 0.23125497996807098, + 0.07207552343606949, + 0.05600113049149513 + ], + "credit_map": [ + 0.003191151423379779, + 0.004731164779514074, + 0.006532831117510796, + 0.004912305623292923, + 0.0036669739056378603, + 0.006584070157259703, + 0.00895562395453453, + 0.008784427307546139, + 0.0034379607532173395, + 0.008022048510611057, + 0.0068868533708155155, + 0.007755856029689312, + 0.005351151805371046, + 0.005986087955534458, + 0.005458250641822815, + 0.003849479602649808, + 0.0021114458795636892, + 0.0021621056366711855, + 0.002821350935846567, + 0.0043618581257760525, + 0.0024097340647131205, + 0.0029973099008202553, + 0.002233633305877447, + 0.002131012035533786, + 0.006620488129556179, + 0.006767482962459326, + 0.007040064316242933, + 0.006057887803763151, + 0.0060118199326097965, + 0.006360955536365509, + 0.007618922274559736, + 0.009485186077654362, + 0.03648143261671066, + 0.024341363459825516, + 0.038343969732522964, + 0.05547340214252472, + 0.07710418850183487, + 0.07221724092960358, + 0.05524984747171402, + 0.042263008654117584, + 0.0023867401760071516, + 0.0015134281711652875, + 0.0018636081367731094, + 0.003904316807165742, + 0.003519935067743063, + 0.0009632974397391081, + 0.00585694657638669, + 0.002204601652920246, + 0.0037546022795140743, + 0.004448829218745232, + 0.0064193736761808395, + 0.009674194268882275, + 0.0043141222558915615, + 0.001935783657245338, + 0.009067660197615623, + 0.0060690026730299, + 0.009902486577630043, + 0.009488381445407867, + 0.015349108725786209, + 0.0828314870595932, + 0.010727806948125362, + 0.007890187203884125, + 0.08056811988353729, + 0.014497394673526287, + 0.006345957051962614, + 0.009853585623204708, + 0.010630263946950436, + 0.010752560570836067, + 0.009196366183459759, + 0.008934340439736843, + 0.009769990108907223, + 0.006592456717044115, + 0.011754019185900688, + 0.0036338954232633114, + 0.00520448200404644, + 0.00518124969676137, + 0.005002693738788366, + 0.0071174209006130695, + 0.009655642323195934, + 0.00845172256231308 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 7, + 8 + ], + "selected_reward": -0.00983016099780798 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a sleek, modern black sedan with glossy paint?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the car's headlights bright white LEDs that slice through the twilight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the car moving steadily from left to right across a wide, multi-lane asphalt road?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a dense cityscape of glass and steel skyscrapers with a sunset sky transitioning to deep blue?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera perform a smooth tracking shot matching the car's speed from a slightly low angle?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.11296098679304123 + ], + "opposite_logprob_tokens": [ + -2.2379610538482666 + ], + "logprob_reward": -0.11296098679304123, + "yes_no_margin_reward": 2.125, + "token_credit_reward": -0.11296098679304123, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10201682150363922, + 0.09690756350755692, + 0.08629415184259415, + 0.02954292483627796, + 0.1667860448360443, + 0.019195185974240303, + 0.058745164424180984, + 0.2633345127105713, + 0.11474107205867767, + 0.06243662163615227 + ], + "credit_map": [ + 0.005153723061084747, + 0.013171498663723469, + 0.012645747512578964, + 0.010082870721817017, + 0.009922822937369347, + 0.014197909273207188, + 0.01883004605770111, + 0.01801219955086708, + 0.006106482818722725, + 0.020153814926743507, + 0.011549119837582111, + 0.015430465340614319, + 0.010263259522616863, + 0.013268481940031052, + 0.013022641651332378, + 0.007113297935575247, + 0.005614706315100193, + 0.008132390677928925, + 0.011050336062908173, + 0.013160943984985352, + 0.008297956548631191, + 0.011680097319185734, + 0.01483188010752201, + 0.013525834307074547, + 0.003373687155544758, + 0.003252059919759631, + 0.0035387722309678793, + 0.0035582114942371845, + 0.0031424963381141424, + 0.003289105137810111, + 0.005729289259761572, + 0.0036593046970665455, + 0.016169266775250435, + 0.009432670660316944, + 0.018032794818282127, + 0.022270740941166878, + 0.030786260962486267, + 0.029843924567103386, + 0.022137757390737534, + 0.018112631514668465, + 0.0015330727910622954, + 0.0010968957794830203, + 0.0018663338851183653, + 0.003619229421019554, + 0.004107515327632427, + 0.001845235819928348, + 0.0033887308090925217, + 0.0017381722573190928, + 0.005470470525324345, + 0.008505274541676044, + 0.010097065940499306, + 0.010160233825445175, + 0.005453040357679129, + 0.002980908378958702, + 0.010248011909425259, + 0.005830156616866589, + 0.011753306724131107, + 0.014444021508097649, + 0.013407169841229916, + 0.08528649061918259, + 0.012448279187083244, + 0.015374870970845222, + 0.09075598418712616, + 0.019864380359649658, + 0.009432395920157433, + 0.01585039310157299, + 0.014796639792621136, + 0.01786854676902294, + 0.013779545202851295, + 0.01437951996922493, + 0.016664324328303337, + 0.01196970883756876, + 0.01980532705783844, + 0.005094412248581648, + 0.005781388841569424, + 0.005449729971587658, + 0.004779192619025707, + 0.006291297730058432, + 0.006629651412367821, + 0.00860561989247799 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 4, + 8 + ], + "selected_reward": -0.11296098679304123 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a sleek, modern black sedan with glossy paint?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the car's headlights bright white LEDs that slice through the twilight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the car moving steadily from left to right across a wide, multi-lane asphalt road?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a dense cityscape of glass and steel skyscrapers with a sunset sky transitioning to deep blue?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera perform a smooth tracking shot matching the car's speed from a slightly low angle?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.020573314279317856 + ], + "opposite_logprob_tokens": [ + -3.895573377609253 + ], + "logprob_reward": -0.020573314279317856, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.020573314279317856, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11201599985361099, + 0.0787968710064888, + 0.037964336574077606, + 0.038795821368694305, + 0.1022915169596672, + 0.05134701356291771, + 0.0363486148416996, + 0.2069772332906723, + 0.10687749832868576, + 0.22858507931232452 + ], + "credit_map": [ + 0.006961209699511528, + 0.018643531948328018, + 0.011327128857374191, + 0.011345132254064083, + 0.012893934734165668, + 0.01788058876991272, + 0.018970591947436333, + 0.013993886299431324, + 0.006454020272940397, + 0.01665400341153145, + 0.010141555219888687, + 0.010142459534108639, + 0.010837467387318611, + 0.01006034854799509, + 0.007956461049616337, + 0.00655056769028306, + 0.003109477926045656, + 0.004195434041321278, + 0.004776341840624809, + 0.006642654072493315, + 0.004379748832434416, + 0.005272184498608112, + 0.005586883053183556, + 0.004001616034656763, + 0.004153698682785034, + 0.004830840043723583, + 0.003195006400346756, + 0.005337692331522703, + 0.004098075442016125, + 0.004839123692363501, + 0.0073316581547260284, + 0.00500972755253315, + 0.010190065950155258, + 0.006668969988822937, + 0.00973602943122387, + 0.014008303172886372, + 0.019512690603733063, + 0.0189431793987751, + 0.013402636162936687, + 0.009829651564359665, + 0.008440625853836536, + 0.0053570205345749855, + 0.006025178357958794, + 0.011262526735663414, + 0.00500254612416029, + 0.00238512153737247, + 0.0076921540312469006, + 0.005181844346225262, + 0.00310778827406466, + 0.0022195312194526196, + 0.00752108683809638, + 0.006145177409052849, + 0.002827004762366414, + 0.0018798054661601782, + 0.0071905809454619884, + 0.005457642022520304, + 0.010327870026230812, + 0.01040845550596714, + 0.011179591529071331, + 0.07057443261146545, + 0.010060987435281277, + 0.011270683258771896, + 0.06790296733379364, + 0.015252260491251945, + 0.009038329124450684, + 0.012374543584883213, + 0.010774598456919193, + 0.01802937313914299, + 0.012813500128686428, + 0.014156973920762539, + 0.018523776903748512, + 0.011166415177285671, + 0.020541993901133537, + 0.01736695133149624, + 0.02977527119219303, + 0.017669586464762688, + 0.03947991132736206, + 0.0398329421877861, + 0.026886478066444397, + 0.03703194856643677 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 7, + 0 + ], + "selected_reward": -0.020573314279317856 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a sleek, modern black sedan with glossy paint?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the car's headlights bright white LEDs that slice through the twilight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the car moving steadily from left to right across a wide, multi-lane asphalt road?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a dense cityscape of glass and steel skyscrapers with a sunset sky transitioning to deep blue?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera perform a smooth tracking shot matching the car's speed from a slightly low angle?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02328609488904476 + ], + "opposite_logprob_tokens": [ + -3.7732861042022705 + ], + "logprob_reward": -0.02328609488904476, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.02328609488904476, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.05974945053458214, + 0.023434903472661972, + 0.019640469923615456, + 0.10296856611967087, + 0.15534722805023193, + 0.03545861318707466, + 0.024493684992194176, + 0.453196257352829, + 0.0653131976723671, + 0.060397543013095856 + ], + "credit_map": [ + 0.003894479013979435, + 0.006087168585509062, + 0.011061009019613266, + 0.006227509118616581, + 0.005351841449737549, + 0.008587825112044811, + 0.010064829140901566, + 0.008474791422486305, + 0.0017056287033483386, + 0.003242294769734144, + 0.004145815037190914, + 0.002645672531798482, + 0.00334947369992733, + 0.0027701363433152437, + 0.003162626875564456, + 0.0024132556281983852, + 0.0016807392239570618, + 0.002740331459790468, + 0.0035223793238401413, + 0.002286787610501051, + 0.002247918862849474, + 0.0023775752633810043, + 0.0029306174255907536, + 0.0018541206372901797, + 0.0112574752420187, + 0.015765687450766563, + 0.010668371804058552, + 0.014650983735918999, + 0.014102521352469921, + 0.012581581249833107, + 0.013284449465572834, + 0.010657493025064468, + 0.01403587032109499, + 0.008738625794649124, + 0.015378592535853386, + 0.019486937671899796, + 0.033212460577487946, + 0.02940240688621998, + 0.019608844071626663, + 0.015483487397432327, + 0.004426725674420595, + 0.0028200659435242414, + 0.004703449085354805, + 0.00667569600045681, + 0.004661147017031908, + 0.002108278451487422, + 0.0064152865670621395, + 0.003647966543212533, + 0.0022649087477475405, + 0.0032614776864647865, + 0.003267835360020399, + 0.004252211190760136, + 0.0027369381859898567, + 0.0010149246081709862, + 0.004768955055624247, + 0.002926434623077512, + 0.02028946951031685, + 0.01915491372346878, + 0.01988142728805542, + 0.16539530456066132, + 0.019525738433003426, + 0.019011063501238823, + 0.16129520535469055, + 0.028643159195780754, + 0.005608346778899431, + 0.008441200479865074, + 0.0076277367770671844, + 0.009064032696187496, + 0.00843946821987629, + 0.009585336782038212, + 0.008770034648478031, + 0.007777044549584389, + 0.012374729849398136, + 0.004853708669543266, + 0.0071291676722466946, + 0.005099677946418524, + 0.007133325096219778, + 0.008619730360805988, + 0.005918543320149183, + 0.00926866102963686 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 4, + 3 + ], + "selected_reward": -0.02328609488904476 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a sleek, modern black sedan with glossy paint?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the car's headlights bright white LEDs that slice through the twilight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the car moving steadily from left to right across a wide, multi-lane asphalt road?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a dense cityscape of glass and steel skyscrapers with a sunset sky transitioning to deep blue?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera perform a smooth tracking shot matching the car's speed from a slightly low angle?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004631743300706148 + ], + "opposite_logprob_tokens": [ + -5.379631519317627 + ], + "logprob_reward": -0.004631743300706148, + "yes_no_margin_reward": 5.375, + "token_credit_reward": -0.004631743300706148, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.20239947736263275, + 0.13770826160907745, + 0.04941333457827568, + 0.06178392842411995, + 0.03584738448262215, + 0.04740313068032265, + 0.04766000807285309, + 0.18075738847255707, + 0.1336544156074524, + 0.10337270051240921 + ], + "credit_map": [ + 0.009430097416043282, + 0.03923660144209862, + 0.01978127285838127, + 0.026406707242131233, + 0.01979793980717659, + 0.027907926589250565, + 0.0361555814743042, + 0.02368333376944065, + 0.010742010548710823, + 0.030009938403964043, + 0.018103573471307755, + 0.0211926456540823, + 0.01605500839650631, + 0.01779823936522007, + 0.013371993787586689, + 0.010434849187731743, + 0.0033696703612804413, + 0.007020335644483566, + 0.006967446301132441, + 0.010045159608125687, + 0.006723836064338684, + 0.006429136265069246, + 0.004737324547022581, + 0.0041204229928553104, + 0.008022657595574856, + 0.005578896496444941, + 0.00784036424010992, + 0.006794586777687073, + 0.006686261855065823, + 0.007761645596474409, + 0.009515335783362389, + 0.009584182873368263, + 0.0025718887336552143, + 0.0024706434924155474, + 0.005242826882749796, + 0.007134142331779003, + 0.006667913869023323, + 0.0033199265599250793, + 0.005275465082377195, + 0.0031645765993744135, + 0.0063398415222764015, + 0.004466683603823185, + 0.0059807682409882545, + 0.009939812123775482, + 0.004644994158297777, + 0.002336209174245596, + 0.008658143691718578, + 0.005036677233874798, + 0.00410508643835783, + 0.0046729715541005135, + 0.006357984151691198, + 0.008810732513666153, + 0.003863999852910638, + 0.0030942377634346485, + 0.00819099135696888, + 0.008564007468521595, + 0.013329106383025646, + 0.018321391195058823, + 0.027120916172862053, + 0.024063512682914734, + 0.0146793806925416, + 0.026089515537023544, + 0.033944662660360336, + 0.02320891246199608, + 0.008737219497561455, + 0.017024008557200432, + 0.017452165484428406, + 0.02447742223739624, + 0.012738510966300964, + 0.013649172149598598, + 0.027404295280575752, + 0.012171615846455097, + 0.020128881558775902, + 0.007878038100898266, + 0.011720230802893639, + 0.012188524939119816, + 0.010997551493346691, + 0.016364721581339836, + 0.011745886877179146, + 0.012348865158855915 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 7, + 1 + ], + "selected_reward": -0.004631743300706148 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000095/group_9/index_tensor([3384, 3384])_group_9_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000095/group_9/index_tensor([3384, 3384])_group_9_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..e0a30104090511edffa443264373e3d43fd2ec38 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000095/group_9/index_tensor([3384, 3384])_group_9_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "dynamic tracking shot, shot from a slightly upward angle, pans smoothly to the left, capturing a playful aerial chase between two large-eyed mosquitoes, rendered in a high-quality, vibrant anime style. The leading mosquito, with its oversized, gleaming emerald eyes and a hint of slyness, deftly executes a playful aerial maneuver. Following closely behind is its companion, also sporting enormous sapphire eyes and a playful yet determined expression, tilting its round, cartoonish body through a sharp turn. Their delicate, translucent wings flap rapidly, shimmering in the light with a soap-bubble-like iridescence, while their slender, segmented limbs trace graceful arcs through the air. The chase takes place deep within a sun-drenched enchanted garden, against a backdrop of giant, dew-drenched clover leaves and vibrantly colored giant wildflowers, creating a soft bokeh effect. A beam of warm golden afternoon sunlight streams in from just outside the frame, creating a soft, ethereal lens flare that sweeps across the screen in orange and gold hues, accentuating their flight path while adding a dreamlike quality to this vibrant, highly saturated color palette, creating an atmosphere of lightheartedness, wonder, and joyful energy.", + "seed": 31291, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000095/group_9/index_3384_seed_31291_dp_18_rank_36_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 9, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.07481988272629678, + "vlm_logprob": -0.07481988272629678, + "vlm_yes_no_margin": 4.1, + "vlm_token_credit": -0.07481988272629678, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are two mosquitoes with large eyes visible in the video?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the leading mosquito's eyes emerald green and the following mosquito's eyes sapphire blue?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the mosquitoes' wings translucent with a shimmering, soap-bubble-like iridescence?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the setting a sun-drenched enchanted garden featuring giant dew-drenched clover leaves and wildflowers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera pan smoothly to the left while capturing the chase?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.010705304332077503 + ], + "opposite_logprob_tokens": [ + -4.635705471038818 + ], + "logprob_reward": -0.010705304332077503, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.010705304332077503, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13302002847194672, + 0.07429829239845276, + 0.32692575454711914, + 0.0761750265955925, + 0.11001303046941757, + 0.018105555325746536, + 0.03256146237254143, + 0.11972218006849289, + 0.04757371544837952, + 0.06160493567585945 + ], + "credit_map": [ + 0.009199644438922405, + 0.014773267321288586, + 0.017757736146450043, + 0.01838505081832409, + 0.021709421649575233, + 0.017439251765608788, + 0.015300123952329159, + 0.01845553331077099, + 0.004740048665553331, + 0.006913500372320414, + 0.010695785284042358, + 0.009052618406713009, + 0.012417228892445564, + 0.01143887173384428, + 0.01102363970130682, + 0.008016597479581833, + 0.027128268033266068, + 0.03411839157342911, + 0.05677876994013786, + 0.028551841154694557, + 0.057027604430913925, + 0.03718322888016701, + 0.05271342396736145, + 0.033424217253923416, + 0.009022897109389305, + 0.0077437665313482285, + 0.00889058317989111, + 0.008539686910808086, + 0.008550687693059444, + 0.011735362000763416, + 0.007461765315383673, + 0.01423027366399765, + 0.009644703008234501, + 0.010031833313405514, + 0.019929012283682823, + 0.015186267904937267, + 0.009230700321495533, + 0.011145349591970444, + 0.01583021692931652, + 0.01901494525372982, + 0.002075095195323229, + 0.0023076203651726246, + 0.0026022023521363735, + 0.0023046962451189756, + 0.0024461885914206505, + 0.002568375552073121, + 0.0020810000132769346, + 0.0017203772440552711, + 0.003970304038375616, + 0.004090326838195324, + 0.003993528429418802, + 0.004926444962620735, + 0.0034625455737113953, + 0.003830155124887824, + 0.004308194387704134, + 0.003979961387813091, + 0.01074173953384161, + 0.011400497518479824, + 0.0189443938434124, + 0.02647755667567253, + 0.013321219943463802, + 0.012219319120049477, + 0.015762487426400185, + 0.010854964144527912, + 0.004791724029928446, + 0.003779304213821888, + 0.008735865354537964, + 0.0048011657781898975, + 0.004314177669584751, + 0.004375701304525137, + 0.01322641596198082, + 0.0035493599716573954, + 0.004726665560156107, + 0.0041158865205943584, + 0.007156376261264086, + 0.007850762456655502, + 0.011377683840692043, + 0.007569580338895321, + 0.008556774817407131, + 0.010251211002469063 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 0, + 7 + ], + "selected_reward": -0.010705304332077503 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are two mosquitoes with large eyes visible in the video?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the leading mosquito's eyes emerald green and the following mosquito's eyes sapphire blue?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the mosquitoes' wings translucent with a shimmering, soap-bubble-like iridescence?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the setting a sun-drenched enchanted garden featuring giant dew-drenched clover leaves and wildflowers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera pan smoothly to the left while capturing the chase?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.17986764013767242 + ], + "opposite_logprob_tokens": [ + -1.8048676252365112 + ], + "logprob_reward": -0.17986764013767242, + "yes_no_margin_reward": 1.625, + "token_credit_reward": -0.17986764013767242, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13827089965343475, + 0.08802644908428192, + 0.16972589492797852, + 0.07161212712526321, + 0.04223393276333809, + 0.049949150532484055, + 0.07781830430030823, + 0.1425018459558487, + 0.1604284942150116, + 0.05943292751908302 + ], + "credit_map": [ + 0.00904112309217453, + 0.020982826128602028, + 0.019933069124817848, + 0.02447054535150528, + 0.013297191821038723, + 0.014059081673622131, + 0.021088585257530212, + 0.015398475341498852, + 0.005813835188746452, + 0.00796138308942318, + 0.01338909287005663, + 0.010767487809062004, + 0.012313945218920708, + 0.016625460237264633, + 0.011192095465958118, + 0.009963151067495346, + 0.009725864045321941, + 0.023189213126897812, + 0.015692751854658127, + 0.0216378103941679, + 0.02487380988895893, + 0.027178768068552017, + 0.027566049247980118, + 0.01986163668334484, + 0.00987713411450386, + 0.008219982497394085, + 0.00945267453789711, + 0.010005081072449684, + 0.008130705915391445, + 0.007298195734620094, + 0.008781440556049347, + 0.009846911765635014, + 0.0038714329712092876, + 0.002518356777727604, + 0.011118561029434204, + 0.00377279962413013, + 0.0031969361007213593, + 0.004394488409161568, + 0.004419128410518169, + 0.008942228741943836, + 0.002897870959714055, + 0.004744919016957283, + 0.008427049964666367, + 0.008032460696995258, + 0.0052334307692945, + 0.005579287186264992, + 0.008037074469029903, + 0.006997060962021351, + 0.006906215567141771, + 0.008030232042074203, + 0.009105078876018524, + 0.01064058393239975, + 0.007442202419042587, + 0.015212769620120525, + 0.009253181517124176, + 0.011228037066757679, + 0.013154939748346806, + 0.012464386411011219, + 0.020939474925398827, + 0.03500264883041382, + 0.01447649672627449, + 0.015056385658681393, + 0.01808433048427105, + 0.01332318875938654, + 0.01258518360555172, + 0.009217076934874058, + 0.03630361333489418, + 0.013104991987347603, + 0.013065950945019722, + 0.017992762848734856, + 0.049282968044281006, + 0.008875939063727856, + 0.007666362449526787, + 0.0072470251470804214, + 0.010476363822817802, + 0.0077111320570111275, + 0.00809900276362896, + 0.0071343169547617435, + 0.005589366424828768, + 0.005509357433766127 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 8, + 7 + ], + "selected_reward": -0.17986764013767242 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are two mosquitoes with large eyes visible in the video?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the leading mosquito's eyes emerald green and the following mosquito's eyes sapphire blue?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the mosquitoes' wings translucent with a shimmering, soap-bubble-like iridescence?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the setting a sun-drenched enchanted garden featuring giant dew-drenched clover leaves and wildflowers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera pan smoothly to the left while capturing the chase?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0021949741058051586 + ], + "opposite_logprob_tokens": [ + -6.127194881439209 + ], + "logprob_reward": -0.0021949741058051586, + "yes_no_margin_reward": 6.125, + "token_credit_reward": -0.0021949741058051586, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12412790209054947, + 0.06906739622354507, + 0.2581774592399597, + 0.045718964189291, + 0.10184166580438614, + 0.03516440466046333, + 0.048711832612752914, + 0.12589547038078308, + 0.09702232480049133, + 0.09427256882190704 + ], + "credit_map": [ + 0.00817603524774313, + 0.01007797010242939, + 0.02313450537621975, + 0.018599271774291992, + 0.019690774381160736, + 0.014361169189214706, + 0.01582484319806099, + 0.01426334585994482, + 0.006428422871977091, + 0.009329459629952908, + 0.01077800989151001, + 0.010856312699615955, + 0.009032905101776123, + 0.0073982407338917255, + 0.008434947580099106, + 0.006809103302657604, + 0.024857427924871445, + 0.026790058240294456, + 0.033462487161159515, + 0.015181242488324642, + 0.028605448082089424, + 0.029002800583839417, + 0.035011857748031616, + 0.06526614725589752, + 0.0062821609899401665, + 0.006079056765884161, + 0.008382802829146385, + 0.006688186898827553, + 0.004511401057243347, + 0.0054557062685489655, + 0.0037756264209747314, + 0.0045440285466611385, + 0.012073379941284657, + 0.012354445643723011, + 0.013979116454720497, + 0.014400436542928219, + 0.010797499679028988, + 0.011406184174120426, + 0.0140165314078331, + 0.012814073823392391, + 0.002707058796659112, + 0.0028365086764097214, + 0.005945297889411449, + 0.007281562779098749, + 0.003992127720266581, + 0.0030800311360508204, + 0.005119240842759609, + 0.004202582407742739, + 0.004257197491824627, + 0.00504012405872345, + 0.00547269731760025, + 0.007901350036263466, + 0.005015567410737276, + 0.008505092933773994, + 0.006928814575076103, + 0.005590988788753748, + 0.012323579750955105, + 0.011829433962702751, + 0.024262098595499992, + 0.021647021174430847, + 0.013499177992343903, + 0.012283701449632645, + 0.018440542742609978, + 0.011609911918640137, + 0.008622020483016968, + 0.010326728224754333, + 0.011547829024493694, + 0.012740260921418667, + 0.015342583879828453, + 0.013651787303388119, + 0.014837122522294521, + 0.009953996166586876, + 0.005717117339372635, + 0.00543992081657052, + 0.00929062720388174, + 0.01297046523541212, + 0.013943471945822239, + 0.00977967120707035, + 0.02742825821042061, + 0.009703034535050392 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 7, + 0 + ], + "selected_reward": -0.0021949741058051586 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are two mosquitoes with large eyes visible in the video?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the leading mosquito's eyes emerald green and the following mosquito's eyes sapphire blue?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the mosquitoes' wings translucent with a shimmering, soap-bubble-like iridescence?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the setting a sun-drenched enchanted garden featuring giant dew-drenched clover leaves and wildflowers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera pan smoothly to the left while capturing the chase?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.001508170971646905 + ], + "opposite_logprob_tokens": [ + -6.5015082359313965 + ], + "logprob_reward": -0.001508170971646905, + "yes_no_margin_reward": 6.5, + "token_credit_reward": -0.001508170971646905, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14407283067703247, + 0.03906267508864403, + 0.12759749591350555, + 0.08116789162158966, + 0.2032279372215271, + 0.031252697110176086, + 0.1366606056690216, + 0.11896952241659164, + 0.06279424577951431, + 0.055194076150655746 + ], + "credit_map": [ + 0.011230559088289738, + 0.014318643137812614, + 0.019860144704580307, + 0.022428805008530617, + 0.018128667026758194, + 0.015262933447957039, + 0.026713771745562553, + 0.016129322350025177, + 0.002631863346323371, + 0.0032654735259711742, + 0.004822878632694483, + 0.004316430073231459, + 0.006642134394496679, + 0.0074877655133605, + 0.005296764895319939, + 0.00459936261177063, + 0.011821147054433823, + 0.014025552198290825, + 0.010134352371096611, + 0.010648577474057674, + 0.015306036919355392, + 0.018297577276825905, + 0.023579787462949753, + 0.02378445863723755, + 0.011597821488976479, + 0.01008276641368866, + 0.018545813858509064, + 0.010345743037760258, + 0.005998663604259491, + 0.011168872006237507, + 0.005947460420429707, + 0.007480749394744635, + 0.024524956941604614, + 0.020933395251631737, + 0.03456638380885124, + 0.026467828080058098, + 0.019185680896043777, + 0.019165657460689545, + 0.027191724628210068, + 0.03119230829179287, + 0.003728642361238599, + 0.0030757638160139322, + 0.0049676005728542805, + 0.0054563432931900024, + 0.003529681358486414, + 0.0028067666571587324, + 0.00489550456404686, + 0.002792394021525979, + 0.01187159400433302, + 0.012271415442228317, + 0.018369406461715698, + 0.019112372770905495, + 0.012644086964428425, + 0.0302504301071167, + 0.01391227450221777, + 0.018229028210043907, + 0.010826935060322285, + 0.011061707511544228, + 0.01937444880604744, + 0.030056903138756752, + 0.012620986439287663, + 0.012473981827497482, + 0.013296232558786869, + 0.009258322417736053, + 0.006095780059695244, + 0.006601282395422459, + 0.007799748796969652, + 0.007832061499357224, + 0.011257251724600792, + 0.006793631706386805, + 0.00930012110620737, + 0.007114369887858629, + 0.005533016752451658, + 0.004735449329018593, + 0.006845556665211916, + 0.007384739350527525, + 0.006459063850343227, + 0.006083407439291477, + 0.012894469313323498, + 0.0052583725191652775 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 0, + 6 + ], + "selected_reward": -0.001508170971646905 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are two mosquitoes with large eyes visible in the video?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the leading mosquito's eyes emerald green and the following mosquito's eyes sapphire blue?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the mosquitoes' wings translucent with a shimmering, soap-bubble-like iridescence?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the setting a sun-drenched enchanted garden featuring giant dew-drenched clover leaves and wildflowers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera pan smoothly to the left while capturing the chase?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.17982332408428192 + ], + "opposite_logprob_tokens": [ + -1.8048232793807983 + ], + "logprob_reward": -0.17982332408428192, + "yes_no_margin_reward": 1.625, + "token_credit_reward": -0.17982332408428192, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08266832679510117, + 0.06168453395366669, + 0.18539589643478394, + 0.08208399266004562, + 0.06640391051769257, + 0.03810328617691994, + 0.03154699131846428, + 0.1317622810602188, + 0.2651362121105194, + 0.055214643478393555 + ], + "credit_map": [ + 0.00784401036798954, + 0.01217497419565916, + 0.010202351957559586, + 0.013296854682266712, + 0.011965697631239891, + 0.008714926429092884, + 0.010778055526316166, + 0.007691455539315939, + 0.006312092300504446, + 0.004827371332794428, + 0.00902693159878254, + 0.0068302941508591175, + 0.008785128593444824, + 0.009611032903194427, + 0.008627277798950672, + 0.007664399687200785, + 0.018061038106679916, + 0.018955018371343613, + 0.016585683450102806, + 0.01443510502576828, + 0.018412237986922264, + 0.01921779289841652, + 0.033644288778305054, + 0.04608473181724548, + 0.008405271917581558, + 0.009966959245502949, + 0.008148999884724617, + 0.009924700483679771, + 0.011489098891615868, + 0.011201733723282814, + 0.0104907201603055, + 0.012456509284675121, + 0.009642597287893295, + 0.009243817999958992, + 0.009422977454960346, + 0.009125297889113426, + 0.006832174025475979, + 0.006426473148167133, + 0.008285065181553364, + 0.007425508927553892, + 0.0055184816010296345, + 0.0033825435675680637, + 0.006233883555978537, + 0.007065913639962673, + 0.0032628763001412153, + 0.0031646054703742266, + 0.004676015581935644, + 0.004798965994268656, + 0.00289576081559062, + 0.0052046217024326324, + 0.004478283226490021, + 0.005320369731634855, + 0.0033107607159763575, + 0.004211060237139463, + 0.003235090058296919, + 0.002891044830903411, + 0.010234437882900238, + 0.01302239391952753, + 0.01609310507774353, + 0.018061935901641846, + 0.015451307408511639, + 0.018918544054031372, + 0.021917670965194702, + 0.018062885850667953, + 0.025067895650863647, + 0.024800047278404236, + 0.036634013056755066, + 0.033174511045217514, + 0.040766697376966476, + 0.02568408101797104, + 0.052937012165784836, + 0.026071952655911446, + 0.006561779882758856, + 0.0037912081461399794, + 0.006094909738749266, + 0.007931471802294254, + 0.0056455861777067184, + 0.005992588587105274, + 0.008216939866542816, + 0.010980159044265747 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 2, + 7 + ], + "selected_reward": -0.17982332408428192 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "dynamic tracking shot, shot from a slightly upward angle, pans smoothly to the left, capturing a playful aerial chase between two large-eyed mosquitoes, rendered in a high-quality, vibrant anime style. The leading mosquito, with its oversized, gleaming emerald eyes and a hint of slyness, deftly executes a playful aerial maneuver. Following closely behind is its companion, also sporting enormous sapphire eyes and a playful yet determined expression, tilting its round, cartoonish body through a sharp turn. Their delicate, translucent wings flap rapidly, shimmering in the light with a soap-bubble-like iridescence, while their slender, segmented limbs trace graceful arcs through the air. The chase takes place deep within a sun-drenched enchanted garden, against a backdrop of giant, dew-drenched clover leaves and vibrantly colored giant wildflowers, creating a soft bokeh effect. A beam of warm golden afternoon sunlight streams in from just outside the frame, creating a soft, ethereal lens flare that sweeps across the screen in orange and gold hues, accentuating their flight path while adding a dreamlike quality to this vibrant, highly saturated color palette, creating an atmosphere of lightheartedness, wonder, and joyful energy.", + "seed": 31292, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000095/group_9/index_3384_seed_31292_dp_18_rank_36_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 9, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.07678843634203077, + "vlm_logprob": -0.07678843634203077, + "vlm_yes_no_margin": 3.525, + "vlm_token_credit": -0.07678843634203077, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are two mosquitoes with large eyes visible in the video?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the leading mosquito's eyes emerald green and the following mosquito's eyes sapphire blue?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the mosquitoes' wings translucent with a shimmering, soap-bubble-like iridescence?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the setting a sun-drenched enchanted garden featuring giant dew-drenched clover leaves and wildflowers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera pan smoothly to the left while capturing the chase?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.035036053508520126 + ], + "opposite_logprob_tokens": [ + -3.410036087036133 + ], + "logprob_reward": -0.035036053508520126, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.035036053508520126, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1932094544172287, + 0.1210784763097763, + 0.1899661123752594, + 0.16648295521736145, + 0.03882464021444321, + 0.06354480981826782, + 0.11668064445257187, + 0.044287554919719696, + 0.02145520970225334, + 0.0444701686501503 + ], + "credit_map": [ + 0.018039369955658913, + 0.013972168788313866, + 0.052483245730400085, + 0.01934026926755905, + 0.017647897824645042, + 0.014278335496783257, + 0.015699464827775955, + 0.04174870252609253, + 0.009658410213887691, + 0.009047477506101131, + 0.02096417360007763, + 0.014533594250679016, + 0.011938415467739105, + 0.014215442351996899, + 0.013695981353521347, + 0.027024976909160614, + 0.01103675365447998, + 0.008198224008083344, + 0.03239034488797188, + 0.024606844410300255, + 0.011281493119895458, + 0.0070111192762851715, + 0.0771971046924591, + 0.01824423484504223, + 0.011767476797103882, + 0.022071925923228264, + 0.019379520788788795, + 0.02131401002407074, + 0.01733531802892685, + 0.016413452103734016, + 0.03826073929667473, + 0.01994052156805992, + 0.003720025299116969, + 0.003551853820681572, + 0.006295137107372284, + 0.005942527204751968, + 0.003183256136253476, + 0.005846587475389242, + 0.005315979942679405, + 0.0049692741595208645, + 0.009566727094352245, + 0.00679450947791338, + 0.006240216549485922, + 0.007704362738877535, + 0.0062764789909124374, + 0.008675310760736465, + 0.006137197371572256, + 0.012150007300078869, + 0.006563782691955566, + 0.0022111511789262295, + 0.005195440258830786, + 0.007573116570711136, + 0.005327593069523573, + 0.03505861759185791, + 0.041107095777988434, + 0.013643844984471798, + 0.0031687654554843903, + 0.0047809355892241, + 0.005720025394111872, + 0.005631930194795132, + 0.003923436626791954, + 0.004661327227950096, + 0.008039579726755619, + 0.008361553773283958, + 0.002804960822686553, + 0.0017402172088623047, + 0.00308893877081573, + 0.002408777829259634, + 0.0025830070953816175, + 0.0030650622211396694, + 0.003295858623459935, + 0.0024683885276317596, + 0.0037187945563346148, + 0.002790349069982767, + 0.007004548795521259, + 0.012960233725607395, + 0.005258191376924515, + 0.0024922736920416355, + 0.005864822305738926, + 0.004380953963845968 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 3 + ], + "selected_reward": -0.035036053508520126 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are two mosquitoes with large eyes visible in the video?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the leading mosquito's eyes emerald green and the following mosquito's eyes sapphire blue?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the mosquitoes' wings translucent with a shimmering, soap-bubble-like iridescence?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the setting a sun-drenched enchanted garden featuring giant dew-drenched clover leaves and wildflowers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera pan smoothly to the left while capturing the chase?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.16035683453083038 + ], + "opposite_logprob_tokens": [ + -1.910356879234314 + ], + "logprob_reward": -0.16035683453083038, + "yes_no_margin_reward": 1.75, + "token_credit_reward": -0.16035683453083038, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11220528930425644, + 0.5019462704658508, + 0.04626075178384781, + 0.06834247708320618, + 0.06800853461027145, + 0.02255968004465103, + 0.0979681983590126, + 0.015453989617526531, + 0.016427507624030113, + 0.0508272685110569 + ], + "credit_map": [ + 0.011374800466001034, + 0.011559780687093735, + 0.013529515825212002, + 0.017176415771245956, + 0.012089562602341175, + 0.010916869156062603, + 0.018389753997325897, + 0.017168598249554634, + 0.04669474810361862, + 0.04454205557703972, + 0.07701631635427475, + 0.06099624186754227, + 0.045717351138591766, + 0.05079270899295807, + 0.06943953037261963, + 0.10674735903739929, + 0.003895900910720229, + 0.003470436669886112, + 0.005233439616858959, + 0.005954409018158913, + 0.005072236526757479, + 0.003218672703951597, + 0.014634068123996258, + 0.004781590308994055, + 0.0051969485357403755, + 0.010174566879868507, + 0.007292683701962233, + 0.0087891835719347, + 0.007490586023777723, + 0.007850104942917824, + 0.012942294590175152, + 0.008606118150055408, + 0.006916241254657507, + 0.008543785661458969, + 0.00886949710547924, + 0.009513910859823227, + 0.0069703878834843636, + 0.00946664810180664, + 0.008592911995947361, + 0.009135160595178604, + 0.002362340223044157, + 0.002088252454996109, + 0.0018408403266221285, + 0.0037450287491083145, + 0.0036397872027009726, + 0.0022293308284133673, + 0.0018104129703715444, + 0.004843689501285553, + 0.011122453026473522, + 0.00976325199007988, + 0.007400228176265955, + 0.011934989131987095, + 0.008848953060805798, + 0.020589495077729225, + 0.018782800063490868, + 0.009526043199002743, + 0.001295548863708973, + 0.002692741109058261, + 0.0022734093945473433, + 0.001795566757209599, + 0.0023028645664453506, + 0.0017525121802464128, + 0.0017313354182988405, + 0.001610012841410935, + 0.0018015280365943909, + 0.003011238994076848, + 0.0023022822570055723, + 0.0020031589083373547, + 0.00139388267416507, + 0.0020825322717428207, + 0.001961162779480219, + 0.001871724147349596, + 0.004549170378595591, + 0.0031993333250284195, + 0.007573537528514862, + 0.012971525080502033, + 0.007461769040673971, + 0.0036639028694480658, + 0.006705010775476694, + 0.004703023005276918 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 6 + ], + "selected_reward": -0.16035683453083038 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are two mosquitoes with large eyes visible in the video?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the leading mosquito's eyes emerald green and the following mosquito's eyes sapphire blue?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the mosquitoes' wings translucent with a shimmering, soap-bubble-like iridescence?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the setting a sun-drenched enchanted garden featuring giant dew-drenched clover leaves and wildflowers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera pan smoothly to the left while capturing the chase?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004092295654118061 + ], + "opposite_logprob_tokens": [ + -5.504092216491699 + ], + "logprob_reward": -0.004092295654118061, + "yes_no_margin_reward": 5.5, + "token_credit_reward": -0.004092295654118061, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.05191916227340698, + 0.047385185956954956, + 0.09809672832489014, + 0.06657041609287262, + 0.027714012190699577, + 0.058903589844703674, + 0.5855715870857239, + 0.013595513999462128, + 0.021417830139398575, + 0.028826028108596802 + ], + "credit_map": [ + 0.005054001230746508, + 0.0036244301591068506, + 0.012860276736319065, + 0.0036368542350828648, + 0.004095973912626505, + 0.0056405202485620975, + 0.006236996967345476, + 0.010770109482109547, + 0.004696783144026995, + 0.0058662546798586845, + 0.005541772115975618, + 0.004791276529431343, + 0.004662389401346445, + 0.004811835940927267, + 0.007412154693156481, + 0.00960271805524826, + 0.00580652104690671, + 0.006762677803635597, + 0.01437661424279213, + 0.012007143348455429, + 0.006668328307569027, + 0.004717675503343344, + 0.037538256496191025, + 0.010219515301287174, + 0.005313006695359945, + 0.005873451475054026, + 0.010435315780341625, + 0.011259187012910843, + 0.005502752494066954, + 0.005295255221426487, + 0.015087708830833435, + 0.007803736720234156, + 0.0024389817845076323, + 0.002104891696944833, + 0.0034919213503599167, + 0.003404878079891205, + 0.0025236960500478745, + 0.006304721347987652, + 0.003593417117372155, + 0.0038515047635883093, + 0.009238844737410545, + 0.006803798954933882, + 0.0057960739359259605, + 0.006839272100478411, + 0.005827358458191156, + 0.007563692517578602, + 0.005997026339173317, + 0.01083751954138279, + 0.03642496466636658, + 0.0115381870418787, + 0.020551862195134163, + 0.042916979640722275, + 0.031521305441856384, + 0.21093711256980896, + 0.18673257529735565, + 0.04494860768318176, + 0.001038872404024005, + 0.0012161118211224675, + 0.0018485730979591608, + 0.0014786639949306846, + 0.001042598276399076, + 0.0015046955086290836, + 0.0023994974326342344, + 0.0030665008816868067, + 0.0030292579904198647, + 0.0033058493863791227, + 0.0030643886420875788, + 0.0022472254931926727, + 0.001763037871569395, + 0.002505829557776451, + 0.0025047885719686747, + 0.0029974535573273897, + 0.0018351689213886857, + 0.001971143065020442, + 0.00429946044459939, + 0.008211820386350155, + 0.0033234793227165937, + 0.0025131418369710445, + 0.004197183530777693, + 0.0024746281560510397 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 2, + 3 + ], + "selected_reward": -0.004092295654118061 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are two mosquitoes with large eyes visible in the video?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the leading mosquito's eyes emerald green and the following mosquito's eyes sapphire blue?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the mosquitoes' wings translucent with a shimmering, soap-bubble-like iridescence?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the setting a sun-drenched enchanted garden featuring giant dew-drenched clover leaves and wildflowers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera pan smoothly to the left while capturing the chase?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004633167292922735 + ], + "opposite_logprob_tokens": [ + -5.379632949829102 + ], + "logprob_reward": -0.004633167292922735, + "yes_no_margin_reward": 5.375, + "token_credit_reward": -0.004633167292922735, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0680137351155281, + 0.12425766885280609, + 0.08661101758480072, + 0.07752387225627899, + 0.09374816715717316, + 0.07511533051729202, + 0.35130131244659424, + 0.041893839836120605, + 0.04085077717900276, + 0.04068431630730629 + ], + "credit_map": [ + 0.005486053414642811, + 0.006398117169737816, + 0.018121229484677315, + 0.006423760671168566, + 0.007872367277741432, + 0.005362673196941614, + 0.005787150003015995, + 0.012562371790409088, + 0.013782786205410957, + 0.009146643802523613, + 0.01604011468589306, + 0.01641208492219448, + 0.007981576956808567, + 0.014966844581067562, + 0.021957678720355034, + 0.023969925940036774, + 0.004213991109281778, + 0.003929278813302517, + 0.01918746344745159, + 0.009684128686785698, + 0.005668486002832651, + 0.0035111585166305304, + 0.033130448311567307, + 0.007286056876182556, + 0.006669557187706232, + 0.008823325857520103, + 0.015102910809218884, + 0.011472105048596859, + 0.00702657038345933, + 0.009928056970238686, + 0.010845636948943138, + 0.007655702065676451, + 0.009061713702976704, + 0.007425766438245773, + 0.015075432136654854, + 0.01470303162932396, + 0.007416136562824249, + 0.01609679125249386, + 0.01147598959505558, + 0.012493294663727283, + 0.01563296467065811, + 0.006508800201117992, + 0.006870252080261707, + 0.01081029698252678, + 0.004792815074324608, + 0.005172458477318287, + 0.005241201259195805, + 0.020086538046598434, + 0.018752440810203552, + 0.006478622555732727, + 0.017662694677710533, + 0.021974442526698112, + 0.020324552431702614, + 0.09812479466199875, + 0.12602847814559937, + 0.04195524752140045, + 0.0025961995124816895, + 0.004726107697933912, + 0.006199920084327459, + 0.004454962443560362, + 0.00366862746886909, + 0.005193982273340225, + 0.006900094915181398, + 0.008153941482305527, + 0.0067181712947785854, + 0.0035514056216925383, + 0.005865985061973333, + 0.004267855081707239, + 0.004301650915294886, + 0.005189745686948299, + 0.004848388954997063, + 0.006107573863118887, + 0.0026379069313406944, + 0.003311102045699954, + 0.005945893935859203, + 0.00971872266381979, + 0.005493734497576952, + 0.0050557577051222324, + 0.005346253514289856, + 0.0031749429181218147 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 1, + 4 + ], + "selected_reward": -0.004633167292922735 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are two mosquitoes with large eyes visible in the video?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the leading mosquito's eyes emerald green and the following mosquito's eyes sapphire blue?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the mosquitoes' wings translucent with a shimmering, soap-bubble-like iridescence?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the setting a sun-drenched enchanted garden featuring giant dew-drenched clover leaves and wildflowers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera pan smoothly to the left while capturing the chase?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.1798238307237625 + ], + "opposite_logprob_tokens": [ + -1.804823875427246 + ], + "logprob_reward": -0.1798238307237625, + "yes_no_margin_reward": 1.625, + "token_credit_reward": -0.1798238307237625, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07924406230449677, + 0.11918769031763077, + 0.03684915229678154, + 0.027517491951584816, + 0.05745350942015648, + 0.040477264672517776, + 0.5642358660697937, + 0.018384860828518867, + 0.02714739553630352, + 0.029502684250473976 + ], + "credit_map": [ + 0.004950840026140213, + 0.004392486531287432, + 0.026108860969543457, + 0.007482617627829313, + 0.006371765397489071, + 0.005808301270008087, + 0.0064392550848424435, + 0.017689932137727737, + 0.011918358504772186, + 0.011209605261683464, + 0.016148915514349937, + 0.01593046635389328, + 0.00976820569485426, + 0.01292235404253006, + 0.01964898221194744, + 0.02164079248905182, + 0.002945686224848032, + 0.003701146226376295, + 0.004526027012616396, + 0.0047994195483624935, + 0.0032146458979696035, + 0.0037230681627988815, + 0.010781518183648586, + 0.0031576340552419424, + 0.0023338806349784136, + 0.0025063862558454275, + 0.005072580650448799, + 0.003241173457354307, + 0.0021484477911144495, + 0.003284406615421176, + 0.004931105766445398, + 0.003999507986009121, + 0.005737425293773413, + 0.003766085719689727, + 0.008753681555390358, + 0.008550342172384262, + 0.004125980660319328, + 0.010186613537371159, + 0.007480444852262735, + 0.00885293260216713, + 0.004881041124463081, + 0.0047582220286130905, + 0.006999598350375891, + 0.007355696056038141, + 0.0032975568901747465, + 0.004359032027423382, + 0.003698545740917325, + 0.005127566866576672, + 0.03430354595184326, + 0.009125873446464539, + 0.019650904461741447, + 0.040628861635923386, + 0.025852341204881668, + 0.19831223785877228, + 0.1892048865556717, + 0.047157224267721176, + 0.002190566621720791, + 0.003978493623435497, + 0.002447521546855569, + 0.0016732066869735718, + 0.0030117749702185392, + 0.001326234545558691, + 0.0021945093758404255, + 0.0015625521773472428, + 0.0028660171665251255, + 0.005456161219626665, + 0.0032748158555477858, + 0.003381218994036317, + 0.0036568979267030954, + 0.002567741321399808, + 0.003484899876639247, + 0.002459641546010971, + 0.003775401972234249, + 0.0030309970024973154, + 0.0035751224495470524, + 0.006489413324743509, + 0.0024391617625951767, + 0.00303534884005785, + 0.004552932921797037, + 0.002604306908324361 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 1, + 0 + ], + "selected_reward": -0.1798238307237625 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "dynamic tracking shot, shot from a slightly upward angle, pans smoothly to the left, capturing a playful aerial chase between two large-eyed mosquitoes, rendered in a high-quality, vibrant anime style. The leading mosquito, with its oversized, gleaming emerald eyes and a hint of slyness, deftly executes a playful aerial maneuver. Following closely behind is its companion, also sporting enormous sapphire eyes and a playful yet determined expression, tilting its round, cartoonish body through a sharp turn. Their delicate, translucent wings flap rapidly, shimmering in the light with a soap-bubble-like iridescence, while their slender, segmented limbs trace graceful arcs through the air. The chase takes place deep within a sun-drenched enchanted garden, against a backdrop of giant, dew-drenched clover leaves and vibrantly colored giant wildflowers, creating a soft bokeh effect. A beam of warm golden afternoon sunlight streams in from just outside the frame, creating a soft, ethereal lens flare that sweeps across the screen in orange and gold hues, accentuating their flight path while adding a dreamlike quality to this vibrant, highly saturated color palette, creating an atmosphere of lightheartedness, wonder, and joyful energy.", + "seed": 31293, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000095/group_9/index_3384_seed_31293_dp_19_rank_38_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 9, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.05505991606041789, + "vlm_logprob": -0.05505991606041789, + "vlm_yes_no_margin": 3.9, + "vlm_token_credit": -0.05505991606041789, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are two mosquitoes with large eyes visible in the video?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the leading mosquito's eyes emerald green and the following mosquito's eyes sapphire blue?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the mosquitoes' wings translucent with a shimmering, soap-bubble-like iridescence?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the setting a sun-drenched enchanted garden featuring giant dew-drenched clover leaves and wildflowers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera pan smoothly to the left while capturing the chase?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02726570889353752 + ], + "opposite_logprob_tokens": [ + -3.652265787124634 + ], + "logprob_reward": -0.02726570889353752, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.02726570889353752, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0975654125213623, + 0.03303883597254753, + 0.045920368283987045, + 0.037030305713415146, + 0.039020150899887085, + 0.12034916132688522, + 0.05932438001036644, + 0.09886683523654938, + 0.3413541316986084, + 0.12753038108348846 + ], + "credit_map": [ + 0.006651176139712334, + 0.009777862578630447, + 0.011902148835361004, + 0.024037567898631096, + 0.007650498766452074, + 0.015117965638637543, + 0.011676081456243992, + 0.010752118192613125, + 0.006170628126710653, + 0.002828263444826007, + 0.0044562299735844135, + 0.004311784170567989, + 0.0021553293336182833, + 0.003122274298220873, + 0.0047728801146149635, + 0.0052214497700333595, + 0.003462750930339098, + 0.004373721778392792, + 0.005918856244534254, + 0.007055025547742844, + 0.0050772069953382015, + 0.003190484829246998, + 0.011979472823441029, + 0.004862852394580841, + 0.0038698441348969936, + 0.004784843884408474, + 0.003310452913865447, + 0.004730306565761566, + 0.004242120310664177, + 0.005409081932157278, + 0.005341808311641216, + 0.005341853480786085, + 0.0038203985895961523, + 0.0036433348432183266, + 0.002717741299420595, + 0.005114100873470306, + 0.009499621577560902, + 0.004352994728833437, + 0.005394707899540663, + 0.004477254580706358, + 0.0220754686743021, + 0.008824948221445084, + 0.0074546379037201405, + 0.010229986160993576, + 0.00769779970869422, + 0.009032039903104305, + 0.01279307808727026, + 0.042241211980581284, + 0.0068011051043868065, + 0.006057795137166977, + 0.007067271042615175, + 0.007111002691090107, + 0.006859594024717808, + 0.007028637453913689, + 0.008549349382519722, + 0.009849635884165764, + 0.018736256286501884, + 0.0087070157751441, + 0.009556903503835201, + 0.010151568800210953, + 0.008544392883777618, + 0.011617655865848064, + 0.008543342351913452, + 0.023009715601801872, + 0.03580822795629501, + 0.08232443034648895, + 0.03592236712574959, + 0.051336366683244705, + 0.032174739986658096, + 0.03676064684987068, + 0.04215086251497269, + 0.024876506999135017, + 0.0077039990574121475, + 0.028194431215524673, + 0.036770034581422806, + 0.013923185877501965, + 0.012868097051978111, + 0.0059797195717692375, + 0.015934117138385773, + 0.00615681242197752 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 9, + 5 + ], + "selected_reward": -0.02726570889353752 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are two mosquitoes with large eyes visible in the video?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the leading mosquito's eyes emerald green and the following mosquito's eyes sapphire blue?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the mosquitoes' wings translucent with a shimmering, soap-bubble-like iridescence?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the setting a sun-drenched enchanted garden featuring giant dew-drenched clover leaves and wildflowers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera pan smoothly to the left while capturing the chase?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.038109809160232544 + ], + "opposite_logprob_tokens": [ + -3.28810977935791 + ], + "logprob_reward": -0.038109809160232544, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.038109809160232544, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.20415283739566803, + 0.027319103479385376, + 0.041501838713884354, + 0.02995089255273342, + 0.016164805740118027, + 0.08070053905248642, + 0.05220268666744232, + 0.055897075682878494, + 0.31008291244506836, + 0.18202726542949677 + ], + "credit_map": [ + 0.012002332136034966, + 0.02262009121477604, + 0.023839550092816353, + 0.05231179669499397, + 0.017360331490635872, + 0.029606923460960388, + 0.029254673048853874, + 0.017157135531306267, + 0.003923480398952961, + 0.0023558277171105146, + 0.003597303992137313, + 0.00392568064853549, + 0.0019872498232871294, + 0.0023111142218112946, + 0.005884929560124874, + 0.003333518747240305, + 0.002975684590637684, + 0.0043311407789587975, + 0.005413250997662544, + 0.008607902564108372, + 0.0045759305357933044, + 0.0027305821422487497, + 0.008466459810733795, + 0.004400886595249176, + 0.0027647025417536497, + 0.002600350882858038, + 0.002884509740397334, + 0.008292867802083492, + 0.005106621887534857, + 0.0023219333961606026, + 0.003667194861918688, + 0.0023127098102122545, + 0.0017119398107752204, + 0.0018785619176924229, + 0.0016441436018794775, + 0.0024428428150713444, + 0.0024118658620864153, + 0.002140129916369915, + 0.0020271276589483023, + 0.0019081936916336417, + 0.01346762664616108, + 0.0058378675021231174, + 0.005722582805901766, + 0.007925315760076046, + 0.007486785762012005, + 0.0072648790664970875, + 0.008100701496005058, + 0.024894777685403824, + 0.00521333422511816, + 0.008292471058666706, + 0.007719592656940222, + 0.008048372343182564, + 0.006593853235244751, + 0.005507530644536018, + 0.006577407941222191, + 0.0042501273564994335, + 0.012318603694438934, + 0.0037870120722800493, + 0.00380520336329937, + 0.003990135621279478, + 0.003012256231158972, + 0.007062072400003672, + 0.004358034580945969, + 0.017563756555318832, + 0.03477349877357483, + 0.08151908963918686, + 0.030741600319743156, + 0.04034104198217392, + 0.026608305051922798, + 0.03523213788866997, + 0.0402056947350502, + 0.020661527290940285, + 0.01364849228411913, + 0.030473699793219566, + 0.03213910013437271, + 0.035853464156389236, + 0.02318265475332737, + 0.013278590515255928, + 0.021490933373570442, + 0.01196032389998436 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 0, + 9 + ], + "selected_reward": -0.038109809160232544 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are two mosquitoes with large eyes visible in the video?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the leading mosquito's eyes emerald green and the following mosquito's eyes sapphire blue?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the mosquitoes' wings translucent with a shimmering, soap-bubble-like iridescence?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the setting a sun-drenched enchanted garden featuring giant dew-drenched clover leaves and wildflowers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera pan smoothly to the left while capturing the chase?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.00594549672678113 + ], + "opposite_logprob_tokens": [ + -5.130945682525635 + ], + "logprob_reward": -0.00594549672678113, + "yes_no_margin_reward": 5.125, + "token_credit_reward": -0.00594549672678113, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.20688952505588531, + 0.17717395722866058, + 0.05252069607377052, + 0.043577272444963455, + 0.04044477269053459, + 0.05101444944739342, + 0.03518223017454147, + 0.10586271435022354, + 0.21337850391864777, + 0.07395592331886292 + ], + "credit_map": [ + 0.012812650762498379, + 0.0217487420886755, + 0.01973523199558258, + 0.03162775933742523, + 0.014987384900450706, + 0.052534449845552444, + 0.022011609748005867, + 0.03143170475959778, + 0.030871860682964325, + 0.013432610780000687, + 0.019784927368164062, + 0.016257748007774353, + 0.0107621094211936, + 0.015365076251327991, + 0.022113529965281487, + 0.048586100339889526, + 0.003725107293576002, + 0.004050510004162788, + 0.006541532464325428, + 0.007514973171055317, + 0.005064731929451227, + 0.0029224720783531666, + 0.01737431436777115, + 0.0053270491771399975, + 0.004636876750737429, + 0.004324610810726881, + 0.002878376515582204, + 0.008289189077913761, + 0.006946717854589224, + 0.00524845952168107, + 0.006428963504731655, + 0.004824077244848013, + 0.003579138545319438, + 0.003403878305107355, + 0.002284300746396184, + 0.003761382307857275, + 0.013253964483737946, + 0.002652378985658288, + 0.00827503576874733, + 0.003234696574509144, + 0.007580469828099012, + 0.0047812988050282, + 0.003930582664906979, + 0.00621378468349576, + 0.004407254047691822, + 0.004762474447488785, + 0.005612057633697987, + 0.013726526871323586, + 0.005938410293310881, + 0.0038046562112867832, + 0.003922954201698303, + 0.0031547006219625473, + 0.0037377316039055586, + 0.0033904279116541147, + 0.00497762905433774, + 0.00625572120770812, + 0.014819078147411346, + 0.009155360981822014, + 0.013972228392958641, + 0.013681357726454735, + 0.010785705409944057, + 0.011182880960404873, + 0.014618896879255772, + 0.01764720492064953, + 0.03370240703225136, + 0.03315006569027901, + 0.036715880036354065, + 0.029339225962758064, + 0.018875110894441605, + 0.017841124907135963, + 0.027692116796970367, + 0.016062570735812187, + 0.013074642978608608, + 0.012890906073153019, + 0.010275059379637241, + 0.012955068610608578, + 0.007673144806176424, + 0.005356230773031712, + 0.007479153573513031, + 0.004251719452440739 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 0, + 1 + ], + "selected_reward": -0.00594549672678113 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are two mosquitoes with large eyes visible in the video?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the leading mosquito's eyes emerald green and the following mosquito's eyes sapphire blue?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the mosquitoes' wings translucent with a shimmering, soap-bubble-like iridescence?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the setting a sun-drenched enchanted garden featuring giant dew-drenched clover leaves and wildflowers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera pan smoothly to the left while capturing the chase?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.002483856864273548 + ], + "opposite_logprob_tokens": [ + -6.00248384475708 + ], + "logprob_reward": -0.002483856864273548, + "yes_no_margin_reward": 6.0, + "token_credit_reward": -0.002483856864273548, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0684577152132988, + 0.08795251697301865, + 0.11080030351877213, + 0.07138366252183914, + 0.1304663121700287, + 0.05007997527718544, + 0.07252201437950134, + 0.1358342468738556, + 0.10553500056266785, + 0.16696825623512268 + ], + "credit_map": [ + 0.005449038930237293, + 0.009230701252818108, + 0.007541255559772253, + 0.010653886944055557, + 0.007740935310721397, + 0.010278415866196156, + 0.009360872209072113, + 0.008202613331377506, + 0.00810183770954609, + 0.006191628519445658, + 0.010494385845959187, + 0.009828055277466774, + 0.005568757653236389, + 0.019084936007857323, + 0.008458337746560574, + 0.020224589854478836, + 0.009213294833898544, + 0.006786609999835491, + 0.013057705946266651, + 0.022691335529088974, + 0.009891432709991932, + 0.014114982448518276, + 0.02396552637219429, + 0.011079426854848862, + 0.006615061778575182, + 0.006280547473579645, + 0.00774609437212348, + 0.009907953441143036, + 0.010985279455780983, + 0.012243935838341713, + 0.009129955433309078, + 0.008474842645227909, + 0.012960819527506828, + 0.011350477114319801, + 0.009612169116735458, + 0.01483414601534605, + 0.03894057124853134, + 0.00813496857881546, + 0.02317066304385662, + 0.01146251242607832, + 0.006087685935199261, + 0.005043400451540947, + 0.004887108225375414, + 0.0075988504104316235, + 0.005650941748172045, + 0.007570577785372734, + 0.008386638015508652, + 0.004854780156165361, + 0.010122692212462425, + 0.010460232384502888, + 0.016764739528298378, + 0.004880571737885475, + 0.00651896744966507, + 0.005480621475726366, + 0.007159087806940079, + 0.011135108768939972, + 0.026051504537463188, + 0.008890151046216488, + 0.01351229939609766, + 0.010358400642871857, + 0.006439683493226767, + 0.020097339525818825, + 0.012473349459469318, + 0.03801152482628822, + 0.010403330437839031, + 0.023510098457336426, + 0.013244292698800564, + 0.02214214578270912, + 0.007464539259672165, + 0.010114402510225773, + 0.011347129009664059, + 0.00730907404795289, + 0.012451505288481712, + 0.02689933218061924, + 0.04392118379473686, + 0.02787480317056179, + 0.020600056275725365, + 0.014885845594108105, + 0.011255662888288498, + 0.009079881012439728 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 7, + 4 + ], + "selected_reward": -0.002483856864273548 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are two mosquitoes with large eyes visible in the video?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the leading mosquito's eyes emerald green and the following mosquito's eyes sapphire blue?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the mosquitoes' wings translucent with a shimmering, soap-bubble-like iridescence?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the setting a sun-drenched enchanted garden featuring giant dew-drenched clover leaves and wildflowers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera pan smoothly to the left while capturing the chase?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.2014947086572647 + ], + "opposite_logprob_tokens": [ + -1.7014946937561035 + ], + "logprob_reward": -0.2014947086572647, + "yes_no_margin_reward": 1.5, + "token_credit_reward": -0.2014947086572647, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1266343593597412, + 0.032565124332904816, + 0.05693964287638664, + 0.09608110040426254, + 0.14235396683216095, + 0.16808445751667023, + 0.05505216494202614, + 0.10427579283714294, + 0.15175217390060425, + 0.06626126915216446 + ], + "credit_map": [ + 0.009459860622882843, + 0.01330492552369833, + 0.014361904002726078, + 0.02842130698263645, + 0.009744496084749699, + 0.01861620880663395, + 0.0197566207498312, + 0.012969041243195534, + 0.003028166014701128, + 0.0030127812642604113, + 0.003148915944620967, + 0.004619630519300699, + 0.0028311575297266245, + 0.0035564634017646313, + 0.00725925387814641, + 0.005108756013214588, + 0.004841356072574854, + 0.0043523795902729034, + 0.006925919093191624, + 0.009211587719619274, + 0.005071776919066906, + 0.012576338835060596, + 0.008813870139420033, + 0.005146410781890154, + 0.006589001510292292, + 0.0091836117208004, + 0.007532555144280195, + 0.016475984826683998, + 0.012444699183106422, + 0.01966123655438423, + 0.011780709959566593, + 0.012413304299116135, + 0.01398959755897522, + 0.02315492369234562, + 0.014121412299573421, + 0.024592595174908638, + 0.015929153189063072, + 0.016739681363105774, + 0.01346345990896225, + 0.02036314085125923, + 0.031434159725904465, + 0.01222989708185196, + 0.01020669937133789, + 0.009856414049863815, + 0.009637665003538132, + 0.0152738643810153, + 0.016298707574605942, + 0.06314705312252045, + 0.006953578907996416, + 0.005114696454256773, + 0.009176861494779587, + 0.0044706338085234165, + 0.006756670773029327, + 0.00475971819832921, + 0.010038723237812519, + 0.007781283929944038, + 0.021973945200443268, + 0.008284917101264, + 0.008442264050245285, + 0.007853320799767971, + 0.004574055317789316, + 0.014187220484018326, + 0.007320965174585581, + 0.03163909912109375, + 0.015047525987029076, + 0.025778379291296005, + 0.019855719059705734, + 0.02745964005589485, + 0.014585082419216633, + 0.011436174623668194, + 0.0264936201274395, + 0.011096028611063957, + 0.0048130229115486145, + 0.007549148518592119, + 0.012775358743965626, + 0.014043296687304974, + 0.008813581429421902, + 0.006912711542099714, + 0.00681935902684927, + 0.004534789826720953 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 8, + 4 + ], + "selected_reward": -0.2014947086572647 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "dynamic tracking shot, shot from a slightly upward angle, pans smoothly to the left, capturing a playful aerial chase between two large-eyed mosquitoes, rendered in a high-quality, vibrant anime style. The leading mosquito, with its oversized, gleaming emerald eyes and a hint of slyness, deftly executes a playful aerial maneuver. Following closely behind is its companion, also sporting enormous sapphire eyes and a playful yet determined expression, tilting its round, cartoonish body through a sharp turn. Their delicate, translucent wings flap rapidly, shimmering in the light with a soap-bubble-like iridescence, while their slender, segmented limbs trace graceful arcs through the air. The chase takes place deep within a sun-drenched enchanted garden, against a backdrop of giant, dew-drenched clover leaves and vibrantly colored giant wildflowers, creating a soft bokeh effect. A beam of warm golden afternoon sunlight streams in from just outside the frame, creating a soft, ethereal lens flare that sweeps across the screen in orange and gold hues, accentuating their flight path while adding a dreamlike quality to this vibrant, highly saturated color palette, creating an atmosphere of lightheartedness, wonder, and joyful energy.", + "seed": 31294, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000095/group_9/index_3384_seed_31294_dp_19_rank_38_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 9, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.12670409344136715, + "vlm_logprob": -0.12670409344136715, + "vlm_yes_no_margin": 3.050000011920929, + "vlm_token_credit": -0.12670409344136715, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are two mosquitoes with large eyes visible in the video?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the leading mosquito's eyes emerald green and the following mosquito's eyes sapphire blue?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the mosquitoes' wings translucent with a shimmering, soap-bubble-like iridescence?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the setting a sun-drenched enchanted garden featuring giant dew-drenched clover leaves and wildflowers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera pan smoothly to the left while capturing the chase?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07116320729255676 + ], + "opposite_logprob_tokens": [ + -2.6961631774902344 + ], + "logprob_reward": -0.07116320729255676, + "yes_no_margin_reward": 2.625, + "token_credit_reward": -0.07116320729255676, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1986730992794037, + 0.1213698759675026, + 0.11744111031293869, + 0.10744412243366241, + 0.2595863938331604, + 0.0475764274597168, + 0.06902121007442474, + 0.045220572501420975, + 0.018591320142149925, + 0.01507585495710373 + ], + "credit_map": [ + 0.009985549375414848, + 0.010302457958459854, + 0.045665204524993896, + 0.07886594533920288, + 0.007721535861492157, + 0.0069134896621108055, + 0.02927066758275032, + 0.009948255494236946, + 0.013212845660746098, + 0.026341192424297333, + 0.007305673789232969, + 0.008052629418671131, + 0.009916458278894424, + 0.018929824233055115, + 0.026088660582900047, + 0.011522593908011913, + 0.006319614127278328, + 0.0084104984998703, + 0.0031960438936948776, + 0.0038525457493960857, + 0.045163389295339584, + 0.012279793620109558, + 0.005730569362640381, + 0.032488662749528885, + 0.006114928051829338, + 0.007716191466897726, + 0.012478376738727093, + 0.024402586743235588, + 0.009345789439976215, + 0.015949945896863937, + 0.020167071372270584, + 0.01126923318952322, + 0.027962572872638702, + 0.05824365094304085, + 0.02976473979651928, + 0.026635892689228058, + 0.030398642644286156, + 0.026345161721110344, + 0.03390161693096161, + 0.0263341274112463, + 0.00451001338660717, + 0.005966722033917904, + 0.006349333096295595, + 0.009022900834679604, + 0.0060841855593025684, + 0.006381148938089609, + 0.005351400002837181, + 0.003910723142325878, + 0.005846809595823288, + 0.009198606014251709, + 0.00561328511685133, + 0.006800977513194084, + 0.007022842299193144, + 0.011560067534446716, + 0.012010332196950912, + 0.010968286544084549, + 0.007083978038281202, + 0.0035305358469486237, + 0.006882897578179836, + 0.0027107561472803354, + 0.004893315490335226, + 0.005450806114822626, + 0.008637751452624798, + 0.006030534394085407, + 0.002085441956296563, + 0.0030921485740691423, + 0.0032940441742539406, + 0.0032378293108195066, + 0.0018178406171500683, + 0.00189715507440269, + 0.0017686942592263222, + 0.0013981679221615195, + 0.0013669671025127172, + 0.0018064422765746713, + 0.001545072067528963, + 0.0015968178631737828, + 0.002301576780155301, + 0.002198467031121254, + 0.0017532053170725703, + 0.002507308265194297 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 0, + 1 + ], + "selected_reward": -0.07116320729255676 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are two mosquitoes with large eyes visible in the video?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the leading mosquito's eyes emerald green and the following mosquito's eyes sapphire blue?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the mosquitoes' wings translucent with a shimmering, soap-bubble-like iridescence?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the setting a sun-drenched enchanted garden featuring giant dew-drenched clover leaves and wildflowers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera pan smoothly to the left while capturing the chase?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.34860795736312866 + ], + "opposite_logprob_tokens": [ + -1.2236080169677734 + ], + "logprob_reward": -0.34860795736312866, + "yes_no_margin_reward": 0.8750000596046448, + "token_credit_reward": -0.34860795736312866, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.030695611611008644, + 0.04997938498854637, + 0.5814648270606995, + 0.06484256684780121, + 0.04353351518511772, + 0.0960274264216423, + 0.06139848381280899, + 0.04730036109685898, + 0.007272145710885525, + 0.017485639080405235 + ], + "credit_map": [ + 0.00293622724711895, + 0.0029149577021598816, + 0.0069890255108475685, + 0.00555175356566906, + 0.0030501100700348616, + 0.003004683181643486, + 0.004084593616425991, + 0.002164257690310478, + 0.005745632108300924, + 0.011482901871204376, + 0.0035644862800836563, + 0.0037525808438658714, + 0.004839619621634483, + 0.006891658063977957, + 0.009632600471377373, + 0.00406990572810173, + 0.03641444072127342, + 0.03963472321629524, + 0.012297856621444225, + 0.012307743541896343, + 0.24385246634483337, + 0.04046139121055603, + 0.024997757747769356, + 0.17149846255779266, + 0.005345083773136139, + 0.005387964192777872, + 0.007459538988769054, + 0.011820771731436253, + 0.008043266832828522, + 0.008943760767579079, + 0.012359076179564, + 0.005483102053403854, + 0.004592657554894686, + 0.0045285653322935104, + 0.005871532950550318, + 0.0046847108751535416, + 0.006117808632552624, + 0.005285308696329594, + 0.00650823162868619, + 0.0059446911327540874, + 0.010928290896117687, + 0.012765971012413502, + 0.01317816972732544, + 0.02028965950012207, + 0.010703608393669128, + 0.010904618538916111, + 0.010585702955722809, + 0.006671401206403971, + 0.005153476260602474, + 0.009670836851000786, + 0.007169731426984072, + 0.009719735011458397, + 0.005978357512503862, + 0.009107178077101707, + 0.006561526097357273, + 0.008037632331252098, + 0.004893469158560038, + 0.003426722716540098, + 0.009872276335954666, + 0.004103555344045162, + 0.004517737310379744, + 0.005968542769551277, + 0.006207852158695459, + 0.008310199715197086, + 0.0008141351863741875, + 0.001009534578770399, + 0.001019408111460507, + 0.0013675520895048976, + 0.0008997200056910515, + 0.000758503912948072, + 0.0007902331999503076, + 0.0006130579276941717, + 0.001500381389632821, + 0.0016379334265366197, + 0.002182276453822851, + 0.0019073273288086057, + 0.002751057967543602, + 0.0027032161597162485, + 0.002008152427151799, + 0.002795293228700757 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 5, + 3 + ], + "selected_reward": -0.34860795736312866 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are two mosquitoes with large eyes visible in the video?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the leading mosquito's eyes emerald green and the following mosquito's eyes sapphire blue?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the mosquitoes' wings translucent with a shimmering, soap-bubble-like iridescence?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the setting a sun-drenched enchanted garden featuring giant dew-drenched clover leaves and wildflowers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera pan smoothly to the left while capturing the chase?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.007623862475156784 + ], + "opposite_logprob_tokens": [ + -4.882623672485352 + ], + "logprob_reward": -0.007623862475156784, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.007623862475156784, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.17108194530010223, + 0.049017131328582764, + 0.23518693447113037, + 0.12946303188800812, + 0.06249199062585831, + 0.1296042799949646, + 0.09542949497699738, + 0.037529848515987396, + 0.04933946579694748, + 0.04085594788193703 + ], + "credit_map": [ + 0.009463414549827576, + 0.007532284129410982, + 0.03601479157805443, + 0.07122708857059479, + 0.0076832096092402935, + 0.004658720456063747, + 0.027961263433098793, + 0.0065411776304244995, + 0.0043624574318528175, + 0.006670426111668348, + 0.003564676968380809, + 0.004924135748296976, + 0.00575385382398963, + 0.008970862254500389, + 0.009066164493560791, + 0.005704554263502359, + 0.016000017523765564, + 0.01606261543929577, + 0.010289906524121761, + 0.00823415257036686, + 0.08531366288661957, + 0.02377581223845482, + 0.015848305076360703, + 0.05966246500611305, + 0.00804024375975132, + 0.009229924529790878, + 0.01726711355149746, + 0.026472218334674835, + 0.010522411204874516, + 0.018108174204826355, + 0.022731058299541473, + 0.017091885209083557, + 0.006791993975639343, + 0.007232711650431156, + 0.009312308393418789, + 0.00632614828646183, + 0.00868946686387062, + 0.007190357428044081, + 0.009437544271349907, + 0.007511457893997431, + 0.014830118045210838, + 0.016349587589502335, + 0.01982223242521286, + 0.030850639566779137, + 0.012060292065143585, + 0.013266094960272312, + 0.013223146088421345, + 0.009202172048389912, + 0.008701562881469727, + 0.012477829121053219, + 0.011449594050645828, + 0.013261943124234676, + 0.010330636985599995, + 0.014105631969869137, + 0.013481266796588898, + 0.011621031910181046, + 0.0046134889125823975, + 0.0032619016710668802, + 0.0047110081650316715, + 0.0025281694252043962, + 0.0033070764038711786, + 0.005915557965636253, + 0.007868361659348011, + 0.005324285943061113, + 0.005862537305802107, + 0.00864576268941164, + 0.010213813744485378, + 0.008697462268173695, + 0.004543060902506113, + 0.0039092134684324265, + 0.004284541588276625, + 0.0031830722000449896, + 0.004299575928598642, + 0.0036264352966099977, + 0.011275534518063068, + 0.005488019902259111, + 0.004750611260533333, + 0.003569040447473526, + 0.003656054148450494, + 0.0041906763799488544 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 0, + 5 + ], + "selected_reward": -0.007623862475156784 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are two mosquitoes with large eyes visible in the video?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the leading mosquito's eyes emerald green and the following mosquito's eyes sapphire blue?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the mosquitoes' wings translucent with a shimmering, soap-bubble-like iridescence?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the setting a sun-drenched enchanted garden featuring giant dew-drenched clover leaves and wildflowers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera pan smoothly to the left while capturing the chase?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0046315062791109085 + ], + "opposite_logprob_tokens": [ + -5.379631519317627 + ], + "logprob_reward": -0.0046315062791109085, + "yes_no_margin_reward": 5.375, + "token_credit_reward": -0.0046315062791109085, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1665402203798294, + 0.048026636242866516, + 0.3683150112628937, + 0.0460720919072628, + 0.07973074167966843, + 0.1393967866897583, + 0.04037047550082207, + 0.03778833523392677, + 0.04337674379348755, + 0.03038294054567814 + ], + "credit_map": [ + 0.017477909103035927, + 0.00814093742519617, + 0.010450662113726139, + 0.07109569013118744, + 0.006983643863350153, + 0.004139573313295841, + 0.04188820719718933, + 0.006363609805703163, + 0.005258255638182163, + 0.0060379523783922195, + 0.003491475246846676, + 0.004229830112308264, + 0.005827613640576601, + 0.00958545133471489, + 0.007673653773963451, + 0.005922409705817699, + 0.021804098039865494, + 0.02599143236875534, + 0.007537307683378458, + 0.00909414142370224, + 0.15254822373390198, + 0.025694508105516434, + 0.018251314759254456, + 0.10739400237798691, + 0.004386413376778364, + 0.005571585614234209, + 0.0055274032056331635, + 0.00807819701731205, + 0.004873059689998627, + 0.006856507156044245, + 0.0061537581495940685, + 0.0046251676976680756, + 0.011007504537701607, + 0.013616764917969704, + 0.009120981208980083, + 0.010538983158767223, + 0.010053858160972595, + 0.010088043287396431, + 0.008095826022326946, + 0.007208786904811859, + 0.01650851219892502, + 0.01781294494867325, + 0.02013821341097355, + 0.0333319716155529, + 0.013903452083468437, + 0.013176721520721912, + 0.014426997862756252, + 0.010097973048686981, + 0.002996023977175355, + 0.0037502730265259743, + 0.003998230677098036, + 0.004603323061019182, + 0.003760013496503234, + 0.00816696509718895, + 0.005236747208982706, + 0.007858900353312492, + 0.005551277194172144, + 0.0031087244860827923, + 0.005484288092702627, + 0.0024776963982731104, + 0.0032317866571247578, + 0.005075845867395401, + 0.007292722351849079, + 0.005566000938415527, + 0.0049379016272723675, + 0.007301849313080311, + 0.009363318793475628, + 0.007622756063938141, + 0.003965694922953844, + 0.0038590373005717993, + 0.0033154787961393595, + 0.0030107107013463974, + 0.004486407153308392, + 0.0026480657979846, + 0.008809524588286877, + 0.004794336389750242, + 0.002226563636213541, + 0.002302181674167514, + 0.002184527227655053, + 0.002931336872279644 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 0, + 5 + ], + "selected_reward": -0.0046315062791109085 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are two mosquitoes with large eyes visible in the video?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the leading mosquito's eyes emerald green and the following mosquito's eyes sapphire blue?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the mosquitoes' wings translucent with a shimmering, soap-bubble-like iridescence?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the setting a sun-drenched enchanted garden featuring giant dew-drenched clover leaves and wildflowers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera pan smoothly to the left while capturing the chase?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.20149393379688263 + ], + "opposite_logprob_tokens": [ + -1.7014939785003662 + ], + "logprob_reward": -0.20149393379688263, + "yes_no_margin_reward": 1.5, + "token_credit_reward": -0.20149393379688263, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06967627257108688, + 0.04617742821574211, + 0.6153438091278076, + 0.04595838114619255, + 0.028383543714880943, + 0.023140566423535347, + 0.03978915885090828, + 0.07678408175706863, + 0.03713088110089302, + 0.017615875229239464 + ], + "credit_map": [ + 0.0051454827189445496, + 0.003650064580142498, + 0.01632353477180004, + 0.02486512064933777, + 0.0032600173726677895, + 0.003381449030712247, + 0.009230689145624638, + 0.003819920588284731, + 0.006195484660565853, + 0.00920209288597107, + 0.002807224402204156, + 0.004738219082355499, + 0.005589939653873444, + 0.007087144535034895, + 0.005680615082383156, + 0.004876712337136269, + 0.0327041931450367, + 0.0441608764231205, + 0.012247377075254917, + 0.01341075450181961, + 0.2553655505180359, + 0.04096594452857971, + 0.025128982961177826, + 0.19136016070842743, + 0.004695276264101267, + 0.005725439637899399, + 0.0058080111630260944, + 0.008741799741983414, + 0.004472479689866304, + 0.00697337556630373, + 0.005908239167183638, + 0.003633764572441578, + 0.004756541457027197, + 0.0035625034943223, + 0.003756321733817458, + 0.0038537418004125357, + 0.003731892444193363, + 0.002543696667999029, + 0.003477863734588027, + 0.0027009849436581135, + 0.0036774491891264915, + 0.0029847228433936834, + 0.0047212825156748295, + 0.004177948459982872, + 0.0019839832093566656, + 0.001812527421861887, + 0.0020741219632327557, + 0.0017085329163819551, + 0.004954630043357611, + 0.004004741553217173, + 0.005204969085752964, + 0.005563960410654545, + 0.0025505044031888247, + 0.006253062281757593, + 0.005372803192585707, + 0.0058844927698373795, + 0.012126387096941471, + 0.00677133584395051, + 0.01513326819986105, + 0.008674724027514458, + 0.007106401026248932, + 0.009200303815305233, + 0.008910369127988815, + 0.008861302398145199, + 0.003262873739004135, + 0.004566113464534283, + 0.005517026875168085, + 0.004907171241939068, + 0.00526383426040411, + 0.003928478807210922, + 0.005762487184256315, + 0.003922899719327688, + 0.0020451063755899668, + 0.0014666973147541285, + 0.0035155012737959623, + 0.0020750926341861486, + 0.0027253672014921904, + 0.001863386482000351, + 0.0017056853976100683, + 0.002219041343778372 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 7, + 0 + ], + "selected_reward": -0.20149393379688263 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000101/group_0/index_tensor([6886, 6886])_group_0_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000101/group_0/index_tensor([6886, 6886])_group_0_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..d2dfdc75518c94dae4afedda5cc8dcf8a37a3937 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000101/group_0/index_tensor([6886, 6886])_group_0_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "static, extremely low-angle wide shot gazes across the vast, towering bookshelves of an ancient, cathedral-like library, their tops disappearing into the vast darkness of a vaulted ceiling. Crafted from dark, polished oak, with intricate Gothic carvings, the shelves are filled with thousands of ancient, leather-bound books of varying sizes. The scene is bathed in a gloomy black-and-white light, with a single, feeble source of light emanating from outside the camera casting long, dramatic shadows and highlighting a layer of fine dust in the still air. The impossible movement begins as the rigid structure of the bookshelf begins to tremble uncomfortably; the solid wood bends and sags as if made of wax. Simultaneously, the ancient books begin to melt, their rigid leather covers bubbling and dissolving into the yellowing pages beneath. The entire mass transforms into a black, viscous, gooey mass that initially drips slowly, then flows in a heavy, relentless stream down the shelf's surface, the inky liquid shimmering in the dim light. These streams cascade onto the dark marble floor below, forming an ever-growing, shimmering pool and filling the frame with a nightmarish atmosphere of silence, surreality, and magnificent decay.", + "seed": 28497, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000101/group_0/index_6886_seed_28497_dp_0_rank_0_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 0, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.07734167653834448, + "vlm_logprob": -0.07734167653834448, + "vlm_yes_no_margin": 3.85, + "vlm_token_credit": -0.07734167653834448, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a vast, cathedral-like library with towering bookshelves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera positioned at an extremely low angle looking up at the shelves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the bookshelves and books transform from solid wood/leather into a black, viscous, gooey mass?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene rendered in black-and-white with a single light source casting long shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the gooey mass flow down the shelves and form a shimmering pool on the dark marble floor?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07404492795467377 + ], + "opposite_logprob_tokens": [ + -2.699044942855835 + ], + "logprob_reward": -0.07404492795467377, + "yes_no_margin_reward": 2.625, + "token_credit_reward": -0.07404492795467377, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1978245973587036, + 0.15958963334560394, + 0.10374537855386734, + 0.09185170382261276, + 0.1011882796883583, + 0.033557627350091934, + 0.08613593131303787, + 0.04430866241455078, + 0.07888501137495041, + 0.10291319340467453 + ], + "credit_map": [ + 0.01596108265221119, + 0.03289758414030075, + 0.027483521029353142, + 0.02606230229139328, + 0.02882843278348446, + 0.028109800070524216, + 0.016586793586611748, + 0.02189508266746998, + 0.010905066505074501, + 0.01715448684990406, + 0.013761267066001892, + 0.024990420788526535, + 0.024562425911426544, + 0.014259585179388523, + 0.032157622277736664, + 0.02179875038564205, + 0.007823260501027107, + 0.021811021491885185, + 0.007752487435936928, + 0.0063076866790652275, + 0.024095304310321808, + 0.023076185956597328, + 0.005027937702834606, + 0.007851491682231426, + 0.009950638748705387, + 0.012082613073289394, + 0.006257760338485241, + 0.007521058898419142, + 0.020142467692494392, + 0.013003169558942318, + 0.01044534146785736, + 0.012448655441403389, + 0.013368580490350723, + 0.011419208720326424, + 0.010635078884661198, + 0.010366364382207394, + 0.01184866577386856, + 0.010732589289546013, + 0.021480632945895195, + 0.011337152682244778, + 0.006518216338008642, + 0.0029760883189737797, + 0.004281063564121723, + 0.002639313694089651, + 0.0031511480920016766, + 0.00382851785980165, + 0.0048714568838477135, + 0.00529182143509388, + 0.013904795050621033, + 0.008863424882292747, + 0.008373287506401539, + 0.0120890773832798, + 0.012269595637917519, + 0.010307527147233486, + 0.008906546048820019, + 0.011421672999858856, + 0.0047771381214261055, + 0.003303233766928315, + 0.009011662565171719, + 0.004196238704025745, + 0.004630844108760357, + 0.0060883150435984135, + 0.005461676046252251, + 0.006839552894234657, + 0.005607972387224436, + 0.0051939100958406925, + 0.008479785174131393, + 0.009814396500587463, + 0.011255110614001751, + 0.013425885699689388, + 0.012470303103327751, + 0.012637646868824959, + 0.005721111316233873, + 0.007428554818034172, + 0.012459523044526577, + 0.014642620459198952, + 0.01747400313615799, + 0.01658179797232151, + 0.01617732085287571, + 0.01242826133966446 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 2 + ], + "selected_reward": -0.07404492795467377 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a vast, cathedral-like library with towering bookshelves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera positioned at an extremely low angle looking up at the shelves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the bookshelves and books transform from solid wood/leather into a black, viscous, gooey mass?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene rendered in black-and-white with a single light source casting long shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the gooey mass flow down the shelves and form a shimmering pool on the dark marble floor?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.05501488596200943 + ], + "opposite_logprob_tokens": [ + -2.9300148487091064 + ], + "logprob_reward": -0.05501488596200943, + "yes_no_margin_reward": 2.875, + "token_credit_reward": -0.05501488596200943, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14343540370464325, + 0.13956394791603088, + 0.08412390202283859, + 0.0828968957066536, + 0.23805150389671326, + 0.03124157525599003, + 0.07694536447525024, + 0.036420438438653946, + 0.0706125795841217, + 0.09670836478471756 + ], + "credit_map": [ + 0.014035595580935478, + 0.021107010543346405, + 0.018418774008750916, + 0.021902933716773987, + 0.023487884551286697, + 0.018347004428505898, + 0.012130157090723515, + 0.014006058685481548, + 0.011314883828163147, + 0.013167626224458218, + 0.015050077810883522, + 0.022475959733128548, + 0.02256336249411106, + 0.013940504752099514, + 0.02383161336183548, + 0.017219940200448036, + 0.008231712505221367, + 0.013580973260104656, + 0.0063979788683354855, + 0.0055490294471383095, + 0.018029605969786644, + 0.021305713802576065, + 0.005506779998540878, + 0.005522117484360933, + 0.010586888529360294, + 0.009405290707945824, + 0.0075027733109891415, + 0.009920523501932621, + 0.016158269718289375, + 0.011351996101439, + 0.010398192331194878, + 0.0075729708187282085, + 0.03787952661514282, + 0.02529662474989891, + 0.02571220137178898, + 0.02596975862979889, + 0.03347565606236458, + 0.025506388396024704, + 0.03916512802243233, + 0.025046225637197495, + 0.004122058395296335, + 0.0035542917903512716, + 0.004175188485532999, + 0.00248579285107553, + 0.003108206670731306, + 0.005719576496630907, + 0.004034964367747307, + 0.004041499923914671, + 0.009046902880072594, + 0.009750999510288239, + 0.008027290925383568, + 0.010598796419799328, + 0.012229161337018013, + 0.01054239273071289, + 0.00718705914914608, + 0.009562769904732704, + 0.004544598050415516, + 0.0035821660421788692, + 0.0043562049977481365, + 0.005906154401600361, + 0.00425383634865284, + 0.004271036945283413, + 0.005141043104231358, + 0.004365399479866028, + 0.004776060115545988, + 0.005561988800764084, + 0.0101500628516078, + 0.007599823642522097, + 0.008110849186778069, + 0.01243300549685955, + 0.007152165286242962, + 0.014828634448349476, + 0.00542786018922925, + 0.007256152108311653, + 0.015421558171510696, + 0.01601051725447178, + 0.01059003546833992, + 0.01604924537241459, + 0.01459130086004734, + 0.011361709795892239 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 0, + 1 + ], + "selected_reward": -0.05501488596200943 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a vast, cathedral-like library with towering bookshelves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera positioned at an extremely low angle looking up at the shelves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the bookshelves and books transform from solid wood/leather into a black, viscous, gooey mass?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene rendered in black-and-white with a single light source casting long shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the gooey mass flow down the shelves and form a shimmering pool on the dark marble floor?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004657373297959566 + ], + "opposite_logprob_tokens": [ + -5.37965726852417 + ], + "logprob_reward": -0.004657373297959566, + "yes_no_margin_reward": 5.375, + "token_credit_reward": -0.004657373297959566, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09997251629829407, + 0.08426342159509659, + 0.09882744401693344, + 0.06181805580854416, + 0.10143335908651352, + 0.04760461673140526, + 0.09850229322910309, + 0.06782132387161255, + 0.11843445152044296, + 0.22132253646850586 + ], + "credit_map": [ + 0.011879823170602322, + 0.014352644793689251, + 0.015031869523227215, + 0.012186015024781227, + 0.012736092321574688, + 0.014035324566066265, + 0.008217288181185722, + 0.011533455923199654, + 0.005874314345419407, + 0.010179804638028145, + 0.0048971027135849, + 0.008679152466356754, + 0.013344807550311089, + 0.008323184214532375, + 0.013955372385680676, + 0.019009685143828392, + 0.007008007261902094, + 0.01642172411084175, + 0.005413326900452375, + 0.005361561197787523, + 0.032764092087745667, + 0.019171839579939842, + 0.0062491209246218204, + 0.006437772884964943, + 0.008033608086407185, + 0.009779505431652069, + 0.005879969801753759, + 0.004752727225422859, + 0.01087216567248106, + 0.010833792388439178, + 0.0056352983228862286, + 0.006030987482517958, + 0.010741209611296654, + 0.008370290510356426, + 0.008483439683914185, + 0.006758247036486864, + 0.011114970780909061, + 0.013689368031919003, + 0.028448086231946945, + 0.013827739283442497, + 0.007244945969432592, + 0.004275599494576454, + 0.006487635895609856, + 0.0034894845448434353, + 0.0046105533838272095, + 0.006892066914588213, + 0.007047394756227732, + 0.007556939497590065, + 0.00905944686383009, + 0.01250777393579483, + 0.012455056421458721, + 0.010104313492774963, + 0.0154457101598382, + 0.01544970739632845, + 0.011142418719828129, + 0.012337875552475452, + 0.008218148723244667, + 0.005625119432806969, + 0.012684959918260574, + 0.013077601790428162, + 0.005898676346987486, + 0.008546793833374977, + 0.006684788037091494, + 0.007085232529789209, + 0.011012591421604156, + 0.009713999927043915, + 0.011769214645028114, + 0.007935942150652409, + 0.021352075040340424, + 0.02049248106777668, + 0.016946619376540184, + 0.019211528822779655, + 0.015231302939355373, + 0.014271829277276993, + 0.05431295558810234, + 0.02760288305580616, + 0.021094614639878273, + 0.025722093880176544, + 0.03250560536980629, + 0.030581248924136162 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 8, + 4 + ], + "selected_reward": -0.004657373297959566 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a vast, cathedral-like library with towering bookshelves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera positioned at an extremely low angle looking up at the shelves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the bookshelves and books transform from solid wood/leather into a black, viscous, gooey mass?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene rendered in black-and-white with a single light source casting long shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the gooey mass flow down the shelves and form a shimmering pool on the dark marble floor?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.2521766126155853 + ], + "opposite_logprob_tokens": [ + -1.5021766424179077 + ], + "logprob_reward": -0.2521766126155853, + "yes_no_margin_reward": 1.25, + "token_credit_reward": -0.2521766126155853, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0832400843501091, + 0.14621177315711975, + 0.03515579178929329, + 0.09894032031297684, + 0.032066863030195236, + 0.1537702977657318, + 0.15058156847953796, + 0.06939224898815155, + 0.1678260862827301, + 0.06281491369009018 + ], + "credit_map": [ + 0.00956045463681221, + 0.009858147241175175, + 0.015295688062906265, + 0.009026129730045795, + 0.011567488312721252, + 0.008506279438734055, + 0.010061198845505714, + 0.009364709258079529, + 0.012422005645930767, + 0.014518658630549908, + 0.00918625108897686, + 0.0160042867064476, + 0.020706769078969955, + 0.013440091162919998, + 0.028718460351228714, + 0.03121527098119259, + 0.004298107232898474, + 0.0054866354912519455, + 0.002659376710653305, + 0.0033294991590082645, + 0.005322434473782778, + 0.005834993440657854, + 0.003419222543016076, + 0.004805524367839098, + 0.009678458794951439, + 0.019443165510892868, + 0.00827945489436388, + 0.010371366515755653, + 0.012996550649404526, + 0.02030736394226551, + 0.009077564813196659, + 0.008786410093307495, + 0.005301730707287788, + 0.0032572827767580748, + 0.003038756549358368, + 0.0022587021812796593, + 0.0027448409236967564, + 0.0035362960770726204, + 0.005144882947206497, + 0.006784375756978989, + 0.022158047184348106, + 0.015467885881662369, + 0.02026892639696598, + 0.01171722449362278, + 0.012688904069364071, + 0.02590443193912506, + 0.015092594549059868, + 0.030472302809357643, + 0.022575167939066887, + 0.01013045385479927, + 0.014879118651151657, + 0.011430871672928333, + 0.02188255824148655, + 0.01664292998611927, + 0.025987859815359116, + 0.027052633464336395, + 0.007997434586286545, + 0.006473960820585489, + 0.009711001999676228, + 0.0123237119987607, + 0.0070758783258497715, + 0.008872760459780693, + 0.007742169313132763, + 0.00919533520936966, + 0.018374668434262276, + 0.010997514240443707, + 0.026431648060679436, + 0.021386634558439255, + 0.014683094806969166, + 0.02280733361840248, + 0.020670229569077492, + 0.03247499093413353, + 0.003645209828391671, + 0.0035325870849192142, + 0.017476893961429596, + 0.008354272693395615, + 0.0055056167766451836, + 0.007116318214684725, + 0.008362486027181149, + 0.008821533992886543 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 5, + 6 + ], + "selected_reward": -0.2521766126155853 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a vast, cathedral-like library with towering bookshelves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera positioned at an extremely low angle looking up at the shelves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the bookshelves and books transform from solid wood/leather into a black, viscous, gooey mass?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene rendered in black-and-white with a single light source casting long shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the gooey mass flow down the shelves and form a shimmering pool on the dark marble floor?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.000814582861494273 + ], + "opposite_logprob_tokens": [ + -7.125814437866211 + ], + "logprob_reward": -0.000814582861494273, + "yes_no_margin_reward": 7.125, + "token_credit_reward": -0.000814582861494273, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07902969419956207, + 0.05124911293387413, + 0.07765214145183563, + 0.04136086255311966, + 0.05123565346002579, + 0.030691826716065407, + 0.06620907038450241, + 0.08413533866405487, + 0.3464936912059784, + 0.1719426065683365 + ], + "credit_map": [ + 0.008211967535316944, + 0.009760916233062744, + 0.009225654415786266, + 0.011115632019937038, + 0.01085473783314228, + 0.010234316810965538, + 0.009362435899674892, + 0.010264032520353794, + 0.0046832263469696045, + 0.006441631820052862, + 0.003940335940569639, + 0.0071283127181231976, + 0.006535885389894247, + 0.006215481553226709, + 0.009320666082203388, + 0.006983574479818344, + 0.004904186353087425, + 0.014626399613916874, + 0.004114131443202496, + 0.004555977415293455, + 0.02438405342400074, + 0.017323223873972893, + 0.0032053026370704174, + 0.004538868088275194, + 0.005153728649020195, + 0.00552251935005188, + 0.0032725001219660044, + 0.0033098289277404547, + 0.008497592993080616, + 0.00632291566580534, + 0.004525679163634777, + 0.004756097216159105, + 0.006782021839171648, + 0.004844969604164362, + 0.005494819488376379, + 0.004621205385774374, + 0.005358475260436535, + 0.006104003172367811, + 0.012719622813165188, + 0.00531053589656949, + 0.0035611731000244617, + 0.0028632094617933035, + 0.00479850172996521, + 0.002503241878002882, + 0.0027836174704134464, + 0.004838454537093639, + 0.005319476593285799, + 0.004024152178317308, + 0.00862139742821455, + 0.007235088851302862, + 0.0075505818240344524, + 0.009570235386490822, + 0.010495693422853947, + 0.007080715615302324, + 0.007500358857214451, + 0.008154992945492268, + 0.008972513489425182, + 0.008652848191559315, + 0.010095478966832161, + 0.019005367532372475, + 0.00856838095933199, + 0.010070500895380974, + 0.009244460612535477, + 0.00952579453587532, + 0.026773974299430847, + 0.027737274765968323, + 0.046943146735429764, + 0.0534578375518322, + 0.028180889785289764, + 0.05333195626735687, + 0.04850349202752113, + 0.061565104871988297, + 0.011142607778310776, + 0.008973506279289722, + 0.04340032860636711, + 0.023738272488117218, + 0.013797194696962833, + 0.020886506885290146, + 0.02812880463898182, + 0.02187538705766201 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 9, + 7 + ], + "selected_reward": -0.000814582861494273 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "static, extremely low-angle wide shot gazes across the vast, towering bookshelves of an ancient, cathedral-like library, their tops disappearing into the vast darkness of a vaulted ceiling. Crafted from dark, polished oak, with intricate Gothic carvings, the shelves are filled with thousands of ancient, leather-bound books of varying sizes. The scene is bathed in a gloomy black-and-white light, with a single, feeble source of light emanating from outside the camera casting long, dramatic shadows and highlighting a layer of fine dust in the still air. The impossible movement begins as the rigid structure of the bookshelf begins to tremble uncomfortably; the solid wood bends and sags as if made of wax. Simultaneously, the ancient books begin to melt, their rigid leather covers bubbling and dissolving into the yellowing pages beneath. The entire mass transforms into a black, viscous, gooey mass that initially drips slowly, then flows in a heavy, relentless stream down the shelf's surface, the inky liquid shimmering in the dim light. These streams cascade onto the dark marble floor below, forming an ever-growing, shimmering pool and filling the frame with a nightmarish atmosphere of silence, surreality, and magnificent decay.", + "seed": 28498, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000101/group_0/index_6886_seed_28498_dp_0_rank_0_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 0, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.12598122982308269, + "vlm_logprob": -0.12598122982308269, + "vlm_yes_no_margin": 3.7749999880790712, + "vlm_token_credit": -0.12598122982308269, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a vast, cathedral-like library with towering bookshelves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera positioned at an extremely low angle looking up at the shelves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the bookshelves and books transform from solid wood/leather into a black, viscous, gooey mass?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene rendered in black-and-white with a single light source casting long shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the gooey mass flow down the shelves and form a shimmering pool on the dark marble floor?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.12847024202346802 + ], + "opposite_logprob_tokens": [ + -2.1284701824188232 + ], + "logprob_reward": -0.12847024202346802, + "yes_no_margin_reward": 2.0, + "token_credit_reward": -0.12847024202346802, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.21862614154815674, + 0.049102798104286194, + 0.003835667623206973, + 0.012643720023334026, + 0.07675745338201523, + 0.024905947968363762, + 0.013081652112305164, + 0.028256254270672798, + 0.5657098889350891, + 0.007080432027578354 + ], + "credit_map": [ + 0.05565788224339485, + 0.0064389342442154884, + 0.013766922056674957, + 0.01242836657911539, + 0.025285717099905014, + 0.009329663589596748, + 0.07383660227060318, + 0.02188204787671566, + 0.017941223457455635, + 0.0019318892154842615, + 0.0026825815439224243, + 0.0014201953308656812, + 0.0011475242208689451, + 0.003459885250777006, + 0.0022707004100084305, + 0.018248800188302994, + 0.0005245457286946476, + 0.00028872230905108154, + 0.0005178667488507926, + 0.0005185704212635756, + 0.0006335924263112247, + 0.0004327210190240294, + 0.00038587121525779366, + 0.0005337774055078626, + 0.0015169442631304264, + 0.0018903360469266772, + 0.0018683961825445294, + 0.0009193015284836292, + 0.0020867150742560625, + 0.0015477911802008748, + 0.0014981176936998963, + 0.001316117704845965, + 0.005198798608034849, + 0.001565565005876124, + 0.029811378568410873, + 0.0018959814915433526, + 0.001432719873264432, + 0.004194069188088179, + 0.0018859575502574444, + 0.0307729821652174, + 0.00184179178904742, + 0.0015083319740369916, + 0.005289854947477579, + 0.002315335674211383, + 0.0016225114231929183, + 0.0030032508075237274, + 0.004537589382380247, + 0.00478728162124753, + 0.0007901402423158288, + 0.0008664228953421116, + 0.004542496521025896, + 0.0020974925719201565, + 0.000664962746668607, + 0.0006819738773629069, + 0.0015854740049690008, + 0.0018526891944929957, + 0.0010495701571926475, + 0.01013559103012085, + 0.00163917220197618, + 0.0015577939338982105, + 0.0056763519532978535, + 0.004814777988940477, + 0.0013019616017118096, + 0.0020810349378734827, + 0.02112925983965397, + 0.06851141899824142, + 0.0149914575740695, + 0.007930267602205276, + 0.18145492672920227, + 0.23978902399539948, + 0.010749390348792076, + 0.02115410938858986, + 0.0009269058355130255, + 0.0014104106230661273, + 0.0004815284046344459, + 0.0003997981548309326, + 0.001210318529047072, + 0.0014929614262655377, + 0.0003580207994673401, + 0.0008004877017810941 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 0, + 4 + ], + "selected_reward": -0.12847024202346802 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a vast, cathedral-like library with towering bookshelves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera positioned at an extremely low angle looking up at the shelves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the bookshelves and books transform from solid wood/leather into a black, viscous, gooey mass?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene rendered in black-and-white with a single light source casting long shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the gooey mass flow down the shelves and form a shimmering pool on the dark marble floor?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07007816433906555 + ], + "opposite_logprob_tokens": [ + -2.695078134536743 + ], + "logprob_reward": -0.07007816433906555, + "yes_no_margin_reward": 2.625, + "token_credit_reward": -0.07007816433906555, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.17086666822433472, + 0.210127592086792, + 0.016308510676026344, + 0.01765693910419941, + 0.021617701277136803, + 0.018549051135778427, + 0.04710088670253754, + 0.056449417024850845, + 0.4243369400501251, + 0.016986314207315445 + ], + "credit_map": [ + 0.03970727324485779, + 0.0055437819100916386, + 0.0122460862621665, + 0.010216420516371727, + 0.02062929980456829, + 0.007162009831517935, + 0.05804767832159996, + 0.0173141211271286, + 0.07577203214168549, + 0.007844485342502594, + 0.010132690891623497, + 0.0058545274659991264, + 0.00430480157956481, + 0.015032324008643627, + 0.007450373377650976, + 0.083736352622509, + 0.0021379778627306223, + 0.0018663732334971428, + 0.0015097978757694364, + 0.002630820032209158, + 0.002496252069249749, + 0.0019428277155384421, + 0.0020798880141228437, + 0.0016445738729089499, + 0.0022513254079967737, + 0.002423718338832259, + 0.0033254565205425024, + 0.0011624312028288841, + 0.0027034722734242678, + 0.002183704636991024, + 0.0015432999935001135, + 0.002063531195744872, + 0.0029468657448887825, + 0.0015207214746624231, + 0.002518778434023261, + 0.002121642930433154, + 0.003397301072254777, + 0.00247128214687109, + 0.003812354989349842, + 0.0028287519235163927, + 0.0012140343897044659, + 0.0017055072821676731, + 0.0035227900370955467, + 0.0015660173958167434, + 0.0018510898808017373, + 0.0028123727533966303, + 0.0025947189424186945, + 0.0032825202215462923, + 0.003440455999225378, + 0.002384401625022292, + 0.01587272249162197, + 0.005807044915854931, + 0.0028276240918785334, + 0.0038453771267086267, + 0.0063359662890434265, + 0.006587288808077574, + 0.004532504826784134, + 0.008933854289352894, + 0.004744257777929306, + 0.005541213322430849, + 0.015220293775200844, + 0.007161748595535755, + 0.004822080954909325, + 0.0054934644140303135, + 0.01586509495973587, + 0.05101272836327553, + 0.0106252022087574, + 0.00656913360580802, + 0.13638779520988464, + 0.17845948040485382, + 0.008529058657586575, + 0.016888460144400597, + 0.0020929512102156878, + 0.0030211317352950573, + 0.0011909854365512729, + 0.0014353902079164982, + 0.002392325783148408, + 0.0031755671370774508, + 0.0010748489294201136, + 0.0026031138841062784 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 1, + 0 + ], + "selected_reward": -0.07007816433906555 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a vast, cathedral-like library with towering bookshelves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera positioned at an extremely low angle looking up at the shelves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the bookshelves and books transform from solid wood/leather into a black, viscous, gooey mass?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene rendered in black-and-white with a single light source casting long shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the gooey mass flow down the shelves and form a shimmering pool on the dark marble floor?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0015211451100185513 + ], + "opposite_logprob_tokens": [ + -6.501521110534668 + ], + "logprob_reward": -0.0015211451100185513, + "yes_no_margin_reward": 6.5, + "token_credit_reward": -0.0015211451100185513, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07596246898174286, + 0.02485152706503868, + 0.00971368607133627, + 0.015220994129776955, + 0.10356730222702026, + 0.023245707154273987, + 0.08255040645599365, + 0.10772991180419922, + 0.5244317650794983, + 0.03272620216012001 + ], + "credit_map": [ + 0.01671014353632927, + 0.002502236980944872, + 0.004901847802102566, + 0.00415749242529273, + 0.008845679461956024, + 0.003949552308768034, + 0.026399480178952217, + 0.008496030233800411, + 0.007705938071012497, + 0.0010045692324638367, + 0.002439243486151099, + 0.0011309789260849357, + 0.0005679056630469859, + 0.0016702223801985383, + 0.0013537673512473702, + 0.008978902362287045, + 0.0014246570644900203, + 0.0010213990462943912, + 0.001146327587775886, + 0.001156229991465807, + 0.0014963027788326144, + 0.0012576582375913858, + 0.0012024076422676444, + 0.0010087031405419111, + 0.0017532309284433722, + 0.0009010653011500835, + 0.002169984858483076, + 0.000669357308652252, + 0.0015214645536616445, + 0.004831917118281126, + 0.0015263446839526296, + 0.0018476289696991444, + 0.007031728979200125, + 0.002119972137734294, + 0.040301479399204254, + 0.0025534657761454582, + 0.001785877626389265, + 0.005408217664808035, + 0.0025065108202397823, + 0.041860055178403854, + 0.0024824959691613913, + 0.002386184874922037, + 0.0034880817402154207, + 0.0022918947506695986, + 0.0030053032096475363, + 0.0037480362225323915, + 0.0033990307711064816, + 0.002444678219035268, + 0.006365838460624218, + 0.008320128545165062, + 0.0188469048589468, + 0.009499650448560715, + 0.010595007799565792, + 0.006945229135453701, + 0.0126491105183959, + 0.009328529238700867, + 0.0030026258900761604, + 0.04737461358308792, + 0.001904274569824338, + 0.0029682698659598827, + 0.03524911403656006, + 0.01161894015967846, + 0.0022767262998968363, + 0.003335351822897792, + 0.019558601081371307, + 0.06249485909938812, + 0.013701548799872398, + 0.007120104972273111, + 0.17037257552146912, + 0.22250337898731232, + 0.009788394905626774, + 0.018892275169491768, + 0.002121824538335204, + 0.002482062904164195, + 0.0015449189813807607, + 0.0011026342399418354, + 0.0119846286252141, + 0.010444597341120243, + 0.0009915896225720644, + 0.0020539427641779184 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 7, + 4 + ], + "selected_reward": -0.0015211451100185513 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a vast, cathedral-like library with towering bookshelves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera positioned at an extremely low angle looking up at the shelves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the bookshelves and books transform from solid wood/leather into a black, viscous, gooey mass?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene rendered in black-and-white with a single light source casting long shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the gooey mass flow down the shelves and form a shimmering pool on the dark marble floor?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.4290228486061096 + ], + "opposite_logprob_tokens": [ + -1.0540227890014648 + ], + "logprob_reward": -0.4290228486061096, + "yes_no_margin_reward": 0.6249999403953552, + "token_credit_reward": -0.4290228486061096, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.010375816375017166, + 0.024272693321108818, + 0.009024256840348244, + 0.029542705044150352, + 0.054007623344659805, + 0.025188522413372993, + 0.08701228350400925, + 0.1891920119524002, + 0.5469785332679749, + 0.024405572563409805 + ], + "credit_map": [ + 0.0017189426580443978, + 0.0010030680568888783, + 0.0013488443801179528, + 0.0010826500365510583, + 0.0008117211400531232, + 0.0007721122237853706, + 0.0024630289990454912, + 0.0011754482984542847, + 0.007617690600454807, + 0.001054229331202805, + 0.0016529950080439448, + 0.0008705923100933433, + 0.0005421555833891034, + 0.001664710114710033, + 0.0013979519717395306, + 0.009472365491092205, + 0.0012432655785232782, + 0.0007252880604937673, + 0.0014308528043329716, + 0.0007361075840890408, + 0.0017726928927004337, + 0.0012354234931990504, + 0.0009630841086618602, + 0.0009175422019325197, + 0.0039619989693164825, + 0.0023546512238681316, + 0.00440038088709116, + 0.0015570031246170402, + 0.004034343175590038, + 0.007828004658222198, + 0.0025879747699946165, + 0.002818346954882145, + 0.003784241620451212, + 0.0011396247427910566, + 0.02083873748779297, + 0.0014297509333118796, + 0.0009216743637807667, + 0.002760691102594137, + 0.001260028569959104, + 0.0218728706240654, + 0.0022631746251136065, + 0.001434877049177885, + 0.005673945881426334, + 0.0022667620796710253, + 0.001794475712813437, + 0.002916039200499654, + 0.003191067837178707, + 0.005648178979754448, + 0.00840417668223381, + 0.011831258423626423, + 0.01073053851723671, + 0.01010200660675764, + 0.014806775376200676, + 0.009570159018039703, + 0.013372885063290596, + 0.008194477297365665, + 0.0038691209629178047, + 0.0748366266489029, + 0.0022759942803531885, + 0.0034396525006741285, + 0.08401452749967575, + 0.015263734385371208, + 0.0026644314639270306, + 0.00282791955396533, + 0.021584633737802505, + 0.06928504258394241, + 0.015962587669491768, + 0.00963997095823288, + 0.16896390914916992, + 0.22909484803676605, + 0.011492070741951466, + 0.020955463871359825, + 0.0014376682229340076, + 0.0057146186009049416, + 0.0009106452926062047, + 0.0012150472030043602, + 0.007568825501948595, + 0.0044851056300103664, + 0.0009729028679430485, + 0.002100757323205471 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 7, + 6 + ], + "selected_reward": -0.4290228486061096 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a vast, cathedral-like library with towering bookshelves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera positioned at an extremely low angle looking up at the shelves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the bookshelves and books transform from solid wood/leather into a black, viscous, gooey mass?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene rendered in black-and-white with a single light source casting long shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the gooey mass flow down the shelves and form a shimmering pool on the dark marble floor?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0008137490367516875 + ], + "opposite_logprob_tokens": [ + -7.125813961029053 + ], + "logprob_reward": -0.0008137490367516875, + "yes_no_margin_reward": 7.125, + "token_credit_reward": -0.0008137490367516875, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.006723128724843264, + 0.00458151288330555, + 0.0037099586334079504, + 0.007101728115230799, + 0.025287790223956108, + 0.013569653034210205, + 0.021655835211277008, + 0.25578707456588745, + 0.5521493554115295, + 0.10943400859832764 + ], + "credit_map": [ + 0.0012435417156666517, + 0.0008376293117180467, + 0.0009262236417271197, + 0.0008124823798425496, + 0.00040810482460074127, + 0.0009877608390524983, + 0.0005706856027245522, + 0.000936700205784291, + 0.0013425617944449186, + 0.0002679573663044721, + 0.0002467096201144159, + 0.00019266076560597867, + 0.00014255153655540198, + 0.00032959814416244626, + 0.0003225026302970946, + 0.0017369710840284824, + 0.0005559811252169311, + 0.00040608522249385715, + 0.000607646768912673, + 0.0005089773912914097, + 0.0003751804761122912, + 0.0003529855457600206, + 0.0004734853282570839, + 0.00042961694998666644, + 0.0007857157615944743, + 0.0006219710921868682, + 0.0011984901502728462, + 0.0003560714249033481, + 0.0008963585714809597, + 0.0020003854297101498, + 0.0004282653098925948, + 0.0008144707535393536, + 0.0016904764343053102, + 0.000547208241187036, + 0.009851361624896526, + 0.0006425883038900793, + 0.0004879343614447862, + 0.0013774444814771414, + 0.0005767758120782673, + 0.010114002041518688, + 0.0015622208593413234, + 0.00117810582742095, + 0.00209479290060699, + 0.0018705545226112008, + 0.0013281157007440925, + 0.0015024433378130198, + 0.002320145023986697, + 0.0017132754437625408, + 0.001180084072984755, + 0.0014313143910840154, + 0.006774049252271652, + 0.00273732072673738, + 0.0018687212141230702, + 0.0011961093405261636, + 0.0036836029030382633, + 0.0027846337761729956, + 0.005265286657959223, + 0.10517808794975281, + 0.0022031357511878014, + 0.004102004691958427, + 0.1132267415523529, + 0.017955301329493523, + 0.003995415288954973, + 0.003861098550260067, + 0.02015990950167179, + 0.06407276540994644, + 0.01394618395715952, + 0.007466757204383612, + 0.18151487410068512, + 0.234783336520195, + 0.010013967752456665, + 0.020191578194499016, + 0.004871700890362263, + 0.025256900116801262, + 0.0026250367518514395, + 0.0027960105799138546, + 0.03786888346076012, + 0.0293120089918375, + 0.0018946581985801458, + 0.0048088026233017445 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 7, + 9 + ], + "selected_reward": -0.0008137490367516875 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "static, extremely low-angle wide shot gazes across the vast, towering bookshelves of an ancient, cathedral-like library, their tops disappearing into the vast darkness of a vaulted ceiling. Crafted from dark, polished oak, with intricate Gothic carvings, the shelves are filled with thousands of ancient, leather-bound books of varying sizes. The scene is bathed in a gloomy black-and-white light, with a single, feeble source of light emanating from outside the camera casting long, dramatic shadows and highlighting a layer of fine dust in the still air. The impossible movement begins as the rigid structure of the bookshelf begins to tremble uncomfortably; the solid wood bends and sags as if made of wax. Simultaneously, the ancient books begin to melt, their rigid leather covers bubbling and dissolving into the yellowing pages beneath. The entire mass transforms into a black, viscous, gooey mass that initially drips slowly, then flows in a heavy, relentless stream down the shelf's surface, the inky liquid shimmering in the dim light. These streams cascade onto the dark marble floor below, forming an ever-growing, shimmering pool and filling the frame with a nightmarish atmosphere of silence, surreality, and magnificent decay.", + "seed": 28499, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000101/group_0/index_6886_seed_28499_dp_1_rank_2_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 0, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.027593023248482496, + "vlm_logprob": -0.027593023248482496, + "vlm_yes_no_margin": 4.925, + "vlm_token_credit": -0.027593023248482496, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a vast, cathedral-like library with towering bookshelves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera positioned at an extremely low angle looking up at the shelves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the bookshelves and books transform from solid wood/leather into a black, viscous, gooey mass?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene rendered in black-and-white with a single light source casting long shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the gooey mass flow down the shelves and form a shimmering pool on the dark marble floor?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.012286923825740814 + ], + "opposite_logprob_tokens": [ + -4.637287139892578 + ], + "logprob_reward": -0.012286923825740814, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.012286923825740814, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1299714297056198, + 0.07487013936042786, + 0.05162363499403, + 0.04826187714934349, + 0.07161330431699753, + 0.03886309266090393, + 0.2676052153110504, + 0.16984570026397705, + 0.03931104764342308, + 0.10803455114364624 + ], + "credit_map": [ + 0.024051588028669357, + 0.013558789156377316, + 0.011979479342699051, + 0.006546551361680031, + 0.024105459451675415, + 0.016397317871451378, + 0.018484950065612793, + 0.014847300946712494, + 0.01220829039812088, + 0.007715641986578703, + 0.0075411107391119, + 0.006327198352664709, + 0.013899854384362698, + 0.011863955296576023, + 0.006699741818010807, + 0.008614351972937584, + 0.00850657932460308, + 0.007119282148778439, + 0.0042935204692184925, + 0.002714098198339343, + 0.0078349644318223, + 0.0067611392587423325, + 0.00680589908733964, + 0.007588157895952463, + 0.005891040433198214, + 0.0052362047135829926, + 0.0018172061536461115, + 0.00373849063180387, + 0.01158987171947956, + 0.0066873361356556416, + 0.006614323239773512, + 0.006687408313155174, + 0.005556988995522261, + 0.005452184472233057, + 0.005981596186757088, + 0.00930213462561369, + 0.009565677493810654, + 0.010262591764330864, + 0.014814071357250214, + 0.010678066872060299, + 0.0031948103569447994, + 0.0029289519879966974, + 0.00461978605017066, + 0.004278198350220919, + 0.005926691927015781, + 0.005553930066525936, + 0.0063374582678079605, + 0.006023271940648556, + 0.00818870309740305, + 0.010988203808665276, + 0.012260799296200275, + 0.022054770961403847, + 0.019559519365429878, + 0.038899555802345276, + 0.08021576702594757, + 0.07543791085481644, + 0.011339174583554268, + 0.019289258867502213, + 0.013200811110436916, + 0.023306559771299362, + 0.033541396260261536, + 0.02501341700553894, + 0.02558501623570919, + 0.01857009530067444, + 0.006281516049057245, + 0.00323928939178586, + 0.002881266875192523, + 0.004607576411217451, + 0.005269899964332581, + 0.005442780442535877, + 0.006169559899717569, + 0.005419161636382341, + 0.018484627828001976, + 0.015343347564339638, + 0.009646556340157986, + 0.015378575772047043, + 0.013554449193179607, + 0.013128070160746574, + 0.010931331664323807, + 0.011567605659365654 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 7, + 0 + ], + "selected_reward": -0.012286923825740814 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a vast, cathedral-like library with towering bookshelves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera positioned at an extremely low angle looking up at the shelves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the bookshelves and books transform from solid wood/leather into a black, viscous, gooey mass?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene rendered in black-and-white with a single light source casting long shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the gooey mass flow down the shelves and form a shimmering pool on the dark marble floor?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.023327788338065147 + ], + "opposite_logprob_tokens": [ + -3.7733278274536133 + ], + "logprob_reward": -0.023327788338065147, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.023327788338065147, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09996368736028671, + 0.07464655488729477, + 0.06492804735898972, + 0.0684654638171196, + 0.09118825197219849, + 0.04979129135608673, + 0.08813811093568802, + 0.17103804647922516, + 0.06557311862707138, + 0.22626732289791107 + ], + "credit_map": [ + 0.01314559206366539, + 0.013834061101078987, + 0.011747979559004307, + 0.0069917794317007065, + 0.01597847230732441, + 0.014041634276509285, + 0.010947265662252903, + 0.013276911340653896, + 0.01209267508238554, + 0.008665556088089943, + 0.00893007405102253, + 0.008414567448198795, + 0.011121386662125587, + 0.011287330649793148, + 0.005673267878592014, + 0.008461697027087212, + 0.013038182631134987, + 0.00688540656119585, + 0.004773094318807125, + 0.004111540503799915, + 0.010445507243275642, + 0.008082452230155468, + 0.00892874225974083, + 0.008663124404847622, + 0.009661123156547546, + 0.005282469559460878, + 0.0034239289816468954, + 0.005430488847196102, + 0.014852646738290787, + 0.009414948523044586, + 0.010624165646731853, + 0.009775688871741295, + 0.009705972857773304, + 0.009200016036629677, + 0.00935544166713953, + 0.00883873924612999, + 0.014060120098292828, + 0.014016712084412575, + 0.013232923112809658, + 0.012778333388268948, + 0.003968174569308758, + 0.0035185315646231174, + 0.005274611990898848, + 0.00528813386335969, + 0.006655413657426834, + 0.007414691150188446, + 0.008892424404621124, + 0.008779313415288925, + 0.003383795265108347, + 0.005281724967062473, + 0.006640021689236164, + 0.006510893814265728, + 0.014439980499446392, + 0.011178207583725452, + 0.02088284306228161, + 0.019820645451545715, + 0.010997087694704533, + 0.01481905858963728, + 0.015950171276926994, + 0.020577766001224518, + 0.03403627872467041, + 0.022472543641924858, + 0.03310296684503555, + 0.01908217929303646, + 0.009492742829024792, + 0.006933046970516443, + 0.005627952516078949, + 0.00771314138546586, + 0.011325289495289326, + 0.010115261189639568, + 0.0075416299514472485, + 0.006824056152254343, + 0.031088057905435562, + 0.029429836198687553, + 0.01931227743625641, + 0.04725584387779236, + 0.02827843278646469, + 0.027039624750614166, + 0.022737715393304825, + 0.021125556901097298 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 7, + 0 + ], + "selected_reward": -0.023327788338065147 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a vast, cathedral-like library with towering bookshelves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera positioned at an extremely low angle looking up at the shelves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the bookshelves and books transform from solid wood/leather into a black, viscous, gooey mass?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene rendered in black-and-white with a single light source casting long shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the gooey mass flow down the shelves and form a shimmering pool on the dark marble floor?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.001348182326182723 + ], + "opposite_logprob_tokens": [ + -6.62634801864624 + ], + "logprob_reward": -0.001348182326182723, + "yes_no_margin_reward": 6.625, + "token_credit_reward": -0.001348182326182723, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.05941281467676163, + 0.04105725884437561, + 0.04183966666460037, + 0.05288230627775192, + 0.05593341588973999, + 0.09518461674451828, + 0.21192903816699982, + 0.17723014950752258, + 0.15995697677135468, + 0.10457369685173035 + ], + "credit_map": [ + 0.007418150547891855, + 0.007812413387000561, + 0.00918169692158699, + 0.005542101804167032, + 0.007824837230145931, + 0.008025200106203556, + 0.007296995725482702, + 0.006311420816928148, + 0.004261589143425226, + 0.004252178128808737, + 0.0046913414262235165, + 0.004080557730048895, + 0.008561195805668831, + 0.00541703961789608, + 0.004815646447241306, + 0.00497771380469203, + 0.005454944912344217, + 0.005125537980347872, + 0.004126979038119316, + 0.002444630488753319, + 0.0071926056407392025, + 0.005111268721520901, + 0.006515111308544874, + 0.005868592299520969, + 0.009886236861348152, + 0.005623909179121256, + 0.0040079038590192795, + 0.0058048078790307045, + 0.008278897032141685, + 0.0059196967631578445, + 0.00526122422888875, + 0.008099636062979698, + 0.004947240464389324, + 0.005461886990815401, + 0.003934680484235287, + 0.0060117547400295734, + 0.00951937586069107, + 0.009367143735289574, + 0.007949979975819588, + 0.008741364814341068, + 0.004583198577165604, + 0.005577792413532734, + 0.0065525006502866745, + 0.009340319782495499, + 0.012549509294331074, + 0.015380876138806343, + 0.02320772409439087, + 0.017992710694670677, + 0.014553332701325417, + 0.013828244060277939, + 0.014796440489590168, + 0.025266241282224655, + 0.017327820882201195, + 0.04143328592181206, + 0.04731355607509613, + 0.037410132586956024, + 0.01265583373606205, + 0.016872717067599297, + 0.015038524754345417, + 0.024983026087284088, + 0.03467072173953056, + 0.023027092218399048, + 0.028228523209691048, + 0.02175373025238514, + 0.02301790565252304, + 0.016965413466095924, + 0.013184502720832825, + 0.011770520359277725, + 0.032336339354515076, + 0.025372633710503578, + 0.017779085785150528, + 0.019530603662133217, + 0.011892538517713547, + 0.014802582561969757, + 0.010342373512685299, + 0.019413288682699203, + 0.016032105311751366, + 0.012581322342157364, + 0.00938000250607729, + 0.01012949738651514 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 7, + 8 + ], + "selected_reward": -0.001348182326182723 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a vast, cathedral-like library with towering bookshelves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera positioned at an extremely low angle looking up at the shelves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the bookshelves and books transform from solid wood/leather into a black, viscous, gooey mass?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene rendered in black-and-white with a single light source casting long shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the gooey mass flow down the shelves and form a shimmering pool on the dark marble floor?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.10036477446556091 + ], + "opposite_logprob_tokens": [ + -2.3503646850585938 + ], + "logprob_reward": -0.10036477446556091, + "yes_no_margin_reward": 2.25, + "token_credit_reward": -0.10036477446556091, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09724542498588562, + 0.0946485847234726, + 0.0840822085738182, + 0.0546686127781868, + 0.04960791394114494, + 0.02770133689045906, + 0.2565717101097107, + 0.07854696363210678, + 0.08682569116353989, + 0.17010153830051422 + ], + "credit_map": [ + 0.014519057236611843, + 0.01109246164560318, + 0.01223547663539648, + 0.011659701354801655, + 0.01928076520562172, + 0.0112239308655262, + 0.009259012527763844, + 0.007975021377205849, + 0.010720187798142433, + 0.015090679749846458, + 0.01433086022734642, + 0.015785755589604378, + 0.009921750985085964, + 0.01071883738040924, + 0.00905657559633255, + 0.009023942053318024, + 0.015022112056612968, + 0.014640336856245995, + 0.005971004255115986, + 0.005783818196505308, + 0.011372747831046581, + 0.010963501408696175, + 0.008381988853216171, + 0.011946706101298332, + 0.009625528939068317, + 0.005764857865869999, + 0.0037190192379057407, + 0.005160567816346884, + 0.008723592385649681, + 0.006544801872223616, + 0.006568435113877058, + 0.008561811409890652, + 0.008997933939099312, + 0.003563320031389594, + 0.004501907154917717, + 0.004145421087741852, + 0.006243420299142599, + 0.005549550522118807, + 0.00983724370598793, + 0.0067691202275455, + 0.0013004379579797387, + 0.0013097695773467422, + 0.001559206284582615, + 0.0023142213467508554, + 0.0034520442131906748, + 0.004708458203822374, + 0.007696703542023897, + 0.005360499490052462, + 0.007578723598271608, + 0.010205683298408985, + 0.011518207378685474, + 0.03170193359255791, + 0.01801694557070732, + 0.03310936689376831, + 0.07751431316137314, + 0.06692655384540558, + 0.006096427328884602, + 0.008890904486179352, + 0.006698303390294313, + 0.012060927227139473, + 0.015042239800095558, + 0.010990376584231853, + 0.0113534452393651, + 0.007414348889142275, + 0.012098844163119793, + 0.009351124987006187, + 0.006510787643492222, + 0.009388911537826061, + 0.014208936132490635, + 0.012770884670317173, + 0.009420143440365791, + 0.01307606603950262, + 0.02059762552380562, + 0.017596660181879997, + 0.014489900320768356, + 0.020162222906947136, + 0.021482884883880615, + 0.025670094415545464, + 0.02287406288087368, + 0.0272280965000391 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 9, + 0 + ], + "selected_reward": -0.10036477446556091 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a vast, cathedral-like library with towering bookshelves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera positioned at an extremely low angle looking up at the shelves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the bookshelves and books transform from solid wood/leather into a black, viscous, gooey mass?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene rendered in black-and-white with a single light source casting long shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the gooey mass flow down the shelves and form a shimmering pool on the dark marble floor?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.00063744728686288 + ], + "opposite_logprob_tokens": [ + -7.375637531280518 + ], + "logprob_reward": -0.00063744728686288, + "yes_no_margin_reward": 7.375, + "token_credit_reward": -0.00063744728686288, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.05757798254489899, + 0.03482677787542343, + 0.030425816774368286, + 0.05131733790040016, + 0.05893547460436821, + 0.03314770758152008, + 0.210381418466568, + 0.06527451425790787, + 0.235471710562706, + 0.222641259431839 + ], + "credit_map": [ + 0.006462708581238985, + 0.007116870954632759, + 0.008333234116435051, + 0.003774309530854225, + 0.011225173249840736, + 0.008754180744290352, + 0.005466023925691843, + 0.006445480976253748, + 0.0045137288980185986, + 0.0035433154553174973, + 0.0035725284833461046, + 0.002684918697923422, + 0.006686751265078783, + 0.005045021418482065, + 0.004600911401212215, + 0.004179601091891527, + 0.0031166905537247658, + 0.004181067459285259, + 0.0018507300410419703, + 0.0017630148213356733, + 0.005255332216620445, + 0.003818053286522627, + 0.005975778680294752, + 0.004465151112526655, + 0.010060208849608898, + 0.003949484787881374, + 0.003818066557869315, + 0.006033353041857481, + 0.00905579887330532, + 0.0060884663835167885, + 0.004968659952282906, + 0.007343298755586147, + 0.0031344785820692778, + 0.0038147554732859135, + 0.00348352431319654, + 0.005011489614844322, + 0.011863736435770988, + 0.009250720962882042, + 0.013736998662352562, + 0.008639768697321415, + 0.0020601688884198666, + 0.0026478287763893604, + 0.002243688330054283, + 0.0029876886401325464, + 0.00429308507591486, + 0.005474378354847431, + 0.008070913143455982, + 0.005369956139475107, + 0.007679307833313942, + 0.009509235620498657, + 0.008269078098237514, + 0.020308414474129677, + 0.01626814343035221, + 0.026197591796517372, + 0.0648665502667427, + 0.05728311836719513, + 0.004908642265945673, + 0.004599299281835556, + 0.006776802707463503, + 0.00731298653408885, + 0.014812473207712173, + 0.009590289555490017, + 0.00997178629040718, + 0.007302233017981052, + 0.02914317138493061, + 0.022400282323360443, + 0.0297932680696249, + 0.028663234785199165, + 0.030523965135216713, + 0.03290509432554245, + 0.03488179296255112, + 0.027160895988345146, + 0.028263889253139496, + 0.023276934400200844, + 0.025674069300293922, + 0.05081603303551674, + 0.018978184089064598, + 0.028354942798614502, + 0.02786417119204998, + 0.01941302977502346 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 9, + 6 + ], + "selected_reward": -0.00063744728686288 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "static, extremely low-angle wide shot gazes across the vast, towering bookshelves of an ancient, cathedral-like library, their tops disappearing into the vast darkness of a vaulted ceiling. Crafted from dark, polished oak, with intricate Gothic carvings, the shelves are filled with thousands of ancient, leather-bound books of varying sizes. The scene is bathed in a gloomy black-and-white light, with a single, feeble source of light emanating from outside the camera casting long, dramatic shadows and highlighting a layer of fine dust in the still air. The impossible movement begins as the rigid structure of the bookshelf begins to tremble uncomfortably; the solid wood bends and sags as if made of wax. Simultaneously, the ancient books begin to melt, their rigid leather covers bubbling and dissolving into the yellowing pages beneath. The entire mass transforms into a black, viscous, gooey mass that initially drips slowly, then flows in a heavy, relentless stream down the shelf's surface, the inky liquid shimmering in the dim light. These streams cascade onto the dark marble floor below, forming an ever-growing, shimmering pool and filling the frame with a nightmarish atmosphere of silence, surreality, and magnificent decay.", + "seed": 28500, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000101/group_0/index_6886_seed_28500_dp_1_rank_2_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 0, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.07593211410567165, + "vlm_logprob": -0.07593211410567165, + "vlm_yes_no_margin": 4.2, + "vlm_token_credit": -0.07593211410567165, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a vast, cathedral-like library with towering bookshelves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera positioned at an extremely low angle looking up at the shelves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the bookshelves and books transform from solid wood/leather into a black, viscous, gooey mass?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene rendered in black-and-white with a single light source casting long shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the gooey mass flow down the shelves and form a shimmering pool on the dark marble floor?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.015037275850772858 + ], + "opposite_logprob_tokens": [ + -4.3900370597839355 + ], + "logprob_reward": -0.015037275850772858, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.015037275850772858, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.5810104608535767, + 0.07307130098342896, + 0.06485355645418167, + 0.03992534801363945, + 0.05762134864926338, + 0.05266110971570015, + 0.0856132060289383, + 0.014980509877204895, + 0.013977601192891598, + 0.016285594552755356 + ], + "credit_map": [ + 0.04844639450311661, + 0.05573331564664841, + 0.06630057096481323, + 0.04239920899271965, + 0.07677870243787766, + 0.12212661653757095, + 0.07893668115139008, + 0.09028894454240799, + 0.005394131410866976, + 0.008070354349911213, + 0.009395644068717957, + 0.008400528691709042, + 0.01122176181524992, + 0.011577845551073551, + 0.010153142735362053, + 0.00885789468884468, + 0.006301395129412413, + 0.006415256764739752, + 0.016118649393320084, + 0.004533673170953989, + 0.006830633617937565, + 0.008233008906245232, + 0.0041993907652795315, + 0.01222155150026083, + 0.004459242802113295, + 0.0031333237420767546, + 0.005588578060269356, + 0.005648110993206501, + 0.005622123368084431, + 0.0050455681048333645, + 0.006260467693209648, + 0.004167931620031595, + 0.006782177370041609, + 0.004046836402267218, + 0.007534919306635857, + 0.010247576050460339, + 0.005578737240284681, + 0.0052422937005758286, + 0.010369927622377872, + 0.007818879559636116, + 0.00828731618821621, + 0.004715246614068747, + 0.004264091141521931, + 0.004375937860459089, + 0.005962878931313753, + 0.008502695709466934, + 0.008852134458720684, + 0.00770080741494894, + 0.004730905406177044, + 0.003038425464183092, + 0.0038991826586425304, + 0.003755509853363037, + 0.0036972176749259233, + 0.03026372566819191, + 0.004594996105879545, + 0.03163323923945427, + 0.002299221698194742, + 0.001505532767623663, + 0.0013669334584847093, + 0.0017015003832057118, + 0.0027193711139261723, + 0.00177850853651762, + 0.001728143310174346, + 0.0018812976777553558, + 0.0019894761499017477, + 0.001530436216853559, + 0.001613272470422089, + 0.0017436266643926501, + 0.0019851909019052982, + 0.0017728618113324046, + 0.0014670678647235036, + 0.0018756706267595291, + 0.002254795515909791, + 0.0025466002989560366, + 0.0013885751832276583, + 0.0011802577646449208, + 0.002749919891357422, + 0.002143454970791936, + 0.00209649745374918, + 0.0019254933577030897 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 6, + 1 + ], + "selected_reward": -0.015037275850772858 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a vast, cathedral-like library with towering bookshelves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera positioned at an extremely low angle looking up at the shelves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the bookshelves and books transform from solid wood/leather into a black, viscous, gooey mass?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene rendered in black-and-white with a single light source casting long shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the gooey mass flow down the shelves and form a shimmering pool on the dark marble floor?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.2016417533159256 + ], + "opposite_logprob_tokens": [ + -1.7016417980194092 + ], + "logprob_reward": -0.2016417533159256, + "yes_no_margin_reward": 1.5, + "token_credit_reward": -0.2016417533159256, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.388315349817276, + 0.08073017001152039, + 0.044865649193525314, + 0.03504156321287155, + 0.06408688426017761, + 0.08547544479370117, + 0.14949075877666473, + 0.07633870840072632, + 0.04630183055996895, + 0.029353605583310127 + ], + "credit_map": [ + 0.03356584906578064, + 0.04263010993599892, + 0.03788761422038078, + 0.031621962785720825, + 0.059746235609054565, + 0.07481124252080917, + 0.050559308379888535, + 0.057493072003126144, + 0.007922171615064144, + 0.011225446127355099, + 0.009593483991920948, + 0.011142889969050884, + 0.010369433090090752, + 0.012785837054252625, + 0.0096058938652277, + 0.008085017092525959, + 0.00776270404458046, + 0.005036745220422745, + 0.0057906596921384335, + 0.004015868995338678, + 0.005047448445111513, + 0.0065550366416573524, + 0.004311373457312584, + 0.006345818284898996, + 0.004423415753990412, + 0.003054705448448658, + 0.005736885592341423, + 0.004599438514560461, + 0.0044068326242268085, + 0.004837673157453537, + 0.004753075540065765, + 0.0032295407727360725, + 0.00558509910479188, + 0.005266185849905014, + 0.004846588242799044, + 0.0054191178642213345, + 0.008484757505357265, + 0.010739048011600971, + 0.014948805794119835, + 0.00879729725420475, + 0.011598712764680386, + 0.008616325445473194, + 0.008784336969256401, + 0.014462679624557495, + 0.00837738998234272, + 0.010781394317746162, + 0.01191357895731926, + 0.010941039770841599, + 0.016759539023041725, + 0.008637082763016224, + 0.008581452071666718, + 0.011085735633969307, + 0.015299677848815918, + 0.03763504698872566, + 0.015174581669270992, + 0.03631765395402908, + 0.01046417560428381, + 0.011241260915994644, + 0.005027490202337503, + 0.006204601842910051, + 0.015444350428879261, + 0.010029558092355728, + 0.009947656653821468, + 0.007979623973369598, + 0.007889941334724426, + 0.005040970630943775, + 0.004663009196519852, + 0.004113006871193647, + 0.007574018556624651, + 0.005858730059117079, + 0.005033860914409161, + 0.006128298584371805, + 0.004850947298109531, + 0.002409145003184676, + 0.0035370835103094578, + 0.0030400657560676336, + 0.0034519066102802753, + 0.004121125675737858, + 0.003023070516064763, + 0.004920267034322023 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 6, + 5 + ], + "selected_reward": -0.2016417533159256 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a vast, cathedral-like library with towering bookshelves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera positioned at an extremely low angle looking up at the shelves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the bookshelves and books transform from solid wood/leather into a black, viscous, gooey mass?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene rendered in black-and-white with a single light source casting long shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the gooey mass flow down the shelves and form a shimmering pool on the dark marble floor?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0015278107021003962 + ], + "opposite_logprob_tokens": [ + -6.501527786254883 + ], + "logprob_reward": -0.0015278107021003962, + "yes_no_margin_reward": 6.5, + "token_credit_reward": -0.0015278107021003962, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.4813046455383301, + 0.03964456915855408, + 0.029330043122172356, + 0.030228372663259506, + 0.21564561128616333, + 0.04949723556637764, + 0.05774092674255371, + 0.02887674793601036, + 0.027786361053586006, + 0.039945486932992935 + ], + "credit_map": [ + 0.03710697218775749, + 0.052165158092975616, + 0.05072672292590141, + 0.03999817371368408, + 0.07101508229970932, + 0.09323690086603165, + 0.06607940047979355, + 0.07097624987363815, + 0.002929734531790018, + 0.0056835380382835865, + 0.006082116160541773, + 0.004616962745785713, + 0.0047213309444487095, + 0.006564296316355467, + 0.004748018924146891, + 0.004298575688153505, + 0.0032613729126751423, + 0.0026263415347784758, + 0.004474041052162647, + 0.0052414545789361, + 0.0033253561705350876, + 0.004403282422572374, + 0.002643072046339512, + 0.00335512263700366, + 0.0035427799448370934, + 0.0028750489000231028, + 0.0035378129687160254, + 0.005281704943627119, + 0.004276100546121597, + 0.0035336597356945276, + 0.00399819714948535, + 0.003183069173246622, + 0.01873246394097805, + 0.018734563142061234, + 0.021983876824378967, + 0.034352514892816544, + 0.03559635207056999, + 0.022204767912626266, + 0.03839841112494469, + 0.025642678141593933, + 0.006739545613527298, + 0.004205912351608276, + 0.005481631029397249, + 0.004708847496658564, + 0.004943628795444965, + 0.007899167016148567, + 0.00822468288242817, + 0.007293825503438711, + 0.010162794962525368, + 0.005523875821381807, + 0.005452833138406277, + 0.004252381157130003, + 0.006748311221599579, + 0.01003122515976429, + 0.005811644252389669, + 0.009757865220308304, + 0.003617517650127411, + 0.0033512392546981573, + 0.0019838239531964064, + 0.0027889003977179527, + 0.005233594216406345, + 0.0038145629223436117, + 0.0047950600273907185, + 0.0032920511439442635, + 0.004374576732516289, + 0.004045342560857534, + 0.00256277434527874, + 0.002410722430795431, + 0.004783786833286285, + 0.0036509342025965452, + 0.00274365465156734, + 0.003214570926502347, + 0.004488612059503794, + 0.004630933981388807, + 0.004647572059184313, + 0.005760528147220612, + 0.005028926767408848, + 0.004846187774091959, + 0.004159866366535425, + 0.006382861640304327 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 6 + ], + "selected_reward": -0.0015278107021003962 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a vast, cathedral-like library with towering bookshelves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera positioned at an extremely low angle looking up at the shelves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the bookshelves and books transform from solid wood/leather into a black, viscous, gooey mass?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene rendered in black-and-white with a single light source casting long shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the gooey mass flow down the shelves and form a shimmering pool on the dark marble floor?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.16041024029254913 + ], + "opposite_logprob_tokens": [ + -1.9104102849960327 + ], + "logprob_reward": -0.16041024029254913, + "yes_no_margin_reward": 1.75, + "token_credit_reward": -0.16041024029254913, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.30371418595314026, + 0.11206625401973724, + 0.06690084934234619, + 0.07234202325344086, + 0.19699892401695251, + 0.03828289359807968, + 0.1293998658657074, + 0.035563766956329346, + 0.010501466691493988, + 0.0342298224568367 + ], + "credit_map": [ + 0.022877603769302368, + 0.034418392926454544, + 0.03135357052087784, + 0.023502366617321968, + 0.04471142217516899, + 0.061251766979694366, + 0.04301253706216812, + 0.04258650168776512, + 0.008915990591049194, + 0.011653135530650616, + 0.012194476090371609, + 0.011012823320925236, + 0.015018866397440434, + 0.017937980592250824, + 0.019641445949673653, + 0.015691522508859634, + 0.005749003030359745, + 0.005151329096406698, + 0.01747300662100315, + 0.005254287738353014, + 0.006757082883268595, + 0.008575690910220146, + 0.005551878362894058, + 0.012388567440211773, + 0.01214404497295618, + 0.005989462602883577, + 0.010101712308824062, + 0.009050720371305943, + 0.007549213245511055, + 0.009240612387657166, + 0.009311342611908913, + 0.008954913355410099, + 0.019029811024665833, + 0.015231349505484104, + 0.02135545201599598, + 0.026271797716617584, + 0.03483707830309868, + 0.019408194348216057, + 0.037117209285497665, + 0.02374800108373165, + 0.005679941736161709, + 0.0033043590374290943, + 0.0049461377784609795, + 0.0068180980160832405, + 0.0036210750695317984, + 0.00419713044539094, + 0.004476518370211124, + 0.005239629652351141, + 0.006849056109786034, + 0.003263325896114111, + 0.004056247416883707, + 0.005537642631679773, + 0.005164654925465584, + 0.0476360060274601, + 0.005729409866034985, + 0.05116351693868637, + 0.004590494092553854, + 0.0037245387211441994, + 0.003217994002625346, + 0.005365218035876751, + 0.004945665132254362, + 0.004424886777997017, + 0.004948302172124386, + 0.004346667788922787, + 0.0016557332128286362, + 0.001288923085667193, + 0.0010886501986533403, + 0.0014695917489007115, + 0.0012706019915640354, + 0.0013928432017564774, + 0.0010982219828292727, + 0.0012369012692943215, + 0.005304811056703329, + 0.0038716094568371773, + 0.004520160611718893, + 0.0037456604186445475, + 0.0041814749129116535, + 0.00431880634278059, + 0.003208721289411187, + 0.005078574176877737 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 6 + ], + "selected_reward": -0.16041024029254913 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a vast, cathedral-like library with towering bookshelves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera positioned at an extremely low angle looking up at the shelves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the bookshelves and books transform from solid wood/leather into a black, viscous, gooey mass?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene rendered in black-and-white with a single light source casting long shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the gooey mass flow down the shelves and form a shimmering pool on the dark marble floor?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0010434903670102358 + ], + "opposite_logprob_tokens": [ + -6.876043319702148 + ], + "logprob_reward": -0.0010434903670102358, + "yes_no_margin_reward": 6.875, + "token_credit_reward": -0.0010434903670102358, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08859485387802124, + 0.04986437037587166, + 0.04168366268277168, + 0.09725779294967651, + 0.0838048905134201, + 0.0865190178155899, + 0.06517694145441055, + 0.15923868119716644, + 0.08457625657320023, + 0.2432834804058075 + ], + "credit_map": [ + 0.011537441983819008, + 0.00658089853823185, + 0.009275464341044426, + 0.010048218071460724, + 0.00846130307763815, + 0.01585906744003296, + 0.012318684719502926, + 0.014513769187033176, + 0.004357290919870138, + 0.007178901229053736, + 0.00537277664989233, + 0.005503109190613031, + 0.007076445035636425, + 0.007661757059395313, + 0.006836783606559038, + 0.005877310410141945, + 0.0060176970437169075, + 0.003030683845281601, + 0.0073797390796244144, + 0.003185519715771079, + 0.004898419138044119, + 0.005728668067604303, + 0.004127946682274342, + 0.007314987014979124, + 0.0093736806884408, + 0.006002402864396572, + 0.010677224025130272, + 0.013104583136737347, + 0.012327177450060844, + 0.011532221920788288, + 0.02139526978135109, + 0.012845235876739025, + 0.012589339166879654, + 0.005773911252617836, + 0.007781649939715862, + 0.008565095253288746, + 0.013390550389885902, + 0.008327716030180454, + 0.0157642625272274, + 0.011612365022301674, + 0.01313629187643528, + 0.006813253276050091, + 0.018269775435328484, + 0.010868008248507977, + 0.0077797346748411655, + 0.009446068666875362, + 0.008303280919790268, + 0.011902605183422565, + 0.013395609334111214, + 0.006769497878849506, + 0.004798930138349533, + 0.00502220680937171, + 0.008289914578199387, + 0.010083251632750034, + 0.007408295292407274, + 0.009409239515662193, + 0.021153612062335014, + 0.019795406609773636, + 0.014687871560454369, + 0.0152625422924757, + 0.024380957707762718, + 0.01942538470029831, + 0.02459922432899475, + 0.019933676347136497, + 0.01395376306027174, + 0.007540387101471424, + 0.008683751337230206, + 0.007747462019324303, + 0.012153052724897861, + 0.011450470425188541, + 0.007963024079799652, + 0.01508434396237135, + 0.02817249856889248, + 0.026917144656181335, + 0.026580555364489555, + 0.03144212067127228, + 0.02977457456290722, + 0.030793100595474243, + 0.028364643454551697, + 0.04123883321881294 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 7, + 3 + ], + "selected_reward": -0.0010434903670102358 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000101/group_1/index_tensor([3162, 3162])_group_1_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000101/group_1/index_tensor([3162, 3162])_group_1_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..b0c35cf4c7cf0cdaa7f5dfa5de34b6da2be4a202 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000101/group_1/index_tensor([3162, 3162])_group_1_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "In a scene steeped in the aesthetics of a PlayStation 1 game, a crane shot begins in a medium shot, capturing the decisive blow, then slowly and steadily advances, ultimately focusing on a close-up of the victor. The scene depicts a heroic turtle, wearing a signature purple mask, his entire body composed of jagged, angular polygons, his skin covered in a low-resolution green texture. He holds a powerful, retracting stance, his hands tightly gripping a long, low-polygon brown staff. At his feet, his recently defeated opponent falls backward in a stiff, canned animation, his body erupting in a shower of bright, pixelated orange sparks that flicker and quickly fade across the low-resolution concrete floor. The scene is set in a dimly lit industrial area, with buildings clad in a blurry, repetitive brickwork. Polygonal seams and slight texture distortion are evident throughout, and hard shadows cast by a single, cool-toned light source create a strong, retro feel. As the camera moves inexorably closer, the frame gradually tightens, blurring the background until it becomes a tight close-up, completely focused on the resolute facial expression of the hero turtle. His face, composed of simple geometric shapes, his eyes reveal an unwavering determination through the openings of his purple mask, and his mouth is tightly closed into a straight line, perfectly conveying the calm, focused, and powerful atmosphere after a victorious battle.", + "seed": 4004, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000101/group_1/index_3162_seed_4004_dp_2_rank_4_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 1, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.016290284041315318, + "vlm_logprob": -0.016290284041315318, + "vlm_yes_no_margin": 4.35, + "vlm_token_credit": -0.016290284041315318, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a heroic turtle wearing a purple mask with a low-resolution green texture?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the turtle holding a long, low-polygon brown staff in a powerful, retracting stance?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a defeated opponent visible at the turtle's feet, falling backward with pixelated orange sparks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the scene take place in a dimly lit industrial area with blurry, repetitive brickwork buildings?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start as a medium shot and slowly advance to a tight close-up of the turtle's face?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005983179900795221 + ], + "opposite_logprob_tokens": [ + -5.380983352661133 + ], + "logprob_reward": -0.005983179900795221, + "yes_no_margin_reward": 5.375, + "token_credit_reward": -0.005983179900795221, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07360076904296875, + 0.02612549439072609, + 0.04866626113653183, + 0.07311239838600159, + 0.06895950436592102, + 0.04527287557721138, + 0.02437727525830269, + 0.5439455509185791, + 0.042912088334560394, + 0.05302775651216507 + ], + "credit_map": [ + 0.005758069455623627, + 0.00916205532848835, + 0.009295179508626461, + 0.00785754807293415, + 0.013382813893258572, + 0.008675623685121536, + 0.008680292405188084, + 0.010789178311824799, + 0.004416474141180515, + 0.002766032936051488, + 0.0027053579688072205, + 0.0021467639598995447, + 0.0028705967124551535, + 0.0036037710960954428, + 0.0034297555685043335, + 0.004186739679425955, + 0.004601078573614359, + 0.006446885410696268, + 0.007290365640074015, + 0.005071955733001232, + 0.005137893836945295, + 0.004683893173933029, + 0.007823171094059944, + 0.007611013483256102, + 0.0058219837956130505, + 0.00924803875386715, + 0.005741466302424669, + 0.009213174693286419, + 0.009993982501327991, + 0.009300991892814636, + 0.012181427329778671, + 0.011611321941018105, + 0.007894495502114296, + 0.004690530709922314, + 0.012919163331389427, + 0.005836081225425005, + 0.006219550501555204, + 0.014316136948764324, + 0.006409031804651022, + 0.010674513876438141, + 0.005677288863807917, + 0.003344112541526556, + 0.009980529546737671, + 0.004703632555902004, + 0.0034660929813981056, + 0.007351630832999945, + 0.0035937391221523285, + 0.007155843079090118, + 0.003394046099856496, + 0.003298757132142782, + 0.003085772506892681, + 0.0022132149897515774, + 0.003042660653591156, + 0.003976621199399233, + 0.0024656038731336594, + 0.0029005948454141617, + 0.08993270248174667, + 0.09926550835371017, + 0.054862476885318756, + 0.04408706724643707, + 0.06619197875261307, + 0.08078626543283463, + 0.046944353729486465, + 0.06187518313527107, + 0.003467156318947673, + 0.002799761248752475, + 0.005127451848238707, + 0.005307856947183609, + 0.003961462527513504, + 0.006399168632924557, + 0.007408253848552704, + 0.008440972305834293, + 0.005282376892864704, + 0.006804431788623333, + 0.009394979104399681, + 0.005498406942933798, + 0.00950924213975668, + 0.0055703045800328255, + 0.006005092058330774, + 0.004962915554642677 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 0, + 3 + ], + "selected_reward": -0.005983179900795221 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a heroic turtle wearing a purple mask with a low-resolution green texture?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the turtle holding a long, low-polygon brown staff in a powerful, retracting stance?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a defeated opponent visible at the turtle's feet, falling backward with pixelated orange sparks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the scene take place in a dimly lit industrial area with blurry, repetitive brickwork buildings?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start as a medium shot and slowly advance to a tight close-up of the turtle's face?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.026346080005168915 + ], + "opposite_logprob_tokens": [ + -3.65134596824646 + ], + "logprob_reward": -0.026346080005168915, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.026346080005168915, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07279131561517715, + 0.07430953532457352, + 0.05317061021924019, + 0.1562839299440384, + 0.07827369123697281, + 0.15280453860759735, + 0.08518237620592117, + 0.10073521733283997, + 0.09487753361463547, + 0.13157127797603607 + ], + "credit_map": [ + 0.010423585772514343, + 0.01044816430658102, + 0.011479896493256092, + 0.007698791101574898, + 0.009748997166752815, + 0.0069775390438735485, + 0.007203655783087015, + 0.008810683153569698, + 0.006731490138918161, + 0.006446610204875469, + 0.00825040228664875, + 0.00580177828669548, + 0.010036659426987171, + 0.014204750768840313, + 0.011095923371613026, + 0.011741917580366135, + 0.00945490412414074, + 0.0066130138002336025, + 0.005813708063215017, + 0.00513631384819746, + 0.005540636368095875, + 0.00713062658905983, + 0.00552971800789237, + 0.007951687090098858, + 0.012719524092972279, + 0.027155015617609024, + 0.01207215990871191, + 0.016862723976373672, + 0.01916145347058773, + 0.018774554133415222, + 0.02638675644993782, + 0.023151738569140434, + 0.011766011826694012, + 0.005697231274098158, + 0.009100427851080894, + 0.00694007845595479, + 0.00905580073595047, + 0.011295825242996216, + 0.012316743843257427, + 0.012101579457521439, + 0.016311651095747948, + 0.011685291305184364, + 0.033883724361658096, + 0.015862949192523956, + 0.013243685476481915, + 0.02652522549033165, + 0.013107874430716038, + 0.022184140980243683, + 0.011358504183590412, + 0.009336544200778008, + 0.010045739822089672, + 0.007083736825734377, + 0.011121002025902271, + 0.014058124274015427, + 0.011605828069150448, + 0.010572889819741249, + 0.013748771511018276, + 0.022615840658545494, + 0.011373946443200111, + 0.007895149290561676, + 0.013586552813649178, + 0.014213457703590393, + 0.008208206854760647, + 0.00909330416470766, + 0.015101654455065727, + 0.010579274967312813, + 0.01157430186867714, + 0.011102032847702503, + 0.013790979981422424, + 0.013024413958191872, + 0.008469867520034313, + 0.011235000565648079, + 0.015659809112548828, + 0.007862551137804985, + 0.013716624118387699, + 0.010987126268446445, + 0.012913567945361137, + 0.04279732331633568, + 0.011327924206852913, + 0.016306357458233833 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 5, + 9 + ], + "selected_reward": -0.026346080005168915 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a heroic turtle wearing a purple mask with a low-resolution green texture?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the turtle holding a long, low-polygon brown staff in a powerful, retracting stance?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a defeated opponent visible at the turtle's feet, falling backward with pixelated orange sparks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the scene take place in a dimly lit industrial area with blurry, repetitive brickwork buildings?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start as a medium shot and slowly advance to a tight close-up of the turtle's face?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.016059335321187973 + ], + "opposite_logprob_tokens": [ + -4.141059398651123 + ], + "logprob_reward": -0.016059335321187973, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.016059335321187973, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06096993014216423, + 0.08491494506597519, + 0.34424301981925964, + 0.09973184019327164, + 0.08915779739618301, + 0.0670432597398758, + 0.044344183057546616, + 0.08250018209218979, + 0.06886038184165955, + 0.05823444575071335 + ], + "credit_map": [ + 0.00728664081543684, + 0.0052533457055687904, + 0.007831493392586708, + 0.0068434965796768665, + 0.007832765579223633, + 0.012239373289048672, + 0.007079758681356907, + 0.006603057496249676, + 0.00994111131876707, + 0.007705678232014179, + 0.009895545430481434, + 0.0066197654232382774, + 0.013009846210479736, + 0.01490382943302393, + 0.011268935166299343, + 0.011570232920348644, + 0.027015749365091324, + 0.04176315665245056, + 0.05831799656152725, + 0.025378579273819923, + 0.03440992161631584, + 0.0453934408724308, + 0.05680899694561958, + 0.05515517294406891, + 0.009833033196628094, + 0.01272724661976099, + 0.013079805299639702, + 0.009046148508787155, + 0.012317344546318054, + 0.015321954153478146, + 0.011500570923089981, + 0.015905743464827538, + 0.005845988634973764, + 0.008326290175318718, + 0.010224777273833752, + 0.008840969763696194, + 0.013464979827404022, + 0.013412642292678356, + 0.01570112630724907, + 0.013341018930077553, + 0.007831821218132973, + 0.006605497095733881, + 0.011475098319351673, + 0.009541786275804043, + 0.006612409371882677, + 0.008580190129578114, + 0.006328112445771694, + 0.010068344883620739, + 0.004658391699194908, + 0.005724822636693716, + 0.005048800725489855, + 0.003301527351140976, + 0.0057715256698429585, + 0.009257632307708263, + 0.004450925160199404, + 0.0061305551789700985, + 0.014289474114775658, + 0.013906614854931831, + 0.006616278551518917, + 0.006619613151997328, + 0.011685235425829887, + 0.014231255277991295, + 0.006667328532785177, + 0.00848438311368227, + 0.006635990925133228, + 0.008044255897402763, + 0.007199929095804691, + 0.00977657362818718, + 0.008734622970223427, + 0.00961674191057682, + 0.00876579899340868, + 0.010086472146213055, + 0.007632130291312933, + 0.004059354774653912, + 0.006559453438967466, + 0.004433373920619488, + 0.006324636284261942, + 0.01769193634390831, + 0.004911385010927916, + 0.006622173357754946 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 3, + 4 + ], + "selected_reward": -0.016059335321187973 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a heroic turtle wearing a purple mask with a low-resolution green texture?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the turtle holding a long, low-polygon brown staff in a powerful, retracting stance?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a defeated opponent visible at the turtle's feet, falling backward with pixelated orange sparks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the scene take place in a dimly lit industrial area with blurry, repetitive brickwork buildings?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start as a medium shot and slowly advance to a tight close-up of the turtle's face?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02632935903966427 + ], + "opposite_logprob_tokens": [ + -3.651329278945923 + ], + "logprob_reward": -0.02632935903966427, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.02632935903966427, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.054081954061985016, + 0.04651818424463272, + 0.17372964322566986, + 0.3345765769481659, + 0.11019665002822876, + 0.018903888761997223, + 0.03559799864888191, + 0.15820187330245972, + 0.021751511842012405, + 0.04644177109003067 + ], + "credit_map": [ + 0.007176398765295744, + 0.004713321570307016, + 0.006957584526389837, + 0.00536634074524045, + 0.009767413139343262, + 0.008385070599615574, + 0.005726140923798084, + 0.005989676341414452, + 0.005514658521860838, + 0.004218930844217539, + 0.006530360318720341, + 0.00325891375541687, + 0.007152941077947617, + 0.007501691114157438, + 0.006275748368352652, + 0.006064935587346554, + 0.020593907684087753, + 0.023207923397421837, + 0.015434063971042633, + 0.016848161816596985, + 0.01563272252678871, + 0.03912538290023804, + 0.01319865696132183, + 0.029688803479075432, + 0.055555835366249084, + 0.029831623658537865, + 0.0610225684940815, + 0.023428644984960556, + 0.03467923775315285, + 0.04226374253630638, + 0.031241802498698235, + 0.05655311793088913, + 0.018952010199427605, + 0.007128095719963312, + 0.015400957316160202, + 0.007480328902602196, + 0.01034598983824253, + 0.021504342555999756, + 0.009564469568431377, + 0.019820446148514748, + 0.0016049318946897984, + 0.002434699796140194, + 0.003910105209797621, + 0.0020482311956584454, + 0.001851922832429409, + 0.0026542863342911005, + 0.0017050050664693117, + 0.0026947034057229757, + 0.005792330019176006, + 0.0031809345819056034, + 0.0045992471277713776, + 0.0026999867986887693, + 0.004517713561654091, + 0.0061011891812086105, + 0.003981274552643299, + 0.004725322593003511, + 0.020126909017562866, + 0.03887936845421791, + 0.017114175483584404, + 0.011084593832492828, + 0.022267840802669525, + 0.021403953433036804, + 0.011064511723816395, + 0.01626051589846611, + 0.002511262893676758, + 0.0019448564853519201, + 0.0022846716456115246, + 0.002900100080296397, + 0.003003861987963319, + 0.0029899098444730043, + 0.003175741760060191, + 0.002941105281934142, + 0.004130209796130657, + 0.004731021821498871, + 0.008904664777219296, + 0.005127720069140196, + 0.00609486224129796, + 0.00802669208496809, + 0.004444556310772896, + 0.004982038401067257 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 2, + 7 + ], + "selected_reward": -0.02632935903966427 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a heroic turtle wearing a purple mask with a low-resolution green texture?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the turtle holding a long, low-polygon brown staff in a powerful, retracting stance?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a defeated opponent visible at the turtle's feet, falling backward with pixelated orange sparks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the scene take place in a dimly lit industrial area with blurry, repetitive brickwork buildings?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start as a medium shot and slowly advance to a tight close-up of the turtle's face?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.006733465939760208 + ], + "opposite_logprob_tokens": [ + -5.006733417510986 + ], + "logprob_reward": -0.006733465939760208, + "yes_no_margin_reward": 5.0, + "token_credit_reward": -0.006733465939760208, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.041833385825157166, + 0.05616551637649536, + 0.07315132766962051, + 0.07949982583522797, + 0.1407875269651413, + 0.08711463958024979, + 0.09086459875106812, + 0.18884171545505524, + 0.08552709221839905, + 0.1562144160270691 + ], + "credit_map": [ + 0.0048184446059167385, + 0.005076981149613857, + 0.005511217750608921, + 0.004856553860008717, + 0.006898293737322092, + 0.006348776165395975, + 0.004129904787987471, + 0.004193213768303394, + 0.005900619085878134, + 0.005764784757047892, + 0.005553049501031637, + 0.004419878125190735, + 0.006730383727699518, + 0.011477034538984299, + 0.006040643900632858, + 0.010279121808707714, + 0.006641881074756384, + 0.008021299727261066, + 0.007158775348216295, + 0.00550680048763752, + 0.009789126925170422, + 0.013790368102490902, + 0.009241270832717419, + 0.013001806102693081, + 0.007418375927954912, + 0.009670077823102474, + 0.008783357217907906, + 0.00914123747497797, + 0.009684999473392963, + 0.01008643675595522, + 0.011623277328908443, + 0.01309206336736679, + 0.01244955975562334, + 0.008868998847901821, + 0.01905222237110138, + 0.013107393868267536, + 0.018671955913305283, + 0.018070721998810768, + 0.027352385222911835, + 0.023214280605316162, + 0.008702682331204414, + 0.008619082160294056, + 0.01492479257285595, + 0.013086890801787376, + 0.0071748169139027596, + 0.011882632970809937, + 0.010189329273998737, + 0.012534411624073982, + 0.013188259676098824, + 0.012214803136885166, + 0.010659486055374146, + 0.007982376031577587, + 0.013933981768786907, + 0.012665116228163242, + 0.008647415786981583, + 0.011573164723813534, + 0.022121384739875793, + 0.04485365003347397, + 0.016429925337433815, + 0.013850835151970387, + 0.02844330668449402, + 0.028596125543117523, + 0.014990466646850109, + 0.019556038081645966, + 0.010669362731277943, + 0.008832406252622604, + 0.010288488119840622, + 0.015346145257353783, + 0.01023588515818119, + 0.011738897301256657, + 0.010077358223497868, + 0.008338550105690956, + 0.016239752992987633, + 0.011226220056414604, + 0.026262201368808746, + 0.020824270322918892, + 0.014295444823801517, + 0.03866071254014969, + 0.013165709562599659, + 0.015540112741291523 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 9, + 4 + ], + "selected_reward": -0.006733465939760208 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "In a scene steeped in the aesthetics of a PlayStation 1 game, a crane shot begins in a medium shot, capturing the decisive blow, then slowly and steadily advances, ultimately focusing on a close-up of the victor. The scene depicts a heroic turtle, wearing a signature purple mask, his entire body composed of jagged, angular polygons, his skin covered in a low-resolution green texture. He holds a powerful, retracting stance, his hands tightly gripping a long, low-polygon brown staff. At his feet, his recently defeated opponent falls backward in a stiff, canned animation, his body erupting in a shower of bright, pixelated orange sparks that flicker and quickly fade across the low-resolution concrete floor. The scene is set in a dimly lit industrial area, with buildings clad in a blurry, repetitive brickwork. Polygonal seams and slight texture distortion are evident throughout, and hard shadows cast by a single, cool-toned light source create a strong, retro feel. As the camera moves inexorably closer, the frame gradually tightens, blurring the background until it becomes a tight close-up, completely focused on the resolute facial expression of the hero turtle. His face, composed of simple geometric shapes, his eyes reveal an unwavering determination through the openings of his purple mask, and his mouth is tightly closed into a straight line, perfectly conveying the calm, focused, and powerful atmosphere after a victorious battle.", + "seed": 4005, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000101/group_1/index_3162_seed_4005_dp_2_rank_4_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 1, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.025067953392863274, + "vlm_logprob": -0.025067953392863274, + "vlm_yes_no_margin": 3.924999952316284, + "vlm_token_credit": -0.025067953392863274, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a heroic turtle wearing a purple mask with a low-resolution green texture?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the turtle holding a long, low-polygon brown staff in a powerful, retracting stance?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a defeated opponent visible at the turtle's feet, falling backward with pixelated orange sparks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the scene take place in a dimly lit industrial area with blurry, repetitive brickwork buildings?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start as a medium shot and slowly advance to a tight close-up of the turtle's face?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.020682966336607933 + ], + "opposite_logprob_tokens": [ + -4.0206828117370605 + ], + "logprob_reward": -0.020682966336607933, + "yes_no_margin_reward": 3.999999761581421, + "token_credit_reward": -0.020682966336607933, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.04559694230556488, + 0.023935984820127487, + 0.03610435128211975, + 0.23220326006412506, + 0.05537475645542145, + 0.028502317145466805, + 0.23775412142276764, + 0.17192284762859344, + 0.106801837682724, + 0.06180362030863762 + ], + "credit_map": [ + 0.004944982938468456, + 0.005016656126827002, + 0.009388231672346592, + 0.004505674354732037, + 0.005212964490056038, + 0.00827951543033123, + 0.003949131816625595, + 0.004299786873161793, + 0.004849819932132959, + 0.0031465147621929646, + 0.0035398448817431927, + 0.001971353543922305, + 0.0028530897106975317, + 0.0021123706828802824, + 0.003540206700563431, + 0.0019227847224101424, + 0.005020913202315569, + 0.0038513769395649433, + 0.006591595243662596, + 0.003457884071394801, + 0.0036198911257088184, + 0.005418519955128431, + 0.0035463254898786545, + 0.004597845952957869, + 0.03658991679549217, + 0.025642290711402893, + 0.03623460978269577, + 0.032851044088602066, + 0.029941650107502937, + 0.019845500588417053, + 0.019444728270173073, + 0.0316535159945488, + 0.006908579263836145, + 0.006485067307949066, + 0.008176447823643684, + 0.007288322318345308, + 0.006681364960968494, + 0.005980161018669605, + 0.007396336644887924, + 0.006458476185798645, + 0.003794207237660885, + 0.0040202965028584, + 0.005294409114867449, + 0.003484422340989113, + 0.002922602230682969, + 0.00241571175865829, + 0.003273113863542676, + 0.0032975547946989536, + 0.03118106722831726, + 0.030706966295838356, + 0.03194117918610573, + 0.021462399512529373, + 0.041671689599752426, + 0.030298151075839996, + 0.024629881605505943, + 0.02586277388036251, + 0.013897724449634552, + 0.015919355675578117, + 0.013547979295253754, + 0.018271474167704582, + 0.021633410826325417, + 0.023704243823885918, + 0.028646498918533325, + 0.03630216047167778, + 0.007525989785790443, + 0.008882055059075356, + 0.012431030161678791, + 0.01657257415354252, + 0.009989558719098568, + 0.012467960827052593, + 0.020820703357458115, + 0.01811196468770504, + 0.005361597985029221, + 0.005183287430554628, + 0.00583217479288578, + 0.006606753915548325, + 0.004555701278150082, + 0.015766886994242668, + 0.007302447687834501, + 0.011194771155714989 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 3, + 7 + ], + "selected_reward": -0.020682966336607933 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a heroic turtle wearing a purple mask with a low-resolution green texture?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the turtle holding a long, low-polygon brown staff in a powerful, retracting stance?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a defeated opponent visible at the turtle's feet, falling backward with pixelated orange sparks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the scene take place in a dimly lit industrial area with blurry, repetitive brickwork buildings?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start as a medium shot and slowly advance to a tight close-up of the turtle's face?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03370201960206032 + ], + "opposite_logprob_tokens": [ + -3.4087021350860596 + ], + "logprob_reward": -0.03370201960206032, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.03370201960206032, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07109484076499939, + 0.08874615281820297, + 0.06107071414589882, + 0.23492278158664703, + 0.08314819633960724, + 0.07198052853345871, + 0.09148140996694565, + 0.16621264815330505, + 0.10053842514753342, + 0.030804280191659927 + ], + "credit_map": [ + 0.008290868252515793, + 0.009227612987160683, + 0.013815849088132381, + 0.0077759581618011, + 0.010364603251218796, + 0.008682104758918285, + 0.006487345788627863, + 0.0064504933543503284, + 0.011509038507938385, + 0.012162331491708755, + 0.009768493473529816, + 0.012440882623195648, + 0.013318859972059727, + 0.007762037683278322, + 0.013484648428857327, + 0.008299868553876877, + 0.008669080212712288, + 0.005518955644220114, + 0.01367244403809309, + 0.007910500280559063, + 0.005604871548712254, + 0.007855535484850407, + 0.005484358873218298, + 0.006354969926178455, + 0.03248041123151779, + 0.027114396914839745, + 0.038162197917699814, + 0.035316307097673416, + 0.03193141147494316, + 0.01977626420557499, + 0.021762467920780182, + 0.028379319235682487, + 0.009804882109165192, + 0.01212986372411251, + 0.01034556608647108, + 0.010031837038695812, + 0.011088991537690163, + 0.009114286862313747, + 0.012613951228559017, + 0.008018815889954567, + 0.011020960286259651, + 0.0067299483343958855, + 0.010303481481969357, + 0.007235087454319, + 0.010491977445781231, + 0.010644212365150452, + 0.00575305987149477, + 0.009801802225410938, + 0.00852685235440731, + 0.014062969945371151, + 0.011572632938623428, + 0.014952056109905243, + 0.013645880855619907, + 0.009054948575794697, + 0.009857415221631527, + 0.00980865303426981, + 0.011677207425236702, + 0.021335285156965256, + 0.012395631521940231, + 0.022292152047157288, + 0.03168269246816635, + 0.02030005306005478, + 0.024035952985286713, + 0.022493664175271988, + 0.008887588046491146, + 0.0077332085929811, + 0.010194488801062107, + 0.015995794907212257, + 0.009024771861732006, + 0.017884699627757072, + 0.013886597938835621, + 0.016931265592575073, + 0.0024167245719581842, + 0.003127950942143798, + 0.0027970836963504553, + 0.002952542155981064, + 0.0028364474419504404, + 0.0075637297704815865, + 0.0035909865982830524, + 0.0055188145488500595 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 7, + 8 + ], + "selected_reward": -0.03370201960206032 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a heroic turtle wearing a purple mask with a low-resolution green texture?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the turtle holding a long, low-polygon brown staff in a powerful, retracting stance?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a defeated opponent visible at the turtle's feet, falling backward with pixelated orange sparks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the scene take place in a dimly lit industrial area with blurry, repetitive brickwork buildings?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start as a medium shot and slowly advance to a tight close-up of the turtle's face?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.023273982107639313 + ], + "opposite_logprob_tokens": [ + -3.7732739448547363 + ], + "logprob_reward": -0.023273982107639313, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.023273982107639313, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14213652908802032, + 0.076981320977211, + 0.03848102688789368, + 0.10362672060728073, + 0.14896510541439056, + 0.02961788699030876, + 0.09370771050453186, + 0.12047368288040161, + 0.10545375198125839, + 0.1405562460422516 + ], + "credit_map": [ + 0.011473103426396847, + 0.016472872346639633, + 0.029812920838594437, + 0.01692536287009716, + 0.014129911549389362, + 0.02241724357008934, + 0.018322043120861053, + 0.012583059258759022, + 0.009183329530060291, + 0.010079659521579742, + 0.008361409418284893, + 0.00946741458028555, + 0.011945825070142746, + 0.006900759879499674, + 0.01156170479953289, + 0.009481221437454224, + 0.00486860703676939, + 0.00439341738820076, + 0.007317767012864351, + 0.005080754868686199, + 0.0029245803598314524, + 0.005832300521433353, + 0.003516219789162278, + 0.004547379910945892, + 0.021544551476836205, + 0.008756552822887897, + 0.02167164906859398, + 0.010289167985320091, + 0.009455428458750248, + 0.010364634916186333, + 0.00803424697369337, + 0.01351049356162548, + 0.019483471289277077, + 0.013955469243228436, + 0.018780047073960304, + 0.01986137218773365, + 0.016244586557149887, + 0.015618285164237022, + 0.028104547411203384, + 0.01691731996834278, + 0.0038541911635547876, + 0.002813217230141163, + 0.0038723717443645, + 0.004777497611939907, + 0.004307864233851433, + 0.003497052239254117, + 0.0037698745727539062, + 0.0027258179616183043, + 0.01134080346673727, + 0.010799696668982506, + 0.008860931731760502, + 0.01210014894604683, + 0.015356827527284622, + 0.011269824579358101, + 0.011199720203876495, + 0.012779759243130684, + 0.009632265195250511, + 0.011799737811088562, + 0.012029211036860943, + 0.017816023901104927, + 0.015973305329680443, + 0.01680579222738743, + 0.018953433260321617, + 0.017463909462094307, + 0.009786352515220642, + 0.010098010301589966, + 0.0103919031098485, + 0.017357172444462776, + 0.007110234349966049, + 0.021169375628232956, + 0.010736033320426941, + 0.018804673105478287, + 0.019403336569666862, + 0.013397370465099812, + 0.01501344982534647, + 0.014919176697731018, + 0.011915690265595913, + 0.022669775411486626, + 0.020079810172319412, + 0.023157630115747452 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 0, + 9 + ], + "selected_reward": -0.023273982107639313 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a heroic turtle wearing a purple mask with a low-resolution green texture?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the turtle holding a long, low-polygon brown staff in a powerful, retracting stance?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a defeated opponent visible at the turtle's feet, falling backward with pixelated orange sparks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the scene take place in a dimly lit industrial area with blurry, repetitive brickwork buildings?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start as a medium shot and slowly advance to a tight close-up of the turtle's face?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04304490238428116 + ], + "opposite_logprob_tokens": [ + -3.1680448055267334 + ], + "logprob_reward": -0.04304490238428116, + "yes_no_margin_reward": 3.125, + "token_credit_reward": -0.04304490238428116, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1188528835773468, + 0.22821998596191406, + 0.07076504826545715, + 0.12422208487987518, + 0.07348685711622238, + 0.05224473029375076, + 0.11455931514501572, + 0.06905052810907364, + 0.11094161123037338, + 0.0376569926738739 + ], + "credit_map": [ + 0.017774954438209534, + 0.01736639067530632, + 0.02132166363298893, + 0.009322128258645535, + 0.013073299080133438, + 0.01849398948252201, + 0.010907364077866077, + 0.010593081824481487, + 0.033376503735780716, + 0.05068497732281685, + 0.036695633083581924, + 0.026583056896924973, + 0.018736127763986588, + 0.024540342390537262, + 0.020476456731557846, + 0.017126871272921562, + 0.009471259079873562, + 0.015432641841471195, + 0.013692856766283512, + 0.007701409514993429, + 0.006104825064539909, + 0.0076373121701180935, + 0.004973404109477997, + 0.005751336924731731, + 0.023937134072184563, + 0.010488692671060562, + 0.026294533163309097, + 0.012920829467475414, + 0.01180172711610794, + 0.013411745429039001, + 0.008719719015061855, + 0.016647694632411003, + 0.010442052036523819, + 0.009968417696654797, + 0.009638169780373573, + 0.007916289381682873, + 0.008153633214533329, + 0.00858398899435997, + 0.009204632602632046, + 0.009579665027558804, + 0.008088118396699429, + 0.007350454572588205, + 0.006432226859033108, + 0.007480457425117493, + 0.008249850012362003, + 0.005157179199159145, + 0.0049976385198533535, + 0.004488802049309015, + 0.014078744687139988, + 0.01735372468829155, + 0.01215757243335247, + 0.02115366794168949, + 0.012977982871234417, + 0.009848669171333313, + 0.0156566109508276, + 0.01133232843130827, + 0.004267869517207146, + 0.006366025190800428, + 0.005550793372094631, + 0.007025576196610928, + 0.008888348005712032, + 0.014934021979570389, + 0.009929673746228218, + 0.012088220566511154, + 0.008218536153435707, + 0.010130397975444794, + 0.010791519656777382, + 0.018729718402028084, + 0.011975607834756374, + 0.02048368938267231, + 0.013132764026522636, + 0.01747937686741352, + 0.0028270555194467306, + 0.0026165114250034094, + 0.004125781357288361, + 0.004507321398705244, + 0.0028080816846340895, + 0.00878075696527958, + 0.005411387886852026, + 0.0065800962038338184 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 3, + 0 + ], + "selected_reward": -0.04304490238428116 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a heroic turtle wearing a purple mask with a low-resolution green texture?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the turtle holding a long, low-polygon brown staff in a powerful, retracting stance?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a defeated opponent visible at the turtle's feet, falling backward with pixelated orange sparks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the scene take place in a dimly lit industrial area with blurry, repetitive brickwork buildings?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start as a medium shot and slowly advance to a tight close-up of the turtle's face?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004635896533727646 + ], + "opposite_logprob_tokens": [ + -5.379635810852051 + ], + "logprob_reward": -0.004635896533727646, + "yes_no_margin_reward": 5.375, + "token_credit_reward": -0.004635896533727646, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.04858997464179993, + 0.03630530834197998, + 0.048185016959905624, + 0.26831966638565063, + 0.29030221700668335, + 0.11898168176412582, + 0.05940759554505348, + 0.0685565248131752, + 0.04050131514668465, + 0.02085079625248909 + ], + "credit_map": [ + 0.006744679529219866, + 0.005039464682340622, + 0.008945059962570667, + 0.005734962876886129, + 0.0076143513433635235, + 0.005287108942866325, + 0.00493792537599802, + 0.004286422394216061, + 0.007103331852704287, + 0.005516627337783575, + 0.0045381346717476845, + 0.004000932443886995, + 0.004204023629426956, + 0.0037696491926908493, + 0.004228223115205765, + 0.002944385167211294, + 0.0061283293180167675, + 0.005549474619328976, + 0.005788212176412344, + 0.008953684940934181, + 0.004941566381603479, + 0.006211480125784874, + 0.00630377558991313, + 0.004308491945266724, + 0.04560161381959915, + 0.024940330535173416, + 0.05180756747722626, + 0.038860782980918884, + 0.02899918332695961, + 0.02293160930275917, + 0.02349844016134739, + 0.03168012574315071, + 0.031576599925756454, + 0.0349607951939106, + 0.03313818573951721, + 0.03711973875761032, + 0.035740867257118225, + 0.03173409402370453, + 0.045175950974226, + 0.04085596650838852, + 0.019624954089522362, + 0.016303231939673424, + 0.01444964949041605, + 0.01796087995171547, + 0.016680635511875153, + 0.010709201917052269, + 0.012587142176926136, + 0.010665984824299812, + 0.007569488137960434, + 0.004334172699600458, + 0.007720788940787315, + 0.009065094403922558, + 0.007061753887683153, + 0.00775910634547472, + 0.007977823726832867, + 0.0079193701967597, + 0.006506050005555153, + 0.007346338126808405, + 0.006417352706193924, + 0.00832446850836277, + 0.008213169872760773, + 0.00822814367711544, + 0.010932015255093575, + 0.0125889852643013, + 0.0031385819893330336, + 0.00435213465243578, + 0.004096135497093201, + 0.005541038233786821, + 0.004898648243397474, + 0.005464451853185892, + 0.006036417558789253, + 0.0069739073514938354, + 0.0019900063052773476, + 0.0024647489190101624, + 0.002511310623958707, + 0.002064300701022148, + 0.0022462198976427317, + 0.004015893675386906, + 0.0022026856895536184, + 0.0033556311391294003 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 3, + 5 + ], + "selected_reward": -0.004635896533727646 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "In a scene steeped in the aesthetics of a PlayStation 1 game, a crane shot begins in a medium shot, capturing the decisive blow, then slowly and steadily advances, ultimately focusing on a close-up of the victor. The scene depicts a heroic turtle, wearing a signature purple mask, his entire body composed of jagged, angular polygons, his skin covered in a low-resolution green texture. He holds a powerful, retracting stance, his hands tightly gripping a long, low-polygon brown staff. At his feet, his recently defeated opponent falls backward in a stiff, canned animation, his body erupting in a shower of bright, pixelated orange sparks that flicker and quickly fade across the low-resolution concrete floor. The scene is set in a dimly lit industrial area, with buildings clad in a blurry, repetitive brickwork. Polygonal seams and slight texture distortion are evident throughout, and hard shadows cast by a single, cool-toned light source create a strong, retro feel. As the camera moves inexorably closer, the frame gradually tightens, blurring the background until it becomes a tight close-up, completely focused on the resolute facial expression of the hero turtle. His face, composed of simple geometric shapes, his eyes reveal an unwavering determination through the openings of his purple mask, and his mouth is tightly closed into a straight line, perfectly conveying the calm, focused, and powerful atmosphere after a victorious battle.", + "seed": 4006, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000101/group_1/index_3162_seed_4006_dp_3_rank_6_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 1, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.035145122837275264, + "vlm_logprob": -0.035145122837275264, + "vlm_yes_no_margin": 3.75, + "vlm_token_credit": -0.035145122837275264, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a heroic turtle wearing a purple mask with a low-resolution green texture?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the turtle holding a long, low-polygon brown staff in a powerful, retracting stance?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a defeated opponent visible at the turtle's feet, falling backward with pixelated orange sparks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the scene take place in a dimly lit industrial area with blurry, repetitive brickwork buildings?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start as a medium shot and slowly advance to a tight close-up of the turtle's face?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.046720776706933975 + ], + "opposite_logprob_tokens": [ + -3.1717207431793213 + ], + "logprob_reward": -0.046720776706933975, + "yes_no_margin_reward": 3.125, + "token_credit_reward": -0.046720776706933975, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.05004456266760826, + 0.030897408723831177, + 0.03197142109274864, + 0.06492036581039429, + 0.0806618183851242, + 0.09437598288059235, + 0.07785732299089432, + 0.12032314389944077, + 0.30928605794906616, + 0.13966195285320282 + ], + "credit_map": [ + 0.008746636100113392, + 0.0067043728195130825, + 0.007162886671721935, + 0.0066446238197386265, + 0.005559736397117376, + 0.005281155928969383, + 0.005739188753068447, + 0.004205961246043444, + 0.004962291102856398, + 0.005510002840310335, + 0.003347622463479638, + 0.00253407028503716, + 0.004100818652659655, + 0.0028826252091675997, + 0.0035922115202993155, + 0.003967765718698502, + 0.00439946074038744, + 0.003087772522121668, + 0.005431475583463907, + 0.003080722875893116, + 0.0060559469275176525, + 0.003072247840464115, + 0.00404762476682663, + 0.00279616960324347, + 0.011117976158857346, + 0.005199055653065443, + 0.007341314107179642, + 0.00752784451469779, + 0.007377219386398792, + 0.0077692363411188126, + 0.008661012165248394, + 0.009926706552505493, + 0.011477701365947723, + 0.010010392405092716, + 0.010210731066763401, + 0.006209079176187515, + 0.009146089665591717, + 0.014316989108920097, + 0.008043205365538597, + 0.011247629299759865, + 0.011128951795399189, + 0.009504583664238453, + 0.009055346250534058, + 0.009024749509990215, + 0.02016700804233551, + 0.008937244303524494, + 0.013997490517795086, + 0.012560611590743065, + 0.007949396967887878, + 0.009548863396048546, + 0.007477994542568922, + 0.006716847885400057, + 0.015443463809788227, + 0.010712912306189537, + 0.009440234862267971, + 0.010567604564130306, + 0.011870334856212139, + 0.019422873854637146, + 0.011499056592583656, + 0.010083715431392193, + 0.02228478342294693, + 0.013154312036931515, + 0.01473588403314352, + 0.017272179946303368, + 0.026782965287566185, + 0.04895734041929245, + 0.029415542259812355, + 0.03752987086772919, + 0.05355815961956978, + 0.03461747244000435, + 0.03454058617353439, + 0.04388415440917015, + 0.013605737127363682, + 0.02119775302708149, + 0.011151201091706753, + 0.01170437503606081, + 0.01713702082633972, + 0.019484488293528557, + 0.02356928586959839, + 0.021812086924910545 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 9, + 7 + ], + "selected_reward": -0.046720776706933975 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a heroic turtle wearing a purple mask with a low-resolution green texture?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the turtle holding a long, low-polygon brown staff in a powerful, retracting stance?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a defeated opponent visible at the turtle's feet, falling backward with pixelated orange sparks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the scene take place in a dimly lit industrial area with blurry, repetitive brickwork buildings?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start as a medium shot and slowly advance to a tight close-up of the turtle's face?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07899396121501923 + ], + "opposite_logprob_tokens": [ + -2.578994035720825 + ], + "logprob_reward": -0.07899396121501923, + "yes_no_margin_reward": 2.5, + "token_credit_reward": -0.07899396121501923, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1614457219839096, + 0.17637819051742554, + 0.055379774421453476, + 0.03685810789465904, + 0.16177886724472046, + 0.047327328473329544, + 0.07190869748592377, + 0.08921096473932266, + 0.14405706524848938, + 0.055655330419540405 + ], + "credit_map": [ + 0.027180476114153862, + 0.02047811821103096, + 0.023845328018069267, + 0.021825486794114113, + 0.020072724670171738, + 0.018671328201889992, + 0.014236360788345337, + 0.01513588521629572, + 0.030689025297760963, + 0.018145358189940453, + 0.014993146993219852, + 0.017479749396443367, + 0.022002803161740303, + 0.03145631030201912, + 0.02024480141699314, + 0.02136697806417942, + 0.007159566041082144, + 0.005153093487024307, + 0.007298742420971394, + 0.0072378721088171005, + 0.007257041055709124, + 0.007929892279207706, + 0.0082339346408844, + 0.0051096281968057156, + 0.0062973033636808395, + 0.0037924526259303093, + 0.0028952322900295258, + 0.00387569610029459, + 0.00537186348810792, + 0.004135471303015947, + 0.005546101834625006, + 0.004943982232362032, + 0.022147178649902344, + 0.016409141942858696, + 0.02061476930975914, + 0.011945731937885284, + 0.02163059636950493, + 0.029092304408550262, + 0.01710132323205471, + 0.022837791591882706, + 0.006466243881732225, + 0.005153683014214039, + 0.005391324870288372, + 0.00549694849178195, + 0.0071718511171638966, + 0.004794327076524496, + 0.006539992056787014, + 0.006312952376902103, + 0.00664657074958086, + 0.0073906295001506805, + 0.007319879252463579, + 0.005821371451020241, + 0.01465250551700592, + 0.010950259864330292, + 0.00871536135673523, + 0.010412117466330528, + 0.007568970322608948, + 0.010018421337008476, + 0.011384351179003716, + 0.008297051303088665, + 0.015472061932086945, + 0.010842117480933666, + 0.013649939559400082, + 0.011978044174611568, + 0.012514441274106503, + 0.019453389570116997, + 0.012783807702362537, + 0.01249745860695839, + 0.020408356562256813, + 0.018319858238101006, + 0.027118384838104248, + 0.020961347967386246, + 0.0042603048495948315, + 0.008894731290638447, + 0.0051638903096318245, + 0.0070738294161856174, + 0.010635181330144405, + 0.005902326665818691, + 0.0066475109197199345, + 0.007077550981193781 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 4, + 0 + ], + "selected_reward": -0.07899396121501923 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a heroic turtle wearing a purple mask with a low-resolution green texture?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the turtle holding a long, low-polygon brown staff in a powerful, retracting stance?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a defeated opponent visible at the turtle's feet, falling backward with pixelated orange sparks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the scene take place in a dimly lit industrial area with blurry, repetitive brickwork buildings?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start as a medium shot and slowly advance to a tight close-up of the turtle's face?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.011073091067373753 + ], + "opposite_logprob_tokens": [ + -4.511073112487793 + ], + "logprob_reward": -0.011073091067373753, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.011073091067373753, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08401170372962952, + 0.5608865022659302, + 0.04119594395160675, + 0.039673831313848495, + 0.1262560486793518, + 0.03136802464723587, + 0.028062783181667328, + 0.036298442631959915, + 0.02503182925283909, + 0.02721482887864113 + ], + "credit_map": [ + 0.012860162183642387, + 0.011688762344419956, + 0.008607695810496807, + 0.011097499169409275, + 0.012174285016953945, + 0.00895694363862276, + 0.009593695402145386, + 0.00903266854584217, + 0.07653962075710297, + 0.12039650976657867, + 0.05134417116641998, + 0.06356867402791977, + 0.07969145476818085, + 0.05074189230799675, + 0.05808174982666969, + 0.06052246689796448, + 0.004956613760441542, + 0.0043276650831103325, + 0.006066984962671995, + 0.00586276687681675, + 0.00590473460033536, + 0.005324742291122675, + 0.005290077533572912, + 0.003462362103164196, + 0.0082999924197793, + 0.0034243157133460045, + 0.0041792141273617744, + 0.003970609046518803, + 0.0034979337360709906, + 0.004686011932790279, + 0.005699440836906433, + 0.00591631792485714, + 0.015547550283372402, + 0.015299050137400627, + 0.015466428361833096, + 0.00938283372670412, + 0.01863362081348896, + 0.0200737826526165, + 0.013331708498299122, + 0.01852109469473362, + 0.00423079589381814, + 0.002971040550619364, + 0.0029696389101445675, + 0.0032436614856123924, + 0.0046320208348333836, + 0.005225149914622307, + 0.0034594654571264982, + 0.0046362546272575855, + 0.0034248591400682926, + 0.003368859179317951, + 0.0033043536823242903, + 0.00280145276337862, + 0.004154151771217585, + 0.0037364703603088856, + 0.003946790471673012, + 0.003325847676023841, + 0.0033623212948441505, + 0.005756905302405357, + 0.0036222869530320168, + 0.00311476388014853, + 0.005019074771553278, + 0.004168269224464893, + 0.005589326843619347, + 0.00566549738869071, + 0.0019429591484367847, + 0.004222241695970297, + 0.002170875668525696, + 0.0019415414426475763, + 0.0042192367836833, + 0.0032439162023365498, + 0.0036240373738110065, + 0.003667021868750453, + 0.0025264862924814224, + 0.004039610270410776, + 0.0026179421693086624, + 0.003045799909159541, + 0.004207381047308445, + 0.003414900740608573, + 0.0036581612657755613, + 0.0037045483477413654 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 4, + 0 + ], + "selected_reward": -0.011073091067373753 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a heroic turtle wearing a purple mask with a low-resolution green texture?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the turtle holding a long, low-polygon brown staff in a powerful, retracting stance?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a defeated opponent visible at the turtle's feet, falling backward with pixelated orange sparks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the scene take place in a dimly lit industrial area with blurry, repetitive brickwork buildings?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start as a medium shot and slowly advance to a tight close-up of the turtle's face?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.033685650676488876 + ], + "opposite_logprob_tokens": [ + -3.4086856842041016 + ], + "logprob_reward": -0.033685650676488876, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.033685650676488876, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.21827934682369232, + 0.24438221752643585, + 0.20818224549293518, + 0.08427930623292923, + 0.04769636318087578, + 0.04059574753046036, + 0.030778935179114342, + 0.029555417597293854, + 0.04285828396677971, + 0.053392138332128525 + ], + "credit_map": [ + 0.035158511251211166, + 0.033891480416059494, + 0.03240760788321495, + 0.022193854674696922, + 0.022357478737831116, + 0.021481426432728767, + 0.028844444081187248, + 0.021944545209407806, + 0.038876812905073166, + 0.03993576765060425, + 0.030100738629698753, + 0.025179019197821617, + 0.027081802487373352, + 0.03137844055891037, + 0.028751060366630554, + 0.023078572005033493, + 0.037402182817459106, + 0.019954020157456398, + 0.03305308148264885, + 0.022538168355822563, + 0.020661640912294388, + 0.04130471125245094, + 0.015955336391925812, + 0.01731310784816742, + 0.016584156081080437, + 0.005289667751640081, + 0.007574328687041998, + 0.009676269255578518, + 0.008700482547283173, + 0.0118855070322752, + 0.01371214259415865, + 0.010856743901968002, + 0.005784750450402498, + 0.006303231231868267, + 0.004599315579980612, + 0.005722750909626484, + 0.007141302339732647, + 0.005847541149705648, + 0.006813581567257643, + 0.0054838890209794044, + 0.004484270233660936, + 0.004466081038117409, + 0.003036794951185584, + 0.0035113526973873377, + 0.009028174914419651, + 0.004596096929162741, + 0.00579722598195076, + 0.005675751715898514, + 0.002935902215540409, + 0.00393147487193346, + 0.0029280262533575296, + 0.0022680286783725023, + 0.006507208105176687, + 0.004467502236366272, + 0.0035028234124183655, + 0.004237969871610403, + 0.00251014600507915, + 0.0036298574414104223, + 0.0036408784799277782, + 0.002974637085571885, + 0.005112892482429743, + 0.00383851770311594, + 0.0034727249294519424, + 0.0043757641687989235, + 0.0033374850172549486, + 0.006883702706545591, + 0.003962609451264143, + 0.005075371358543634, + 0.008106115274131298, + 0.004940585698932409, + 0.005016416311264038, + 0.005535999312996864, + 0.004710506182163954, + 0.007155308965593576, + 0.004433193244040012, + 0.004579903092235327, + 0.006565687246620655, + 0.007578217424452305, + 0.010108469054102898, + 0.008260855451226234 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 2 + ], + "selected_reward": -0.033685650676488876 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a heroic turtle wearing a purple mask with a low-resolution green texture?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the turtle holding a long, low-polygon brown staff in a powerful, retracting stance?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a defeated opponent visible at the turtle's feet, falling backward with pixelated orange sparks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the scene take place in a dimly lit industrial area with blurry, repetitive brickwork buildings?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start as a medium shot and slowly advance to a tight close-up of the turtle's face?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005252134520560503 + ], + "opposite_logprob_tokens": [ + -5.255252361297607 + ], + "logprob_reward": -0.005252134520560503, + "yes_no_margin_reward": 5.25, + "token_credit_reward": -0.005252134520560503, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14947733283042908, + 0.14292553067207336, + 0.09302621334791183, + 0.06459485739469528, + 0.24630431830883026, + 0.10995321720838547, + 0.05589345470070839, + 0.039958544075489044, + 0.04996175318956375, + 0.047904692590236664 + ], + "credit_map": [ + 0.020357858389616013, + 0.013953262940049171, + 0.019883889704942703, + 0.023164665326476097, + 0.018555449321866035, + 0.020069545134902, + 0.01939271204173565, + 0.01409995649009943, + 0.025116274133324623, + 0.026934653520584106, + 0.019484689459204674, + 0.015454787760972977, + 0.01697728782892227, + 0.01696396805346012, + 0.012414230033755302, + 0.009579644538462162, + 0.013820338994264603, + 0.011718904599547386, + 0.012222153134644032, + 0.009561010636389256, + 0.013666408136487007, + 0.012744114734232426, + 0.011672047898173332, + 0.00762123242020607, + 0.010503701865673065, + 0.005329588893800974, + 0.006010717712342739, + 0.008853642269968987, + 0.009019679389894009, + 0.007863818667829037, + 0.009412392042577267, + 0.0076013170182704926, + 0.0321844108402729, + 0.02526823990046978, + 0.03199675306677818, + 0.019400369375944138, + 0.03450401872396469, + 0.040351852774620056, + 0.028416072949767113, + 0.0341825969517231, + 0.014905380085110664, + 0.00859522819519043, + 0.013405315577983856, + 0.012708913534879684, + 0.017470087856054306, + 0.013297447003424168, + 0.014901737682521343, + 0.014669107273221016, + 0.00830706674605608, + 0.005003225523978472, + 0.007170313503593206, + 0.005591112654656172, + 0.006111626513302326, + 0.008159185759723186, + 0.007710125297307968, + 0.007840796373784542, + 0.004139157477766275, + 0.004788375459611416, + 0.0046850102953612804, + 0.003611968131735921, + 0.005333687178790569, + 0.004415849689394236, + 0.007495667785406113, + 0.005488827358931303, + 0.003478572703897953, + 0.007123107090592384, + 0.0041292705573141575, + 0.006613891571760178, + 0.009058166295289993, + 0.005738012492656708, + 0.00698880897834897, + 0.006831926293671131, + 0.004180161748081446, + 0.007983969524502754, + 0.003989176359027624, + 0.005218728445470333, + 0.009112567640841007, + 0.005669711623340845, + 0.005007064435631037, + 0.006743315607309341 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 0, + 1 + ], + "selected_reward": -0.005252134520560503 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "In a scene steeped in the aesthetics of a PlayStation 1 game, a crane shot begins in a medium shot, capturing the decisive blow, then slowly and steadily advances, ultimately focusing on a close-up of the victor. The scene depicts a heroic turtle, wearing a signature purple mask, his entire body composed of jagged, angular polygons, his skin covered in a low-resolution green texture. He holds a powerful, retracting stance, his hands tightly gripping a long, low-polygon brown staff. At his feet, his recently defeated opponent falls backward in a stiff, canned animation, his body erupting in a shower of bright, pixelated orange sparks that flicker and quickly fade across the low-resolution concrete floor. The scene is set in a dimly lit industrial area, with buildings clad in a blurry, repetitive brickwork. Polygonal seams and slight texture distortion are evident throughout, and hard shadows cast by a single, cool-toned light source create a strong, retro feel. As the camera moves inexorably closer, the frame gradually tightens, blurring the background until it becomes a tight close-up, completely focused on the resolute facial expression of the hero turtle. His face, composed of simple geometric shapes, his eyes reveal an unwavering determination through the openings of his purple mask, and his mouth is tightly closed into a straight line, perfectly conveying the calm, focused, and powerful atmosphere after a victorious battle.", + "seed": 4007, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000101/group_1/index_3162_seed_4007_dp_3_rank_6_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 1, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.016853915620595218, + "vlm_logprob": -0.016853915620595218, + "vlm_yes_no_margin": 4.85, + "vlm_token_credit": -0.016853915620595218, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a heroic turtle wearing a purple mask with a low-resolution green texture?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the turtle holding a long, low-polygon brown staff in a powerful, retracting stance?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a defeated opponent visible at the turtle's feet, falling backward with pixelated orange sparks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the scene take place in a dimly lit industrial area with blurry, repetitive brickwork buildings?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start as a medium shot and slowly advance to a tight close-up of the turtle's face?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005318539682775736 + ], + "opposite_logprob_tokens": [ + -6.005318641662598 + ], + "logprob_reward": -0.005318539682775736, + "yes_no_margin_reward": 6.0, + "token_credit_reward": -0.005318539682775736, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08952364325523376, + 0.03208494931459427, + 0.049637701362371445, + 0.09766291081905365, + 0.25666165351867676, + 0.03899892792105675, + 0.1705685555934906, + 0.04108817130327225, + 0.13293901085853577, + 0.09083452075719833 + ], + "credit_map": [ + 0.009452782571315765, + 0.012785164639353752, + 0.01617123931646347, + 0.012921859510242939, + 0.011186476796865463, + 0.009693109430372715, + 0.009336675517261028, + 0.007976330816745758, + 0.004883109591901302, + 0.003445313312113285, + 0.005601714365184307, + 0.0026129051111638546, + 0.00323494547046721, + 0.003944696392863989, + 0.0028318394906818867, + 0.005530426278710365, + 0.008662230335175991, + 0.0033256933093070984, + 0.00784672424197197, + 0.004517669789493084, + 0.0035219050478190184, + 0.014892552979290485, + 0.0038514856714755297, + 0.0030194337014108896, + 0.020106500014662743, + 0.004393452312797308, + 0.015136959962546825, + 0.005950909573584795, + 0.004712222144007683, + 0.033757325261831284, + 0.0053398385643959045, + 0.008265689015388489, + 0.08846581727266312, + 0.00907563790678978, + 0.013270217925310135, + 0.009205601178109646, + 0.007540133316069841, + 0.03357623517513275, + 0.009510522708296776, + 0.08601745963096619, + 0.005292482208460569, + 0.004575800616294146, + 0.004764540120959282, + 0.006181810516864061, + 0.003374603809788823, + 0.004799904767423868, + 0.005164314992725849, + 0.0048454697243869305, + 0.015041941776871681, + 0.020167704671621323, + 0.023606566712260246, + 0.02338811196386814, + 0.030082594603300095, + 0.018487850204110146, + 0.023177163675427437, + 0.01661660708487034, + 0.004241880960762501, + 0.004005913157016039, + 0.004957818426191807, + 0.004866956267505884, + 0.0066519989632070065, + 0.004953180905431509, + 0.004844754468649626, + 0.006565665826201439, + 0.015366634353995323, + 0.016380079090595245, + 0.016717515885829926, + 0.012321267277002335, + 0.01671244390308857, + 0.01955876499414444, + 0.02319810725748539, + 0.012684201821684837, + 0.010946336202323437, + 0.0032717017456889153, + 0.02286187745630741, + 0.013861272484064102, + 0.004664140287786722, + 0.004487977828830481, + 0.00959781650453806, + 0.021143389865756035 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 6, + 8 + ], + "selected_reward": -0.005318539682775736 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a heroic turtle wearing a purple mask with a low-resolution green texture?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the turtle holding a long, low-polygon brown staff in a powerful, retracting stance?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a defeated opponent visible at the turtle's feet, falling backward with pixelated orange sparks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the scene take place in a dimly lit industrial area with blurry, repetitive brickwork buildings?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start as a medium shot and slowly advance to a tight close-up of the turtle's face?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.029803644865751266 + ], + "opposite_logprob_tokens": [ + -3.52980375289917 + ], + "logprob_reward": -0.029803644865751266, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.029803644865751266, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15955065190792084, + 0.20417895913124084, + 0.2127538025379181, + 0.07585200667381287, + 0.06734839081764221, + 0.05298501253128052, + 0.07266587018966675, + 0.03530263528227806, + 0.0818294957280159, + 0.037533171474933624 + ], + "credit_map": [ + 0.02087445557117462, + 0.022926757112145424, + 0.034276332706213, + 0.018142322078347206, + 0.015485447831451893, + 0.021571779623627663, + 0.012467041611671448, + 0.013806519098579884, + 0.03661402687430382, + 0.020558791235089302, + 0.03387189283967018, + 0.017533984035253525, + 0.028740091249346733, + 0.02244035340845585, + 0.017188476398587227, + 0.02723134309053421, + 0.02762530744075775, + 0.014153941534459591, + 0.029032854363322258, + 0.015342936851084232, + 0.012576695531606674, + 0.05509345978498459, + 0.03844822570681572, + 0.020480377599596977, + 0.016660861670970917, + 0.0035973647609353065, + 0.011248958297073841, + 0.004161238204687834, + 0.0032367114908993244, + 0.02722655050456524, + 0.004522209521383047, + 0.00519810663536191, + 0.018971532583236694, + 0.002927850466221571, + 0.006178510840982199, + 0.004839216358959675, + 0.002691596746444702, + 0.007695004343986511, + 0.003560760524123907, + 0.020483922213315964, + 0.0077674961648881435, + 0.005423737224191427, + 0.007923471741378307, + 0.006429511122405529, + 0.006241968367248774, + 0.006689426489174366, + 0.00578369339928031, + 0.006725708954036236, + 0.006587067618966103, + 0.006542355753481388, + 0.013152510859072208, + 0.01120516937226057, + 0.008546276949346066, + 0.009373738430440426, + 0.009013614617288113, + 0.0082451356574893, + 0.00364699843339622, + 0.002905462635681033, + 0.005037587136030197, + 0.004938655532896519, + 0.004472316708415747, + 0.003565431572496891, + 0.005779777653515339, + 0.004956405609846115, + 0.012645029462873936, + 0.009596801362931728, + 0.009924406185746193, + 0.008465683087706566, + 0.008582833223044872, + 0.012159635312855244, + 0.011634833179414272, + 0.008820271119475365, + 0.004387903027236462, + 0.0018294909968972206, + 0.007916607894003391, + 0.00463388254866004, + 0.0023694110568612814, + 0.002847617259249091, + 0.005352956708520651, + 0.008195302449166775 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 1, + 0 + ], + "selected_reward": -0.029803644865751266 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a heroic turtle wearing a purple mask with a low-resolution green texture?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the turtle holding a long, low-polygon brown staff in a powerful, retracting stance?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a defeated opponent visible at the turtle's feet, falling backward with pixelated orange sparks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the scene take place in a dimly lit industrial area with blurry, repetitive brickwork buildings?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start as a medium shot and slowly advance to a tight close-up of the turtle's face?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0024877809919416904 + ], + "opposite_logprob_tokens": [ + -6.002487659454346 + ], + "logprob_reward": -0.0024877809919416904, + "yes_no_margin_reward": 6.0, + "token_credit_reward": -0.0024877809919416904, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.16794253885746002, + 0.26053521037101746, + 0.05858495086431503, + 0.07950834184885025, + 0.10619590431451797, + 0.021897856146097183, + 0.05675555020570755, + 0.028891613706946373, + 0.11773260682821274, + 0.1019553542137146 + ], + "credit_map": [ + 0.016851065680384636, + 0.009878608398139477, + 0.019886916503310204, + 0.011556568555533886, + 0.011764363385736942, + 0.020693259313702583, + 0.03975750133395195, + 0.037554264068603516, + 0.03072204440832138, + 0.020853837952017784, + 0.044205110520124435, + 0.025069812312722206, + 0.01982014626264572, + 0.030433010309934616, + 0.03399569168686867, + 0.055435582995414734, + 0.010038390755653381, + 0.004374242387712002, + 0.009872401133179665, + 0.003328503342345357, + 0.004200355615466833, + 0.010436447337269783, + 0.008264819160103798, + 0.008069795556366444, + 0.016947805881500244, + 0.0058551961556077, + 0.01262811291962862, + 0.00589302321895957, + 0.004855258855968714, + 0.020853158086538315, + 0.005106158554553986, + 0.007369633298367262, + 0.03344346210360527, + 0.0042880503460764885, + 0.006618355866521597, + 0.004952627699822187, + 0.003760094754397869, + 0.015498086810112, + 0.004785166587680578, + 0.03285006061196327, + 0.003249660599976778, + 0.0026315231807529926, + 0.0032050679437816143, + 0.0027017162647098303, + 0.00154806487262249, + 0.002762160263955593, + 0.0022403362672775984, + 0.003559327917173505, + 0.00610136054456234, + 0.008510136976838112, + 0.007149162702262402, + 0.00652439147233963, + 0.009969322010874748, + 0.00627515185624361, + 0.006774522364139557, + 0.005451506935060024, + 0.0030154376290738583, + 0.0030518723651766777, + 0.003131947247311473, + 0.005592913366854191, + 0.004308801144361496, + 0.002427788684144616, + 0.004048315808176994, + 0.0033145404886454344, + 0.014860865660011768, + 0.01446797326207161, + 0.019627882167696953, + 0.011141838505864143, + 0.016399728134274483, + 0.012931397184729576, + 0.018770232796669006, + 0.00953269936144352, + 0.013404903002083302, + 0.0030371430329978466, + 0.02575995959341526, + 0.013935226015746593, + 0.003514777170494199, + 0.005368842743337154, + 0.010312151163816452, + 0.026622358709573746 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 8 + ], + "selected_reward": -0.0024877809919416904 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a heroic turtle wearing a purple mask with a low-resolution green texture?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the turtle holding a long, low-polygon brown staff in a powerful, retracting stance?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a defeated opponent visible at the turtle's feet, falling backward with pixelated orange sparks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the scene take place in a dimly lit industrial area with blurry, repetitive brickwork buildings?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start as a medium shot and slowly advance to a tight close-up of the turtle's face?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04304410144686699 + ], + "opposite_logprob_tokens": [ + -3.168044090270996 + ], + "logprob_reward": -0.04304410144686699, + "yes_no_margin_reward": 3.125, + "token_credit_reward": -0.04304410144686699, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12420937418937683, + 0.10175029933452606, + 0.27316445112228394, + 0.17051425576210022, + 0.1415209025144577, + 0.02074671909213066, + 0.04239874705672264, + 0.01927587389945984, + 0.04312669858336449, + 0.06329271197319031 + ], + "credit_map": [ + 0.01599215716123581, + 0.01154671423137188, + 0.013479596935212612, + 0.010455107316374779, + 0.010859296657145023, + 0.012838154099881649, + 0.024378743022680283, + 0.02465960569679737, + 0.016155889257788658, + 0.009904509410262108, + 0.020681532099843025, + 0.009303968399763107, + 0.009372350759804249, + 0.018355581909418106, + 0.006585902534425259, + 0.011390567757189274, + 0.06004592403769493, + 0.017469754442572594, + 0.04622782766819, + 0.018009215593338013, + 0.014056684449315071, + 0.0422075130045414, + 0.04799817129969597, + 0.027149366214871407, + 0.03800807148218155, + 0.012614315375685692, + 0.027446886524558067, + 0.014662213623523712, + 0.010182617232203484, + 0.0374581441283226, + 0.011844717897474766, + 0.01829729601740837, + 0.04721986502408981, + 0.005046560894697905, + 0.007822631858289242, + 0.004334255587309599, + 0.004545752890408039, + 0.023202629759907722, + 0.005072175990790129, + 0.04427703097462654, + 0.0027306429110467434, + 0.0019401450408622622, + 0.001943244249559939, + 0.004409197252243757, + 0.002170947380363941, + 0.0025350044015794992, + 0.0027140681631863117, + 0.0023034699261188507, + 0.004174952395260334, + 0.005266885738819838, + 0.00445434357970953, + 0.00547615485265851, + 0.008246533572673798, + 0.0054590473882853985, + 0.005348528269678354, + 0.003972302656620741, + 0.002082821214571595, + 0.00214094715192914, + 0.0021255488973110914, + 0.0021803141571581364, + 0.0028811574447900057, + 0.002951179165393114, + 0.0021958020515739918, + 0.00271810544654727, + 0.004688306711614132, + 0.004405538085848093, + 0.006209060084074736, + 0.005029045045375824, + 0.0052091521210968494, + 0.004788992926478386, + 0.008338618092238903, + 0.004457985982298851, + 0.010129217989742756, + 0.0031331952195614576, + 0.012442699633538723, + 0.010671829804778099, + 0.003858062671497464, + 0.003789277747273445, + 0.008783425204455853, + 0.010485001839697361 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 3, + 4 + ], + "selected_reward": -0.04304410144686699 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a heroic turtle wearing a purple mask with a low-resolution green texture?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the turtle holding a long, low-polygon brown staff in a powerful, retracting stance?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a defeated opponent visible at the turtle's feet, falling backward with pixelated orange sparks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the scene take place in a dimly lit industrial area with blurry, repetitive brickwork buildings?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start as a medium shot and slowly advance to a tight close-up of the turtle's face?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.003615511115640402 + ], + "opposite_logprob_tokens": [ + -5.628615379333496 + ], + "logprob_reward": -0.003615511115640402, + "yes_no_margin_reward": 5.625, + "token_credit_reward": -0.003615511115640402, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1220960021018982, + 0.050428878515958786, + 0.05098629742860794, + 0.05131090432405472, + 0.10028654336929321, + 0.06595518440008163, + 0.1208977922797203, + 0.14454761147499084, + 0.1254510134458542, + 0.16803978383541107 + ], + "credit_map": [ + 0.01426673773676157, + 0.02025909349322319, + 0.018987398594617844, + 0.01819470338523388, + 0.02116241306066513, + 0.010160377249121666, + 0.00928883720189333, + 0.00977644044905901, + 0.00739588076248765, + 0.005078159272670746, + 0.007936608977615833, + 0.006897930055856705, + 0.004522360395640135, + 0.0036714717280119658, + 0.00727430684491992, + 0.007652162108570337, + 0.009698920883238316, + 0.004454689100384712, + 0.008573634549975395, + 0.005424931645393372, + 0.004240082111209631, + 0.006448991131037474, + 0.007420697249472141, + 0.004724353086203337, + 0.0066855852492153645, + 0.00499140378087759, + 0.008876943960785866, + 0.005548400338739157, + 0.004795759450644255, + 0.01169563364237547, + 0.003917560912668705, + 0.004799618385732174, + 0.026134619489312172, + 0.005885407794266939, + 0.007165121845901012, + 0.005663401447236538, + 0.005755119491368532, + 0.024446716532111168, + 0.0045951963402330875, + 0.020640959963202477, + 0.01037030853331089, + 0.006127572618424892, + 0.00684311194345355, + 0.009447705000638962, + 0.00682363286614418, + 0.00816835556179285, + 0.007119548041373491, + 0.011054949834942818, + 0.00930007454007864, + 0.015651367604732513, + 0.015020059421658516, + 0.018952306360006332, + 0.020957911387085915, + 0.011603279039263725, + 0.01818423718214035, + 0.011228558607399464, + 0.014036658219993114, + 0.011452283710241318, + 0.018322456628084183, + 0.01683172956109047, + 0.01703658141195774, + 0.018838170915842056, + 0.017187388613820076, + 0.030842339619994164, + 0.019731251522898674, + 0.014553321525454521, + 0.016445962712168694, + 0.013850684277713299, + 0.01712643727660179, + 0.01753556914627552, + 0.014975178055465221, + 0.011232607066631317, + 0.02133535034954548, + 0.007629341911524534, + 0.03795420378446579, + 0.02640511840581894, + 0.007223356049507856, + 0.009453949518501759, + 0.020885156467556953, + 0.03715330734848976 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 7, + 8 + ], + "selected_reward": -0.003615511115640402 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000101/group_10/index_tensor([5828, 5828])_group_10_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000101/group_10/index_tensor([5828, 5828])_group_10_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..e3d5de2aa74e1e3c329814da2f2d1cc91db68342 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000101/group_10/index_tensor([5828, 5828])_group_10_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "dynamic, low-altitude drone tracking shot closely follows a steed, jet-black, shimmering satin, from the side and rear. It gallops vigorously across the vast, jagged summit of the ancient Great Wall, braving a fierce, strong wind. The horse's muscles bulge, and each powerful kick sends tiny pebbles and dust from the weathered gray masonry flung backward, carving brief arcs through the air. Its long black mane and tail dance wildly in the strong wind, its nostrils flaring with intense breathing, its eyes revealing an unyielding determination. The Great Wall, as captured, resembles a giant gray dragon, winding its way across rolling green mountains, stretching endlessly into the distance until it disappears beneath a gloomy, cloud-covered skyline. The entire scene is captured in an even, cool, diffuse light, dominated by a palette of somber grays, blacks, and dark greens, creating an atmosphere of desolation, loneliness, and epic power.", + "seed": 31691, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000101/group_10/index_5828_seed_31691_dp_20_rank_40_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 10, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.0714488860219717, + "vlm_logprob": -0.0714488860219717, + "vlm_yes_no_margin": 3.1, + "vlm_token_credit": -0.0714488860219717, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a jet-black steed with shimmering satin fur?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the horse galloping vigorously across the jagged summit of the Great Wall?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera follow the horse from the side and rear in a low-altitude tracking shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the horse's muscles bulging and is dust/pebbles flung backward by its kicks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit with even, cool, diffuse light dominated by somber grays, blacks, and dark greens?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0752096027135849 + ], + "opposite_logprob_tokens": [ + -2.700209617614746 + ], + "logprob_reward": -0.0752096027135849, + "yes_no_margin_reward": 2.625, + "token_credit_reward": -0.0752096027135849, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2215995043516159, + 0.13746589422225952, + 0.04411538690328598, + 0.14207305014133453, + 0.05901813507080078, + 0.23961469531059265, + 0.03200055658817291, + 0.044226065278053284, + 0.04064246267080307, + 0.039244212210178375 + ], + "credit_map": [ + 0.022048618644475937, + 0.027445241808891296, + 0.03333913907408714, + 0.032946228981018066, + 0.029677409678697586, + 0.022877700626850128, + 0.03179007023572922, + 0.021475140005350113, + 0.021208887919783592, + 0.024943647906184196, + 0.015737801790237427, + 0.012868995778262615, + 0.0103071928024292, + 0.03472810983657837, + 0.0074241794645786285, + 0.010247081518173218, + 0.007498165592551231, + 0.006443180609494448, + 0.004506823606789112, + 0.005240598693490028, + 0.0058136675506830215, + 0.005683919880539179, + 0.0041712261736392975, + 0.0047578103840351105, + 0.019945483654737473, + 0.036768149584531784, + 0.012385611422359943, + 0.016736123710870743, + 0.01607239805161953, + 0.020412277430295944, + 0.011593935079872608, + 0.008159077726304531, + 0.007721072528511286, + 0.011561782099306583, + 0.0072976709343492985, + 0.0062280818819999695, + 0.006854753941297531, + 0.007035483606159687, + 0.006565918680280447, + 0.005753377918154001, + 0.03160790726542473, + 0.05677381530404091, + 0.039020389318466187, + 0.01831950433552265, + 0.025804711505770683, + 0.02399544045329094, + 0.025837436318397522, + 0.018255513161420822, + 0.004578935913741589, + 0.004453572910279036, + 0.004287043120712042, + 0.004168747924268246, + 0.0018253716407343745, + 0.0036317480262368917, + 0.004620530642569065, + 0.004434608854353428, + 0.004532990977168083, + 0.005006326828151941, + 0.003891166765242815, + 0.008034255355596542, + 0.006108521018177271, + 0.003926090896129608, + 0.00843100342899561, + 0.004295718856155872, + 0.004582069348543882, + 0.005018604453653097, + 0.004850680474191904, + 0.0073111760430037975, + 0.0032753306441009045, + 0.005408638622611761, + 0.004584289155900478, + 0.0056116776540875435, + 0.00487069645896554, + 0.003063790500164032, + 0.004362535662949085, + 0.0046507567167282104, + 0.004801905248314142, + 0.009339958429336548, + 0.0048353662714362144, + 0.003319205017760396 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 0, + 3 + ], + "selected_reward": -0.0752096027135849 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a jet-black steed with shimmering satin fur?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the horse galloping vigorously across the jagged summit of the Great Wall?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera follow the horse from the side and rear in a low-altitude tracking shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the horse's muscles bulging and is dust/pebbles flung backward by its kicks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit with even, cool, diffuse light dominated by somber grays, blacks, and dark greens?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.2015863060951233 + ], + "opposite_logprob_tokens": [ + -1.7015862464904785 + ], + "logprob_reward": -0.2015863060951233, + "yes_no_margin_reward": 1.5, + "token_credit_reward": -0.2015863060951233, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07934398949146271, + 0.16849824786186218, + 0.052452392876148224, + 0.14035595953464508, + 0.13031524419784546, + 0.10644852370023727, + 0.06195240467786789, + 0.05493085831403732, + 0.13866516947746277, + 0.06703713536262512 + ], + "credit_map": [ + 0.00837645772844553, + 0.01743185892701149, + 0.008011153899133205, + 0.01184636540710926, + 0.010869169607758522, + 0.009970192797482014, + 0.006732642650604248, + 0.006106154061853886, + 0.020491989329457283, + 0.0380401536822319, + 0.016102729365229607, + 0.018992528319358826, + 0.016120582818984985, + 0.03312929347157478, + 0.013411306776106358, + 0.012209688313305378, + 0.0052170949056744576, + 0.008006791584193707, + 0.006881556939333677, + 0.00827736221253872, + 0.006223391275852919, + 0.006350663956254721, + 0.00694512203335762, + 0.004550413694232702, + 0.019844500347971916, + 0.03902224451303482, + 0.011708914302289486, + 0.016112254932522774, + 0.01482432708144188, + 0.020557910203933716, + 0.010714976117014885, + 0.0075708478689193726, + 0.019071616232395172, + 0.02731539122760296, + 0.01772058568894863, + 0.014005378820002079, + 0.01114764716476202, + 0.018540997058153152, + 0.012443107552826405, + 0.010070545598864555, + 0.017437534406781197, + 0.036343518644571304, + 0.020279226824641228, + 0.007532294373959303, + 0.006196652539074421, + 0.00856610108166933, + 0.005962505470961332, + 0.004130697343498468, + 0.014352076686918736, + 0.01367010734975338, + 0.007321518380194902, + 0.008296959102153778, + 0.004140977282077074, + 0.004636706784367561, + 0.005986764095723629, + 0.0035473050083965063, + 0.008345047011971474, + 0.008435587398707867, + 0.006394749507308006, + 0.008891027420759201, + 0.005607327446341515, + 0.004112056456506252, + 0.008349362760782242, + 0.00479570496827364, + 0.016810469329357147, + 0.01553378812968731, + 0.019466737285256386, + 0.01869351975619793, + 0.012163987383246422, + 0.019389266148209572, + 0.012892554514110088, + 0.02371484972536564, + 0.007678487803786993, + 0.007791436742991209, + 0.010991617105901241, + 0.010839560069143772, + 0.007938881404697895, + 0.0068042040802538395, + 0.006501385476440191, + 0.008491567336022854 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 3, + 8 + ], + "selected_reward": -0.2015863060951233 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a jet-black steed with shimmering satin fur?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the horse galloping vigorously across the jagged summit of the Great Wall?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera follow the horse from the side and rear in a low-altitude tracking shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the horse's muscles bulging and is dust/pebbles flung backward by its kicks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit with even, cool, diffuse light dominated by somber grays, blacks, and dark greens?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.008642295375466347 + ], + "opposite_logprob_tokens": [ + -4.758642196655273 + ], + "logprob_reward": -0.008642295375466347, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.008642295375466347, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13870152831077576, + 0.18697026371955872, + 0.06503071635961533, + 0.0834319069981575, + 0.09475275129079819, + 0.11369814723730087, + 0.09114176034927368, + 0.07447685301303864, + 0.07720944285392761, + 0.07458662986755371 + ], + "credit_map": [ + 0.013340736739337444, + 0.0161555428057909, + 0.03218960016965866, + 0.017969001084566116, + 0.01288510113954544, + 0.018216390162706375, + 0.014023744501173496, + 0.013921401463449001, + 0.03318353742361069, + 0.02731221728026867, + 0.023169303312897682, + 0.02127082645893097, + 0.01411698292940855, + 0.03573064133524895, + 0.01472834125161171, + 0.01745840534567833, + 0.007874339818954468, + 0.008066282607614994, + 0.007118756417185068, + 0.01012682355940342, + 0.007445627357810736, + 0.00898069143295288, + 0.009972086176276207, + 0.0054461113177239895, + 0.011867041699588299, + 0.013713785447180271, + 0.007820087485015392, + 0.00977475754916668, + 0.00906850304454565, + 0.014909299090504646, + 0.009048816747963428, + 0.007229619659483433, + 0.007306456565856934, + 0.010957025922834873, + 0.011425092816352844, + 0.011733856983482838, + 0.011895399540662766, + 0.013660155236721039, + 0.0149861304089427, + 0.012788627296686172, + 0.013513429090380669, + 0.01926892250776291, + 0.01699182763695717, + 0.011967760510742664, + 0.0134360883384943, + 0.012746033258736134, + 0.015324879437685013, + 0.010449210181832314, + 0.02518177404999733, + 0.02518131211400032, + 0.007762248162180185, + 0.00788510125130415, + 0.006127653643488884, + 0.005255366209894419, + 0.00882031861692667, + 0.004927978850901127, + 0.008226195350289345, + 0.008697805926203728, + 0.01067977212369442, + 0.013852635398507118, + 0.0071059465408325195, + 0.005764903966337442, + 0.011962903663516045, + 0.008186688646674156, + 0.009410036727786064, + 0.008431433700025082, + 0.00797545537352562, + 0.010588652454316616, + 0.008428490720689297, + 0.014913683757185936, + 0.009634212590754032, + 0.00782747007906437, + 0.008773640729486942, + 0.007944775745272636, + 0.009130438789725304, + 0.009343680925667286, + 0.009119246155023575, + 0.013863369822502136, + 0.009108702652156353, + 0.0073027778416872025 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 5 + ], + "selected_reward": -0.008642295375466347 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a jet-black steed with shimmering satin fur?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the horse galloping vigorously across the jagged summit of the Great Wall?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera follow the horse from the side and rear in a low-altitude tracking shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the horse's muscles bulging and is dust/pebbles flung backward by its kicks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit with even, cool, diffuse light dominated by somber grays, blacks, and dark greens?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03370732069015503 + ], + "opposite_logprob_tokens": [ + -3.4087073802948 + ], + "logprob_reward": -0.03370732069015503, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.03370732069015503, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13788241147994995, + 0.1808709353208542, + 0.054025162011384964, + 0.10854630172252655, + 0.12229561805725098, + 0.16031856834888458, + 0.04161575809121132, + 0.09026682376861572, + 0.05552014708518982, + 0.04865827411413193 + ], + "credit_map": [ + 0.015697946771979332, + 0.020691586658358574, + 0.036035969853401184, + 0.01419264730066061, + 0.010142868384718895, + 0.014902913011610508, + 0.01256790291517973, + 0.013650567270815372, + 0.025423407554626465, + 0.033531900495290756, + 0.02290528640151024, + 0.01628774404525757, + 0.013298653066158295, + 0.04538979381322861, + 0.011310701258480549, + 0.01272345520555973, + 0.011986271478235722, + 0.011162049137055874, + 0.005061660893261433, + 0.006137467920780182, + 0.0048282565549016, + 0.006046159192919731, + 0.003739690873771906, + 0.0050636110827326775, + 0.01449846476316452, + 0.028945697471499443, + 0.009407728910446167, + 0.012206579558551311, + 0.01167329028248787, + 0.016135215759277344, + 0.009798531420528889, + 0.005880787502974272, + 0.016912449151277542, + 0.020581228658556938, + 0.016011884436011314, + 0.0165401641279459, + 0.01227552443742752, + 0.016995031386613846, + 0.013403961434960365, + 0.0095753725618124, + 0.023654600605368614, + 0.04651791602373123, + 0.01805298402905464, + 0.014074251055717468, + 0.015716325491666794, + 0.021211011335253716, + 0.011518285609781742, + 0.00957319512963295, + 0.0077317566610872746, + 0.007703972049057484, + 0.00504680909216404, + 0.005001200828701258, + 0.0038089139852672815, + 0.005473258439451456, + 0.0037932712584733963, + 0.003056575544178486, + 0.011298002675175667, + 0.015263539738953114, + 0.011499103158712387, + 0.013924826867878437, + 0.010057837702333927, + 0.010627157986164093, + 0.009844103828072548, + 0.007752247620373964, + 0.006213872227817774, + 0.0054657417349517345, + 0.007114216219633818, + 0.009631818160414696, + 0.00506718922406435, + 0.006761513184756041, + 0.005454016383737326, + 0.009811781346797943, + 0.008384200744330883, + 0.004321777727454901, + 0.005931224208325148, + 0.007843179628252983, + 0.005475964397192001, + 0.005004404578357935, + 0.006211820058524609, + 0.005485704634338617 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 5, + 0 + ], + "selected_reward": -0.03370732069015503 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a jet-black steed with shimmering satin fur?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the horse galloping vigorously across the jagged summit of the Great Wall?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera follow the horse from the side and rear in a low-altitude tracking shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the horse's muscles bulging and is dust/pebbles flung backward by its kicks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit with even, cool, diffuse light dominated by somber grays, blacks, and dark greens?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.038098905235528946 + ], + "opposite_logprob_tokens": [ + -3.2880988121032715 + ], + "logprob_reward": -0.038098905235528946, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.038098905235528946, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09491481631994247, + 0.056043196469545364, + 0.05822736397385597, + 0.08009492605924606, + 0.06595548242330551, + 0.3443715572357178, + 0.06470464915037155, + 0.1300676167011261, + 0.04109816253185272, + 0.06452224403619766 + ], + "credit_map": [ + 0.010230077430605888, + 0.008599699474871159, + 0.022642318159341812, + 0.013219727203249931, + 0.010073642246425152, + 0.008684813044965267, + 0.01186835765838623, + 0.009596182033419609, + 0.006494679022580385, + 0.014239244163036346, + 0.0056870970875024796, + 0.0068673850037157536, + 0.0045691682025790215, + 0.009833220392465591, + 0.0037164390087127686, + 0.004635961260646582, + 0.013248953968286514, + 0.011857324279844761, + 0.005864923819899559, + 0.00644690589979291, + 0.00491698831319809, + 0.006003551185131073, + 0.004119574557989836, + 0.005769137293100357, + 0.010556480847299099, + 0.017513617873191833, + 0.011491016484797001, + 0.008201823569834232, + 0.007293331436812878, + 0.011273783631622791, + 0.006667259149253368, + 0.007097611203789711, + 0.007405887357890606, + 0.010381550528109074, + 0.009776853956282139, + 0.006088372319936752, + 0.005836669821292162, + 0.009213224984705448, + 0.009587421081960201, + 0.007665496319532394, + 0.05790043994784355, + 0.12129877507686615, + 0.06602571159601212, + 0.026112312451004982, + 0.018636351451277733, + 0.02628409117460251, + 0.017701659351587296, + 0.010412224568426609, + 0.007457282859832048, + 0.008747592568397522, + 0.009272637777030468, + 0.011759724467992783, + 0.0069920397363603115, + 0.005853448994457722, + 0.010397368110716343, + 0.00422455882653594, + 0.022486209869384766, + 0.020298786461353302, + 0.018429744988679886, + 0.02067340724170208, + 0.012217860668897629, + 0.01168268546462059, + 0.010980558581650257, + 0.013298356905579567, + 0.005703079514205456, + 0.004853247199207544, + 0.004091090522706509, + 0.007030259817838669, + 0.003236879361793399, + 0.005176186561584473, + 0.004474456887692213, + 0.006532961968332529, + 0.009799334220588207, + 0.011247694492340088, + 0.0072114719077944756, + 0.006287022959440947, + 0.009376192465424538, + 0.00765601871535182, + 0.007393321022391319, + 0.005551190115511417 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 7, + 0 + ], + "selected_reward": -0.038098905235528946 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "dynamic, low-altitude drone tracking shot closely follows a steed, jet-black, shimmering satin, from the side and rear. It gallops vigorously across the vast, jagged summit of the ancient Great Wall, braving a fierce, strong wind. The horse's muscles bulge, and each powerful kick sends tiny pebbles and dust from the weathered gray masonry flung backward, carving brief arcs through the air. Its long black mane and tail dance wildly in the strong wind, its nostrils flaring with intense breathing, its eyes revealing an unyielding determination. The Great Wall, as captured, resembles a giant gray dragon, winding its way across rolling green mountains, stretching endlessly into the distance until it disappears beneath a gloomy, cloud-covered skyline. The entire scene is captured in an even, cool, diffuse light, dominated by a palette of somber grays, blacks, and dark greens, creating an atmosphere of desolation, loneliness, and epic power.", + "seed": 31692, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000101/group_10/index_5828_seed_31692_dp_20_rank_40_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 10, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.071532316878438, + "vlm_logprob": -0.071532316878438, + "vlm_yes_no_margin": 3.15, + "vlm_token_credit": -0.071532316878438, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a jet-black steed with shimmering satin fur?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the horse galloping vigorously across the jagged summit of the Great Wall?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera follow the horse from the side and rear in a low-altitude tracking shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the horse's muscles bulging and is dust/pebbles flung backward by its kicks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit with even, cool, diffuse light dominated by somber grays, blacks, and dark greens?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.10543405264616013 + ], + "opposite_logprob_tokens": [ + -2.355433940887451 + ], + "logprob_reward": -0.10543405264616013, + "yes_no_margin_reward": 2.25, + "token_credit_reward": -0.10543405264616013, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08184332400560379, + 0.05112573504447937, + 0.06367286294698715, + 0.07896999269723892, + 0.17505069077014923, + 0.026640506461262703, + 0.19257019460201263, + 0.08971282839775085, + 0.1420547217130661, + 0.09835919737815857 + ], + "credit_map": [ + 0.009511207230389118, + 0.011278772726655006, + 0.01229885220527649, + 0.010782945901155472, + 0.010029810480773449, + 0.00782441534101963, + 0.010229489766061306, + 0.00988782662898302, + 0.0049637989141047, + 0.0057366229593753815, + 0.006197414360940456, + 0.0053076171316206455, + 0.011819404549896717, + 0.0056882514618337154, + 0.007008606567978859, + 0.004404018167406321, + 0.006406623870134354, + 0.005626286845654249, + 0.007033596746623516, + 0.012757956050336361, + 0.005183718167245388, + 0.013471446931362152, + 0.006071337964385748, + 0.007121890317648649, + 0.00876894872635603, + 0.00920071266591549, + 0.009903033263981342, + 0.013521593064069748, + 0.009811966679990292, + 0.009119469672441483, + 0.009993123821914196, + 0.00865114200860262, + 0.026476040482521057, + 0.015993976965546608, + 0.037277255207300186, + 0.015945257619023323, + 0.011875485070049763, + 0.03597066178917885, + 0.01124994270503521, + 0.02026207372546196, + 0.002184231299906969, + 0.002935928525403142, + 0.0040754284709692, + 0.00382889318279922, + 0.004863415844738483, + 0.0024637870956212282, + 0.0029712908435612917, + 0.0033175304997712374, + 0.02904582768678665, + 0.03160742670297623, + 0.021397044882178307, + 0.02707904577255249, + 0.020638760179281235, + 0.02445119433104992, + 0.021473221480846405, + 0.016877681016921997, + 0.01007140800356865, + 0.0065895491279661655, + 0.009140018373727798, + 0.008001395501196384, + 0.012540479190647602, + 0.016806501895189285, + 0.01210036501288414, + 0.014463108032941818, + 0.02131708525121212, + 0.010946446098387241, + 0.015482340939342976, + 0.012451539747416973, + 0.018980376422405243, + 0.01653643324971199, + 0.016067948192358017, + 0.030272549018263817, + 0.00873488374054432, + 0.00860434677451849, + 0.010123448446393013, + 0.017215166240930557, + 0.012867296114563942, + 0.008751078508794308, + 0.018233731389045715, + 0.01382924523204565 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 4, + 8 + ], + "selected_reward": -0.10543405264616013 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a jet-black steed with shimmering satin fur?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the horse galloping vigorously across the jagged summit of the Great Wall?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera follow the horse from the side and rear in a low-altitude tracking shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the horse's muscles bulging and is dust/pebbles flung backward by its kicks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit with even, cool, diffuse light dominated by somber grays, blacks, and dark greens?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.17993874847888947 + ], + "opposite_logprob_tokens": [ + -1.804938793182373 + ], + "logprob_reward": -0.17993874847888947, + "yes_no_margin_reward": 1.625, + "token_credit_reward": -0.17993874847888947, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08924607932567596, + 0.03915388882160187, + 0.15868109464645386, + 0.06692974269390106, + 0.17541784048080444, + 0.0637703686952591, + 0.052126381546258926, + 0.12988437712192535, + 0.13378331065177917, + 0.09100685268640518 + ], + "credit_map": [ + 0.009885602630674839, + 0.011222917586565018, + 0.010709220543503761, + 0.010064602829515934, + 0.015213307924568653, + 0.009896642528474331, + 0.010287756100296974, + 0.011966037563979626, + 0.007306661922484636, + 0.007929597049951553, + 0.003715598490089178, + 0.004202579148113728, + 0.003386306343600154, + 0.005000206641852856, + 0.0035699077416211367, + 0.004043037537485361, + 0.014837894588708878, + 0.013644412159919739, + 0.02004249207675457, + 0.033058397471904755, + 0.013050374574959278, + 0.030394868925213814, + 0.015872249379754066, + 0.017780404537916183, + 0.006495041772723198, + 0.006414680276066065, + 0.007106697652488947, + 0.009871297515928745, + 0.007478737737983465, + 0.007166552357375622, + 0.010968214832246304, + 0.011428524740040302, + 0.0209454707801342, + 0.01465096790343523, + 0.04656155779957771, + 0.01676938869059086, + 0.012153753079473972, + 0.022726615890860558, + 0.011458531953394413, + 0.030151575803756714, + 0.00399188045412302, + 0.006338400300592184, + 0.00996010284870863, + 0.008280591107904911, + 0.01373316626995802, + 0.003899865783751011, + 0.011892285197973251, + 0.005674084182828665, + 0.00779971806332469, + 0.00795166939496994, + 0.005582117475569248, + 0.005275324918329716, + 0.004225554410368204, + 0.008763044141232967, + 0.00482613779604435, + 0.0077028232626616955, + 0.012643144465982914, + 0.013604119420051575, + 0.012150615453720093, + 0.010165632702410221, + 0.018202394247055054, + 0.032062679529190063, + 0.015969211235642433, + 0.015086594969034195, + 0.013288108631968498, + 0.014952907338738441, + 0.014755677431821823, + 0.02010928839445114, + 0.014053432270884514, + 0.015359187498688698, + 0.018486378714442253, + 0.022778345271945, + 0.00814883317798376, + 0.008745639584958553, + 0.008703281171619892, + 0.013408460654318333, + 0.011243524961173534, + 0.01367534976452589, + 0.016668254509568214, + 0.01041351817548275 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 2, + 8 + ], + "selected_reward": -0.17993874847888947 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a jet-black steed with shimmering satin fur?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the horse galloping vigorously across the jagged summit of the Great Wall?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera follow the horse from the side and rear in a low-altitude tracking shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the horse's muscles bulging and is dust/pebbles flung backward by its kicks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit with even, cool, diffuse light dominated by somber grays, blacks, and dark greens?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.011082876473665237 + ], + "opposite_logprob_tokens": [ + -4.511082649230957 + ], + "logprob_reward": -0.011082876473665237, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.011082876473665237, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06533799320459366, + 0.05117809399962425, + 0.07234857976436615, + 0.07973656058311462, + 0.15804217755794525, + 0.08128761500120163, + 0.07130811363458633, + 0.07602584362030029, + 0.29681167006492615, + 0.04792332276701927 + ], + "credit_map": [ + 0.007501852232962847, + 0.007665782235562801, + 0.009797798469662666, + 0.008632040582597256, + 0.009403114207088947, + 0.007589161861687899, + 0.00795089639723301, + 0.006797359324991703, + 0.0040313140489161015, + 0.006200537551194429, + 0.0067521617747843266, + 0.00746350921690464, + 0.01015389896929264, + 0.005148385651409626, + 0.006177282426506281, + 0.0052510094828903675, + 0.006478303112089634, + 0.008919215761125088, + 0.010061726905405521, + 0.013311995193362236, + 0.009788366965949535, + 0.006922524888068438, + 0.008388935588300228, + 0.00847751833498478, + 0.006289136130362749, + 0.004462523385882378, + 0.006623593159019947, + 0.010314119048416615, + 0.005291039124131203, + 0.014003816992044449, + 0.006973183713853359, + 0.025779159739613533, + 0.02546173520386219, + 0.01444319263100624, + 0.036624133586883545, + 0.011218663305044174, + 0.008973606862127781, + 0.03318518400192261, + 0.011323746293783188, + 0.016811929643154144, + 0.01343523059040308, + 0.015432639978826046, + 0.007710078731179237, + 0.007609157357364893, + 0.013549285009503365, + 0.009030278772115707, + 0.007870126515626907, + 0.006650822702795267, + 0.011282086372375488, + 0.012396447360515594, + 0.005785888526588678, + 0.00687010120600462, + 0.009474378079175949, + 0.008290812373161316, + 0.007360285613685846, + 0.009848123416304588, + 0.007913722656667233, + 0.006307484116405249, + 0.007798308972269297, + 0.007654746063053608, + 0.00992483738809824, + 0.01549038477241993, + 0.00958330649882555, + 0.011353064328432083, + 0.0529191717505455, + 0.025471236556768417, + 0.03932249918580055, + 0.029467381536960602, + 0.019763367250561714, + 0.027029214426875114, + 0.02189750038087368, + 0.08094131946563721, + 0.00425258232280612, + 0.00404881127178669, + 0.005332778207957745, + 0.0062917727045714855, + 0.006000986322760582, + 0.0054046353325247765, + 0.01020659040659666, + 0.006385172717273235 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 4, + 5 + ], + "selected_reward": -0.011082876473665237 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a jet-black steed with shimmering satin fur?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the horse galloping vigorously across the jagged summit of the Great Wall?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera follow the horse from the side and rear in a low-altitude tracking shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the horse's muscles bulging and is dust/pebbles flung backward by its kicks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit with even, cool, diffuse light dominated by somber grays, blacks, and dark greens?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.012547949329018593 + ], + "opposite_logprob_tokens": [ + -4.387547969818115 + ], + "logprob_reward": -0.012547949329018593, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.012547949329018593, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.17652186751365662, + 0.13089442253112793, + 0.0738406851887703, + 0.06198594719171524, + 0.18314366042613983, + 0.06993428617715836, + 0.05697992444038391, + 0.03250754252076149, + 0.16130560636520386, + 0.052886102348566055 + ], + "credit_map": [ + 0.021298576146364212, + 0.02413908950984478, + 0.024277158081531525, + 0.019905027002096176, + 0.023155633360147476, + 0.018528247252106667, + 0.02488149143755436, + 0.020336640998721123, + 0.01131226122379303, + 0.01255226880311966, + 0.01403135247528553, + 0.016925137490034103, + 0.01916523650288582, + 0.01613015867769718, + 0.01710856519639492, + 0.02366943098604679, + 0.008904695510864258, + 0.00875120609998703, + 0.008678434416651726, + 0.010574047453701496, + 0.007622695993632078, + 0.013388331048190594, + 0.00805783923715353, + 0.007863430306315422, + 0.006045326124876738, + 0.005024086683988571, + 0.005902205128222704, + 0.007083697710186243, + 0.006686098873615265, + 0.009068243205547333, + 0.007723890244960785, + 0.01445239782333374, + 0.03254241868853569, + 0.019329266622662544, + 0.031809002161026, + 0.01451281737536192, + 0.011787905357778072, + 0.04318637773394585, + 0.011477513238787651, + 0.018498331308364868, + 0.006285487674176693, + 0.009266958571970463, + 0.010118146426975727, + 0.009787431918084621, + 0.0093923294916749, + 0.005274582654237747, + 0.012876685708761215, + 0.006932658143341541, + 0.006580972112715244, + 0.006601085886359215, + 0.005788260605186224, + 0.007241013925522566, + 0.009135298430919647, + 0.007459699641913176, + 0.008562983945012093, + 0.005610607098788023, + 0.0025149877183139324, + 0.003320080693811178, + 0.0036678125616163015, + 0.004751625005155802, + 0.003451989032328129, + 0.004715203307569027, + 0.00427119480445981, + 0.005814647767692804, + 0.026917589828372, + 0.013463973999023438, + 0.020401693880558014, + 0.016847040504217148, + 0.011311141774058342, + 0.017656680196523666, + 0.012822948396205902, + 0.04188453406095505, + 0.005568415857851505, + 0.004771839827299118, + 0.0058889081701636314, + 0.006733815185725689, + 0.006019728258252144, + 0.005301603116095066, + 0.011987930163741112, + 0.00661385478451848 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 0, + 8 + ], + "selected_reward": -0.012547949329018593 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a jet-black steed with shimmering satin fur?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the horse galloping vigorously across the jagged summit of the Great Wall?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera follow the horse from the side and rear in a low-altitude tracking shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the horse's muscles bulging and is dust/pebbles flung backward by its kicks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit with even, cool, diffuse light dominated by somber grays, blacks, and dark greens?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04865795746445656 + ], + "opposite_logprob_tokens": [ + -3.0486578941345215 + ], + "logprob_reward": -0.04865795746445656, + "yes_no_margin_reward": 3.0, + "token_credit_reward": -0.04865795746445656, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1499490886926651, + 0.06808070093393326, + 0.048790331929922104, + 0.06586646288633347, + 0.3940645456314087, + 0.04326661676168442, + 0.09825489670038223, + 0.03900101035833359, + 0.04630818963050842, + 0.04641811549663544 + ], + "credit_map": [ + 0.016250645741820335, + 0.013408049941062927, + 0.020470980554819107, + 0.019106246531009674, + 0.029074501246213913, + 0.01656608283519745, + 0.020292164757847786, + 0.014780416153371334, + 0.007572644390165806, + 0.012317491695284843, + 0.009309155866503716, + 0.01000889204442501, + 0.00499885156750679, + 0.00850890763103962, + 0.007918382994830608, + 0.0074463714845478535, + 0.00422424403950572, + 0.004270894452929497, + 0.0062807598151266575, + 0.010138895362615585, + 0.0057031842879951, + 0.0061530196107923985, + 0.005423818714916706, + 0.006595516111701727, + 0.0061571658588945866, + 0.006913041230291128, + 0.007405397016555071, + 0.010819497518241405, + 0.007821165956556797, + 0.007583604659885168, + 0.009164175018668175, + 0.01000241469591856, + 0.05572913587093353, + 0.03818926587700844, + 0.0640491470694542, + 0.029564427211880684, + 0.022059036418795586, + 0.06707502156496048, + 0.028328238055109978, + 0.08907029777765274, + 0.0048459176905453205, + 0.007118613459169865, + 0.007472412660717964, + 0.0039376080967485905, + 0.0044599599204957485, + 0.004775360692292452, + 0.0037943306379020214, + 0.006862411741167307, + 0.016050420701503754, + 0.01838483102619648, + 0.010388840921223164, + 0.010920055210590363, + 0.010892186313867569, + 0.014422120526432991, + 0.008227373473346233, + 0.008969072252511978, + 0.005706600844860077, + 0.003505150554701686, + 0.004205257166177034, + 0.003314943751320243, + 0.0046224407851696014, + 0.007018732372671366, + 0.003894976805895567, + 0.006732905749231577, + 0.00506634684279561, + 0.0060436297208070755, + 0.005546749569475651, + 0.004323807079344988, + 0.006745344493538141, + 0.007140369154512882, + 0.005513485055416822, + 0.005928460042923689, + 0.005312301684170961, + 0.005903151817619801, + 0.005226392298936844, + 0.005860698875039816, + 0.006685291416943073, + 0.006165866740047932, + 0.00550961634144187, + 0.0057547930628061295 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 0, + 6 + ], + "selected_reward": -0.04865795746445656 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "dynamic, low-altitude drone tracking shot closely follows a steed, jet-black, shimmering satin, from the side and rear. It gallops vigorously across the vast, jagged summit of the ancient Great Wall, braving a fierce, strong wind. The horse's muscles bulge, and each powerful kick sends tiny pebbles and dust from the weathered gray masonry flung backward, carving brief arcs through the air. Its long black mane and tail dance wildly in the strong wind, its nostrils flaring with intense breathing, its eyes revealing an unyielding determination. The Great Wall, as captured, resembles a giant gray dragon, winding its way across rolling green mountains, stretching endlessly into the distance until it disappears beneath a gloomy, cloud-covered skyline. The entire scene is captured in an even, cool, diffuse light, dominated by a palette of somber grays, blacks, and dark greens, creating an atmosphere of desolation, loneliness, and epic power.", + "seed": 31693, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000101/group_10/index_5828_seed_31693_dp_21_rank_42_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 10, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.1375745328143239, + "vlm_logprob": -0.1375745328143239, + "vlm_yes_no_margin": 2.775, + "vlm_token_credit": -0.1375745328143239, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a jet-black steed with shimmering satin fur?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the horse galloping vigorously across the jagged summit of the Great Wall?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera follow the horse from the side and rear in a low-altitude tracking shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the horse's muscles bulging and is dust/pebbles flung backward by its kicks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit with even, cool, diffuse light dominated by somber grays, blacks, and dark greens?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.13434137403964996 + ], + "opposite_logprob_tokens": [ + -2.1343414783477783 + ], + "logprob_reward": -0.13434137403964996, + "yes_no_margin_reward": 2.0, + "token_credit_reward": -0.13434137403964996, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15432947874069214, + 0.031650859862565994, + 0.14771801233291626, + 0.13879536092281342, + 0.12089576572179794, + 0.10897146910429001, + 0.04123673960566521, + 0.08920033276081085, + 0.0680399090051651, + 0.09916200488805771 + ], + "credit_map": [ + 0.015668826177716255, + 0.02834039367735386, + 0.015382430516183376, + 0.019121503457427025, + 0.022631105035543442, + 0.02216523140668869, + 0.013992116786539555, + 0.01702788658440113, + 0.0031375500839203596, + 0.005438189022243023, + 0.003110770834609866, + 0.004309847019612789, + 0.003215556964278221, + 0.004106903448700905, + 0.00438866438344121, + 0.003943375777453184, + 0.011786838993430138, + 0.018063681200146675, + 0.018899358808994293, + 0.021296711638569832, + 0.018535539507865906, + 0.020645800977945328, + 0.019975032657384872, + 0.018515044823288918, + 0.030030611902475357, + 0.022307025268673897, + 0.013028004206717014, + 0.011057683266699314, + 0.016219589859247208, + 0.024027355015277863, + 0.009206749498844147, + 0.012918352149426937, + 0.010412008501589298, + 0.011435134336352348, + 0.024350585415959358, + 0.017151549458503723, + 0.010034476406872272, + 0.019785849377512932, + 0.014379819855093956, + 0.013346338644623756, + 0.010152236558496952, + 0.008528631180524826, + 0.01729438826441765, + 0.016162941232323647, + 0.01456419751048088, + 0.012599831447005272, + 0.01527367066591978, + 0.014395570382475853, + 0.00414159195497632, + 0.0046631041914224625, + 0.0062262131832540035, + 0.004463302902877331, + 0.004803972784429789, + 0.005647967103868723, + 0.00540276849642396, + 0.005887819454073906, + 0.007772521581500769, + 0.011015219613909721, + 0.017875561490654945, + 0.012229771353304386, + 0.013381225056946278, + 0.007171733770519495, + 0.011684487573802471, + 0.008069814182817936, + 0.004794247914105654, + 0.007536976132541895, + 0.010998285375535488, + 0.007933715358376503, + 0.010519267991185188, + 0.010032233782112598, + 0.010624177753925323, + 0.005601002834737301, + 0.01187947392463684, + 0.011995907872915268, + 0.010204433463513851, + 0.010695711709558964, + 0.00873738992959261, + 0.01790783926844597, + 0.01648421585559845, + 0.011257031001150608 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 3 + ], + "selected_reward": -0.13434137403964996 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a jet-black steed with shimmering satin fur?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the horse galloping vigorously across the jagged summit of the Great Wall?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera follow the horse from the side and rear in a low-altitude tracking shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the horse's muscles bulging and is dust/pebbles flung backward by its kicks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit with even, cool, diffuse light dominated by somber grays, blacks, and dark greens?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.4743175208568573 + ], + "opposite_logprob_tokens": [ + -0.9743175506591797 + ], + "logprob_reward": -0.4743175208568573, + "yes_no_margin_reward": 0.5, + "token_credit_reward": -0.4743175208568573, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11150780320167542, + 0.044099703431129456, + 0.08673866838216782, + 0.3095715045928955, + 0.17327389121055603, + 0.04034600034356117, + 0.03538928180932999, + 0.05683640390634537, + 0.0805201455950737, + 0.06171661987900734 + ], + "credit_map": [ + 0.01997040957212448, + 0.013408660888671875, + 0.014950484968721867, + 0.009062794037163258, + 0.012690510600805283, + 0.02107204869389534, + 0.008627495728433132, + 0.011725393123924732, + 0.004902211017906666, + 0.007826080545783043, + 0.005075865890830755, + 0.0033999988809227943, + 0.0048260935582220554, + 0.00681761559098959, + 0.004950464703142643, + 0.006301372777670622, + 0.01077903900295496, + 0.01943790912628174, + 0.014696047641336918, + 0.007978116162121296, + 0.009638040326535702, + 0.01148897409439087, + 0.006208972539752722, + 0.006511562503874302, + 0.07537755370140076, + 0.053850606083869934, + 0.028142178431153297, + 0.027247613295912743, + 0.03193257749080658, + 0.04936537891626358, + 0.02102035842835903, + 0.0226352009922266, + 0.016331268474459648, + 0.014853391796350479, + 0.03252548351883888, + 0.023525778204202652, + 0.014472206123173237, + 0.03381431847810745, + 0.021703626960515976, + 0.016047805547714233, + 0.0036800005473196507, + 0.002186424098908901, + 0.005672176368534565, + 0.00493084080517292, + 0.003840723307803273, + 0.006099480204284191, + 0.009109091944992542, + 0.004827258177101612, + 0.0031122476793825626, + 0.002870941534638405, + 0.0048898011445999146, + 0.0036482661962509155, + 0.004029419738799334, + 0.007098724599927664, + 0.005189866758882999, + 0.004550011828541756, + 0.007631996180862188, + 0.005517061334103346, + 0.009382626973092556, + 0.006791666615754366, + 0.00733207818120718, + 0.005437182728201151, + 0.006085611879825592, + 0.008658178150653839, + 0.005276406183838844, + 0.007670118939131498, + 0.01584901288151741, + 0.01075050700455904, + 0.013158184476196766, + 0.009269158355891705, + 0.012271478772163391, + 0.00627527479082346, + 0.007727731019258499, + 0.007047311402857304, + 0.00579160125926137, + 0.005980130285024643, + 0.006467287428677082, + 0.01223078928887844, + 0.00946246087551117, + 0.007009301334619522 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 4, + 0 + ], + "selected_reward": -0.4743175208568573 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a jet-black steed with shimmering satin fur?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the horse galloping vigorously across the jagged summit of the Great Wall?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera follow the horse from the side and rear in a low-altitude tracking shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the horse's muscles bulging and is dust/pebbles flung backward by its kicks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit with even, cool, diffuse light dominated by somber grays, blacks, and dark greens?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.009787782095372677 + ], + "opposite_logprob_tokens": [ + -4.634787559509277 + ], + "logprob_reward": -0.009787782095372677, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.009787782095372677, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08637388795614243, + 0.028728434816002846, + 0.07320992648601532, + 0.3175736665725708, + 0.07373356819152832, + 0.0522431880235672, + 0.07069827616214752, + 0.09781219065189362, + 0.12209051102399826, + 0.07753635197877884 + ], + "credit_map": [ + 0.012205241248011589, + 0.008819404058158398, + 0.01027403399348259, + 0.00957865547388792, + 0.009477578103542328, + 0.01649565063416958, + 0.008273683488368988, + 0.011249640956521034, + 0.00265308516100049, + 0.0036084677558392286, + 0.0027965535409748554, + 0.0038330168463289738, + 0.003889823332428932, + 0.00336453621275723, + 0.004460795316845179, + 0.004122157581150532, + 0.008477839641273022, + 0.016158178448677063, + 0.013632574118673801, + 0.006779795978218317, + 0.007833405397832394, + 0.009115965105593204, + 0.004537410102784634, + 0.006674764212220907, + 0.07337945699691772, + 0.053226765245199203, + 0.025118598714470863, + 0.02397102303802967, + 0.03410516306757927, + 0.06624754518270493, + 0.0202422346919775, + 0.021282868459820747, + 0.008253869600594044, + 0.00555068114772439, + 0.010822637006640434, + 0.009626884944736958, + 0.005765404086560011, + 0.01420557126402855, + 0.010895284824073315, + 0.008613234385848045, + 0.004613396245986223, + 0.0035626485478132963, + 0.0054817283526062965, + 0.005668602883815765, + 0.004476760979741812, + 0.006057452410459518, + 0.013965618796646595, + 0.00841697957366705, + 0.005341843236237764, + 0.0071525247767567635, + 0.012986522167921066, + 0.008782342076301575, + 0.008546102792024612, + 0.007744028232991695, + 0.00987450871616602, + 0.010270405560731888, + 0.010976909659802914, + 0.010806865990161896, + 0.018256481736898422, + 0.01206658873707056, + 0.013665747828781605, + 0.008093263022601604, + 0.011539174243807793, + 0.012407157570123672, + 0.008240672759711742, + 0.012783581390976906, + 0.026635844260454178, + 0.020070374011993408, + 0.01892818510532379, + 0.013252555392682552, + 0.013884874060750008, + 0.008294434286653996, + 0.010255991481244564, + 0.012026400305330753, + 0.009565949440002441, + 0.008045626804232597, + 0.008261872455477715, + 0.012848539277911186, + 0.009016389958560467, + 0.007515584584325552 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 8, + 7 + ], + "selected_reward": -0.009787782095372677 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a jet-black steed with shimmering satin fur?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the horse galloping vigorously across the jagged summit of the Great Wall?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera follow the horse from the side and rear in a low-altitude tracking shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the horse's muscles bulging and is dust/pebbles flung backward by its kicks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit with even, cool, diffuse light dominated by somber grays, blacks, and dark greens?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02635699324309826 + ], + "opposite_logprob_tokens": [ + -3.6513569355010986 + ], + "logprob_reward": -0.02635699324309826, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.02635699324309826, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08026119321584702, + 0.029434503987431526, + 0.06635518372058868, + 0.1307050734758377, + 0.05390702933073044, + 0.03458820655941963, + 0.041032303124666214, + 0.03448149561882019, + 0.11499810218811035, + 0.41423699259757996 + ], + "credit_map": [ + 0.01085339393466711, + 0.007321328390389681, + 0.011635497212409973, + 0.010084475390613079, + 0.010661435313522816, + 0.01024549175053835, + 0.008438310585916042, + 0.011021257378160954, + 0.003030013060197234, + 0.00572741124778986, + 0.0030580877792090178, + 0.0026677753776311874, + 0.0029570695478469133, + 0.0030847738962620497, + 0.003580240998417139, + 0.005329128820449114, + 0.0075698839500546455, + 0.011845497414469719, + 0.010413870215415955, + 0.0057836491614580154, + 0.008021709509193897, + 0.009286746382713318, + 0.0069008017890155315, + 0.0065330262295901775, + 0.023942194879055023, + 0.017219096422195435, + 0.013388780876994133, + 0.010094603523612022, + 0.011554664000868797, + 0.032659128308296204, + 0.008149676956236362, + 0.013696925714612007, + 0.004944820422679186, + 0.005760796833783388, + 0.008390781469643116, + 0.005930508486926556, + 0.004712620750069618, + 0.0063897850923240185, + 0.008150607347488403, + 0.009627102874219418, + 0.0036398740485310555, + 0.0033870632760226727, + 0.003938854672014713, + 0.004091905429959297, + 0.003985928371548653, + 0.0057977549731731415, + 0.005525851622223854, + 0.004220969509333372, + 0.005002964287996292, + 0.0031874210108071566, + 0.004831232596188784, + 0.00452717998996377, + 0.003018140560016036, + 0.006467444356530905, + 0.007582800462841988, + 0.0064151170663535595, + 0.003665200201794505, + 0.0028547951951622963, + 0.007603735197335482, + 0.00474202586337924, + 0.005249264650046825, + 0.0037286151200532913, + 0.003723087487742305, + 0.0029147653840482235, + 0.0073379226960241795, + 0.011613051407039165, + 0.0226840041577816, + 0.015220091678202152, + 0.01754249818623066, + 0.01535032782703638, + 0.016042327508330345, + 0.009207861497998238, + 0.0380934402346611, + 0.04252126067876816, + 0.0420093797147274, + 0.050290998071432114, + 0.052190136164426804, + 0.07724898308515549, + 0.06961727887392044, + 0.042265474796295166 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 3, + 8 + ], + "selected_reward": -0.02635699324309826 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a jet-black steed with shimmering satin fur?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the horse galloping vigorously across the jagged summit of the Great Wall?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera follow the horse from the side and rear in a low-altitude tracking shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the horse's muscles bulging and is dust/pebbles flung backward by its kicks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit with even, cool, diffuse light dominated by somber grays, blacks, and dark greens?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04306899383664131 + ], + "opposite_logprob_tokens": [ + -3.1680688858032227 + ], + "logprob_reward": -0.04306899383664131, + "yes_no_margin_reward": 3.125, + "token_credit_reward": -0.04306899383664131, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.18493299186229706, + 0.0709829106926918, + 0.10920266807079315, + 0.16713424026966095, + 0.09004363417625427, + 0.05498592555522919, + 0.05193144083023071, + 0.10074470192193985, + 0.07089010626077652, + 0.09915141016244888 + ], + "credit_map": [ + 0.025151748210191727, + 0.015340493060648441, + 0.01895342953503132, + 0.014634071849286556, + 0.019902046769857407, + 0.04487045854330063, + 0.013496619649231434, + 0.03258413448929787, + 0.006599494256079197, + 0.013468670658767223, + 0.007984448224306107, + 0.005937281996011734, + 0.007229025475680828, + 0.012756533920764923, + 0.006390789523720741, + 0.010616668500006199, + 0.011538143269717693, + 0.01403259951621294, + 0.019731959328055382, + 0.01032725628465414, + 0.01740392856299877, + 0.01737612672150135, + 0.007611763197928667, + 0.011180894449353218, + 0.04224315658211708, + 0.030662478879094124, + 0.015520231798291206, + 0.014350871555507183, + 0.01860811933875084, + 0.021722009405493736, + 0.01210195291787386, + 0.011925425380468369, + 0.011974797584116459, + 0.007814396172761917, + 0.009813463315367699, + 0.006988417357206345, + 0.005895807407796383, + 0.021919794380664825, + 0.012634347192943096, + 0.013002610765397549, + 0.00400047330185771, + 0.0050082337111234665, + 0.006132744252681732, + 0.006795541848987341, + 0.007654638960957527, + 0.008704866282641888, + 0.009424577467143536, + 0.007264850661158562, + 0.0037704857531934977, + 0.003149403026327491, + 0.007879926823079586, + 0.005689085461199284, + 0.005452622659504414, + 0.007648024242371321, + 0.007808889728039503, + 0.010532999411225319, + 0.008547299541532993, + 0.012799429707229137, + 0.018035732209682465, + 0.012142941355705261, + 0.014603658579289913, + 0.011004769243299961, + 0.011272624135017395, + 0.012338249944150448, + 0.005244425497949123, + 0.007051064167171717, + 0.013931818306446075, + 0.013827537186443806, + 0.008722292259335518, + 0.008519407361745834, + 0.008544070646166801, + 0.005049490369856358, + 0.010031980462372303, + 0.011790186166763306, + 0.01082723680883646, + 0.010354732163250446, + 0.01480082981288433, + 0.01484809909015894, + 0.015138515271246433, + 0.01135982759296894 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 2 + ], + "selected_reward": -0.04306899383664131 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "dynamic, low-altitude drone tracking shot closely follows a steed, jet-black, shimmering satin, from the side and rear. It gallops vigorously across the vast, jagged summit of the ancient Great Wall, braving a fierce, strong wind. The horse's muscles bulge, and each powerful kick sends tiny pebbles and dust from the weathered gray masonry flung backward, carving brief arcs through the air. Its long black mane and tail dance wildly in the strong wind, its nostrils flaring with intense breathing, its eyes revealing an unyielding determination. The Great Wall, as captured, resembles a giant gray dragon, winding its way across rolling green mountains, stretching endlessly into the distance until it disappears beneath a gloomy, cloud-covered skyline. The entire scene is captured in an even, cool, diffuse light, dominated by a palette of somber grays, blacks, and dark greens, creating an atmosphere of desolation, loneliness, and epic power.", + "seed": 31694, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000101/group_10/index_5828_seed_31694_dp_21_rank_42_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 10, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.15973448939621449, + "vlm_logprob": -0.15973448939621449, + "vlm_yes_no_margin": 2.45, + "vlm_token_credit": -0.15973448939621449, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a jet-black steed with shimmering satin fur?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the horse galloping vigorously across the jagged summit of the Great Wall?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera follow the horse from the side and rear in a low-altitude tracking shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the horse's muscles bulging and is dust/pebbles flung backward by its kicks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit with even, cool, diffuse light dominated by somber grays, blacks, and dark greens?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.18906651437282562 + ], + "opposite_logprob_tokens": [ + -1.8140665292739868 + ], + "logprob_reward": -0.18906651437282562, + "yes_no_margin_reward": 1.625, + "token_credit_reward": -0.18906651437282562, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13641321659088135, + 0.0200764462351799, + 0.023989928886294365, + 0.03583177551627159, + 0.25795385241508484, + 0.0879770964384079, + 0.03811879828572273, + 0.04782280698418617, + 0.03534731641411781, + 0.3164688050746918 + ], + "credit_map": [ + 0.013441144488751888, + 0.010119589045643806, + 0.02376469597220421, + 0.025777459144592285, + 0.016622599214315414, + 0.007575794123113155, + 0.015630822628736496, + 0.023481104522943497, + 0.0023882242385298014, + 0.0030850344337522984, + 0.002674276940524578, + 0.0020180183928459883, + 0.001534230774268508, + 0.0032812070567160845, + 0.0017057437216863036, + 0.0033897114917635918, + 0.0023963998537510633, + 0.0025784377939999104, + 0.004122539889067411, + 0.003153459634631872, + 0.0029176187235862017, + 0.0026066217105835676, + 0.002936720382422209, + 0.003278132528066635, + 0.0027731608133763075, + 0.004000958986580372, + 0.00795009732246399, + 0.004515879321843386, + 0.003657410852611065, + 0.0036842019762843847, + 0.004763832315802574, + 0.004486232064664364, + 0.033462777733802795, + 0.040037427097558975, + 0.03390493616461754, + 0.03928171470761299, + 0.027747567743062973, + 0.0192672498524189, + 0.020125482231378555, + 0.04412669315934181, + 0.008321478962898254, + 0.006833280436694622, + 0.010705058462917805, + 0.005990974605083466, + 0.008890022523701191, + 0.008719746954739094, + 0.02633873000741005, + 0.01217779889702797, + 0.003131559584289789, + 0.004998063202947378, + 0.005375272128731012, + 0.006082472391426563, + 0.008291594684123993, + 0.0027053989470005035, + 0.0032463176175951958, + 0.0042881169356405735, + 0.005848768632858992, + 0.004976876545697451, + 0.003975890576839447, + 0.002967779291793704, + 0.016420746222138405, + 0.0025372563395649195, + 0.003005191683769226, + 0.008090299554169178, + 0.005056025460362434, + 0.002842885907739401, + 0.004300061147660017, + 0.010440422222018242, + 0.002117832889780402, + 0.0023812411818653345, + 0.0046983580105006695, + 0.0035104930866509676, + 0.03175913542509079, + 0.016359718516469002, + 0.03673820570111275, + 0.10207993537187576, + 0.013359434902668, + 0.011136184446513653, + 0.019401462748646736, + 0.08563472330570221 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 4, + 0 + ], + "selected_reward": -0.18906651437282562 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a jet-black steed with shimmering satin fur?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the horse galloping vigorously across the jagged summit of the Great Wall?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera follow the horse from the side and rear in a low-altitude tracking shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the horse's muscles bulging and is dust/pebbles flung backward by its kicks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit with even, cool, diffuse light dominated by somber grays, blacks, and dark greens?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.474298894405365 + ], + "opposite_logprob_tokens": [ + -0.974298894405365 + ], + "logprob_reward": -0.474298894405365, + "yes_no_margin_reward": 0.5, + "token_credit_reward": -0.474298894405365, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09027714282274246, + 0.1485498696565628, + 0.08308684080839157, + 0.08310107886791229, + 0.23157255351543427, + 0.040604118257761, + 0.02323022298514843, + 0.19010710716247559, + 0.04032539576292038, + 0.06914566457271576 + ], + "credit_map": [ + 0.009133189916610718, + 0.00810105912387371, + 0.011783577501773834, + 0.016254546120762825, + 0.015122352167963982, + 0.006754548288881779, + 0.011493696831166744, + 0.01163417287170887, + 0.015520558692514896, + 0.01648530177772045, + 0.0278912503272295, + 0.013595796190202236, + 0.008932134136557579, + 0.019956320524215698, + 0.011950577609241009, + 0.03421792760491371, + 0.007206853944808245, + 0.010377072729170322, + 0.011361012235283852, + 0.014810094609856606, + 0.008950830437242985, + 0.007767973933368921, + 0.012292974628508091, + 0.01032002829015255, + 0.00801145937293768, + 0.011511576361954212, + 0.008530530147254467, + 0.015890950337052345, + 0.006784500554203987, + 0.008818010799586773, + 0.010737866163253784, + 0.012816186062991619, + 0.024385055527091026, + 0.026420703157782555, + 0.05116613209247589, + 0.02729874849319458, + 0.019838737323880196, + 0.026835573837161064, + 0.014693121425807476, + 0.0409344844520092, + 0.002765555400401354, + 0.002926533343270421, + 0.005472543649375439, + 0.0022929054684937, + 0.004203298594802618, + 0.0030965525656938553, + 0.013813629746437073, + 0.006033094134181738, + 0.0030528802890330553, + 0.002443415578454733, + 0.002778585534542799, + 0.003224229672923684, + 0.004186841659247875, + 0.002539469860494137, + 0.00236930837854743, + 0.0026354920119047165, + 0.014916768297553062, + 0.018042592331767082, + 0.044924452900886536, + 0.016865067183971405, + 0.049235451966524124, + 0.01333321537822485, + 0.016949612647294998, + 0.0158399548381567, + 0.005067711696028709, + 0.004099304787814617, + 0.004959918092936277, + 0.005708811338990927, + 0.003539232537150383, + 0.003262801095843315, + 0.005723463837057352, + 0.00796415563672781, + 0.011874907650053501, + 0.005206150934100151, + 0.011321394704282284, + 0.012465939857065678, + 0.0036864811554551125, + 0.005124170333147049, + 0.005936103407293558, + 0.013530519790947437 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 7, + 1 + ], + "selected_reward": -0.474298894405365 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a jet-black steed with shimmering satin fur?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the horse galloping vigorously across the jagged summit of the Great Wall?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera follow the horse from the side and rear in a low-altitude tracking shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the horse's muscles bulging and is dust/pebbles flung backward by its kicks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit with even, cool, diffuse light dominated by somber grays, blacks, and dark greens?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01254771463572979 + ], + "opposite_logprob_tokens": [ + -4.387547492980957 + ], + "logprob_reward": -0.01254771463572979, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.01254771463572979, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.061171114444732666, + 0.07277307659387589, + 0.08689326047897339, + 0.09590199589729309, + 0.2733747661113739, + 0.11741752177476883, + 0.032034605741500854, + 0.09679697453975677, + 0.070503830909729, + 0.09313283860683441 + ], + "credit_map": [ + 0.006054953671991825, + 0.007516961544752121, + 0.0076955812983214855, + 0.010643985122442245, + 0.00823630765080452, + 0.005016940645873547, + 0.007594633847475052, + 0.008411753922700882, + 0.00955120474100113, + 0.006194685120135546, + 0.01520484033972025, + 0.006396070122718811, + 0.004910310264676809, + 0.00585528090596199, + 0.006824333220720291, + 0.017836350947618484, + 0.006625843234360218, + 0.011800309643149376, + 0.010410758666694164, + 0.01553570106625557, + 0.009410104714334011, + 0.009390505030751228, + 0.012906859628856182, + 0.010813174769282341, + 0.009212515316903591, + 0.008409202098846436, + 0.01546945609152317, + 0.014700478874146938, + 0.011443068273365498, + 0.009662565775215626, + 0.013324500992894173, + 0.013680205680429935, + 0.035103436559438705, + 0.040080636739730835, + 0.04071846604347229, + 0.038570743054151535, + 0.02791772596538067, + 0.022043494507670403, + 0.02143215388059616, + 0.0475081205368042, + 0.006772332359105349, + 0.0045963251031935215, + 0.017285821959376335, + 0.006000634282827377, + 0.011899224482476711, + 0.0070432317443192005, + 0.04431252181529999, + 0.019507430493831635, + 0.002627347130328417, + 0.0036405546125024557, + 0.0038786467630416155, + 0.005569850094616413, + 0.0048040724359452724, + 0.003693454433232546, + 0.003961425274610519, + 0.0038592552300542593, + 0.009865493513643742, + 0.012566418386995792, + 0.009677034802734852, + 0.008250657469034195, + 0.023892993107438087, + 0.0054357098415493965, + 0.009205177426338196, + 0.01790349930524826, + 0.009907549247145653, + 0.008307230658829212, + 0.007880458608269691, + 0.017442267388105392, + 0.004768498707562685, + 0.006658061407506466, + 0.007555426098406315, + 0.007984333671629429, + 0.016525061801075935, + 0.0096767358481884, + 0.0160334762185812, + 0.014509248547255993, + 0.0044995443895459175, + 0.007814663462340832, + 0.007647303864359856, + 0.016426801681518555 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 5, + 7 + ], + "selected_reward": -0.01254771463572979 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a jet-black steed with shimmering satin fur?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the horse galloping vigorously across the jagged summit of the Great Wall?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera follow the horse from the side and rear in a low-altitude tracking shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the horse's muscles bulging and is dust/pebbles flung backward by its kicks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit with even, cool, diffuse light dominated by somber grays, blacks, and dark greens?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.033713314682245255 + ], + "opposite_logprob_tokens": [ + -3.4087133407592773 + ], + "logprob_reward": -0.033713314682245255, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.033713314682245255, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.05918189883232117, + 0.21153320372104645, + 0.08257579803466797, + 0.0638291984796524, + 0.2974390387535095, + 0.03998124599456787, + 0.015624798834323883, + 0.08480779826641083, + 0.07198889553546906, + 0.07303811609745026 + ], + "credit_map": [ + 0.005464853718876839, + 0.0060044205747544765, + 0.009569719433784485, + 0.007808326277881861, + 0.00703212758526206, + 0.005844818893820047, + 0.005088911857455969, + 0.012368720956146717, + 0.027301345020532608, + 0.018968923017382622, + 0.05110595375299454, + 0.016419827938079834, + 0.013816493563354015, + 0.019987471401691437, + 0.019091632217168808, + 0.04484153911471367, + 0.00682635884732008, + 0.012125588022172451, + 0.009377231821417809, + 0.017286358401179314, + 0.007986979559063911, + 0.007505161687731743, + 0.011885326355695724, + 0.009582794271409512, + 0.007169180549681187, + 0.006397082004696131, + 0.01109793595969677, + 0.009625749662518501, + 0.00637470930814743, + 0.0063090878538787365, + 0.00948315393179655, + 0.007372301537543535, + 0.03979232907295227, + 0.04638117924332619, + 0.037848640233278275, + 0.044519539922475815, + 0.033652693033218384, + 0.02376100979745388, + 0.024993153288960457, + 0.04649048298597336, + 0.004443163517862558, + 0.0046056341379880905, + 0.006204684264957905, + 0.005393693223595619, + 0.004600010812282562, + 0.004837901331484318, + 0.0061375838704407215, + 0.0037585734389722347, + 0.0017171305371448398, + 0.0016782209277153015, + 0.0019886463414877653, + 0.002823054790496826, + 0.0024515907280147076, + 0.0016113425372168422, + 0.00153625151142478, + 0.0018185615772381425, + 0.007658445741981268, + 0.010622072964906693, + 0.01573001593351364, + 0.010756860487163067, + 0.012447197921574116, + 0.004016516730189323, + 0.007356336805969477, + 0.016220349818468094, + 0.010404938831925392, + 0.00723382318392396, + 0.00903363898396492, + 0.01793544925749302, + 0.004000317770987749, + 0.0044684261083602905, + 0.007173237856477499, + 0.011739065870642662, + 0.00789086613804102, + 0.004480286501348019, + 0.00981038250029087, + 0.021552875638008118, + 0.003277439856901765, + 0.003804985899478197, + 0.0051109944470226765, + 0.01711028814315796 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 1, + 7 + ], + "selected_reward": -0.033713314682245255 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a jet-black steed with shimmering satin fur?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the horse galloping vigorously across the jagged summit of the Great Wall?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera follow the horse from the side and rear in a low-altitude tracking shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the horse's muscles bulging and is dust/pebbles flung backward by its kicks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit with even, cool, diffuse light dominated by somber grays, blacks, and dark greens?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.08904600888490677 + ], + "opposite_logprob_tokens": [ + -2.464046001434326 + ], + "logprob_reward": -0.08904600888490677, + "yes_no_margin_reward": 2.375, + "token_credit_reward": -0.08904600888490677, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.16270650923252106, + 0.07255398482084274, + 0.050110526382923126, + 0.026537764817476273, + 0.163889542222023, + 0.25033891201019287, + 0.08091424405574799, + 0.04019741714000702, + 0.04926434904336929, + 0.10348676145076752 + ], + "credit_map": [ + 0.021364931017160416, + 0.026053469628095627, + 0.021457219496369362, + 0.020533939823508263, + 0.016844574362039566, + 0.020933546125888824, + 0.014516961760818958, + 0.021001866087317467, + 0.008814942091703415, + 0.006965273525565863, + 0.015520934946835041, + 0.006136095616966486, + 0.004991612397134304, + 0.007004707120358944, + 0.006861996836960316, + 0.016258426010608673, + 0.004329120274633169, + 0.0055145681835711, + 0.009083696641027927, + 0.006852617487311363, + 0.005581301636993885, + 0.005315863527357578, + 0.006017998792231083, + 0.007415357045829296, + 0.003435359336435795, + 0.0027873984072357416, + 0.0053244843147695065, + 0.0031553683802485466, + 0.0024666183162480593, + 0.003586236387491226, + 0.0024861632846295834, + 0.0032961368560791016, + 0.022287126630544662, + 0.02580067329108715, + 0.027960538864135742, + 0.017573507502675056, + 0.01411083061248064, + 0.016537925228476524, + 0.011128569953143597, + 0.028490379452705383, + 0.023954324424266815, + 0.022898204624652863, + 0.030901527032256126, + 0.023892076686024666, + 0.02769755758345127, + 0.026638230308890343, + 0.06372654438018799, + 0.030630439519882202, + 0.013276365585625172, + 0.009755158796906471, + 0.007950185798108578, + 0.006935860496014357, + 0.00961452815681696, + 0.008652301505208015, + 0.012724687345325947, + 0.012005154974758625, + 0.0036758221685886383, + 0.004185945726931095, + 0.005372407380491495, + 0.004370792303234339, + 0.010292316786944866, + 0.0021310627926141024, + 0.0027693838346749544, + 0.007399685215204954, + 0.0057655512355268, + 0.006599422078579664, + 0.006448434200137854, + 0.007678093854337931, + 0.00464391615241766, + 0.004195431713014841, + 0.008290758356451988, + 0.005642742849886417, + 0.013052407652139664, + 0.007653326727449894, + 0.01586897484958172, + 0.02525520697236061, + 0.005823545157909393, + 0.006757473573088646, + 0.007175416685640812, + 0.021900413557887077 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 4, + 0 + ], + "selected_reward": -0.08904600888490677 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000101/group_11/index_tensor([732, 732])_group_11_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000101/group_11/index_tensor([732, 732])_group_11_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..aebca53d156722c8998be6c60ab3c5b848b01b21 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000101/group_11/index_tensor([732, 732])_group_11_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "high-fidelity CGI animation, rendered in extreme close-up, focuses on a massive, ghostly-faced hawk moth, its thick, hairy thorax rendered with a disturbing, bone-white skull pattern. Sculpted from a pitch-black, void-like background by a single, harsh overhead light, the giant moth unfolds its massive, velvety wings, covered in an intricate mosaic of deep chocolate brown, dusty ochre, and jet black, each microscopic scale visible with crystalline clarity. Its segmented abdomen contracts violently, its feathery antennae quivering with vibrations as it lets out a piercing, almost electronic howl. In sync with the scream, the unstable camera also trembled violently, as if hit by sound waves, causing every hyper-realistic detail in the picture - from its huge compound eyes flashing with highlights to the few scales floating in the air due to violent vibrations - to quickly switch between out of focus and clarity, creating a suffocating sense of fear from biological instinct and a strong visual and auditory impact.", + "seed": 15321, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000101/group_11/index_732_seed_15321_dp_22_rank_44_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 11, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.40544563680887225, + "vlm_logprob": -0.40544563680887225, + "vlm_yes_no_margin": 1.875, + "vlm_token_credit": -0.40544563680887225, + "vlm_yes_ratio": 0.6, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the moth's thorax thick, hairy, and patterned with a bone-white skull design?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the moth's wings massive, velvety, and covered in a mosaic of deep chocolate brown, dusty ochre, and jet black?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the background pitch-black and void-like, illuminated by a single harsh overhead light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the moth's segmented abdomen contract violently while its feathery antennae quiver?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera tremble violently, causing the image to switch rapidly between out of focus and clarity?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.053890034556388855 + ], + "opposite_logprob_tokens": [ + -3.0538899898529053 + ], + "logprob_reward": -0.053890034556388855, + "yes_no_margin_reward": 3.0, + "token_credit_reward": -0.053890034556388855, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2357524186372757, + 0.05353642255067825, + 0.043418530374765396, + 0.09887409210205078, + 0.22369278967380524, + 0.058565180748701096, + 0.03540312871336937, + 0.1458325833082199, + 0.05183376744389534, + 0.053091034293174744 + ], + "credit_map": [ + 0.026050489395856857, + 0.045458242297172546, + 0.037074897438287735, + 0.023927392438054085, + 0.033680763095617294, + 0.01949416659772396, + 0.029598407447338104, + 0.02046806365251541, + 0.005704200826585293, + 0.0056702313013374805, + 0.007069685030728579, + 0.006076551508158445, + 0.007111195474863052, + 0.004294624552130699, + 0.00619059894233942, + 0.011419330723583698, + 0.004309375304728746, + 0.005613588262349367, + 0.007920154370367527, + 0.003759032115340233, + 0.0044401404447853565, + 0.0068667265586555, + 0.004615379497408867, + 0.0058941347524523735, + 0.013932602480053902, + 0.008103322237730026, + 0.015270781703293324, + 0.007656486704945564, + 0.008737519383430481, + 0.01207311637699604, + 0.010864460840821266, + 0.022235799580812454, + 0.016849178820848465, + 0.01723361387848854, + 0.04213688150048256, + 0.04728890210390091, + 0.012176093645393848, + 0.025997519493103027, + 0.031027236953377724, + 0.03098336234688759, + 0.005875809583812952, + 0.006121364887803793, + 0.010561271570622921, + 0.008694472722709179, + 0.005345259793102741, + 0.007082284893840551, + 0.0069842878729105, + 0.007900431752204895, + 0.005117329303175211, + 0.005151429679244757, + 0.004223613068461418, + 0.0037905594799667597, + 0.004248409532010555, + 0.0032785076182335615, + 0.005051672458648682, + 0.004541607107967138, + 0.018364708870649338, + 0.013467979617416859, + 0.015075846575200558, + 0.01802504062652588, + 0.01747976802289486, + 0.027923064306378365, + 0.015035407617688179, + 0.020460771396756172, + 0.005386420525610447, + 0.006418329197913408, + 0.004408649168908596, + 0.005759336519986391, + 0.007367460522800684, + 0.006966762710362673, + 0.008976995013654232, + 0.00654981704428792, + 0.004951175767928362, + 0.006086468696594238, + 0.005164520815014839, + 0.006322816945612431, + 0.009608427062630653, + 0.005172640085220337, + 0.008408761583268642, + 0.007376221474260092 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 7 + ], + "selected_reward": -0.053890034556388855 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the moth's thorax thick, hairy, and patterned with a bone-white skull design?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the moth's wings massive, velvety, and covered in a mosaic of deep chocolate brown, dusty ochre, and jet black?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the background pitch-black and void-like, illuminated by a single harsh overhead light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the moth's segmented abdomen contract violently while its feathery antennae quiver?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera tremble violently, causing the image to switch rapidly between out of focus and clarity?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01819268986582756 + ], + "opposite_logprob_tokens": [ + -4.018192768096924 + ], + "logprob_reward": -0.01819268986582756, + "yes_no_margin_reward": 4.0, + "token_credit_reward": -0.01819268986582756, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2121862918138504, + 0.09256593137979507, + 0.07768187671899796, + 0.11814044415950775, + 0.09235930442810059, + 0.05065328627824783, + 0.09150252491235733, + 0.06706563383340836, + 0.09924930334091187, + 0.09859543293714523 + ], + "credit_map": [ + 0.021589472889900208, + 0.036075472831726074, + 0.03645937889814377, + 0.016680322587490082, + 0.034267257899045944, + 0.021060090512037277, + 0.020556554198265076, + 0.025497743859887123, + 0.01041532400995493, + 0.012321548536419868, + 0.013010415248572826, + 0.009498540312051773, + 0.012296237051486969, + 0.012709424830973148, + 0.007490635849535465, + 0.014823795296251774, + 0.00735275074839592, + 0.009979940950870514, + 0.0157473124563694, + 0.007598243188112974, + 0.010411828756332397, + 0.008243080228567123, + 0.007351804990321398, + 0.010996912606060505, + 0.019336111843585968, + 0.013339572586119175, + 0.014077420346438885, + 0.01056702621281147, + 0.013539604842662811, + 0.012274512089788914, + 0.0177261084318161, + 0.017280083149671555, + 0.014233123511075974, + 0.008170735090970993, + 0.01896597258746624, + 0.009275552816689014, + 0.005638019647449255, + 0.014933677390217781, + 0.008433793671429157, + 0.01270841620862484, + 0.007297832984477282, + 0.004736799746751785, + 0.010807136073708534, + 0.0034392683301120996, + 0.004237320274114609, + 0.008361462503671646, + 0.005629391875118017, + 0.006144069600850344, + 0.010176904499530792, + 0.021213369444012642, + 0.007305392064154148, + 0.006499674171209335, + 0.017682049423456192, + 0.00872027687728405, + 0.006713686976581812, + 0.0131911626085639, + 0.007350414525717497, + 0.00735179940238595, + 0.004690615925937891, + 0.006861059460788965, + 0.009746565483510494, + 0.013133984059095383, + 0.0059037599712610245, + 0.012027429416775703, + 0.009830396622419357, + 0.010513553395867348, + 0.010463463142514229, + 0.009744401089847088, + 0.01369410753250122, + 0.013362504541873932, + 0.017179809510707855, + 0.014461062848567963, + 0.010192317888140678, + 0.01207774505019188, + 0.010936729609966278, + 0.009616554714739323, + 0.015224013477563858, + 0.012221463955938816, + 0.015249538235366344, + 0.01307706069201231 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 8 + ], + "selected_reward": -0.01819268986582756 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the moth's thorax thick, hairy, and patterned with a bone-white skull design?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the moth's wings massive, velvety, and covered in a mosaic of deep chocolate brown, dusty ochre, and jet black?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the background pitch-black and void-like, illuminated by a single harsh overhead light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the moth's segmented abdomen contract violently while its feathery antennae quiver?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera tremble violently, causing the image to switch rapidly between out of focus and clarity?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.038084905594587326 + ], + "opposite_logprob_tokens": [ + -3.2880849838256836 + ], + "logprob_reward": -0.038084905594587326, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.038084905594587326, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.21937385201454163, + 0.044203657656908035, + 0.06351195275783539, + 0.11543408781290054, + 0.13951890170574188, + 0.13903410732746124, + 0.07304990291595459, + 0.05025823041796684, + 0.11282172799110413, + 0.04279360547661781 + ], + "credit_map": [ + 0.02606535330414772, + 0.029510386288166046, + 0.03041345439851284, + 0.01879878342151642, + 0.03592371940612793, + 0.022722508758306503, + 0.024157250300049782, + 0.031782396137714386, + 0.004248254466801882, + 0.005297713913023472, + 0.005907549522817135, + 0.004888960625976324, + 0.005654162727296352, + 0.004995265509933233, + 0.005705609917640686, + 0.007506142370402813, + 0.007335055619478226, + 0.011796302162110806, + 0.007524976972490549, + 0.008258070796728134, + 0.008473595604300499, + 0.008229691535234451, + 0.005257139448076487, + 0.006637115031480789, + 0.015512357465922832, + 0.01378627773374319, + 0.01618063636124134, + 0.010568752884864807, + 0.0171916875988245, + 0.011789602227509022, + 0.01639251410961151, + 0.014012268744409084, + 0.014764438383281231, + 0.011692419648170471, + 0.02920915186405182, + 0.017857875674962997, + 0.00908452458679676, + 0.022983601316809654, + 0.015972647815942764, + 0.017954248934984207, + 0.024026023223996162, + 0.023894954472780228, + 0.015318808145821095, + 0.016595685854554176, + 0.013866262510418892, + 0.01339699700474739, + 0.011610845103859901, + 0.020324524492025375, + 0.007830396294593811, + 0.018672311678528786, + 0.006166747771203518, + 0.0043001300655305386, + 0.016360748559236526, + 0.0065504820086061954, + 0.006238028407096863, + 0.0069310544058680534, + 0.006406151223927736, + 0.0067243417724967, + 0.0029549687169492245, + 0.0064179678447544575, + 0.00887352041900158, + 0.006361509207636118, + 0.005156611558049917, + 0.007363161072134972, + 0.015483729541301727, + 0.01777058281004429, + 0.009503301233053207, + 0.010521264746785164, + 0.01567121222615242, + 0.012570898048579693, + 0.01403985358774662, + 0.017260883003473282, + 0.004373541101813316, + 0.0034714345820248127, + 0.006465023849159479, + 0.00466589629650116, + 0.00461221020668745, + 0.007247077766805887, + 0.005823446903377771, + 0.0061349752359092236 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 5 + ], + "selected_reward": -0.038084905594587326 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the moth's thorax thick, hairy, and patterned with a bone-white skull design?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the moth's wings massive, velvety, and covered in a mosaic of deep chocolate brown, dusty ochre, and jet black?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the background pitch-black and void-like, illuminated by a single harsh overhead light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the moth's segmented abdomen contract violently while its feathery antennae quiver?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera tremble violently, causing the image to switch rapidly between out of focus and clarity?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -1.2236320972442627 + ], + "opposite_logprob_tokens": [ + -0.3486320972442627 + ], + "logprob_reward": -1.2236320972442627, + "yes_no_margin_reward": -0.875, + "token_credit_reward": -1.2236320972442627, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.14409920573234558, + 0.1100018322467804, + 0.14335885643959045, + 0.1608455330133438, + 0.054949693381786346, + 0.07135991007089615, + 0.151855006814003, + 0.07092998921871185, + 0.05188422650098801, + 0.0407157838344574 + ], + "credit_map": [ + 0.015208831056952477, + 0.022565627470612526, + 0.01973535493016243, + 0.015372910536825657, + 0.020301152020692825, + 0.016290783882141113, + 0.018497291952371597, + 0.016127245500683784, + 0.01168400701135397, + 0.011252142488956451, + 0.015602663159370422, + 0.011364255100488663, + 0.014453340321779251, + 0.01220274530351162, + 0.011761650443077087, + 0.02168101817369461, + 0.012456350028514862, + 0.010507838800549507, + 0.03737160563468933, + 0.009525774978101254, + 0.017302826046943665, + 0.015060982666909695, + 0.011601266451179981, + 0.029532207176089287, + 0.015900949016213417, + 0.012054702267050743, + 0.022670499980449677, + 0.015039918012917042, + 0.015609413385391235, + 0.02301979809999466, + 0.02725810557603836, + 0.029292121529579163, + 0.009152280166745186, + 0.004804983735084534, + 0.010014163330197334, + 0.004684098064899445, + 0.003971514757722616, + 0.008121130056679249, + 0.006607531104236841, + 0.007593989837914705, + 0.007034586276859045, + 0.00826056394726038, + 0.012659868225455284, + 0.00832374207675457, + 0.006802524905651808, + 0.011303297244012356, + 0.007700182497501373, + 0.009275143966078758, + 0.01695336401462555, + 0.03669700399041176, + 0.009668135084211826, + 0.012878057546913624, + 0.032975099980831146, + 0.01136026531457901, + 0.012414940632879734, + 0.018908126279711723, + 0.007488733157515526, + 0.010649709030985832, + 0.00540766678750515, + 0.006304359063506126, + 0.011591285467147827, + 0.010217978619039059, + 0.008897578343749046, + 0.01037267129868269, + 0.006824549753218889, + 0.006424893159419298, + 0.004258384462445974, + 0.005941327661275864, + 0.008488285355269909, + 0.00451362319290638, + 0.007183389738202095, + 0.008249771781265736, + 0.005295373033732176, + 0.0036135201808065176, + 0.004201735369861126, + 0.007301820907741785, + 0.0035068311262875795, + 0.006697661243379116, + 0.005854435730725527, + 0.004244402050971985 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 6, + 0 + ], + "selected_reward": -1.2236320972442627 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the moth's thorax thick, hairy, and patterned with a bone-white skull design?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the moth's wings massive, velvety, and covered in a mosaic of deep chocolate brown, dusty ochre, and jet black?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the background pitch-black and void-like, illuminated by a single harsh overhead light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the moth's segmented abdomen contract violently while its feathery antennae quiver?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera tremble violently, causing the image to switch rapidly between out of focus and clarity?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -0.6934284567832947 + ], + "opposite_logprob_tokens": [ + -0.6934284567832947 + ], + "logprob_reward": -0.6934284567832947, + "yes_no_margin_reward": 0.0, + "token_credit_reward": -0.6934284567832947, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.09424825757741928, + 0.07560621201992035, + 0.12682248651981354, + 0.06343237310647964, + 0.08435969054698944, + 0.08148551732301712, + 0.32347214221954346, + 0.053725048899650574, + 0.07578562945127487, + 0.02106272615492344 + ], + "credit_map": [ + 0.007462840061634779, + 0.014209074899554253, + 0.01082586869597435, + 0.011105348356068134, + 0.016475193202495575, + 0.010831191204488277, + 0.012899658642709255, + 0.010439072735607624, + 0.006869051139801741, + 0.010347951203584671, + 0.010746720246970654, + 0.00702825840562582, + 0.01010082382708788, + 0.00811082310974598, + 0.006804536562412977, + 0.015598051249980927, + 0.012404109351336956, + 0.013243108987808228, + 0.03150027245283127, + 0.006419970653951168, + 0.01779096946120262, + 0.014508355408906937, + 0.008294649422168732, + 0.022661060094833374, + 0.007485857233405113, + 0.005688648205250502, + 0.010044545866549015, + 0.005328007973730564, + 0.007170730736106634, + 0.006944678258150816, + 0.009682425297796726, + 0.011087479069828987, + 0.009998200461268425, + 0.007122306618839502, + 0.01613778993487358, + 0.009480498731136322, + 0.006275026593357325, + 0.013936178758740425, + 0.010212098248302937, + 0.011197598651051521, + 0.008737063966691494, + 0.0087361391633749, + 0.01289680227637291, + 0.010075095109641552, + 0.008438700810074806, + 0.010772916488349438, + 0.009983927942812443, + 0.01184486411511898, + 0.031941842287778854, + 0.07023247331380844, + 0.02742735855281353, + 0.02876308001577854, + 0.0654677078127861, + 0.026262152940034866, + 0.036439742892980576, + 0.036937762051820755, + 0.007654921151697636, + 0.008620115928351879, + 0.0026349846739321947, + 0.004673943389207125, + 0.010774782858788967, + 0.006630145013332367, + 0.005829447880387306, + 0.006906706374138594, + 0.00894148275256157, + 0.010055731050670147, + 0.007327140774577856, + 0.0060906121507287025, + 0.013081982731819153, + 0.007987959310412407, + 0.008117547258734703, + 0.014183171093463898, + 0.0016346047632396221, + 0.002440792042762041, + 0.003268689149990678, + 0.0035858748015016317, + 0.0024372420739382505, + 0.0020810433197766542, + 0.0031837422866374254, + 0.0024307379499077797 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 2, + 0 + ], + "selected_reward": -0.6934284567832947 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "No", + "No" + ] + }, + { + "prompt": "high-fidelity CGI animation, rendered in extreme close-up, focuses on a massive, ghostly-faced hawk moth, its thick, hairy thorax rendered with a disturbing, bone-white skull pattern. Sculpted from a pitch-black, void-like background by a single, harsh overhead light, the giant moth unfolds its massive, velvety wings, covered in an intricate mosaic of deep chocolate brown, dusty ochre, and jet black, each microscopic scale visible with crystalline clarity. Its segmented abdomen contracts violently, its feathery antennae quivering with vibrations as it lets out a piercing, almost electronic howl. In sync with the scream, the unstable camera also trembled violently, as if hit by sound waves, causing every hyper-realistic detail in the picture - from its huge compound eyes flashing with highlights to the few scales floating in the air due to violent vibrations - to quickly switch between out of focus and clarity, creating a suffocating sense of fear from biological instinct and a strong visual and auditory impact.", + "seed": 15322, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000101/group_11/index_732_seed_15322_dp_22_rank_44_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 11, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.24486231952905654, + "vlm_logprob": -0.24486231952905654, + "vlm_yes_no_margin": 2.175, + "vlm_token_credit": -0.24486231952905654, + "vlm_yes_ratio": 0.8, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the moth's thorax thick, hairy, and patterned with a bone-white skull design?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the moth's wings massive, velvety, and covered in a mosaic of deep chocolate brown, dusty ochre, and jet black?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the background pitch-black and void-like, illuminated by a single harsh overhead light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the moth's segmented abdomen contract violently while its feathery antennae quiver?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera tremble violently, causing the image to switch rapidly between out of focus and clarity?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04485540837049484 + ], + "opposite_logprob_tokens": [ + -3.2948553562164307 + ], + "logprob_reward": -0.04485540837049484, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.04485540837049484, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14097793400287628, + 0.06547350436449051, + 0.07360158115625381, + 0.06190330907702446, + 0.13482630252838135, + 0.035346128046512604, + 0.24820014834403992, + 0.049927812069654465, + 0.07775848358869553, + 0.11198484897613525 + ], + "credit_map": [ + 0.014625719748437405, + 0.029643412679433823, + 0.016468429937958717, + 0.013971704058349133, + 0.023110006004571915, + 0.020885799080133438, + 0.01324114203453064, + 0.009031714871525764, + 0.010079236701130867, + 0.010096394456923008, + 0.005915107671171427, + 0.006407332140952349, + 0.009402136318385601, + 0.011443207040429115, + 0.005497905891388655, + 0.006632174830883741, + 0.010502755641937256, + 0.008970196358859539, + 0.00612195860594511, + 0.00997452437877655, + 0.014672727324068546, + 0.008145598694682121, + 0.008656135760247707, + 0.006557680666446686, + 0.015267346985638142, + 0.015114485286176205, + 0.007771431468427181, + 0.003154020756483078, + 0.00517882127314806, + 0.004651093855500221, + 0.004489358048886061, + 0.006276749074459076, + 0.01538122072815895, + 0.016792548820376396, + 0.009744195267558098, + 0.013318978250026703, + 0.03492121770977974, + 0.0207907035946846, + 0.013291439041495323, + 0.010585988871753216, + 0.0031564149539917707, + 0.00271730893291533, + 0.0023806970566511154, + 0.0025932800490409136, + 0.0038082727696746588, + 0.010284587740898132, + 0.008251500315964222, + 0.002154064364731312, + 0.04332831874489784, + 0.01069214940071106, + 0.014732209965586662, + 0.025597726926207542, + 0.01086242776364088, + 0.0209059976041317, + 0.040392424911260605, + 0.08168888837099075, + 0.004343648906797171, + 0.006046632304787636, + 0.006949455011636019, + 0.006631582975387573, + 0.005277465097606182, + 0.009280093014240265, + 0.0050923023372888565, + 0.006306629162281752, + 0.007706036325544119, + 0.006561310961842537, + 0.009526856243610382, + 0.007613157853484154, + 0.009715796448290348, + 0.012034634128212929, + 0.012286242097616196, + 0.01231444627046585, + 0.011610829271376133, + 0.010673833079636097, + 0.009084422141313553, + 0.01696181483566761, + 0.012199885211884975, + 0.01679222658276558, + 0.01907614804804325, + 0.015585682354867458 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 0, + 4 + ], + "selected_reward": -0.04485540837049484 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the moth's thorax thick, hairy, and patterned with a bone-white skull design?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the moth's wings massive, velvety, and covered in a mosaic of deep chocolate brown, dusty ochre, and jet black?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the background pitch-black and void-like, illuminated by a single harsh overhead light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the moth's segmented abdomen contract violently while its feathery antennae quiver?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera tremble violently, causing the image to switch rapidly between out of focus and clarity?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02329517900943756 + ], + "opposite_logprob_tokens": [ + -3.7732951641082764 + ], + "logprob_reward": -0.02329517900943756, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.02329517900943756, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15994873642921448, + 0.11862694472074509, + 0.10002031177282333, + 0.12485875934362411, + 0.07363726943731308, + 0.03343966603279114, + 0.12615327537059784, + 0.030422991141676903, + 0.05531030520796776, + 0.1775817573070526 + ], + "credit_map": [ + 0.01667342148721218, + 0.02905750460922718, + 0.022789467126131058, + 0.017385629937052727, + 0.0215329360216856, + 0.022091520950198174, + 0.01603550836443901, + 0.014382747933268547, + 0.01623959094285965, + 0.01234907191246748, + 0.013234681449830532, + 0.013583473861217499, + 0.018969180062413216, + 0.014437025412917137, + 0.012784712947905064, + 0.017029210925102234, + 0.015594442375004292, + 0.012949702329933643, + 0.009905689395964146, + 0.00846911408007145, + 0.017519710585474968, + 0.014028375968337059, + 0.009454525075852871, + 0.012098751030862331, + 0.016196122393012047, + 0.01824273355305195, + 0.01700695790350437, + 0.006932833697646856, + 0.013786811381578445, + 0.012301977723836899, + 0.012104634195566177, + 0.028286689892411232, + 0.007074545603245497, + 0.007633980829268694, + 0.005851649679243565, + 0.0074457284063100815, + 0.020094363018870354, + 0.011345657519996166, + 0.009225831367075443, + 0.004965517669916153, + 0.0027826414443552494, + 0.002432482549920678, + 0.0022325466852635145, + 0.0024656548630446196, + 0.0036849298048764467, + 0.009156420826911926, + 0.008008627220988274, + 0.0026763638015836477, + 0.019955286756157875, + 0.005436406005173922, + 0.00629641767591238, + 0.01314800325781107, + 0.006099934224039316, + 0.010986718349158764, + 0.017350459471344948, + 0.04688005894422531, + 0.0027813445776700974, + 0.003507449524477124, + 0.004078980535268784, + 0.0037868234794586897, + 0.0034188320860266685, + 0.00554047292098403, + 0.0036831635516136885, + 0.003625923302024603, + 0.005869465414434671, + 0.004712957888841629, + 0.005700974725186825, + 0.005227698478847742, + 0.006886867806315422, + 0.008968503214418888, + 0.009209873154759407, + 0.008733967319130898, + 0.013301213271915913, + 0.014378363266587257, + 0.014263295568525791, + 0.022310012951493263, + 0.027594134211540222, + 0.035639919340610504, + 0.025584444403648376, + 0.024510378018021584 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 0, + 6 + ], + "selected_reward": -0.02329517900943756 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the moth's thorax thick, hairy, and patterned with a bone-white skull design?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the moth's wings massive, velvety, and covered in a mosaic of deep chocolate brown, dusty ochre, and jet black?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the background pitch-black and void-like, illuminated by a single harsh overhead light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the moth's segmented abdomen contract violently while its feathery antennae quiver?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera tremble violently, causing the image to switch rapidly between out of focus and clarity?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.048647284507751465 + ], + "opposite_logprob_tokens": [ + -3.048647403717041 + ], + "logprob_reward": -0.048647284507751465, + "yes_no_margin_reward": 3.0, + "token_credit_reward": -0.048647284507751465, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.38109833002090454, + 0.19613242149353027, + 0.06947145611047745, + 0.040057554841041565, + 0.06235538050532341, + 0.04866422712802887, + 0.0583375059068203, + 0.017206842079758644, + 0.06882935017347336, + 0.057846978306770325 + ], + "credit_map": [ + 0.03704133257269859, + 0.09030426293611526, + 0.04362258315086365, + 0.03484765812754631, + 0.05810307338833809, + 0.05711319297552109, + 0.028791051357984543, + 0.03127514198422432, + 0.03472896292805672, + 0.03789439797401428, + 0.012791050598025322, + 0.01878320425748825, + 0.025748949497938156, + 0.03890899941325188, + 0.013377114199101925, + 0.013899753801524639, + 0.010637815110385418, + 0.008334112353622913, + 0.006978444289416075, + 0.005500003695487976, + 0.012426926754415035, + 0.008141093887388706, + 0.008041576482355595, + 0.009411484934389591, + 0.005810968112200499, + 0.006376182660460472, + 0.0059920065104961395, + 0.0032041803933680058, + 0.005825564730912447, + 0.004130224697291851, + 0.0034063830971717834, + 0.005312045104801655, + 0.006613271776586771, + 0.007233533076941967, + 0.007474875543266535, + 0.00605423841625452, + 0.009888493455946445, + 0.010089680552482605, + 0.0055049872025847435, + 0.009496296755969524, + 0.004271534271538258, + 0.0023732429835945368, + 0.0036755064502358437, + 0.0027277555782347918, + 0.004062793683260679, + 0.015387923456728458, + 0.013052399270236492, + 0.00311306887306273, + 0.009750006720423698, + 0.0029761919286102057, + 0.004711084067821503, + 0.005780654959380627, + 0.0032662127632647753, + 0.0049195969477295876, + 0.007980981841683388, + 0.018952777609229088, + 0.001233415212482214, + 0.0019428149098530412, + 0.0028783006127923727, + 0.0024714977480471134, + 0.0014053168706595898, + 0.003544637467712164, + 0.001510409638285637, + 0.0022204495035111904, + 0.009339055977761745, + 0.005651248153299093, + 0.0081398393958807, + 0.005611867178231478, + 0.00924726203083992, + 0.012685664929449558, + 0.0071113393642008305, + 0.011043073609471321, + 0.005568487104028463, + 0.00496519124135375, + 0.0040278504602611065, + 0.006808901205658913, + 0.009476191364228725, + 0.011403392069041729, + 0.007348847575485706, + 0.008248113095760345 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 2 + ], + "selected_reward": -0.048647284507751465 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the moth's thorax thick, hairy, and patterned with a bone-white skull design?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the moth's wings massive, velvety, and covered in a mosaic of deep chocolate brown, dusty ochre, and jet black?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the background pitch-black and void-like, illuminated by a single harsh overhead light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the moth's segmented abdomen contract violently while its feathery antennae quiver?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera tremble violently, causing the image to switch rapidly between out of focus and clarity?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -0.8261575698852539 + ], + "opposite_logprob_tokens": [ + -0.5761575698852539 + ], + "logprob_reward": -0.8261575698852539, + "yes_no_margin_reward": -0.25, + "token_credit_reward": -0.8261575698852539, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.18271081149578094, + 0.10860507935285568, + 0.10360526293516159, + 0.19518287479877472, + 0.08553681522607803, + 0.0862724632024765, + 0.04829437658190727, + 0.032507967203855515, + 0.045533422380685806, + 0.11175093054771423 + ], + "credit_map": [ + 0.015154504217207432, + 0.03671928495168686, + 0.022759124636650085, + 0.02079976163804531, + 0.02698901668190956, + 0.026298746466636658, + 0.01960468478500843, + 0.014385698363184929, + 0.015540770255029202, + 0.01826823502779007, + 0.014517635107040405, + 0.01358798984438181, + 0.015551405027508736, + 0.012454315088689327, + 0.008984619751572609, + 0.009700112976133823, + 0.015465867705643177, + 0.011629078537225723, + 0.010779103264212608, + 0.012432495132088661, + 0.021107669919729233, + 0.011549675837159157, + 0.009732464328408241, + 0.010908904485404491, + 0.05411909893155098, + 0.05418669804930687, + 0.013542333617806435, + 0.009058590047061443, + 0.020086107775568962, + 0.01605243794620037, + 0.013312913477420807, + 0.01482469029724598, + 0.008958445861935616, + 0.011916016228497028, + 0.01044928003102541, + 0.0071135093457996845, + 0.017532505095005035, + 0.012940318323671818, + 0.007426337338984013, + 0.009200400672852993, + 0.00732340756803751, + 0.006708653178066015, + 0.0036419215612113476, + 0.00583959324285388, + 0.005597066134214401, + 0.02705986797809601, + 0.02572452835738659, + 0.004377423785626888, + 0.007135732099413872, + 0.0029637666884809732, + 0.004503775388002396, + 0.011471244506537914, + 0.003873629728332162, + 0.005110907834023237, + 0.007300740107893944, + 0.005934576969593763, + 0.0033047872129827738, + 0.005222881678491831, + 0.0037050170358270407, + 0.0037825170438736677, + 0.0035239264834672213, + 0.00404832698404789, + 0.005001915153115988, + 0.003918598406016827, + 0.004857318475842476, + 0.004942554980516434, + 0.006528682075440884, + 0.006039384752511978, + 0.004640009719878435, + 0.006951868534088135, + 0.005329090170562267, + 0.00624451506882906, + 0.011667224578559399, + 0.009364776313304901, + 0.009545081295073032, + 0.015284797176718712, + 0.010505331680178642, + 0.02137617953121662, + 0.016916081309318542, + 0.017091458663344383 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 0, + 9 + ], + "selected_reward": -0.8261575698852539 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the moth's thorax thick, hairy, and patterned with a bone-white skull design?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the moth's wings massive, velvety, and covered in a mosaic of deep chocolate brown, dusty ochre, and jet black?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the background pitch-black and void-like, illuminated by a single harsh overhead light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the moth's segmented abdomen contract violently while its feathery antennae quiver?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera tremble violently, causing the image to switch rapidly between out of focus and clarity?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.28135615587234497 + ], + "opposite_logprob_tokens": [ + -1.4063560962677002 + ], + "logprob_reward": -0.28135615587234497, + "yes_no_margin_reward": 1.125, + "token_credit_reward": -0.28135615587234497, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.17518608272075653, + 0.16739630699157715, + 0.16905725002288818, + 0.1420111507177353, + 0.04295559227466583, + 0.060406994074583054, + 0.10496875643730164, + 0.03679514676332474, + 0.04957290366292, + 0.0516497902572155 + ], + "credit_map": [ + 0.01936890184879303, + 0.029416518285870552, + 0.02963765524327755, + 0.0191859919577837, + 0.02141023427248001, + 0.02459750324487686, + 0.015662886202335358, + 0.015906384214758873, + 0.02199077419936657, + 0.03746549412608147, + 0.016771411523222923, + 0.01900378055870533, + 0.02221875637769699, + 0.022363130003213882, + 0.015162898227572441, + 0.012420051731169224, + 0.02387937530875206, + 0.023632530122995377, + 0.014277813024818897, + 0.019356293603777885, + 0.027550671249628067, + 0.022740023210644722, + 0.016369571909308434, + 0.02125096507370472, + 0.04247988760471344, + 0.03779007866978645, + 0.01617118902504444, + 0.006099981721490622, + 0.009816955775022507, + 0.007969607599079609, + 0.009593675844371319, + 0.012089788913726807, + 0.003836155403405428, + 0.00605808338150382, + 0.0039017782546579838, + 0.004703111946582794, + 0.008969947695732117, + 0.0055222287774086, + 0.005719294305890799, + 0.004244992509484291, + 0.0055824569426476955, + 0.0047929356805980206, + 0.004070356488227844, + 0.006354869809001684, + 0.005701933056116104, + 0.016775935888290405, + 0.013338664546608925, + 0.0037898398004472256, + 0.019177937880158424, + 0.006522095762193203, + 0.008960051462054253, + 0.01991734839975834, + 0.008143824525177479, + 0.009888208471238613, + 0.013859913684427738, + 0.018499376252293587, + 0.0031390441581606865, + 0.006482934113591909, + 0.004573252983391285, + 0.0040321433916687965, + 0.004793477710336447, + 0.0054013775661587715, + 0.004325583577156067, + 0.004047333262860775, + 0.004835518077015877, + 0.003846685867756605, + 0.006293075159192085, + 0.004919474478811026, + 0.005552157759666443, + 0.00929784495383501, + 0.006868663243949413, + 0.007959484122693539, + 0.0048131076619029045, + 0.0038658196572214365, + 0.003971054218709469, + 0.006293693091720343, + 0.005722097586840391, + 0.010050312615931034, + 0.008391092531383038, + 0.008542612195014954 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 1 + ], + "selected_reward": -0.28135615587234497 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "No", + "Yes" + ] + }, + { + "prompt": "high-fidelity CGI animation, rendered in extreme close-up, focuses on a massive, ghostly-faced hawk moth, its thick, hairy thorax rendered with a disturbing, bone-white skull pattern. Sculpted from a pitch-black, void-like background by a single, harsh overhead light, the giant moth unfolds its massive, velvety wings, covered in an intricate mosaic of deep chocolate brown, dusty ochre, and jet black, each microscopic scale visible with crystalline clarity. Its segmented abdomen contracts violently, its feathery antennae quivering with vibrations as it lets out a piercing, almost electronic howl. In sync with the scream, the unstable camera also trembled violently, as if hit by sound waves, causing every hyper-realistic detail in the picture - from its huge compound eyes flashing with highlights to the few scales floating in the air due to violent vibrations - to quickly switch between out of focus and clarity, creating a suffocating sense of fear from biological instinct and a strong visual and auditory impact.", + "seed": 15323, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000101/group_11/index_732_seed_15323_dp_23_rank_46_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 11, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.3938585026189685, + "vlm_logprob": -0.3938585026189685, + "vlm_yes_no_margin": 1.875, + "vlm_token_credit": -0.3938585026189685, + "vlm_yes_ratio": 0.6, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the moth's thorax thick, hairy, and patterned with a bone-white skull design?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the moth's wings massive, velvety, and covered in a mosaic of deep chocolate brown, dusty ochre, and jet black?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the background pitch-black and void-like, illuminated by a single harsh overhead light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the moth's segmented abdomen contract violently while its feathery antennae quiver?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera tremble violently, causing the image to switch rapidly between out of focus and clarity?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01782587170600891 + ], + "opposite_logprob_tokens": [ + -4.392826080322266 + ], + "logprob_reward": -0.01782587170600891, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.01782587170600891, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13146807253360748, + 0.0855865553021431, + 0.06391457468271255, + 0.14150117337703705, + 0.03756726533174515, + 0.08072337508201599, + 0.10601714253425598, + 0.09615825116634369, + 0.18140237033367157, + 0.07566113770008087 + ], + "credit_map": [ + 0.020694006234407425, + 0.014064904302358627, + 0.017831534147262573, + 0.01363279391080141, + 0.019993938505649567, + 0.019618621096014977, + 0.012880893424153328, + 0.012751398608088493, + 0.01885395497083664, + 0.006776486989110708, + 0.010577059350907803, + 0.007234814111143351, + 0.011055370792746544, + 0.01401527225971222, + 0.005936611909419298, + 0.011136979795992374, + 0.012369927018880844, + 0.006588535383343697, + 0.0062543731182813644, + 0.007656710688024759, + 0.007326854858547449, + 0.010641459375619888, + 0.00596980657428503, + 0.007106906268745661, + 0.044666558504104614, + 0.007994625717401505, + 0.013014053925871849, + 0.007805626839399338, + 0.006914179772138596, + 0.009607061743736267, + 0.011361194774508476, + 0.040137879550457, + 0.006066320464015007, + 0.0034259373787790537, + 0.005223399959504604, + 0.003462434746325016, + 0.003907585050910711, + 0.0043266103602945805, + 0.0048591941595077515, + 0.006295784842222929, + 0.009734835475683212, + 0.006792683620005846, + 0.015504182316362858, + 0.008860200643539429, + 0.007556782104074955, + 0.010078868828713894, + 0.009776508435606956, + 0.01241931039839983, + 0.01460421085357666, + 0.010232214815914631, + 0.016764359548687935, + 0.007118863053619862, + 0.010511604137718678, + 0.015351067297160625, + 0.015440731309354305, + 0.015994088724255562, + 0.014679303392767906, + 0.00894983857870102, + 0.017891986295580864, + 0.012258767150342464, + 0.00814793724566698, + 0.010278998874127865, + 0.013196048326790333, + 0.010755371302366257, + 0.029650356620550156, + 0.01633838750422001, + 0.021850721910595894, + 0.017822502180933952, + 0.023535694926977158, + 0.025359122082591057, + 0.02593117766082287, + 0.020914414897561073, + 0.007464507594704628, + 0.01190477330237627, + 0.010114844888448715, + 0.011448710225522518, + 0.009402827359735966, + 0.007772394455969334, + 0.008901613764464855, + 0.008651472628116608 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 3, + 0 + ], + "selected_reward": -0.01782587170600891 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the moth's thorax thick, hairy, and patterned with a bone-white skull design?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the moth's wings massive, velvety, and covered in a mosaic of deep chocolate brown, dusty ochre, and jet black?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the background pitch-black and void-like, illuminated by a single harsh overhead light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the moth's segmented abdomen contract violently while its feathery antennae quiver?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera tremble violently, causing the image to switch rapidly between out of focus and clarity?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01108264084905386 + ], + "opposite_logprob_tokens": [ + -4.511082649230957 + ], + "logprob_reward": -0.01108264084905386, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.01108264084905386, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09673529118299484, + 0.10762333869934082, + 0.07369045913219452, + 0.1640448421239853, + 0.09811890870332718, + 0.08459187299013138, + 0.13470759987831116, + 0.0847092866897583, + 0.06785021722316742, + 0.08792810142040253 + ], + "credit_map": [ + 0.01630312204360962, + 0.00897328183054924, + 0.013350349850952625, + 0.010573722422122955, + 0.012594867497682571, + 0.014108721166849136, + 0.00941983051598072, + 0.01141140516847372, + 0.016833586618304253, + 0.012025526724755764, + 0.015492401085793972, + 0.008666588924825191, + 0.01673722453415394, + 0.016399256885051727, + 0.01003945805132389, + 0.01142930518835783, + 0.012083824723958969, + 0.0115357032045722, + 0.007043569348752499, + 0.012218854390084743, + 0.0074664270505309105, + 0.008101831190288067, + 0.007845869287848473, + 0.007394388318061829, + 0.0570615753531456, + 0.008111684583127499, + 0.016355689615011215, + 0.0070415884256362915, + 0.007564974017441273, + 0.009486914612352848, + 0.011484037153422832, + 0.04693838208913803, + 0.020162321627140045, + 0.008241807110607624, + 0.014426428824663162, + 0.009285886771976948, + 0.008037269115447998, + 0.010034700855612755, + 0.015084576793015003, + 0.012845931574702263, + 0.01058971881866455, + 0.008060717023909092, + 0.01210738718509674, + 0.008846851997077465, + 0.009274653159081936, + 0.013014720752835274, + 0.010007196106016636, + 0.01269063726067543, + 0.022908173501491547, + 0.010710232891142368, + 0.018638981506228447, + 0.007016679737716913, + 0.011479631997644901, + 0.026428567245602608, + 0.01765608787536621, + 0.019869264215230942, + 0.013379612937569618, + 0.00969904288649559, + 0.016521429643034935, + 0.009935741312801838, + 0.006266838870942593, + 0.009180409833788872, + 0.009192096069455147, + 0.010534117929637432, + 0.009364533238112926, + 0.007202046923339367, + 0.008798143826425076, + 0.007190450094640255, + 0.009598976001143456, + 0.007538315374404192, + 0.009145411662757397, + 0.009012348018586636, + 0.009607626125216484, + 0.015452324412763119, + 0.011188858188688755, + 0.007889739237725735, + 0.017237670719623566, + 0.007929903455078602, + 0.009809838607907295, + 0.008812147192656994 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 6, + 1 + ], + "selected_reward": -0.01108264084905386 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the moth's thorax thick, hairy, and patterned with a bone-white skull design?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the moth's wings massive, velvety, and covered in a mosaic of deep chocolate brown, dusty ochre, and jet black?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the background pitch-black and void-like, illuminated by a single harsh overhead light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the moth's segmented abdomen contract violently while its feathery antennae quiver?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera tremble violently, causing the image to switch rapidly between out of focus and clarity?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.3486303389072418 + ], + "opposite_logprob_tokens": [ + -1.2236303091049194 + ], + "logprob_reward": -0.3486303389072418, + "yes_no_margin_reward": 0.875, + "token_credit_reward": -0.3486303389072418, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.20607756078243256, + 0.04926695302128792, + 0.08334442228078842, + 0.12249606847763062, + 0.050240423530340195, + 0.03533082455396652, + 0.07834439724683762, + 0.10285834968090057, + 0.06775689125061035, + 0.20428410172462463 + ], + "credit_map": [ + 0.03761335834860802, + 0.013335702940821648, + 0.032883740961551666, + 0.011613103561103344, + 0.031139569357037544, + 0.028402401134371758, + 0.017116915434598923, + 0.03397276997566223, + 0.007181781809777021, + 0.004171589855104685, + 0.00783462356775999, + 0.004095545969903469, + 0.008352924138307571, + 0.0056421165354549885, + 0.00622823741286993, + 0.005760136526077986, + 0.016679011285305023, + 0.007163474336266518, + 0.00849917158484459, + 0.012969061732292175, + 0.008903407491743565, + 0.012086762115359306, + 0.008413492701947689, + 0.008630039170384407, + 0.030405165627598763, + 0.007145459298044443, + 0.014830012805759907, + 0.009283989667892456, + 0.008061706088483334, + 0.012646656483411789, + 0.010148044675588608, + 0.029975034296512604, + 0.007708700839430094, + 0.004719293676316738, + 0.007857859134674072, + 0.00512916361913085, + 0.004684607032686472, + 0.004448925144970417, + 0.008075634948909283, + 0.007616239134222269, + 0.0033927944023162127, + 0.0028265113942325115, + 0.0077029927633702755, + 0.0033518588170409203, + 0.003962777089327574, + 0.004003844223916531, + 0.0038866447284817696, + 0.006203401368111372, + 0.015041201375424862, + 0.006312108598649502, + 0.011174248531460762, + 0.0038332082331180573, + 0.005869319196790457, + 0.012631279416382313, + 0.009718653745949268, + 0.013764383271336555, + 0.014714574441313744, + 0.012132822535932064, + 0.02051238901913166, + 0.010527128353714943, + 0.009005953557789326, + 0.015428160317242146, + 0.008790215477347374, + 0.01174710039049387, + 0.010883225128054619, + 0.005940506234765053, + 0.008463910780847073, + 0.0074525391682982445, + 0.009466925635933876, + 0.008692019619047642, + 0.00968408863991499, + 0.007173680700361729, + 0.02161289192736149, + 0.04179587587714195, + 0.024262063205242157, + 0.016024351119995117, + 0.04179389029741287, + 0.018592728301882744, + 0.02144950069487095, + 0.018752798438072205 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 3 + ], + "selected_reward": -0.3486303389072418 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the moth's thorax thick, hairy, and patterned with a bone-white skull design?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the moth's wings massive, velvety, and covered in a mosaic of deep chocolate brown, dusty ochre, and jet black?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the background pitch-black and void-like, illuminated by a single harsh overhead light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the moth's segmented abdomen contract violently while its feathery antennae quiver?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera tremble violently, causing the image to switch rapidly between out of focus and clarity?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -0.8983432054519653 + ], + "opposite_logprob_tokens": [ + -0.5233432054519653 + ], + "logprob_reward": -0.8983432054519653, + "yes_no_margin_reward": -0.375, + "token_credit_reward": -0.8983432054519653, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.055421821773052216, + 0.05605807527899742, + 0.06260277330875397, + 0.23305043578147888, + 0.0467175617814064, + 0.041988957673311234, + 0.06988102942705154, + 0.25713586807250977, + 0.06435036659240723, + 0.11279304325580597 + ], + "credit_map": [ + 0.008524524047970772, + 0.005952760111540556, + 0.007254438009113073, + 0.004864070564508438, + 0.00789120513945818, + 0.0071203685365617275, + 0.006382784806191921, + 0.007431679405272007, + 0.011278771795332432, + 0.006072999909520149, + 0.007273240014910698, + 0.005316643510013819, + 0.00869922898709774, + 0.00715743750333786, + 0.005180202890187502, + 0.0050795599818229675, + 0.009395187720656395, + 0.008445726707577705, + 0.0084824338555336, + 0.010400354862213135, + 0.007135781925171614, + 0.00669974135234952, + 0.006247187033295631, + 0.005796361714601517, + 0.06926165521144867, + 0.017420928925275803, + 0.020400311797857285, + 0.023779328912496567, + 0.013477854430675507, + 0.015348784625530243, + 0.017673257738351822, + 0.05568832531571388, + 0.009830152615904808, + 0.0033324051182717085, + 0.006881671492010355, + 0.0044779288582503796, + 0.0036563887260854244, + 0.006012909114360809, + 0.005887968465685844, + 0.006638142745941877, + 0.006159285549074411, + 0.0026094159111380577, + 0.008178615011274815, + 0.004145212937146425, + 0.002797066932544112, + 0.005425529088824987, + 0.004007978364825249, + 0.008665859699249268, + 0.008552315644919872, + 0.009369445033371449, + 0.010552763938903809, + 0.006449067033827305, + 0.0069722807966172695, + 0.008389619179069996, + 0.01024720724672079, + 0.009348343126475811, + 0.031908608973026276, + 0.030902864411473274, + 0.03206794336438179, + 0.05988699570298195, + 0.021255943924188614, + 0.023165730759501457, + 0.028644977137446404, + 0.029302822425961494, + 0.008082158863544464, + 0.009391412138938904, + 0.008273538202047348, + 0.008275328204035759, + 0.008500236086547375, + 0.007766841445118189, + 0.007283465005457401, + 0.006777393165975809, + 0.012312016449868679, + 0.025270532816648483, + 0.012659337371587753, + 0.011384297162294388, + 0.021248532459139824, + 0.009858249686658382, + 0.008945273235440254, + 0.011114816181361675 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 3, + 9 + ], + "selected_reward": -0.8983432054519653 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the moth's thorax thick, hairy, and patterned with a bone-white skull design?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the moth's wings massive, velvety, and covered in a mosaic of deep chocolate brown, dusty ochre, and jet black?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the background pitch-black and void-like, illuminated by a single harsh overhead light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the moth's segmented abdomen contract violently while its feathery antennae quiver?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera tremble violently, causing the image to switch rapidly between out of focus and clarity?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -0.6934104561805725 + ], + "opposite_logprob_tokens": [ + -0.6934104561805725 + ], + "logprob_reward": -0.6934104561805725, + "yes_no_margin_reward": 0.0, + "token_credit_reward": -0.6934104561805725, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.05388193577528, + 0.048613253980875015, + 0.06989239156246185, + 0.08164981752634048, + 0.05614417791366577, + 0.09704036265611649, + 0.05133315920829773, + 0.13192583620548248, + 0.3160363435745239, + 0.09348271787166595 + ], + "credit_map": [ + 0.007562714163213968, + 0.006476305425167084, + 0.007737098727375269, + 0.005521560087800026, + 0.0071426876820623875, + 0.006691135466098785, + 0.006798414047807455, + 0.005952021107077599, + 0.006804729346185923, + 0.0065844194032251835, + 0.007040798664093018, + 0.0060676271095871925, + 0.006185414735227823, + 0.005662240087985992, + 0.004751549568027258, + 0.005516478326171637, + 0.010292282328009605, + 0.007398573681712151, + 0.008556214161217213, + 0.011872474104166031, + 0.010717425495386124, + 0.00725935772061348, + 0.007520313840359449, + 0.006275748834013939, + 0.01613927073776722, + 0.00765636982396245, + 0.01017427071928978, + 0.011264902539551258, + 0.00653116125613451, + 0.009268589317798615, + 0.009077413007616997, + 0.011537839658558369, + 0.008372683078050613, + 0.004355453420430422, + 0.008444488979876041, + 0.004817121662199497, + 0.006295058876276016, + 0.00723447697237134, + 0.008415380492806435, + 0.008209512569010258, + 0.014994810335338116, + 0.006706725340336561, + 0.019757764413952827, + 0.009328741580247879, + 0.009117759764194489, + 0.010603734292089939, + 0.00856603030115366, + 0.017964793369174004, + 0.00881139189004898, + 0.004040627274662256, + 0.007332280278205872, + 0.003262947779148817, + 0.004630279261618853, + 0.00640610046684742, + 0.006026197224855423, + 0.010823335498571396, + 0.018118197098374367, + 0.01206347718834877, + 0.020309465005993843, + 0.019613241776823997, + 0.012257634662091732, + 0.01846383512020111, + 0.013718691654503345, + 0.017381293699145317, + 0.054295554757118225, + 0.0262870192527771, + 0.03618022799491882, + 0.03369646891951561, + 0.04259264096617699, + 0.04348132386803627, + 0.047398705035448074, + 0.032104410231113434, + 0.009478291496634483, + 0.017114512622356415, + 0.011940228752791882, + 0.008864283561706543, + 0.017364321276545525, + 0.008479274809360504, + 0.009465260431170464, + 0.010776548646390438 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 7, + 5 + ], + "selected_reward": -0.6934104561805725 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "No", + "No" + ] + }, + { + "prompt": "high-fidelity CGI animation, rendered in extreme close-up, focuses on a massive, ghostly-faced hawk moth, its thick, hairy thorax rendered with a disturbing, bone-white skull pattern. Sculpted from a pitch-black, void-like background by a single, harsh overhead light, the giant moth unfolds its massive, velvety wings, covered in an intricate mosaic of deep chocolate brown, dusty ochre, and jet black, each microscopic scale visible with crystalline clarity. Its segmented abdomen contracts violently, its feathery antennae quivering with vibrations as it lets out a piercing, almost electronic howl. In sync with the scream, the unstable camera also trembled violently, as if hit by sound waves, causing every hyper-realistic detail in the picture - from its huge compound eyes flashing with highlights to the few scales floating in the air due to violent vibrations - to quickly switch between out of focus and clarity, creating a suffocating sense of fear from biological instinct and a strong visual and auditory impact.", + "seed": 15324, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000101/group_11/index_732_seed_15324_dp_23_rank_46_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 11, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.3449770241975784, + "vlm_logprob": -0.3449770241975784, + "vlm_yes_no_margin": 1.574999988079071, + "vlm_token_credit": -0.3449770241975784, + "vlm_yes_ratio": 0.8, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the moth's thorax thick, hairy, and patterned with a bone-white skull design?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the moth's wings massive, velvety, and covered in a mosaic of deep chocolate brown, dusty ochre, and jet black?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the background pitch-black and void-like, illuminated by a single harsh overhead light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the moth's segmented abdomen contract violently while its feathery antennae quiver?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera tremble violently, causing the image to switch rapidly between out of focus and clarity?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.060880061239004135 + ], + "opposite_logprob_tokens": [ + -2.935879945755005 + ], + "logprob_reward": -0.060880061239004135, + "yes_no_margin_reward": 2.875, + "token_credit_reward": -0.060880061239004135, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11210691928863525, + 0.09546389430761337, + 0.08410899341106415, + 0.08313237130641937, + 0.05265289917588234, + 0.10445141047239304, + 0.09145487844944, + 0.20624661445617676, + 0.07946896553039551, + 0.09091304242610931 + ], + "credit_map": [ + 0.012778205797076225, + 0.013520428910851479, + 0.010594042018055916, + 0.018545003607869148, + 0.012585530057549477, + 0.01388569362461567, + 0.013431486673653126, + 0.01676652953028679, + 0.009965804405510426, + 0.01582539640367031, + 0.009254951030015945, + 0.012890676967799664, + 0.009153138846158981, + 0.010781806893646717, + 0.011740697547793388, + 0.015851417556405067, + 0.006434179842472076, + 0.007482927292585373, + 0.006602487061172724, + 0.010228882543742657, + 0.01249733753502369, + 0.018340356647968292, + 0.010438553057610989, + 0.01208427082747221, + 0.007102319039404392, + 0.00884752906858921, + 0.005466280039399862, + 0.011465592309832573, + 0.014516950584948063, + 0.01319018006324768, + 0.009787512011826038, + 0.01275600679218769, + 0.004191330634057522, + 0.004366189241409302, + 0.0038312424439936876, + 0.003842354519292712, + 0.004829497076570988, + 0.009465028531849384, + 0.003801633371040225, + 0.018325624987483025, + 0.007775279693305492, + 0.012260057963430882, + 0.00709177553653717, + 0.009746978990733624, + 0.02519231289625168, + 0.013824179768562317, + 0.013369731605052948, + 0.015191089361906052, + 0.008350308053195477, + 0.007897511124610901, + 0.022872572764754295, + 0.009471295401453972, + 0.008371763862669468, + 0.008916687220335007, + 0.009925260208547115, + 0.01564948447048664, + 0.03565600886940956, + 0.02518751285970211, + 0.036671653389930725, + 0.016378074884414673, + 0.021353483200073242, + 0.02139965631067753, + 0.014762728475034237, + 0.0348375029861927, + 0.013160431757569313, + 0.011481637135148048, + 0.009844955988228321, + 0.007597827818244696, + 0.008457357995212078, + 0.0072525786235928535, + 0.008042133413255215, + 0.013632041402161121, + 0.01147057767957449, + 0.02492484450340271, + 0.023747896775603294, + 0.005689515732228756, + 0.007468812167644501, + 0.0055992938578128815, + 0.006418662611395121, + 0.005593437701463699 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 0, + 5 + ], + "selected_reward": -0.060880061239004135 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the moth's thorax thick, hairy, and patterned with a bone-white skull design?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the moth's wings massive, velvety, and covered in a mosaic of deep chocolate brown, dusty ochre, and jet black?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the background pitch-black and void-like, illuminated by a single harsh overhead light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the moth's segmented abdomen contract violently while its feathery antennae quiver?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera tremble violently, causing the image to switch rapidly between out of focus and clarity?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.023291219025850296 + ], + "opposite_logprob_tokens": [ + -3.7732911109924316 + ], + "logprob_reward": -0.023291219025850296, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.023291219025850296, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07050290703773499, + 0.14801500737667084, + 0.052414603531360626, + 0.058445122092962265, + 0.04623977839946747, + 0.08117669075727463, + 0.04337623715400696, + 0.3493867814540863, + 0.05966596305370331, + 0.09077692776918411 + ], + "credit_map": [ + 0.007400323636829853, + 0.008174427784979343, + 0.007801028434187174, + 0.009840413928031921, + 0.007024485617876053, + 0.012660626322031021, + 0.007500008679926395, + 0.010101590305566788, + 0.01389643270522356, + 0.021756144240498543, + 0.013575233519077301, + 0.018719999119639397, + 0.016862450167536736, + 0.020237693563103676, + 0.02000354789197445, + 0.0229635052382946, + 0.004168000537902117, + 0.005527992267161608, + 0.0038386997766792774, + 0.009414255619049072, + 0.005396249238401651, + 0.01095964852720499, + 0.00446291035041213, + 0.008646847680211067, + 0.004783043172210455, + 0.00648599024862051, + 0.003936965949833393, + 0.009722398594021797, + 0.010229508392512798, + 0.008481456898152828, + 0.007228198926895857, + 0.007577560376375914, + 0.00432282080873847, + 0.004922573454678059, + 0.004746335558593273, + 0.0037944267969578505, + 0.004894502926617861, + 0.008142451755702496, + 0.0027560063172131777, + 0.012660657986998558, + 0.00943394098430872, + 0.010438931174576283, + 0.01116290781646967, + 0.012237690389156342, + 0.009754094295203686, + 0.009253613650798798, + 0.006572592072188854, + 0.012322921305894852, + 0.004506874829530716, + 0.004262117203325033, + 0.009782997891306877, + 0.005068924743682146, + 0.0041239880956709385, + 0.004048849921673536, + 0.003962655086070299, + 0.007619828451424837, + 0.06435414403676987, + 0.03888791427016258, + 0.06439744681119919, + 0.024130340665578842, + 0.039536621421575546, + 0.03758475184440613, + 0.021147120743989944, + 0.0593484491109848, + 0.008364887908101082, + 0.007960014045238495, + 0.007044874131679535, + 0.005349542945623398, + 0.005409633740782738, + 0.006267043296247721, + 0.005703429691493511, + 0.013566538691520691, + 0.010611136443912983, + 0.022478213533759117, + 0.022600144147872925, + 0.006604558322578669, + 0.007516123820096254, + 0.006155036855489016, + 0.006187980528920889, + 0.008623729459941387 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 1, + 9 + ], + "selected_reward": -0.023291219025850296 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the moth's thorax thick, hairy, and patterned with a bone-white skull design?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the moth's wings massive, velvety, and covered in a mosaic of deep chocolate brown, dusty ochre, and jet black?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the background pitch-black and void-like, illuminated by a single harsh overhead light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the moth's segmented abdomen contract violently while its feathery antennae quiver?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera tremble violently, causing the image to switch rapidly between out of focus and clarity?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.4288979470729828 + ], + "opposite_logprob_tokens": [ + -1.0538979768753052 + ], + "logprob_reward": -0.4288979470729828, + "yes_no_margin_reward": 0.625, + "token_credit_reward": -0.4288979470729828, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.04542163014411926, + 0.05618956312537193, + 0.055389463901519775, + 0.08546766638755798, + 0.01946706511080265, + 0.023024918511509895, + 0.03772662207484245, + 0.6089410781860352, + 0.04339353367686272, + 0.024978429079055786 + ], + "credit_map": [ + 0.004490986932069063, + 0.007300110068172216, + 0.004297667182981968, + 0.005959112197160721, + 0.005047124810516834, + 0.007944365963339806, + 0.004792896099388599, + 0.005589371081441641, + 0.005805690307170153, + 0.01132079679518938, + 0.004307604860514402, + 0.007270971313118935, + 0.0053970650769770145, + 0.00721430080011487, + 0.005704713054001331, + 0.009168419055640697, + 0.005554307717829943, + 0.005516760051250458, + 0.003996654413640499, + 0.005811655893921852, + 0.005247734487056732, + 0.015418748371303082, + 0.003751894226297736, + 0.010091705247759819, + 0.008216961286962032, + 0.008536064997315407, + 0.006208694074302912, + 0.017302434891462326, + 0.012399806641042233, + 0.011469719931483269, + 0.00918921735137701, + 0.01214477326720953, + 0.0026103274431079626, + 0.002723714103922248, + 0.0028102740179747343, + 0.002770185237750411, + 0.0020801802165806293, + 0.0029727488290518522, + 0.0014192115049809217, + 0.0020804251544177532, + 0.0031382632441818714, + 0.0027010152116417885, + 0.0034189755097031593, + 0.003738271538168192, + 0.0020352639257907867, + 0.0023108473978936672, + 0.002184585900977254, + 0.0034976983442902565, + 0.004080598242580891, + 0.0034673241898417473, + 0.009633216075599194, + 0.0038838170003145933, + 0.0039977929554879665, + 0.0033095937687903643, + 0.0029842217918485403, + 0.006370058748871088, + 0.11514133214950562, + 0.06906630843877792, + 0.11104359477758408, + 0.04201111942529678, + 0.06574855744838715, + 0.06573106348514557, + 0.036928631365299225, + 0.10327044129371643, + 0.007431132718920708, + 0.005684142932295799, + 0.005985572002828121, + 0.003121260553598404, + 0.0029129900503903627, + 0.004365147091448307, + 0.003703799331560731, + 0.010189488530158997, + 0.0034781123977154493, + 0.004229348618537188, + 0.004316436126828194, + 0.002808579243719578, + 0.002862188033759594, + 0.0020489119924604893, + 0.0024989370722323656, + 0.002735914895310998 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 3, + 1 + ], + "selected_reward": -0.4288979470729828 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the moth's thorax thick, hairy, and patterned with a bone-white skull design?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the moth's wings massive, velvety, and covered in a mosaic of deep chocolate brown, dusty ochre, and jet black?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the background pitch-black and void-like, illuminated by a single harsh overhead light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the moth's segmented abdomen contract violently while its feathery antennae quiver?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera tremble violently, causing the image to switch rapidly between out of focus and clarity?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -0.8983364105224609 + ], + "opposite_logprob_tokens": [ + -0.5233364105224609 + ], + "logprob_reward": -0.8983364105224609, + "yes_no_margin_reward": -0.375, + "token_credit_reward": -0.8983364105224609, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.053575485944747925, + 0.037854962050914764, + 0.04560594633221626, + 0.02770116738975048, + 0.04318338632583618, + 0.031480614095926285, + 0.058093126863241196, + 0.5482480525970459, + 0.09851797670125961, + 0.05573926493525505 + ], + "credit_map": [ + 0.004734898917376995, + 0.006113253068178892, + 0.005948233418166637, + 0.006387355737388134, + 0.004728743340820074, + 0.009818917140364647, + 0.006657402962446213, + 0.009186680428683758, + 0.0032738817390054464, + 0.006235952954739332, + 0.003959037363529205, + 0.004254658706486225, + 0.003827557433396578, + 0.006675563752651215, + 0.003973448183387518, + 0.005654862616211176, + 0.0032760377507656813, + 0.005246189888566732, + 0.003043822944164276, + 0.007651491090655327, + 0.005352483596652746, + 0.00840601697564125, + 0.005910337902605534, + 0.006719567347317934, + 0.0023373186122626066, + 0.0027368890587240458, + 0.002324078930541873, + 0.004427689593285322, + 0.0030247948598116636, + 0.0043201446533203125, + 0.0033846553415060043, + 0.005145595874637365, + 0.0045932200737297535, + 0.004359794314950705, + 0.004632945172488689, + 0.004363786894828081, + 0.0066742366179823875, + 0.00584074342623353, + 0.005356654990464449, + 0.007362003438174725, + 0.003792583243921399, + 0.0037674244958907366, + 0.004974473733454943, + 0.003985797055065632, + 0.003154423553496599, + 0.002953964052721858, + 0.004536248743534088, + 0.004315700847655535, + 0.005808642134070396, + 0.007069048006087542, + 0.01190366130322218, + 0.008156423456966877, + 0.004992120899260044, + 0.005439690314233303, + 0.006783104036003351, + 0.007940439507365227, + 0.1047583520412445, + 0.06175996735692024, + 0.10032385587692261, + 0.03739313408732414, + 0.05815158039331436, + 0.059113845229148865, + 0.033413272351026535, + 0.09333407133817673, + 0.016832221299409866, + 0.012439165264368057, + 0.010944156907498837, + 0.006084604654461145, + 0.007294178940355778, + 0.010407859459519386, + 0.007382159121334553, + 0.027133632451295853, + 0.007569388952106237, + 0.010867752134799957, + 0.013031904585659504, + 0.005258029792457819, + 0.006194768473505974, + 0.004126977175474167, + 0.003909785766154528, + 0.004780658055096865 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 8, + 6 + ], + "selected_reward": -0.8983364105224609 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the moth's thorax thick, hairy, and patterned with a bone-white skull design?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the moth's wings massive, velvety, and covered in a mosaic of deep chocolate brown, dusty ochre, and jet black?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the background pitch-black and void-like, illuminated by a single harsh overhead light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the moth's segmented abdomen contract violently while its feathery antennae quiver?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera tremble violently, causing the image to switch rapidly between out of focus and clarity?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.313479483127594 + ], + "opposite_logprob_tokens": [ + -1.3134794235229492 + ], + "logprob_reward": -0.313479483127594, + "yes_no_margin_reward": 0.9999999403953552, + "token_credit_reward": -0.313479483127594, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.04407034069299698, + 0.20278072357177734, + 0.049654144793748856, + 0.10523980110883713, + 0.13129901885986328, + 0.049626752734184265, + 0.03846481442451477, + 0.19044801592826843, + 0.1276848316192627, + 0.060731545090675354 + ], + "credit_map": [ + 0.004497490357607603, + 0.00518913846462965, + 0.004320833366364241, + 0.006344052962958813, + 0.005125066265463829, + 0.0058199334889650345, + 0.006587109994143248, + 0.006186711601912975, + 0.01538052223622799, + 0.0309571772813797, + 0.01614009588956833, + 0.018832644447684288, + 0.024127794429659843, + 0.03718781843781471, + 0.03164413943886757, + 0.02851054258644581, + 0.0037119549233466387, + 0.006095555145293474, + 0.004002423491328955, + 0.00498697767034173, + 0.005048198625445366, + 0.01162667479366064, + 0.005372926592826843, + 0.008809433318674564, + 0.007866721600294113, + 0.012556444853544235, + 0.008170240558683872, + 0.014560000039637089, + 0.020487239584326744, + 0.0158999003469944, + 0.013193761929869652, + 0.012505481019616127, + 0.011202567256987095, + 0.012219471856951714, + 0.011972111649811268, + 0.012126674875617027, + 0.0131437499076128, + 0.02454308792948723, + 0.006819380447268486, + 0.039271991699934006, + 0.006950256414711475, + 0.006741760298609734, + 0.009194421581923962, + 0.004755443427711725, + 0.004652880597859621, + 0.00556820398196578, + 0.0038166658487170935, + 0.007947123609483242, + 0.0040803756564855576, + 0.0035021791700273752, + 0.008959794417023659, + 0.004878529347479343, + 0.0031266347505152225, + 0.0028336532413959503, + 0.0030995537526905537, + 0.007984094321727753, + 0.03740439936518669, + 0.020931728184223175, + 0.034251660108566284, + 0.013328335247933865, + 0.019011877477169037, + 0.020633945241570473, + 0.012629739008843899, + 0.032256316393613815, + 0.02151591144502163, + 0.01570887491106987, + 0.013366416096687317, + 0.008212345652282238, + 0.008026020601391792, + 0.014327019453048706, + 0.009993590414524078, + 0.03653464466333389, + 0.012055151164531708, + 0.008374316617846489, + 0.010568303987383842, + 0.006931575946509838, + 0.005061289761215448, + 0.005880849435925484, + 0.005854771006852388, + 0.00600528484210372 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 7, + 4 + ], + "selected_reward": -0.313479483127594 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "No", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000101/group_12/index_tensor([1054, 1054])_group_12_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000101/group_12/index_tensor([1054, 1054])_group_12_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..e3a3dcbcecdb573b52d12ea1f07dc3e3cb5bc72c --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000101/group_12/index_tensor([1054, 1054])_group_12_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "In charming Disney-Pixar style, a locked, static, eye-level close-up captures a tender underwater moment, focusing on the heartwarming interaction between two creatures. A tiny crab, with a glossy, warm orange shell and glowing black eyes filled with innocent wonder, carefully nestles into its new home: a vibrant sea anemone. Anchored to a patch of sun-dappled coral, the anemone is vibrantly colored, with a deep purplish-red base from which sprouts a dense cluster of soft, fleshy tentacles. Gently swaying in the calm current, these tentacles are translucent, their interiors glowing with gentle bioluminescence, their tips emitting a brilliant azure blue. With slow, careful movements, the crab's slender legs gently move, tucking its tiny body into the anemone's embrace; in response, the tentacles gently part as if in welcome, then gently curl up, protectively surrounding its shell. Sunlight filtering through the crystal-clear turquoise water casts shimmering, corroded patterns on the sandy seafloor and illuminates the scene with a warm, ethereal glow, creating a tranquil and magical atmosphere of finding a safe, comforting home.", + "seed": 32581, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000101/group_12/index_1054_seed_32581_dp_24_rank_48_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 12, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.034848814085125926, + "vlm_logprob": -0.034848814085125926, + "vlm_yes_no_margin": 3.825, + "vlm_token_credit": -0.034848814085125926, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video shot as a locked, static, eye-level close-up?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a tiny crab with a glossy warm orange shell and glowing black eyes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the crab nesting into a vibrant sea anemone anchored to sun-dappled coral?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the anemone's tentacles sway gently, part to welcome the crab, and curl up to protect it?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by sunlight filtering through turquoise water, casting shimmering patterns on the sandy seafloor?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02473457343876362 + ], + "opposite_logprob_tokens": [ + -4.1497344970703125 + ], + "logprob_reward": -0.02473457343876362, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.02473457343876362, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.17216800153255463, + 0.047788314521312714, + 0.05590508133172989, + 0.047631122171878815, + 0.05288372561335564, + 0.05127037316560745, + 0.07131963223218918, + 0.19752365350723267, + 0.06883876025676727, + 0.23467132449150085 + ], + "credit_map": [ + 0.02090730518102646, + 0.02225593850016594, + 0.028762521222233772, + 0.0235605351626873, + 0.019936271011829376, + 0.02043936401605606, + 0.02020622231066227, + 0.016099853441119194, + 0.004764336161315441, + 0.007199829909950495, + 0.0069721899926662445, + 0.005662902723997831, + 0.006092499941587448, + 0.006932004354894161, + 0.005548377521336079, + 0.004616170655936003, + 0.00600242055952549, + 0.010561115108430386, + 0.004864101763814688, + 0.006304508075118065, + 0.008374074473977089, + 0.006210733205080032, + 0.00626943027600646, + 0.007318696938455105, + 0.0055802795104682446, + 0.007333679124712944, + 0.006235713139176369, + 0.006927042733877897, + 0.005931775085628033, + 0.004949776455760002, + 0.00617205211892724, + 0.004500803537666798, + 0.009027070365846157, + 0.0036027596797794104, + 0.015564857982099056, + 0.005841768346726894, + 0.0034185955300927162, + 0.006643441040068865, + 0.004692436195909977, + 0.004092796705663204, + 0.006129522807896137, + 0.0061202989891171455, + 0.010188557207584381, + 0.005905429366976023, + 0.007460433524101973, + 0.006235914304852486, + 0.0045503731817007065, + 0.004679843783378601, + 0.009908610954880714, + 0.0055665625259280205, + 0.013939399272203445, + 0.008364219218492508, + 0.00577833317220211, + 0.00984342023730278, + 0.008600917644798756, + 0.009318172931671143, + 0.027793556451797485, + 0.02681499719619751, + 0.0228702574968338, + 0.023205500096082687, + 0.0251651331782341, + 0.02998981811106205, + 0.017843425273895264, + 0.02384096384048462, + 0.007466443814337254, + 0.006882295943796635, + 0.012394670397043228, + 0.0069177658297121525, + 0.009318712167441845, + 0.008415549993515015, + 0.009487790055572987, + 0.00795553345233202, + 0.026553550735116005, + 0.028342347592115402, + 0.04104110598564148, + 0.024311784654855728, + 0.029030857607722282, + 0.02725226990878582, + 0.023505456745624542, + 0.03463396057486534 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 7, + 0 + ], + "selected_reward": -0.02473457343876362 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video shot as a locked, static, eye-level close-up?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a tiny crab with a glossy warm orange shell and glowing black eyes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the crab nesting into a vibrant sea anemone anchored to sun-dappled coral?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the anemone's tentacles sway gently, part to welcome the crab, and curl up to protect it?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by sunlight filtering through turquoise water, casting shimmering patterns on the sandy seafloor?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.007637940347194672 + ], + "opposite_logprob_tokens": [ + -4.882637977600098 + ], + "logprob_reward": -0.007637940347194672, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.007637940347194672, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12562702596187592, + 0.05463219806551933, + 0.023834669962525368, + 0.06059996038675308, + 0.04247628524899483, + 0.1626814901828766, + 0.05221378430724144, + 0.0837630107998848, + 0.05369133874773979, + 0.3404802083969116 + ], + "credit_map": [ + 0.012767977081239223, + 0.01709219254553318, + 0.020433764904737473, + 0.01259137224406004, + 0.01739325560629368, + 0.017294129356741905, + 0.01529015228152275, + 0.01276418473571539, + 0.006511774845421314, + 0.008299808017909527, + 0.011324694380164146, + 0.006780992727726698, + 0.005870593246072531, + 0.006126981228590012, + 0.005332442000508308, + 0.004384910687804222, + 0.002217075088992715, + 0.003177859354764223, + 0.004392792005091906, + 0.0023919513914734125, + 0.0029854874592274427, + 0.0030059197451919317, + 0.0027628883253782988, + 0.0029006965924054384, + 0.013367465697228909, + 0.004503306467086077, + 0.017818337306380272, + 0.0042456393130123615, + 0.004237092565745115, + 0.006763495970517397, + 0.0049781156703829765, + 0.004686509259045124, + 0.005883084610104561, + 0.007335857022553682, + 0.007692879997193813, + 0.004476924426853657, + 0.004829034674912691, + 0.005144150462001562, + 0.0037139710038900375, + 0.0034003828186541796, + 0.04495605081319809, + 0.005707058124244213, + 0.06810121238231659, + 0.008851677179336548, + 0.00508889788761735, + 0.019610555842518806, + 0.005682408809661865, + 0.004683628678321838, + 0.005189258139580488, + 0.00773957883939147, + 0.007871217094361782, + 0.0061025964096188545, + 0.007696566637605429, + 0.007298115640878677, + 0.00525264348834753, + 0.005063808988779783, + 0.010161367245018482, + 0.00937123317271471, + 0.014207578264176846, + 0.010400206781923771, + 0.010329395532608032, + 0.010080083273351192, + 0.009479141794145107, + 0.009734007529914379, + 0.007639843970537186, + 0.005875121336430311, + 0.008495197631418705, + 0.004610852804034948, + 0.005224672146141529, + 0.0069038476794958115, + 0.007248103618621826, + 0.007693699095398188, + 0.040366385132074356, + 0.0486794076859951, + 0.06053686514496803, + 0.038605574518442154, + 0.044161420315504074, + 0.039953991770744324, + 0.027277039363980293, + 0.040899522602558136 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 5, + 0 + ], + "selected_reward": -0.007637940347194672 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video shot as a locked, static, eye-level close-up?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a tiny crab with a glossy warm orange shell and glowing black eyes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the crab nesting into a vibrant sea anemone anchored to sun-dappled coral?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the anemone's tentacles sway gently, part to welcome the crab, and curl up to protect it?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by sunlight filtering through turquoise water, casting shimmering patterns on the sandy seafloor?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.08903084695339203 + ], + "opposite_logprob_tokens": [ + -2.4640307426452637 + ], + "logprob_reward": -0.08903084695339203, + "yes_no_margin_reward": 2.375, + "token_credit_reward": -0.08903084695339203, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.19494342803955078, + 0.056468233466148376, + 0.026522506028413773, + 0.11529726535081863, + 0.12362456321716309, + 0.14072701334953308, + 0.04569099098443985, + 0.09456522762775421, + 0.024271270260214806, + 0.17788954079151154 + ], + "credit_map": [ + 0.031690407544374466, + 0.018413683399558067, + 0.04208788275718689, + 0.016459478065371513, + 0.021039217710494995, + 0.024267930537462234, + 0.01907426491379738, + 0.021910542622208595, + 0.012205558829009533, + 0.003869263920933008, + 0.01927581988275051, + 0.004481419920921326, + 0.002727432642132044, + 0.0047848462127149105, + 0.004564513918012381, + 0.004559372551739216, + 0.0025607773568481207, + 0.005072788801044226, + 0.0035418968182057142, + 0.0023291059769690037, + 0.003503810614347458, + 0.0034064447972923517, + 0.0025541335344314575, + 0.0035535464994609356, + 0.03050106205046177, + 0.007509313058108091, + 0.04199691489338875, + 0.00806538388133049, + 0.00506481621414423, + 0.012383311986923218, + 0.004759100265800953, + 0.005017347168177366, + 0.031931474804878235, + 0.005640183575451374, + 0.04978692904114723, + 0.008060711435973644, + 0.004945908673107624, + 0.014060309156775475, + 0.004874326754361391, + 0.004324712324887514, + 0.038882382214069366, + 0.004902279004454613, + 0.060069773346185684, + 0.007464101072400808, + 0.003611618187278509, + 0.017950894311070442, + 0.0041814944706857204, + 0.0036644816864281893, + 0.0058034248650074005, + 0.004906836431473494, + 0.008985691703855991, + 0.00574192451313138, + 0.005153812933713198, + 0.005824262276291847, + 0.00509659992530942, + 0.004178435541689396, + 0.013621297664940357, + 0.012351092882454395, + 0.011711960658431053, + 0.0118141183629632, + 0.014217803254723549, + 0.01235323864966631, + 0.008581419475376606, + 0.009914291091263294, + 0.002747263992205262, + 0.003462248481810093, + 0.0024748959112912416, + 0.002364277606830001, + 0.003885361598804593, + 0.0037188855931162834, + 0.003209307324141264, + 0.002409029984846711, + 0.0206521674990654, + 0.024112332612276077, + 0.035250261425971985, + 0.019197026267647743, + 0.02417316846549511, + 0.02080339938402176, + 0.015508642420172691, + 0.018192531540989876 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 5 + ], + "selected_reward": -0.08903084695339203 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video shot as a locked, static, eye-level close-up?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a tiny crab with a glossy warm orange shell and glowing black eyes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the crab nesting into a vibrant sea anemone anchored to sun-dappled coral?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the anemone's tentacles sway gently, part to welcome the crab, and curl up to protect it?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by sunlight filtering through turquoise water, casting shimmering patterns on the sandy seafloor?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0097922682762146 + ], + "opposite_logprob_tokens": [ + -4.634792327880859 + ], + "logprob_reward": -0.0097922682762146, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.0097922682762146, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12797750532627106, + 0.07266975194215775, + 0.046466778963804245, + 0.10201384872198105, + 0.24886848032474518, + 0.16233579814434052, + 0.044456154108047485, + 0.06539301574230194, + 0.03423038870096207, + 0.09558821469545364 + ], + "credit_map": [ + 0.02285931073129177, + 0.01910955272614956, + 0.023378461599349976, + 0.009561622515320778, + 0.01666698046028614, + 0.014939903281629086, + 0.00997535977512598, + 0.011486325412988663, + 0.017060628160834312, + 0.006876386236399412, + 0.02529088966548443, + 0.004630255978554487, + 0.004763761535286903, + 0.006532474886626005, + 0.0038480034563690424, + 0.0036673604045063257, + 0.0044046007096767426, + 0.006759690586477518, + 0.00795693602412939, + 0.0046288748271763325, + 0.005923562217503786, + 0.0047165523283183575, + 0.006541871000081301, + 0.005534692667424679, + 0.014093518257141113, + 0.015439721755683422, + 0.022090129554271698, + 0.008087147027254105, + 0.012895525433123112, + 0.011655100621283054, + 0.009624470956623554, + 0.008128238841891289, + 0.058010343462228775, + 0.010560990311205387, + 0.09228555113077164, + 0.01864706166088581, + 0.012672552838921547, + 0.029721373692154884, + 0.015962393954396248, + 0.011008252389729023, + 0.042712993919849396, + 0.00604321900755167, + 0.06558536738157272, + 0.00929876510053873, + 0.007290728855878115, + 0.01938580349087715, + 0.006891279947012663, + 0.005127655807882547, + 0.005593711510300636, + 0.004639086779206991, + 0.008026191033422947, + 0.005993439815938473, + 0.004148500505834818, + 0.006786879617720842, + 0.004221509676426649, + 0.005046839825809002, + 0.008060934022068977, + 0.007318378426134586, + 0.010135279037058353, + 0.010116906836628914, + 0.0073078954592347145, + 0.008288489654660225, + 0.007937682792544365, + 0.006227452773600817, + 0.003674950683489442, + 0.005102583672851324, + 0.004799230024218559, + 0.0033251261338591576, + 0.004920747131109238, + 0.005189899355173111, + 0.0040970779955387115, + 0.0031207764986902475, + 0.010986984707415104, + 0.011227311566472054, + 0.011876202188432217, + 0.011769403703510761, + 0.011101325042545795, + 0.01519234199076891, + 0.010032118298113346, + 0.013402529060840607 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 5, + 0 + ], + "selected_reward": -0.0097922682762146 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video shot as a locked, static, eye-level close-up?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a tiny crab with a glossy warm orange shell and glowing black eyes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the crab nesting into a vibrant sea anemone anchored to sun-dappled coral?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the anemone's tentacles sway gently, part to welcome the crab, and curl up to protect it?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by sunlight filtering through turquoise water, casting shimmering patterns on the sandy seafloor?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0430484414100647 + ], + "opposite_logprob_tokens": [ + -3.16804838180542 + ], + "logprob_reward": -0.0430484414100647, + "yes_no_margin_reward": 3.125, + "token_credit_reward": -0.0430484414100647, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.24218255281448364, + 0.08363260328769684, + 0.03458536043763161, + 0.13597306609153748, + 0.04195297136902809, + 0.16999953985214233, + 0.04987242817878723, + 0.07386382669210434, + 0.041988175362348557, + 0.12594950199127197 + ], + "credit_map": [ + 0.027418917044997215, + 0.028500908985733986, + 0.03363294154405594, + 0.02622394450008869, + 0.03428731858730316, + 0.03260998800396919, + 0.030616002157330513, + 0.0288925189524889, + 0.019341280683875084, + 0.004851461388170719, + 0.03042406588792801, + 0.005718836095184088, + 0.004020758904516697, + 0.007728236727416515, + 0.005834716372191906, + 0.00571323512122035, + 0.0040877144783735275, + 0.003822351573035121, + 0.005838071461766958, + 0.003325258381664753, + 0.003241672646254301, + 0.00508370716124773, + 0.004340003710240126, + 0.004846576135605574, + 0.03596040606498718, + 0.007350338622927666, + 0.0478028729557991, + 0.008822095580399036, + 0.006187975872308016, + 0.01487115491181612, + 0.006961605977267027, + 0.008016618900001049, + 0.0040161097422242165, + 0.005877348128706217, + 0.0044300551526248455, + 0.006174213252961636, + 0.004460900090634823, + 0.005702280905097723, + 0.0062178876250982285, + 0.005074177403002977, + 0.046953488141298294, + 0.006275933235883713, + 0.07162577658891678, + 0.008977350778877735, + 0.005119089502841234, + 0.02126036398112774, + 0.005284424871206284, + 0.004503100644797087, + 0.006382099352777004, + 0.0053960103541612625, + 0.011176453903317451, + 0.0049217743799090385, + 0.006408157758414745, + 0.005958341062068939, + 0.005187081638723612, + 0.0044425055384635925, + 0.011050521396100521, + 0.007847470231354237, + 0.011026773601770401, + 0.01030938234180212, + 0.011393911205232143, + 0.007948728278279305, + 0.007722648326307535, + 0.006564384326338768, + 0.0054089780896902084, + 0.006282664369791746, + 0.004676213953644037, + 0.003568967804312706, + 0.005691999103873968, + 0.0063607473857700825, + 0.005185930989682674, + 0.00481267087161541, + 0.015377212315797806, + 0.01382360327988863, + 0.026432091370224953, + 0.013353541493415833, + 0.014631488360464573, + 0.01680734194815159, + 0.012505553662776947, + 0.013018665835261345 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 5, + 3 + ], + "selected_reward": -0.0430484414100647 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "In charming Disney-Pixar style, a locked, static, eye-level close-up captures a tender underwater moment, focusing on the heartwarming interaction between two creatures. A tiny crab, with a glossy, warm orange shell and glowing black eyes filled with innocent wonder, carefully nestles into its new home: a vibrant sea anemone. Anchored to a patch of sun-dappled coral, the anemone is vibrantly colored, with a deep purplish-red base from which sprouts a dense cluster of soft, fleshy tentacles. Gently swaying in the calm current, these tentacles are translucent, their interiors glowing with gentle bioluminescence, their tips emitting a brilliant azure blue. With slow, careful movements, the crab's slender legs gently move, tucking its tiny body into the anemone's embrace; in response, the tentacles gently part as if in welcome, then gently curl up, protectively surrounding its shell. Sunlight filtering through the crystal-clear turquoise water casts shimmering, corroded patterns on the sandy seafloor and illuminates the scene with a warm, ethereal glow, creating a tranquil and magical atmosphere of finding a safe, comforting home.", + "seed": 32582, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000101/group_12/index_1054_seed_32582_dp_24_rank_48_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 12, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.04988127164542675, + "vlm_logprob": -0.04988127164542675, + "vlm_yes_no_margin": 3.475, + "vlm_token_credit": -0.04988127164542675, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video shot as a locked, static, eye-level close-up?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a tiny crab with a glossy warm orange shell and glowing black eyes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the crab nesting into a vibrant sea anemone anchored to sun-dappled coral?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the anemone's tentacles sway gently, part to welcome the crab, and curl up to protect it?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by sunlight filtering through turquoise water, casting shimmering patterns on the sandy seafloor?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03433021157979965 + ], + "opposite_logprob_tokens": [ + -3.784330129623413 + ], + "logprob_reward": -0.03433021157979965, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.03433021157979965, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09495294094085693, + 0.06333598494529724, + 0.03094240091741085, + 0.09971681982278824, + 0.10280541330575943, + 0.20942750573158264, + 0.10701268166303635, + 0.1675206869840622, + 0.04808107763528824, + 0.07620452344417572 + ], + "credit_map": [ + 0.012495080009102821, + 0.011388876475393772, + 0.014374618418514729, + 0.007951141335070133, + 0.023849530145525932, + 0.008377998135983944, + 0.007179074455052614, + 0.009336622431874275, + 0.006528240162879229, + 0.009082052856683731, + 0.00829031877219677, + 0.00807117111980915, + 0.013272633776068687, + 0.005326943472027779, + 0.006065216846764088, + 0.006699403282254934, + 0.005133816972374916, + 0.003377254121005535, + 0.004651047755032778, + 0.003388029523193836, + 0.0037940973415970802, + 0.004760094452649355, + 0.0025647825095802546, + 0.0032732777763158083, + 0.011194214224815369, + 0.004355350509285927, + 0.02495717816054821, + 0.004778893198817968, + 0.038898468017578125, + 0.006998648401349783, + 0.0038409989792853594, + 0.0046930694952607155, + 0.011404929682612419, + 0.02075899764895439, + 0.015032649040222168, + 0.012356617487967014, + 0.016356663778424263, + 0.009864428080618382, + 0.009172256104648113, + 0.007858872413635254, + 0.02592111937701702, + 0.018418196588754654, + 0.059566132724285126, + 0.021308818832039833, + 0.02299286611378193, + 0.02210375666618347, + 0.023776644840836525, + 0.015339966863393784, + 0.010132679715752602, + 0.01678168959915638, + 0.009462139569222927, + 0.013653415255248547, + 0.014453317038714886, + 0.01125496905297041, + 0.01359124481678009, + 0.017683228477835655, + 0.017864009365439415, + 0.012345541268587112, + 0.019987473264336586, + 0.008748798631131649, + 0.06944723427295685, + 0.013763695023953915, + 0.009625503793358803, + 0.015738435089588165, + 0.004459534306079149, + 0.008250640705227852, + 0.004113426897674799, + 0.007115479558706284, + 0.00541361328214407, + 0.005636928137391806, + 0.006017695181071758, + 0.007073758635669947, + 0.005598579067736864, + 0.015589023008942604, + 0.007970405742526054, + 0.009809820912778378, + 0.009325283579528332, + 0.00693195266649127, + 0.00878677237778902, + 0.012192689813673496 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 7, + 6 + ], + "selected_reward": -0.03433021157979965 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video shot as a locked, static, eye-level close-up?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a tiny crab with a glossy warm orange shell and glowing black eyes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the crab nesting into a vibrant sea anemone anchored to sun-dappled coral?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the anemone's tentacles sway gently, part to welcome the crab, and curl up to protect it?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by sunlight filtering through turquoise water, casting shimmering patterns on the sandy seafloor?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.020588496699929237 + ], + "opposite_logprob_tokens": [ + -3.8955883979797363 + ], + "logprob_reward": -0.020588496699929237, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.020588496699929237, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1530771255493164, + 0.08446389436721802, + 0.0505860298871994, + 0.12376740574836731, + 0.12204959243535995, + 0.06999875605106354, + 0.10927306115627289, + 0.14016194641590118, + 0.08143895864486694, + 0.06518325954675674 + ], + "credit_map": [ + 0.015928810462355614, + 0.01662776805460453, + 0.028662731871008873, + 0.01205897331237793, + 0.04078251123428345, + 0.011538139544427395, + 0.013381797820329666, + 0.014096383936703205, + 0.010639175772666931, + 0.010072845034301281, + 0.012463605031371117, + 0.010871988721191883, + 0.015383942052721977, + 0.008408556692302227, + 0.009213930927217007, + 0.00740984408184886, + 0.005672241095453501, + 0.009364427998661995, + 0.008273886516690254, + 0.004610465839505196, + 0.007029444444924593, + 0.005195993930101395, + 0.004409900866448879, + 0.006029672920703888, + 0.01718227192759514, + 0.008618032559752464, + 0.028770271688699722, + 0.0043900893069803715, + 0.03982570767402649, + 0.013345199637115002, + 0.005322875455021858, + 0.006312956567853689, + 0.017310861498117447, + 0.011254456825554371, + 0.022480113431811333, + 0.011546047404408455, + 0.017070308327674866, + 0.014409796334803104, + 0.009605358354747295, + 0.01837265118956566, + 0.008025366812944412, + 0.010775748640298843, + 0.009441052563488483, + 0.0068513378500938416, + 0.007391930557787418, + 0.006999006494879723, + 0.012245495803654194, + 0.008268817327916622, + 0.010565885342657566, + 0.01601351425051689, + 0.012096784077584743, + 0.016270386055111885, + 0.016899509355425835, + 0.017281917855143547, + 0.011343397200107574, + 0.008801661431789398, + 0.014325941912829876, + 0.011786047369241714, + 0.021765096113085747, + 0.00755164073780179, + 0.055346421897411346, + 0.01252411212772131, + 0.009317435324192047, + 0.007545256055891514, + 0.007746736519038677, + 0.01708865351974964, + 0.007406636141240597, + 0.013663717545568943, + 0.008997127413749695, + 0.008726369589567184, + 0.008726808242499828, + 0.009082911536097527, + 0.005661775358021259, + 0.013357726857066154, + 0.007906454615294933, + 0.008491197600960732, + 0.007603827398270369, + 0.004516042303293943, + 0.008077023550868034, + 0.00956921000033617 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 7, + 3 + ], + "selected_reward": -0.020588496699929237 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video shot as a locked, static, eye-level close-up?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a tiny crab with a glossy warm orange shell and glowing black eyes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the crab nesting into a vibrant sea anemone anchored to sun-dappled coral?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the anemone's tentacles sway gently, part to welcome the crab, and curl up to protect it?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by sunlight filtering through turquoise water, casting shimmering patterns on the sandy seafloor?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.14279083907604218 + ], + "opposite_logprob_tokens": [ + -2.0177907943725586 + ], + "logprob_reward": -0.14279083907604218, + "yes_no_margin_reward": 1.875, + "token_credit_reward": -0.14279083907604218, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2589995861053467, + 0.07283443957567215, + 0.03177174925804138, + 0.06460752338171005, + 0.18392327427864075, + 0.07348130643367767, + 0.09993916004896164, + 0.12431585043668747, + 0.03611975908279419, + 0.05400730296969414 + ], + "credit_map": [ + 0.02776951529085636, + 0.02413412369787693, + 0.02810731902718544, + 0.02131473645567894, + 0.07109633833169937, + 0.03415718674659729, + 0.021978840231895447, + 0.0304415263235569, + 0.011261838488280773, + 0.009124363772571087, + 0.00827101618051529, + 0.008568466641008854, + 0.010647756047546864, + 0.011928260326385498, + 0.006781969219446182, + 0.006250772625207901, + 0.006069152615964413, + 0.004470479674637318, + 0.00514086103066802, + 0.002101236954331398, + 0.003995003644376993, + 0.004479362163692713, + 0.0024352893233299255, + 0.0030803640838712454, + 0.007449342403560877, + 0.003905930556356907, + 0.012683471664786339, + 0.005013261456042528, + 0.022698966786265373, + 0.005186363123357296, + 0.003588506719097495, + 0.004081676248461008, + 0.028494039550423622, + 0.032509464770555496, + 0.019440725445747375, + 0.013359619304537773, + 0.028280489146709442, + 0.034053362905979156, + 0.013292280957102776, + 0.014493301510810852, + 0.00685214065015316, + 0.00693132821470499, + 0.008165312930941582, + 0.009220991283655167, + 0.00812055915594101, + 0.005692490842193365, + 0.018483709543943405, + 0.01001477986574173, + 0.009837601333856583, + 0.009760210290551186, + 0.014115936122834682, + 0.009479203261435032, + 0.01362831611186266, + 0.017964286729693413, + 0.014901939779520035, + 0.010251673869788647, + 0.011484706774353981, + 0.00890388060361147, + 0.016446953639388084, + 0.00743610505014658, + 0.039947573095560074, + 0.011597693897783756, + 0.016254493966698647, + 0.012244459241628647, + 0.0029425742104649544, + 0.0027638725005090237, + 0.003866150975227356, + 0.003910012077540159, + 0.005192220211029053, + 0.005506676621735096, + 0.006383371539413929, + 0.00555488420650363, + 0.004809185862541199, + 0.004210151731967926, + 0.005735299549996853, + 0.004893919453024864, + 0.007119274232536554, + 0.007381614297628403, + 0.008301842026412487, + 0.011556020006537437 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 7 + ], + "selected_reward": -0.14279083907604218 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video shot as a locked, static, eye-level close-up?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a tiny crab with a glossy warm orange shell and glowing black eyes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the crab nesting into a vibrant sea anemone anchored to sun-dappled coral?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the anemone's tentacles sway gently, part to welcome the crab, and curl up to protect it?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by sunlight filtering through turquoise water, casting shimmering patterns on the sandy seafloor?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.008646313101053238 + ], + "opposite_logprob_tokens": [ + -4.758646488189697 + ], + "logprob_reward": -0.008646313101053238, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.008646313101053238, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1872761845588684, + 0.07472218573093414, + 0.05967002734541893, + 0.09338807314634323, + 0.10714409500360489, + 0.12402693182229996, + 0.08533007651567459, + 0.16697022318840027, + 0.037969380617141724, + 0.06350288540124893 + ], + "credit_map": [ + 0.0208649393171072, + 0.022596614435315132, + 0.03087858483195305, + 0.012384110130369663, + 0.050299279391765594, + 0.0186478141695261, + 0.016226407140493393, + 0.01537841372191906, + 0.00870164018124342, + 0.011394315399229527, + 0.010103435255587101, + 0.006720732431858778, + 0.014507995918393135, + 0.0071029821410775185, + 0.0077271354384720325, + 0.008463937789201736, + 0.010690751485526562, + 0.0064861224964261055, + 0.00878105591982603, + 0.005890949163585901, + 0.007016776595264673, + 0.008027281612157822, + 0.005798422731459141, + 0.0069786617532372475, + 0.009580825455486774, + 0.004679932724684477, + 0.023013176396489143, + 0.00587020767852664, + 0.036206599324941635, + 0.005642645061016083, + 0.0044580865651369095, + 0.003936584573239088, + 0.012219281867146492, + 0.018403051421046257, + 0.016432316973805428, + 0.013820004649460316, + 0.017268743366003036, + 0.01244307216256857, + 0.009370081126689911, + 0.0071875350549817085, + 0.015016489662230015, + 0.016354257240891457, + 0.024916743859648705, + 0.011505049653351307, + 0.013252438977360725, + 0.011459789238870144, + 0.018184537068009377, + 0.013337609358131886, + 0.008602461777627468, + 0.011738565750420094, + 0.007484160829335451, + 0.010410221293568611, + 0.01123372744768858, + 0.01485471986234188, + 0.010520420968532562, + 0.010485795326530933, + 0.017838314175605774, + 0.010552775114774704, + 0.029557980597019196, + 0.007450134959071875, + 0.06843870133161545, + 0.014496936462819576, + 0.010621413588523865, + 0.008013957180082798, + 0.002889521885663271, + 0.004288483876734972, + 0.003950746264308691, + 0.005993187893182039, + 0.00454200292006135, + 0.004470734391361475, + 0.006003545597195625, + 0.005831154063344002, + 0.005183266010135412, + 0.011176234111189842, + 0.005518683232367039, + 0.009312527254223824, + 0.007090611383318901, + 0.006412374787032604, + 0.010024976916611195, + 0.008784200064837933 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 7, + 5 + ], + "selected_reward": -0.008646313101053238 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video shot as a locked, static, eye-level close-up?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a tiny crab with a glossy warm orange shell and glowing black eyes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the crab nesting into a vibrant sea anemone anchored to sun-dappled coral?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the anemone's tentacles sway gently, part to welcome the crab, and curl up to protect it?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by sunlight filtering through turquoise water, casting shimmering patterns on the sandy seafloor?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04305049777030945 + ], + "opposite_logprob_tokens": [ + -3.168050527572632 + ], + "logprob_reward": -0.04305049777030945, + "yes_no_margin_reward": 3.125, + "token_credit_reward": -0.04305049777030945, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.27330565452575684, + 0.06256133317947388, + 0.03387165814638138, + 0.03152339160442352, + 0.056218672543764114, + 0.08933228999376297, + 0.11518249660730362, + 0.2816832959651947, + 0.023506462574005127, + 0.03281474858522415 + ], + "credit_map": [ + 0.030361978337168694, + 0.023229099810123444, + 0.04186789691448212, + 0.014621862210333347, + 0.10255439579486847, + 0.029375694692134857, + 0.01700722984969616, + 0.014287487603724003, + 0.010370872914791107, + 0.010570370592176914, + 0.006898706313222647, + 0.005805410910397768, + 0.008929599076509476, + 0.008738163858652115, + 0.005336208269000053, + 0.005912001244723797, + 0.006794643122702837, + 0.00427375640720129, + 0.004061950370669365, + 0.0023370280396193266, + 0.0055453055538237095, + 0.004938074387609959, + 0.0026985404547303915, + 0.003222361207008362, + 0.003815834643319249, + 0.0025002446491271257, + 0.004888087976723909, + 0.0019108654232695699, + 0.009080308489501476, + 0.0032916120253503323, + 0.002430627355352044, + 0.003605811158195138, + 0.008557800203561783, + 0.0064548649825155735, + 0.00885863322764635, + 0.005418650805950165, + 0.006893782410770655, + 0.0054123313166201115, + 0.006275685038417578, + 0.008346925489604473, + 0.013255797326564789, + 0.006219504866749048, + 0.026929164305329323, + 0.005824827589094639, + 0.009481832385063171, + 0.011064360849559307, + 0.00928175263106823, + 0.00727505050599575, + 0.012122534215450287, + 0.009314369410276413, + 0.012874836102128029, + 0.008872194215655327, + 0.024424314498901367, + 0.017730336636304855, + 0.01261151023209095, + 0.01723240315914154, + 0.028176449239253998, + 0.018807094544172287, + 0.03175507113337517, + 0.01570054516196251, + 0.10549069941043854, + 0.024851834401488304, + 0.028810976073145866, + 0.028090616688132286, + 0.0019144683610647917, + 0.0027024387381970882, + 0.002449361141771078, + 0.0028586741536855698, + 0.0032413199078291655, + 0.0029671904630959034, + 0.0040863268077373505, + 0.0032866820693016052, + 0.0029840131755918264, + 0.003698016284033656, + 0.003602467244490981, + 0.003938139416277409, + 0.004243774339556694, + 0.0036050276830792427, + 0.00442226929590106, + 0.006321038585156202 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 0, + 6 + ], + "selected_reward": -0.04305049777030945 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "In charming Disney-Pixar style, a locked, static, eye-level close-up captures a tender underwater moment, focusing on the heartwarming interaction between two creatures. A tiny crab, with a glossy, warm orange shell and glowing black eyes filled with innocent wonder, carefully nestles into its new home: a vibrant sea anemone. Anchored to a patch of sun-dappled coral, the anemone is vibrantly colored, with a deep purplish-red base from which sprouts a dense cluster of soft, fleshy tentacles. Gently swaying in the calm current, these tentacles are translucent, their interiors glowing with gentle bioluminescence, their tips emitting a brilliant azure blue. With slow, careful movements, the crab's slender legs gently move, tucking its tiny body into the anemone's embrace; in response, the tentacles gently part as if in welcome, then gently curl up, protectively surrounding its shell. Sunlight filtering through the crystal-clear turquoise water casts shimmering, corroded patterns on the sandy seafloor and illuminates the scene with a warm, ethereal glow, creating a tranquil and magical atmosphere of finding a safe, comforting home.", + "seed": 32583, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000101/group_12/index_1054_seed_32583_dp_25_rank_50_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 12, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.05882075112313032, + "vlm_logprob": -0.05882075112313032, + "vlm_yes_no_margin": 3.25, + "vlm_token_credit": -0.05882075112313032, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video shot as a locked, static, eye-level close-up?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a tiny crab with a glossy warm orange shell and glowing black eyes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the crab nesting into a vibrant sea anemone anchored to sun-dappled coral?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the anemone's tentacles sway gently, part to welcome the crab, and curl up to protect it?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by sunlight filtering through turquoise water, casting shimmering patterns on the sandy seafloor?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.035820648074150085 + ], + "opposite_logprob_tokens": [ + -3.785820722579956 + ], + "logprob_reward": -0.035820648074150085, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.035820648074150085, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.146302729845047, + 0.06184260547161102, + 0.0917532667517662, + 0.36457502841949463, + 0.08954901248216629, + 0.03249402344226837, + 0.054466575384140015, + 0.06774037331342697, + 0.027783237397670746, + 0.06349309533834457 + ], + "credit_map": [ + 0.019617393612861633, + 0.020614083856344223, + 0.01945754885673523, + 0.015809061005711555, + 0.019899481907486916, + 0.0227656252682209, + 0.01216563954949379, + 0.015973897650837898, + 0.0077238501980900764, + 0.010180888697504997, + 0.008091418072581291, + 0.00727103790268302, + 0.010447224602103233, + 0.008220125921070576, + 0.0050961836241185665, + 0.0048118773847818375, + 0.011053880676627159, + 0.017291897907853127, + 0.01019507460296154, + 0.01136085856705904, + 0.0108428243547678, + 0.01340707577764988, + 0.007874589413404465, + 0.009727062657475471, + 0.12324483692646027, + 0.01600421965122223, + 0.028622504323720932, + 0.013535473495721817, + 0.11020486801862717, + 0.023439301177859306, + 0.02002079226076603, + 0.029503026977181435, + 0.023331735283136368, + 0.008217153139412403, + 0.010800247080624104, + 0.007131104823201895, + 0.005068009719252586, + 0.010060486383736134, + 0.005180457606911659, + 0.01975982077419758, + 0.003903248580172658, + 0.005613900721073151, + 0.003505089320242405, + 0.006416233256459236, + 0.0035179562401026487, + 0.0042163473553955555, + 0.00299178808927536, + 0.0023294584825634956, + 0.013225596398115158, + 0.004800064954906702, + 0.0070193856954574585, + 0.00515711447224021, + 0.0055805230513215065, + 0.006271503400057554, + 0.004506710451096296, + 0.00790567509829998, + 0.009317991323769093, + 0.008388547226786613, + 0.007725392933934927, + 0.008584841154515743, + 0.009027334861457348, + 0.008732236921787262, + 0.009116385132074356, + 0.006847647950053215, + 0.003459707833826542, + 0.002214042702689767, + 0.00318423705175519, + 0.004465286619961262, + 0.003922626376152039, + 0.003401524852961302, + 0.0040033068507909775, + 0.003132506739348173, + 0.015201952308416367, + 0.007636252790689468, + 0.008597276173532009, + 0.004561000503599644, + 0.005727046635001898, + 0.007831287570297718, + 0.007089734077453613, + 0.006848541088402271 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 0, + 2 + ], + "selected_reward": -0.035820648074150085 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video shot as a locked, static, eye-level close-up?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a tiny crab with a glossy warm orange shell and glowing black eyes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the crab nesting into a vibrant sea anemone anchored to sun-dappled coral?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the anemone's tentacles sway gently, part to welcome the crab, and curl up to protect it?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by sunlight filtering through turquoise water, casting shimmering patterns on the sandy seafloor?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.014209521003067493 + ], + "opposite_logprob_tokens": [ + -4.264209747314453 + ], + "logprob_reward": -0.014209521003067493, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.014209521003067493, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10314668715000153, + 0.435761958360672, + 0.0943404957652092, + 0.03939612954854965, + 0.10697068274021149, + 0.024182934314012527, + 0.04651809111237526, + 0.045930080115795135, + 0.06273359060287476, + 0.04101931303739548 + ], + "credit_map": [ + 0.010359184816479683, + 0.010854875668883324, + 0.01647130586206913, + 0.016088049858808517, + 0.018329372629523277, + 0.012023688293993473, + 0.009617829695343971, + 0.009402402676641941, + 0.07204224169254303, + 0.01674993894994259, + 0.028376221656799316, + 0.01635909266769886, + 0.18094055354595184, + 0.0364820659160614, + 0.023612389340996742, + 0.061199478805065155, + 0.016514157876372337, + 0.017448991537094116, + 0.009393789805471897, + 0.005884713958948851, + 0.011530031450092793, + 0.014898493885993958, + 0.006271532271057367, + 0.012398790568113327, + 0.008277966640889645, + 0.00457256194204092, + 0.004875033628195524, + 0.0029263331089168787, + 0.0038908878341317177, + 0.004749965853989124, + 0.0040261950343847275, + 0.006077194586396217, + 0.03060341253876686, + 0.009648248553276062, + 0.010640304535627365, + 0.010224947705864906, + 0.006984546780586243, + 0.009409117512404919, + 0.0059698899276554585, + 0.023490235209465027, + 0.0031666532158851624, + 0.0037390831857919693, + 0.002468693070113659, + 0.004408311098814011, + 0.003024398349225521, + 0.003165891394019127, + 0.0022361462470144033, + 0.001973760547116399, + 0.0072649382054805756, + 0.004924653563648462, + 0.0048174867406487465, + 0.006950381677597761, + 0.004477180540561676, + 0.005736777558922768, + 0.007824243046343327, + 0.004522434901446104, + 0.005934291984885931, + 0.005325526464730501, + 0.006311721168458462, + 0.00616877805441618, + 0.0066226087510585785, + 0.005422778893262148, + 0.005880436860024929, + 0.004263945389539003, + 0.010984590277075768, + 0.0058428505435585976, + 0.007244960870593786, + 0.006158583331853151, + 0.006922196596860886, + 0.012096264399588108, + 0.006769038271158934, + 0.00671512121334672, + 0.007434461265802383, + 0.005016825161874294, + 0.005190761759877205, + 0.004094998352229595, + 0.003530171699821949, + 0.005500270985066891, + 0.005829682573676109, + 0.0044221472926437855 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 4, + 0 + ], + "selected_reward": -0.014209521003067493 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video shot as a locked, static, eye-level close-up?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a tiny crab with a glossy warm orange shell and glowing black eyes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the crab nesting into a vibrant sea anemone anchored to sun-dappled coral?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the anemone's tentacles sway gently, part to welcome the crab, and curl up to protect it?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by sunlight filtering through turquoise water, casting shimmering patterns on the sandy seafloor?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.11294502019882202 + ], + "opposite_logprob_tokens": [ + -2.237945079803467 + ], + "logprob_reward": -0.11294502019882202, + "yes_no_margin_reward": 2.125, + "token_credit_reward": -0.11294502019882202, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13973861932754517, + 0.13254912197589874, + 0.08460556715726852, + 0.2060212641954422, + 0.10588451474905014, + 0.03836226835846901, + 0.07494796067476273, + 0.12448444217443466, + 0.03060832992196083, + 0.0627979263663292 + ], + "credit_map": [ + 0.01345696672797203, + 0.01487788651138544, + 0.01959392987191677, + 0.01969330757856369, + 0.023796197026968002, + 0.01889129914343357, + 0.01596580259501934, + 0.013463214971125126, + 0.011561267077922821, + 0.006330719217658043, + 0.010715407319366932, + 0.005315274000167847, + 0.04432045668363571, + 0.012991354800760746, + 0.008820162154734135, + 0.032494477927684784, + 0.014132587239146233, + 0.009386547841131687, + 0.00789437536150217, + 0.005895807407796383, + 0.01649058423936367, + 0.013793724589049816, + 0.008082905784249306, + 0.008929035626351833, + 0.027973325923085213, + 0.014074328355491161, + 0.014290720224380493, + 0.0153348408639431, + 0.07273374497890472, + 0.02001294679939747, + 0.012030825018882751, + 0.02957053855061531, + 0.03290176764130592, + 0.0074375648982822895, + 0.01217135414481163, + 0.004840285517275333, + 0.005938808899372816, + 0.010827440768480301, + 0.006781810894608498, + 0.024985486641526222, + 0.005629991181194782, + 0.004591572564095259, + 0.0037559166084975004, + 0.0052753048948943615, + 0.004275428131222725, + 0.0065024662762880325, + 0.004033103119581938, + 0.004298486281186342, + 0.023012330755591393, + 0.006519146729260683, + 0.009549052454531193, + 0.00827628094702959, + 0.004450042266398668, + 0.0052522532641887665, + 0.005041217431426048, + 0.012847639620304108, + 0.016976594924926758, + 0.01689332164824009, + 0.023307273164391518, + 0.012984136119484901, + 0.015206766314804554, + 0.015546640381217003, + 0.013138570822775364, + 0.010431136935949326, + 0.0037414601538330317, + 0.0028407806530594826, + 0.004093641880899668, + 0.0045615085400640965, + 0.003895734902471304, + 0.003998816478997469, + 0.0037132189609110355, + 0.003763168817386031, + 0.011953226290643215, + 0.006241009570658207, + 0.010484055615961552, + 0.006267609540373087, + 0.004468701779842377, + 0.008357551880180836, + 0.007885908707976341, + 0.007139858324080706 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 0, + 1 + ], + "selected_reward": -0.11294502019882202 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video shot as a locked, static, eye-level close-up?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a tiny crab with a glossy warm orange shell and glowing black eyes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the crab nesting into a vibrant sea anemone anchored to sun-dappled coral?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the anemone's tentacles sway gently, part to welcome the crab, and curl up to protect it?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by sunlight filtering through turquoise water, casting shimmering patterns on the sandy seafloor?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.018206970766186714 + ], + "opposite_logprob_tokens": [ + -4.01820707321167 + ], + "logprob_reward": -0.018206970766186714, + "yes_no_margin_reward": 4.0, + "token_credit_reward": -0.018206970766186714, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1809041053056717, + 0.24765516817569733, + 0.12813043594360352, + 0.17098872363567352, + 0.06658005714416504, + 0.03894798457622528, + 0.06156131252646446, + 0.04453355073928833, + 0.02436337061226368, + 0.03633522987365723 + ], + "credit_map": [ + 0.018899783492088318, + 0.021545199677348137, + 0.02330494299530983, + 0.02315678261220455, + 0.03253421559929848, + 0.029024524614214897, + 0.016641320660710335, + 0.015797359868884087, + 0.03543373942375183, + 0.013511305674910545, + 0.018350422382354736, + 0.011453874409198761, + 0.09455003589391708, + 0.022355636581778526, + 0.01712411269545555, + 0.03487606719136238, + 0.018630506470799446, + 0.015113326720893383, + 0.0128211984410882, + 0.008314107544720173, + 0.027244918048381805, + 0.015445959754288197, + 0.013767647556960583, + 0.016792789101600647, + 0.03469138219952583, + 0.009274715557694435, + 0.010783740319311619, + 0.010441439226269722, + 0.06620030105113983, + 0.01319157425314188, + 0.009347546845674515, + 0.017058026045560837, + 0.009916899725794792, + 0.007384256459772587, + 0.007039086893200874, + 0.010552482679486275, + 0.006523937918245792, + 0.007830792106688023, + 0.0070778317749500275, + 0.010254772379994392, + 0.006715264171361923, + 0.00631747068837285, + 0.003901823190972209, + 0.004519120790064335, + 0.00376470061019063, + 0.005138427019119263, + 0.004649315029382706, + 0.003941865172237158, + 0.013023347593843937, + 0.005610145162791014, + 0.00713513744994998, + 0.008427363820374012, + 0.004228261299431324, + 0.006564565468579531, + 0.008203523233532906, + 0.00836897548288107, + 0.006911974400281906, + 0.0046264296397566795, + 0.007342672441154718, + 0.0049737608060240746, + 0.004291946534067392, + 0.0069007123820483685, + 0.0049121566116809845, + 0.0045739030465483665, + 0.00362213933840394, + 0.003149676602333784, + 0.0033497666008770466, + 0.0023206425830721855, + 0.002641181694343686, + 0.003971227910369635, + 0.002941229846328497, + 0.0023675078991800547, + 0.0058047096244990826, + 0.003604529658332467, + 0.004945266991853714, + 0.003933868836611509, + 0.0026250199880450964, + 0.004925383720546961, + 0.006086694076657295, + 0.004409759771078825 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 3 + ], + "selected_reward": -0.018206970766186714 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video shot as a locked, static, eye-level close-up?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a tiny crab with a glossy warm orange shell and glowing black eyes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the crab nesting into a vibrant sea anemone anchored to sun-dappled coral?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the anemone's tentacles sway gently, part to welcome the crab, and curl up to protect it?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by sunlight filtering through turquoise water, casting shimmering patterns on the sandy seafloor?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.11292159557342529 + ], + "opposite_logprob_tokens": [ + -2.237921714782715 + ], + "logprob_reward": -0.11292159557342529, + "yes_no_margin_reward": 2.125, + "token_credit_reward": -0.11292159557342529, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0940682515501976, + 0.21974313259124756, + 0.06125962361693382, + 0.2538621425628662, + 0.06616859138011932, + 0.02169097773730755, + 0.0949583649635315, + 0.10090640932321548, + 0.030163872987031937, + 0.057178694754838943 + ], + "credit_map": [ + 0.009482672438025475, + 0.009311563335359097, + 0.013447091914713383, + 0.010418681427836418, + 0.016307739540934563, + 0.013550972566008568, + 0.011664137244224548, + 0.00988539308309555, + 0.03026021085679531, + 0.007940971292555332, + 0.014889269135892391, + 0.008185486309230328, + 0.0912749245762825, + 0.01866133324801922, + 0.010869830846786499, + 0.0376611202955246, + 0.01270811352878809, + 0.006108745001256466, + 0.0067659057676792145, + 0.0038362026680260897, + 0.009558211080729961, + 0.009952436201274395, + 0.005163012072443962, + 0.00716699892655015, + 0.026398032903671265, + 0.018428290262818336, + 0.018103517591953278, + 0.018725551664829254, + 0.08836781233549118, + 0.02536698803305626, + 0.013625193387269974, + 0.04484674334526062, + 0.019107915461063385, + 0.0051451087929308414, + 0.006844052113592625, + 0.004375964868813753, + 0.003987960983067751, + 0.009188090451061726, + 0.004673149436712265, + 0.012846353463828564, + 0.003307158360257745, + 0.0023279415909200907, + 0.0028691620100289583, + 0.0027664154767990112, + 0.0017855126643553376, + 0.003776909550651908, + 0.0023369507398456335, + 0.002520926296710968, + 0.026627719402313232, + 0.008110709488391876, + 0.011154034174978733, + 0.011069456115365028, + 0.00493737543001771, + 0.007361819967627525, + 0.00914863869547844, + 0.016548609361052513, + 0.014014959335327148, + 0.01384508702903986, + 0.01856905221939087, + 0.010568703524768353, + 0.012070898897945881, + 0.012990057468414307, + 0.010191813111305237, + 0.008655838668346405, + 0.0036564753390848637, + 0.002754250541329384, + 0.0037807088810950518, + 0.004504543263465166, + 0.003771549789234996, + 0.004115645308047533, + 0.003970318008214235, + 0.0036103827878832817, + 0.011733085848391056, + 0.004982075188308954, + 0.008573023602366447, + 0.005979903507977724, + 0.0038494470063596964, + 0.007838635705411434, + 0.008354407735168934, + 0.005868114531040192 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 1, + 7 + ], + "selected_reward": -0.11292159557342529 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "In charming Disney-Pixar style, a locked, static, eye-level close-up captures a tender underwater moment, focusing on the heartwarming interaction between two creatures. A tiny crab, with a glossy, warm orange shell and glowing black eyes filled with innocent wonder, carefully nestles into its new home: a vibrant sea anemone. Anchored to a patch of sun-dappled coral, the anemone is vibrantly colored, with a deep purplish-red base from which sprouts a dense cluster of soft, fleshy tentacles. Gently swaying in the calm current, these tentacles are translucent, their interiors glowing with gentle bioluminescence, their tips emitting a brilliant azure blue. With slow, careful movements, the crab's slender legs gently move, tucking its tiny body into the anemone's embrace; in response, the tentacles gently part as if in welcome, then gently curl up, protectively surrounding its shell. Sunlight filtering through the crystal-clear turquoise water casts shimmering, corroded patterns on the sandy seafloor and illuminates the scene with a warm, ethereal glow, creating a tranquil and magical atmosphere of finding a safe, comforting home.", + "seed": 32584, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000101/group_12/index_1054_seed_32584_dp_25_rank_50_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 12, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.0722994316369295, + "vlm_logprob": -0.0722994316369295, + "vlm_yes_no_margin": 2.95, + "vlm_token_credit": -0.0722994316369295, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video shot as a locked, static, eye-level close-up?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a tiny crab with a glossy warm orange shell and glowing black eyes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the crab nesting into a vibrant sea anemone anchored to sun-dappled coral?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the anemone's tentacles sway gently, part to welcome the crab, and curl up to protect it?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by sunlight filtering through turquoise water, casting shimmering patterns on the sandy seafloor?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0966527909040451 + ], + "opposite_logprob_tokens": [ + -2.5966527462005615 + ], + "logprob_reward": -0.0966527909040451, + "yes_no_margin_reward": 2.5, + "token_credit_reward": -0.0966527909040451, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10665936022996902, + 0.0877418965101242, + 0.1134648472070694, + 0.10317550599575043, + 0.06173865124583244, + 0.037788327783346176, + 0.2629123628139496, + 0.13993364572525024, + 0.05171773210167885, + 0.034867677837610245 + ], + "credit_map": [ + 0.012902124784886837, + 0.013968991115689278, + 0.014858132228255272, + 0.010698716156184673, + 0.013920687139034271, + 0.011053833179175854, + 0.016585057601332664, + 0.012671818025410175, + 0.015073507092893124, + 0.013978841714560986, + 0.009315838105976582, + 0.006839428562670946, + 0.01147542241960764, + 0.010410936549305916, + 0.009325647726655006, + 0.011322272941470146, + 0.01753254421055317, + 0.01434934139251709, + 0.013898462988436222, + 0.011277364566922188, + 0.0177658349275589, + 0.013926640152931213, + 0.012185538187623024, + 0.012529121711850166, + 0.011793073266744614, + 0.016004085540771484, + 0.014744752086699009, + 0.008359731175005436, + 0.01666661724448204, + 0.011217126622796059, + 0.01252369862049818, + 0.011866413056850433, + 0.00893853884190321, + 0.01016707718372345, + 0.00839962437748909, + 0.004878712818026543, + 0.00743387034162879, + 0.006975133903324604, + 0.00502222403883934, + 0.009923473000526428, + 0.005132126156240702, + 0.0038169100880622864, + 0.006102938205003738, + 0.005125205032527447, + 0.0030380692332983017, + 0.003483495907858014, + 0.004470114130526781, + 0.006619472522288561, + 0.0396900400519371, + 0.038272302597761154, + 0.0315152071416378, + 0.028268562629818916, + 0.03587627783417702, + 0.026176469400525093, + 0.02381138876080513, + 0.039302125573158264, + 0.022194204851984978, + 0.017620254307985306, + 0.02340669184923172, + 0.01571050100028515, + 0.013825065456330776, + 0.015192859806120396, + 0.01649796962738037, + 0.015486100688576698, + 0.0074423193000257015, + 0.005159350112080574, + 0.008903227746486664, + 0.0067074717953801155, + 0.004354716744273901, + 0.004724088590592146, + 0.008549005724489689, + 0.005877552088350058, + 0.005876720417290926, + 0.004237235523760319, + 0.0031438141595572233, + 0.0049957693554461, + 0.003763374872505665, + 0.003964577801525593, + 0.005199025385081768, + 0.0036871577613055706 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 7, + 2 + ], + "selected_reward": -0.0966527909040451 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video shot as a locked, static, eye-level close-up?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a tiny crab with a glossy warm orange shell and glowing black eyes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the crab nesting into a vibrant sea anemone anchored to sun-dappled coral?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the anemone's tentacles sway gently, part to welcome the crab, and curl up to protect it?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by sunlight filtering through turquoise water, casting shimmering patterns on the sandy seafloor?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.026364192366600037 + ], + "opposite_logprob_tokens": [ + -3.6513640880584717 + ], + "logprob_reward": -0.026364192366600037, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.026364192366600037, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.23784592747688293, + 0.1170627772808075, + 0.0253863874822855, + 0.03235683962702751, + 0.15566928684711456, + 0.03893505409359932, + 0.10678372532129288, + 0.11134298890829086, + 0.049963876605033875, + 0.1246531531214714 + ], + "credit_map": [ + 0.03184177726507187, + 0.028944384306669235, + 0.032173484563827515, + 0.019823109731078148, + 0.03685155138373375, + 0.02777557447552681, + 0.03305591270327568, + 0.027380136772990227, + 0.022099904716014862, + 0.01661515422165394, + 0.012571045197546482, + 0.009010271169245243, + 0.0158473439514637, + 0.014765510335564613, + 0.011937443166971207, + 0.014216099865734577, + 0.0037440068554133177, + 0.0033565268386155367, + 0.0035423319786787033, + 0.002315690740942955, + 0.003414844162762165, + 0.002843231428414583, + 0.003068363294005394, + 0.0031013907864689827, + 0.005686767864972353, + 0.004976906813681126, + 0.0033179996535182, + 0.002159496070817113, + 0.004739155527204275, + 0.003790447721257806, + 0.002906148787587881, + 0.004779913928359747, + 0.019654056057333946, + 0.029099591076374054, + 0.0166306272149086, + 0.014343592338263988, + 0.018190940842032433, + 0.016597267240285873, + 0.012977772392332554, + 0.028175435960292816, + 0.008245816454291344, + 0.00644764956086874, + 0.003911648876965046, + 0.003577746683731675, + 0.0038001483771950006, + 0.004441319964826107, + 0.003434606594964862, + 0.005076118744909763, + 0.013529916293919086, + 0.01648828387260437, + 0.014836723916232586, + 0.014250189997255802, + 0.014770771376788616, + 0.010498300194740295, + 0.009193887934088707, + 0.013215660117566586, + 0.02329845540225506, + 0.014791272580623627, + 0.015058309771120548, + 0.01170464139431715, + 0.013194985687732697, + 0.011896414682269096, + 0.012706016190350056, + 0.008692892268300056, + 0.006094167474657297, + 0.006890559569001198, + 0.008154724724590778, + 0.007030584383755922, + 0.005970566533505917, + 0.004545243922621012, + 0.006712506525218487, + 0.004565523937344551, + 0.016314134001731873, + 0.017489539459347725, + 0.010975334793329239, + 0.021858878433704376, + 0.015327728353440762, + 0.01327542494982481, + 0.01741524040699005, + 0.011996875517070293 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 9 + ], + "selected_reward": -0.026364192366600037 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video shot as a locked, static, eye-level close-up?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a tiny crab with a glossy warm orange shell and glowing black eyes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the crab nesting into a vibrant sea anemone anchored to sun-dappled coral?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the anemone's tentacles sway gently, part to welcome the crab, and curl up to protect it?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by sunlight filtering through turquoise water, casting shimmering patterns on the sandy seafloor?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.16035784780979156 + ], + "opposite_logprob_tokens": [ + -1.9103578329086304 + ], + "logprob_reward": -0.16035784780979156, + "yes_no_margin_reward": 1.75, + "token_credit_reward": -0.16035784780979156, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2904820740222931, + 0.08516897261142731, + 0.0719919502735138, + 0.11563047021627426, + 0.12396346032619476, + 0.047462739050388336, + 0.08245648443698883, + 0.07014565169811249, + 0.051996320486068726, + 0.06070194020867348 + ], + "credit_map": [ + 0.047547347843647, + 0.03742389380931854, + 0.036764632910490036, + 0.021631747484207153, + 0.04567413404583931, + 0.03482159227132797, + 0.034583110362291336, + 0.03203555941581726, + 0.017919139936566353, + 0.010833162814378738, + 0.008982658386230469, + 0.005970186088234186, + 0.01094875205308199, + 0.010835803113877773, + 0.008900117129087448, + 0.010779141448438168, + 0.013247611001133919, + 0.00886312685906887, + 0.008606338873505592, + 0.00645090825855732, + 0.008777732029557228, + 0.009030518122017384, + 0.008227294310927391, + 0.008788413368165493, + 0.016053365543484688, + 0.0200909785926342, + 0.012850700877606869, + 0.008773240260779858, + 0.020184412598609924, + 0.01379938330501318, + 0.013183481059968472, + 0.010694890283048153, + 0.019514579325914383, + 0.016746021807193756, + 0.013701491057872772, + 0.010180399753153324, + 0.009651588276028633, + 0.01240764744579792, + 0.01325193326920271, + 0.028509793803095818, + 0.006292796693742275, + 0.007592164911329746, + 0.005928609985858202, + 0.005856429226696491, + 0.004429251421242952, + 0.004904261790215969, + 0.005839593708515167, + 0.006619627587497234, + 0.013214652426540852, + 0.012333680875599384, + 0.009290301240980625, + 0.011157983914017677, + 0.011447392404079437, + 0.008214469999074936, + 0.00820502731949091, + 0.008592969737946987, + 0.013513335958123207, + 0.009999828413128853, + 0.00912998616695404, + 0.0079729575663805, + 0.008698985911905766, + 0.0072749946266412735, + 0.008085104636847973, + 0.005470450967550278, + 0.005966430064290762, + 0.006945058237761259, + 0.005608793813735247, + 0.007971308194100857, + 0.007043361663818359, + 0.005628966726362705, + 0.0083586685359478, + 0.004473725333809853, + 0.008247917518019676, + 0.008484986610710621, + 0.0062149157747626305, + 0.010282617062330246, + 0.007359932642430067, + 0.006507591810077429, + 0.008023563772439957, + 0.0055804080329835415 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 3 + ], + "selected_reward": -0.16035784780979156 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video shot as a locked, static, eye-level close-up?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a tiny crab with a glossy warm orange shell and glowing black eyes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the crab nesting into a vibrant sea anemone anchored to sun-dappled coral?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the anemone's tentacles sway gently, part to welcome the crab, and curl up to protect it?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by sunlight filtering through turquoise water, casting shimmering patterns on the sandy seafloor?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01608244515955448 + ], + "opposite_logprob_tokens": [ + -4.141082286834717 + ], + "logprob_reward": -0.01608244515955448, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.01608244515955448, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.21234750747680664, + 0.09017835557460785, + 0.07902263104915619, + 0.06114235147833824, + 0.10664136707782745, + 0.04928536340594292, + 0.13923689723014832, + 0.11457894742488861, + 0.035260193049907684, + 0.1123063862323761 + ], + "credit_map": [ + 0.027293501421809196, + 0.024807238951325417, + 0.029310647398233414, + 0.01869281567633152, + 0.03360217437148094, + 0.024217307567596436, + 0.030636440962553024, + 0.0237873662263155, + 0.015004927292466164, + 0.010126993060112, + 0.011013075709342957, + 0.00843286793678999, + 0.010832984000444412, + 0.010849660262465477, + 0.010521558113396168, + 0.013396281749010086, + 0.011450325138866901, + 0.012322810478508472, + 0.009916652925312519, + 0.0068906135857105255, + 0.013284374959766865, + 0.009713070467114449, + 0.008439668454229832, + 0.007005103398114443, + 0.008207788690924644, + 0.007645054254680872, + 0.008498242124915123, + 0.004774291068315506, + 0.007683842908591032, + 0.00654584588482976, + 0.009131073951721191, + 0.008656207472085953, + 0.013027090579271317, + 0.020447181537747383, + 0.011101415380835533, + 0.011895982548594475, + 0.011781729757785797, + 0.011637127958238125, + 0.010518553666770458, + 0.01623227633535862, + 0.005258129443973303, + 0.009089385159313679, + 0.005975812673568726, + 0.006438134703785181, + 0.006050813477486372, + 0.005333915818482637, + 0.004932736046612263, + 0.006206430960446596, + 0.021331628784537315, + 0.018217185512185097, + 0.01933174580335617, + 0.018680622801184654, + 0.012383361347019672, + 0.015093386173248291, + 0.013241048902273178, + 0.02095790207386017, + 0.022727960720658302, + 0.016233479604125023, + 0.01440420188009739, + 0.01273286622017622, + 0.015622060745954514, + 0.011816564947366714, + 0.012348447926342487, + 0.008693361654877663, + 0.004819310735911131, + 0.003910138737410307, + 0.004138010088354349, + 0.005324574187397957, + 0.0050484067760407925, + 0.0034913853742182255, + 0.004966360051184893, + 0.0035620066337287426, + 0.01746678166091442, + 0.014107386581599712, + 0.010633873753249645, + 0.017221787944436073, + 0.011504093185067177, + 0.012069672346115112, + 0.018668299540877342, + 0.01063448190689087 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 6, + 7 + ], + "selected_reward": -0.01608244515955448 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video shot as a locked, static, eye-level close-up?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a tiny crab with a glossy warm orange shell and glowing black eyes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the crab nesting into a vibrant sea anemone anchored to sun-dappled coral?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the anemone's tentacles sway gently, part to welcome the crab, and curl up to protect it?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by sunlight filtering through turquoise water, casting shimmering patterns on the sandy seafloor?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.06203988194465637 + ], + "opposite_logprob_tokens": [ + -2.812039852142334 + ], + "logprob_reward": -0.06203988194465637, + "yes_no_margin_reward": 2.75, + "token_credit_reward": -0.06203988194465637, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.28394073247909546, + 0.09854496270418167, + 0.05003876984119415, + 0.05360342562198639, + 0.12740404903888702, + 0.058677706867456436, + 0.106606625020504, + 0.08850334584712982, + 0.04251085966825485, + 0.0901695266366005 + ], + "credit_map": [ + 0.037562642246484756, + 0.03493814915418625, + 0.036963097751140594, + 0.022132262587547302, + 0.04628799855709076, + 0.033389780670404434, + 0.039346229285001755, + 0.03332056105136871, + 0.017707059159874916, + 0.014516359195113182, + 0.010052607394754887, + 0.0078501645475626, + 0.013811163604259491, + 0.01269195880740881, + 0.010482467710971832, + 0.011433186940848827, + 0.0067413123324513435, + 0.006835387088358402, + 0.006373315118253231, + 0.00454295426607132, + 0.008220488205552101, + 0.005986198782920837, + 0.0055667320266366005, + 0.005772377829998732, + 0.008085590787231922, + 0.009632381610572338, + 0.006141867022961378, + 0.003918684087693691, + 0.007906930521130562, + 0.006215158384293318, + 0.005994404666125774, + 0.005708410404622555, + 0.017657114192843437, + 0.02217998541891575, + 0.013641614466905594, + 0.011294287629425526, + 0.01371195912361145, + 0.014146982692182064, + 0.011302107945084572, + 0.023469991981983185, + 0.0076797413639724255, + 0.009469892829656601, + 0.007126845419406891, + 0.006933076772838831, + 0.008803293108940125, + 0.006489783525466919, + 0.006003507878631353, + 0.00617156783118844, + 0.01731775887310505, + 0.0148153780028224, + 0.013413535431027412, + 0.014240570366382599, + 0.013406403362751007, + 0.010905907489359379, + 0.010790913365781307, + 0.011716152541339397, + 0.017654338851571083, + 0.011984385550022125, + 0.012174873612821102, + 0.00937677826732397, + 0.011497879400849342, + 0.00937272422015667, + 0.00899040512740612, + 0.007451958954334259, + 0.0047872052527964115, + 0.00632877741008997, + 0.005788165144622326, + 0.006095026154071093, + 0.005477698519825935, + 0.004208491649478674, + 0.006165182217955589, + 0.0036603149492293596, + 0.013485993258655071, + 0.012568255886435509, + 0.007888616062700748, + 0.014580957591533661, + 0.01142911147326231, + 0.009900229051709175, + 0.010852825827896595, + 0.009463543072342873 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 6 + ], + "selected_reward": -0.06203988194465637 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000101/group_13/index_tensor([1275, 1275])_group_13_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000101/group_13/index_tensor([1275, 1275])_group_13_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..c888a29ebfd814fd3115b1a90606c6c8aee5feb1 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000101/group_13/index_tensor([1275, 1275])_group_13_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "The film begins with a gentle medium shot. Deep in a tranquil bamboo forest, bathed in cool, silver moonlight, stands a healer of sagelike demeanor. His long, moon-like silver hair is loosely tied with a simple wooden hairpin. He wears a flowing, moon-white robe made of natural linen, its sleeves sweeping gently in the night breeze. With exquisite grace, he kneels on the soft, mossy ground. Before him, a fawn, its fur speckled with white, appears exceptionally fragile. Its large, moist eyes are filled with utter trust. The healer carefully tends to a small wound on its foreleg. With his slender fingertips, he scoops up a herbal solution from a small ceramic bowl beside him and gently, calmly, applies it to the wound. His focused and compassionate expression appears remarkably sacred in the soft light. Towering, emerald green bamboos sway gently in the night breeze. Moonlight filters through the leaves, casting dappled, flickering shadows on the ground. A few fireflies flicker in the air, adding a touch of dreamy vitality to the tranquil scene. The entire scene is imbued with the distinctive hand-painted aesthetic of Ghibli animation, with soft colors and flowing lines, imbued with a sense of tranquility and tenderness. As the camera steadily and slowly tilts upward, the intimate interaction between the healer and the deer gradually gives way to a broader landscape, ultimately revealing a grand canvas: a boundless sea of bamboo undulating like waves under the moonlight, with the inky outlines of rolling mountains in the distance. A bright moon hangs high in the tranquil night sky, creating a vast and healing dreamlike atmosphere.", + "seed": 11046, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000101/group_13/index_1275_seed_11046_dp_26_rank_52_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 13, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.08228041613474488, + "vlm_logprob": -0.08228041613474488, + "vlm_yes_no_margin": 3.7, + "vlm_token_credit": -0.08228041613474488, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a bamboo forest at night with a bright moon in the sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a healer with long silver hair tied by a wooden hairpin and wearing a moon-white linen robe visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the healer kneel on mossy ground and apply a herbal solution from a ceramic bowl to a fawn's foreleg wound?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the video feature emerald green bamboo swaying, dappled shadows, and flickering fireflies?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start with a medium shot and steadily tilt upward to reveal a vast landscape of bamboo and mountains?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.009950675070285797 + ], + "opposite_logprob_tokens": [ + -5.009950637817383 + ], + "logprob_reward": -0.009950675070285797, + "yes_no_margin_reward": 5.0, + "token_credit_reward": -0.009950675070285797, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.16157537698745728, + 0.06880777329206467, + 0.03556577488780022, + 0.03224990889430046, + 0.03931377828121185, + 0.06998182088136673, + 0.33080971240997314, + 0.07017941027879715, + 0.05175141245126724, + 0.13976500928401947 + ], + "credit_map": [ + 0.032228078693151474, + 0.012959771789610386, + 0.02987920492887497, + 0.018645726144313812, + 0.011537918820977211, + 0.014287302270531654, + 0.02011583372950554, + 0.021921543404459953, + 0.006252474617213011, + 0.008256413042545319, + 0.008789191953837872, + 0.012253417633473873, + 0.0050120302475988865, + 0.00785538274794817, + 0.009863009676337242, + 0.010525855235755444, + 0.0029497144278138876, + 0.005631543695926666, + 0.003635840257629752, + 0.00379932951182127, + 0.004973812494426966, + 0.003060439834371209, + 0.008174547925591469, + 0.0033405479043722153, + 0.0023409384302794933, + 0.0029754559509456158, + 0.003684538882225752, + 0.004910305608063936, + 0.006763780023902655, + 0.0021992945112288, + 0.004563020076602697, + 0.00481257401406765, + 0.002731656190007925, + 0.0051902118138968945, + 0.004216916859149933, + 0.005224819760769606, + 0.008615857921540737, + 0.0029217146802693605, + 0.004289328586310148, + 0.006123274099081755, + 0.005366487428545952, + 0.009970405139029026, + 0.009968781843781471, + 0.006048211362212896, + 0.00682695209980011, + 0.0047991108149290085, + 0.00569871487095952, + 0.021303152665495872, + 0.026960400864481926, + 0.02402493916451931, + 0.10895824432373047, + 0.024512508884072304, + 0.02029215171933174, + 0.016427356749773026, + 0.0413055382668972, + 0.06832855939865112, + 0.007242849096655846, + 0.01299320813268423, + 0.007908218540251255, + 0.012505915015935898, + 0.007349198684096336, + 0.006078081671148539, + 0.010810339823365211, + 0.0052916002459824085, + 0.006024863105267286, + 0.006704174913465977, + 0.005478609818965197, + 0.008262732066214085, + 0.006969989277422428, + 0.004018235951662064, + 0.010554618202149868, + 0.00373818795196712, + 0.01706097461283207, + 0.021105775609612465, + 0.00730900140479207, + 0.018397435545921326, + 0.028378579765558243, + 0.016829226166009903, + 0.020943844690918922, + 0.009740166366100311 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 0, + 9 + ], + "selected_reward": -0.009950675070285797 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a bamboo forest at night with a bright moon in the sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a healer with long silver hair tied by a wooden hairpin and wearing a moon-white linen robe visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the healer kneel on mossy ground and apply a herbal solution from a ceramic bowl to a fawn's foreleg wound?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the video feature emerald green bamboo swaying, dappled shadows, and flickering fireflies?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start with a medium shot and steadily tilt upward to reveal a vast landscape of bamboo and mountains?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005955806467682123 + ], + "opposite_logprob_tokens": [ + -5.130955696105957 + ], + "logprob_reward": -0.005955806467682123, + "yes_no_margin_reward": 5.125, + "token_credit_reward": -0.005955806467682123, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1194087564945221, + 0.04680937901139259, + 0.03780153766274452, + 0.04953261464834213, + 0.03321436792612076, + 0.20918089151382446, + 0.2324185073375702, + 0.15899959206581116, + 0.044586777687072754, + 0.06804756820201874 + ], + "credit_map": [ + 0.015316251665353775, + 0.009139448404312134, + 0.02445460483431816, + 0.017649659886956215, + 0.008702174760401249, + 0.00975426472723484, + 0.015219295397400856, + 0.01917305774986744, + 0.004081642720848322, + 0.005737637635320425, + 0.004495599772781134, + 0.006966530345380306, + 0.00565290218219161, + 0.004657594952732325, + 0.007430586498230696, + 0.007786881644278765, + 0.0032958548981696367, + 0.00621040491387248, + 0.004489743150770664, + 0.004130829591304064, + 0.0059170788154006, + 0.003658998990431428, + 0.006868680939078331, + 0.0032299484591931105, + 0.0032875165343284607, + 0.006228760350495577, + 0.004957897588610649, + 0.006252686958760023, + 0.012860455550253391, + 0.003237252589315176, + 0.007687810342758894, + 0.005020237527787685, + 0.0024473622906953096, + 0.004443504847586155, + 0.0039000948891043663, + 0.004052802920341492, + 0.00636283727362752, + 0.003514793235808611, + 0.004029891919344664, + 0.004463080782443285, + 0.011909808963537216, + 0.025642603635787964, + 0.024889709427952766, + 0.01658838428556919, + 0.04490021988749504, + 0.014163058251142502, + 0.02278420701622963, + 0.048302896320819855, + 0.0174565427005291, + 0.016774756833910942, + 0.07701990753412247, + 0.01650797389447689, + 0.01281178742647171, + 0.011184753850102425, + 0.035646870732307434, + 0.04501592740416527, + 0.014721132814884186, + 0.027334999293088913, + 0.018962150439620018, + 0.03377196937799454, + 0.014440389350056648, + 0.013078616932034492, + 0.02464207448065281, + 0.012048253789544106, + 0.005137264728546143, + 0.008147872984409332, + 0.004007422365248203, + 0.007835512049496174, + 0.005325807724148035, + 0.0025722647551447153, + 0.008357132785022259, + 0.0032034986652433872, + 0.007196478080004454, + 0.008144267834722996, + 0.0037745528388768435, + 0.009370220825076103, + 0.009854670614004135, + 0.009037056006491184, + 0.01249956339597702, + 0.008170757442712784 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 5, + 7 + ], + "selected_reward": -0.005955806467682123 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a bamboo forest at night with a bright moon in the sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a healer with long silver hair tied by a wooden hairpin and wearing a moon-white linen robe visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the healer kneel on mossy ground and apply a herbal solution from a ceramic bowl to a fawn's foreleg wound?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the video feature emerald green bamboo swaying, dappled shadows, and flickering fireflies?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start with a medium shot and steadily tilt upward to reveal a vast landscape of bamboo and mountains?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02634584717452526 + ], + "opposite_logprob_tokens": [ + -3.651345729827881 + ], + "logprob_reward": -0.02634584717452526, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.02634584717452526, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09535466134548187, + 0.06971898674964905, + 0.05485105141997337, + 0.13134443759918213, + 0.02601766027510166, + 0.13822072744369507, + 0.19307734072208405, + 0.20410273969173431, + 0.03257744014263153, + 0.05473495274782181 + ], + "credit_map": [ + 0.012534016743302345, + 0.008971715345978737, + 0.020258059725165367, + 0.013464964926242828, + 0.006658636033535004, + 0.007027352228760719, + 0.010584039613604546, + 0.01585587114095688, + 0.005894447676837444, + 0.005177369806915522, + 0.007492705248296261, + 0.01021676417440176, + 0.008328458294272423, + 0.0059966533444821835, + 0.016552703455090523, + 0.010059884749352932, + 0.006664725486189127, + 0.010239094495773315, + 0.006179200019687414, + 0.007384175434708595, + 0.005702299065887928, + 0.005439079832285643, + 0.006985332816839218, + 0.0062571438029408455, + 0.00840385165065527, + 0.018080364912748337, + 0.013056793250143528, + 0.021560093387961388, + 0.028160013258457184, + 0.008748393505811691, + 0.021469544619321823, + 0.011865389533340931, + 0.0024140370078384876, + 0.004024195484817028, + 0.003049488179385662, + 0.005378225818276405, + 0.0034787317272275686, + 0.002476937836036086, + 0.002457077382132411, + 0.0027389659080654383, + 0.00677297031506896, + 0.014853166416287422, + 0.01617591083049774, + 0.02599819377064705, + 0.02278718538582325, + 0.007279273588210344, + 0.01750371791422367, + 0.02685031294822693, + 0.01413737889379263, + 0.014637446962296963, + 0.06667692214250565, + 0.015719406306743622, + 0.008918976411223412, + 0.0074626789428293705, + 0.025965092703700066, + 0.03955943509936333, + 0.01568230614066124, + 0.03153199329972267, + 0.021857360377907753, + 0.05656244233250618, + 0.021205665543675423, + 0.015678564086556435, + 0.02768726833164692, + 0.013897133059799671, + 0.0038189017213881016, + 0.004196276422590017, + 0.0031209501903504133, + 0.005719748791307211, + 0.004607087932527065, + 0.003571383887901902, + 0.004970873706042767, + 0.002572218421846628, + 0.006559853907674551, + 0.008401885628700256, + 0.0029938644729554653, + 0.006581496447324753, + 0.0067397309467196465, + 0.0075868116691708565, + 0.010582642629742622, + 0.005288667976856232 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 6, + 5 + ], + "selected_reward": -0.02634584717452526 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a bamboo forest at night with a bright moon in the sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a healer with long silver hair tied by a wooden hairpin and wearing a moon-white linen robe visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the healer kneel on mossy ground and apply a herbal solution from a ceramic bowl to a fawn's foreleg wound?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the video feature emerald green bamboo swaying, dappled shadows, and flickering fireflies?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start with a medium shot and steadily tilt upward to reveal a vast landscape of bamboo and mountains?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.020572848618030548 + ], + "opposite_logprob_tokens": [ + -3.8955729007720947 + ], + "logprob_reward": -0.020572848618030548, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.020572848618030548, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3680908977985382, + 0.07926427572965622, + 0.11049225926399231, + 0.04495735466480255, + 0.046093765646219254, + 0.07153855264186859, + 0.060043394565582275, + 0.0623074471950531, + 0.043398115783929825, + 0.11381393671035767 + ], + "credit_map": [ + 0.057832878082990646, + 0.02201213501393795, + 0.0729442834854126, + 0.035666752606630325, + 0.04085028916597366, + 0.051504723727703094, + 0.044893987476825714, + 0.04238586500287056, + 0.008514237590134144, + 0.007663433905690908, + 0.009817549027502537, + 0.011925364844501019, + 0.008850237354636192, + 0.009936963208019733, + 0.011765243485569954, + 0.010791238397359848, + 0.006473492365330458, + 0.004794013220816851, + 0.007900039665400982, + 0.004172658547759056, + 0.017328202724456787, + 0.006185630336403847, + 0.04284539446234703, + 0.020792828872799873, + 0.0034238279331475496, + 0.005225161090493202, + 0.004974618088454008, + 0.00948240328580141, + 0.0048187798820436, + 0.0029500345699489117, + 0.009464511647820473, + 0.004618021193891764, + 0.0033232546411454678, + 0.0052701374515891075, + 0.004402705933898687, + 0.010032059624791145, + 0.00871468335390091, + 0.004008177202194929, + 0.005332855973392725, + 0.005009888671338558, + 0.004038849845528603, + 0.008851634338498116, + 0.009314409457147121, + 0.013426118530333042, + 0.0110420361161232, + 0.0038883460219949484, + 0.00864978414028883, + 0.012327378615736961, + 0.00743822380900383, + 0.005270854104310274, + 0.014602866023778915, + 0.006394531577825546, + 0.007566493470221758, + 0.0034910719841718674, + 0.008205292746424675, + 0.007074060849845409, + 0.009317201562225819, + 0.006564945448189974, + 0.007500096689909697, + 0.012917418032884598, + 0.009561698883771896, + 0.006472696550190449, + 0.005681177135556936, + 0.004292213823646307, + 0.005493497941643, + 0.004871997516602278, + 0.00499684875831008, + 0.00712378416210413, + 0.006592989433556795, + 0.00627580052241683, + 0.0048285918310284615, + 0.0032146042212843895, + 0.011969005689024925, + 0.013118601404130459, + 0.006610586307942867, + 0.020047420635819435, + 0.022405186668038368, + 0.01119860541075468, + 0.018268849700689316, + 0.010195684619247913 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 2 + ], + "selected_reward": -0.020572848618030548 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a bamboo forest at night with a bright moon in the sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a healer with long silver hair tied by a wooden hairpin and wearing a moon-white linen robe visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the healer kneel on mossy ground and apply a herbal solution from a ceramic bowl to a fawn's foreleg wound?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the video feature emerald green bamboo swaying, dappled shadows, and flickering fireflies?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start with a medium shot and steadily tilt upward to reveal a vast landscape of bamboo and mountains?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.3485769033432007 + ], + "opposite_logprob_tokens": [ + -1.2235769033432007 + ], + "logprob_reward": -0.3485769033432007, + "yes_no_margin_reward": 0.875, + "token_credit_reward": -0.3485769033432007, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12347161769866943, + 0.08505608886480331, + 0.07090526074171066, + 0.06967097520828247, + 0.040197838097810745, + 0.126290962100029, + 0.23191258311271667, + 0.09137919545173645, + 0.03393189609050751, + 0.12718358635902405 + ], + "credit_map": [ + 0.014008126221597195, + 0.019243016839027405, + 0.011792930774390697, + 0.020284583792090416, + 0.007203462067991495, + 0.021612202748656273, + 0.014288138598203659, + 0.015039156191051006, + 0.00707267178222537, + 0.011064700782299042, + 0.009507136419415474, + 0.016704430803656578, + 0.008593179285526276, + 0.0073979077860713005, + 0.013752107508480549, + 0.010963955894112587, + 0.004346547648310661, + 0.008296548388898373, + 0.005707253701984882, + 0.005585018545389175, + 0.005870356690138578, + 0.004196212626993656, + 0.0216965414583683, + 0.015206779353320599, + 0.004960021004080772, + 0.01039552316069603, + 0.008403554558753967, + 0.012066487222909927, + 0.01156881544739008, + 0.004483956843614578, + 0.007442675530910492, + 0.010349937714636326, + 0.003260706551373005, + 0.006263297516852617, + 0.004643242806196213, + 0.005908699240535498, + 0.007833660580217838, + 0.003574001369997859, + 0.004188234452158213, + 0.004525993950664997, + 0.009552499279379845, + 0.02107534371316433, + 0.01599440909922123, + 0.00935761071741581, + 0.01490830909460783, + 0.007559601217508316, + 0.011763481423258781, + 0.03607971593737602, + 0.016810784116387367, + 0.01649247668683529, + 0.07944149523973465, + 0.016359694302082062, + 0.012627250514924526, + 0.009313903748989105, + 0.03463554382324219, + 0.046231430023908615, + 0.009602434001863003, + 0.010417990386486053, + 0.009061368182301521, + 0.030824394896626472, + 0.00788428820669651, + 0.008054773323237896, + 0.007555875461548567, + 0.007978069595992565, + 0.002898952690884471, + 0.003821078920736909, + 0.003414630889892578, + 0.00527513911947608, + 0.005034246481955051, + 0.004511120263487101, + 0.005840284284204245, + 0.0031364434398710728, + 0.01402131374925375, + 0.015701178461313248, + 0.006644091103225946, + 0.01936355233192444, + 0.021802591159939766, + 0.01751641184091568, + 0.02069156989455223, + 0.011442885734140873 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 9, + 5 + ], + "selected_reward": -0.3485769033432007 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "The film begins with a gentle medium shot. Deep in a tranquil bamboo forest, bathed in cool, silver moonlight, stands a healer of sagelike demeanor. His long, moon-like silver hair is loosely tied with a simple wooden hairpin. He wears a flowing, moon-white robe made of natural linen, its sleeves sweeping gently in the night breeze. With exquisite grace, he kneels on the soft, mossy ground. Before him, a fawn, its fur speckled with white, appears exceptionally fragile. Its large, moist eyes are filled with utter trust. The healer carefully tends to a small wound on its foreleg. With his slender fingertips, he scoops up a herbal solution from a small ceramic bowl beside him and gently, calmly, applies it to the wound. His focused and compassionate expression appears remarkably sacred in the soft light. Towering, emerald green bamboos sway gently in the night breeze. Moonlight filters through the leaves, casting dappled, flickering shadows on the ground. A few fireflies flicker in the air, adding a touch of dreamy vitality to the tranquil scene. The entire scene is imbued with the distinctive hand-painted aesthetic of Ghibli animation, with soft colors and flowing lines, imbued with a sense of tranquility and tenderness. As the camera steadily and slowly tilts upward, the intimate interaction between the healer and the deer gradually gives way to a broader landscape, ultimately revealing a grand canvas: a boundless sea of bamboo undulating like waves under the moonlight, with the inky outlines of rolling mountains in the distance. A bright moon hangs high in the tranquil night sky, creating a vast and healing dreamlike atmosphere.", + "seed": 11047, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000101/group_13/index_1275_seed_11047_dp_26_rank_52_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 13, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.0567441332153976, + "vlm_logprob": -0.0567441332153976, + "vlm_yes_no_margin": 3.65, + "vlm_token_credit": -0.0567441332153976, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a bamboo forest at night with a bright moon in the sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a healer with long silver hair tied by a wooden hairpin and wearing a moon-white linen robe visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the healer kneel on mossy ground and apply a herbal solution from a ceramic bowl to a fawn's foreleg wound?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the video feature emerald green bamboo swaying, dappled shadows, and flickering fireflies?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start with a medium shot and steadily tilt upward to reveal a vast landscape of bamboo and mountains?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01618555188179016 + ], + "opposite_logprob_tokens": [ + -4.391185760498047 + ], + "logprob_reward": -0.01618555188179016, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.01618555188179016, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0720386728644371, + 0.02497466281056404, + 0.040569908916950226, + 0.5021317005157471, + 0.019639959558844566, + 0.09522192925214767, + 0.04397162050008774, + 0.024034233763813972, + 0.10796479880809784, + 0.06945261359214783 + ], + "credit_map": [ + 0.010730392299592495, + 0.014187253080308437, + 0.014702065847814083, + 0.006313307676464319, + 0.006073204334825277, + 0.007747670169919729, + 0.004936473909765482, + 0.007348305080085993, + 0.004126223735511303, + 0.0042326427064836025, + 0.004816538654267788, + 0.002000583801418543, + 0.0020879493094980717, + 0.0029382884968072176, + 0.0014165467582643032, + 0.003355889581143856, + 0.005698388442397118, + 0.004802255891263485, + 0.005350382532924414, + 0.004707623273134232, + 0.0035345209762454033, + 0.007396620232611895, + 0.0037525552324950695, + 0.00532755721360445, + 0.0359206348657608, + 0.1104009822010994, + 0.20364564657211304, + 0.020489023998379707, + 0.01981850527226925, + 0.020771101117134094, + 0.009455367922782898, + 0.08163038641214371, + 0.0019077409524470568, + 0.002132737310603261, + 0.003357664216309786, + 0.004319704603403807, + 0.0019003330962732434, + 0.0018268536077812314, + 0.0017456964123994112, + 0.002449226099997759, + 0.011160679161548615, + 0.02534915693104267, + 0.009453297592699528, + 0.007908510975539684, + 0.0067457472905516624, + 0.007516002282500267, + 0.004890819080173969, + 0.02219770848751068, + 0.008600213564932346, + 0.005670940037816763, + 0.005439490545541048, + 0.00881961826235056, + 0.0031291197519749403, + 0.003821679623797536, + 0.003376537701115012, + 0.005114017985761166, + 0.0021721133962273598, + 0.003350459737703204, + 0.003719286061823368, + 0.0034197401255369186, + 0.0031353470403701067, + 0.0027104122564196587, + 0.0026504595298320055, + 0.00287641491740942, + 0.007197611033916473, + 0.012139357626438141, + 0.009079712443053722, + 0.02002309262752533, + 0.02118927612900734, + 0.012067742645740509, + 0.014629959128797054, + 0.011638046242296696, + 0.008475535549223423, + 0.009528403170406818, + 0.008149450644850731, + 0.006694378796964884, + 0.010202852077782154, + 0.011671530082821846, + 0.007756821811199188, + 0.006973633076995611 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 8, + 5 + ], + "selected_reward": -0.01618555188179016 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a bamboo forest at night with a bright moon in the sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a healer with long silver hair tied by a wooden hairpin and wearing a moon-white linen robe visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the healer kneel on mossy ground and apply a herbal solution from a ceramic bowl to a fawn's foreleg wound?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the video feature emerald green bamboo swaying, dappled shadows, and flickering fireflies?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start with a medium shot and steadily tilt upward to reveal a vast landscape of bamboo and mountains?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004646813031286001 + ], + "opposite_logprob_tokens": [ + -5.3796467781066895 + ], + "logprob_reward": -0.004646813031286001, + "yes_no_margin_reward": 5.375, + "token_credit_reward": -0.004646813031286001, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09396073967218399, + 0.04531119763851166, + 0.031456924974918365, + 0.48668286204338074, + 0.03110937401652336, + 0.06111470237374306, + 0.0645478367805481, + 0.08658035844564438, + 0.050137076526880264, + 0.0490989051759243 + ], + "credit_map": [ + 0.013498682528734207, + 0.013682442717254162, + 0.014616262167692184, + 0.009245230816304684, + 0.012771357782185078, + 0.008283768780529499, + 0.009513208642601967, + 0.012349783442914486, + 0.006460373755544424, + 0.005312816705554724, + 0.0071042669005692005, + 0.0047034090384840965, + 0.004886054899543524, + 0.0076189725659787655, + 0.0035111261531710625, + 0.005714179016649723, + 0.004844943061470985, + 0.00314129120670259, + 0.005283275153487921, + 0.0044513666070997715, + 0.0034765598829835653, + 0.002952667884528637, + 0.002912236610427499, + 0.004394584335386753, + 0.034720927476882935, + 0.10703382641077042, + 0.19714482128620148, + 0.02022498846054077, + 0.01938716135919094, + 0.02032500132918358, + 0.00904255360364914, + 0.07880360633134842, + 0.002746892860159278, + 0.003075114218518138, + 0.004695729352533817, + 0.003546125255525112, + 0.004778007976710796, + 0.002927748952060938, + 0.004048321861773729, + 0.00529143400490284, + 0.006005443166941404, + 0.012601891532540321, + 0.007348319049924612, + 0.009335053153336048, + 0.006524765398353338, + 0.005347623489797115, + 0.006957062985748053, + 0.006994541734457016, + 0.005542587488889694, + 0.006960808299481869, + 0.009041565470397472, + 0.011033239774405956, + 0.013226871378719807, + 0.004204974975436926, + 0.009948237799108028, + 0.004589550197124481, + 0.006809602957218885, + 0.014171821065247059, + 0.005964548792690039, + 0.01223917305469513, + 0.01968587003648281, + 0.006442783400416374, + 0.014902880415320396, + 0.006363682448863983, + 0.0032042632810771465, + 0.007845154963433743, + 0.004924951586872339, + 0.00648911576718092, + 0.009161562658846378, + 0.007459812331944704, + 0.006337007973343134, + 0.004715209361165762, + 0.007377996109426022, + 0.008734527975320816, + 0.0043660919182002544, + 0.0043056681752204895, + 0.005808716639876366, + 0.008666560985147953, + 0.005358642898499966, + 0.004480702802538872 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 0, + 7 + ], + "selected_reward": -0.004646813031286001 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a bamboo forest at night with a bright moon in the sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a healer with long silver hair tied by a wooden hairpin and wearing a moon-white linen robe visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the healer kneel on mossy ground and apply a herbal solution from a ceramic bowl to a fawn's foreleg wound?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the video feature emerald green bamboo swaying, dappled shadows, and flickering fireflies?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start with a medium shot and steadily tilt upward to reveal a vast landscape of bamboo and mountains?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.038099367171525955 + ], + "opposite_logprob_tokens": [ + -3.2880992889404297 + ], + "logprob_reward": -0.038099367171525955, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.038099367171525955, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11528494954109192, + 0.07402394711971283, + 0.07215359061956406, + 0.1814463585615158, + 0.047586191445589066, + 0.0719861164689064, + 0.05427515134215355, + 0.2140907198190689, + 0.09812665730714798, + 0.0710262581706047 + ], + "credit_map": [ + 0.01741347834467888, + 0.016303202137351036, + 0.01950555108487606, + 0.010617966763675213, + 0.015987053513526917, + 0.011238730512559414, + 0.00939385499805212, + 0.014825098216533661, + 0.013287384063005447, + 0.009938733652234077, + 0.012476466596126556, + 0.007215206511318684, + 0.007416511885821819, + 0.008870357647538185, + 0.005479053594172001, + 0.00934023316949606, + 0.011200297623872757, + 0.009002072736620903, + 0.01201280951499939, + 0.00886395014822483, + 0.009702332317829132, + 0.006744681391865015, + 0.00748758390545845, + 0.007139865774661303, + 0.014000954106450081, + 0.04196345806121826, + 0.06605328619480133, + 0.013758954592049122, + 0.008572686463594437, + 0.009493418969213963, + 0.006472524255514145, + 0.021131083369255066, + 0.004604860674589872, + 0.005002355668693781, + 0.004632802680134773, + 0.006008293945342302, + 0.010004235431551933, + 0.004783253651112318, + 0.008179754950106144, + 0.004370634909719229, + 0.007164419628679752, + 0.019126029685139656, + 0.0066239843145012856, + 0.005987764801830053, + 0.005414722021669149, + 0.005122043192386627, + 0.005022143479436636, + 0.01752501167356968, + 0.008174169808626175, + 0.00714335311204195, + 0.010824272409081459, + 0.009147153235971928, + 0.0045560263097286224, + 0.004063835367560387, + 0.004296964965760708, + 0.00606937613338232, + 0.01160766277462244, + 0.03431912511587143, + 0.011565093882381916, + 0.030891045928001404, + 0.04982062056660652, + 0.013725925236940384, + 0.04436207562685013, + 0.017799172550439835, + 0.0071156444028019905, + 0.01508097443729639, + 0.009301816113293171, + 0.014574509114027023, + 0.016460616141557693, + 0.014593875035643578, + 0.011335078626871109, + 0.009664141573011875, + 0.01037343218922615, + 0.010202115401625633, + 0.008871586062014103, + 0.007424689829349518, + 0.006699846591800451, + 0.01062654610723257, + 0.00765974260866642, + 0.009168297983705997 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 3, + 0 + ], + "selected_reward": -0.038099367171525955 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a bamboo forest at night with a bright moon in the sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a healer with long silver hair tied by a wooden hairpin and wearing a moon-white linen robe visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the healer kneel on mossy ground and apply a herbal solution from a ceramic bowl to a fawn's foreleg wound?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the video feature emerald green bamboo swaying, dappled shadows, and flickering fireflies?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start with a medium shot and steadily tilt upward to reveal a vast landscape of bamboo and mountains?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02327677607536316 + ], + "opposite_logprob_tokens": [ + -3.7732768058776855 + ], + "logprob_reward": -0.02327677607536316, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.02327677607536316, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12997935712337494, + 0.042430102825164795, + 0.0665738508105278, + 0.2389759123325348, + 0.04049218073487282, + 0.07945562154054642, + 0.06013336032629013, + 0.16085435450077057, + 0.09251310676336288, + 0.08859215676784515 + ], + "credit_map": [ + 0.022223057225346565, + 0.02247820980846882, + 0.02747044898569584, + 0.013961650431156158, + 0.010435725562274456, + 0.012250839732587337, + 0.009285266511142254, + 0.01187415886670351, + 0.005396044347435236, + 0.006352559197694063, + 0.005701703950762749, + 0.00281207705847919, + 0.0057542999275028706, + 0.0059095872566103935, + 0.004495547618716955, + 0.006008281372487545, + 0.012419088743627071, + 0.008214117959141731, + 0.011630178429186344, + 0.008574586361646652, + 0.005536653101444244, + 0.007614692207425833, + 0.006181260105222464, + 0.006403280422091484, + 0.01713268831372261, + 0.05193234980106354, + 0.09486222267150879, + 0.0121178412809968, + 0.009531581774353981, + 0.010016607120633125, + 0.0056063337251544, + 0.03777630999684334, + 0.003351928200572729, + 0.004770026542246342, + 0.0058828662149608135, + 0.005889471620321274, + 0.006735638715326786, + 0.003692639758810401, + 0.005744467489421368, + 0.00442514568567276, + 0.008604883216321468, + 0.017724718898534775, + 0.00881459005177021, + 0.012828506529331207, + 0.006436964962631464, + 0.007294122129678726, + 0.005611338187009096, + 0.012140505947172642, + 0.011484852060675621, + 0.007142027374356985, + 0.007070961873978376, + 0.010373644530773163, + 0.006326712667942047, + 0.0049957334995269775, + 0.005752860568463802, + 0.006986575201153755, + 0.012314820662140846, + 0.02740348130464554, + 0.007771860342472792, + 0.023625347763299942, + 0.03694648668169975, + 0.010888302698731422, + 0.029402902349829674, + 0.012501165270805359, + 0.00524020753800869, + 0.013030195608735085, + 0.0066019874066114426, + 0.013743013143539429, + 0.022640898823738098, + 0.008714808151125908, + 0.013139252550899982, + 0.009402749128639698, + 0.008986360393464565, + 0.012886248528957367, + 0.007708176039159298, + 0.008936921134591103, + 0.017306990921497345, + 0.010645028203725815, + 0.014806408435106277, + 0.0073160273022949696 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 7, + 0 + ], + "selected_reward": -0.02327677607536316 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a bamboo forest at night with a bright moon in the sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a healer with long silver hair tied by a wooden hairpin and wearing a moon-white linen robe visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the healer kneel on mossy ground and apply a herbal solution from a ceramic bowl to a fawn's foreleg wound?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the video feature emerald green bamboo swaying, dappled shadows, and flickering fireflies?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start with a medium shot and steadily tilt upward to reveal a vast landscape of bamboo and mountains?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.2015121579170227 + ], + "opposite_logprob_tokens": [ + -1.701512098312378 + ], + "logprob_reward": -0.2015121579170227, + "yes_no_margin_reward": 1.5, + "token_credit_reward": -0.2015121579170227, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.04584223031997681, + 0.025387383997440338, + 0.023342283442616463, + 0.5108674168586731, + 0.02558673359453678, + 0.0462317168712616, + 0.04612239450216293, + 0.11037054657936096, + 0.13204431533813477, + 0.034205030649900436 + ], + "credit_map": [ + 0.007846195250749588, + 0.006888872943818569, + 0.007634717505425215, + 0.0046949139796197414, + 0.005285685881972313, + 0.0054320176132023335, + 0.002939172089099884, + 0.005120652262121439, + 0.0038831892888993025, + 0.003206169232726097, + 0.003722380381077528, + 0.0037276851944625378, + 0.0033357818610966206, + 0.0031072639394551516, + 0.002406660234555602, + 0.001998254330828786, + 0.002586584771052003, + 0.002540411427617073, + 0.0025103946682065725, + 0.003378985682502389, + 0.0028671352192759514, + 0.003553347662091255, + 0.002814149484038353, + 0.003091274993494153, + 0.03663810342550278, + 0.11099640280008316, + 0.2075076699256897, + 0.021281886845827103, + 0.019837986677885056, + 0.02067478932440281, + 0.009525629691779613, + 0.08440492302179337, + 0.0022685821168124676, + 0.0027294119354337454, + 0.0030928640626370907, + 0.004179288633167744, + 0.00482841907069087, + 0.0017932213377207518, + 0.004222731105983257, + 0.0024722155649214983, + 0.0034909674432128668, + 0.012621372938156128, + 0.0038083037361502647, + 0.005008684005588293, + 0.004142178688198328, + 0.00333798723295331, + 0.0036743751261383295, + 0.010147844441235065, + 0.009414315223693848, + 0.005566880106925964, + 0.007411415688693523, + 0.006238909903913736, + 0.004730151500552893, + 0.0034948450047522783, + 0.004100212827324867, + 0.005165663547813892, + 0.006053101737052202, + 0.018469465896487236, + 0.0067627690732479095, + 0.01633356139063835, + 0.023557931184768677, + 0.008658968843519688, + 0.02132873423397541, + 0.009206020273268223, + 0.009043511934578419, + 0.02072801999747753, + 0.010629008524119854, + 0.01945655047893524, + 0.028351591899991035, + 0.013737933710217476, + 0.01838911883533001, + 0.011708583682775497, + 0.004001948516815901, + 0.004556011408567429, + 0.00394679419696331, + 0.003550080582499504, + 0.0054969871416687965, + 0.003564311657100916, + 0.004995670635253191, + 0.00409322464838624 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 8, + 7 + ], + "selected_reward": -0.2015121579170227 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "The film begins with a gentle medium shot. Deep in a tranquil bamboo forest, bathed in cool, silver moonlight, stands a healer of sagelike demeanor. His long, moon-like silver hair is loosely tied with a simple wooden hairpin. He wears a flowing, moon-white robe made of natural linen, its sleeves sweeping gently in the night breeze. With exquisite grace, he kneels on the soft, mossy ground. Before him, a fawn, its fur speckled with white, appears exceptionally fragile. Its large, moist eyes are filled with utter trust. The healer carefully tends to a small wound on its foreleg. With his slender fingertips, he scoops up a herbal solution from a small ceramic bowl beside him and gently, calmly, applies it to the wound. His focused and compassionate expression appears remarkably sacred in the soft light. Towering, emerald green bamboos sway gently in the night breeze. Moonlight filters through the leaves, casting dappled, flickering shadows on the ground. A few fireflies flicker in the air, adding a touch of dreamy vitality to the tranquil scene. The entire scene is imbued with the distinctive hand-painted aesthetic of Ghibli animation, with soft colors and flowing lines, imbued with a sense of tranquility and tenderness. As the camera steadily and slowly tilts upward, the intimate interaction between the healer and the deer gradually gives way to a broader landscape, ultimately revealing a grand canvas: a boundless sea of bamboo undulating like waves under the moonlight, with the inky outlines of rolling mountains in the distance. A bright moon hangs high in the tranquil night sky, creating a vast and healing dreamlike atmosphere.", + "seed": 11048, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000101/group_13/index_1275_seed_11048_dp_27_rank_54_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 13, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.09295876072719693, + "vlm_logprob": -0.09295876072719693, + "vlm_yes_no_margin": 3.475, + "vlm_token_credit": -0.09295876072719693, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a bamboo forest at night with a bright moon in the sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a healer with long silver hair tied by a wooden hairpin and wearing a moon-white linen robe visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the healer kneel on mossy ground and apply a herbal solution from a ceramic bowl to a fawn's foreleg wound?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the video feature emerald green bamboo swaying, dappled shadows, and flickering fireflies?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start with a medium shot and steadily tilt upward to reveal a vast landscape of bamboo and mountains?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01555406954139471 + ], + "opposite_logprob_tokens": [ + -4.265553951263428 + ], + "logprob_reward": -0.01555406954139471, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.01555406954139471, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2520548701286316, + 0.1388583481311798, + 0.11113552004098892, + 0.10279478877782822, + 0.07721719890832901, + 0.07606663554906845, + 0.0368499681353569, + 0.046448443084955215, + 0.07710294425487518, + 0.08147125691175461 + ], + "credit_map": [ + 0.028506046161055565, + 0.03519278019666672, + 0.028434736654162407, + 0.03593572974205017, + 0.03464420884847641, + 0.031114788725972176, + 0.031208302825689316, + 0.02701828069984913, + 0.01149960421025753, + 0.01853642426431179, + 0.016223540529608727, + 0.028939979150891304, + 0.017822053283452988, + 0.008635174483060837, + 0.021747449412941933, + 0.015454129315912724, + 0.012117061764001846, + 0.020559055730700493, + 0.011211574077606201, + 0.019947564229369164, + 0.010539468377828598, + 0.007699858397245407, + 0.016870079562067986, + 0.012190851382911205, + 0.010217170231044292, + 0.01563892513513565, + 0.009249716997146606, + 0.014136040583252907, + 0.015583721920847893, + 0.006703641265630722, + 0.019927801564335823, + 0.0113377645611763, + 0.006367092952132225, + 0.010259789414703846, + 0.010317039676010609, + 0.015661075711250305, + 0.010950476862490177, + 0.006491785869002342, + 0.010948363691568375, + 0.006221575662493706, + 0.006815968547016382, + 0.01847788691520691, + 0.006478979717940092, + 0.015150590799748898, + 0.008374074473977089, + 0.005936449859291315, + 0.008535480126738548, + 0.006297209300100803, + 0.005794293712824583, + 0.005017672665417194, + 0.003947818651795387, + 0.0038576610386371613, + 0.005697245243936777, + 0.003536358242854476, + 0.004950609989464283, + 0.004048308823257685, + 0.0059981620870530605, + 0.006122926715761423, + 0.004160603042691946, + 0.008835443295538425, + 0.006139449309557676, + 0.005214662291109562, + 0.006649121642112732, + 0.0033280737698078156, + 0.010043920949101448, + 0.007474288810044527, + 0.011620963923633099, + 0.011185809969902039, + 0.007256940472871065, + 0.008537108078598976, + 0.011745561845600605, + 0.00923834927380085, + 0.01323707401752472, + 0.007070209830999374, + 0.011929799802601337, + 0.011231152340769768, + 0.0077859931625425816, + 0.013148539699614048, + 0.009599601849913597, + 0.0074688890017569065 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 2 + ], + "selected_reward": -0.01555406954139471 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a bamboo forest at night with a bright moon in the sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a healer with long silver hair tied by a wooden hairpin and wearing a moon-white linen robe visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the healer kneel on mossy ground and apply a herbal solution from a ceramic bowl to a fawn's foreleg wound?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the video feature emerald green bamboo swaying, dappled shadows, and flickering fireflies?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start with a medium shot and steadily tilt upward to reveal a vast landscape of bamboo and mountains?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005259012337774038 + ], + "opposite_logprob_tokens": [ + -5.255259037017822 + ], + "logprob_reward": -0.005259012337774038, + "yes_no_margin_reward": 5.25, + "token_credit_reward": -0.005259012337774038, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09605339169502258, + 0.10620101541280746, + 0.046930138021707535, + 0.11220551282167435, + 0.14409081637859344, + 0.048600342124700546, + 0.18542197346687317, + 0.08450473845005035, + 0.07613424956798553, + 0.09985781461000443 + ], + "credit_map": [ + 0.01136059407144785, + 0.013658450916409492, + 0.009329340420663357, + 0.015299621038138866, + 0.010488111525774002, + 0.010273940861225128, + 0.014714902266860008, + 0.010928433388471603, + 0.009312557987868786, + 0.013172035105526447, + 0.012470509856939316, + 0.020948659628629684, + 0.015904365107417107, + 0.0073207663372159, + 0.015912286937236786, + 0.011159829795360565, + 0.005072711501270533, + 0.00865393690764904, + 0.005880204029381275, + 0.008102226071059704, + 0.004894579295068979, + 0.003392496844753623, + 0.005872488021850586, + 0.0050614969804883, + 0.011221515946090221, + 0.016135815531015396, + 0.009560796432197094, + 0.016054287552833557, + 0.017804665490984917, + 0.008509796112775803, + 0.0211265180259943, + 0.011792118661105633, + 0.012251685373485088, + 0.019593331962823868, + 0.013433895073831081, + 0.028708435595035553, + 0.02439058944582939, + 0.009070167317986488, + 0.022172339260578156, + 0.014470383524894714, + 0.004461072385311127, + 0.0064589581452310085, + 0.006293643731623888, + 0.0075146970339119434, + 0.007135441992431879, + 0.003721883986145258, + 0.007985416799783707, + 0.005029224790632725, + 0.039009708911180496, + 0.03006107732653618, + 0.013382143341004848, + 0.03360345959663391, + 0.025607680901885033, + 0.019851528108119965, + 0.016237599775195122, + 0.007668779231607914, + 0.009833092801272869, + 0.01074684876948595, + 0.010159764438867569, + 0.009861528873443604, + 0.014318916015326977, + 0.007370790001004934, + 0.016315631568431854, + 0.005898160859942436, + 0.010779854841530323, + 0.008515555411577225, + 0.011236828751862049, + 0.00871431827545166, + 0.008915990591049194, + 0.007306517567485571, + 0.012950778007507324, + 0.00771440751850605, + 0.017166098579764366, + 0.007447436917573214, + 0.015992600470781326, + 0.009299300611019135, + 0.007673881947994232, + 0.017855709418654442, + 0.014927910640835762, + 0.009494880214333534 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 4, + 3 + ], + "selected_reward": -0.005259012337774038 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a bamboo forest at night with a bright moon in the sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a healer with long silver hair tied by a wooden hairpin and wearing a moon-white linen robe visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the healer kneel on mossy ground and apply a herbal solution from a ceramic bowl to a fawn's foreleg wound?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the video feature emerald green bamboo swaying, dappled shadows, and flickering fireflies?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start with a medium shot and steadily tilt upward to reveal a vast landscape of bamboo and mountains?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02328353188931942 + ], + "opposite_logprob_tokens": [ + -3.7732834815979004 + ], + "logprob_reward": -0.02328353188931942, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.02328353188931942, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15020832419395447, + 0.09287527948617935, + 0.08071985840797424, + 0.04617039114236832, + 0.10406635701656342, + 0.054572783410549164, + 0.05308343470096588, + 0.21452681720256805, + 0.09423230588436127, + 0.10954446345567703 + ], + "credit_map": [ + 0.012570386752486229, + 0.02046032063663006, + 0.010827647522091866, + 0.017244575545191765, + 0.02419445291161537, + 0.011219747364521027, + 0.03428340703248978, + 0.019407793879508972, + 0.007201795000582933, + 0.011720413342118263, + 0.009871208108961582, + 0.014354247599840164, + 0.013887081295251846, + 0.0077186403796076775, + 0.016504190862178802, + 0.011617710813879967, + 0.010388520546257496, + 0.02006003074347973, + 0.00795652437955141, + 0.013505811803042889, + 0.008225911296904087, + 0.005389696452766657, + 0.0093662915751338, + 0.0058270650915801525, + 0.004925454035401344, + 0.006198577582836151, + 0.004056781996041536, + 0.006790188606828451, + 0.006627256516367197, + 0.004482083022594452, + 0.007918868213891983, + 0.0051711807027459145, + 0.009631599299609661, + 0.012772753834724426, + 0.012073863297700882, + 0.019452206790447235, + 0.017044957727193832, + 0.006010132376104593, + 0.017044922336935997, + 0.010035916231572628, + 0.004292042460292578, + 0.009295793250203133, + 0.005593797191977501, + 0.01038284134119749, + 0.006763134151697159, + 0.0039853802882134914, + 0.009382939897477627, + 0.004876856692135334, + 0.006811692379415035, + 0.010709826834499836, + 0.004488730803132057, + 0.0069993650540709496, + 0.00767302792519331, + 0.006097577046602964, + 0.0056064617820084095, + 0.004696758463978767, + 0.02186940610408783, + 0.026678165420889854, + 0.02730730175971985, + 0.028258204460144043, + 0.035802438855171204, + 0.018569594249129295, + 0.04056045413017273, + 0.015481253154575825, + 0.011544310487806797, + 0.009135122410953045, + 0.0136407520622015, + 0.013141739182174206, + 0.010296027176082134, + 0.011146580800414085, + 0.017146367579698563, + 0.00818141084164381, + 0.018136844038963318, + 0.008968010544776917, + 0.01739499345421791, + 0.009395316243171692, + 0.009130444377660751, + 0.01937972940504551, + 0.017082253471016884, + 0.010056867264211178 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 0, + 9 + ], + "selected_reward": -0.02328353188931942 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a bamboo forest at night with a bright moon in the sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a healer with long silver hair tied by a wooden hairpin and wearing a moon-white linen robe visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the healer kneel on mossy ground and apply a herbal solution from a ceramic bowl to a fawn's foreleg wound?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the video feature emerald green bamboo swaying, dappled shadows, and flickering fireflies?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start with a medium shot and steadily tilt upward to reveal a vast landscape of bamboo and mountains?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.033692337572574615 + ], + "opposite_logprob_tokens": [ + -3.4086923599243164 + ], + "logprob_reward": -0.033692337572574615, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.033692337572574615, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.16846881806850433, + 0.07307007908821106, + 0.09922618418931961, + 0.07548238337039948, + 0.060861751437187195, + 0.10285531729459763, + 0.08219563961029053, + 0.08934585005044937, + 0.10497129708528519, + 0.14352262020111084 + ], + "credit_map": [ + 0.01965217851102352, + 0.020075391978025436, + 0.01976574771106243, + 0.02378469705581665, + 0.027006028220057487, + 0.021506424993276596, + 0.02103535644710064, + 0.01564301922917366, + 0.00642047543078661, + 0.009808032773435116, + 0.010188799351453781, + 0.01948685199022293, + 0.00800410471856594, + 0.005686517804861069, + 0.007111621554940939, + 0.0063636815175414085, + 0.01115915086120367, + 0.020601972937583923, + 0.010553460568189621, + 0.01799635961651802, + 0.011027161031961441, + 0.007768526207655668, + 0.011688887141644955, + 0.008430678397417068, + 0.0061721173115074635, + 0.008294089697301388, + 0.006935945712029934, + 0.012882153503596783, + 0.008572043851017952, + 0.006675299257040024, + 0.015551791526377201, + 0.01039894763380289, + 0.006713169626891613, + 0.007227139547467232, + 0.007222887594252825, + 0.011360693722963333, + 0.006753795780241489, + 0.005190533120185137, + 0.0098814582452178, + 0.006512081250548363, + 0.009231762029230595, + 0.01949000358581543, + 0.008781857788562775, + 0.016469020396471024, + 0.013496465981006622, + 0.009368130937218666, + 0.01456982921808958, + 0.011448260396718979, + 0.015306053683161736, + 0.010720279067754745, + 0.0076509080827236176, + 0.009832476265728474, + 0.011592083610594273, + 0.009414942003786564, + 0.009872810915112495, + 0.007806092966347933, + 0.00810242723673582, + 0.011777380481362343, + 0.008769809268414974, + 0.015861015766859055, + 0.012499233707785606, + 0.010530572384595871, + 0.01564975269138813, + 0.006155666895210743, + 0.012896094471216202, + 0.0074922931380569935, + 0.01394183561205864, + 0.01491085346788168, + 0.010045240633189678, + 0.018171770498156548, + 0.011492472141981125, + 0.016020746901631355, + 0.019932879135012627, + 0.016898436471819878, + 0.020168036222457886, + 0.017886832356452942, + 0.013368351384997368, + 0.016970597207546234, + 0.02555793523788452, + 0.012739571742713451 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 8 + ], + "selected_reward": -0.033692337572574615 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a bamboo forest at night with a bright moon in the sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a healer with long silver hair tied by a wooden hairpin and wearing a moon-white linen robe visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the healer kneel on mossy ground and apply a herbal solution from a ceramic bowl to a fawn's foreleg wound?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the video feature emerald green bamboo swaying, dappled shadows, and flickering fireflies?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start with a medium shot and steadily tilt upward to reveal a vast landscape of bamboo and mountains?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.3870048522949219 + ], + "opposite_logprob_tokens": [ + -1.1370048522949219 + ], + "logprob_reward": -0.3870048522949219, + "yes_no_margin_reward": 0.75, + "token_credit_reward": -0.3870048522949219, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10954172164201736, + 0.09815379977226257, + 0.11615894734859467, + 0.05155221000313759, + 0.11530022323131561, + 0.09115368872880936, + 0.07317569851875305, + 0.05240058898925781, + 0.19288505613803864, + 0.09967799484729767 + ], + "credit_map": [ + 0.010595838539302349, + 0.01608811877667904, + 0.007437351159751415, + 0.012296191416680813, + 0.022741077467799187, + 0.011770700104534626, + 0.016285311430692673, + 0.01232713833451271, + 0.009153773076832294, + 0.012851264327764511, + 0.011166960000991821, + 0.014754164032638073, + 0.01668737269937992, + 0.010163909755647182, + 0.012506085447967052, + 0.010870276018977165, + 0.011608806438744068, + 0.019487449899315834, + 0.013633288443088531, + 0.019213275983929634, + 0.015254859812557697, + 0.008383315056562424, + 0.016140110790729523, + 0.012437845580279827, + 0.0062095061875879765, + 0.007358742877840996, + 0.00515639316290617, + 0.007312352303415537, + 0.006773349829018116, + 0.005968607030808926, + 0.007263549137860537, + 0.005509712267667055, + 0.009140660054981709, + 0.015136348083615303, + 0.01347286719828844, + 0.02570345625281334, + 0.015408186241984367, + 0.009389151819050312, + 0.0156923308968544, + 0.01135722454637289, + 0.008472089655697346, + 0.013185313902795315, + 0.012857634574174881, + 0.011714481748640537, + 0.011154607869684696, + 0.009700796566903591, + 0.014388514682650566, + 0.009680253453552723, + 0.008030155673623085, + 0.012321241199970245, + 0.007598137483000755, + 0.012003283016383648, + 0.008556688204407692, + 0.007867090404033661, + 0.008803857490420341, + 0.007995248772203922, + 0.006129970774054527, + 0.008819455280900002, + 0.0041623180732131, + 0.006597032770514488, + 0.009288606233894825, + 0.00525777880102396, + 0.007336013950407505, + 0.004809418227523565, + 0.03592244163155556, + 0.014171533286571503, + 0.024346744641661644, + 0.021458294242620468, + 0.017214389517903328, + 0.025062251836061478, + 0.021080337464809418, + 0.03362906351685524, + 0.012669784016907215, + 0.016351301223039627, + 0.01368642970919609, + 0.017505479976534843, + 0.011245716363191605, + 0.011915909126400948, + 0.010900582186877728, + 0.005402787122875452 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 2, + 4 + ], + "selected_reward": -0.3870048522949219 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "The film begins with a gentle medium shot. Deep in a tranquil bamboo forest, bathed in cool, silver moonlight, stands a healer of sagelike demeanor. His long, moon-like silver hair is loosely tied with a simple wooden hairpin. He wears a flowing, moon-white robe made of natural linen, its sleeves sweeping gently in the night breeze. With exquisite grace, he kneels on the soft, mossy ground. Before him, a fawn, its fur speckled with white, appears exceptionally fragile. Its large, moist eyes are filled with utter trust. The healer carefully tends to a small wound on its foreleg. With his slender fingertips, he scoops up a herbal solution from a small ceramic bowl beside him and gently, calmly, applies it to the wound. His focused and compassionate expression appears remarkably sacred in the soft light. Towering, emerald green bamboos sway gently in the night breeze. Moonlight filters through the leaves, casting dappled, flickering shadows on the ground. A few fireflies flicker in the air, adding a touch of dreamy vitality to the tranquil scene. The entire scene is imbued with the distinctive hand-painted aesthetic of Ghibli animation, with soft colors and flowing lines, imbued with a sense of tranquility and tenderness. As the camera steadily and slowly tilts upward, the intimate interaction between the healer and the deer gradually gives way to a broader landscape, ultimately revealing a grand canvas: a boundless sea of bamboo undulating like waves under the moonlight, with the inky outlines of rolling mountains in the distance. A bright moon hangs high in the tranquil night sky, creating a vast and healing dreamlike atmosphere.", + "seed": 11049, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000101/group_13/index_1275_seed_11049_dp_27_rank_54_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 13, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.09366043964400887, + "vlm_logprob": -0.09366043964400887, + "vlm_yes_no_margin": 3.375, + "vlm_token_credit": -0.09366043964400887, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a bamboo forest at night with a bright moon in the sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a healer with long silver hair tied by a wooden hairpin and wearing a moon-white linen robe visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the healer kneel on mossy ground and apply a herbal solution from a ceramic bowl to a fawn's foreleg wound?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the video feature emerald green bamboo swaying, dappled shadows, and flickering fireflies?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start with a medium shot and steadily tilt upward to reveal a vast landscape of bamboo and mountains?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.019384237006306648 + ], + "opposite_logprob_tokens": [ + -4.019384384155273 + ], + "logprob_reward": -0.019384237006306648, + "yes_no_margin_reward": 4.0, + "token_credit_reward": -0.019384237006306648, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06410515308380127, + 0.02355002611875534, + 0.04687746986746788, + 0.03974509239196777, + 0.22264963388442993, + 0.04517277702689171, + 0.21737819910049438, + 0.06342680752277374, + 0.0986143946647644, + 0.17848052084445953 + ], + "credit_map": [ + 0.009479847736656666, + 0.00514577142894268, + 0.008582375943660736, + 0.01124770101159811, + 0.0063251713290810585, + 0.008088375441730022, + 0.006221975665539503, + 0.009013926610350609, + 0.003256283700466156, + 0.0029568998143076897, + 0.0017298756865784526, + 0.002683935221284628, + 0.0037841673474758863, + 0.0029032446909695864, + 0.002976966556161642, + 0.0032586499582976103, + 0.0075186169706285, + 0.004782746080309153, + 0.00714058568701148, + 0.008342101238667965, + 0.0038490232545882463, + 0.006091693881899118, + 0.0043643140234053135, + 0.004788386169821024, + 0.005234795156866312, + 0.0031300082337111235, + 0.006233705673366785, + 0.007036231458187103, + 0.005282032769173384, + 0.0042753019370138645, + 0.004943180363625288, + 0.0036098335403949022, + 0.03515757620334625, + 0.01982579566538334, + 0.036310482770204544, + 0.021412035450339317, + 0.020187625661492348, + 0.04075246676802635, + 0.016744835302233696, + 0.03225880116224289, + 0.006389313377439976, + 0.004487541038542986, + 0.006683962419629097, + 0.004313184414058924, + 0.006765696685761213, + 0.007269877940416336, + 0.0046628788113594055, + 0.004600317683070898, + 0.03711645305156708, + 0.012516272254288197, + 0.034425899386405945, + 0.014884098432958126, + 0.013266785070300102, + 0.07461746782064438, + 0.014784432016313076, + 0.015766775235533714, + 0.011113454587757587, + 0.004934858996421099, + 0.010811960324645042, + 0.0037289028987288475, + 0.005703897215425968, + 0.018305428326129913, + 0.003173881908878684, + 0.00565441744402051, + 0.01940799318253994, + 0.00714820995926857, + 0.023819370195269585, + 0.007599078584462404, + 0.006936341989785433, + 0.014565291814506054, + 0.012206189334392548, + 0.006931910756975412, + 0.028098516166210175, + 0.01603701338171959, + 0.038377441465854645, + 0.01577160507440567, + 0.01783691719174385, + 0.03005279414355755, + 0.015275698155164719, + 0.017030514776706696 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 6, + 9 + ], + "selected_reward": -0.019384237006306648 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a bamboo forest at night with a bright moon in the sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a healer with long silver hair tied by a wooden hairpin and wearing a moon-white linen robe visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the healer kneel on mossy ground and apply a herbal solution from a ceramic bowl to a fawn's foreleg wound?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the video feature emerald green bamboo swaying, dappled shadows, and flickering fireflies?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start with a medium shot and steadily tilt upward to reveal a vast landscape of bamboo and mountains?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.007634746376425028 + ], + "opposite_logprob_tokens": [ + -4.88263463973999 + ], + "logprob_reward": -0.007634746376425028, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.007634746376425028, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1455836147069931, + 0.0628410205245018, + 0.09567800909280777, + 0.12073768675327301, + 0.10765153914690018, + 0.060751982033252716, + 0.23315580189228058, + 0.0349649116396904, + 0.0693037286400795, + 0.06933175027370453 + ], + "credit_map": [ + 0.01862187311053276, + 0.010844205506145954, + 0.01702827401459217, + 0.024542706087231636, + 0.017270199954509735, + 0.017523687332868576, + 0.017537331208586693, + 0.02221531793475151, + 0.008670288138091564, + 0.0045363521203398705, + 0.005884310696274042, + 0.015220606699585915, + 0.006729073356837034, + 0.008183151483535767, + 0.006244235206395388, + 0.0073729935102164745, + 0.012586166150867939, + 0.009720510803163052, + 0.013963727280497551, + 0.011033810675144196, + 0.014043401926755905, + 0.010900797322392464, + 0.012882836163043976, + 0.010546745732426643, + 0.020467642694711685, + 0.00716383708640933, + 0.014594201929867268, + 0.035163719207048416, + 0.006982059683650732, + 0.016782665625214577, + 0.010062433779239655, + 0.009521112777292728, + 0.015183739364147186, + 0.009892542846500874, + 0.019888747483491898, + 0.011655827052891254, + 0.010156775824725628, + 0.016643695533275604, + 0.00836631003767252, + 0.015863897278904915, + 0.01059622224420309, + 0.005628714803606272, + 0.00849104579538107, + 0.007402926217764616, + 0.01012603472918272, + 0.006811986677348614, + 0.006301403045654297, + 0.005393647588789463, + 0.03100801259279251, + 0.013405876234173775, + 0.03742361068725586, + 0.01762138307094574, + 0.019690241664648056, + 0.0783659964799881, + 0.015306316316127777, + 0.020334342494606972, + 0.00558179197832942, + 0.0034649979788810015, + 0.004597912076860666, + 0.0029513095505535603, + 0.00323012238368392, + 0.005508395377546549, + 0.004973543342202902, + 0.004656835459172726, + 0.013111108914017677, + 0.0052301776595413685, + 0.017940759658813477, + 0.0057165478356182575, + 0.0048256004229187965, + 0.010555428452789783, + 0.005952466279268265, + 0.005971631966531277, + 0.009242800995707512, + 0.0070055709220469, + 0.011586720123887062, + 0.0068185641430318356, + 0.009573379531502724, + 0.010970066301524639, + 0.006343410350382328, + 0.0077912285923957825 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 0, + 3 + ], + "selected_reward": -0.007634746376425028 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a bamboo forest at night with a bright moon in the sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a healer with long silver hair tied by a wooden hairpin and wearing a moon-white linen robe visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the healer kneel on mossy ground and apply a herbal solution from a ceramic bowl to a fawn's foreleg wound?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the video feature emerald green bamboo swaying, dappled shadows, and flickering fireflies?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start with a medium shot and steadily tilt upward to reveal a vast landscape of bamboo and mountains?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03369844704866409 + ], + "opposite_logprob_tokens": [ + -3.408698558807373 + ], + "logprob_reward": -0.03369844704866409, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.03369844704866409, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.20961090922355652, + 0.06079639121890068, + 0.08424914628267288, + 0.13627871870994568, + 0.08035948127508163, + 0.06879301369190216, + 0.1300308108329773, + 0.07057985663414001, + 0.07120422273874283, + 0.0880974680185318 + ], + "credit_map": [ + 0.031230781227350235, + 0.016114190220832825, + 0.02780868113040924, + 0.04233177378773689, + 0.016052426770329475, + 0.026620149612426758, + 0.020508239045739174, + 0.028944678604602814, + 0.009145358577370644, + 0.004251756239682436, + 0.006535680033266544, + 0.015334933064877987, + 0.007529237307608128, + 0.007233139127492905, + 0.004898149985820055, + 0.005868136417120695, + 0.012497897259891033, + 0.008233350701630116, + 0.01467953808605671, + 0.011748182587325573, + 0.009996430017054081, + 0.009611289016902447, + 0.009021163918077946, + 0.008461295627057552, + 0.023352745920419693, + 0.00841565802693367, + 0.01732482574880123, + 0.043028928339481354, + 0.006625699810683727, + 0.01701931655406952, + 0.010970300994813442, + 0.00954124704003334, + 0.011874253861606121, + 0.006009465083479881, + 0.0061273351311683655, + 0.012205061502754688, + 0.009056362323462963, + 0.016423949971795082, + 0.009608346037566662, + 0.009054705500602722, + 0.009263032115995884, + 0.0070914411917328835, + 0.010210826061666012, + 0.007707135286182165, + 0.012877373956143856, + 0.006729837507009506, + 0.006566859781742096, + 0.008346508257091045, + 0.022528082132339478, + 0.009565919637680054, + 0.019113535061478615, + 0.0110397320240736, + 0.00883312989026308, + 0.038475289940834045, + 0.011111916042864323, + 0.009363201446831226, + 0.01027756743133068, + 0.0064311851747334, + 0.010454824194312096, + 0.006306326482445002, + 0.010210593231022358, + 0.009899329394102097, + 0.006696303375065327, + 0.010303720831871033, + 0.008559615351259708, + 0.006866251118481159, + 0.009246650151908398, + 0.010037682019174099, + 0.010134439915418625, + 0.008635886013507843, + 0.010929633863270283, + 0.006794061977416277, + 0.006297735031694174, + 0.008060745894908905, + 0.009178943932056427, + 0.009333204478025436, + 0.009931981563568115, + 0.015291355550289154, + 0.022247284650802612, + 0.0077562229707837105 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 6 + ], + "selected_reward": -0.03369844704866409 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a bamboo forest at night with a bright moon in the sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a healer with long silver hair tied by a wooden hairpin and wearing a moon-white linen robe visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the healer kneel on mossy ground and apply a herbal solution from a ceramic bowl to a fawn's foreleg wound?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the video feature emerald green bamboo swaying, dappled shadows, and flickering fireflies?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start with a medium shot and steadily tilt upward to reveal a vast landscape of bamboo and mountains?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02057693526148796 + ], + "opposite_logprob_tokens": [ + -3.8955769538879395 + ], + "logprob_reward": -0.02057693526148796, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.02057693526148796, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09412822872400284, + 0.05671651288866997, + 0.06739112734794617, + 0.049282923340797424, + 0.27000510692596436, + 0.0654248371720314, + 0.15557581186294556, + 0.04628497734665871, + 0.10261644423007965, + 0.09257406741380692 + ], + "credit_map": [ + 0.012838458642363548, + 0.006353397853672504, + 0.014210791327059269, + 0.018914951011538506, + 0.009507114067673683, + 0.010677119717001915, + 0.008336766622960567, + 0.013289626687765121, + 0.010688680224120617, + 0.005583863705396652, + 0.009099991992115974, + 0.006051241420209408, + 0.006326787639409304, + 0.008480978198349476, + 0.004994620103389025, + 0.005490349140018225, + 0.011574086733162403, + 0.007701835595071316, + 0.009359964169561863, + 0.012709809467196465, + 0.004597509279847145, + 0.009417342022061348, + 0.005354498978704214, + 0.006676084361970425, + 0.008617258630692959, + 0.003477903315797448, + 0.005933995358645916, + 0.00782052706927061, + 0.005284860730171204, + 0.008093687705695629, + 0.0053093209862709045, + 0.00474536744877696, + 0.03974420577287674, + 0.027127429842948914, + 0.04492569714784622, + 0.02585853450000286, + 0.02537279948592186, + 0.048567064106464386, + 0.020740089938044548, + 0.037669289857149124, + 0.010761580429971218, + 0.007658082526177168, + 0.009212239645421505, + 0.006897244602441788, + 0.010852377861738205, + 0.006849664729088545, + 0.006381137762218714, + 0.006812509149312973, + 0.020108310505747795, + 0.010476472787559032, + 0.026860004290938377, + 0.013695836998522282, + 0.010402738116681576, + 0.05027489736676216, + 0.01236767414957285, + 0.011389881372451782, + 0.009906807914376259, + 0.005990079138427973, + 0.005798062775284052, + 0.0034912775736302137, + 0.006389041896909475, + 0.006776854395866394, + 0.004209441598504782, + 0.0037234125193208456, + 0.01853826455771923, + 0.009098326787352562, + 0.02253681793808937, + 0.0072062816470861435, + 0.007361010182648897, + 0.014794839546084404, + 0.014937499538064003, + 0.00814339704811573, + 0.01127166859805584, + 0.015340523794293404, + 0.013890925794839859, + 0.009648828767240047, + 0.01143850851804018, + 0.010586591437458992, + 0.011343063786625862, + 0.009053953923285007 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 6, + 8 + ], + "selected_reward": -0.02057693526148796 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a bamboo forest at night with a bright moon in the sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a healer with long silver hair tied by a wooden hairpin and wearing a moon-white linen robe visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the healer kneel on mossy ground and apply a herbal solution from a ceramic bowl to a fawn's foreleg wound?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the video feature emerald green bamboo swaying, dappled shadows, and flickering fireflies?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start with a medium shot and steadily tilt upward to reveal a vast landscape of bamboo and mountains?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.38700783252716064 + ], + "opposite_logprob_tokens": [ + -1.1370078325271606 + ], + "logprob_reward": -0.38700783252716064, + "yes_no_margin_reward": 0.75, + "token_credit_reward": -0.38700783252716064, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.055924925953149796, + 0.044710513204336166, + 0.03609755262732506, + 0.034503135830163956, + 0.08884551376104355, + 0.06186550855636597, + 0.321542352437973, + 0.08945857733488083, + 0.15073071420192719, + 0.11632120609283447 + ], + "credit_map": [ + 0.007873492315411568, + 0.004646564833819866, + 0.00869375467300415, + 0.010580929927527905, + 0.005729450844228268, + 0.006403704173862934, + 0.004509285558015108, + 0.007487742230296135, + 0.006624875124543905, + 0.004759516566991806, + 0.005851045250892639, + 0.0063156429678201675, + 0.005885970313102007, + 0.005881980061531067, + 0.004614387638866901, + 0.004777092952281237, + 0.005287129897624254, + 0.0037533293943852186, + 0.00507454015314579, + 0.005659340415149927, + 0.0036601671017706394, + 0.00489777559414506, + 0.0034241992980241776, + 0.004341068211942911, + 0.005792161449790001, + 0.0024562592152506113, + 0.005043970886617899, + 0.00541755510494113, + 0.004192977678030729, + 0.005117607302963734, + 0.0032886897679418325, + 0.0031939158216118813, + 0.01222385372966528, + 0.008442558348178864, + 0.013253548182547092, + 0.01145575474947691, + 0.010088326409459114, + 0.012873655185103416, + 0.0076428744941949844, + 0.012864946387708187, + 0.009141508489847183, + 0.0059822979383170605, + 0.01031097024679184, + 0.0062651252374053, + 0.008168510161340237, + 0.007788456045091152, + 0.006636640056967735, + 0.007571996655315161, + 0.05144655331969261, + 0.019427169114351273, + 0.051364701241254807, + 0.022630728781223297, + 0.021024778485298157, + 0.10731367021799088, + 0.02429109252989292, + 0.024043692275881767, + 0.012004622258245945, + 0.010531567968428135, + 0.013019360601902008, + 0.008348120376467705, + 0.013965288177132607, + 0.011437525972723961, + 0.00884411670267582, + 0.011307971552014351, + 0.020137960091233253, + 0.015347162261605263, + 0.02313665859401226, + 0.01692892238497734, + 0.01399211771786213, + 0.017751554027199745, + 0.02950405701994896, + 0.013932278379797935, + 0.013557536527514458, + 0.012068635784089565, + 0.015792161226272583, + 0.01383993960916996, + 0.012781338766217232, + 0.014302353374660015, + 0.02063431218266487, + 0.013344929553568363 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 8, + 9 + ], + "selected_reward": -0.38700783252716064 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000101/group_14/index_tensor([7182, 7182])_group_14_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000101/group_14/index_tensor([7182, 7182])_group_14_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..297fb2b634c3fcc1b9507c5aebfe9c8537bcbc42 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000101/group_14/index_tensor([7182, 7182])_group_14_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "The camera pans in a fluid, unbroken 360-degree pan, capturing a fierce confrontation deep within an abandoned factory from a slightly upward angle. Two biomechanical creatures, a fusion of organic form and cold machinery, engage in a desperate struggle on a corroded metal floor covered in rust and oil. Their bodies are composed of mirror-polished chrome, their intricate internal structures visible through their skeletal shells. Countless precisely meshing gears whirr as they move, emitting a cold metallic gleam. One of the creatures resembles a scorpion, its massive hydraulic jaws snapping shut with immense force. The other, more like a multi-legged spider, its forelimbs terminating in arc cutters that sizzle and flash with a menacing blue-white electric light. As the claws and cutters collide violently, a blinding storm of sparks erupts. These flaming particles briefly illuminate the surrounding darkness, casting a frenzy of light and shadow across their mirror-polished chrome shells and illuminating the massive silhouettes of rusted chains and abandoned machinery hanging in the background. The entire scene is full of primitive violent aesthetics and iconic biomechanical art style, creating a tense, futuristic and destructive atmosphere of conflict amidst the harsh screams of metal and the crackling of electric arcs.", + "seed": 8512, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000101/group_14/index_7182_seed_8512_dp_28_rank_56_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 14, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.08617508709430695, + "vlm_logprob": -0.08617508709430695, + "vlm_yes_no_margin": 3.374999976158142, + "vlm_token_credit": -0.08617508709430695, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are two biomechanical creatures visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the environment an abandoned factory with a corroded metal floor covered in rust and oil?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the creatures have bodies composed of mirror-polished chrome with visible internal structures?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is one creature resembling a scorpion with snapping hydraulic jaws and the other resembling a spider with arc cutters emitting blue-white electric light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera perform a fluid, unbroken 360-degree pan?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.011134984903037548 + ], + "opposite_logprob_tokens": [ + -4.761135101318359 + ], + "logprob_reward": -0.011134984903037548, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.011134984903037548, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.23958145081996918, + 0.04964533448219299, + 0.17391526699066162, + 0.02715410478413105, + 0.05633572116494179, + 0.04416220635175705, + 0.038708414882421494, + 0.18217967450618744, + 0.129073366522789, + 0.05924442782998085 + ], + "credit_map": [ + 0.033919185400009155, + 0.033258501440286636, + 0.016213135793805122, + 0.018352843821048737, + 0.033166322857141495, + 0.0228666253387928, + 0.03617189824581146, + 0.04563293978571892, + 0.005402261391282082, + 0.00754670100286603, + 0.006431179586797953, + 0.00590828200802207, + 0.005918465554714203, + 0.006332624703645706, + 0.007336141541600227, + 0.004769682418555021, + 0.023263225331902504, + 0.02266559563577175, + 0.019780030474066734, + 0.018569862470030785, + 0.019930129870772362, + 0.02942119725048542, + 0.011799768544733524, + 0.028485450893640518, + 0.005133498460054398, + 0.00223845848813653, + 0.0038674380630254745, + 0.0020584699232131243, + 0.0027676664758473635, + 0.004800848197191954, + 0.0026542500127106905, + 0.0036334742326289415, + 0.009694064036011696, + 0.005582221783697605, + 0.009184762835502625, + 0.004180989693850279, + 0.0062463488429784775, + 0.006712747272104025, + 0.006163858808577061, + 0.00857072975486517, + 0.00626755366101861, + 0.005823459941893816, + 0.005614631809294224, + 0.0034472544211894274, + 0.006239228416234255, + 0.005937278736382723, + 0.0037850127555429935, + 0.00704778591170907, + 0.003795204684138298, + 0.0050041754730045795, + 0.004474949557334185, + 0.005013967398554087, + 0.004697435535490513, + 0.005177020560950041, + 0.005494358018040657, + 0.005051301792263985, + 0.03234880045056343, + 0.014074447564780712, + 0.014882329851388931, + 0.01661381870508194, + 0.03877891227602959, + 0.02420678175985813, + 0.024278530851006508, + 0.016996052116155624, + 0.032664090394973755, + 0.009757054038345814, + 0.010106907226145267, + 0.006127715576440096, + 0.009717783890664577, + 0.01042818371206522, + 0.012276743538677692, + 0.03799488767981529, + 0.006333353463560343, + 0.006029585842043161, + 0.009037680923938751, + 0.009145296178758144, + 0.006781575735658407, + 0.007027753628790379, + 0.006779246497899294, + 0.008109934628009796 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 7, + 2 + ], + "selected_reward": -0.011134984903037548 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are two biomechanical creatures visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the environment an abandoned factory with a corroded metal floor covered in rust and oil?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the creatures have bodies composed of mirror-polished chrome with visible internal structures?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is one creature resembling a scorpion with snapping hydraulic jaws and the other resembling a spider with arc cutters emitting blue-white electric light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera perform a fluid, unbroken 360-degree pan?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.1427878439426422 + ], + "opposite_logprob_tokens": [ + -2.0177879333496094 + ], + "logprob_reward": -0.1427878439426422, + "yes_no_margin_reward": 1.8750001192092896, + "token_credit_reward": -0.1427878439426422, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06729628145694733, + 0.09927958250045776, + 0.06917987763881683, + 0.051786456257104874, + 0.05867637321352959, + 0.1265316605567932, + 0.049746330827474594, + 0.22539743781089783, + 0.18860840797424316, + 0.06349755823612213 + ], + "credit_map": [ + 0.009812897071242332, + 0.009773323312401772, + 0.006670283619314432, + 0.00527055561542511, + 0.008215134963393211, + 0.0076956553384661674, + 0.010803977027535439, + 0.009054453112185001, + 0.01765424571931362, + 0.012362562119960785, + 0.007780381944030523, + 0.007771007716655731, + 0.029063737019896507, + 0.008424925617873669, + 0.008315623737871647, + 0.007907100953161716, + 0.00870757270604372, + 0.010779458098113537, + 0.009176348336040974, + 0.006551187951117754, + 0.009191764518618584, + 0.009181197732686996, + 0.005436284467577934, + 0.010156062431633472, + 0.008288891986012459, + 0.007062725257128477, + 0.006788672413676977, + 0.0048288460820913315, + 0.005277604795992374, + 0.008488436229526997, + 0.004695620387792587, + 0.006355662830173969, + 0.010783662088215351, + 0.007328244857490063, + 0.009884906932711601, + 0.0035049873404204845, + 0.007355824112892151, + 0.007077764254063368, + 0.005516103934496641, + 0.0072248829528689384, + 0.032368261367082596, + 0.014531751163303852, + 0.02395128272473812, + 0.010546545498073101, + 0.00874237809330225, + 0.018000606447458267, + 0.00799038540571928, + 0.01040043868124485, + 0.005670891609042883, + 0.004934993572533131, + 0.00681473221629858, + 0.005906805861741304, + 0.007870576344430447, + 0.006524750031530857, + 0.005657038651406765, + 0.0063665397465229034, + 0.02965926192700863, + 0.024450846016407013, + 0.025028157979249954, + 0.02436518482863903, + 0.04435276985168457, + 0.02927165850996971, + 0.018798669800162315, + 0.029470890760421753, + 0.05548657849431038, + 0.014699005521833897, + 0.012184781022369862, + 0.008281709626317024, + 0.020580122247338295, + 0.020628785714507103, + 0.014977785758674145, + 0.04176963493227959, + 0.006621557287871838, + 0.006686901208013296, + 0.009972483851015568, + 0.007951402105391026, + 0.0076020085252821445, + 0.004736291244626045, + 0.011610415764153004, + 0.008316492661833763 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 8, + 5 + ], + "selected_reward": -0.1427878439426422 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are two biomechanical creatures visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the environment an abandoned factory with a corroded metal floor covered in rust and oil?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the creatures have bodies composed of mirror-polished chrome with visible internal structures?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is one creature resembling a scorpion with snapping hydraulic jaws and the other resembling a spider with arc cutters emitting blue-white electric light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera perform a fluid, unbroken 360-degree pan?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.018183909356594086 + ], + "opposite_logprob_tokens": [ + -4.018183708190918 + ], + "logprob_reward": -0.018183909356594086, + "yes_no_margin_reward": 3.999999761581421, + "token_credit_reward": -0.018183909356594086, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11544885486364365, + 0.0696268305182457, + 0.10510845482349396, + 0.1304253339767456, + 0.07354696094989777, + 0.10983254760503769, + 0.1955346167087555, + 0.07646123319864273, + 0.08817324042320251, + 0.035841912031173706 + ], + "credit_map": [ + 0.016101660206913948, + 0.022488519549369812, + 0.009215401485562325, + 0.011526543647050858, + 0.014003653079271317, + 0.010795419104397297, + 0.019511016085743904, + 0.011806639842689037, + 0.011749960482120514, + 0.011167118325829506, + 0.007738067302852869, + 0.0045784832909703255, + 0.013114124536514282, + 0.008053697645664215, + 0.007261671591550112, + 0.005963707342743874, + 0.013194018043577671, + 0.013955643400549889, + 0.009806991554796696, + 0.009378057904541492, + 0.011620379984378815, + 0.0188363716006279, + 0.008279724046587944, + 0.020037269219756126, + 0.04126839339733124, + 0.008390062488615513, + 0.014326955191791058, + 0.007379285525530577, + 0.010503504425287247, + 0.026077235117554665, + 0.01013217307627201, + 0.012347724288702011, + 0.010953851975500584, + 0.010468238964676857, + 0.009251476265490055, + 0.006605643313378096, + 0.008359243161976337, + 0.008042359724640846, + 0.01083981804549694, + 0.009026333689689636, + 0.021542062982916832, + 0.013369984924793243, + 0.01859702169895172, + 0.00881020724773407, + 0.009340211749076843, + 0.014253487810492516, + 0.008567877113819122, + 0.015351695008575916, + 0.019374504685401917, + 0.02822139486670494, + 0.015538746491074562, + 0.020412549376487732, + 0.030235232785344124, + 0.027703609317541122, + 0.02604815922677517, + 0.02800043858587742, + 0.013713409192860126, + 0.006948132533580065, + 0.005491437390446663, + 0.009213456884026527, + 0.013838008046150208, + 0.008759173564612865, + 0.009550441056489944, + 0.008947175927460194, + 0.015106010250747204, + 0.015307296067476273, + 0.009181387722492218, + 0.005564853083342314, + 0.017237365245819092, + 0.00717146135866642, + 0.0074620964005589485, + 0.01114276610314846, + 0.003925213124603033, + 0.004676726181060076, + 0.004775349050760269, + 0.006606557406485081, + 0.004371371120214462, + 0.002283894456923008, + 0.004464256577193737, + 0.0047385417856276035 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 3, + 0 + ], + "selected_reward": -0.018183909356594086 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are two biomechanical creatures visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the environment an abandoned factory with a corroded metal floor covered in rust and oil?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the creatures have bodies composed of mirror-polished chrome with visible internal structures?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is one creature resembling a scorpion with snapping hydraulic jaws and the other resembling a spider with arc cutters emitting blue-white electric light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera perform a fluid, unbroken 360-degree pan?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.006737728603184223 + ], + "opposite_logprob_tokens": [ + -5.00673770904541 + ], + "logprob_reward": -0.006737728603184223, + "yes_no_margin_reward": 5.0, + "token_credit_reward": -0.006737728603184223, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11014264822006226, + 0.11425558477640152, + 0.08579215407371521, + 0.08629722893238068, + 0.05457659065723419, + 0.04740994796156883, + 0.053136736154556274, + 0.12520398199558258, + 0.22251775860786438, + 0.10066744685173035 + ], + "credit_map": [ + 0.02315700799226761, + 0.015142614021897316, + 0.009535929188132286, + 0.008135153912007809, + 0.01543524768203497, + 0.013719601556658745, + 0.01169014535844326, + 0.013326939195394516, + 0.014567341655492783, + 0.013681629672646523, + 0.011907944455742836, + 0.011523005552589893, + 0.02389555424451828, + 0.01705590821802616, + 0.012428315356373787, + 0.00919587817043066, + 0.011895625852048397, + 0.01583193615078926, + 0.010026169009506702, + 0.010112940333783627, + 0.009185641072690487, + 0.01016288623213768, + 0.0060358415357768536, + 0.01254111435264349, + 0.01568106934428215, + 0.010176529176533222, + 0.011643927544355392, + 0.0066150035709142685, + 0.0075484588742256165, + 0.01593313179910183, + 0.007386165205389261, + 0.011312930844724178, + 0.007413432467728853, + 0.0064374287612736225, + 0.010247497819364071, + 0.004337762948125601, + 0.004997034557163715, + 0.006714996881783009, + 0.006097692996263504, + 0.008330742828547955, + 0.009692930616438389, + 0.004351400304585695, + 0.008337770588696003, + 0.002939340891316533, + 0.006015529856085777, + 0.006100863683968782, + 0.003947305027395487, + 0.006024801172316074, + 0.0044813454151153564, + 0.006669696886092424, + 0.0061340900138020515, + 0.0059316325932741165, + 0.010378059931099415, + 0.0075908927246928215, + 0.006227860227227211, + 0.005723155569285154, + 0.01551035512238741, + 0.014562640339136124, + 0.014620650559663773, + 0.015231396071612835, + 0.024315936490893364, + 0.01728743687272072, + 0.010369416326284409, + 0.013306139037013054, + 0.04452471807599068, + 0.03554113209247589, + 0.018175136297941208, + 0.013980619609355927, + 0.03888881579041481, + 0.01804296486079693, + 0.018729278817772865, + 0.03463507071137428, + 0.013440234586596489, + 0.011973164044320583, + 0.020613305270671844, + 0.01165858469903469, + 0.015587883070111275, + 0.008244642056524754, + 0.008014673367142677, + 0.011134948581457138 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 7, + 1 + ], + "selected_reward": -0.006737728603184223 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are two biomechanical creatures visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the environment an abandoned factory with a corroded metal floor covered in rust and oil?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the creatures have bodies composed of mirror-polished chrome with visible internal structures?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is one creature resembling a scorpion with snapping hydraulic jaws and the other resembling a spider with arc cutters emitting blue-white electric light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera perform a fluid, unbroken 360-degree pan?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.25203096866607666 + ], + "opposite_logprob_tokens": [ + -1.5020309686660767 + ], + "logprob_reward": -0.25203096866607666, + "yes_no_margin_reward": 1.25, + "token_credit_reward": -0.25203096866607666, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11179523169994354, + 0.08248575031757355, + 0.08296395093202591, + 0.09588222205638885, + 0.06838175654411316, + 0.29109063744544983, + 0.043931443244218826, + 0.0632176399230957, + 0.10402573645114899, + 0.05622566491365433 + ], + "credit_map": [ + 0.016661418601870537, + 0.013339856639504433, + 0.009412267245352268, + 0.011921592056751251, + 0.011226166039705276, + 0.010706311091780663, + 0.020761245861649513, + 0.017766373232007027, + 0.011513312347233295, + 0.013947825878858566, + 0.006754505913704634, + 0.006723265629261732, + 0.012242762371897697, + 0.011147460900247097, + 0.01226872019469738, + 0.007887902669608593, + 0.011292031966149807, + 0.013753366656601429, + 0.007880241610109806, + 0.007170910481363535, + 0.01076732948422432, + 0.014817408286035061, + 0.0054220822639763355, + 0.011860574595630169, + 0.014522489160299301, + 0.012037258595228195, + 0.013632822781801224, + 0.008609461598098278, + 0.0105101577937603, + 0.014304249547421932, + 0.008330234326422215, + 0.013935545459389687, + 0.009113481268286705, + 0.007884194143116474, + 0.0085277846083045, + 0.005964584182947874, + 0.011637880466878414, + 0.007194806355983019, + 0.009631572291254997, + 0.008427456021308899, + 0.0756903886795044, + 0.025575928390026093, + 0.05739526078104973, + 0.01519384142011404, + 0.01598544791340828, + 0.04831473156809807, + 0.017339864745736122, + 0.03559517115354538, + 0.004857973661273718, + 0.005231325980275869, + 0.004871834069490433, + 0.004284283146262169, + 0.007983040064573288, + 0.006232955027371645, + 0.004699163604527712, + 0.005770871881395578, + 0.012806103564798832, + 0.006006012205034494, + 0.006013268139213324, + 0.007493334822356701, + 0.009416336193680763, + 0.0067162527702748775, + 0.008896837942302227, + 0.0058694928884506226, + 0.014301302842795849, + 0.023430969566106796, + 0.011075757443904877, + 0.008730707690119743, + 0.01664092391729355, + 0.010185526683926582, + 0.008668424561619759, + 0.010992113500833511, + 0.007449757773429155, + 0.008858962915837765, + 0.009584929794073105, + 0.00638941116631031, + 0.006161073222756386, + 0.004012214485555887, + 0.0071964566595852375, + 0.006572859361767769 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 0, + 8 + ], + "selected_reward": -0.25203096866607666 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "The camera pans in a fluid, unbroken 360-degree pan, capturing a fierce confrontation deep within an abandoned factory from a slightly upward angle. Two biomechanical creatures, a fusion of organic form and cold machinery, engage in a desperate struggle on a corroded metal floor covered in rust and oil. Their bodies are composed of mirror-polished chrome, their intricate internal structures visible through their skeletal shells. Countless precisely meshing gears whirr as they move, emitting a cold metallic gleam. One of the creatures resembles a scorpion, its massive hydraulic jaws snapping shut with immense force. The other, more like a multi-legged spider, its forelimbs terminating in arc cutters that sizzle and flash with a menacing blue-white electric light. As the claws and cutters collide violently, a blinding storm of sparks erupts. These flaming particles briefly illuminate the surrounding darkness, casting a frenzy of light and shadow across their mirror-polished chrome shells and illuminating the massive silhouettes of rusted chains and abandoned machinery hanging in the background. The entire scene is full of primitive violent aesthetics and iconic biomechanical art style, creating a tense, futuristic and destructive atmosphere of conflict amidst the harsh screams of metal and the crackling of electric arcs.", + "seed": 8513, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000101/group_14/index_7182_seed_8513_dp_28_rank_56_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 14, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.1558406513184309, + "vlm_logprob": -0.1558406513184309, + "vlm_yes_no_margin": 2.874999976158142, + "vlm_token_credit": -0.1558406513184309, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are two biomechanical creatures visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the environment an abandoned factory with a corroded metal floor covered in rust and oil?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the creatures have bodies composed of mirror-polished chrome with visible internal structures?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is one creature resembling a scorpion with snapping hydraulic jaws and the other resembling a spider with arc cutters emitting blue-white electric light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera perform a fluid, unbroken 360-degree pan?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.020733410492539406 + ], + "opposite_logprob_tokens": [ + -4.14573335647583 + ], + "logprob_reward": -0.020733410492539406, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.020733410492539406, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15628427267074585, + 0.07865741848945618, + 0.17521080374717712, + 0.10015951097011566, + 0.036326706409454346, + 0.049710407853126526, + 0.11051666736602783, + 0.06473088264465332, + 0.12987549602985382, + 0.09852787852287292 + ], + "credit_map": [ + 0.019463812932372093, + 0.016376348212361336, + 0.015895701944828033, + 0.016098491847515106, + 0.026469485834240913, + 0.01700342632830143, + 0.02510308474302292, + 0.019873913377523422, + 0.007733388338238001, + 0.008030280470848083, + 0.008851919323205948, + 0.011374983936548233, + 0.009805685840547085, + 0.014962558634579182, + 0.008737575262784958, + 0.009161029011011124, + 0.022293049842119217, + 0.017267568036913872, + 0.026431068778038025, + 0.023448023945093155, + 0.017393147572875023, + 0.026667078956961632, + 0.024824874475598335, + 0.016885999590158463, + 0.01609647274017334, + 0.012048767879605293, + 0.012366726994514465, + 0.014180346392095089, + 0.01126444898545742, + 0.012109750881791115, + 0.011700586415827274, + 0.010392408818006516, + 0.005907718557864428, + 0.004335372243076563, + 0.004785927943885326, + 0.0038896468468010426, + 0.003999172709882259, + 0.0036791558377444744, + 0.0046196384355425835, + 0.00511007197201252, + 0.00597990769892931, + 0.006206645164638758, + 0.006385075859725475, + 0.006648332346230745, + 0.007531430572271347, + 0.00428664218634367, + 0.0066496566869318485, + 0.006022719666361809, + 0.015473537147045135, + 0.02275739423930645, + 0.010324631817638874, + 0.013563902117311954, + 0.017664536833763123, + 0.006760706659406424, + 0.010852553881704807, + 0.013119406066834927, + 0.00595445791259408, + 0.0055915070697665215, + 0.011383874341845512, + 0.016995195299386978, + 0.006912220269441605, + 0.0057151406072080135, + 0.007286035921424627, + 0.004892445635050535, + 0.015837552025914192, + 0.007176969666033983, + 0.022518130019307137, + 0.03212179243564606, + 0.007606728468090296, + 0.017618980258703232, + 0.014274518936872482, + 0.012720810249447823, + 0.015840701758861542, + 0.011133684776723385, + 0.02167065627872944, + 0.009614560753107071, + 0.0070742471143603325, + 0.009651320986449718, + 0.014489552937448025, + 0.009053157642483711 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 0, + 8 + ], + "selected_reward": -0.020733410492539406 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are two biomechanical creatures visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the environment an abandoned factory with a corroded metal floor covered in rust and oil?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the creatures have bodies composed of mirror-polished chrome with visible internal structures?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is one creature resembling a scorpion with snapping hydraulic jaws and the other resembling a spider with arc cutters emitting blue-white electric light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera perform a fluid, unbroken 360-degree pan?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.1427968442440033 + ], + "opposite_logprob_tokens": [ + -2.017796754837036 + ], + "logprob_reward": -0.1427968442440033, + "yes_no_margin_reward": 1.8749998807907104, + "token_credit_reward": -0.1427968442440033, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.17003974318504333, + 0.21336445212364197, + 0.057037416845560074, + 0.02808516100049019, + 0.040657684206962585, + 0.08504154533147812, + 0.08265803754329681, + 0.19134773313999176, + 0.0986144170165062, + 0.033153753727674484 + ], + "credit_map": [ + 0.018733670935034752, + 0.020414095371961594, + 0.01698973961174488, + 0.0152570940554142, + 0.019011298194527626, + 0.03201637417078018, + 0.019539009779691696, + 0.028078466653823853, + 0.020009193569421768, + 0.01710158959031105, + 0.017305057495832443, + 0.016802020370960236, + 0.015671810135245323, + 0.05817946046590805, + 0.014620492234826088, + 0.0536748468875885, + 0.008961623534560204, + 0.004453081171959639, + 0.008188375271856785, + 0.005840353202074766, + 0.005448529496788979, + 0.009967184625566006, + 0.0081800427287817, + 0.005998231004923582, + 0.005292841698974371, + 0.002981777535751462, + 0.003203397151082754, + 0.0028226638678461313, + 0.0031296759843826294, + 0.004061545245349407, + 0.0034175219479948282, + 0.0031757415272295475, + 0.005212616641074419, + 0.004350859671831131, + 0.005373375955969095, + 0.006809913087636232, + 0.004743329714983702, + 0.004525013733655214, + 0.004737886134535074, + 0.004904692992568016, + 0.012748056091368198, + 0.013206600211560726, + 0.01573873683810234, + 0.014075402170419693, + 0.006813399028033018, + 0.006500702816992998, + 0.006620560307055712, + 0.00933810044080019, + 0.017344802618026733, + 0.010422268882393837, + 0.007521369494497776, + 0.013200956396758556, + 0.01053201500326395, + 0.0048127556219697, + 0.008186453022062778, + 0.010637428611516953, + 0.018989278003573418, + 0.03178609907627106, + 0.031357474625110626, + 0.04283227026462555, + 0.01954878866672516, + 0.013303178362548351, + 0.018151700496673584, + 0.01537897065281868, + 0.014106608927249908, + 0.007324456237256527, + 0.010757507756352425, + 0.017012467607855797, + 0.009311985224485397, + 0.018892165273427963, + 0.010935937985777855, + 0.010273301973938942, + 0.0038746583741158247, + 0.003852998139336705, + 0.003936159890145063, + 0.0025242113042622805, + 0.0038251879159361124, + 0.004329017363488674, + 0.004984382074326277, + 0.005827142391353846 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 7, + 0 + ], + "selected_reward": -0.1427968442440033 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are two biomechanical creatures visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the environment an abandoned factory with a corroded metal floor covered in rust and oil?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the creatures have bodies composed of mirror-polished chrome with visible internal structures?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is one creature resembling a scorpion with snapping hydraulic jaws and the other resembling a spider with arc cutters emitting blue-white electric light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera perform a fluid, unbroken 360-degree pan?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.029802024364471436 + ], + "opposite_logprob_tokens": [ + -3.529802083969116 + ], + "logprob_reward": -0.029802024364471436, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.029802024364471436, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1494651436805725, + 0.29294246435165405, + 0.08051830530166626, + 0.09469907730817795, + 0.05640124902129173, + 0.036053117364645004, + 0.04479001462459564, + 0.08497121930122375, + 0.045757126063108444, + 0.11440235376358032 + ], + "credit_map": [ + 0.016771728172898293, + 0.020104041323065758, + 0.01773492805659771, + 0.01685606688261032, + 0.022445671260356903, + 0.026428120210766792, + 0.014018494635820389, + 0.015106101520359516, + 0.044763773679733276, + 0.049549032002687454, + 0.032120488584041595, + 0.023202000185847282, + 0.03275642916560173, + 0.05814697965979576, + 0.019911859184503555, + 0.032491911202669144, + 0.007685162127017975, + 0.009976746514439583, + 0.010177500545978546, + 0.008767565712332726, + 0.010982654057443142, + 0.011303070932626724, + 0.011981521733105183, + 0.009644088335335255, + 0.014390537515282631, + 0.013085559010505676, + 0.011243983171880245, + 0.01325142476707697, + 0.012237020768225193, + 0.011341306380927563, + 0.010636438615620136, + 0.008512808941304684, + 0.008081698790192604, + 0.007318434771150351, + 0.007707221899181604, + 0.007134548854082823, + 0.00717973243445158, + 0.005325098987668753, + 0.006473364774137735, + 0.00718114385381341, + 0.0037070014514029026, + 0.003818462835624814, + 0.004285338334739208, + 0.005578596144914627, + 0.005406504031270742, + 0.0036379361990839243, + 0.004899663385003805, + 0.0047196149826049805, + 0.008883995935320854, + 0.004753137473016977, + 0.004802942741662264, + 0.006882327608764172, + 0.006328903138637543, + 0.004393553826957941, + 0.004759952891618013, + 0.003985202871263027, + 0.007129324600100517, + 0.01096404530107975, + 0.010265122167766094, + 0.014877678826451302, + 0.008326627314090729, + 0.008345715701580048, + 0.01280675083398819, + 0.012255948968231678, + 0.005630385130643845, + 0.004036464262753725, + 0.00648453226312995, + 0.010037458501756191, + 0.005029087886214256, + 0.0059754252433776855, + 0.0049363914877176285, + 0.003627379657700658, + 0.020913586020469666, + 0.009924915619194508, + 0.016786059364676476, + 0.011088180355727673, + 0.009730761870741844, + 0.009172536432743073, + 0.016180530190467834, + 0.02060578390955925 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 9 + ], + "selected_reward": -0.029802024364471436 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are two biomechanical creatures visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the environment an abandoned factory with a corroded metal floor covered in rust and oil?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the creatures have bodies composed of mirror-polished chrome with visible internal structures?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is one creature resembling a scorpion with snapping hydraulic jaws and the other resembling a spider with arc cutters emitting blue-white electric light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera perform a fluid, unbroken 360-degree pan?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.009786602109670639 + ], + "opposite_logprob_tokens": [ + -4.634786605834961 + ], + "logprob_reward": -0.009786602109670639, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.009786602109670639, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0685574859380722, + 0.08667965978384018, + 0.07809028029441833, + 0.056689150631427765, + 0.0319344736635685, + 0.11185246706008911, + 0.03315577656030655, + 0.3654138743877411, + 0.06420959532260895, + 0.10341726243495941 + ], + "credit_map": [ + 0.009334804490208626, + 0.007268406916409731, + 0.009578855708241463, + 0.008751039393246174, + 0.006689575035125017, + 0.008363256230950356, + 0.008413695730268955, + 0.01015784777700901, + 0.013655267655849457, + 0.013031947426497936, + 0.011963480152189732, + 0.01126439031213522, + 0.009616069495677948, + 0.010911252349615097, + 0.008292991667985916, + 0.00794425792992115, + 0.01075134240090847, + 0.006398203317075968, + 0.012288532219827175, + 0.008979440666735172, + 0.011180085130035877, + 0.009985202923417091, + 0.010153976269066334, + 0.008353491313755512, + 0.008678846061229706, + 0.0060315802693367004, + 0.006737979594618082, + 0.006362310145050287, + 0.009128349833190441, + 0.005912003573030233, + 0.00731687992811203, + 0.006521195638924837, + 0.003626385936513543, + 0.0039901710115373135, + 0.003956574015319347, + 0.003589475294575095, + 0.004137233830988407, + 0.0027367372531443834, + 0.0052915229462087154, + 0.004606370348483324, + 0.015878600999712944, + 0.015098619274795055, + 0.016325701028108597, + 0.011654206551611423, + 0.022138362750411034, + 0.007597840391099453, + 0.013681543059647083, + 0.009477588348090649, + 0.0030019728001207113, + 0.004006297793239355, + 0.004821332637220621, + 0.005753565113991499, + 0.0042230174876749516, + 0.0030714853201061487, + 0.005184563808143139, + 0.003093540668487549, + 0.03754644840955734, + 0.05298168212175369, + 0.07414790242910385, + 0.07172755151987076, + 0.04086196422576904, + 0.027136703953146935, + 0.028302175924181938, + 0.03270943462848663, + 0.006399785168468952, + 0.006525612436234951, + 0.009243781678378582, + 0.006981492508202791, + 0.009089960716664791, + 0.009776301681995392, + 0.006636281032115221, + 0.009556379169225693, + 0.023086467757821083, + 0.005644882097840309, + 0.029542071744799614, + 0.00835347454994917, + 0.006013217847794294, + 0.007056638598442078, + 0.011325732804834843, + 0.012394771911203861 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 5, + 9 + ], + "selected_reward": -0.009786602109670639 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are two biomechanical creatures visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the environment an abandoned factory with a corroded metal floor covered in rust and oil?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the creatures have bodies composed of mirror-polished chrome with visible internal structures?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is one creature resembling a scorpion with snapping hydraulic jaws and the other resembling a spider with arc cutters emitting blue-white electric light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera perform a fluid, unbroken 360-degree pan?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.5760843753814697 + ], + "opposite_logprob_tokens": [ + -0.8260843753814697 + ], + "logprob_reward": -0.5760843753814697, + "yes_no_margin_reward": 0.25, + "token_credit_reward": -0.5760843753814697, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10626620054244995, + 0.09172965586185455, + 0.07162162661552429, + 0.047566141933202744, + 0.0383961983025074, + 0.1214541643857956, + 0.14779254794120789, + 0.15590859949588776, + 0.17411570250988007, + 0.04514916241168976 + ], + "credit_map": [ + 0.012973223812878132, + 0.01162869855761528, + 0.011469363234937191, + 0.010944168083369732, + 0.012880328111350536, + 0.012863281182944775, + 0.01397367101162672, + 0.01953347586095333, + 0.011403760872781277, + 0.011562218889594078, + 0.009853427298367023, + 0.01015574112534523, + 0.00991694163531065, + 0.014841210097074509, + 0.006788965314626694, + 0.017207395285367966, + 0.007212119642645121, + 0.008175482042133808, + 0.009916013106703758, + 0.007951640523970127, + 0.00882104691118002, + 0.010375967249274254, + 0.01058921031653881, + 0.008580151945352554, + 0.010295525193214417, + 0.0052729398012161255, + 0.006616897881031036, + 0.005922774318605661, + 0.004294008947908878, + 0.006109177600592375, + 0.005122536327689886, + 0.003932281397283077, + 0.004572134464979172, + 0.004246763419359922, + 0.004367141053080559, + 0.005366871599107981, + 0.005081217736005783, + 0.0035671188961714506, + 0.005734426900744438, + 0.0054605258628726006, + 0.01714043691754341, + 0.01797965168952942, + 0.015676256269216537, + 0.01460344810038805, + 0.023184331133961678, + 0.008225704543292522, + 0.014587599784135818, + 0.010056738741695881, + 0.018046194687485695, + 0.013439422473311424, + 0.019684622064232826, + 0.021051591262221336, + 0.025886695832014084, + 0.014444738626480103, + 0.019532907754182816, + 0.015706392005085945, + 0.015094133093953133, + 0.02026406303048134, + 0.0286411140114069, + 0.03582116588950157, + 0.015057282522320747, + 0.015124619007110596, + 0.012148109264671803, + 0.013758119195699692, + 0.025827987119555473, + 0.011190683580935001, + 0.027045099064707756, + 0.05042409524321556, + 0.012022316455841064, + 0.02162548527121544, + 0.014474469237029552, + 0.01150557678192854, + 0.007715091109275818, + 0.005895309150218964, + 0.009426345117390156, + 0.005803458392620087, + 0.004115815740078688, + 0.0032986097503453493, + 0.005723918788135052, + 0.003170615993440151 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 7, + 6 + ], + "selected_reward": -0.5760843753814697 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "The camera pans in a fluid, unbroken 360-degree pan, capturing a fierce confrontation deep within an abandoned factory from a slightly upward angle. Two biomechanical creatures, a fusion of organic form and cold machinery, engage in a desperate struggle on a corroded metal floor covered in rust and oil. Their bodies are composed of mirror-polished chrome, their intricate internal structures visible through their skeletal shells. Countless precisely meshing gears whirr as they move, emitting a cold metallic gleam. One of the creatures resembles a scorpion, its massive hydraulic jaws snapping shut with immense force. The other, more like a multi-legged spider, its forelimbs terminating in arc cutters that sizzle and flash with a menacing blue-white electric light. As the claws and cutters collide violently, a blinding storm of sparks erupts. These flaming particles briefly illuminate the surrounding darkness, casting a frenzy of light and shadow across their mirror-polished chrome shells and illuminating the massive silhouettes of rusted chains and abandoned machinery hanging in the background. The entire scene is full of primitive violent aesthetics and iconic biomechanical art style, creating a tense, futuristic and destructive atmosphere of conflict amidst the harsh screams of metal and the crackling of electric arcs.", + "seed": 8514, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000101/group_14/index_7182_seed_8514_dp_29_rank_58_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 14, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.11531778499484062, + "vlm_logprob": -0.11531778499484062, + "vlm_yes_no_margin": 2.825, + "vlm_token_credit": -0.11531778499484062, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are two biomechanical creatures visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the environment an abandoned factory with a corroded metal floor covered in rust and oil?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the creatures have bodies composed of mirror-polished chrome with visible internal structures?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is one creature resembling a scorpion with snapping hydraulic jaws and the other resembling a spider with arc cutters emitting blue-white electric light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera perform a fluid, unbroken 360-degree pan?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.041719112545251846 + ], + "opposite_logprob_tokens": [ + -3.2917191982269287 + ], + "logprob_reward": -0.041719112545251846, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.041719112545251846, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.23058182001113892, + 0.04144306853413582, + 0.053765919059515, + 0.21712298691272736, + 0.1026342511177063, + 0.05682769790291786, + 0.0537387952208519, + 0.06447356939315796, + 0.11262568086385727, + 0.0667862743139267 + ], + "credit_map": [ + 0.019274549558758736, + 0.030072053894400597, + 0.030046267434954643, + 0.03202908858656883, + 0.03424002602696419, + 0.024188756942749023, + 0.029856715351343155, + 0.030874349176883698, + 0.005007983185350895, + 0.005880285985767841, + 0.0032939722295850515, + 0.004687184933573008, + 0.007999803870916367, + 0.005068647209554911, + 0.004533829167485237, + 0.004971357993781567, + 0.005238322541117668, + 0.0057132490910589695, + 0.006677009165287018, + 0.006417513359338045, + 0.006418247241526842, + 0.005193660035729408, + 0.009232454933226109, + 0.008875459432601929, + 0.02842492237687111, + 0.024850906804203987, + 0.033419135957956314, + 0.027395742014050484, + 0.028088316321372986, + 0.03259457275271416, + 0.021849192678928375, + 0.020500192418694496, + 0.015683991834521294, + 0.007008085958659649, + 0.014222042635083199, + 0.013571544550359249, + 0.009443298913538456, + 0.017136478796601295, + 0.006298011634498835, + 0.019270790740847588, + 0.005579750519245863, + 0.004981652367860079, + 0.006144058424979448, + 0.005071524530649185, + 0.010595542378723621, + 0.005453881341964006, + 0.006517047993838787, + 0.0124842319637537, + 0.007570004556328058, + 0.006921298336237669, + 0.005205128341913223, + 0.005858318880200386, + 0.006411299109458923, + 0.005119496025145054, + 0.009487317875027657, + 0.007165929302573204, + 0.012606063857674599, + 0.008071496151387691, + 0.006829271093010902, + 0.006066412199288607, + 0.009040444158017635, + 0.009904079139232635, + 0.007665650453418493, + 0.004290150478482246, + 0.02041289210319519, + 0.006144446786493063, + 0.013800188899040222, + 0.008639082312583923, + 0.013675982132554054, + 0.010111358016729355, + 0.03516647592186928, + 0.004675250966101885, + 0.006974140647798777, + 0.012599463574588299, + 0.00980494823306799, + 0.008423713967204094, + 0.008442454971373081, + 0.008783089928328991, + 0.007286871317774057, + 0.004471589345484972 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 8 + ], + "selected_reward": -0.041719112545251846 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are two biomechanical creatures visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the environment an abandoned factory with a corroded metal floor covered in rust and oil?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the creatures have bodies composed of mirror-polished chrome with visible internal structures?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is one creature resembling a scorpion with snapping hydraulic jaws and the other resembling a spider with arc cutters emitting blue-white electric light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera perform a fluid, unbroken 360-degree pan?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.2520889639854431 + ], + "opposite_logprob_tokens": [ + -1.502089023590088 + ], + "logprob_reward": -0.2520889639854431, + "yes_no_margin_reward": 1.25, + "token_credit_reward": -0.2520889639854431, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06457201391458511, + 0.04985750466585159, + 0.04336094483733177, + 0.2869647443294525, + 0.06169081851840019, + 0.04004466533660889, + 0.06928262114524841, + 0.06921835243701935, + 0.14566436409950256, + 0.16934403777122498 + ], + "credit_map": [ + 0.006070996169000864, + 0.0082640890032053, + 0.004897295497357845, + 0.007640824653208256, + 0.007409528363496065, + 0.007033699192106724, + 0.01101514883339405, + 0.012240425683557987, + 0.006388891022652388, + 0.008190431632101536, + 0.0040961806662380695, + 0.006166614126414061, + 0.00875505618751049, + 0.006137617398053408, + 0.0048577082343399525, + 0.005265002138912678, + 0.003989763557910919, + 0.00509409187361598, + 0.004676017444580793, + 0.009465856477618217, + 0.0043616327457129955, + 0.004525582306087017, + 0.006655813194811344, + 0.0045921895653009415, + 0.03740853816270828, + 0.03282840922474861, + 0.04240959882736206, + 0.03633308410644531, + 0.03684653714299202, + 0.04226672649383545, + 0.030693043023347855, + 0.028178809210658073, + 0.011165084317326546, + 0.005303855519741774, + 0.0052502877078950405, + 0.009035762399435043, + 0.00878144707530737, + 0.011097545735538006, + 0.004368837922811508, + 0.006687996909022331, + 0.003970881924033165, + 0.003483148291707039, + 0.004289868753403425, + 0.004311034455895424, + 0.006929890252649784, + 0.004412278067320585, + 0.0047673010267317295, + 0.007880262099206448, + 0.009489339776337147, + 0.007626038510352373, + 0.007207963149994612, + 0.010630113072693348, + 0.009417183697223663, + 0.005681204609572887, + 0.01133015938103199, + 0.007900618016719818, + 0.015338131226599216, + 0.008568517863750458, + 0.011139024049043655, + 0.0061328900046646595, + 0.006946013309061527, + 0.006903108209371567, + 0.007909908890724182, + 0.006280757952481508, + 0.02952384762465954, + 0.006321777123957872, + 0.017490629106760025, + 0.010308129712939262, + 0.015098214149475098, + 0.010954286903142929, + 0.048973288387060165, + 0.006994199939072132, + 0.02090560458600521, + 0.031624771654605865, + 0.01920238696038723, + 0.01957874745130539, + 0.021257543936371803, + 0.02210002765059471, + 0.018457094207406044, + 0.01621786504983902 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 9, + 8 + ], + "selected_reward": -0.2520889639854431 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are two biomechanical creatures visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the environment an abandoned factory with a corroded metal floor covered in rust and oil?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the creatures have bodies composed of mirror-polished chrome with visible internal structures?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is one creature resembling a scorpion with snapping hydraulic jaws and the other resembling a spider with arc cutters emitting blue-white electric light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera perform a fluid, unbroken 360-degree pan?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.018193742260336876 + ], + "opposite_logprob_tokens": [ + -4.01819372177124 + ], + "logprob_reward": -0.018193742260336876, + "yes_no_margin_reward": 4.0, + "token_credit_reward": -0.018193742260336876, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07918426394462585, + 0.05071005970239639, + 0.058852292597293854, + 0.10450802743434906, + 0.14654020965099335, + 0.0427534244954586, + 0.09102732688188553, + 0.06006324663758278, + 0.29101231694221497, + 0.07534882426261902 + ], + "credit_map": [ + 0.013758271932601929, + 0.009455107152462006, + 0.008455571718513966, + 0.009896054863929749, + 0.009896899573504925, + 0.0073148831725120544, + 0.009757925756275654, + 0.010649556294083595, + 0.0049131219275295734, + 0.005644075572490692, + 0.0054894257336854935, + 0.008668950758874416, + 0.005751985125243664, + 0.005302155390381813, + 0.007352791260927916, + 0.007587553933262825, + 0.005035306792706251, + 0.005930321291089058, + 0.0067300801165401936, + 0.013450521044433117, + 0.005522898864001036, + 0.005278111435472965, + 0.009719731286168098, + 0.007185325492173433, + 0.013709680177271366, + 0.01519144605845213, + 0.02025921829044819, + 0.015070337802171707, + 0.009336831979453564, + 0.011144503951072693, + 0.008687152527272701, + 0.011108862236142159, + 0.027175839990377426, + 0.010976802557706833, + 0.01393001526594162, + 0.018487123772501945, + 0.0183990690857172, + 0.029970914125442505, + 0.010879662819206715, + 0.01672079786658287, + 0.003511162707582116, + 0.0029465865809470415, + 0.005110561847686768, + 0.004117591772228479, + 0.008590222336351871, + 0.0035307309590280056, + 0.0027181331533938646, + 0.012228438630700111, + 0.008713419549167156, + 0.011653032153844833, + 0.007710704579949379, + 0.009954455308616161, + 0.020597493276000023, + 0.00917621236294508, + 0.013453912921249866, + 0.009768098592758179, + 0.011273047886788845, + 0.006661977153271437, + 0.006929271854460239, + 0.005522055551409721, + 0.008469751104712486, + 0.010275265201926231, + 0.006368166301399469, + 0.0045637148432433605, + 0.05596839264035225, + 0.013136508874595165, + 0.0379762277007103, + 0.021017538383603096, + 0.025256944820284843, + 0.020129473879933357, + 0.1044948548078537, + 0.013032350689172745, + 0.009934429079294205, + 0.008688130415976048, + 0.00775088369846344, + 0.007303973659873009, + 0.01011946052312851, + 0.01411482971161604, + 0.008224776946008205, + 0.009212348610162735 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 4, + 3 + ], + "selected_reward": -0.018193742260336876 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are two biomechanical creatures visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the environment an abandoned factory with a corroded metal floor covered in rust and oil?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the creatures have bodies composed of mirror-polished chrome with visible internal structures?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is one creature resembling a scorpion with snapping hydraulic jaws and the other resembling a spider with arc cutters emitting blue-white electric light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera perform a fluid, unbroken 360-degree pan?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.012543829157948494 + ], + "opposite_logprob_tokens": [ + -4.387543678283691 + ], + "logprob_reward": -0.012543829157948494, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.012543829157948494, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1392037570476532, + 0.02777256816625595, + 0.08591146767139435, + 0.1320292204618454, + 0.1351805478334427, + 0.028924092650413513, + 0.0791362002491951, + 0.08046529442071915, + 0.23911266028881073, + 0.05226423218846321 + ], + "credit_map": [ + 0.01470741257071495, + 0.01959661953151226, + 0.016072362661361694, + 0.017613759264349937, + 0.021974094212055206, + 0.014377640560269356, + 0.016679568216204643, + 0.0181823018938303, + 0.0033324265386909246, + 0.002746348734945059, + 0.002701570512726903, + 0.004847164731472731, + 0.003920366987586021, + 0.003230924718081951, + 0.003635717323049903, + 0.0033580486197024584, + 0.007737151347100735, + 0.00878957100212574, + 0.010383340530097485, + 0.00916989054530859, + 0.009663446806371212, + 0.007597469724714756, + 0.01576521061360836, + 0.016805386170744896, + 0.015695199370384216, + 0.019034532830119133, + 0.02199905924499035, + 0.017785020172595978, + 0.01417635940015316, + 0.016913529485464096, + 0.012849105522036552, + 0.013576407916843891, + 0.014683899469673634, + 0.006989107467234135, + 0.026514412835240364, + 0.019516294822096825, + 0.009220310486853123, + 0.015970291569828987, + 0.006662219297140837, + 0.03562401235103607, + 0.0025563070084899664, + 0.0026295785792171955, + 0.003578880801796913, + 0.004777288064360619, + 0.00472251046448946, + 0.003635152941569686, + 0.0028646846767514944, + 0.004159688483923674, + 0.008042051456868649, + 0.014976751059293747, + 0.0048119137063622475, + 0.004731518682092428, + 0.02308613434433937, + 0.0076125250197947025, + 0.00948562752455473, + 0.0063896807841956615, + 0.011410899460315704, + 0.009826049208641052, + 0.007774257101118565, + 0.007953339256346226, + 0.012185770086944103, + 0.012892646715044975, + 0.009982939809560776, + 0.008439391851425171, + 0.052688222378492355, + 0.01019246969372034, + 0.026887094601988792, + 0.017409544438123703, + 0.017341524362564087, + 0.015035558491945267, + 0.08885154873132706, + 0.010706694796681404, + 0.006316327955573797, + 0.00972026027739048, + 0.005975120235234499, + 0.005838074721395969, + 0.0067389653995633125, + 0.008291545324027538, + 0.005648056976497173, + 0.0037358838599175215 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 0, + 4 + ], + "selected_reward": -0.012543829157948494 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are two biomechanical creatures visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the environment an abandoned factory with a corroded metal floor covered in rust and oil?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the creatures have bodies composed of mirror-polished chrome with visible internal structures?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is one creature resembling a scorpion with snapping hydraulic jaws and the other resembling a spider with arc cutters emitting blue-white electric light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera perform a fluid, unbroken 360-degree pan?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.2520432770252228 + ], + "opposite_logprob_tokens": [ + -1.5020432472229004 + ], + "logprob_reward": -0.2520432770252228, + "yes_no_margin_reward": 1.25, + "token_credit_reward": -0.2520432770252228, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11299489438533783, + 0.07046234607696533, + 0.06131592020392418, + 0.032250311225652695, + 0.19255581498146057, + 0.0877729207277298, + 0.06035640090703964, + 0.09950632601976395, + 0.20507387816905975, + 0.07771120220422745 + ], + "credit_map": [ + 0.011251064017415047, + 0.017029820010066032, + 0.013194720260798931, + 0.015525732189416885, + 0.01311211846768856, + 0.013359026052057743, + 0.014851978980004787, + 0.014670439064502716, + 0.009113313630223274, + 0.007983515039086342, + 0.007352161221206188, + 0.010037331841886044, + 0.011557037010788918, + 0.007494059856981039, + 0.00764040369540453, + 0.009284528903663158, + 0.00708881625905633, + 0.0064963698387146, + 0.008327820338308811, + 0.011839784681797028, + 0.008397866040468216, + 0.00565387262031436, + 0.007388741709291935, + 0.006122644990682602, + 0.004257394000887871, + 0.004475853871554136, + 0.003995607607066631, + 0.0038911502342671156, + 0.003953985869884491, + 0.0041810497641563416, + 0.004418616183102131, + 0.0030766536947339773, + 0.025548456236720085, + 0.012219858355820179, + 0.03384507820010185, + 0.013122210279107094, + 0.016189562156796455, + 0.036699097603559494, + 0.012243454344570637, + 0.042688097804784775, + 0.008466212078928947, + 0.0074219550006091595, + 0.010940080508589745, + 0.016602667048573494, + 0.011062911711633205, + 0.009343902580440044, + 0.014804646372795105, + 0.009130542166531086, + 0.005688507109880447, + 0.012865014374256134, + 0.005706846714019775, + 0.006054708268493414, + 0.013958442956209183, + 0.004350090399384499, + 0.008084115572273731, + 0.0036486689932644367, + 0.013310416601598263, + 0.012571175582706928, + 0.016854241490364075, + 0.009190799668431282, + 0.011996353976428509, + 0.011610559187829494, + 0.011029821820557117, + 0.012942954897880554, + 0.04345235228538513, + 0.007974553853273392, + 0.025610074400901794, + 0.015279542654752731, + 0.014805382117629051, + 0.011674032546579838, + 0.07569772750139236, + 0.010580205358564854, + 0.00975237786769867, + 0.012846077792346478, + 0.011402442120015621, + 0.007057661190629005, + 0.009755007922649384, + 0.010727142915129662, + 0.008028906770050526, + 0.008141581900417805 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 4, + 0 + ], + "selected_reward": -0.2520432770252228 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "The camera pans in a fluid, unbroken 360-degree pan, capturing a fierce confrontation deep within an abandoned factory from a slightly upward angle. Two biomechanical creatures, a fusion of organic form and cold machinery, engage in a desperate struggle on a corroded metal floor covered in rust and oil. Their bodies are composed of mirror-polished chrome, their intricate internal structures visible through their skeletal shells. Countless precisely meshing gears whirr as they move, emitting a cold metallic gleam. One of the creatures resembles a scorpion, its massive hydraulic jaws snapping shut with immense force. The other, more like a multi-legged spider, its forelimbs terminating in arc cutters that sizzle and flash with a menacing blue-white electric light. As the claws and cutters collide violently, a blinding storm of sparks erupts. These flaming particles briefly illuminate the surrounding darkness, casting a frenzy of light and shadow across their mirror-polished chrome shells and illuminating the massive silhouettes of rusted chains and abandoned machinery hanging in the background. The entire scene is full of primitive violent aesthetics and iconic biomechanical art style, creating a tense, futuristic and destructive atmosphere of conflict amidst the harsh screams of metal and the crackling of electric arcs.", + "seed": 8515, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000101/group_14/index_7182_seed_8515_dp_29_rank_58_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 14, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.16941669546067714, + "vlm_logprob": -0.16941669546067714, + "vlm_yes_no_margin": 2.975, + "vlm_token_credit": -0.16941669546067714, + "vlm_yes_ratio": 0.8, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are two biomechanical creatures visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the environment an abandoned factory with a corroded metal floor covered in rust and oil?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the creatures have bodies composed of mirror-polished chrome with visible internal structures?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is one creature resembling a scorpion with snapping hydraulic jaws and the other resembling a spider with arc cutters emitting blue-white electric light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera perform a fluid, unbroken 360-degree pan?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.021831780672073364 + ], + "opposite_logprob_tokens": [ + -4.27183198928833 + ], + "logprob_reward": -0.021831780672073364, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.021831780672073364, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15392164885997772, + 0.04250504449009895, + 0.04977889358997345, + 0.07424061000347137, + 0.05438129976391792, + 0.04764823243021965, + 0.14247000217437744, + 0.02421383745968342, + 0.049339767545461655, + 0.36150071024894714 + ], + "credit_map": [ + 0.016680166125297546, + 0.029478833079338074, + 0.01421500276774168, + 0.015008860267698765, + 0.015793869271874428, + 0.020617738366127014, + 0.019957341253757477, + 0.02216983586549759, + 0.004441309720277786, + 0.008401378057897091, + 0.004412911832332611, + 0.004076669458299875, + 0.0072036925703287125, + 0.005355993285775185, + 0.004985760431736708, + 0.0036273275036364794, + 0.0107799107208848, + 0.005627086386084557, + 0.008145244792103767, + 0.0044237170368433, + 0.004901430569589138, + 0.006415079347789288, + 0.004027631599456072, + 0.0054587931372225285, + 0.009759179316461086, + 0.010477893985807896, + 0.00770477345213294, + 0.013687808997929096, + 0.009709318168461323, + 0.008619844913482666, + 0.008633055724203587, + 0.0056487396359443665, + 0.008339601568877697, + 0.00827983021736145, + 0.005751246120780706, + 0.005913582630455494, + 0.009156299754977226, + 0.008327463641762733, + 0.0037845680490136147, + 0.004828709177672863, + 0.006731570698320866, + 0.008331838995218277, + 0.004188256338238716, + 0.0039438181556761265, + 0.004622805863618851, + 0.008164992555975914, + 0.004252803977578878, + 0.007412146311253309, + 0.013830234296619892, + 0.017644425854086876, + 0.013612598180770874, + 0.015477120876312256, + 0.0294235497713089, + 0.01851646415889263, + 0.01713954098522663, + 0.016826072707772255, + 0.0033607243094593287, + 0.003004611935466528, + 0.0019762369338423014, + 0.001968943979591131, + 0.004026343114674091, + 0.0037042913027107716, + 0.0030714336317032576, + 0.0031012529507279396, + 0.007862521335482597, + 0.005121151451021433, + 0.004253535997122526, + 0.005718317348510027, + 0.0054626320488750935, + 0.005928723141551018, + 0.00915559008717537, + 0.0058372970670461655, + 0.03141556680202484, + 0.03398534283041954, + 0.036475952714681625, + 0.03257399797439575, + 0.056108083575963974, + 0.05070815980434418, + 0.08189024776220322, + 0.038343336433172226 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 0, + 6 + ], + "selected_reward": -0.021831780672073364 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are two biomechanical creatures visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the environment an abandoned factory with a corroded metal floor covered in rust and oil?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the creatures have bodies composed of mirror-polished chrome with visible internal structures?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is one creature resembling a scorpion with snapping hydraulic jaws and the other resembling a spider with arc cutters emitting blue-white electric light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera perform a fluid, unbroken 360-degree pan?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.10030320286750793 + ], + "opposite_logprob_tokens": [ + -2.3503031730651855 + ], + "logprob_reward": -0.10030320286750793, + "yes_no_margin_reward": 2.25, + "token_credit_reward": -0.10030320286750793, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11037340015172958, + 0.18513882160186768, + 0.09912315011024475, + 0.05157032981514931, + 0.15176081657409668, + 0.03152916580438614, + 0.036377351731061935, + 0.04207366332411766, + 0.05252829194068909, + 0.2395249754190445 + ], + "credit_map": [ + 0.018369050696492195, + 0.0074365390464663506, + 0.02180217020213604, + 0.013353216461837292, + 0.010989170521497726, + 0.01628652587532997, + 0.01051369309425354, + 0.011623038910329342, + 0.019162381067872047, + 0.026900453492999077, + 0.017303241416811943, + 0.017292581498622894, + 0.037748079746961594, + 0.030757462605834007, + 0.02562023513019085, + 0.010354396887123585, + 0.015427302569150925, + 0.015316897071897984, + 0.01518816128373146, + 0.008876485750079155, + 0.014861787669360638, + 0.01554047130048275, + 0.00711073586717248, + 0.006801310461014509, + 0.009183753281831741, + 0.0067085870541632175, + 0.0055520241148769855, + 0.005479930434376001, + 0.0058897798880934715, + 0.0103225689381361, + 0.003577715717256069, + 0.004855972249060869, + 0.024580327793955803, + 0.014921324327588081, + 0.01672874018549919, + 0.01626317948102951, + 0.029792554676532745, + 0.027402427047491074, + 0.008787735365331173, + 0.01328453328460455, + 0.0034589136485010386, + 0.0045409975573420525, + 0.003973506856709719, + 0.0033852679189294577, + 0.003610375104472041, + 0.004981797654181719, + 0.0036472517531365156, + 0.003931055776774883, + 0.003986184019595385, + 0.004900326021015644, + 0.003376588225364685, + 0.004048108123242855, + 0.00683291582390666, + 0.004783719312399626, + 0.0038348252419382334, + 0.004614683333784342, + 0.006629663519561291, + 0.004348040092736483, + 0.003923819866031408, + 0.0032847414258867502, + 0.006369445007294416, + 0.005672353319823742, + 0.004384496249258518, + 0.0074611022137105465, + 0.004170539788901806, + 0.004701321944594383, + 0.005812830291688442, + 0.004593761172145605, + 0.006586082279682159, + 0.006191513035446405, + 0.011315290816128254, + 0.009156951680779457, + 0.01900300942361355, + 0.02611379511654377, + 0.028056733310222626, + 0.020853634923696518, + 0.028704481199383736, + 0.031177472323179245, + 0.04332859069108963, + 0.04228724539279938 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 1, + 4 + ], + "selected_reward": -0.10030320286750793 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are two biomechanical creatures visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the environment an abandoned factory with a corroded metal floor covered in rust and oil?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the creatures have bodies composed of mirror-polished chrome with visible internal structures?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is one creature resembling a scorpion with snapping hydraulic jaws and the other resembling a spider with arc cutters emitting blue-white electric light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera perform a fluid, unbroken 360-degree pan?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.020580438897013664 + ], + "opposite_logprob_tokens": [ + -3.895580530166626 + ], + "logprob_reward": -0.020580438897013664, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.020580438897013664, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.21032921969890594, + 0.08980001509189606, + 0.09178711473941803, + 0.057829003781080246, + 0.06694263964891434, + 0.06825004518032074, + 0.06794828176498413, + 0.12048641592264175, + 0.09399573504924774, + 0.13263149559497833 + ], + "credit_map": [ + 0.02684359811246395, + 0.02892794832587242, + 0.02201218716800213, + 0.022887635976076126, + 0.025113528594374657, + 0.031945280730724335, + 0.024357452988624573, + 0.028241578489542007, + 0.009570452384650707, + 0.014154912903904915, + 0.012776446528732777, + 0.009377655573189259, + 0.010527577251195908, + 0.012096293270587921, + 0.011155012995004654, + 0.010141681879758835, + 0.013656760565936565, + 0.012129568494856358, + 0.013724613934755325, + 0.00888112559914589, + 0.00981918815523386, + 0.016343066468834877, + 0.008397608995437622, + 0.008835189044475555, + 0.007329313084483147, + 0.008685038425028324, + 0.004879173822700977, + 0.0076098693534731865, + 0.0064835138618946075, + 0.00805926974862814, + 0.005633693188428879, + 0.009149134159088135, + 0.012694903649389744, + 0.006856875494122505, + 0.007130156736820936, + 0.007583478465676308, + 0.006615631747990847, + 0.013409212231636047, + 0.00522812083363533, + 0.007424268871545792, + 0.007392771542072296, + 0.009045479819178581, + 0.007422263268381357, + 0.005793343763798475, + 0.008945888839662075, + 0.013048475608229637, + 0.007956051267683506, + 0.008645772002637386, + 0.007588159292936325, + 0.013448614627122879, + 0.005115917418152094, + 0.007623740471899509, + 0.010431796312332153, + 0.00812321063131094, + 0.006658133585005999, + 0.008958709426224232, + 0.01582086645066738, + 0.012537609785795212, + 0.011011284776031971, + 0.008490607142448425, + 0.02078983001410961, + 0.017929155379533768, + 0.013012903742492199, + 0.020894169807434082, + 0.009695060551166534, + 0.008950470946729183, + 0.00961250625550747, + 0.011553240939974785, + 0.012328648939728737, + 0.012440088205039501, + 0.017054347321391106, + 0.012361375615000725, + 0.014622310176491737, + 0.013114366680383682, + 0.013455658219754696, + 0.012172943912446499, + 0.02125203050673008, + 0.01950198784470558, + 0.025533076375722885, + 0.01297912746667862 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 7 + ], + "selected_reward": -0.020580438897013664 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are two biomechanical creatures visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the environment an abandoned factory with a corroded metal floor covered in rust and oil?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the creatures have bodies composed of mirror-polished chrome with visible internal structures?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is one creature resembling a scorpion with snapping hydraulic jaws and the other resembling a spider with arc cutters emitting blue-white electric light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera perform a fluid, unbroken 360-degree pan?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.011076509952545166 + ], + "opposite_logprob_tokens": [ + -4.5110764503479 + ], + "logprob_reward": -0.011076509952545166, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.011076509952545166, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.18153108656406403, + 0.03717781975865364, + 0.032315876334905624, + 0.05614234134554863, + 0.062301140278577805, + 0.048557452857494354, + 0.04653235897421837, + 0.03428920358419418, + 0.06427456438541412, + 0.43687811493873596 + ], + "credit_map": [ + 0.013848775066435337, + 0.025219367817044258, + 0.019994376227259636, + 0.020988356322050095, + 0.02203880250453949, + 0.025687893852591515, + 0.027276037260890007, + 0.026477470993995667, + 0.004403665661811829, + 0.007892186753451824, + 0.0036044300068169832, + 0.00372407422401011, + 0.005818216595798731, + 0.005851357709616423, + 0.003453408367931843, + 0.0024304811377078295, + 0.004556872881948948, + 0.0037654894404113293, + 0.003720410168170929, + 0.0034598212223500013, + 0.003993805032223463, + 0.004603953566402197, + 0.0034695174545049667, + 0.0047460091300308704, + 0.006831469014286995, + 0.007214836776256561, + 0.005242126062512398, + 0.009150346741080284, + 0.0075097400695085526, + 0.008875270374119282, + 0.006404065527021885, + 0.004914489109069109, + 0.009718303568661213, + 0.007474424783140421, + 0.00823281891644001, + 0.009488758631050587, + 0.009710187092423439, + 0.008793000131845474, + 0.003814386436715722, + 0.005069267004728317, + 0.005391156300902367, + 0.005024292040616274, + 0.004790863487869501, + 0.005281249992549419, + 0.005286107771098614, + 0.009075487963855267, + 0.005717868451029062, + 0.007990428246557713, + 0.005233167205005884, + 0.00801281351596117, + 0.004107248038053513, + 0.004532018210738897, + 0.00715707428753376, + 0.005950325168669224, + 0.005616463720798492, + 0.005923252087086439, + 0.004652324598282576, + 0.0036144829355180264, + 0.0038979132659733295, + 0.002945252461358905, + 0.0047817351296544075, + 0.0048464396968483925, + 0.00386073999106884, + 0.005690313410013914, + 0.007835454307496548, + 0.007763713598251343, + 0.006577301304787397, + 0.007205360569059849, + 0.010752331465482712, + 0.00829606968909502, + 0.008599814958870411, + 0.007244514767080545, + 0.03829764202237129, + 0.04263903200626373, + 0.04212507605552673, + 0.039187293499708176, + 0.06890621781349182, + 0.061725325882434845, + 0.09864713251590729, + 0.04535038396716118 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 0, + 8 + ], + "selected_reward": -0.011076509952545166 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are two biomechanical creatures visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the environment an abandoned factory with a corroded metal floor covered in rust and oil?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the creatures have bodies composed of mirror-polished chrome with visible internal structures?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is one creature resembling a scorpion with snapping hydraulic jaws and the other resembling a spider with arc cutters emitting blue-white electric light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera perform a fluid, unbroken 360-degree pan?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -0.6932915449142456 + ], + "opposite_logprob_tokens": [ + -0.6932915449142456 + ], + "logprob_reward": -0.6932915449142456, + "yes_no_margin_reward": 0.0, + "token_credit_reward": -0.6932915449142456, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.19652888178825378, + 0.07490795105695724, + 0.09873780608177185, + 0.04986598342657089, + 0.09844518452882767, + 0.06696173548698425, + 0.03915223851799965, + 0.09877586364746094, + 0.14934247732162476, + 0.12728190422058105 + ], + "credit_map": [ + 0.026677748188376427, + 0.020511183887720108, + 0.02282533422112465, + 0.021700050681829453, + 0.02037055231630802, + 0.03134899213910103, + 0.026856012642383575, + 0.02623898908495903, + 0.0073533011600375175, + 0.0073584397323429585, + 0.00871837604790926, + 0.008828860707581043, + 0.015071420930325985, + 0.012406822293996811, + 0.008622907102108002, + 0.00654782447963953, + 0.01414440292865038, + 0.014158564619719982, + 0.009587759152054787, + 0.009541084989905357, + 0.012485027313232422, + 0.017841774970293045, + 0.010028119198977947, + 0.010951078496873379, + 0.00836646556854248, + 0.006907913368195295, + 0.004305030684918165, + 0.005553253460675478, + 0.006050078663975, + 0.0078009567223489285, + 0.004286903887987137, + 0.00659538060426712, + 0.014003737829625607, + 0.011569137685000896, + 0.014340545050799847, + 0.014854956418275833, + 0.011585072614252567, + 0.015546182170510292, + 0.006480340845882893, + 0.010065215639770031, + 0.005916749592870474, + 0.008178584277629852, + 0.008354517631232738, + 0.005714119877666235, + 0.007774813566356897, + 0.012914326041936874, + 0.008969249203801155, + 0.009139377623796463, + 0.004273728933185339, + 0.006099450867623091, + 0.004609917290508747, + 0.00502249039709568, + 0.004690668545663357, + 0.00463504996150732, + 0.004505772143602371, + 0.005315160378813744, + 0.012966763228178024, + 0.010840792208909988, + 0.009159333072602749, + 0.0075638252310454845, + 0.019552189856767654, + 0.014510301873087883, + 0.009163999930024147, + 0.015018654055893421, + 0.022713763639330864, + 0.01827511005103588, + 0.013219385407865047, + 0.017861954867839813, + 0.024844259023666382, + 0.019727347418665886, + 0.018544849008321762, + 0.014155794866383076, + 0.011701730079948902, + 0.0124176274985075, + 0.01180867850780487, + 0.011234316974878311, + 0.019697563722729683, + 0.017747510224580765, + 0.02825835347175598, + 0.014416136778891087 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 9 + ], + "selected_reward": -0.6932915449142456 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "No" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000101/group_15/index_tensor([4454, 4454])_group_15_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000101/group_15/index_tensor([4454, 4454])_group_15_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..8f7a6ca7a68e57362382b37647c0fed31d884cc9 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000101/group_15/index_tensor([4454, 4454])_group_15_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "shot that begins in the medium shot, gently tilting upward and slowly zooming out, transports the viewer to the summit of a vast, ever-bending sea of clouds. A sacred phoenix perches majestically atop a lone peak piercing the clouds. Its feathers, like flowing flames, interwoven with threads of pure gold, gleam in the golden light of dawn. Its crest is as resplendent as gemstones, and its deep eyes gleam with ancient wisdom. Cautiously, it lowers its head, and with a beak that seems carved from warm jade, it gently but steadily grasps one end of a scroll made of aged, yellowed silk. The scroll's ends are framed by dark wooden shafts intricately carved with cloud patterns. As the phoenix's unhurried movements unfold, the scroll slowly unfurls, soft, multicolored streams of light spilling from its surface. The fantastical Eastern world depicted—tiered mountain ranges, winding rivers, and mist-shrouded pavilions—doesn't simply appear, but gradually coalesces and takes shape within these shimmering particles of light, as if a dormant dimension is awakening. The entire scene is bathed in a magnificent beam of light that pierces the clouds, gilding the phoenix's silhouette with a sacred gilt. The dominant tones of the painting are warm gold and calming cloud white, punctuated by the phoenix's fiery hues and the vibrant brilliance of the scroll, creating an atmosphere of grandeur, tranquility, and boundless wonder—an Eastern fantasy.", + "seed": 26988, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000101/group_15/index_4454_seed_26988_dp_30_rank_60_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 15, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.0342142466455698, + "vlm_logprob": -0.0342142466455698, + "vlm_yes_no_margin": 3.95, + "vlm_token_credit": -0.0342142466455698, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the video begin with a medium shot of a phoenix perched on a peak above a sea of clouds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera gently tilt upward and slowly zoom out from the beginning?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the phoenix have feathers resembling flowing flames interwoven with threads of pure gold?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the phoenix lower its head to grasp a scroll made of aged, yellowed silk with dark wooden shafts?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do soft, multicolored streams of light spill from the scroll as it unfurls to reveal a fantastical Eastern world?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.008151122368872166 + ], + "opposite_logprob_tokens": [ + -5.133151054382324 + ], + "logprob_reward": -0.008151122368872166, + "yes_no_margin_reward": 5.125, + "token_credit_reward": -0.008151122368872166, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10703089833259583, + 0.05162398889660835, + 0.05221118405461311, + 0.09499476104974747, + 0.14314334094524384, + 0.06919649243354797, + 0.0732816755771637, + 0.21091172099113464, + 0.14905549585819244, + 0.04855044558644295 + ], + "credit_map": [ + 0.018065769225358963, + 0.014081008732318878, + 0.011961347423493862, + 0.01129424013197422, + 0.01748872548341751, + 0.011689645238220692, + 0.012106196023523808, + 0.010343967005610466, + 0.007235456723719835, + 0.00673253508284688, + 0.00674647931009531, + 0.004779004026204348, + 0.008299775421619415, + 0.006710069719702005, + 0.005743823945522308, + 0.005376847460865974, + 0.00808587484061718, + 0.008580083958804607, + 0.010265004821121693, + 0.0029749313835054636, + 0.003794089425355196, + 0.008883941918611526, + 0.003414538223296404, + 0.0062127127312123775, + 0.015764810144901276, + 0.012763244099915028, + 0.010845431126654148, + 0.007989323697984219, + 0.009149790741503239, + 0.014840221963822842, + 0.009097939357161522, + 0.014543995261192322, + 0.028149571269750595, + 0.02459304966032505, + 0.018801188096404076, + 0.008198928087949753, + 0.0099313510581851, + 0.024777868762612343, + 0.010283125564455986, + 0.018408270552754402, + 0.008903256617486477, + 0.007934394292533398, + 0.010037976317107677, + 0.005980683024972677, + 0.008407090790569782, + 0.007108164019882679, + 0.009475800208747387, + 0.011349131353199482, + 0.010614045895636082, + 0.007086183410137892, + 0.008722479455173016, + 0.006203101482242346, + 0.00747031020000577, + 0.015806201845407486, + 0.009733065962791443, + 0.007646289188414812, + 0.026828939095139503, + 0.023912139236927032, + 0.018169015645980835, + 0.022731713950634003, + 0.04107155278325081, + 0.03563840687274933, + 0.02348930388689041, + 0.019070642068982124, + 0.011991164647042751, + 0.01192628312855959, + 0.009295533411204815, + 0.03758483752608299, + 0.01328345201909542, + 0.0471784770488739, + 0.008677125908434391, + 0.009118625894188881, + 0.005517581943422556, + 0.007589962333440781, + 0.004171346314251423, + 0.006007961463183165, + 0.007721065543591976, + 0.006767876446247101, + 0.005869959481060505, + 0.0049046906642615795 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 8, + 4 + ], + "selected_reward": -0.008151122368872166 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the video begin with a medium shot of a phoenix perched on a peak above a sea of clouds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera gently tilt upward and slowly zoom out from the beginning?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the phoenix have feathers resembling flowing flames interwoven with threads of pure gold?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the phoenix lower its head to grasp a scroll made of aged, yellowed silk with dark wooden shafts?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do soft, multicolored streams of light spill from the scroll as it unfurls to reveal a fantastical Eastern world?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.10030622780323029 + ], + "opposite_logprob_tokens": [ + -2.350306272506714 + ], + "logprob_reward": -0.10030622780323029, + "yes_no_margin_reward": 2.25, + "token_credit_reward": -0.10030622780323029, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12198112905025482, + 0.05336013808846474, + 0.022774705663323402, + 0.03110816515982151, + 0.0796823799610138, + 0.049340471625328064, + 0.06552290171384811, + 0.06106201931834221, + 0.41989779472351074, + 0.0952703207731247 + ], + "credit_map": [ + 0.016472872346639633, + 0.021597955375909805, + 0.017615055665373802, + 0.009933370165526867, + 0.025351759046316147, + 0.012582561001181602, + 0.0095286900177598, + 0.008898861706256866, + 0.007587716449052095, + 0.007835171185433865, + 0.010647324845194817, + 0.004791766870766878, + 0.002918141195550561, + 0.00861575361341238, + 0.0048483628779649734, + 0.0061159017495810986, + 0.0028515427839010954, + 0.003294450929388404, + 0.0036640025209635496, + 0.0026006426196545362, + 0.002866677939891815, + 0.003188728354871273, + 0.0019751484505832195, + 0.0023335108999162912, + 0.0044833761639893055, + 0.003858453594148159, + 0.004517833702266216, + 0.002904962981119752, + 0.003844866529107094, + 0.004955971613526344, + 0.0029846273828297853, + 0.0035580722615122795, + 0.006391582079231739, + 0.017903286963701248, + 0.01932520791888237, + 0.005801417399197817, + 0.007121680770069361, + 0.00707363523542881, + 0.006467888131737709, + 0.009597686119377613, + 0.007248678710311651, + 0.005229454953223467, + 0.005225120112299919, + 0.004417353309690952, + 0.006007006391882896, + 0.011106793768703938, + 0.005400181747972965, + 0.004705886822193861, + 0.0124427555128932, + 0.006554336752742529, + 0.0078096636570990086, + 0.006059897597879171, + 0.008259027265012264, + 0.01086066197603941, + 0.006609139032661915, + 0.006927420850843191, + 0.007790527306497097, + 0.008959321305155754, + 0.0064365314319729805, + 0.006334913428872824, + 0.0078534921631217, + 0.009976222179830074, + 0.005963012110441923, + 0.007747997529804707, + 0.0775141566991806, + 0.0208570696413517, + 0.020456882193684578, + 0.0686241015791893, + 0.022916242480278015, + 0.16269467771053314, + 0.021168675273656845, + 0.02566598169505596, + 0.010560085065662861, + 0.010366613045334816, + 0.006566097028553486, + 0.014818893745541573, + 0.013965523801743984, + 0.013656536117196083, + 0.016732115298509598, + 0.008604456670582294 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 0, + 9 + ], + "selected_reward": -0.10030622780323029 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the video begin with a medium shot of a phoenix perched on a peak above a sea of clouds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera gently tilt upward and slowly zoom out from the beginning?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the phoenix have feathers resembling flowing flames interwoven with threads of pure gold?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the phoenix lower its head to grasp a scroll made of aged, yellowed silk with dark wooden shafts?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do soft, multicolored streams of light spill from the scroll as it unfurls to reveal a fantastical Eastern world?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.009785421192646027 + ], + "opposite_logprob_tokens": [ + -4.6347856521606445 + ], + "logprob_reward": -0.009785421192646027, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.009785421192646027, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1742635816335678, + 0.05980315059423447, + 0.06468650698661804, + 0.08245723694562912, + 0.11119828373193741, + 0.06359241157770157, + 0.15647011995315552, + 0.1338551789522171, + 0.10002293437719345, + 0.05365060269832611 + ], + "credit_map": [ + 0.02347126603126526, + 0.027001624926924706, + 0.02529803104698658, + 0.015006999485194683, + 0.03516080975532532, + 0.018010295927524567, + 0.0148923359811306, + 0.015422221273183823, + 0.013977582566440105, + 0.008390707895159721, + 0.00932165514677763, + 0.005072240252047777, + 0.006082975771278143, + 0.006425034254789352, + 0.0051366123370826244, + 0.005396339576691389, + 0.006997319404035807, + 0.009905662387609482, + 0.00877284724265337, + 0.0073552378453314304, + 0.008352935314178467, + 0.010850661434233189, + 0.005893460474908352, + 0.006558379158377647, + 0.01746392622590065, + 0.011904740706086159, + 0.011894123628735542, + 0.004081309773027897, + 0.00739310123026371, + 0.010320828296244144, + 0.004701858386397362, + 0.014697347767651081, + 0.01915600337088108, + 0.019969120621681213, + 0.025146521627902985, + 0.006989033427089453, + 0.009509644471108913, + 0.009836827404797077, + 0.008152390830218792, + 0.012438738718628883, + 0.009687402285635471, + 0.006082546431571245, + 0.005295887589454651, + 0.0049350680783391, + 0.00765526806935668, + 0.0145245511084795, + 0.008218762464821339, + 0.007192931137979031, + 0.02262716554105282, + 0.014331714250147343, + 0.01653456874191761, + 0.014566680416464806, + 0.02199685387313366, + 0.028102857992053032, + 0.0190002229064703, + 0.019310051575303078, + 0.016012301668524742, + 0.018143506720662117, + 0.012081804685294628, + 0.014274833723902702, + 0.02950556017458439, + 0.017491891980171204, + 0.014441072009503841, + 0.01190420612692833, + 0.013168168254196644, + 0.00684918649494648, + 0.010777008719742298, + 0.020303063094615936, + 0.007154460065066814, + 0.025404132902622223, + 0.007460158318281174, + 0.008906758390367031, + 0.005553571041673422, + 0.007840771228075027, + 0.007535484153777361, + 0.008526107296347618, + 0.0070436736568808556, + 0.005027958191931248, + 0.0078657865524292, + 0.00425725057721138 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 6, + 7 + ], + "selected_reward": -0.009785421192646027 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the video begin with a medium shot of a phoenix perched on a peak above a sea of clouds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera gently tilt upward and slowly zoom out from the beginning?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the phoenix have feathers resembling flowing flames interwoven with threads of pure gold?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the phoenix lower its head to grasp a scroll made of aged, yellowed silk with dark wooden shafts?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do soft, multicolored streams of light spill from the scroll as it unfurls to reveal a fantastical Eastern world?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04305106773972511 + ], + "opposite_logprob_tokens": [ + -3.16805100440979 + ], + "logprob_reward": -0.04305106773972511, + "yes_no_margin_reward": 3.125, + "token_credit_reward": -0.04305106773972511, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.17381037771701813, + 0.03781332075595856, + 0.03678881376981735, + 0.04062756896018982, + 0.06505700945854187, + 0.03772446885704994, + 0.18145966529846191, + 0.16406573355197906, + 0.1945040374994278, + 0.06814903765916824 + ], + "credit_map": [ + 0.024505462497472763, + 0.029063040390610695, + 0.025744598358869553, + 0.01474855188280344, + 0.03278007358312607, + 0.019927386194467545, + 0.011649499647319317, + 0.015391766093671322, + 0.006659019738435745, + 0.005716482177376747, + 0.006050345953553915, + 0.0036014006473124027, + 0.004031238611787558, + 0.0048747677356004715, + 0.0031433606054633856, + 0.003736703423783183, + 0.006224432028830051, + 0.005193241406232119, + 0.004973034840077162, + 0.004009999334812164, + 0.00543413357809186, + 0.004290012642741203, + 0.0033876749221235514, + 0.003276281291618943, + 0.006648376118391752, + 0.00553935905918479, + 0.0056030419655144215, + 0.0032706954516470432, + 0.004373092669993639, + 0.0054528964683413506, + 0.003157936967909336, + 0.006582166533917189, + 0.011807975359261036, + 0.009935179725289345, + 0.009365877136588097, + 0.005559523589909077, + 0.004942968022078276, + 0.010037165135145187, + 0.005583485122770071, + 0.007824832573533058, + 0.005487822461873293, + 0.004181267227977514, + 0.005800389684736729, + 0.0028616348281502724, + 0.0025311496574431658, + 0.006953212432563305, + 0.0030864241998642683, + 0.006822564173489809, + 0.02937294915318489, + 0.014618047513067722, + 0.020118894055485725, + 0.015762709081172943, + 0.025333309546113014, + 0.028544845059514046, + 0.022940298542380333, + 0.024768592789769173, + 0.023847442120313644, + 0.02001233585178852, + 0.013702129945158958, + 0.021101491525769234, + 0.02964024245738983, + 0.024573668837547302, + 0.017054500058293343, + 0.014133911579847336, + 0.016042299568653107, + 0.017042314633727074, + 0.01508263684809208, + 0.04304664582014084, + 0.022697098553180695, + 0.05392376706004143, + 0.013338224962353706, + 0.013331039808690548, + 0.007749590557068586, + 0.011833039112389088, + 0.007832221686840057, + 0.010164289735257626, + 0.010537907481193542, + 0.007291881367564201, + 0.00889738742262125, + 0.003842714009806514 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 6, + 0 + ], + "selected_reward": -0.04305106773972511 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the video begin with a medium shot of a phoenix perched on a peak above a sea of clouds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera gently tilt upward and slowly zoom out from the beginning?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the phoenix have feathers resembling flowing flames interwoven with threads of pure gold?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the phoenix lower its head to grasp a scroll made of aged, yellowed silk with dark wooden shafts?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do soft, multicolored streams of light spill from the scroll as it unfurls to reveal a fantastical Eastern world?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.009777394123375416 + ], + "opposite_logprob_tokens": [ + -4.634777545928955 + ], + "logprob_reward": -0.009777394123375416, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.009777394123375416, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14190466701984406, + 0.048196371644735336, + 0.037529490888118744, + 0.07066570967435837, + 0.1619836539030075, + 0.049656834453344345, + 0.07655570656061172, + 0.2423163652420044, + 0.09933856874704361, + 0.07185260206460953 + ], + "credit_map": [ + 0.023628756403923035, + 0.022635359317064285, + 0.017424924299120903, + 0.011544582433998585, + 0.027264341711997986, + 0.015007710084319115, + 0.013918064534664154, + 0.010480918921530247, + 0.0054819923825562, + 0.007584818173199892, + 0.007831874303519726, + 0.00482975784689188, + 0.006524857599288225, + 0.0053253392688930035, + 0.005967005155980587, + 0.004650727845728397, + 0.00609003659337759, + 0.005706119351089001, + 0.004836113192141056, + 0.002953887917101383, + 0.003927623387426138, + 0.006516691762953997, + 0.0036422775592654943, + 0.003856740426272154, + 0.012889686971902847, + 0.009274745360016823, + 0.009831362403929234, + 0.005061537493020296, + 0.007101608440279961, + 0.010254006832838058, + 0.005699741654098034, + 0.010553021915256977, + 0.03025861829519272, + 0.017582526430487633, + 0.029125235974788666, + 0.01093586441129446, + 0.014254649169743061, + 0.027158223092556, + 0.01374063640832901, + 0.01892790198326111, + 0.005236592143774033, + 0.0055698007345199585, + 0.007782131899148226, + 0.003613894572481513, + 0.0066101220436394215, + 0.007107219193130732, + 0.005032639484852552, + 0.00870443694293499, + 0.012161537073552608, + 0.0073611061088740826, + 0.009772821329534054, + 0.006539337802678347, + 0.008078447543084621, + 0.013889540918171406, + 0.008704682812094688, + 0.010048238560557365, + 0.034640055149793625, + 0.015505796298384666, + 0.03293474018573761, + 0.03144923970103264, + 0.026305172592401505, + 0.04849999397993088, + 0.021920105442404747, + 0.031061256304383278, + 0.015260969288647175, + 0.0066466256976127625, + 0.013093731366097927, + 0.014189455658197403, + 0.007379323709756136, + 0.02589656412601471, + 0.006498289760202169, + 0.010373614728450775, + 0.011003178544342518, + 0.004679183941334486, + 0.008825200609862804, + 0.010862000286579132, + 0.0074806129559874535, + 0.013825545087456703, + 0.0068841353058815, + 0.008292743004858494 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 4, + 0 + ], + "selected_reward": -0.009777394123375416 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "shot that begins in the medium shot, gently tilting upward and slowly zooming out, transports the viewer to the summit of a vast, ever-bending sea of clouds. A sacred phoenix perches majestically atop a lone peak piercing the clouds. Its feathers, like flowing flames, interwoven with threads of pure gold, gleam in the golden light of dawn. Its crest is as resplendent as gemstones, and its deep eyes gleam with ancient wisdom. Cautiously, it lowers its head, and with a beak that seems carved from warm jade, it gently but steadily grasps one end of a scroll made of aged, yellowed silk. The scroll's ends are framed by dark wooden shafts intricately carved with cloud patterns. As the phoenix's unhurried movements unfold, the scroll slowly unfurls, soft, multicolored streams of light spilling from its surface. The fantastical Eastern world depicted—tiered mountain ranges, winding rivers, and mist-shrouded pavilions—doesn't simply appear, but gradually coalesces and takes shape within these shimmering particles of light, as if a dormant dimension is awakening. The entire scene is bathed in a magnificent beam of light that pierces the clouds, gilding the phoenix's silhouette with a sacred gilt. The dominant tones of the painting are warm gold and calming cloud white, punctuated by the phoenix's fiery hues and the vibrant brilliance of the scroll, creating an atmosphere of grandeur, tranquility, and boundless wonder—an Eastern fantasy.", + "seed": 26989, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000101/group_15/index_4454_seed_26989_dp_30_rank_60_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 15, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.017737143859267234, + "vlm_logprob": -0.017737143859267234, + "vlm_yes_no_margin": 4.325, + "vlm_token_credit": -0.017737143859267234, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the video begin with a medium shot of a phoenix perched on a peak above a sea of clouds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera gently tilt upward and slowly zoom out from the beginning?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the phoenix have feathers resembling flowing flames interwoven with threads of pure gold?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the phoenix lower its head to grasp a scroll made of aged, yellowed silk with dark wooden shafts?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do soft, multicolored streams of light spill from the scroll as it unfurls to reveal a fantastical Eastern world?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.006583311595022678 + ], + "opposite_logprob_tokens": [ + -5.381583213806152 + ], + "logprob_reward": -0.006583311595022678, + "yes_no_margin_reward": 5.375, + "token_credit_reward": -0.006583311595022678, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.16062940657138824, + 0.0654737576842308, + 0.06335298717021942, + 0.05501674488186836, + 0.05004703626036644, + 0.06780723482370377, + 0.1344364434480667, + 0.1109047383069992, + 0.10090521723031998, + 0.19142641127109528 + ], + "credit_map": [ + 0.02815563604235649, + 0.015049777925014496, + 0.021623726934194565, + 0.017816860228776932, + 0.015096183866262436, + 0.016055835410952568, + 0.02105896547436714, + 0.025772422552108765, + 0.009765916503965855, + 0.006227925419807434, + 0.009440732188522816, + 0.009573027491569519, + 0.007297878619283438, + 0.007892163470387459, + 0.008353397250175476, + 0.006922714877873659, + 0.010438314639031887, + 0.006117863580584526, + 0.006237744353711605, + 0.00790691003203392, + 0.009231535717844963, + 0.009296932257711887, + 0.006535877473652363, + 0.007587801665067673, + 0.011117693968117237, + 0.005701909773051739, + 0.005593417678028345, + 0.004465210717171431, + 0.007853451184928417, + 0.006819905247539282, + 0.004795490298420191, + 0.008669668808579445, + 0.006208470091223717, + 0.004060407169163227, + 0.005480887833982706, + 0.013441135175526142, + 0.004849941469728947, + 0.00746780913323164, + 0.00470713060349226, + 0.0038312519900500774, + 0.0115018580108881, + 0.0070032645016908646, + 0.005780329927802086, + 0.009415077045559883, + 0.009427440352737904, + 0.01034844946116209, + 0.008733266964554787, + 0.0055975476279854774, + 0.026442086324095726, + 0.01329609751701355, + 0.010484904982149601, + 0.01373708713799715, + 0.02011094242334366, + 0.017597323283553123, + 0.018488416448235512, + 0.01427957322448492, + 0.017418768256902695, + 0.011616280302405357, + 0.009270928800106049, + 0.013610306195914745, + 0.019660117104649544, + 0.011430135928094387, + 0.014232346788048744, + 0.013665847480297089, + 0.015455988235771656, + 0.00869964063167572, + 0.01271017361432314, + 0.008390557952225208, + 0.009017681702971458, + 0.014458933845162392, + 0.01069489773362875, + 0.021477345377206802, + 0.029714783653616905, + 0.013878002762794495, + 0.030111322179436684, + 0.026376236230134964, + 0.014762455597519875, + 0.02487923391163349, + 0.021454649046063423, + 0.030249733477830887 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 0, + 6 + ], + "selected_reward": -0.006583311595022678 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the video begin with a medium shot of a phoenix perched on a peak above a sea of clouds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera gently tilt upward and slowly zoom out from the beginning?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the phoenix have feathers resembling flowing flames interwoven with threads of pure gold?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the phoenix lower its head to grasp a scroll made of aged, yellowed silk with dark wooden shafts?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do soft, multicolored streams of light spill from the scroll as it unfurls to reveal a fantastical Eastern world?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.011086530983448029 + ], + "opposite_logprob_tokens": [ + -4.511086463928223 + ], + "logprob_reward": -0.011086530983448029, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.011086530983448029, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.05901976302266121, + 0.06390953063964844, + 0.07444626837968826, + 0.05322905629873276, + 0.06750569492578506, + 0.06055743247270584, + 0.31792452931404114, + 0.09312152862548828, + 0.17741550505161285, + 0.032870642840862274 + ], + "credit_map": [ + 0.010525739751756191, + 0.0056776930578053, + 0.008167972788214684, + 0.006861379370093346, + 0.006354285404086113, + 0.006545900367200375, + 0.00655719731003046, + 0.008329590782523155, + 0.009198412299156189, + 0.008278445340692997, + 0.006131135858595371, + 0.0072019342333078384, + 0.011611120775341988, + 0.007714944425970316, + 0.0074773915112018585, + 0.006296153180301189, + 0.012510121800005436, + 0.00852978602051735, + 0.00791151076555252, + 0.008348974399268627, + 0.013961557298898697, + 0.008572629652917385, + 0.007632707245647907, + 0.006978983525186777, + 0.008154390379786491, + 0.005296918097883463, + 0.004354990087449551, + 0.0053594643250107765, + 0.012276736088097095, + 0.007529779337346554, + 0.005004401784390211, + 0.005252378061413765, + 0.010422317311167717, + 0.007163143251091242, + 0.00516537157818675, + 0.01096893846988678, + 0.010181153193116188, + 0.009496980346739292, + 0.00676575256511569, + 0.0073420386761426926, + 0.0071244933642446995, + 0.007890673354268074, + 0.004548775497823954, + 0.0055465614423155785, + 0.013945359736680984, + 0.009662340395152569, + 0.006540828850120306, + 0.005298399366438389, + 0.055303335189819336, + 0.036366529762744904, + 0.019807927310466766, + 0.033162858337163925, + 0.055845100432634354, + 0.03876573592424393, + 0.04668724164366722, + 0.03198578581213951, + 0.02288137748837471, + 0.00955058541148901, + 0.007226263638585806, + 0.00847723800688982, + 0.009981222450733185, + 0.01060701347887516, + 0.012170346453785896, + 0.012227479368448257, + 0.03133383393287659, + 0.016780590638518333, + 0.02027996815741062, + 0.018193349242210388, + 0.01565288007259369, + 0.023984378203749657, + 0.02169957384467125, + 0.02949092723429203, + 0.005923902150243521, + 0.0024982825852930546, + 0.0035348935052752495, + 0.004499782808125019, + 0.003982325550168753, + 0.004098119214177132, + 0.004645727574825287, + 0.0036876073572784662 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 8, + 7 + ], + "selected_reward": -0.011086530983448029 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the video begin with a medium shot of a phoenix perched on a peak above a sea of clouds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera gently tilt upward and slowly zoom out from the beginning?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the phoenix have feathers resembling flowing flames interwoven with threads of pure gold?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the phoenix lower its head to grasp a scroll made of aged, yellowed silk with dark wooden shafts?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do soft, multicolored streams of light spill from the scroll as it unfurls to reveal a fantastical Eastern world?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.00978435855358839 + ], + "opposite_logprob_tokens": [ + -4.63478422164917 + ], + "logprob_reward": -0.00978435855358839, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.00978435855358839, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08985982090234756, + 0.11129403114318848, + 0.05334323272109032, + 0.05775400996208191, + 0.04353876784443855, + 0.06686737388372421, + 0.14736583828926086, + 0.07404183596372604, + 0.2141818255186081, + 0.14175330102443695 + ], + "credit_map": [ + 0.016736529767513275, + 0.00795745849609375, + 0.010277592577040195, + 0.00947505235671997, + 0.010673901997506618, + 0.009052414447069168, + 0.014601143077015877, + 0.011085717007517815, + 0.014644733630120754, + 0.010346956551074982, + 0.014641324989497662, + 0.014169316738843918, + 0.019352329894900322, + 0.012618057429790497, + 0.013211127370595932, + 0.012310178019106388, + 0.007471812888979912, + 0.005563603714108467, + 0.0061031002551317215, + 0.005755335558205843, + 0.009723103605210781, + 0.006070019211620092, + 0.005787787493318319, + 0.006868468131870031, + 0.014410327188670635, + 0.006592260207980871, + 0.0063753193244338036, + 0.004215810913592577, + 0.00853000022470951, + 0.008002432063221931, + 0.003476719604805112, + 0.0061511388048529625, + 0.006596623919904232, + 0.004099476616829634, + 0.005021459423005581, + 0.00585296843200922, + 0.006680142134428024, + 0.005719522945582867, + 0.004145800601691008, + 0.005422770511358976, + 0.010371661745011806, + 0.007674787193536758, + 0.0058710514567792416, + 0.007325940299779177, + 0.012307839468121529, + 0.0102442791685462, + 0.007722813636064529, + 0.005348994862288237, + 0.020244482904672623, + 0.01778576150536537, + 0.010340564884245396, + 0.020082077011466026, + 0.025475088506937027, + 0.01700250431895256, + 0.022252224385738373, + 0.014183133840560913, + 0.011299495585262775, + 0.006674791686236858, + 0.007102868985384703, + 0.009370943531394005, + 0.010548499412834644, + 0.008505992591381073, + 0.008666724897921085, + 0.011872515082359314, + 0.0322461798787117, + 0.018009517341852188, + 0.026221442967653275, + 0.022521281614899635, + 0.01943584717810154, + 0.02900579757988453, + 0.027725283056497574, + 0.039016444236040115, + 0.02136746421456337, + 0.015648631379008293, + 0.016859550029039383, + 0.02140170708298683, + 0.015499475412070751, + 0.017095936462283134, + 0.020049547776579857, + 0.013830980286002159 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 6, + 9 + ], + "selected_reward": -0.00978435855358839 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the video begin with a medium shot of a phoenix perched on a peak above a sea of clouds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera gently tilt upward and slowly zoom out from the beginning?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the phoenix have feathers resembling flowing flames interwoven with threads of pure gold?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the phoenix lower its head to grasp a scroll made of aged, yellowed silk with dark wooden shafts?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do soft, multicolored streams of light spill from the scroll as it unfurls to reveal a fantastical Eastern world?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.043053578585386276 + ], + "opposite_logprob_tokens": [ + -3.16805362701416 + ], + "logprob_reward": -0.043053578585386276, + "yes_no_margin_reward": 3.125, + "token_credit_reward": -0.043053578585386276, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0378018319606781, + 0.06252457946538925, + 0.061825428158044815, + 0.06090936437249184, + 0.07322772592306137, + 0.0591902881860733, + 0.10458219051361084, + 0.18599840998649597, + 0.2176494002342224, + 0.1362908035516739 + ], + "credit_map": [ + 0.007184374146163464, + 0.003277567680925131, + 0.004133258480578661, + 0.003992373123764992, + 0.004760317038744688, + 0.004815042484551668, + 0.005184297915548086, + 0.004454602021723986, + 0.01010467205196619, + 0.005751509685069323, + 0.0069097974337637424, + 0.007708105258643627, + 0.009232494048774242, + 0.0076222531497478485, + 0.008624495938420296, + 0.006571248639374971, + 0.010356117971241474, + 0.007441862020641565, + 0.0064400117844343185, + 0.006532818544656038, + 0.009536031633615494, + 0.007230997085571289, + 0.0062096077017486095, + 0.008077974431216717, + 0.010885557159781456, + 0.005776637699455023, + 0.006826706696301699, + 0.006384255830198526, + 0.011154131963849068, + 0.008584050461649895, + 0.005064882338047028, + 0.0062331389635801315, + 0.009747163392603397, + 0.007082443218678236, + 0.006431312300264835, + 0.014507398940622807, + 0.00864771194756031, + 0.00999438390135765, + 0.006899684201925993, + 0.009917626157402992, + 0.007825375534594059, + 0.007113186176866293, + 0.0050133042968809605, + 0.006005650851875544, + 0.0115994056686759, + 0.008329925127327442, + 0.005472760647535324, + 0.00783067662268877, + 0.015476592816412449, + 0.014263510704040527, + 0.006243205163627863, + 0.011227192357182503, + 0.020175578072667122, + 0.0124121755361557, + 0.014502966776490211, + 0.010280968621373177, + 0.03321788087487221, + 0.019981561228632927, + 0.014054358005523682, + 0.02544250339269638, + 0.025991640985012054, + 0.01900525391101837, + 0.02548959292471409, + 0.022815605625510216, + 0.03439033031463623, + 0.019598323851823807, + 0.02502552419900894, + 0.023531580343842506, + 0.01860366016626358, + 0.02857334353029728, + 0.025744875892996788, + 0.042181745171546936, + 0.022912699729204178, + 0.015047521330416203, + 0.014573410153388977, + 0.017465587705373764, + 0.01676231436431408, + 0.017891323193907738, + 0.01799987629055977, + 0.013638055883347988 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 7, + 9 + ], + "selected_reward": -0.043053578585386276 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the video begin with a medium shot of a phoenix perched on a peak above a sea of clouds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera gently tilt upward and slowly zoom out from the beginning?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the phoenix have feathers resembling flowing flames interwoven with threads of pure gold?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the phoenix lower its head to grasp a scroll made of aged, yellowed silk with dark wooden shafts?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do soft, multicolored streams of light spill from the scroll as it unfurls to reveal a fantastical Eastern world?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0181779395788908 + ], + "opposite_logprob_tokens": [ + -4.0181779861450195 + ], + "logprob_reward": -0.0181779395788908, + "yes_no_margin_reward": 4.0, + "token_credit_reward": -0.0181779395788908, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15739698708057404, + 0.04816141724586487, + 0.03269267454743385, + 0.053996991366147995, + 0.032036684453487396, + 0.08154689520597458, + 0.15909934043884277, + 0.16036787629127502, + 0.22320234775543213, + 0.051498740911483765 + ], + "credit_map": [ + 0.03347466140985489, + 0.015418075025081635, + 0.017478367313742638, + 0.01626662164926529, + 0.01745796576142311, + 0.01757311262190342, + 0.017198406159877777, + 0.022529756650328636, + 0.0081249438226223, + 0.004287390969693661, + 0.006875134538859129, + 0.006561910733580589, + 0.005534849129617214, + 0.005981731694191694, + 0.006073463708162308, + 0.004721992649137974, + 0.005348070524632931, + 0.0036098386626690626, + 0.0036793937906622887, + 0.003090930636972189, + 0.0045263683423399925, + 0.0037144178058952093, + 0.0035957901272922754, + 0.0051278662867844105, + 0.012230935506522655, + 0.005086645483970642, + 0.0062893470749258995, + 0.003487775567919016, + 0.008599520660936832, + 0.007874629460275173, + 0.004091289360076189, + 0.0063368468545377254, + 0.0041447109542787075, + 0.003978373948484659, + 0.003211449831724167, + 0.004792982712388039, + 0.004421830177307129, + 0.004357720725238323, + 0.003001908306032419, + 0.004127709660679102, + 0.009205986745655537, + 0.010912387631833553, + 0.006147373467683792, + 0.007772898767143488, + 0.018700378015637398, + 0.012157506309449673, + 0.007989151403307915, + 0.008661209605634212, + 0.02917069010436535, + 0.013440907001495361, + 0.014579977840185165, + 0.018938247114419937, + 0.02196999080479145, + 0.01765694096684456, + 0.02226245030760765, + 0.0210801362991333, + 0.03180363029241562, + 0.015179075300693512, + 0.020269911736249924, + 0.01652892306447029, + 0.011050837114453316, + 0.020040959119796753, + 0.02221256121993065, + 0.02328198403120041, + 0.03498741611838341, + 0.020839910954236984, + 0.02158103510737419, + 0.035965271294116974, + 0.022940948605537415, + 0.028551148250699043, + 0.02880246751010418, + 0.02953415922820568, + 0.008697191253304482, + 0.005853080656379461, + 0.006410217843949795, + 0.0055357785895466805, + 0.0063888696022331715, + 0.0059069194830954075, + 0.00735849654302001, + 0.005348187871277332 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 7, + 6 + ], + "selected_reward": -0.0181779395788908 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "shot that begins in the medium shot, gently tilting upward and slowly zooming out, transports the viewer to the summit of a vast, ever-bending sea of clouds. A sacred phoenix perches majestically atop a lone peak piercing the clouds. Its feathers, like flowing flames, interwoven with threads of pure gold, gleam in the golden light of dawn. Its crest is as resplendent as gemstones, and its deep eyes gleam with ancient wisdom. Cautiously, it lowers its head, and with a beak that seems carved from warm jade, it gently but steadily grasps one end of a scroll made of aged, yellowed silk. The scroll's ends are framed by dark wooden shafts intricately carved with cloud patterns. As the phoenix's unhurried movements unfold, the scroll slowly unfurls, soft, multicolored streams of light spilling from its surface. The fantastical Eastern world depicted—tiered mountain ranges, winding rivers, and mist-shrouded pavilions—doesn't simply appear, but gradually coalesces and takes shape within these shimmering particles of light, as if a dormant dimension is awakening. The entire scene is bathed in a magnificent beam of light that pierces the clouds, gilding the phoenix's silhouette with a sacred gilt. The dominant tones of the painting are warm gold and calming cloud white, punctuated by the phoenix's fiery hues and the vibrant brilliance of the scroll, creating an atmosphere of grandeur, tranquility, and boundless wonder—an Eastern fantasy.", + "seed": 26990, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000101/group_15/index_4454_seed_26990_dp_31_rank_62_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 15, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.023591971956193447, + "vlm_logprob": -0.023591971956193447, + "vlm_yes_no_margin": 4.2, + "vlm_token_credit": -0.023591971956193447, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the video begin with a medium shot of a phoenix perched on a peak above a sea of clouds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera gently tilt upward and slowly zoom out from the beginning?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the phoenix have feathers resembling flowing flames interwoven with threads of pure gold?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the phoenix lower its head to grasp a scroll made of aged, yellowed silk with dark wooden shafts?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do soft, multicolored streams of light spill from the scroll as it unfurls to reveal a fantastical Eastern world?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.010358854196965694 + ], + "opposite_logprob_tokens": [ + -5.010358810424805 + ], + "logprob_reward": -0.010358854196965694, + "yes_no_margin_reward": 5.0, + "token_credit_reward": -0.010358854196965694, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14579196274280548, + 0.1511308252811432, + 0.10063422471284866, + 0.09586859494447708, + 0.11131061613559723, + 0.10841595381498337, + 0.06611204892396927, + 0.07926183193922043, + 0.09064796566963196, + 0.050825972110033035 + ], + "credit_map": [ + 0.01830614171922207, + 0.02006184495985508, + 0.019089743494987488, + 0.01840805448591709, + 0.015484031289815903, + 0.012472260743379593, + 0.024516768753528595, + 0.01745312474668026, + 0.017705533653497696, + 0.01642065867781639, + 0.01861288584768772, + 0.020021408796310425, + 0.01798613928258419, + 0.029442796483635902, + 0.01617787405848503, + 0.014763522893190384, + 0.010907809250056744, + 0.011293365620076656, + 0.011655628681182861, + 0.017907388508319855, + 0.011992608197033405, + 0.013181323185563087, + 0.012945881113409996, + 0.010750221088528633, + 0.014444306492805481, + 0.009604261256754398, + 0.01015112828463316, + 0.01828235760331154, + 0.013519838452339172, + 0.008320033550262451, + 0.009035171009600163, + 0.012511497363448143, + 0.01465991698205471, + 0.019401945173740387, + 0.01645779050886631, + 0.012379932217299938, + 0.013491062447428703, + 0.012463578023016453, + 0.009747902862727642, + 0.012708484195172787, + 0.028061440214514732, + 0.014887512661516666, + 0.010548523627221584, + 0.0085331741720438, + 0.009989053010940552, + 0.02085316926240921, + 0.008747948333621025, + 0.006795137654989958, + 0.008181777782738209, + 0.012815925292670727, + 0.007509524468332529, + 0.008049244992434978, + 0.00901572871953249, + 0.008565723896026611, + 0.0054409438744187355, + 0.00653317803516984, + 0.007927522994577885, + 0.014803132973611355, + 0.0065006474032998085, + 0.011950545012950897, + 0.011018041521310806, + 0.011511445045471191, + 0.007047583349049091, + 0.00850291084498167, + 0.01066301204264164, + 0.009710120968520641, + 0.007940193638205528, + 0.0090259350836277, + 0.009614727459847927, + 0.018389323726296425, + 0.008561227470636368, + 0.016743425279855728, + 0.0060493131168186665, + 0.0070089977234601974, + 0.005904736928641796, + 0.007874183356761932, + 0.0063107856549322605, + 0.0047878120094537735, + 0.003926216624677181, + 0.008963927626609802 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 4 + ], + "selected_reward": -0.010358854196965694 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the video begin with a medium shot of a phoenix perched on a peak above a sea of clouds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera gently tilt upward and slowly zoom out from the beginning?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the phoenix have feathers resembling flowing flames interwoven with threads of pure gold?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the phoenix lower its head to grasp a scroll made of aged, yellowed silk with dark wooden shafts?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do soft, multicolored streams of light spill from the scroll as it unfurls to reveal a fantastical Eastern world?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.06204257160425186 + ], + "opposite_logprob_tokens": [ + -2.812042474746704 + ], + "logprob_reward": -0.06204257160425186, + "yes_no_margin_reward": 2.75, + "token_credit_reward": -0.06204257160425186, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14098402857780457, + 0.24204157292842865, + 0.05819672346115112, + 0.1737566739320755, + 0.09980481117963791, + 0.05686967074871063, + 0.10034064948558807, + 0.04341894015669823, + 0.016534261405467987, + 0.0680527314543724 + ], + "credit_map": [ + 0.020170044153928757, + 0.014954006299376488, + 0.016590692102909088, + 0.011555314995348454, + 0.01738654263317585, + 0.010300085879862309, + 0.015908055007457733, + 0.03411927819252014, + 0.028847426176071167, + 0.030223198235034943, + 0.03345659002661705, + 0.03914165124297142, + 0.029128776863217354, + 0.03267543762922287, + 0.023075154051184654, + 0.025493310764431953, + 0.007573357317596674, + 0.006765550002455711, + 0.007208429276943207, + 0.008233413100242615, + 0.005430502817034721, + 0.009458632208406925, + 0.0059336107224226, + 0.007593225222080946, + 0.029038328677415848, + 0.015208118595182896, + 0.022176025435328484, + 0.035181675106287, + 0.027081141248345375, + 0.016657892614603043, + 0.012451066635549068, + 0.01596241258084774, + 0.019488565623760223, + 0.014938769862055779, + 0.010099484585225582, + 0.008000216446816921, + 0.016839006915688515, + 0.013654283247888088, + 0.006218396592885256, + 0.010566082783043385, + 0.014263052493333817, + 0.006049095652997494, + 0.005961265414953232, + 0.004557253327220678, + 0.005448151379823685, + 0.01052169967442751, + 0.00393008254468441, + 0.006139066535979509, + 0.013575146906077862, + 0.01830155774950981, + 0.012491661123931408, + 0.011711763218045235, + 0.012672554701566696, + 0.013795177452266216, + 0.00812623929232359, + 0.009666545316576958, + 0.00526810809969902, + 0.006802950520068407, + 0.004277377389371395, + 0.005631959531456232, + 0.0072263614274561405, + 0.006192920263856649, + 0.003971871919929981, + 0.004047386813908815, + 0.0019424038473516703, + 0.0027540037408471107, + 0.0016975431935861707, + 0.002689830493181944, + 0.0019576933700591326, + 0.0018665492534637451, + 0.0017730401596054435, + 0.0018531951354816556, + 0.010248656384646893, + 0.00708219688385725, + 0.0070616514421999454, + 0.007514955475926399, + 0.010249369777739048, + 0.008357033133506775, + 0.004837442189455032, + 0.012701421044766903 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 3, + 0 + ], + "selected_reward": -0.06204257160425186 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the video begin with a medium shot of a phoenix perched on a peak above a sea of clouds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera gently tilt upward and slowly zoom out from the beginning?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the phoenix have feathers resembling flowing flames interwoven with threads of pure gold?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the phoenix lower its head to grasp a scroll made of aged, yellowed silk with dark wooden shafts?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do soft, multicolored streams of light spill from the scroll as it unfurls to reveal a fantastical Eastern world?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004098350182175636 + ], + "opposite_logprob_tokens": [ + -5.504098415374756 + ], + "logprob_reward": -0.004098350182175636, + "yes_no_margin_reward": 5.5, + "token_credit_reward": -0.004098350182175636, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07058390974998474, + 0.31566953659057617, + 0.039833199232816696, + 0.07312429696321487, + 0.12661799788475037, + 0.20686399936676025, + 0.031110407784581184, + 0.048459552228450775, + 0.04383790120482445, + 0.04389914497733116 + ], + "credit_map": [ + 0.008126897737383842, + 0.014377256855368614, + 0.007787372916936874, + 0.009802363812923431, + 0.006696341559290886, + 0.0059165023267269135, + 0.009157485328614712, + 0.008719694800674915, + 0.03748699277639389, + 0.038808424025774, + 0.04447316750884056, + 0.04108308255672455, + 0.03735390678048134, + 0.04826844483613968, + 0.03683217242360115, + 0.03136336803436279, + 0.005032265093177557, + 0.005491620395332575, + 0.004342342261224985, + 0.00681552616879344, + 0.005947702098637819, + 0.004359034821391106, + 0.004025730770081282, + 0.003818980185315013, + 0.012789789587259293, + 0.006751972250640392, + 0.00802223477512598, + 0.014036308042705059, + 0.01332188956439495, + 0.007971702143549919, + 0.004080825485289097, + 0.006149578839540482, + 0.028165174648165703, + 0.010681607760488987, + 0.020324209704995155, + 0.01128408033400774, + 0.027703799307346344, + 0.012851384468376637, + 0.007906369864940643, + 0.00770137645304203, + 0.044497039169073105, + 0.0323740690946579, + 0.02187284082174301, + 0.01840599626302719, + 0.021926671266555786, + 0.033530592918395996, + 0.018181458115577698, + 0.01607535593211651, + 0.004106428008526564, + 0.0039709447883069515, + 0.003028897102922201, + 0.003774537704885006, + 0.0050001260824501514, + 0.004027940332889557, + 0.0027583979535847902, + 0.004443138372153044, + 0.005127361044287682, + 0.009875995106995106, + 0.0043304734863340855, + 0.008147875778377056, + 0.006854776293039322, + 0.005332386586815119, + 0.004719121847301722, + 0.0040715644136071205, + 0.004304295405745506, + 0.005330777261406183, + 0.004748658277094364, + 0.005313176661729813, + 0.005244425497949123, + 0.006941317580640316, + 0.003848888212814927, + 0.00810636579990387, + 0.004421865101903677, + 0.006464964710175991, + 0.004704460501670837, + 0.004070115275681019, + 0.006005499046295881, + 0.005284855142235756, + 0.004620601423084736, + 0.008326789364218712 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 5, + 4 + ], + "selected_reward": -0.004098350182175636 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the video begin with a medium shot of a phoenix perched on a peak above a sea of clouds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera gently tilt upward and slowly zoom out from the beginning?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the phoenix have feathers resembling flowing flames interwoven with threads of pure gold?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the phoenix lower its head to grasp a scroll made of aged, yellowed silk with dark wooden shafts?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do soft, multicolored streams of light spill from the scroll as it unfurls to reveal a fantastical Eastern world?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.018179696053266525 + ], + "opposite_logprob_tokens": [ + -4.018179893493652 + ], + "logprob_reward": -0.018179696053266525, + "yes_no_margin_reward": 4.0, + "token_credit_reward": -0.018179696053266525, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06055426597595215, + 0.3994782865047455, + 0.10827086865901947, + 0.16782701015472412, + 0.05554942041635513, + 0.04599577188491821, + 0.028266483917832375, + 0.051826685667037964, + 0.058081213384866714, + 0.02415001392364502 + ], + "credit_map": [ + 0.009306829422712326, + 0.00564871309325099, + 0.009854239411652088, + 0.005998092237859964, + 0.007187996059656143, + 0.0038373249117285013, + 0.008815652690827847, + 0.009905419312417507, + 0.04806138947606087, + 0.045705970376729965, + 0.057858116924762726, + 0.060275666415691376, + 0.04766405001282692, + 0.058359354734420776, + 0.041324786841869354, + 0.0402289479970932, + 0.02475876919925213, + 0.01052811834961176, + 0.006355982273817062, + 0.007399583235383034, + 0.03298585116863251, + 0.010024329647421837, + 0.00871579721570015, + 0.007502434775233269, + 0.02804475836455822, + 0.014238722622394562, + 0.019033309072256088, + 0.03451191633939743, + 0.029284270480275154, + 0.017202410846948624, + 0.010585247538983822, + 0.01492637675255537, + 0.006565521005541086, + 0.009838414378464222, + 0.006465288810431957, + 0.005235485266894102, + 0.008294334635138512, + 0.006085765548050404, + 0.005826970096677542, + 0.0072376420721411705, + 0.00970142986625433, + 0.00592406140640378, + 0.0043867905624210835, + 0.004002283327281475, + 0.005060707218945026, + 0.008730210363864899, + 0.0032094623893499374, + 0.0049808272160589695, + 0.0031390090007334948, + 0.004161872435361147, + 0.003414188977330923, + 0.0036806659772992134, + 0.004620029591023922, + 0.0038300936575978994, + 0.0025709199253469706, + 0.0028497071471065283, + 0.006319781765341759, + 0.00786780845373869, + 0.006097900681197643, + 0.008925852365791798, + 0.006381522864103317, + 0.004859937354922295, + 0.0057365866377949715, + 0.005637291818857193, + 0.00713727530092001, + 0.009064961224794388, + 0.006576091516762972, + 0.009779701009392738, + 0.00811156164854765, + 0.005943888798356056, + 0.0060123116709291935, + 0.005455420818179846, + 0.002781614661216736, + 0.0030177398584783077, + 0.0030855482909828424, + 0.003898194059729576, + 0.0033995364792644978, + 0.002715420676395297, + 0.002029731869697571, + 0.003222227096557617 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 3, + 2 + ], + "selected_reward": -0.018179696053266525 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the video begin with a medium shot of a phoenix perched on a peak above a sea of clouds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera gently tilt upward and slowly zoom out from the beginning?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the phoenix have feathers resembling flowing flames interwoven with threads of pure gold?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the phoenix lower its head to grasp a scroll made of aged, yellowed silk with dark wooden shafts?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do soft, multicolored streams of light spill from the scroll as it unfurls to reveal a fantastical Eastern world?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.023280387744307518 + ], + "opposite_logprob_tokens": [ + -3.773280382156372 + ], + "logprob_reward": -0.023280387744307518, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.023280387744307518, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.04818718507885933, + 0.48098570108413696, + 0.07928826659917831, + 0.027790186926722527, + 0.044812433421611786, + 0.1262490600347519, + 0.03201646730303764, + 0.024988889694213867, + 0.05138472467660904, + 0.0842970460653305 + ], + "credit_map": [ + 0.008591263554990292, + 0.0050026108510792255, + 0.005709576420485973, + 0.005184438545256853, + 0.005050617270171642, + 0.002873531309887767, + 0.007701321970671415, + 0.008073830045759678, + 0.056440211832523346, + 0.06028725206851959, + 0.0679217278957367, + 0.061575088649988174, + 0.05785910785198212, + 0.07209521532058716, + 0.056575287133455276, + 0.048231858760118484, + 0.016359250992536545, + 0.008037745021283627, + 0.0058867186307907104, + 0.007934405468404293, + 0.022827541455626488, + 0.006563470233231783, + 0.0065140011720359325, + 0.005165134556591511, + 0.003994029946625233, + 0.002505130600184202, + 0.0028416914865374565, + 0.006227307021617889, + 0.0038403167854994535, + 0.0024268953129649162, + 0.002753877779468894, + 0.003200938692316413, + 0.006571796722710133, + 0.005113027058541775, + 0.006728608626872301, + 0.005991717334836721, + 0.006706015206873417, + 0.004847452510148287, + 0.0039709340780973434, + 0.004882884677499533, + 0.02560785599052906, + 0.018981490284204483, + 0.017790833488106728, + 0.011270101182162762, + 0.012447953224182129, + 0.019745387136936188, + 0.010726956650614738, + 0.009678492322564125, + 0.003935026470571756, + 0.004442716483026743, + 0.003583470592275262, + 0.004014852922409773, + 0.005097416695207357, + 0.004030249547213316, + 0.0028477776795625687, + 0.004064960405230522, + 0.002842552261427045, + 0.003935393411666155, + 0.002481095725670457, + 0.0038091130554676056, + 0.003300596959888935, + 0.0026855997275561094, + 0.0028891523834317923, + 0.003045388264581561, + 0.005839657504111528, + 0.007671989966183901, + 0.00603654608130455, + 0.004607990384101868, + 0.006877304520457983, + 0.006952032446861267, + 0.004987598862498999, + 0.008411606773734093, + 0.010911890305578709, + 0.00822184607386589, + 0.008356627076864243, + 0.009650617837905884, + 0.010961848311126232, + 0.00866224430501461, + 0.009891757741570473, + 0.017640218138694763 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 5, + 9 + ], + "selected_reward": -0.023280387744307518 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "shot that begins in the medium shot, gently tilting upward and slowly zooming out, transports the viewer to the summit of a vast, ever-bending sea of clouds. A sacred phoenix perches majestically atop a lone peak piercing the clouds. Its feathers, like flowing flames, interwoven with threads of pure gold, gleam in the golden light of dawn. Its crest is as resplendent as gemstones, and its deep eyes gleam with ancient wisdom. Cautiously, it lowers its head, and with a beak that seems carved from warm jade, it gently but steadily grasps one end of a scroll made of aged, yellowed silk. The scroll's ends are framed by dark wooden shafts intricately carved with cloud patterns. As the phoenix's unhurried movements unfold, the scroll slowly unfurls, soft, multicolored streams of light spilling from its surface. The fantastical Eastern world depicted—tiered mountain ranges, winding rivers, and mist-shrouded pavilions—doesn't simply appear, but gradually coalesces and takes shape within these shimmering particles of light, as if a dormant dimension is awakening. The entire scene is bathed in a magnificent beam of light that pierces the clouds, gilding the phoenix's silhouette with a sacred gilt. The dominant tones of the painting are warm gold and calming cloud white, punctuated by the phoenix's fiery hues and the vibrant brilliance of the scroll, creating an atmosphere of grandeur, tranquility, and boundless wonder—an Eastern fantasy.", + "seed": 26991, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000101/group_15/index_4454_seed_26991_dp_31_rank_62_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 15, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.02404705323278904, + "vlm_logprob": -0.02404705323278904, + "vlm_yes_no_margin": 4.2, + "vlm_token_credit": -0.02404705323278904, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the video begin with a medium shot of a phoenix perched on a peak above a sea of clouds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera gently tilt upward and slowly zoom out from the beginning?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the phoenix have feathers resembling flowing flames interwoven with threads of pure gold?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the phoenix lower its head to grasp a scroll made of aged, yellowed silk with dark wooden shafts?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do soft, multicolored streams of light spill from the scroll as it unfurls to reveal a fantastical Eastern world?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0082594258710742 + ], + "opposite_logprob_tokens": [ + -5.383259296417236 + ], + "logprob_reward": -0.0082594258710742, + "yes_no_margin_reward": 5.375, + "token_credit_reward": -0.0082594258710742, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11990754306316376, + 0.04659076780080795, + 0.07496760785579681, + 0.05421501770615578, + 0.03477151319384575, + 0.038975078612565994, + 0.1258479207754135, + 0.24931618571281433, + 0.17689496278762817, + 0.07851346582174301 + ], + "credit_map": [ + 0.014787844382226467, + 0.00737716956064105, + 0.021481232717633247, + 0.009200120344758034, + 0.009094564244151115, + 0.03776968643069267, + 0.00962566677480936, + 0.010571240447461605, + 0.004740556236356497, + 0.006141478195786476, + 0.0056669460609555244, + 0.004363334737718105, + 0.007635567337274551, + 0.008596493862569332, + 0.004466715734452009, + 0.004979672376066446, + 0.00962262786924839, + 0.012843643315136433, + 0.009177674539387226, + 0.008953453041613102, + 0.01050100289285183, + 0.008669784292578697, + 0.006471031811088324, + 0.008728385902941227, + 0.008883780799806118, + 0.0050214733928442, + 0.008328204043209553, + 0.004281535744667053, + 0.00485324626788497, + 0.014543106779456139, + 0.00367330783046782, + 0.004630356561392546, + 0.003540720557793975, + 0.005177398677915335, + 0.0054036215879023075, + 0.003189773764461279, + 0.00638644490391016, + 0.005545494146645069, + 0.0031461575999855995, + 0.002381897997111082, + 0.004116250202059746, + 0.0067687141709029675, + 0.003635064233094454, + 0.0055824811570346355, + 0.005594166927039623, + 0.005228547845035791, + 0.0034992604050785303, + 0.004550589714199305, + 0.018996404483914375, + 0.016019806265830994, + 0.023720957338809967, + 0.008401505649089813, + 0.00946994312107563, + 0.03914724662899971, + 0.004330798517912626, + 0.005761245731264353, + 0.03374367579817772, + 0.0060739777982234955, + 0.07093283534049988, + 0.015444298274815083, + 0.011808129958808422, + 0.09490905702114105, + 0.007276289165019989, + 0.009127906523644924, + 0.0470387227833271, + 0.023840241134166718, + 0.03770497068762779, + 0.016001947224140167, + 0.014785470440983772, + 0.01747516728937626, + 0.008701165206730366, + 0.011347263120114803, + 0.010043500922620296, + 0.01197215635329485, + 0.008922474458813667, + 0.011531227268278599, + 0.008762670680880547, + 0.012069564312696457, + 0.006405474618077278, + 0.008806388825178146 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 8, + 6 + ], + "selected_reward": -0.0082594258710742 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the video begin with a medium shot of a phoenix perched on a peak above a sea of clouds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera gently tilt upward and slowly zoom out from the beginning?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the phoenix have feathers resembling flowing flames interwoven with threads of pure gold?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the phoenix lower its head to grasp a scroll made of aged, yellowed silk with dark wooden shafts?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do soft, multicolored streams of light spill from the scroll as it unfurls to reveal a fantastical Eastern world?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.06204436346888542 + ], + "opposite_logprob_tokens": [ + -2.812044382095337 + ], + "logprob_reward": -0.06204436346888542, + "yes_no_margin_reward": 2.75, + "token_credit_reward": -0.06204436346888542, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0733523890376091, + 0.04924573004245758, + 0.10821816325187683, + 0.05580848455429077, + 0.03693563491106033, + 0.05673786252737045, + 0.06080587953329086, + 0.2662886381149292, + 0.19060301780700684, + 0.10200417041778564 + ], + "credit_map": [ + 0.007169557735323906, + 0.00588845694437623, + 0.01693565770983696, + 0.0059842560440301895, + 0.006774062756448984, + 0.018605509772896767, + 0.00541213434189558, + 0.0065827518701553345, + 0.005198244005441666, + 0.005048718769103289, + 0.005162427201867104, + 0.006998089607805014, + 0.003811589442193508, + 0.011075666174292564, + 0.005931277759373188, + 0.006019718945026398, + 0.012118393555283546, + 0.00996095035225153, + 0.017991093918681145, + 0.007071962114423513, + 0.009098384529352188, + 0.019458608701825142, + 0.00891905464231968, + 0.023599717766046524, + 0.010528307408094406, + 0.00604168139398098, + 0.005776693113148212, + 0.004000082146376371, + 0.008387225680053234, + 0.011767764575779438, + 0.00459402846172452, + 0.004712708294391632, + 0.004221626557409763, + 0.006791734602302313, + 0.0045286077074706554, + 0.003217516001313925, + 0.006645765621215105, + 0.004771808162331581, + 0.0036098400596529245, + 0.0031487413216382265, + 0.006463785655796528, + 0.0073325312696397305, + 0.00931042991578579, + 0.008216741494834423, + 0.004849559161812067, + 0.011617415584623814, + 0.0031013265252113342, + 0.005846072919666767, + 0.009987345896661282, + 0.012316775508224964, + 0.010380007326602936, + 0.006319046951830387, + 0.005928501486778259, + 0.007072955369949341, + 0.003698270069435239, + 0.005102979019284248, + 0.030008655041456223, + 0.007806156761944294, + 0.08217281848192215, + 0.01695368066430092, + 0.01443412248045206, + 0.09734451770782471, + 0.007298801094293594, + 0.010269908234477043, + 0.04318224638700485, + 0.012858557514846325, + 0.05008867010474205, + 0.01931930147111416, + 0.009503350593149662, + 0.022255385294556618, + 0.01644144207239151, + 0.0169540848582983, + 0.014279382303357124, + 0.01521253027021885, + 0.010301783680915833, + 0.015015451237559319, + 0.012118169106543064, + 0.01333235576748848, + 0.01125425472855568, + 0.01049024611711502 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 8, + 2 + ], + "selected_reward": -0.06204436346888542 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the video begin with a medium shot of a phoenix perched on a peak above a sea of clouds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera gently tilt upward and slowly zoom out from the beginning?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the phoenix have feathers resembling flowing flames interwoven with threads of pure gold?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the phoenix lower its head to grasp a scroll made of aged, yellowed silk with dark wooden shafts?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do soft, multicolored streams of light spill from the scroll as it unfurls to reveal a fantastical Eastern world?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.014196240343153477 + ], + "opposite_logprob_tokens": [ + -4.264196395874023 + ], + "logprob_reward": -0.014196240343153477, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.014196240343153477, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.05721097066998482, + 0.04544251412153244, + 0.04837042838335037, + 0.07146661728620529, + 0.0443454347550869, + 0.04136432707309723, + 0.09365671873092651, + 0.414400190114975, + 0.10935354232788086, + 0.07438922673463821 + ], + "credit_map": [ + 0.006279422901570797, + 0.004356340505182743, + 0.011595632880926132, + 0.005151670891791582, + 0.005015651695430279, + 0.018549958243966103, + 0.0032835397869348526, + 0.0029787609819322824, + 0.004457319155335426, + 0.004510405473411083, + 0.004886711481958628, + 0.006986518390476704, + 0.00876951776444912, + 0.005803691688925028, + 0.0045155067928135395, + 0.005512850824743509, + 0.0061089955270290375, + 0.005010024644434452, + 0.005364647600799799, + 0.0037806187756359577, + 0.0064172763377428055, + 0.009017701260745525, + 0.004899417050182819, + 0.007771752774715424, + 0.011777106672525406, + 0.005622463300824165, + 0.008772709406912327, + 0.004869912285357714, + 0.00845927931368351, + 0.019978415220975876, + 0.004173515364527702, + 0.007813220843672752, + 0.005493674427270889, + 0.00820255558937788, + 0.0057199448347091675, + 0.00408152537420392, + 0.006277392152696848, + 0.008236442692577839, + 0.003005518112331629, + 0.003328384831547737, + 0.0038676620461046696, + 0.0034313364885747433, + 0.009160607121884823, + 0.004638938698917627, + 0.003498281817883253, + 0.011376990005373955, + 0.002433520508930087, + 0.0029569927137345076, + 0.020442280918359756, + 0.019454989582300186, + 0.013811598531901836, + 0.009143989533185959, + 0.008090473711490631, + 0.010251873172819614, + 0.004494796972721815, + 0.007966723293066025, + 0.05443336442112923, + 0.009216202422976494, + 0.12078861892223358, + 0.023715419694781303, + 0.021481167525053024, + 0.16131430864334106, + 0.01038377545773983, + 0.013067328371107578, + 0.02723236382007599, + 0.012622462585568428, + 0.02577698603272438, + 0.010728199034929276, + 0.008361275307834148, + 0.008958607912063599, + 0.006232236046344042, + 0.009441413916647434, + 0.009831331670284271, + 0.00631724763661623, + 0.022347023710608482, + 0.01565246470272541, + 0.003380239475518465, + 0.004686159547418356, + 0.007336039561778307, + 0.004838720429688692 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 8, + 6 + ], + "selected_reward": -0.014196240343153477 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the video begin with a medium shot of a phoenix perched on a peak above a sea of clouds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera gently tilt upward and slowly zoom out from the beginning?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the phoenix have feathers resembling flowing flames interwoven with threads of pure gold?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the phoenix lower its head to grasp a scroll made of aged, yellowed silk with dark wooden shafts?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do soft, multicolored streams of light spill from the scroll as it unfurls to reveal a fantastical Eastern world?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.029790569096803665 + ], + "opposite_logprob_tokens": [ + -3.5297906398773193 + ], + "logprob_reward": -0.029790569096803665, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.029790569096803665, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.014850657433271408, + 0.018877994269132614, + 0.047273341566324234, + 0.05093967914581299, + 0.022934887558221817, + 0.05277562513947487, + 0.15665575861930847, + 0.41733577847480774, + 0.15701620280742645, + 0.061340101063251495 + ], + "credit_map": [ + 0.0016785046318545938, + 0.0013881478225812316, + 0.002385693369433284, + 0.0013977662893012166, + 0.002171423751860857, + 0.0031319663394242525, + 0.0011154005769640207, + 0.0015817545354366302, + 0.0021130479872226715, + 0.002954656956717372, + 0.001973224338144064, + 0.0017671688692644238, + 0.0024574247654527426, + 0.00392623245716095, + 0.0017247082432731986, + 0.0019615287892520428, + 0.006338795181363821, + 0.005439708475023508, + 0.005692936480045319, + 0.003850297536700964, + 0.007911529392004013, + 0.00881289318203926, + 0.004546946845948696, + 0.004680234007537365, + 0.00861730519682169, + 0.003733250079676509, + 0.0070163956843316555, + 0.0028930811677128077, + 0.00631370535120368, + 0.014338454231619835, + 0.0032050684094429016, + 0.004822415765374899, + 0.003958805929869413, + 0.002689163200557232, + 0.0030394853092730045, + 0.0019243856659159064, + 0.002636163029819727, + 0.004194730427116156, + 0.0012486479245126247, + 0.0032435061875730753, + 0.005735494662076235, + 0.004351057577878237, + 0.013033155351877213, + 0.0038351656403392553, + 0.003919282462447882, + 0.01677442155778408, + 0.0020781958010047674, + 0.003048849292099476, + 0.018344247713685036, + 0.0038271863013505936, + 0.045764606446027756, + 0.008116403594613075, + 0.0070192557759583, + 0.062361665070056915, + 0.004704284016042948, + 0.006518124602735043, + 0.05388304963707924, + 0.008170642890036106, + 0.12264806777238846, + 0.024247899651527405, + 0.021549776196479797, + 0.16337858140468597, + 0.010369017720222473, + 0.013088727369904518, + 0.0404643788933754, + 0.015534747391939163, + 0.041613850742578506, + 0.012623529881238937, + 0.01365070603787899, + 0.0135676683858037, + 0.008279814384877682, + 0.011281522922217846, + 0.012537594884634018, + 0.007998228073120117, + 0.013828265480697155, + 0.006639433093369007, + 0.004444169346243143, + 0.007467131596058607, + 0.004628100898116827, + 0.0037971818819642067 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 8, + 6 + ], + "selected_reward": -0.029790569096803665 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the video begin with a medium shot of a phoenix perched on a peak above a sea of clouds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera gently tilt upward and slowly zoom out from the beginning?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the phoenix have feathers resembling flowing flames interwoven with threads of pure gold?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the phoenix lower its head to grasp a scroll made of aged, yellowed silk with dark wooden shafts?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do soft, multicolored streams of light spill from the scroll as it unfurls to reveal a fantastical Eastern world?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005944667384028435 + ], + "opposite_logprob_tokens": [ + -5.130944728851318 + ], + "logprob_reward": -0.005944667384028435, + "yes_no_margin_reward": 5.125, + "token_credit_reward": -0.005944667384028435, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.056762415915727615, + 0.047784265130758286, + 0.06502266228199005, + 0.018992945551872253, + 0.037431199103593826, + 0.18400880694389343, + 0.12598855793476105, + 0.08522479981184006, + 0.14935781061649323, + 0.22942650318145752 + ], + "credit_map": [ + 0.007841002196073532, + 0.0049791778437793255, + 0.008985301479697227, + 0.004523569718003273, + 0.004753545392304659, + 0.017243390902876854, + 0.00478180218487978, + 0.0036546343471854925, + 0.0041235461831092834, + 0.004837753716856241, + 0.004554925952106714, + 0.009537795558571815, + 0.003981537185609341, + 0.008619921281933784, + 0.005752588622272015, + 0.006376202683895826, + 0.00653084646910429, + 0.004843821749091148, + 0.012073605321347713, + 0.004735902883112431, + 0.005712407641112804, + 0.008734649047255516, + 0.005451662465929985, + 0.01693977415561676, + 0.0022910682018846273, + 0.003063300857320428, + 0.0017420988297089934, + 0.002655204152688384, + 0.0026325625367462635, + 0.0022602244280278683, + 0.0020393081940710545, + 0.002309179399162531, + 0.004083708394318819, + 0.00573820760473609, + 0.0037022114265710115, + 0.0046180495992302895, + 0.004170119762420654, + 0.0070726010017097, + 0.0030252495780587196, + 0.005021057557314634, + 0.020235810428857803, + 0.017880484461784363, + 0.0454275943338871, + 0.012379397638142109, + 0.010908597148954868, + 0.061465732753276825, + 0.006643272005021572, + 0.00906794797629118, + 0.026669790968298912, + 0.028208184987306595, + 0.017152128741145134, + 0.011408473365008831, + 0.011355088092386723, + 0.015283760614693165, + 0.005606984719634056, + 0.010304153896868229, + 0.0094588752835989, + 0.008063855580985546, + 0.02343514747917652, + 0.00624405313283205, + 0.006884455215185881, + 0.023618722334504128, + 0.003306414233520627, + 0.004213280044496059, + 0.0360286682844162, + 0.012442272156476974, + 0.037969574332237244, + 0.013645894825458527, + 0.009318807162344456, + 0.01326326560229063, + 0.013627106323838234, + 0.013062240555882454, + 0.039289213716983795, + 0.02005034126341343, + 0.07145267724990845, + 0.03868889436125755, + 0.014771498739719391, + 0.01564643904566765, + 0.015739699825644493, + 0.013787752948701382 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 5, + 8 + ], + "selected_reward": -0.005944667384028435 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000101/group_2/index_tensor([6541, 6541])_group_2_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000101/group_2/index_tensor([6541, 6541])_group_2_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..eb85b2cc1e052cbd532cf5cb716a45a875873e88 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000101/group_2/index_tensor([6541, 6541])_group_2_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "static, close-up shot, shot from a slightly overhead angle, focuses the viewer's attention on a scene illuminated by the flickering flame of an ancient copper oil lamp. At the center of the frame lies a middle-aged strategist, wearing a dark silk scarf and a wide-sleeved black robe. He sits at a heavy, dark wooden desk, his brow furrowed, his eyes sharp as an eagle's, his gaze fixed intently on the object before him. The warm, dancing light casts deep shadows on his chiseled features, highlighting his lips pursed in deep thought. His slender, powerful hands, carefully, steadily, and deliberately, slowly unroll a bamboo map bound by string. As he moves, the polished bamboo strips rub against each other, creating a crisp, subtle rustling sound. The map's ink-drawn mountains and rivers, along with the cinnabar-marked military fortresses, gradually emerge in the dim light, exuding a sense of age and age. Beside him, on the desk, rests a feather fan crafted from white crane feathers, its ebony handle gleaming softly in the light. The painting's warm, deep tones, predominantly dark browns and oranges, create a stark contrast between light and shadow, evoking a sense of intense concentration and tension as one plots and decides a battle in the stillness of the night.", + "seed": 30522, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000101/group_2/index_6541_seed_30522_dp_4_rank_8_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 2, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.061763187404721975, + "vlm_logprob": -0.061763187404721975, + "vlm_yes_no_margin": 3.674999976158142, + "vlm_token_credit": -0.061763187404721975, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static close-up from a slightly overhead angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene illuminated by the flickering flame of an ancient copper oil lamp?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a middle-aged man wearing a dark silk scarf and a wide-sleeved black robe sitting at a heavy dark wooden desk?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the man slowly unrolling a bamboo map bound by string with his hands?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is a feather fan crafted from white crane feathers with an ebony handle visible on the desk beside him?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03844918683171272 + ], + "opposite_logprob_tokens": [ + -4.788449287414551 + ], + "logprob_reward": -0.03844918683171272, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.03844918683171272, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12810185551643372, + 0.04388866946101189, + 0.046031370759010315, + 0.09192437678575516, + 0.08650513738393784, + 0.07686685025691986, + 0.12330307066440582, + 0.19866937398910522, + 0.08627781271934509, + 0.11843153834342957 + ], + "credit_map": [ + 0.012310145422816277, + 0.028316110372543335, + 0.011039157398045063, + 0.01319110207259655, + 0.018346570432186127, + 0.018111655488610268, + 0.018490811809897423, + 0.00829627737402916, + 0.005117545370012522, + 0.006828763522207737, + 0.004228295758366585, + 0.0053490265272557735, + 0.00631804671138525, + 0.00643810024484992, + 0.005926498211920261, + 0.0036823851987719536, + 0.0037824194878339767, + 0.007580848410725594, + 0.006737548857927322, + 0.005823299754410982, + 0.008036852814257145, + 0.007393856067210436, + 0.003908231388777494, + 0.0027683076914399862, + 0.01807759888470173, + 0.017484504729509354, + 0.014498021453619003, + 0.007186876609921455, + 0.0050521125085651875, + 0.017068935558199883, + 0.007741750683635473, + 0.004814566113054752, + 0.008286841213703156, + 0.012273396365344524, + 0.010338320396840572, + 0.014975748024880886, + 0.00926361046731472, + 0.014606163837015629, + 0.0100869694724679, + 0.0066740745678544044, + 0.005871577188372612, + 0.013357001356780529, + 0.006644213572144508, + 0.008480161428451538, + 0.0047225141897797585, + 0.023788556456565857, + 0.009705506265163422, + 0.004297315143048763, + 0.014507404528558254, + 0.01381623838096857, + 0.022746121510863304, + 0.021687019616365433, + 0.015271399170160294, + 0.011459068395197392, + 0.014629882760345936, + 0.009185925126075745, + 0.022851508110761642, + 0.019055165350437164, + 0.02986224554479122, + 0.03871074691414833, + 0.021703269332647324, + 0.0198126882314682, + 0.029212215915322304, + 0.017461514100432396, + 0.009951971471309662, + 0.007582142949104309, + 0.01110224612057209, + 0.01809069514274597, + 0.01071029994636774, + 0.008766040205955505, + 0.012200826779007912, + 0.007873582653701305, + 0.014819469302892685, + 0.013397762551903725, + 0.01147479098290205, + 0.010583815164864063, + 0.0183250792324543, + 0.02941114641726017, + 0.009112635627388954, + 0.011306826025247574 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 0, + 6 + ], + "selected_reward": -0.03844918683171272 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static close-up from a slightly overhead angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene illuminated by the flickering flame of an ancient copper oil lamp?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a middle-aged man wearing a dark silk scarf and a wide-sleeved black robe sitting at a heavy dark wooden desk?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the man slowly unrolling a bamboo map bound by string with his hands?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is a feather fan crafted from white crane feathers with an ebony handle visible on the desk beside him?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.12706433236598969 + ], + "opposite_logprob_tokens": [ + -2.1270642280578613 + ], + "logprob_reward": -0.12706433236598969, + "yes_no_margin_reward": 1.9999998807907104, + "token_credit_reward": -0.12706433236598969, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.19148525595664978, + 0.09041667729616165, + 0.21282894909381866, + 0.15977749228477478, + 0.10200918465852737, + 0.048364970833063126, + 0.07608090341091156, + 0.046619098633527756, + 0.05474479869008064, + 0.017672689631581306 + ], + "credit_map": [ + 0.01588362455368042, + 0.029852526262402534, + 0.02083040028810501, + 0.02888522297143936, + 0.03288890793919563, + 0.026882749050855637, + 0.022466354072093964, + 0.013795463368296623, + 0.00852958858013153, + 0.014849833212792873, + 0.010300303809344769, + 0.012870671227574348, + 0.010671005584299564, + 0.01447182334959507, + 0.010934648104012012, + 0.007788804825395346, + 0.02206728607416153, + 0.028948599472641945, + 0.04108608886599541, + 0.031078116968274117, + 0.023378292098641396, + 0.03480735421180725, + 0.018209699541330338, + 0.013253496959805489, + 0.024400806054472923, + 0.017711536958813667, + 0.024835703894495964, + 0.021033821627497673, + 0.02788599394261837, + 0.022809043526649475, + 0.01132542546838522, + 0.009775158017873764, + 0.009514081291854382, + 0.014043323695659637, + 0.011335518211126328, + 0.01034558191895485, + 0.020821301266551018, + 0.014391303062438965, + 0.00943538174033165, + 0.01212269812822342, + 0.004226439632475376, + 0.007336399983614683, + 0.00433490052819252, + 0.006864831317216158, + 0.005458643659949303, + 0.011008642613887787, + 0.006523278076201677, + 0.002611837349832058, + 0.009898657910525799, + 0.007256595883518457, + 0.011094329878687859, + 0.01335198525339365, + 0.009989448823034763, + 0.007366969715803862, + 0.009530860930681229, + 0.007592055480927229, + 0.005116241052746773, + 0.004923020955175161, + 0.0071293022483587265, + 0.008695117197930813, + 0.005433432757854462, + 0.005752120167016983, + 0.005074577406048775, + 0.004495289642363787, + 0.009831855073571205, + 0.00576288066804409, + 0.008385688997805119, + 0.011555161327123642, + 0.005288764368742704, + 0.004042987711727619, + 0.00661392230540514, + 0.0032635359093546867, + 0.0014056771760806441, + 0.001945237279869616, + 0.003939723130315542, + 0.0028335165698081255, + 0.0018440592102706432, + 0.0017867182614281774, + 0.002502842340618372, + 0.0014149148482829332 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 0, + 3 + ], + "selected_reward": -0.12706433236598969 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static close-up from a slightly overhead angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene illuminated by the flickering flame of an ancient copper oil lamp?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a middle-aged man wearing a dark silk scarf and a wide-sleeved black robe sitting at a heavy dark wooden desk?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the man slowly unrolling a bamboo map bound by string with his hands?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is a feather fan crafted from white crane feathers with an ebony handle visible on the desk beside him?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.007630960550159216 + ], + "opposite_logprob_tokens": [ + -4.882630825042725 + ], + "logprob_reward": -0.007630960550159216, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.007630960550159216, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.21384358406066895, + 0.07095600664615631, + 0.0740237906575203, + 0.14641444385051727, + 0.13728216290473938, + 0.12829241156578064, + 0.06803454458713531, + 0.07538636773824692, + 0.05890672653913498, + 0.026859933510422707 + ], + "credit_map": [ + 0.01993442513048649, + 0.04453391581773758, + 0.021115584298968315, + 0.027024004608392715, + 0.032612212002277374, + 0.03376805782318115, + 0.019463784992694855, + 0.015391623601317406, + 0.007054123096168041, + 0.010287681594491005, + 0.006916454993188381, + 0.011564345099031925, + 0.009651928208768368, + 0.010903113521635532, + 0.009747578762471676, + 0.004830783233046532, + 0.007235814817249775, + 0.009519525803625584, + 0.00846471730619669, + 0.011570692993700504, + 0.013401846401393414, + 0.011189883574843407, + 0.007968291640281677, + 0.004673032555729151, + 0.017631839960813522, + 0.023165369406342506, + 0.012114359997212887, + 0.012316684238612652, + 0.03243377432227135, + 0.02436177246272564, + 0.010776782408356667, + 0.013613862916827202, + 0.014210614375770092, + 0.018066469579935074, + 0.012214453890919685, + 0.014550760388374329, + 0.028994714841246605, + 0.025263618677854538, + 0.008710432797670364, + 0.015271114185452461, + 0.012470269575715065, + 0.021628303453326225, + 0.009588428772985935, + 0.011500628665089607, + 0.01230743620544672, + 0.03706388548016548, + 0.01333654299378395, + 0.010396924801170826, + 0.00789809599518776, + 0.007439837791025639, + 0.010203588753938675, + 0.008570613339543343, + 0.013666655868291855, + 0.008245594799518585, + 0.005598505958914757, + 0.00641165766865015, + 0.011388435959815979, + 0.008818211033940315, + 0.009109294973313808, + 0.01422636117786169, + 0.007406945805996656, + 0.0065423911437392235, + 0.010546514764428139, + 0.0073482259176671505, + 0.010713569819927216, + 0.005446288269013166, + 0.007356849033385515, + 0.009502463042736053, + 0.009031490422785282, + 0.00499009620398283, + 0.007058636751025915, + 0.004807341378182173, + 0.00343835586681962, + 0.0020472363103181124, + 0.004803302697837353, + 0.0032315070275217295, + 0.0036774296313524246, + 0.003085075644776225, + 0.0025986232794821262, + 0.003978405613452196 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 4 + ], + "selected_reward": -0.007630960550159216 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static close-up from a slightly overhead angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene illuminated by the flickering flame of an ancient copper oil lamp?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a middle-aged man wearing a dark silk scarf and a wide-sleeved black robe sitting at a heavy dark wooden desk?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the man slowly unrolling a bamboo map bound by string with his hands?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is a feather fan crafted from white crane feathers with an ebony handle visible on the desk beside him?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.008637567982077599 + ], + "opposite_logprob_tokens": [ + -4.758637428283691 + ], + "logprob_reward": -0.008637567982077599, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.008637567982077599, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11283828318119049, + 0.05121699720621109, + 0.204631507396698, + 0.04948906600475311, + 0.06399023532867432, + 0.09510473906993866, + 0.07668742537498474, + 0.0654735267162323, + 0.21728956699371338, + 0.06327861547470093 + ], + "credit_map": [ + 0.007099440786987543, + 0.01681152731180191, + 0.012039725668728352, + 0.025209441781044006, + 0.008858655579388142, + 0.01756417751312256, + 0.01733175292611122, + 0.007923563942313194, + 0.005322751589119434, + 0.008091319352388382, + 0.005372217856347561, + 0.008042905479669571, + 0.00723332678899169, + 0.007287053856998682, + 0.005804597400128841, + 0.00406282301992178, + 0.02371506206691265, + 0.02643689326941967, + 0.03087204322218895, + 0.03260545805096626, + 0.02387959510087967, + 0.03265955299139023, + 0.018769660964608192, + 0.015693239867687225, + 0.005502447951585054, + 0.007401721086353064, + 0.004520728252828121, + 0.004379429388791323, + 0.011826029978692532, + 0.00804288312792778, + 0.004114546813070774, + 0.003701278008520603, + 0.005363333970308304, + 0.009522920474410057, + 0.005846077110618353, + 0.01525613572448492, + 0.006396291311830282, + 0.008506303653120995, + 0.008276291191577911, + 0.004822880961000919, + 0.013512959703803062, + 0.011482875794172287, + 0.01034670788794756, + 0.015258094295859337, + 0.009693553671240807, + 0.018734559416770935, + 0.009107574820518494, + 0.006968413479626179, + 0.005826277192682028, + 0.008154058828949928, + 0.011808542534708977, + 0.013372239656746387, + 0.006192492321133614, + 0.012849091552197933, + 0.00806716550141573, + 0.010417559184134007, + 0.006364803295582533, + 0.006847599055618048, + 0.00709166657179594, + 0.01438052486628294, + 0.009040419943630695, + 0.008087093010544777, + 0.008290855213999748, + 0.005370556376874447, + 0.0345936119556427, + 0.023685721680521965, + 0.034844182431697845, + 0.03838246315717697, + 0.02523922733962536, + 0.016629422083497047, + 0.0257536843419075, + 0.018161242827773094, + 0.008926940150558949, + 0.006493154913187027, + 0.007085548713803291, + 0.0104863615706563, + 0.006374587304890156, + 0.01275215670466423, + 0.006478958763182163, + 0.004680907819420099 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 2, + 0 + ], + "selected_reward": -0.008637567982077599 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static close-up from a slightly overhead angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene illuminated by the flickering flame of an ancient copper oil lamp?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a middle-aged man wearing a dark silk scarf and a wide-sleeved black robe sitting at a heavy dark wooden desk?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the man slowly unrolling a bamboo map bound by string with his hands?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is a feather fan crafted from white crane feathers with an ebony handle visible on the desk beside him?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.12703388929367065 + ], + "opposite_logprob_tokens": [ + -2.1270339488983154 + ], + "logprob_reward": -0.12703388929367065, + "yes_no_margin_reward": 2.0, + "token_credit_reward": -0.12703388929367065, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13964737951755524, + 0.06617804616689682, + 0.1234835833311081, + 0.14463317394256592, + 0.10934522747993469, + 0.13388730585575104, + 0.10482785850763321, + 0.07210757583379745, + 0.038606323301792145, + 0.06728354841470718 + ], + "credit_map": [ + 0.012553668580949306, + 0.02874492108821869, + 0.013345209881663322, + 0.01871250569820404, + 0.019154401496052742, + 0.019902577623724937, + 0.016203757375478745, + 0.011030328460037708, + 0.006394183728843927, + 0.01100701279938221, + 0.005504409782588482, + 0.008797841146588326, + 0.009797884151339531, + 0.010479884222149849, + 0.00906994566321373, + 0.005126879550516605, + 0.01253900770097971, + 0.016257772222161293, + 0.019508805125951767, + 0.015857821330428123, + 0.018601199612021446, + 0.01988859660923481, + 0.0117079121991992, + 0.00912246573716402, + 0.024343764409422874, + 0.02416084334254265, + 0.014831536449491978, + 0.010734166018664837, + 0.025562047958374023, + 0.024900900200009346, + 0.00798195693641901, + 0.012117958627641201, + 0.01136370375752449, + 0.0172598734498024, + 0.007165343035012484, + 0.008972489275038242, + 0.02366483025252819, + 0.022914763540029526, + 0.0059301406145095825, + 0.012074080295860767, + 0.011774986982345581, + 0.021138709038496017, + 0.008669714443385601, + 0.011283600702881813, + 0.014399168081581593, + 0.03806727007031441, + 0.01860772632062435, + 0.009946132078766823, + 0.014380640350282192, + 0.012224491685628891, + 0.009927838109433651, + 0.014344013296067715, + 0.019781555980443954, + 0.012854277156293392, + 0.009372093714773655, + 0.01194295659661293, + 0.010350091382861137, + 0.00776331964880228, + 0.0098768575116992, + 0.014615661464631557, + 0.00710634607821703, + 0.00634130509570241, + 0.010566726326942444, + 0.005487268324941397, + 0.004199718590825796, + 0.004020052030682564, + 0.007020788732916117, + 0.008627070114016533, + 0.0042351060546934605, + 0.0029369385447353125, + 0.004627267364412546, + 0.002939383964985609, + 0.007824717089533806, + 0.008044864051043987, + 0.0056261527352035046, + 0.005307499319314957, + 0.01057805959135294, + 0.016278216615319252, + 0.005917416885495186, + 0.00770662073045969 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 0, + 5 + ], + "selected_reward": -0.12703388929367065 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "static, close-up shot, shot from a slightly overhead angle, focuses the viewer's attention on a scene illuminated by the flickering flame of an ancient copper oil lamp. At the center of the frame lies a middle-aged strategist, wearing a dark silk scarf and a wide-sleeved black robe. He sits at a heavy, dark wooden desk, his brow furrowed, his eyes sharp as an eagle's, his gaze fixed intently on the object before him. The warm, dancing light casts deep shadows on his chiseled features, highlighting his lips pursed in deep thought. His slender, powerful hands, carefully, steadily, and deliberately, slowly unroll a bamboo map bound by string. As he moves, the polished bamboo strips rub against each other, creating a crisp, subtle rustling sound. The map's ink-drawn mountains and rivers, along with the cinnabar-marked military fortresses, gradually emerge in the dim light, exuding a sense of age and age. Beside him, on the desk, rests a feather fan crafted from white crane feathers, its ebony handle gleaming softly in the light. The painting's warm, deep tones, predominantly dark browns and oranges, create a stark contrast between light and shadow, evoking a sense of intense concentration and tension as one plots and decides a battle in the stillness of the night.", + "seed": 30523, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000101/group_2/index_6541_seed_30523_dp_4_rank_8_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 2, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.03915475308895111, + "vlm_logprob": -0.03915475308895111, + "vlm_yes_no_margin": 3.75, + "vlm_token_credit": -0.03915475308895111, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static close-up from a slightly overhead angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene illuminated by the flickering flame of an ancient copper oil lamp?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a middle-aged man wearing a dark silk scarf and a wide-sleeved black robe sitting at a heavy dark wooden desk?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the man slowly unrolling a bamboo map bound by string with his hands?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is a feather fan crafted from white crane feathers with an ebony handle visible on the desk beside him?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.035746458917856216 + ], + "opposite_logprob_tokens": [ + -5.1607465744018555 + ], + "logprob_reward": -0.035746458917856216, + "yes_no_margin_reward": 5.125, + "token_credit_reward": -0.035746458917856216, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1317519098520279, + 0.04999659210443497, + 0.029296142980456352, + 0.01863052137196064, + 0.34223294258117676, + 0.08491375297307968, + 0.11124590039253235, + 0.10154537111520767, + 0.07358007878065109, + 0.056806743144989014 + ], + "credit_map": [ + 0.012279675342142582, + 0.01534388866275549, + 0.01740514114499092, + 0.020649781450629234, + 0.011630149558186531, + 0.02103549614548683, + 0.011003493331372738, + 0.022404292598366737, + 0.004218835849314928, + 0.006547451950609684, + 0.007364662364125252, + 0.008127870038151741, + 0.0056757694110274315, + 0.007281891535967588, + 0.005156205501407385, + 0.005623898934572935, + 0.0026089686434715986, + 0.004959189333021641, + 0.003906812984496355, + 0.0038335376884788275, + 0.003762909211218357, + 0.0028684099670499563, + 0.0040000369772315025, + 0.003356279106810689, + 0.0018339678645133972, + 0.0022198250517249107, + 0.0023293839767575264, + 0.00290962727740407, + 0.0022449959069490433, + 0.0023007697891443968, + 0.0028553493320941925, + 0.0019366027554497123, + 0.0758698359131813, + 0.03258664906024933, + 0.03898727893829346, + 0.01840732805430889, + 0.11880677193403244, + 0.03234870359301567, + 0.012448125518858433, + 0.012778257019817829, + 0.019012512639164925, + 0.022059964016079903, + 0.008932286873459816, + 0.007846836000680923, + 0.009571204893290997, + 0.005032443441450596, + 0.0076330117881298065, + 0.004825495649129152, + 0.021719997748732567, + 0.02481973171234131, + 0.012408090755343437, + 0.016633134335279465, + 0.006761232856661081, + 0.006380076054483652, + 0.011567463167011738, + 0.01095616165548563, + 0.011455755680799484, + 0.007840891368687153, + 0.011788783594965935, + 0.027201266959309578, + 0.00815113540738821, + 0.008544404990971088, + 0.019075941294431686, + 0.007487193681299686, + 0.009939205832779408, + 0.0060905395075678825, + 0.011613404378294945, + 0.013683026656508446, + 0.005254263058304787, + 0.005252467468380928, + 0.016627443954348564, + 0.005119729321449995, + 0.006901674438267946, + 0.007187661714851856, + 0.007897509261965752, + 0.01096136774867773, + 0.005492242984473705, + 0.00792427733540535, + 0.0060827541165053844, + 0.004359254613518715 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 0, + 6 + ], + "selected_reward": -0.035746458917856216 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static close-up from a slightly overhead angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene illuminated by the flickering flame of an ancient copper oil lamp?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a middle-aged man wearing a dark silk scarf and a wide-sleeved black robe sitting at a heavy dark wooden desk?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the man slowly unrolling a bamboo map bound by string with his hands?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is a feather fan crafted from white crane feathers with an ebony handle visible on the desk beside him?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02983662113547325 + ], + "opposite_logprob_tokens": [ + -3.529836654663086 + ], + "logprob_reward": -0.02983662113547325, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.02983662113547325, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.028481829911470413, + 0.018418986350297928, + 0.014025182463228703, + 0.012439952231943607, + 0.7134297490119934, + 0.025889797136187553, + 0.02345014549791813, + 0.06994450837373734, + 0.05029433220624924, + 0.043625447899103165 + ], + "credit_map": [ + 0.0034027379006147385, + 0.003270274493843317, + 0.004360802471637726, + 0.003471272997558117, + 0.0024223930668085814, + 0.004732290282845497, + 0.00336862588301301, + 0.003453435841947794, + 0.0017465517157688737, + 0.002564341062679887, + 0.0028155019972473383, + 0.003080395283177495, + 0.002202550182119012, + 0.0034004475455731153, + 0.0011985336896032095, + 0.001410666503943503, + 0.0012445051688700914, + 0.002607773058116436, + 0.0020003090612590313, + 0.0021437318064272404, + 0.001725196954794228, + 0.001645122654736042, + 0.001471476978622377, + 0.0011870679445564747, + 0.0009688299032859504, + 0.001973574049770832, + 0.001608505379408598, + 0.002298652194440365, + 0.001826677587814629, + 0.0016656079096719623, + 0.0011877916986122727, + 0.0009103151387535036, + 0.1806950569152832, + 0.06067809462547302, + 0.08212949335575104, + 0.023504076525568962, + 0.28950396180152893, + 0.04422374069690704, + 0.012674770317971706, + 0.02002066932618618, + 0.004110023844987154, + 0.005297175608575344, + 0.0028402134776115417, + 0.002989499131217599, + 0.004144631326198578, + 0.0020247572101652622, + 0.0026014058385044336, + 0.0018820901168510318, + 0.004178899805992842, + 0.004270427860319614, + 0.002448250539600849, + 0.003391454229131341, + 0.0014472261536866426, + 0.0014562291326001287, + 0.0037692226469516754, + 0.0024884359445422888, + 0.009234647266566753, + 0.00708524277433753, + 0.007865897379815578, + 0.013105031102895737, + 0.0062910420820117, + 0.008169743232429028, + 0.010643666610121727, + 0.007549242116510868, + 0.006266709882766008, + 0.005714694038033485, + 0.009380555711686611, + 0.008558486588299274, + 0.0036465313751250505, + 0.004635416436940432, + 0.008777095936238766, + 0.0033148452639579773, + 0.0059920791536569595, + 0.005651008803397417, + 0.006386471446603537, + 0.0053512929007411, + 0.003574717789888382, + 0.00674129044637084, + 0.0060827238485217094, + 0.0038458663038909435 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 7, + 8 + ], + "selected_reward": -0.02983662113547325 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static close-up from a slightly overhead angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene illuminated by the flickering flame of an ancient copper oil lamp?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a middle-aged man wearing a dark silk scarf and a wide-sleeved black robe sitting at a heavy dark wooden desk?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the man slowly unrolling a bamboo map bound by string with his hands?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is a feather fan crafted from white crane feathers with an ebony handle visible on the desk beside him?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02058931440114975 + ], + "opposite_logprob_tokens": [ + -3.8955893516540527 + ], + "logprob_reward": -0.02058931440114975, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.02058931440114975, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10229292511940002, + 0.030273908749222755, + 0.027660207822918892, + 0.03686506301164627, + 0.48280858993530273, + 0.04885635897517204, + 0.09902393817901611, + 0.08966152369976044, + 0.045570582151412964, + 0.03698692470788956 + ], + "credit_map": [ + 0.008289403282105923, + 0.01591525971889496, + 0.015379543416202068, + 0.0189884714782238, + 0.011331220157444477, + 0.011115608736872673, + 0.012003177776932716, + 0.00927023310214281, + 0.0027310652658343315, + 0.0034428725484758615, + 0.004402969032526016, + 0.006149908062070608, + 0.003062541363760829, + 0.004220195114612579, + 0.0030842020642012358, + 0.0031801508739590645, + 0.002231158083304763, + 0.005340610630810261, + 0.0039032811764627695, + 0.004090601112693548, + 0.0039040716364979744, + 0.003764726920053363, + 0.0024042739532887936, + 0.002021480118855834, + 0.003630770603194833, + 0.00555380480363965, + 0.005508032161742449, + 0.006212871056050062, + 0.005156796891242266, + 0.004633767530322075, + 0.0040089902468025684, + 0.002160025527700782, + 0.12173444777727127, + 0.044043656438589096, + 0.051194921135902405, + 0.02307729795575142, + 0.18693512678146362, + 0.030511081218719482, + 0.011524380184710026, + 0.013787651434540749, + 0.005572140216827393, + 0.013131532818078995, + 0.004471567925065756, + 0.004303157329559326, + 0.00953522976487875, + 0.0034616352058947086, + 0.004572839010506868, + 0.0038082494866102934, + 0.014760660007596016, + 0.01527502853423357, + 0.009261700324714184, + 0.019083814695477486, + 0.006250231061130762, + 0.006515659391880035, + 0.017605921253561974, + 0.010270912200212479, + 0.010525899939239025, + 0.007138052489608526, + 0.011760697700083256, + 0.02306031621992588, + 0.006063820794224739, + 0.008806259371340275, + 0.015218859538435936, + 0.007087610196322203, + 0.006327688228338957, + 0.0075265695340931416, + 0.005689249839633703, + 0.007046672515571117, + 0.003559695789590478, + 0.0036931552458554506, + 0.008125989697873592, + 0.0036015580408275127, + 0.004217320121824741, + 0.004282757174223661, + 0.005486807785928249, + 0.006718985270708799, + 0.0035525509156286716, + 0.005291851237416267, + 0.004535915330052376, + 0.002900733146816492 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 0, + 6 + ], + "selected_reward": -0.02058931440114975 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static close-up from a slightly overhead angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene illuminated by the flickering flame of an ancient copper oil lamp?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a middle-aged man wearing a dark silk scarf and a wide-sleeved black robe sitting at a heavy dark wooden desk?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the man slowly unrolling a bamboo map bound by string with his hands?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is a feather fan crafted from white crane feathers with an ebony handle visible on the desk beside him?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02057892084121704 + ], + "opposite_logprob_tokens": [ + -3.8955788612365723 + ], + "logprob_reward": -0.02057892084121704, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.02057892084121704, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.020846497267484665, + 0.01473244372755289, + 0.026352494955062866, + 0.013640515506267548, + 0.806074321269989, + 0.020111778751015663, + 0.04462587460875511, + 0.01137020718306303, + 0.02557929791510105, + 0.016666576266288757 + ], + "credit_map": [ + 0.001959877321496606, + 0.0022369101643562317, + 0.0025565754622220993, + 0.0032391189597547054, + 0.001978898188099265, + 0.0032494000624865294, + 0.002169701736420393, + 0.003456014906987548, + 0.0010360634187236428, + 0.0017441833624616265, + 0.0020471836905926466, + 0.0022487000096589327, + 0.0021409785840660334, + 0.0022493095602840185, + 0.0018128349911421537, + 0.0014531889464706182, + 0.002212466672062874, + 0.0036824881099164486, + 0.0034141356591135263, + 0.004261303227394819, + 0.004839694127440453, + 0.003197083482518792, + 0.002387921093031764, + 0.002357403514906764, + 0.0012110619572922587, + 0.0018409513868391514, + 0.0017867969581857324, + 0.002289022784680128, + 0.0018373033963143826, + 0.0017006773268803954, + 0.002020138083025813, + 0.000954562216065824, + 0.2023114562034607, + 0.06913179904222488, + 0.09164559841156006, + 0.02599690482020378, + 0.3244684338569641, + 0.05612204223871231, + 0.01441764086484909, + 0.021980421617627144, + 0.004608693532645702, + 0.0049738530069589615, + 0.002289806492626667, + 0.0020257318392395973, + 0.0021336840000003576, + 0.0011540622217580676, + 0.0015068118227645755, + 0.0014191375812515616, + 0.005022049881517887, + 0.005655186716467142, + 0.0061188689433038235, + 0.008221999742090702, + 0.002909916453063488, + 0.0029249375220388174, + 0.008929249830543995, + 0.004843664821237326, + 0.001730250078253448, + 0.0014446016866713762, + 0.0011063780402764678, + 0.00220413226634264, + 0.0008823463576845825, + 0.0006794463843107224, + 0.002307416871190071, + 0.001015635090880096, + 0.0032266636844724417, + 0.0034311343915760517, + 0.0056250556372106075, + 0.004057662561535835, + 0.0017412600573152304, + 0.0016485126689076424, + 0.004229556303471327, + 0.0016194517957046628, + 0.002266840310767293, + 0.0015291168820112944, + 0.0030156797729432583, + 0.0033689942210912704, + 0.0012042638845741749, + 0.0014415729092434049, + 0.0022979816421866417, + 0.0015421273419633508 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 6, + 2 + ], + "selected_reward": -0.02057892084121704 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static close-up from a slightly overhead angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene illuminated by the flickering flame of an ancient copper oil lamp?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a middle-aged man wearing a dark silk scarf and a wide-sleeved black robe sitting at a heavy dark wooden desk?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the man slowly unrolling a bamboo map bound by string with his hands?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is a feather fan crafted from white crane feathers with an ebony handle visible on the desk beside him?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0890224501490593 + ], + "opposite_logprob_tokens": [ + -2.464022397994995 + ], + "logprob_reward": -0.0890224501490593, + "yes_no_margin_reward": 2.375, + "token_credit_reward": -0.0890224501490593, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.049786850810050964, + 0.032062362879514694, + 0.024575894698500633, + 0.011486888863146305, + 0.6391185522079468, + 0.01835818588733673, + 0.09568587690591812, + 0.05130685493350029, + 0.04763103649020195, + 0.02998749539256096 + ], + "credit_map": [ + 0.004562389571219683, + 0.005475672427564859, + 0.006612775381654501, + 0.008241132833063602, + 0.00528592336922884, + 0.008616075851023197, + 0.004839991685003042, + 0.006152885500341654, + 0.002549211960285902, + 0.004930231254547834, + 0.005055657587945461, + 0.0045085777528584, + 0.004154598340392113, + 0.005644450429826975, + 0.00235163071192801, + 0.0028680083341896534, + 0.0020017321221530437, + 0.004038412123918533, + 0.0032139502000063658, + 0.0042213378474116325, + 0.00291537307202816, + 0.00316607509739697, + 0.0026301988400518894, + 0.0023888146970421076, + 0.0011873849434778094, + 0.0010874014114961028, + 0.0016059739282354712, + 0.001725401496514678, + 0.0014456750359386206, + 0.0017132856883108616, + 0.001826721359975636, + 0.0008950448245741427, + 0.16120125353336334, + 0.05636320635676384, + 0.06860774755477905, + 0.020851293578743935, + 0.2527608871459961, + 0.04980253800749779, + 0.011741084046661854, + 0.01779053546488285, + 0.002608777955174446, + 0.003893522545695305, + 0.0019769826903939247, + 0.002462795004248619, + 0.0027615935541689396, + 0.0010898303007707, + 0.002053177682682872, + 0.001511506736278534, + 0.011130345985293388, + 0.012734316289424896, + 0.009276198223233223, + 0.019511206075549126, + 0.0064744059927761555, + 0.007612862158566713, + 0.01623525097966194, + 0.012711288407444954, + 0.005982590839266777, + 0.004326539114117622, + 0.00468094926327467, + 0.012683396227657795, + 0.004075644072145224, + 0.005496591329574585, + 0.009718014858663082, + 0.004343126434832811, + 0.005823994055390358, + 0.004586240742355585, + 0.006672360934317112, + 0.008147169835865498, + 0.003925023600459099, + 0.004829473327845335, + 0.010753141716122627, + 0.0028936308808624744, + 0.003925841301679611, + 0.003667577402666211, + 0.00466111907735467, + 0.004326180089265108, + 0.0023002135567367077, + 0.0035169434268027544, + 0.0051944139413535595, + 0.002395207993686199 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 6, + 7 + ], + "selected_reward": -0.0890224501490593 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "static, close-up shot, shot from a slightly overhead angle, focuses the viewer's attention on a scene illuminated by the flickering flame of an ancient copper oil lamp. At the center of the frame lies a middle-aged strategist, wearing a dark silk scarf and a wide-sleeved black robe. He sits at a heavy, dark wooden desk, his brow furrowed, his eyes sharp as an eagle's, his gaze fixed intently on the object before him. The warm, dancing light casts deep shadows on his chiseled features, highlighting his lips pursed in deep thought. His slender, powerful hands, carefully, steadily, and deliberately, slowly unroll a bamboo map bound by string. As he moves, the polished bamboo strips rub against each other, creating a crisp, subtle rustling sound. The map's ink-drawn mountains and rivers, along with the cinnabar-marked military fortresses, gradually emerge in the dim light, exuding a sense of age and age. Beside him, on the desk, rests a feather fan crafted from white crane feathers, its ebony handle gleaming softly in the light. The painting's warm, deep tones, predominantly dark browns and oranges, create a stark contrast between light and shadow, evoking a sense of intense concentration and tension as one plots and decides a battle in the stillness of the night.", + "seed": 30524, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000101/group_2/index_6541_seed_30524_dp_5_rank_10_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 2, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.04748409166932106, + "vlm_logprob": -0.04748409166932106, + "vlm_yes_no_margin": 4.2, + "vlm_token_credit": -0.04748409166932106, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static close-up from a slightly overhead angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene illuminated by the flickering flame of an ancient copper oil lamp?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a middle-aged man wearing a dark silk scarf and a wide-sleeved black robe sitting at a heavy dark wooden desk?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the man slowly unrolling a bamboo map bound by string with his hands?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is a feather fan crafted from white crane feathers with an ebony handle visible on the desk beside him?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03299327194690704 + ], + "opposite_logprob_tokens": [ + -5.782993316650391 + ], + "logprob_reward": -0.03299327194690704, + "yes_no_margin_reward": 5.75, + "token_credit_reward": -0.03299327194690704, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12616491317749023, + 0.04589984938502312, + 0.0685710459947586, + 0.08020220696926117, + 0.07170145958662033, + 0.05985215678811073, + 0.07438333332538605, + 0.11508072167634964, + 0.15241187810897827, + 0.20573243498802185 + ], + "credit_map": [ + 0.02412792295217514, + 0.023969898000359535, + 0.01950785331428051, + 0.00798395462334156, + 0.011594593524932861, + 0.017942387610673904, + 0.014042843133211136, + 0.006995459087193012, + 0.007533876225352287, + 0.008408401161432266, + 0.0073223053477704525, + 0.0031444933265447617, + 0.0054496536031365395, + 0.007048745173960924, + 0.00454713124781847, + 0.0024452453944832087, + 0.007190964184701443, + 0.01049800030887127, + 0.012018436565995216, + 0.006274391897022724, + 0.009029516950249672, + 0.009023167192935944, + 0.009033001028001308, + 0.005503565538674593, + 0.012086101807653904, + 0.016609670594334602, + 0.014346717856824398, + 0.004364704713225365, + 0.006456068251281977, + 0.01805541291832924, + 0.005393458530306816, + 0.0028900739271193743, + 0.009941578842699528, + 0.01615561544895172, + 0.008048037067055702, + 0.00423977617174387, + 0.00931960716843605, + 0.014495650306344032, + 0.00586532149463892, + 0.003635867964476347, + 0.009203300811350346, + 0.00786928366869688, + 0.005432446952909231, + 0.003963501192629337, + 0.009036232717335224, + 0.007846840657293797, + 0.010450920090079308, + 0.006049633491784334, + 0.009821133688092232, + 0.014006268233060837, + 0.008283528499305248, + 0.006452283822000027, + 0.006731859873980284, + 0.016105137765407562, + 0.0070526013150811195, + 0.005930524319410324, + 0.0069199781864881516, + 0.03364956006407738, + 0.009290908463299274, + 0.0052194599993526936, + 0.01117316447198391, + 0.03461569547653198, + 0.007735480088740587, + 0.006476463284343481, + 0.011535657569766045, + 0.03416797146201134, + 0.019823139533400536, + 0.009584090672433376, + 0.02048204094171524, + 0.03025481291115284, + 0.013542372733354568, + 0.013021796010434628, + 0.019624968990683556, + 0.02856607735157013, + 0.05304647609591484, + 0.016261758282780647, + 0.028347253799438477, + 0.025128737092018127, + 0.01627315767109394, + 0.018484016880393028 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 8, + 0 + ], + "selected_reward": -0.03299327194690704 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static close-up from a slightly overhead angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene illuminated by the flickering flame of an ancient copper oil lamp?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a middle-aged man wearing a dark silk scarf and a wide-sleeved black robe sitting at a heavy dark wooden desk?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the man slowly unrolling a bamboo map bound by string with his hands?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is a feather fan crafted from white crane feathers with an ebony handle visible on the desk beside him?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04867101460695267 + ], + "opposite_logprob_tokens": [ + -3.048671007156372 + ], + "logprob_reward": -0.04867101460695267, + "yes_no_margin_reward": 3.0, + "token_credit_reward": -0.04867101460695267, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.26467904448509216, + 0.0691804587841034, + 0.09813845157623291, + 0.03338327258825302, + 0.1421009749174118, + 0.056551698595285416, + 0.05332170054316521, + 0.043721556663513184, + 0.2062385082244873, + 0.0326843298971653 + ], + "credit_map": [ + 0.06649110466241837, + 0.06293237954378128, + 0.038229066878557205, + 0.01277089398354292, + 0.019664272665977478, + 0.039438169449567795, + 0.012831817381083965, + 0.01232131477445364, + 0.009959508664906025, + 0.01287116203457117, + 0.011617517098784447, + 0.005073098000138998, + 0.008476002141833305, + 0.010437283664941788, + 0.006613518577069044, + 0.004132359754294157, + 0.007815150544047356, + 0.014992965385317802, + 0.01975083351135254, + 0.006516657304018736, + 0.012394076213240623, + 0.020667433738708496, + 0.010391980409622192, + 0.005609350278973579, + 0.005264153238385916, + 0.008170957677066326, + 0.005121119786053896, + 0.0018506809137761593, + 0.0028300685808062553, + 0.0069321999326348305, + 0.0020185920875519514, + 0.0011954986257478595, + 0.030363010242581367, + 0.02276756428182125, + 0.0164586640894413, + 0.009711344726383686, + 0.016080088913440704, + 0.0219041109085083, + 0.015598934143781662, + 0.00921724084764719, + 0.006451742723584175, + 0.007743163499981165, + 0.00648133922368288, + 0.004446823615580797, + 0.007781762629747391, + 0.007727962452918291, + 0.008160356432199478, + 0.007758541963994503, + 0.006083469372242689, + 0.008577853441238403, + 0.0074004787020385265, + 0.00649566575884819, + 0.00735202943906188, + 0.006581871770322323, + 0.005123737268149853, + 0.005706590134650469, + 0.004979812074452639, + 0.007114836014807224, + 0.007706965319812298, + 0.004601623862981796, + 0.005589722190052271, + 0.006422223523259163, + 0.00292093469761312, + 0.004385435022413731, + 0.015898292884230614, + 0.04344547912478447, + 0.031130895018577576, + 0.021335430443286896, + 0.026782706379890442, + 0.03845806419849396, + 0.017738347873091698, + 0.011449269950389862, + 0.0028247805312275887, + 0.004657611716538668, + 0.002915129065513611, + 0.0024330331943929195, + 0.0041197375394403934, + 0.0037076708395034075, + 0.006297612097114325, + 0.0057287514209747314 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 4 + ], + "selected_reward": -0.04867101460695267 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static close-up from a slightly overhead angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene illuminated by the flickering flame of an ancient copper oil lamp?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a middle-aged man wearing a dark silk scarf and a wide-sleeved black robe sitting at a heavy dark wooden desk?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the man slowly unrolling a bamboo map bound by string with his hands?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is a feather fan crafted from white crane feathers with an ebony handle visible on the desk beside him?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.009781762026250362 + ], + "opposite_logprob_tokens": [ + -4.634781837463379 + ], + "logprob_reward": -0.009781762026250362, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.009781762026250362, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1834162324666977, + 0.06748133152723312, + 0.09008384495973587, + 0.03196613863110542, + 0.042313672602176666, + 0.04170054569840431, + 0.174835667014122, + 0.2080845981836319, + 0.0648077204823494, + 0.09531022608280182 + ], + "credit_map": [ + 0.031609755009412766, + 0.029571061953902245, + 0.03529877960681915, + 0.011471135541796684, + 0.021266229450702667, + 0.023658905178308487, + 0.01948992908000946, + 0.011050430126488209, + 0.005490218289196491, + 0.005177921149879694, + 0.015236292034387589, + 0.005401642993092537, + 0.01068620290607214, + 0.011914688162505627, + 0.009043160825967789, + 0.004531200043857098, + 0.007749737706035376, + 0.014583412557840347, + 0.014601628296077251, + 0.007870829664170742, + 0.010703071020543575, + 0.015962885692715645, + 0.012500300072133541, + 0.0061119734309613705, + 0.0032306092325598, + 0.0042190756648778915, + 0.006056787446141243, + 0.002447767648845911, + 0.004601804539561272, + 0.007096690591424704, + 0.0026268090587109327, + 0.0016865922370925546, + 0.007938620634377003, + 0.005055469460785389, + 0.004616034682840109, + 0.002988469786942005, + 0.0046914988197386265, + 0.005722758825868368, + 0.008526073768734932, + 0.0027747449930757284, + 0.007734920829534531, + 0.006644539535045624, + 0.007613730616867542, + 0.002901553176343441, + 0.00515340082347393, + 0.006360811647027731, + 0.002830708399415016, + 0.002460875315591693, + 0.019623370841145515, + 0.02310415916144848, + 0.020532019436359406, + 0.02024494670331478, + 0.03219239413738251, + 0.023984145373106003, + 0.015545237809419632, + 0.019609373062849045, + 0.010908281430602074, + 0.06520362943410873, + 0.016875654458999634, + 0.010219192132353783, + 0.014844618737697601, + 0.06671778857707977, + 0.01206507720053196, + 0.011250334791839123, + 0.005379586946219206, + 0.012034693732857704, + 0.009504320099949837, + 0.007385862525552511, + 0.007855861447751522, + 0.01093791052699089, + 0.005905701778829098, + 0.005803776904940605, + 0.008471393957734108, + 0.012218001298606396, + 0.019405553117394447, + 0.00936425942927599, + 0.012683720327913761, + 0.011891082860529423, + 0.009761279448866844, + 0.011514924466609955 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 0, + 6 + ], + "selected_reward": -0.009781762026250362 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static close-up from a slightly overhead angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene illuminated by the flickering flame of an ancient copper oil lamp?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a middle-aged man wearing a dark silk scarf and a wide-sleeved black robe sitting at a heavy dark wooden desk?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the man slowly unrolling a bamboo map bound by string with his hands?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is a feather fan crafted from white crane feathers with an ebony handle visible on the desk beside him?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.003191736526787281 + ], + "opposite_logprob_tokens": [ + -5.753191947937012 + ], + "logprob_reward": -0.003191736526787281, + "yes_no_margin_reward": 5.75, + "token_credit_reward": -0.003191736526787281, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1191382110118866, + 0.055468447506427765, + 0.04718082398176193, + 0.03639712184667587, + 0.04795587435364723, + 0.18885502219200134, + 0.07439620792865753, + 0.21492181718349457, + 0.08843566477298737, + 0.12725083529949188 + ], + "credit_map": [ + 0.013794036582112312, + 0.01409406028687954, + 0.022625043988227844, + 0.008232439868152142, + 0.018751060590147972, + 0.015103092417120934, + 0.015468120574951172, + 0.011070344597101212, + 0.0038660201244056225, + 0.004224902018904686, + 0.009572434239089489, + 0.008604424074292183, + 0.005941472016274929, + 0.008452698588371277, + 0.008876384235918522, + 0.005930109415203333, + 0.004754338413476944, + 0.006847746670246124, + 0.007622858975082636, + 0.00510111590847373, + 0.0052208066917955875, + 0.007108679506927729, + 0.006605721078813076, + 0.003919555805623531, + 0.004290047567337751, + 0.007847501896321774, + 0.004499644041061401, + 0.0025894406717270613, + 0.004532394465059042, + 0.00548130227252841, + 0.005078438203781843, + 0.002078348770737648, + 0.0046304031275212765, + 0.005597262643277645, + 0.005547130014747381, + 0.005508793983608484, + 0.0036936018150299788, + 0.006780071184039116, + 0.011022565886378288, + 0.005176043603569269, + 0.02306368574500084, + 0.028821676969528198, + 0.030245231464505196, + 0.016766253858804703, + 0.029864203184843063, + 0.02502322569489479, + 0.018109653145074844, + 0.016961093991994858, + 0.009991963393986225, + 0.012325240299105644, + 0.011362569406628609, + 0.006233870517462492, + 0.008781308308243752, + 0.011492324061691761, + 0.007912122644484043, + 0.006296804640442133, + 0.012181268073618412, + 0.06450735032558441, + 0.02146696299314499, + 0.012838844209909439, + 0.016142142936587334, + 0.06642936170101166, + 0.010496311821043491, + 0.010859556496143341, + 0.005819899030029774, + 0.022248342633247375, + 0.007980972528457642, + 0.00496139470487833, + 0.008655451238155365, + 0.020495368167757988, + 0.008686928078532219, + 0.009587305597960949, + 0.013240424916148186, + 0.020057689398527145, + 0.015616700053215027, + 0.009354553185403347, + 0.013922575861215591, + 0.01551112812012434, + 0.019273841753602028, + 0.02027391828596592 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 5, + 9 + ], + "selected_reward": -0.003191736526787281 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static close-up from a slightly overhead angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene illuminated by the flickering flame of an ancient copper oil lamp?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a middle-aged man wearing a dark silk scarf and a wide-sleeved black robe sitting at a heavy dark wooden desk?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the man slowly unrolling a bamboo map bound by string with his hands?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is a feather fan crafted from white crane feathers with an ebony handle visible on the desk beside him?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.14278267323970795 + ], + "opposite_logprob_tokens": [ + -2.017782688140869 + ], + "logprob_reward": -0.14278267323970795, + "yes_no_margin_reward": 1.875, + "token_credit_reward": -0.14278267323970795, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13473573327064514, + 0.03410525247454643, + 0.013656405732035637, + 0.020177140831947327, + 0.03261773660778999, + 0.03724541887640953, + 0.27953988313674927, + 0.2710398733615875, + 0.09660448879003525, + 0.08027812093496323 + ], + "credit_map": [ + 0.0320492684841156, + 0.03256947919726372, + 0.0214662104845047, + 0.006805780343711376, + 0.009450389072299004, + 0.019616875797510147, + 0.007651348598301411, + 0.005126388743519783, + 0.0022654433269053698, + 0.0032716512214392424, + 0.006304209586232901, + 0.0039965491741895676, + 0.004401188809424639, + 0.0055839079432189465, + 0.005168289411813021, + 0.003114015329629183, + 0.0014857654459774494, + 0.002285317750647664, + 0.0020969316828995943, + 0.0013434061547741294, + 0.0017580303829163313, + 0.0022087693214416504, + 0.0017462632386013865, + 0.0007319211727008224, + 0.0025502631906419992, + 0.0043098279275000095, + 0.0024408409371972084, + 0.0015540402382612228, + 0.0025465490762144327, + 0.0034604654647409916, + 0.001912592677399516, + 0.0014025624841451645, + 0.006837458349764347, + 0.004991300869733095, + 0.0033487335313111544, + 0.0018449757480993867, + 0.0040872786194086075, + 0.007132195867598057, + 0.0024681980721652508, + 0.0019075957825407386, + 0.0037175484467297792, + 0.005471542477607727, + 0.0035591833293437958, + 0.0034166420809924603, + 0.0061044334433972836, + 0.0050797308795154095, + 0.006986353080719709, + 0.002909984439611435, + 0.03676692396402359, + 0.042606156319379807, + 0.034723103046417236, + 0.032226357609033585, + 0.05329802259802818, + 0.032565440982580185, + 0.020214099436998367, + 0.027139753103256226, + 0.023621192201972008, + 0.051865823566913605, + 0.03968384116888046, + 0.0254011619836092, + 0.032203201204538345, + 0.05066334456205368, + 0.02395930141210556, + 0.023642005398869514, + 0.009021338075399399, + 0.014496728777885437, + 0.016944140195846558, + 0.008533630520105362, + 0.01698550023138523, + 0.012062327936291695, + 0.009233846329152584, + 0.009326980449259281, + 0.007493320386856794, + 0.010728689841926098, + 0.016070444136857986, + 0.006619204767048359, + 0.01281663402915001, + 0.010857397690415382, + 0.00670771487057209, + 0.008984715677797794 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 7, + 0 + ], + "selected_reward": -0.14278267323970795 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "static, close-up shot, shot from a slightly overhead angle, focuses the viewer's attention on a scene illuminated by the flickering flame of an ancient copper oil lamp. At the center of the frame lies a middle-aged strategist, wearing a dark silk scarf and a wide-sleeved black robe. He sits at a heavy, dark wooden desk, his brow furrowed, his eyes sharp as an eagle's, his gaze fixed intently on the object before him. The warm, dancing light casts deep shadows on his chiseled features, highlighting his lips pursed in deep thought. His slender, powerful hands, carefully, steadily, and deliberately, slowly unroll a bamboo map bound by string. As he moves, the polished bamboo strips rub against each other, creating a crisp, subtle rustling sound. The map's ink-drawn mountains and rivers, along with the cinnabar-marked military fortresses, gradually emerge in the dim light, exuding a sense of age and age. Beside him, on the desk, rests a feather fan crafted from white crane feathers, its ebony handle gleaming softly in the light. The painting's warm, deep tones, predominantly dark browns and oranges, create a stark contrast between light and shadow, evoking a sense of intense concentration and tension as one plots and decides a battle in the stillness of the night.", + "seed": 30525, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000101/group_2/index_6541_seed_30525_dp_5_rank_10_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 2, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.05253446772694588, + "vlm_logprob": -0.05253446772694588, + "vlm_yes_no_margin": 3.55, + "vlm_token_credit": -0.05253446772694588, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static close-up from a slightly overhead angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene illuminated by the flickering flame of an ancient copper oil lamp?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a middle-aged man wearing a dark silk scarf and a wide-sleeved black robe sitting at a heavy dark wooden desk?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the man slowly unrolling a bamboo map bound by string with his hands?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is a feather fan crafted from white crane feathers with an ebony handle visible on the desk beside him?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.05382226034998894 + ], + "opposite_logprob_tokens": [ + -5.303822040557861 + ], + "logprob_reward": -0.05382226034998894, + "yes_no_margin_reward": 5.25, + "token_credit_reward": -0.05382226034998894, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.33705392479896545, + 0.18721488118171692, + 0.08012817054986954, + 0.05715854465961456, + 0.03613051027059555, + 0.030332475900650024, + 0.081138476729393, + 0.05233510211110115, + 0.04878823831677437, + 0.0897197276353836 + ], + "credit_map": [ + 0.03604020178318024, + 0.0432407520711422, + 0.028527526184916496, + 0.034414853900671005, + 0.04707469046115875, + 0.04963447153568268, + 0.03481808304786682, + 0.06330329179763794, + 0.015863502398133278, + 0.022606758400797844, + 0.011249086819589138, + 0.01305665448307991, + 0.014819980598986149, + 0.03069406747817993, + 0.03167436644434929, + 0.04725044220685959, + 0.008425336331129074, + 0.01724526472389698, + 0.005277429241687059, + 0.008948620408773422, + 0.010725966654717922, + 0.013382632285356522, + 0.009852902963757515, + 0.006270013749599457, + 0.00491825258359313, + 0.009511561132967472, + 0.00344859529286623, + 0.00500138197094202, + 0.005637177266180515, + 0.009769383817911148, + 0.006279745139181614, + 0.012592446990311146, + 0.003838749835267663, + 0.004949680529534817, + 0.002896354766562581, + 0.003040905576199293, + 0.00575388316065073, + 0.005145966075360775, + 0.0051129539497196674, + 0.0053920135833323, + 0.002922698622569442, + 0.004426382482051849, + 0.0020157943945378065, + 0.0030548335053026676, + 0.003773365868255496, + 0.005185773130506277, + 0.0027433924842625856, + 0.006210232153534889, + 0.011067107319831848, + 0.016725771129131317, + 0.004965252708643675, + 0.005641115829348564, + 0.013050416484475136, + 0.014415247365832329, + 0.00798686034977436, + 0.0072867004200816154, + 0.006623197812587023, + 0.005266846157610416, + 0.00530992029234767, + 0.007211877033114433, + 0.00735781854018569, + 0.007871042937040329, + 0.005073287524282932, + 0.007621110882610083, + 0.006325105670839548, + 0.005000099074095488, + 0.006776719819754362, + 0.005014112684875727, + 0.008393000811338425, + 0.0077115874737501144, + 0.0040877265855669975, + 0.005479881539940834, + 0.010684993118047714, + 0.009281069971621037, + 0.010151667520403862, + 0.014415684156119823, + 0.01029268093407154, + 0.012549585662782192, + 0.009540102444589138, + 0.012803935445845127 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 9 + ], + "selected_reward": -0.05382226034998894 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static close-up from a slightly overhead angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene illuminated by the flickering flame of an ancient copper oil lamp?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a middle-aged man wearing a dark silk scarf and a wide-sleeved black robe sitting at a heavy dark wooden desk?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the man slowly unrolling a bamboo map bound by string with his hands?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is a feather fan crafted from white crane feathers with an ebony handle visible on the desk beside him?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07004882395267487 + ], + "opposite_logprob_tokens": [ + -2.6950488090515137 + ], + "logprob_reward": -0.07004882395267487, + "yes_no_margin_reward": 2.625, + "token_credit_reward": -0.07004882395267487, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3259652256965637, + 0.1318599432706833, + 0.06279926747083664, + 0.04002721980214119, + 0.09755314141511917, + 0.025439849123358727, + 0.034998781979084015, + 0.08445216715335846, + 0.15365581214427948, + 0.04324856027960777 + ], + "credit_map": [ + 0.03143680840730667, + 0.043854136019945145, + 0.02211877331137657, + 0.04184006154537201, + 0.05354355648159981, + 0.050066303461790085, + 0.04984980449080467, + 0.033255793154239655, + 0.013491748832166195, + 0.014956137165427208, + 0.005712795536965132, + 0.010539848357439041, + 0.014995768666267395, + 0.021435517817735672, + 0.0200644563883543, + 0.03066367469727993, + 0.007569680921733379, + 0.008717706426978111, + 0.0035582755226641893, + 0.00831303745508194, + 0.011447030119597912, + 0.0095420703291893, + 0.007149599492549896, + 0.006501868832856417, + 0.0027313020545989275, + 0.006446785293519497, + 0.002337283920496702, + 0.004860192537307739, + 0.007227743044495583, + 0.0070070503279566765, + 0.0038493068423122168, + 0.005567553918808699, + 0.013112462125718594, + 0.014715162105858326, + 0.0059461817145347595, + 0.007437755353748798, + 0.015322481282055378, + 0.015629842877388, + 0.008416947908699512, + 0.016972312703728676, + 0.00286038382910192, + 0.005144753027707338, + 0.0013197694206610322, + 0.0016940587665885687, + 0.0048673017881810665, + 0.004640202037990093, + 0.0018537853611633182, + 0.003059596288949251, + 0.004785306751728058, + 0.00470922514796257, + 0.00246623158454895, + 0.003325792495161295, + 0.005182799883186817, + 0.005753065459430218, + 0.004263227339833975, + 0.00451313192024827, + 0.008650307543575764, + 0.013853504322469234, + 0.007200517691671848, + 0.011055982671678066, + 0.010154152289032936, + 0.01396390050649643, + 0.009188915602862835, + 0.010384880006313324, + 0.016796447336673737, + 0.0155378058552742, + 0.022120798006653786, + 0.020052876323461533, + 0.023286089301109314, + 0.022430667653679848, + 0.013825495727360249, + 0.019605636596679688, + 0.005084151867777109, + 0.004260298330336809, + 0.00334260449744761, + 0.004216271918267012, + 0.006501941476017237, + 0.0060640363954007626, + 0.005166069138795137, + 0.00861318502575159 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 1 + ], + "selected_reward": -0.07004882395267487 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static close-up from a slightly overhead angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene illuminated by the flickering flame of an ancient copper oil lamp?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a middle-aged man wearing a dark silk scarf and a wide-sleeved black robe sitting at a heavy dark wooden desk?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the man slowly unrolling a bamboo map bound by string with his hands?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is a feather fan crafted from white crane feathers with an ebony handle visible on the desk beside him?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.016077518463134766 + ], + "opposite_logprob_tokens": [ + -4.141077518463135 + ], + "logprob_reward": -0.016077518463134766, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.016077518463134766, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3765854835510254, + 0.18119587004184723, + 0.07479119300842285, + 0.042747724801301956, + 0.05120048299431801, + 0.03186766803264618, + 0.08660444617271423, + 0.03228128328919411, + 0.04275256767868996, + 0.07997319847345352 + ], + "credit_map": [ + 0.03461291640996933, + 0.05673946067690849, + 0.021599026396870613, + 0.033944252878427505, + 0.07307779043912888, + 0.06563396751880646, + 0.05369879677891731, + 0.03727927803993225, + 0.014377189800143242, + 0.030445193871855736, + 0.01030641794204712, + 0.01354243140667677, + 0.03255098685622215, + 0.032060958445072174, + 0.021787360310554504, + 0.0261253472417593, + 0.006532491184771061, + 0.010592672042548656, + 0.004476490430533886, + 0.005470765754580498, + 0.015545117668807507, + 0.01342783123254776, + 0.007068164646625519, + 0.011677666567265987, + 0.0038482772652059793, + 0.007567838300019503, + 0.002502080053091049, + 0.004138398449867964, + 0.008778824470937252, + 0.007942397147417068, + 0.004411764908581972, + 0.0035581497941166162, + 0.006788306869566441, + 0.0074082850478589535, + 0.0025474110152572393, + 0.005634985864162445, + 0.00845349207520485, + 0.0069886124692857265, + 0.0061263288371264935, + 0.007253066170960665, + 0.0031606059055775404, + 0.0047769551165401936, + 0.0016951810102909803, + 0.00375299877487123, + 0.005720274988561869, + 0.004560250323265791, + 0.004321615677326918, + 0.003879789961501956, + 0.008430913090705872, + 0.017014699056744576, + 0.00441923039034009, + 0.006996206473559141, + 0.01829400286078453, + 0.016090290620923042, + 0.008018399588763714, + 0.007340712007135153, + 0.003640069393441081, + 0.007114740554243326, + 0.0019418104784563184, + 0.0038077314384281635, + 0.004356487654149532, + 0.005500365514308214, + 0.0031323765870183706, + 0.002787709468975663, + 0.0034942361526191235, + 0.005524784326553345, + 0.002439612289890647, + 0.004307866096496582, + 0.010446708649396896, + 0.006942620035260916, + 0.004203952848911285, + 0.005392791237682104, + 0.007488214876502752, + 0.008415103890001774, + 0.00841005239635706, + 0.007754393853247166, + 0.015720218420028687, + 0.012422936037182808, + 0.006931668147444725, + 0.012830615043640137 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 6 + ], + "selected_reward": -0.016077518463134766 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static close-up from a slightly overhead angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene illuminated by the flickering flame of an ancient copper oil lamp?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a middle-aged man wearing a dark silk scarf and a wide-sleeved black robe sitting at a heavy dark wooden desk?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the man slowly unrolling a bamboo map bound by string with his hands?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is a feather fan crafted from white crane feathers with an ebony handle visible on the desk beside him?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0336964875459671 + ], + "opposite_logprob_tokens": [ + -3.408696413040161 + ], + "logprob_reward": -0.0336964875459671, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.0336964875459671, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3284797668457031, + 0.12270528823137283, + 0.09764479845762253, + 0.05849593132734299, + 0.05405769124627113, + 0.032302070409059525, + 0.050616342574357986, + 0.04674752429127693, + 0.06548906117677689, + 0.14346149563789368 + ], + "credit_map": [ + 0.03251201659440994, + 0.04369625821709633, + 0.021851446479558945, + 0.04142245650291443, + 0.050347257405519485, + 0.05529453605413437, + 0.04442029073834419, + 0.03893553093075752, + 0.012517288327217102, + 0.018323048949241638, + 0.009090988896787167, + 0.01456469390541315, + 0.0157649926841259, + 0.018485700711607933, + 0.02272941544651985, + 0.011229157447814941, + 0.009394404478371143, + 0.015337872318923473, + 0.006704457104206085, + 0.012048098258674145, + 0.013348039239645004, + 0.015170086175203323, + 0.01535410713404417, + 0.010287736542522907, + 0.005805563181638718, + 0.012232490815222263, + 0.002743096323683858, + 0.006338497158139944, + 0.00816342793405056, + 0.010113357566297054, + 0.006059234496206045, + 0.00704026548191905, + 0.009738747030496597, + 0.009105111472308636, + 0.0032388719264417887, + 0.005379186011850834, + 0.007021011784672737, + 0.008128604851663113, + 0.005784871056675911, + 0.005661288276314735, + 0.004288048949092627, + 0.004479759372770786, + 0.0016768084606155753, + 0.003604406723752618, + 0.0054757799953222275, + 0.005224450957030058, + 0.00413178326562047, + 0.0034210323356091976, + 0.0059341066516935825, + 0.010330204851925373, + 0.0023740949109196663, + 0.003368234960362315, + 0.009692645631730556, + 0.009767431765794754, + 0.004224020522087812, + 0.0049256025813519955, + 0.0050765457563102245, + 0.005582040641456842, + 0.0036009671166539192, + 0.0068526435643434525, + 0.008211088366806507, + 0.00682397885248065, + 0.0052574132569134235, + 0.0053428472019732, + 0.0068021719343960285, + 0.007876248098909855, + 0.0081493454053998, + 0.008621800690889359, + 0.008860728703439236, + 0.009066153317689896, + 0.008047565817832947, + 0.008065047673881054, + 0.017352869734168053, + 0.016505779698491096, + 0.011519395746290684, + 0.013097109273076057, + 0.019856087863445282, + 0.019901754334568977, + 0.02342478558421135, + 0.021803714334964752 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 1 + ], + "selected_reward": -0.0336964875459671 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static close-up from a slightly overhead angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene illuminated by the flickering flame of an ancient copper oil lamp?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a middle-aged man wearing a dark silk scarf and a wide-sleeved black robe sitting at a heavy dark wooden desk?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the man slowly unrolling a bamboo map bound by string with his hands?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is a feather fan crafted from white crane feathers with an ebony handle visible on the desk beside him?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.08902724832296371 + ], + "opposite_logprob_tokens": [ + -2.464027166366577 + ], + "logprob_reward": -0.08902724832296371, + "yes_no_margin_reward": 2.375, + "token_credit_reward": -0.08902724832296371, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.35365650057792664, + 0.12388097494840622, + 0.06809300184249878, + 0.07141193747520447, + 0.03393207862973213, + 0.023587847128510475, + 0.042823705822229385, + 0.18283170461654663, + 0.04630295932292938, + 0.05347931757569313 + ], + "credit_map": [ + 0.034119874238967896, + 0.04825962334871292, + 0.024510903283953667, + 0.046608079224824905, + 0.045090679079294205, + 0.05035940930247307, + 0.057557862251996994, + 0.04715008661150932, + 0.010690249502658844, + 0.017207887023687363, + 0.008746478706598282, + 0.01024201326072216, + 0.012146745808422565, + 0.020975809544324875, + 0.013322515413165092, + 0.030549272894859314, + 0.006026349030435085, + 0.008968841284513474, + 0.004122467245906591, + 0.010385618545114994, + 0.011206143535673618, + 0.01136535219848156, + 0.008568650111556053, + 0.00744958221912384, + 0.005913276690989733, + 0.011171262711286545, + 0.004989824257791042, + 0.009759721346199512, + 0.013221140950918198, + 0.012376351282000542, + 0.007292787544429302, + 0.00668756989762187, + 0.0034582403022795916, + 0.004406198859214783, + 0.0019070182461291552, + 0.0062783267349004745, + 0.00531554501503706, + 0.004929035436362028, + 0.0037774676457047462, + 0.003860246390104294, + 0.003092754865065217, + 0.0031874978449195623, + 0.0013767261989414692, + 0.002670709043741226, + 0.003073248779401183, + 0.003432824509218335, + 0.0028976870235055685, + 0.0038563990965485573, + 0.006013000383973122, + 0.008477064780890942, + 0.0037375115789473057, + 0.004380380269140005, + 0.004857281222939491, + 0.007525779772549868, + 0.002954315859824419, + 0.004878371953964233, + 0.019053852185606956, + 0.02476700209081173, + 0.015644267201423645, + 0.023871254175901413, + 0.026417454704642296, + 0.03063124231994152, + 0.016118532046675682, + 0.026328101754188538, + 0.006600068416446447, + 0.004761658608913422, + 0.00720446091145277, + 0.00499340333044529, + 0.005236998200416565, + 0.007150009740144014, + 0.003870557527989149, + 0.006485800724476576, + 0.008706644177436829, + 0.004865074530243874, + 0.00815967470407486, + 0.005495836492627859, + 0.004730340093374252, + 0.00807420164346695, + 0.005711977835744619, + 0.007735568098723888 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 7, + 1 + ], + "selected_reward": -0.08902724832296371 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000101/group_3/index_tensor([6062, 6062])_group_3_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000101/group_3/index_tensor([6062, 6062])_group_3_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..07940f2218a7acd0382ada825ca3422a1bfe1883 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000101/group_3/index_tensor([6062, 6062])_group_3_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "low-angle, medium close-up, with a sharp digital look, remains static, capturing a figure at the wheel of a sleek black motorcycle. The subject, a woman with a determined expression and windswept black hair, wears a well-worn black leather jacket, its wrinkled texture and subtle sheen accentuated by the lighting. With ease, she swings her leg onto the seat, her free hand momentarily resting on the handlebars for balance. Behind her, a massive tidal power station looms, its massive concrete walls spanning the frame, appearing monolithic and imposing. The low, golden glow of the setting sun sweeps across the scene, casting long, dramatic shadows that stretch across the ground and sculpt every surface with intense contrast. This harsh light captures the grain of the leather and the rough, weathered texture of the concrete structure, creating a powerful portrait that emphasizes the stark contrast between the confident, self-assured human subject and the overwhelming industrial scale of her surroundings, all captured with crisp, high-fidelity detail.", + "seed": 24671, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000101/group_3/index_6062_seed_24671_dp_6_rank_12_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 3, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.08729746947064995, + "vlm_logprob": -0.08729746947064995, + "vlm_yes_no_margin": 3.6, + "vlm_token_credit": -0.08729746947064995, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera angle low and the framing a medium close-up?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the subject a woman with windswept black hair wearing a well-worn black leather jacket?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the woman performing the action of swinging her leg onto the seat of a sleek black motorcycle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a massive tidal power station with concrete walls visible in the background behind the motorcycle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit by a low golden glow from a setting sun casting long, dramatic shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.1455429047346115 + ], + "opposite_logprob_tokens": [ + -3.520542860031128 + ], + "logprob_reward": -0.1455429047346115, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.1455429047346115, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.17741696536540985, + 0.07470695674419403, + 0.24284936487674713, + 0.2112414687871933, + 0.037812598049640656, + 0.04542088508605957, + 0.041764989495277405, + 0.04459299147129059, + 0.07900594919919968, + 0.04518785700201988 + ], + "credit_map": [ + 0.016876373440027237, + 0.03499015048146248, + 0.031621046364307404, + 0.021143928170204163, + 0.027586452662944794, + 0.020599085837602615, + 0.009685011580586433, + 0.014914914965629578, + 0.00723986467346549, + 0.012073280289769173, + 0.009934988804161549, + 0.010144523344933987, + 0.014137901365756989, + 0.009478695690631866, + 0.006298283115029335, + 0.0053994194604456425, + 0.0583990141749382, + 0.09785410016775131, + 0.025740787386894226, + 0.017462899908423424, + 0.012760908342897892, + 0.01567285880446434, + 0.009704292751848698, + 0.005254491697996855, + 0.03409671038389206, + 0.020292121917009354, + 0.013071823865175247, + 0.07837449759244919, + 0.014330943115055561, + 0.01272979099303484, + 0.03207624703645706, + 0.006269320845603943, + 0.004268720746040344, + 0.004506440367549658, + 0.006254169158637524, + 0.006605156697332859, + 0.0038564710412174463, + 0.0034239310771226883, + 0.005613502115011215, + 0.0032842070795595646, + 0.00503517035394907, + 0.005253008101135492, + 0.007557870354503393, + 0.004678219091147184, + 0.0061930580995976925, + 0.006112034432590008, + 0.004700330086052418, + 0.005891190376132727, + 0.008643066510558128, + 0.005419482477009296, + 0.007206138223409653, + 0.006274388637393713, + 0.004246264696121216, + 0.003579093609005213, + 0.004191704094409943, + 0.0022048528771847486, + 0.00790473259985447, + 0.005205442663282156, + 0.007365964353084564, + 0.004153538029640913, + 0.007122666575014591, + 0.003951637540012598, + 0.0050597102381289005, + 0.0038293006364256144, + 0.010396016761660576, + 0.014792047441005707, + 0.009604018181562424, + 0.01218077726662159, + 0.010707652196288109, + 0.006361638195812702, + 0.008028296753764153, + 0.00693550193682313, + 0.00717662088572979, + 0.008124311454594135, + 0.006117142736911774, + 0.006162747740745544, + 0.006718292832374573, + 0.003613793756812811, + 0.004338127560913563, + 0.002936821198090911 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 3, + 0 + ], + "selected_reward": -0.1455429047346115 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera angle low and the framing a medium close-up?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the subject a woman with windswept black hair wearing a well-worn black leather jacket?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the woman performing the action of swinging her leg onto the seat of a sleek black motorcycle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a massive tidal power station with concrete walls visible in the background behind the motorcycle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit by a low golden glow from a setting sun casting long, dramatic shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.048705194145441055 + ], + "opposite_logprob_tokens": [ + -3.0487051010131836 + ], + "logprob_reward": -0.048705194145441055, + "yes_no_margin_reward": 3.0, + "token_credit_reward": -0.048705194145441055, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.17908070981502533, + 0.05381326377391815, + 0.1710917204618454, + 0.2552317678928375, + 0.044419750571250916, + 0.07398804277181625, + 0.04986117407679558, + 0.07610232383012772, + 0.07208209484815598, + 0.024329153820872307 + ], + "credit_map": [ + 0.017084818333387375, + 0.03651797026395798, + 0.019523920491337776, + 0.017691541463136673, + 0.03624548390507698, + 0.02579999528825283, + 0.013105075806379318, + 0.013111905194818974, + 0.005811396986246109, + 0.0099320188164711, + 0.008367086760699749, + 0.009366813115775585, + 0.007438749074935913, + 0.005130118224769831, + 0.00464373966678977, + 0.003123340429738164, + 0.031111042946577072, + 0.054876651614904404, + 0.02151404321193695, + 0.021483412012457848, + 0.014492368325591087, + 0.01140642911195755, + 0.010713516734540462, + 0.005494261626154184, + 0.04915952682495117, + 0.016552677378058434, + 0.009606835432350636, + 0.10881447792053223, + 0.013762515038251877, + 0.008255161345005035, + 0.04329740256071091, + 0.00578314857557416, + 0.005893397144973278, + 0.005169011652469635, + 0.00643972447142005, + 0.008746368810534477, + 0.004336722195148468, + 0.004116217605769634, + 0.0068699996918439865, + 0.0028483066707849503, + 0.006251982878893614, + 0.01206508930772543, + 0.008690878748893738, + 0.01398789044469595, + 0.011177181266248226, + 0.006988584529608488, + 0.008556478656828403, + 0.006269956007599831, + 0.009147786535322666, + 0.004389405250549316, + 0.005101613700389862, + 0.011565925553441048, + 0.006400476209819317, + 0.003464148147031665, + 0.005506323184818029, + 0.004285494331270456, + 0.012028911150991917, + 0.00624797772616148, + 0.012037710286676884, + 0.006220733281224966, + 0.013130429200828075, + 0.008456407114863396, + 0.009671058505773544, + 0.008309095166623592, + 0.00648382306098938, + 0.014986439608037472, + 0.005186418537050486, + 0.01718754507601261, + 0.008234602399170399, + 0.0055630928836762905, + 0.007342998869717121, + 0.007097176741808653, + 0.004599584732204676, + 0.003259629476815462, + 0.0037977397441864014, + 0.003731343662366271, + 0.00350696942768991, + 0.0018105525523424149, + 0.0023172323126345873, + 0.0013061023782938719 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 0, + 2 + ], + "selected_reward": -0.048705194145441055 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera angle low and the framing a medium close-up?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the subject a woman with windswept black hair wearing a well-worn black leather jacket?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the woman performing the action of swinging her leg onto the seat of a sleek black motorcycle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a massive tidal power station with concrete walls visible in the background behind the motorcycle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit by a low golden glow from a setting sun casting long, dramatic shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0142021169885993 + ], + "opposite_logprob_tokens": [ + -4.264202117919922 + ], + "logprob_reward": -0.0142021169885993, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.0142021169885993, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2629091441631317, + 0.09137827903032303, + 0.13956715166568756, + 0.18528315424919128, + 0.04738288000226021, + 0.047738704830408096, + 0.021851006895303726, + 0.020321857184171677, + 0.1445741057395935, + 0.03899364918470383 + ], + "credit_map": [ + 0.02532169409096241, + 0.033765628933906555, + 0.047530967742204666, + 0.026798943057656288, + 0.048086002469062805, + 0.03673606738448143, + 0.01809185929596424, + 0.026578010991215706, + 0.008160538040101528, + 0.013919088989496231, + 0.01365902554243803, + 0.01362937968224287, + 0.01573082059621811, + 0.010063721798360348, + 0.01042264699935913, + 0.005793069023638964, + 0.019532330334186554, + 0.0332578681409359, + 0.0163534227758646, + 0.012555519118905067, + 0.02609153650701046, + 0.013025698252022266, + 0.011062956415116787, + 0.007687832228839397, + 0.03898963704705238, + 0.009767269715666771, + 0.008241986855864525, + 0.07551319897174835, + 0.01084849052131176, + 0.009166145697236061, + 0.029107462614774704, + 0.0036489819176495075, + 0.00889892503619194, + 0.0033566222991794348, + 0.003624846925958991, + 0.014608399011194706, + 0.003502576146274805, + 0.0038849730044603348, + 0.007719703484326601, + 0.001786839566193521, + 0.005036481656134129, + 0.006956067867577076, + 0.005382584873586893, + 0.010009055025875568, + 0.006291007157415152, + 0.005351753439754248, + 0.00474529666826129, + 0.00396646186709404, + 0.0036697667092084885, + 0.0038142562843859196, + 0.0045883203856647015, + 0.002279052510857582, + 0.0025432328693568707, + 0.0017795783933252096, + 0.0019528010161593556, + 0.0012240001233294606, + 0.0030194860883057117, + 0.0024767492432147264, + 0.0028172701131552458, + 0.003410808742046356, + 0.0028277768287807703, + 0.0017212232341989875, + 0.0019615981727838516, + 0.002086946275085211, + 0.018305838108062744, + 0.01656172052025795, + 0.007388983387500048, + 0.045032255351543427, + 0.01852581463754177, + 0.009649739600718021, + 0.017290709540247917, + 0.011819057166576385, + 0.00831565074622631, + 0.008117449469864368, + 0.00423998711630702, + 0.0045555029064416885, + 0.005112779326736927, + 0.002697193529456854, + 0.003490926930680871, + 0.0024641621857881546 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 8 + ], + "selected_reward": -0.0142021169885993 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera angle low and the framing a medium close-up?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the subject a woman with windswept black hair wearing a well-worn black leather jacket?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the woman performing the action of swinging her leg onto the seat of a sleek black motorcycle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a massive tidal power station with concrete walls visible in the background behind the motorcycle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit by a low golden glow from a setting sun casting long, dramatic shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.2255474478006363 + ], + "opposite_logprob_tokens": [ + -1.600547432899475 + ], + "logprob_reward": -0.2255474478006363, + "yes_no_margin_reward": 1.375, + "token_credit_reward": -0.2255474478006363, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.03940485790371895, + 0.024570876732468605, + 0.2632559835910797, + 0.24458524584770203, + 0.08035579323768616, + 0.01931586302816868, + 0.011123517528176308, + 0.018154501914978027, + 0.2815362513065338, + 0.017697086557745934 + ], + "credit_map": [ + 0.004625428933650255, + 0.005578347481787205, + 0.007021427154541016, + 0.004339549224823713, + 0.006627053953707218, + 0.006709099747240543, + 0.002254077699035406, + 0.002249875571578741, + 0.0038005842361599207, + 0.0034429607912898064, + 0.003692379454150796, + 0.0030704091768711805, + 0.003980326931923628, + 0.0027418634854257107, + 0.00212513143196702, + 0.0017172206426039338, + 0.06165127828717232, + 0.10507822781801224, + 0.027081526815891266, + 0.025732405483722687, + 0.012195628136396408, + 0.016759298741817474, + 0.010790284723043442, + 0.003967330791056156, + 0.038276709616184235, + 0.022553054615855217, + 0.01395877730101347, + 0.09686212986707687, + 0.01552323903888464, + 0.012032831087708473, + 0.03981630876660347, + 0.005562187638133764, + 0.01541220024228096, + 0.011838767677545547, + 0.013179538771510124, + 0.017027631402015686, + 0.0053010666742920876, + 0.003950423561036587, + 0.009738519787788391, + 0.003907646518200636, + 0.0026360731571912766, + 0.003378142835572362, + 0.004867012146860361, + 0.0023192891385406256, + 0.001641895272769034, + 0.001760471030138433, + 0.001527575310319662, + 0.0011854037875309587, + 0.0016393990954384208, + 0.00237552123144269, + 0.002079627476632595, + 0.0012854448286816478, + 0.0009678414207883179, + 0.0008254711283370852, + 0.001090586301870644, + 0.0008596261031925678, + 0.002960081910714507, + 0.002076013246551156, + 0.0028003347106277943, + 0.0015795715153217316, + 0.0029840567149221897, + 0.001776584074832499, + 0.002293171128258109, + 0.0016846888465806842, + 0.023572541773319244, + 0.05205141752958298, + 0.02577180787920952, + 0.10491395741701126, + 0.021593837067484856, + 0.008680046536028385, + 0.04045286029577255, + 0.004499801434576511, + 0.004141803365200758, + 0.003040876006707549, + 0.0026500485837459564, + 0.0022132336162030697, + 0.001993071287870407, + 0.0010725619504228234, + 0.0014942883281037211, + 0.0010912026045843959 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 2, + 3 + ], + "selected_reward": -0.2255474478006363 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera angle low and the framing a medium close-up?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the subject a woman with windswept black hair wearing a well-worn black leather jacket?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the woman performing the action of swinging her leg onto the seat of a sleek black motorcycle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a massive tidal power station with concrete walls visible in the background behind the motorcycle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit by a low golden glow from a setting sun casting long, dramatic shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.00248968368396163 + ], + "opposite_logprob_tokens": [ + -6.0024895668029785 + ], + "logprob_reward": -0.00248968368396163, + "yes_no_margin_reward": 6.0, + "token_credit_reward": -0.00248968368396163, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.22866691648960114, + 0.06534267961978912, + 0.3727823793888092, + 0.14329563081264496, + 0.03834090381860733, + 0.022136475890874863, + 0.03514422848820686, + 0.030563034117221832, + 0.04019119590520859, + 0.023536527529358864 + ], + "credit_map": [ + 0.01815682277083397, + 0.03377819433808327, + 0.04021873697638512, + 0.02230890467762947, + 0.048398274928331375, + 0.036459021270275116, + 0.010431072674691677, + 0.018915893509984016, + 0.005917351692914963, + 0.009601956233382225, + 0.0055551473051309586, + 0.008947247639298439, + 0.014392368495464325, + 0.009509030729532242, + 0.006904864218086004, + 0.004514714702963829, + 0.08747895807027817, + 0.15491338074207306, + 0.04544934630393982, + 0.02389850839972496, + 0.017231767997145653, + 0.023863056674599648, + 0.014155484735965729, + 0.005791869014501572, + 0.026266183704137802, + 0.008763509802520275, + 0.00776499230414629, + 0.059984371066093445, + 0.007991375401616096, + 0.005944052245467901, + 0.023998934775590897, + 0.002582200337201357, + 0.006591704674065113, + 0.0034052294213324785, + 0.0047642881982028484, + 0.009974355809390545, + 0.003075387328863144, + 0.003104924922809005, + 0.005647034849971533, + 0.0017779790796339512, + 0.0031226580031216145, + 0.0024231495335698128, + 0.003743153065443039, + 0.0056286463513970375, + 0.0027467587497085333, + 0.0015170075930655003, + 0.002009582705795765, + 0.0009455198305658996, + 0.005963741336017847, + 0.006275075487792492, + 0.005902756936848164, + 0.004679354373365641, + 0.00439616059884429, + 0.0027709216810762882, + 0.0029254364781081676, + 0.0022307829931378365, + 0.005341845098882914, + 0.003893284359946847, + 0.004358038306236267, + 0.005554532166570425, + 0.004102598410099745, + 0.002450336702167988, + 0.002903593936935067, + 0.001958805602043867, + 0.01086672767996788, + 0.011329179629683495, + 0.005456376355141401, + 0.0034351558424532413, + 0.0030446432065218687, + 0.0021560746245086193, + 0.002684235107153654, + 0.0012188033433631063, + 0.005248506087809801, + 0.005276998970657587, + 0.0031793969683349133, + 0.00293476739898324, + 0.0021865421440452337, + 0.0015372004127129912, + 0.0018674086313694715, + 0.001305704819969833 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 0, + 3 + ], + "selected_reward": -0.00248968368396163 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "low-angle, medium close-up, with a sharp digital look, remains static, capturing a figure at the wheel of a sleek black motorcycle. The subject, a woman with a determined expression and windswept black hair, wears a well-worn black leather jacket, its wrinkled texture and subtle sheen accentuated by the lighting. With ease, she swings her leg onto the seat, her free hand momentarily resting on the handlebars for balance. Behind her, a massive tidal power station looms, its massive concrete walls spanning the frame, appearing monolithic and imposing. The low, golden glow of the setting sun sweeps across the scene, casting long, dramatic shadows that stretch across the ground and sculpt every surface with intense contrast. This harsh light captures the grain of the leather and the rough, weathered texture of the concrete structure, creating a powerful portrait that emphasizes the stark contrast between the confident, self-assured human subject and the overwhelming industrial scale of her surroundings, all captured with crisp, high-fidelity detail.", + "seed": 24672, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000101/group_3/index_6062_seed_24672_dp_6_rank_12_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 3, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.039451804012060165, + "vlm_logprob": -0.039451804012060165, + "vlm_yes_no_margin": 4.15, + "vlm_token_credit": -0.039451804012060165, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera angle low and the framing a medium close-up?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the subject a woman with windswept black hair wearing a well-worn black leather jacket?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the woman performing the action of swinging her leg onto the seat of a sleek black motorcycle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a massive tidal power station with concrete walls visible in the background behind the motorcycle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit by a low golden glow from a setting sun casting long, dramatic shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.08087814599275589 + ], + "opposite_logprob_tokens": [ + -4.580878257751465 + ], + "logprob_reward": -0.08087814599275589, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.08087814599275589, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10549330711364746, + 0.3111698031425476, + 0.10419103503227234, + 0.06615026295185089, + 0.05769246444106102, + 0.04774957150220871, + 0.03774230554699898, + 0.14942625164985657, + 0.07453712075948715, + 0.04584789648652077 + ], + "credit_map": [ + 0.009954904206097126, + 0.010660328902304173, + 0.013044784776866436, + 0.01731346733868122, + 0.014462776482105255, + 0.02244110219180584, + 0.008511261083185673, + 0.009104689583182335, + 0.0376569963991642, + 0.010736997239291668, + 0.12446673214435577, + 0.023557547479867935, + 0.040367115288972855, + 0.04420452564954758, + 0.019317643716931343, + 0.010862261056900024, + 0.011797754094004631, + 0.005250262562185526, + 0.036852430552244186, + 0.009820911101996899, + 0.010366812348365784, + 0.010752592235803604, + 0.010989709757268429, + 0.008360575884580612, + 0.01093313004821539, + 0.007090672850608826, + 0.01550731435418129, + 0.005736743565648794, + 0.006378042511641979, + 0.0075933486223220825, + 0.006499860435724258, + 0.006411153823137283, + 0.00892667006701231, + 0.0060212924145162106, + 0.013420941308140755, + 0.0051131765358150005, + 0.005218958482146263, + 0.0074313064105808735, + 0.005909254774451256, + 0.005650871898978949, + 0.005431502126157284, + 0.0033144867047667503, + 0.016404110938310623, + 0.004879966843873262, + 0.0045902319252491, + 0.0058546243235468864, + 0.00399671820923686, + 0.0032779360190033913, + 0.006351884454488754, + 0.004432168323546648, + 0.005853794980794191, + 0.0031783925369381905, + 0.003817897289991379, + 0.008742023259401321, + 0.0022965623065829277, + 0.0030695816967636347, + 0.016866914927959442, + 0.00724856648594141, + 0.056918174028396606, + 0.01268446072936058, + 0.015533799305558205, + 0.027004405856132507, + 0.007550754584372044, + 0.005619180388748646, + 0.005818089935928583, + 0.006792799569666386, + 0.020628321915864944, + 0.0063541848212480545, + 0.014913618564605713, + 0.009538238868117332, + 0.006020411383360624, + 0.004471463151276112, + 0.003946154378354549, + 0.003210785100236535, + 0.011551321484148502, + 0.005250648595392704, + 0.005553534720093012, + 0.008178479969501495, + 0.004395224153995514, + 0.0037617499474436045 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 7, + 0 + ], + "selected_reward": -0.08087814599275589 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera angle low and the framing a medium close-up?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the subject a woman with windswept black hair wearing a well-worn black leather jacket?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the woman performing the action of swinging her leg onto the seat of a sleek black motorcycle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a massive tidal power station with concrete walls visible in the background behind the motorcycle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit by a low golden glow from a setting sun casting long, dramatic shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.020607998594641685 + ], + "opposite_logprob_tokens": [ + -3.8956079483032227 + ], + "logprob_reward": -0.020607998594641685, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.020607998594641685, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1416873335838318, + 0.10851240158081055, + 0.133739173412323, + 0.03696882724761963, + 0.02988935261964798, + 0.031844377517700195, + 0.0477820560336113, + 0.331661194562912, + 0.10792889446020126, + 0.029986463487148285 + ], + "credit_map": [ + 0.01370292343199253, + 0.009832754731178284, + 0.01439605187624693, + 0.0204284880310297, + 0.023980509489774704, + 0.029274089261889458, + 0.014175721444189548, + 0.01589680090546608, + 0.010789147578179836, + 0.00962776318192482, + 0.03298335522413254, + 0.011610112152993679, + 0.016216080635786057, + 0.011969546787440777, + 0.010042918846011162, + 0.005273472983390093, + 0.016296302899718285, + 0.010065963491797447, + 0.026003805920481682, + 0.017503302544355392, + 0.020565351471304893, + 0.0165106188505888, + 0.01675138622522354, + 0.010042442008852959, + 0.005636353977024555, + 0.004616166464984417, + 0.006729483138769865, + 0.004741898272186518, + 0.004575673956423998, + 0.005379272159188986, + 0.002956465119495988, + 0.0023335141595453024, + 0.004209140781313181, + 0.002258526161313057, + 0.009695799089968204, + 0.0028511113487184048, + 0.0026726694777607918, + 0.004281104076653719, + 0.0020989852491766214, + 0.0018220162019133568, + 0.0042785340920090675, + 0.003291006665676832, + 0.004516742192208767, + 0.0036818752996623516, + 0.0040314579382538795, + 0.004861520603299141, + 0.004147002007812262, + 0.0030362398829311132, + 0.007179692853242159, + 0.0037048300728201866, + 0.014865953475236893, + 0.0045051611959934235, + 0.004440602380782366, + 0.007460547611117363, + 0.0028221053071320057, + 0.0028031643014401197, + 0.04408149793744087, + 0.010915515944361687, + 0.14241720736026764, + 0.030969824641942978, + 0.03440319001674652, + 0.045917462557554245, + 0.013874913565814495, + 0.009081588126718998, + 0.012582726776599884, + 0.007145772688090801, + 0.04091094434261322, + 0.008986758068203926, + 0.014974911697208881, + 0.014691336080431938, + 0.005407024174928665, + 0.0032294231932610273, + 0.004141021054238081, + 0.0015123074408620596, + 0.011425229720771313, + 0.003245027270168066, + 0.0031245185527950525, + 0.0033665772061794996, + 0.002114668721333146, + 0.0010571129387244582 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 0, + 2 + ], + "selected_reward": -0.020607998594641685 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera angle low and the framing a medium close-up?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the subject a woman with windswept black hair wearing a well-worn black leather jacket?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the woman performing the action of swinging her leg onto the seat of a sleek black motorcycle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a massive tidal power station with concrete walls visible in the background behind the motorcycle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit by a low golden glow from a setting sun casting long, dramatic shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.023288888856768608 + ], + "opposite_logprob_tokens": [ + -3.7732889652252197 + ], + "logprob_reward": -0.023288888856768608, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.023288888856768608, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0733056366443634, + 0.43666142225265503, + 0.27860191464424133, + 0.025519108399748802, + 0.023435577750205994, + 0.027382567524909973, + 0.036274757236242294, + 0.031543247401714325, + 0.04158657044172287, + 0.025689199566841125 + ], + "credit_map": [ + 0.006659571547061205, + 0.007345518097281456, + 0.012059790082275867, + 0.008553924039006233, + 0.012568818405270576, + 0.011992381885647774, + 0.006852090824395418, + 0.007273538038134575, + 0.06521447002887726, + 0.0230559129267931, + 0.17086076736450195, + 0.033351775258779526, + 0.043731920421123505, + 0.05627250671386719, + 0.03152474761009216, + 0.012649327516555786, + 0.04330040514469147, + 0.014680898748338223, + 0.10627671331167221, + 0.024941397830843925, + 0.023472590371966362, + 0.04019578918814659, + 0.015424440614879131, + 0.010309678502380848, + 0.0025080672930926085, + 0.0018421073909848928, + 0.007400049827992916, + 0.003155073616653681, + 0.003209493588656187, + 0.0036185558419674635, + 0.002331506460905075, + 0.001454255310818553, + 0.0032513125333935022, + 0.0015522511675953865, + 0.007386170327663422, + 0.0020276501309126616, + 0.0026967497542500496, + 0.0036157809663563967, + 0.0014957915991544724, + 0.0014098711544647813, + 0.003400531131774187, + 0.0013499725610017776, + 0.01096551027148962, + 0.0024748528376221657, + 0.0026129160542041063, + 0.003576866118237376, + 0.0017081965925171971, + 0.0012937228893861175, + 0.0051981545984745026, + 0.0023753675632178783, + 0.010875937528908253, + 0.0038203843869268894, + 0.004063755739480257, + 0.005202898755669594, + 0.0029527926817536354, + 0.0017854664474725723, + 0.0032225740142166615, + 0.004171616863459349, + 0.003830663627013564, + 0.003806895110756159, + 0.004433739464730024, + 0.005060283932834864, + 0.0036218329332768917, + 0.0033956412225961685, + 0.004282211884856224, + 0.003033439628779888, + 0.014968173578381538, + 0.0033664563670754433, + 0.007950443774461746, + 0.0048896437510848045, + 0.0017132406355813146, + 0.0013829650124534965, + 0.0035006043035537004, + 0.0010611421894282103, + 0.010250148363411427, + 0.002417077776044607, + 0.0027698404155671597, + 0.002918330254033208, + 0.0017487864242866635, + 0.0010232700733467937 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 2, + 0 + ], + "selected_reward": -0.023288888856768608 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera angle low and the framing a medium close-up?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the subject a woman with windswept black hair wearing a well-worn black leather jacket?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the woman performing the action of swinging her leg onto the seat of a sleek black motorcycle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a massive tidal power station with concrete walls visible in the background behind the motorcycle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit by a low golden glow from a setting sun casting long, dramatic shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.06999525427818298 + ], + "opposite_logprob_tokens": [ + -2.694995164871216 + ], + "logprob_reward": -0.06999525427818298, + "yes_no_margin_reward": 2.625, + "token_credit_reward": -0.06999525427818298, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2301936000585556, + 0.05814982205629349, + 0.03764604404568672, + 0.15954381227493286, + 0.11420337855815887, + 0.24025647342205048, + 0.03397880494594574, + 0.06210421398282051, + 0.047053221613168716, + 0.016870645806193352 + ], + "credit_map": [ + 0.027372445911169052, + 0.02471787855029106, + 0.03298280015587807, + 0.025860818102955818, + 0.02720269188284874, + 0.0589994378387928, + 0.01452144980430603, + 0.018536057323217392, + 0.007498344872146845, + 0.003851252840831876, + 0.020988745614886284, + 0.00479931989684701, + 0.006282093469053507, + 0.00836507510393858, + 0.003964302595704794, + 0.0024006860330700874, + 0.0029787553939968348, + 0.0017335372976958752, + 0.00953351054340601, + 0.003364384640008211, + 0.004387907218188047, + 0.0053109172731637955, + 0.005087049212306738, + 0.005249981302767992, + 0.02777714468538761, + 0.011853666044771671, + 0.059195101261138916, + 0.012304605916142464, + 0.014397277496755123, + 0.017111128196120262, + 0.009133697487413883, + 0.00777118606492877, + 0.019426779821515083, + 0.0077590602450072765, + 0.04350242391228676, + 0.009572235867381096, + 0.010129989124834538, + 0.011959974654018879, + 0.00697616022080183, + 0.004876752849668264, + 0.04349829629063606, + 0.016870027408003807, + 0.08579489588737488, + 0.019751736894249916, + 0.021603107452392578, + 0.026891496032476425, + 0.014427355490624905, + 0.01141955517232418, + 0.005071335006505251, + 0.002345356624573469, + 0.010595614090561867, + 0.003141827881336212, + 0.0028606855776160955, + 0.00525866961106658, + 0.0024717573542147875, + 0.0022335555404424667, + 0.0048110163770616055, + 0.003957132808864117, + 0.011754723265767097, + 0.006226107478141785, + 0.006995230447500944, + 0.015891164541244507, + 0.0065963915549218655, + 0.005872441455721855, + 0.0055967699736356735, + 0.003126640571281314, + 0.014791413210332394, + 0.004432495683431625, + 0.0048882849514484406, + 0.006333354860544205, + 0.004843351896852255, + 0.0030409093014895916, + 0.0013348374050110579, + 0.0012579647591337562, + 0.003233749885112047, + 0.0016897051827982068, + 0.0021281111985445023, + 0.0038278161082416773, + 0.0020041184034198523, + 0.0013943416997790337 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 0, + 3 + ], + "selected_reward": -0.06999525427818298 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera angle low and the framing a medium close-up?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the subject a woman with windswept black hair wearing a well-worn black leather jacket?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the woman performing the action of swinging her leg onto the seat of a sleek black motorcycle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a massive tidal power station with concrete walls visible in the background behind the motorcycle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit by a low golden glow from a setting sun casting long, dramatic shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.00248873233795166 + ], + "opposite_logprob_tokens": [ + -6.002488613128662 + ], + "logprob_reward": -0.00248873233795166, + "yes_no_margin_reward": 6.0, + "token_credit_reward": -0.00248873233795166, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14399263262748718, + 0.22780494391918182, + 0.2761651873588562, + 0.06546907126903534, + 0.06331446766853333, + 0.02834964171051979, + 0.03256365284323692, + 0.09935682266950607, + 0.017383642494678497, + 0.04559996724128723 + ], + "credit_map": [ + 0.017299043014645576, + 0.01410826575011015, + 0.023030396550893784, + 0.01715678721666336, + 0.01472933404147625, + 0.022334491834044456, + 0.014885815791785717, + 0.020448507741093636, + 0.03684316575527191, + 0.015818478539586067, + 0.07728329300880432, + 0.018538404256105423, + 0.019128700718283653, + 0.032847218215465546, + 0.01832524873316288, + 0.009020424447953701, + 0.04083995521068573, + 0.015672078356146812, + 0.108911894261837, + 0.024624625220894814, + 0.024251217022538185, + 0.03484107181429863, + 0.014842227101325989, + 0.012182136997580528, + 0.011184193193912506, + 0.007586397230625153, + 0.014829589985311031, + 0.006282394751906395, + 0.007219724822789431, + 0.007456444203853607, + 0.0045658256858587265, + 0.006344497203826904, + 0.01071573793888092, + 0.006407971028238535, + 0.01946692168712616, + 0.005295401904731989, + 0.006120107136666775, + 0.0075017912313342094, + 0.0038652231451123953, + 0.003941320348531008, + 0.004024263005703688, + 0.0036139297299087048, + 0.0035065843258053064, + 0.003891868283972144, + 0.0034405237529426813, + 0.004332547076046467, + 0.00274938833899796, + 0.0027905390597879887, + 0.0030729591380804777, + 0.0022974929306656122, + 0.005629738327115774, + 0.00496010622009635, + 0.003761032596230507, + 0.005763980560004711, + 0.0036810191813856363, + 0.0033973262179642916, + 0.012156355194747448, + 0.009130921214818954, + 0.036416325718164444, + 0.008846897631883621, + 0.011695543304085732, + 0.013194271363317966, + 0.004451191518455744, + 0.0034653230104595423, + 0.0013334719697013497, + 0.0015365496510639787, + 0.0028425108175724745, + 0.0018337481888011098, + 0.0021757925860583782, + 0.004772560205310583, + 0.0014778727199882269, + 0.00141113658901304, + 0.0055574034340679646, + 0.0013888899702578783, + 0.019060729071497917, + 0.003731727134436369, + 0.005412896629422903, + 0.006890701595693827, + 0.002143540885299444, + 0.001414080848917365 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 1, + 0 + ], + "selected_reward": -0.00248873233795166 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "low-angle, medium close-up, with a sharp digital look, remains static, capturing a figure at the wheel of a sleek black motorcycle. The subject, a woman with a determined expression and windswept black hair, wears a well-worn black leather jacket, its wrinkled texture and subtle sheen accentuated by the lighting. With ease, she swings her leg onto the seat, her free hand momentarily resting on the handlebars for balance. Behind her, a massive tidal power station looms, its massive concrete walls spanning the frame, appearing monolithic and imposing. The low, golden glow of the setting sun sweeps across the scene, casting long, dramatic shadows that stretch across the ground and sculpt every surface with intense contrast. This harsh light captures the grain of the leather and the rough, weathered texture of the concrete structure, creating a powerful portrait that emphasizes the stark contrast between the confident, self-assured human subject and the overwhelming industrial scale of her surroundings, all captured with crisp, high-fidelity detail.", + "seed": 24673, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000101/group_3/index_6062_seed_24673_dp_7_rank_14_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 3, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.1390744002535939, + "vlm_logprob": -0.1390744002535939, + "vlm_yes_no_margin": 3.25, + "vlm_token_credit": -0.1390744002535939, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera angle low and the framing a medium close-up?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the subject a woman with windswept black hair wearing a well-worn black leather jacket?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the woman performing the action of swinging her leg onto the seat of a sleek black motorcycle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a massive tidal power station with concrete walls visible in the background behind the motorcycle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit by a low golden glow from a setting sun casting long, dramatic shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.10676366090774536 + ], + "opposite_logprob_tokens": [ + -4.10676383972168 + ], + "logprob_reward": -0.10676366090774536, + "yes_no_margin_reward": 4.0, + "token_credit_reward": -0.10676366090774536, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1656748503446579, + 0.08575230091810226, + 0.046017035841941833, + 0.06385719776153564, + 0.08657694607973099, + 0.057673390954732895, + 0.034418776631355286, + 0.025312038138508797, + 0.12572906911373138, + 0.30898845195770264 + ], + "credit_map": [ + 0.013136831112205982, + 0.016873842105269432, + 0.01161332055926323, + 0.02020331658422947, + 0.04744116961956024, + 0.016016840934753418, + 0.021139511838555336, + 0.01924999989569187, + 0.012169317342340946, + 0.010106934234499931, + 0.008214711211621761, + 0.01061080489307642, + 0.010913900099694729, + 0.008215608075261116, + 0.01637965999543667, + 0.009141367860138416, + 0.004933545365929604, + 0.006429293192923069, + 0.007361989468336105, + 0.007463333662599325, + 0.0053993402980268, + 0.00461459718644619, + 0.00613451050594449, + 0.003680428024381399, + 0.006812593434005976, + 0.006615348160266876, + 0.008109347894787788, + 0.007941510528326035, + 0.013814608566462994, + 0.006207623984664679, + 0.006918734405189753, + 0.007437425199896097, + 0.009480852633714676, + 0.008439750410616398, + 0.009721584618091583, + 0.017343370243906975, + 0.016581062227487564, + 0.009758133441209793, + 0.009776521474123001, + 0.0054756649769842625, + 0.010752517729997635, + 0.005604949779808521, + 0.006752731744199991, + 0.0036630164831876755, + 0.00630181934684515, + 0.016905365511775017, + 0.0031375412363559008, + 0.00455544563010335, + 0.003575769253075123, + 0.003847069339826703, + 0.004503830336034298, + 0.003555628005415201, + 0.005639924202114344, + 0.00643072510138154, + 0.00255956850014627, + 0.004306260962039232, + 0.0026761803310364485, + 0.002867815550416708, + 0.002240202622488141, + 0.0027819836977869272, + 0.0035291777458041906, + 0.0021501933224499226, + 0.006235178560018539, + 0.0028313053771853447, + 0.03146043419837952, + 0.007935396395623684, + 0.013615521602332592, + 0.008306178264319897, + 0.009990919381380081, + 0.042687419801950455, + 0.004866837989538908, + 0.006866362411528826, + 0.06925661116838455, + 0.012506742961704731, + 0.043706901371479034, + 0.016642387956380844, + 0.020388249307870865, + 0.11202546209096909, + 0.021468888968229294, + 0.012993207201361656 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 0, + 8 + ], + "selected_reward": -0.10676366090774536 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera angle low and the framing a medium close-up?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the subject a woman with windswept black hair wearing a well-worn black leather jacket?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the woman performing the action of swinging her leg onto the seat of a sleek black motorcycle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a massive tidal power station with concrete walls visible in the background behind the motorcycle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit by a low golden glow from a setting sun casting long, dramatic shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.016098281368613243 + ], + "opposite_logprob_tokens": [ + -4.141098499298096 + ], + "logprob_reward": -0.016098281368613243, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.016098281368613243, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.05412791669368744, + 0.14114174246788025, + 0.08670952171087265, + 0.10263171792030334, + 0.09953870624303818, + 0.16589447855949402, + 0.09949110448360443, + 0.14913800358772278, + 0.06317953020334244, + 0.03814723342657089 + ], + "credit_map": [ + 0.005790688097476959, + 0.008659186773002148, + 0.00653400644659996, + 0.005953842308372259, + 0.00942256674170494, + 0.004509374964982271, + 0.006491582375019789, + 0.006766667123883963, + 0.02354402281343937, + 0.010477101430296898, + 0.010108234360814095, + 0.013271315023303032, + 0.015448532067239285, + 0.011301995255053043, + 0.021513458341360092, + 0.035477086901664734, + 0.011758370324969292, + 0.00901300460100174, + 0.008110721595585346, + 0.01947111450135708, + 0.01202611904591322, + 0.006160495802760124, + 0.013911832123994827, + 0.006257861852645874, + 0.011748495511710644, + 0.008674431592226028, + 0.013567104935646057, + 0.011997882276773453, + 0.017562545835971832, + 0.011617417447268963, + 0.015034650452435017, + 0.012429195456206799, + 0.007988842204213142, + 0.008361936546862125, + 0.009259325452148914, + 0.018653366714715958, + 0.02205551601946354, + 0.014973873272538185, + 0.012065285816788673, + 0.006180563010275364, + 0.03539338335394859, + 0.012763911858201027, + 0.024655718356370926, + 0.008676601573824883, + 0.016047881916165352, + 0.05480532720685005, + 0.005155573133379221, + 0.008396080695092678, + 0.02017817087471485, + 0.010492081753909588, + 0.013142948970198631, + 0.008466036058962345, + 0.01239782478660345, + 0.022521769627928734, + 0.006342416163533926, + 0.005949853919446468, + 0.013712137937545776, + 0.008525154553353786, + 0.018402278423309326, + 0.01164956297725439, + 0.011257682926952839, + 0.04325980693101883, + 0.035680074244737625, + 0.006651317235082388, + 0.014568260870873928, + 0.0027651546988636255, + 0.005859827157109976, + 0.005412072408944368, + 0.004780652932822704, + 0.024708645418286324, + 0.0017616798868402839, + 0.0033232381101697683, + 0.004745504818856716, + 0.0024431501515209675, + 0.00532574625685811, + 0.00315163959749043, + 0.0036972700618207455, + 0.01206901203840971, + 0.004588620737195015, + 0.0021262909285724163 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 7, + 1 + ], + "selected_reward": -0.016098281368613243 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera angle low and the framing a medium close-up?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the subject a woman with windswept black hair wearing a well-worn black leather jacket?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the woman performing the action of swinging her leg onto the seat of a sleek black motorcycle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a massive tidal power station with concrete walls visible in the background behind the motorcycle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit by a low golden glow from a setting sun casting long, dramatic shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03811095654964447 + ], + "opposite_logprob_tokens": [ + -3.2881109714508057 + ], + "logprob_reward": -0.03811095654964447, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.03811095654964447, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.05423521250486374, + 0.0938258245587349, + 0.11979027092456818, + 0.034524790942668915, + 0.12046036869287491, + 0.11372169852256775, + 0.05644567310810089, + 0.07579538971185684, + 0.22561272978782654, + 0.10558808594942093 + ], + "credit_map": [ + 0.004492965526878834, + 0.008054613135755062, + 0.004961222875863314, + 0.006013147067278624, + 0.013016367331147194, + 0.004482508171349764, + 0.006288930773735046, + 0.006925454828888178, + 0.013815095648169518, + 0.008336932398378849, + 0.007435716688632965, + 0.0101775536313653, + 0.009117973037064075, + 0.006606321781873703, + 0.01463260781019926, + 0.02370361052453518, + 0.013392429798841476, + 0.016916118562221527, + 0.013146751560270786, + 0.018680017441511154, + 0.02080943062901497, + 0.012256050482392311, + 0.012592092156410217, + 0.011997372843325138, + 0.004722765646874905, + 0.004038260318338871, + 0.004111334681510925, + 0.0037263724952936172, + 0.006157048512250185, + 0.003746449714526534, + 0.00447216909378767, + 0.0035503911785781384, + 0.010369664989411831, + 0.00875926110893488, + 0.011564821004867554, + 0.017883023247122765, + 0.027384383603930473, + 0.02565256506204605, + 0.012068141251802444, + 0.006778497248888016, + 0.018737442791461945, + 0.013378898613154888, + 0.023973530158400536, + 0.010093240067362785, + 0.013011670671403408, + 0.023422621190547943, + 0.005526870489120483, + 0.005577412433922291, + 0.00825428031384945, + 0.007739790715277195, + 0.009542899206280708, + 0.004308758769184351, + 0.006143522448837757, + 0.013739579357206821, + 0.0026333669666200876, + 0.004083471372723579, + 0.005967522505670786, + 0.004645040258765221, + 0.008437524549663067, + 0.00650433823466301, + 0.0067825536243617535, + 0.01928156241774559, + 0.020746618509292603, + 0.0034302284475415945, + 0.061321601271629333, + 0.010496407747268677, + 0.019142702221870422, + 0.012304564006626606, + 0.015446938574314117, + 0.08560153841972351, + 0.010118387639522552, + 0.011180585250258446, + 0.017301954329013824, + 0.004610700998455286, + 0.014874969609081745, + 0.006791146472096443, + 0.0077502247877418995, + 0.036416515707969666, + 0.01236026268452406, + 0.005482310429215431 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 4, + 2 + ], + "selected_reward": -0.03811095654964447 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera angle low and the framing a medium close-up?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the subject a woman with windswept black hair wearing a well-worn black leather jacket?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the woman performing the action of swinging her leg onto the seat of a sleek black motorcycle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a massive tidal power station with concrete walls visible in the background behind the motorcycle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit by a low golden glow from a setting sun casting long, dramatic shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.5233235359191895 + ], + "opposite_logprob_tokens": [ + -0.8983235359191895 + ], + "logprob_reward": -0.5233235359191895, + "yes_no_margin_reward": 0.375, + "token_credit_reward": -0.5233235359191895, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14875958859920502, + 0.17016808688640594, + 0.15245622396469116, + 0.09658148139715195, + 0.0911078080534935, + 0.0406377837061882, + 0.04578428715467453, + 0.04804084450006485, + 0.16541291773319244, + 0.041051000356674194 + ], + "credit_map": [ + 0.021134361624717712, + 0.017402861267328262, + 0.019501132890582085, + 0.018561674281954765, + 0.015996592119336128, + 0.023787161335349083, + 0.01733659952878952, + 0.015039204619824886, + 0.02679477259516716, + 0.01541117299348116, + 0.01321116928011179, + 0.015525017865002155, + 0.021229363977909088, + 0.017498819157481194, + 0.022685347124934196, + 0.03781241551041603, + 0.018568815663456917, + 0.0164877288043499, + 0.014933045022189617, + 0.02871112897992134, + 0.0246421005576849, + 0.010087084956467152, + 0.022244470193982124, + 0.016781851649284363, + 0.01242111250758171, + 0.009308741427958012, + 0.015775229781866074, + 0.01091590616852045, + 0.014120606705546379, + 0.014209695160388947, + 0.012684008106589317, + 0.007146175019443035, + 0.010135919786989689, + 0.007448730990290642, + 0.012109114788472652, + 0.012088647112250328, + 0.011972245760262012, + 0.024665284901857376, + 0.007824179716408253, + 0.0048636868596076965, + 0.0050836470909416676, + 0.006221075542271137, + 0.011128196492791176, + 0.004026934504508972, + 0.005152286030352116, + 0.003924182616174221, + 0.002283106790855527, + 0.0028183553367853165, + 0.005685541778802872, + 0.004788021557033062, + 0.008372068405151367, + 0.003373610321432352, + 0.004517547320574522, + 0.012615076266229153, + 0.0022088796831667423, + 0.0042235408909618855, + 0.0071477885358035564, + 0.0031318231485784054, + 0.003551932517439127, + 0.0047443583607673645, + 0.005000116769224405, + 0.00507126422598958, + 0.01658882200717926, + 0.0028047398664057255, + 0.04268921539187431, + 0.007231108844280243, + 0.01442190445959568, + 0.009946669451892376, + 0.012260439805686474, + 0.06417076289653778, + 0.0062547544948756695, + 0.008438060991466045, + 0.008273158222436905, + 0.001958969049155712, + 0.005899081937968731, + 0.002501587150618434, + 0.003079875372350216, + 0.014747434295713902, + 0.0026050638407468796, + 0.0019858272280544043 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 8, + 2 + ], + "selected_reward": -0.5233235359191895 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera angle low and the framing a medium close-up?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the subject a woman with windswept black hair wearing a well-worn black leather jacket?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the woman performing the action of swinging her leg onto the seat of a sleek black motorcycle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a massive tidal power station with concrete walls visible in the background behind the motorcycle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit by a low golden glow from a setting sun casting long, dramatic shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01107556652277708 + ], + "opposite_logprob_tokens": [ + -4.511075496673584 + ], + "logprob_reward": -0.01107556652277708, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.01107556652277708, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13038979470729828, + 0.18368586897850037, + 0.07479912787675858, + 0.08873379975557327, + 0.06974207609891891, + 0.02435988374054432, + 0.15460123121738434, + 0.026141289621591568, + 0.2092970758676529, + 0.03824976459145546 + ], + "credit_map": [ + 0.013165689073503017, + 0.012892135418951511, + 0.016520917415618896, + 0.014247515238821507, + 0.026128480210900307, + 0.01478841807693243, + 0.01698937825858593, + 0.015657275915145874, + 0.02524697035551071, + 0.024916179478168488, + 0.022029250860214233, + 0.018749717622995377, + 0.02563438005745411, + 0.025278259068727493, + 0.022864239290356636, + 0.018966902047395706, + 0.007286430336534977, + 0.011184971779584885, + 0.009963116608560085, + 0.008804850280284882, + 0.014087074436247349, + 0.008746692910790443, + 0.007983200252056122, + 0.006742797326296568, + 0.008973832242190838, + 0.006960021797567606, + 0.01113853044807911, + 0.010988322086632252, + 0.016012685373425484, + 0.00974944792687893, + 0.013456898741424084, + 0.011454072780907154, + 0.010353208519518375, + 0.004398757591843605, + 0.0077143884263932705, + 0.006752928253263235, + 0.009165185503661633, + 0.02233600802719593, + 0.005281280726194382, + 0.0037403199821710587, + 0.0026769088581204414, + 0.003902125870808959, + 0.004027063492685556, + 0.0022222779225558043, + 0.0027801692485809326, + 0.004954321309924126, + 0.0014444324187934399, + 0.0023525848519057035, + 0.043143466114997864, + 0.009039483964443207, + 0.01787690445780754, + 0.010002504102885723, + 0.011282697319984436, + 0.050833869725465775, + 0.004899342078715563, + 0.007522976491600275, + 0.002199167851358652, + 0.004441570490598679, + 0.003436097875237465, + 0.0031216798815876245, + 0.004990020766854286, + 0.0018329507438465953, + 0.0033782694954425097, + 0.0027415319345891476, + 0.05537724122405052, + 0.009018655866384506, + 0.01786043494939804, + 0.01260597724467516, + 0.014378471300005913, + 0.081842802464962, + 0.007809517905116081, + 0.010403999127447605, + 0.007000262383371592, + 0.002019202569499612, + 0.005381320603191853, + 0.002417371142655611, + 0.002670735353603959, + 0.013163169845938683, + 0.003956600092351437, + 0.0016411086544394493 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 1, + 6 + ], + "selected_reward": -0.01107556652277708 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "low-angle, medium close-up, with a sharp digital look, remains static, capturing a figure at the wheel of a sleek black motorcycle. The subject, a woman with a determined expression and windswept black hair, wears a well-worn black leather jacket, its wrinkled texture and subtle sheen accentuated by the lighting. With ease, she swings her leg onto the seat, her free hand momentarily resting on the handlebars for balance. Behind her, a massive tidal power station looms, its massive concrete walls spanning the frame, appearing monolithic and imposing. The low, golden glow of the setting sun sweeps across the scene, casting long, dramatic shadows that stretch across the ground and sculpt every surface with intense contrast. This harsh light captures the grain of the leather and the rough, weathered texture of the concrete structure, creating a powerful portrait that emphasizes the stark contrast between the confident, self-assured human subject and the overwhelming industrial scale of her surroundings, all captured with crisp, high-fidelity detail.", + "seed": 24674, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000101/group_3/index_6062_seed_24674_dp_7_rank_14_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 3, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.05240652384236455, + "vlm_logprob": -0.05240652384236455, + "vlm_yes_no_margin": 3.75, + "vlm_token_credit": -0.05240652384236455, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera angle low and the framing a medium close-up?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the subject a woman with windswept black hair wearing a well-worn black leather jacket?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the woman performing the action of swinging her leg onto the seat of a sleek black motorcycle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a massive tidal power station with concrete walls visible in the background behind the motorcycle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit by a low golden glow from a setting sun casting long, dramatic shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.1117384284734726 + ], + "opposite_logprob_tokens": [ + -3.861738443374634 + ], + "logprob_reward": -0.1117384284734726, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.1117384284734726, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.20190294086933136, + 0.15199069678783417, + 0.07286427170038223, + 0.09416396915912628, + 0.06234673038125038, + 0.13382108509540558, + 0.06040554866194725, + 0.07606405764818192, + 0.08446645736694336, + 0.061974186450242996 + ], + "credit_map": [ + 0.032873231917619705, + 0.02999798022210598, + 0.022794872522354126, + 0.02351268008351326, + 0.02317395992577076, + 0.01900063455104828, + 0.021807214245200157, + 0.028742356225848198, + 0.010134785436093807, + 0.01009004283696413, + 0.022750496864318848, + 0.05837062746286392, + 0.011164252646267414, + 0.005883635487407446, + 0.026520203799009323, + 0.007076658308506012, + 0.009749595075845718, + 0.005701201967895031, + 0.018528791144490242, + 0.0048727854155004025, + 0.013167762197554111, + 0.0066488394513726234, + 0.008265971206128597, + 0.0059293233789503574, + 0.008744781836867332, + 0.007069014478474855, + 0.01872936822474003, + 0.009873404167592525, + 0.014415732584893703, + 0.01203396450728178, + 0.01170707494020462, + 0.011590637266635895, + 0.005543381907045841, + 0.013261117972433567, + 0.019804203882813454, + 0.0034245376009494066, + 0.008230874314904213, + 0.0032800764311105013, + 0.00510578416287899, + 0.003696754341945052, + 0.022307507693767548, + 0.02942964993417263, + 0.033065106719732285, + 0.00932056363672018, + 0.011093576438724995, + 0.012201906181871891, + 0.009651985950767994, + 0.00675079133361578, + 0.0081930011510849, + 0.009496252052485943, + 0.009219937026500702, + 0.005925816483795643, + 0.00710641173645854, + 0.006933475844562054, + 0.005982028786092997, + 0.0075486283749341965, + 0.009232386015355587, + 0.013346856459975243, + 0.01141685526818037, + 0.007907946594059467, + 0.007678620051592588, + 0.00882494542747736, + 0.007460332941263914, + 0.010196108371019363, + 0.01534826960414648, + 0.013166289776563644, + 0.01198946125805378, + 0.009099902585148811, + 0.007754957769066095, + 0.00880619790405035, + 0.009958958253264427, + 0.008342419750988483, + 0.010203632526099682, + 0.012391055934131145, + 0.01070421189069748, + 0.004569804295897484, + 0.007389606907963753, + 0.006668668705970049, + 0.004552179016172886, + 0.00549502857029438 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 5 + ], + "selected_reward": -0.1117384284734726 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera angle low and the framing a medium close-up?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the subject a woman with windswept black hair wearing a well-worn black leather jacket?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the woman performing the action of swinging her leg onto the seat of a sleek black motorcycle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a massive tidal power station with concrete walls visible in the background behind the motorcycle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit by a low golden glow from a setting sun casting long, dramatic shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.012575849890708923 + ], + "opposite_logprob_tokens": [ + -4.387575626373291 + ], + "logprob_reward": -0.012575849890708923, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.012575849890708923, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.20084422826766968, + 0.11708369851112366, + 0.07459668070077896, + 0.192295104265213, + 0.13419312238693237, + 0.055178411304950714, + 0.0804358571767807, + 0.06738361716270447, + 0.04710593447089195, + 0.030883310362696648 + ], + "credit_map": [ + 0.02669401839375496, + 0.026522204279899597, + 0.03343364968895912, + 0.026254307478666306, + 0.02521519362926483, + 0.021750906482338905, + 0.02180585078895092, + 0.01916809380054474, + 0.00954869482666254, + 0.007374138571321964, + 0.016784725710749626, + 0.017157498747110367, + 0.026591937988996506, + 0.00986829400062561, + 0.01892886310815811, + 0.010829553939402103, + 0.006574518978595734, + 0.004872615449130535, + 0.013834570534527302, + 0.007948756217956543, + 0.015566988848149776, + 0.008017940446734428, + 0.009067573584616184, + 0.008713717572391033, + 0.01769503578543663, + 0.014374355785548687, + 0.032777223736047745, + 0.017994824796915054, + 0.030212750658392906, + 0.032428253442049026, + 0.02300192043185234, + 0.023810751736164093, + 0.011371035128831863, + 0.017184240743517876, + 0.01902887597680092, + 0.009706716053187847, + 0.03206898644566536, + 0.011008018627762794, + 0.02078581228852272, + 0.013039447367191315, + 0.008179525844752789, + 0.010458773002028465, + 0.013990159146487713, + 0.0037592556327581406, + 0.005932184401899576, + 0.004779880866408348, + 0.004464455880224705, + 0.0036141781602054834, + 0.010962193831801414, + 0.015604418702423573, + 0.016331559047102928, + 0.007653580512851477, + 0.008511792868375778, + 0.00666391896083951, + 0.00802567508071661, + 0.006682727951556444, + 0.012841826304793358, + 0.014556534588336945, + 0.01517065055668354, + 0.004384356085211039, + 0.005175629165023565, + 0.004271851386874914, + 0.005528721027076244, + 0.0054540494456887245, + 0.008687684312462807, + 0.005834635347127914, + 0.009261011146008968, + 0.005300918128341436, + 0.004384557716548443, + 0.004402880556881428, + 0.0051461681723594666, + 0.004088081419467926, + 0.004975077696144581, + 0.004920802544802427, + 0.004913565702736378, + 0.0032493730541318655, + 0.00421177688986063, + 0.002735830843448639, + 0.003019020427018404, + 0.002857862738892436 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 4 + ], + "selected_reward": -0.012575849890708923 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera angle low and the framing a medium close-up?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the subject a woman with windswept black hair wearing a well-worn black leather jacket?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the woman performing the action of swinging her leg onto the seat of a sleek black motorcycle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a massive tidal power station with concrete walls visible in the background behind the motorcycle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit by a low golden glow from a setting sun casting long, dramatic shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.029804222285747528 + ], + "opposite_logprob_tokens": [ + -3.529804229736328 + ], + "logprob_reward": -0.029804222285747528, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.029804222285747528, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.24246111512184143, + 0.17891591787338257, + 0.05545659363269806, + 0.10850807279348373, + 0.03968954458832741, + 0.06269268691539764, + 0.0982687696814537, + 0.13657385110855103, + 0.04439888149499893, + 0.033034615218639374 + ], + "credit_map": [ + 0.021054347977042198, + 0.020639568567276, + 0.03140358254313469, + 0.034721434116363525, + 0.04664316028356552, + 0.023618169128894806, + 0.03556303307414055, + 0.02881779707968235, + 0.012608136050403118, + 0.011933386325836182, + 0.01582912728190422, + 0.06536933034658432, + 0.01621086895465851, + 0.007693461608141661, + 0.03761889785528183, + 0.01165271271020174, + 0.006926245056092739, + 0.003655147971585393, + 0.013693083077669144, + 0.003865944454446435, + 0.00795652810484171, + 0.006083122454583645, + 0.00889777485281229, + 0.004378752317279577, + 0.010607589036226273, + 0.007611813489347696, + 0.01870446652173996, + 0.011158362030982971, + 0.017887290567159653, + 0.016717921942472458, + 0.01063705887645483, + 0.015183580107986927, + 0.00489228218793869, + 0.005955945234745741, + 0.004731838591396809, + 0.0035535262431949377, + 0.006255289074033499, + 0.0037490571849048138, + 0.005390080623328686, + 0.005161527544260025, + 0.00745105417445302, + 0.01320076547563076, + 0.012516606599092484, + 0.0034829634241759777, + 0.008439365774393082, + 0.006990785710513592, + 0.0057455082423985004, + 0.004865639377385378, + 0.010796713642776012, + 0.02360387146472931, + 0.019137967377901077, + 0.007967524230480194, + 0.008876249194145203, + 0.008590840734541416, + 0.010379163548350334, + 0.00891643762588501, + 0.028151115402579308, + 0.024042194709181786, + 0.029644079506397247, + 0.009552407078444958, + 0.01323014497756958, + 0.011055081151425838, + 0.011574102565646172, + 0.009324709884822369, + 0.004676972981542349, + 0.007843365892767906, + 0.006621074862778187, + 0.004662628751248121, + 0.005976720247417688, + 0.005176555830985308, + 0.003961739130318165, + 0.005479822400957346, + 0.005448367912322283, + 0.0057876803912222385, + 0.0041012149304151535, + 0.0029926870483905077, + 0.004699797369539738, + 0.003475744975730777, + 0.003036899957805872, + 0.0034922221675515175 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 7 + ], + "selected_reward": -0.029804222285747528 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera angle low and the framing a medium close-up?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the subject a woman with windswept black hair wearing a well-worn black leather jacket?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the woman performing the action of swinging her leg onto the seat of a sleek black motorcycle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a massive tidal power station with concrete walls visible in the background behind the motorcycle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit by a low golden glow from a setting sun casting long, dramatic shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.10028670728206635 + ], + "opposite_logprob_tokens": [ + -2.3502867221832275 + ], + "logprob_reward": -0.10028670728206635, + "yes_no_margin_reward": 2.25, + "token_credit_reward": -0.10028670728206635, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15072621405124664, + 0.30438879132270813, + 0.02230260707437992, + 0.0370946042239666, + 0.06146489456295967, + 0.1255941390991211, + 0.046660132706165314, + 0.1717292070388794, + 0.04044334590435028, + 0.03959610313177109 + ], + "credit_map": [ + 0.024265892803668976, + 0.026472415775060654, + 0.02729671262204647, + 0.01668701320886612, + 0.011829792521893978, + 0.012471023015677929, + 0.015132741071283817, + 0.016570627689361572, + 0.019741343334317207, + 0.018245097249746323, + 0.05016964673995972, + 0.12505780160427094, + 0.018979402258992195, + 0.008344455622136593, + 0.0526859425008297, + 0.01116509735584259, + 0.0031849113292992115, + 0.002903049811720848, + 0.004679778125137091, + 0.0016513774171471596, + 0.004144597332924604, + 0.0014760300982743502, + 0.0026357206515967846, + 0.0016271431231871247, + 0.004865910392254591, + 0.004936990328133106, + 0.0043212492018938065, + 0.0038137738592922688, + 0.005922276061028242, + 0.0047550564631819725, + 0.0035270366352051497, + 0.004952308256179094, + 0.006057450547814369, + 0.016315637156367302, + 0.015752673149108887, + 0.0049475268460810184, + 0.005474976263940334, + 0.004685731139034033, + 0.00505117466673255, + 0.0031797289848327637, + 0.016987459734082222, + 0.038034453988075256, + 0.03147031366825104, + 0.006655918434262276, + 0.009044043719768524, + 0.010046936571598053, + 0.007760399486869574, + 0.005594608839601278, + 0.007920963689684868, + 0.008546657860279083, + 0.010034801438450813, + 0.005029580090194941, + 0.004005678929388523, + 0.004036878235638142, + 0.003999263048171997, + 0.0030863070860505104, + 0.03627386316657066, + 0.03007909469306469, + 0.04509549215435982, + 0.011081468313932419, + 0.011598165147006512, + 0.0161671731621027, + 0.012801091186702251, + 0.008632871322333813, + 0.008608750067651272, + 0.006688639521598816, + 0.011968167498707771, + 0.003201826475560665, + 0.0031580401118844748, + 0.0027032066136598587, + 0.0024889190681278706, + 0.0016257953830063343, + 0.006296380423009396, + 0.010646690614521503, + 0.009133582003414631, + 0.0028472619596868753, + 0.003425080794841051, + 0.0025152661837637424, + 0.00232813716866076, + 0.002403705846518278 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 7, + 0 + ], + "selected_reward": -0.10028670728206635 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera angle low and the framing a medium close-up?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the subject a woman with windswept black hair wearing a well-worn black leather jacket?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the woman performing the action of swinging her leg onto the seat of a sleek black motorcycle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a massive tidal power station with concrete walls visible in the background behind the motorcycle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit by a low golden glow from a setting sun casting long, dramatic shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.007627411279827356 + ], + "opposite_logprob_tokens": [ + -4.882627487182617 + ], + "logprob_reward": -0.007627411279827356, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.007627411279827356, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1341969072818756, + 0.06818009912967682, + 0.06630266457796097, + 0.02964738942682743, + 0.0537496954202652, + 0.3050178289413452, + 0.030404480174183846, + 0.18839474022388458, + 0.09904387593269348, + 0.025062354281544685 + ], + "credit_map": [ + 0.02339525707066059, + 0.019098302349448204, + 0.01922684907913208, + 0.015486146323382854, + 0.015183462761342525, + 0.014560461975634098, + 0.014600577764213085, + 0.012645839713513851, + 0.005234017036855221, + 0.003561882534995675, + 0.009351426735520363, + 0.02628716081380844, + 0.006478444207459688, + 0.003261386649683118, + 0.01072127465158701, + 0.0032845118548721075, + 0.01027730479836464, + 0.00589544465765357, + 0.019249478355050087, + 0.003807768924161792, + 0.007575975265353918, + 0.004272708669304848, + 0.011643906123936176, + 0.0035800724290311337, + 0.003593480447307229, + 0.003661923110485077, + 0.006599947344511747, + 0.0025845018681138754, + 0.0037362328730523586, + 0.0024618185125291348, + 0.0032135865185409784, + 0.003795898286625743, + 0.007698782719671726, + 0.00877593643963337, + 0.01299980003386736, + 0.00494043342769146, + 0.00627074483782053, + 0.0034658305812627077, + 0.006013616919517517, + 0.003584550227969885, + 0.04478169605135918, + 0.08193816244602203, + 0.07117830216884613, + 0.016231657937169075, + 0.025765687227249146, + 0.026648186147212982, + 0.022569134831428528, + 0.015904998406767845, + 0.006321118213236332, + 0.006637715268880129, + 0.004989967215806246, + 0.0027192295528948307, + 0.002706709085032344, + 0.0026979297399520874, + 0.0022837077267467976, + 0.0020481052342802286, + 0.03702777624130249, + 0.03556082770228386, + 0.03315812349319458, + 0.012822328135371208, + 0.01666126772761345, + 0.019900476559996605, + 0.017564574256539345, + 0.015699367970228195, + 0.017537705600261688, + 0.021112559363245964, + 0.0197805967181921, + 0.006408995483070612, + 0.008103793486952782, + 0.009700401686131954, + 0.008780356496572495, + 0.007619468495249748, + 0.0035880834329873323, + 0.006695038639008999, + 0.0044122738763689995, + 0.0016304339515045285, + 0.002838634420186281, + 0.0019880933687090874, + 0.0020099878311157227, + 0.0018998084124177694 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 7, + 0 + ], + "selected_reward": -0.007627411279827356 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000101/group_4/index_tensor([7096, 7096])_group_4_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000101/group_4/index_tensor([7096, 7096])_group_4_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..2cc0e314e65ceb459ebc4452a381ae5bac760403 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000101/group_4/index_tensor([7096, 7096])_group_4_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "static, low-angle, medium-long shot frames a young woman amidst a vast, silent, snow-covered landscape, rendered with the soft edges and luminous transparency of a vivid watercolor. Her long, dark hair, gently stirred by the crisp evening air, is adorned with a distinctive, vibrant red ribbon, its saturated color creating a striking accent against the cool palette. She is wrapped in a thick, cream-colored wool coat, its texture appearing soft and painterly. Her head is tilted fully back, gazing upward, her pose completely focused. Her facial expression is one of pure wonder, her wide, sparkling eyes reflecting the celestial scene, her lips slightly parted in silent awe. Above her, the northern lights shimmer and dance across the deep blue sky—not clear rays of light, but magnificent, bleeding emerald green, ethereal violet, and soft mauve, their colors blending seamlessly as if applied with a brush to a damp canvas. This otherworldly light casts a dynamic, colorful glow on her face and the powdery snow at her feet, creating an atmosphere of tranquil, magical contemplation.", + "seed": 3759, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000101/group_4/index_7096_seed_3759_dp_8_rank_16_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 4, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.07444614400155843, + "vlm_logprob": -0.07444614400155843, + "vlm_yes_no_margin": 4.624999988079071, + "vlm_token_credit": -0.07444614400155843, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a vast, snow-covered landscape?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the visual style rendered with soft edges and luminous transparency resembling a vivid watercolor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the woman wearing a thick, cream-colored wool coat with a soft, painterly texture?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the northern lights visible in the deep blue sky with bleeding emerald green, ethereal violet, and soft mauve colors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot framed as a static, low-angle, medium-long shot?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.00887839775532484 + ], + "opposite_logprob_tokens": [ + -5.633878231048584 + ], + "logprob_reward": -0.00887839775532484, + "yes_no_margin_reward": 5.625, + "token_credit_reward": -0.00887839775532484, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07215753942728043, + 0.14975568652153015, + 0.05996344983577728, + 0.04023490101099014, + 0.060459621250629425, + 0.06541971862316132, + 0.07193057984113693, + 0.30070701241493225, + 0.07345574349164963, + 0.10591569542884827 + ], + "credit_map": [ + 0.00889563001692295, + 0.009546554647386074, + 0.0077544692903757095, + 0.006517686881124973, + 0.00961791817098856, + 0.007966483011841774, + 0.010346702300012112, + 0.0115120904520154, + 0.018750524148344994, + 0.01760316826403141, + 0.01308527309447527, + 0.015882031992077827, + 0.02473924309015274, + 0.013830089941620827, + 0.01701473817229271, + 0.02885061874985695, + 0.007319565862417221, + 0.006637066602706909, + 0.004631739109754562, + 0.0058281198143959045, + 0.007900288328528404, + 0.005822211969643831, + 0.006825675256550312, + 0.01499878242611885, + 0.004685258027166128, + 0.007781278807669878, + 0.004501803312450647, + 0.005585829261690378, + 0.0049304356798529625, + 0.0030727782286703587, + 0.004475969821214676, + 0.0052015469409525394, + 0.007449232507497072, + 0.008439083583652973, + 0.0053610047325491905, + 0.007692950777709484, + 0.0077234432101249695, + 0.005749489646404982, + 0.00839509628713131, + 0.009649316780269146, + 0.008828656747937202, + 0.007992013357579708, + 0.005744246765971184, + 0.005470070987939835, + 0.013787970878183842, + 0.007245829328894615, + 0.0060373651795089245, + 0.010313564911484718, + 0.008478399366140366, + 0.00978014338761568, + 0.0063198283314704895, + 0.005950734950602055, + 0.010082521475851536, + 0.008079460822045803, + 0.01100085861980915, + 0.012238634750247002, + 0.07713106274604797, + 0.020615870133042336, + 0.016223115846514702, + 0.00873158872127533, + 0.04602601379156113, + 0.08806917816400528, + 0.017394796013832092, + 0.026515403762459755, + 0.023208513855934143, + 0.0037805039901286364, + 0.0037930288817733526, + 0.002523483010008931, + 0.005714710336178541, + 0.027037225663661957, + 0.003629847429692745, + 0.0037684261333197355, + 0.01400022767484188, + 0.011013983748853207, + 0.007684680633246899, + 0.005592273082584143, + 0.024049464613199234, + 0.011381209827959538, + 0.008964871056377888, + 0.02322898991405964 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 1, + 9 + ], + "selected_reward": -0.00887839775532484 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a vast, snow-covered landscape?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the visual style rendered with soft edges and luminous transparency resembling a vivid watercolor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the woman wearing a thick, cream-colored wool coat with a soft, painterly texture?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the northern lights visible in the deep blue sky with bleeding emerald green, ethereal violet, and soft mauve colors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot framed as a static, low-angle, medium-long shot?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.002506093354895711 + ], + "opposite_logprob_tokens": [ + -6.002506256103516 + ], + "logprob_reward": -0.002506093354895711, + "yes_no_margin_reward": 6.0, + "token_credit_reward": -0.002506093354895711, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11801379173994064, + 0.12679792940616608, + 0.08565546572208405, + 0.04972788691520691, + 0.08235389739274979, + 0.16422028839588165, + 0.11894764006137848, + 0.09474047273397446, + 0.051907166838645935, + 0.10763543099164963 + ], + "credit_map": [ + 0.015855582430958748, + 0.012074519880115986, + 0.01327375415712595, + 0.011857119388878345, + 0.015451100654900074, + 0.013045798055827618, + 0.016592340543866158, + 0.019863585010170937, + 0.015599966049194336, + 0.02481439709663391, + 0.009741959162056446, + 0.012740390375256538, + 0.020653722807765007, + 0.014400893822312355, + 0.013812572695314884, + 0.015034029260277748, + 0.010281453840434551, + 0.01287474762648344, + 0.010245446115732193, + 0.012367731891572475, + 0.010119439102709293, + 0.006970853544771671, + 0.009899771772325039, + 0.01289602555334568, + 0.006836144719272852, + 0.007574404124170542, + 0.005098441150039434, + 0.006054780911654234, + 0.00471857376396656, + 0.005057983566075563, + 0.006333366502076387, + 0.008054192177951336, + 0.010960488580167294, + 0.009785673581063747, + 0.005615048576146364, + 0.010208550840616226, + 0.01764289103448391, + 0.008194761350750923, + 0.008054682984948158, + 0.011891797184944153, + 0.020574800670146942, + 0.020381512120366096, + 0.011058421805500984, + 0.022229066118597984, + 0.028689276427030563, + 0.018224673345685005, + 0.019417723640799522, + 0.02364482171833515, + 0.015863489359617233, + 0.018789125606417656, + 0.0094795236364007, + 0.008496616967022419, + 0.013250762596726418, + 0.020409749820828438, + 0.012602977454662323, + 0.02005540020763874, + 0.02572658844292164, + 0.009157135151326656, + 0.005691226571798325, + 0.004402067046612501, + 0.006503553129732609, + 0.029055550694465637, + 0.007094091270118952, + 0.007110258564352989, + 0.01323363184928894, + 0.00485881045460701, + 0.0032538066152483225, + 0.002562593435868621, + 0.003123018890619278, + 0.01433554757386446, + 0.003966683521866798, + 0.006573076825588942, + 0.013729910366237164, + 0.011055824346840382, + 0.008712832815945148, + 0.004940011538565159, + 0.02318994700908661, + 0.01177196204662323, + 0.009633723646402359, + 0.024601219221949577 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 1, + 6 + ], + "selected_reward": -0.002506093354895711 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a vast, snow-covered landscape?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the visual style rendered with soft edges and luminous transparency resembling a vivid watercolor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the woman wearing a thick, cream-colored wool coat with a soft, painterly texture?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the northern lights visible in the deep blue sky with bleeding emerald green, ethereal violet, and soft mauve colors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot framed as a static, low-angle, medium-long shot?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.009785657748579979 + ], + "opposite_logprob_tokens": [ + -4.6347856521606445 + ], + "logprob_reward": -0.009785657748579979, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.009785657748579979, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.024768929928541183, + 0.021906930953264236, + 0.0339791364967823, + 0.03182227164506912, + 0.09622861444950104, + 0.07587183266878128, + 0.0570765882730484, + 0.5586007833480835, + 0.041453152894973755, + 0.05829174816608429 + ], + "credit_map": [ + 0.003364250995218754, + 0.003418974345549941, + 0.002922812942415476, + 0.0028911211993545294, + 0.002757640788331628, + 0.002934874501079321, + 0.0034313781652599573, + 0.003047878621146083, + 0.0032157213427126408, + 0.002325543435290456, + 0.001708722673356533, + 0.0023242090828716755, + 0.0031996837351471186, + 0.0019933367148041725, + 0.002731231041252613, + 0.004408482927829027, + 0.004481662530452013, + 0.004518999718129635, + 0.0033177402801811695, + 0.004081240389496088, + 0.005215353798121214, + 0.0034032452385872602, + 0.004330843687057495, + 0.0046300520189106464, + 0.003704919945448637, + 0.0036528678610920906, + 0.002906904788687825, + 0.004230303689837456, + 0.003907859791070223, + 0.0032765225041657686, + 0.005744841881096363, + 0.004398050252348185, + 0.010923270136117935, + 0.011106060817837715, + 0.008541559800505638, + 0.015769852325320244, + 0.014062714762985706, + 0.008974446915090084, + 0.01350477710366249, + 0.013345940038561821, + 0.008685245178639889, + 0.010743064805865288, + 0.006207944825291634, + 0.013087732717394829, + 0.00927034206688404, + 0.009266012348234653, + 0.011437971144914627, + 0.00717351958155632, + 0.007825336419045925, + 0.00565052917227149, + 0.00319674308411777, + 0.003729006042703986, + 0.012897067703306675, + 0.012672803364694118, + 0.003623089985921979, + 0.007482009474188089, + 0.17823685705661774, + 0.02649029530584812, + 0.02828216925263405, + 0.01175115630030632, + 0.04262486845254898, + 0.22102496027946472, + 0.02335434779524803, + 0.02683616615831852, + 0.0106663191691041, + 0.004229356534779072, + 0.0027867737226188183, + 0.0022248339373618364, + 0.002690077293664217, + 0.013089410960674286, + 0.0035993363708257675, + 0.002167045371606946, + 0.007545292843133211, + 0.005023354664444923, + 0.004698775243014097, + 0.002259115921333432, + 0.01469045877456665, + 0.006947591435164213, + 0.004257442895323038, + 0.012869719415903091 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 4, + 5 + ], + "selected_reward": -0.009785657748579979 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a vast, snow-covered landscape?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the visual style rendered with soft edges and luminous transparency resembling a vivid watercolor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the woman wearing a thick, cream-colored wool coat with a soft, painterly texture?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the northern lights visible in the deep blue sky with bleeding emerald green, ethereal violet, and soft mauve colors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot framed as a static, low-angle, medium-long shot?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0024880189448595047 + ], + "opposite_logprob_tokens": [ + -6.002488136291504 + ], + "logprob_reward": -0.0024880189448595047, + "yes_no_margin_reward": 6.0, + "token_credit_reward": -0.0024880189448595047, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09826068580150604, + 0.08775416016578674, + 0.08379588276147842, + 0.08557523787021637, + 0.16807110607624054, + 0.06779684871435165, + 0.0717908963561058, + 0.15633663535118103, + 0.09571345895528793, + 0.0849049836397171 + ], + "credit_map": [ + 0.01667727343738079, + 0.012784265913069248, + 0.011603537015616894, + 0.01166284829378128, + 0.007496883627027273, + 0.011710027232766151, + 0.011857486329972744, + 0.014468380250036716, + 0.01092735305428505, + 0.008971202187240124, + 0.006271775811910629, + 0.007674030959606171, + 0.016432536765933037, + 0.008520776405930519, + 0.011355895549058914, + 0.017600594088435173, + 0.01082052942365408, + 0.009040344506502151, + 0.007788992021232843, + 0.009191606193780899, + 0.00714878411963582, + 0.007667036261409521, + 0.010206409730017185, + 0.02193218469619751, + 0.009168672375380993, + 0.01136296708136797, + 0.009476964361965656, + 0.009103013202548027, + 0.01018973346799612, + 0.00922570750117302, + 0.011876607313752174, + 0.015171577222645283, + 0.023735245689749718, + 0.015414069406688213, + 0.01217849925160408, + 0.01375796739012003, + 0.042952314019203186, + 0.01860393024981022, + 0.015197285450994968, + 0.026231804862618446, + 0.008478376083076, + 0.006958845537155867, + 0.005523718893527985, + 0.00561654195189476, + 0.017089078202843666, + 0.007739244494587183, + 0.006304375361651182, + 0.010086677968502045, + 0.009565412066876888, + 0.01035736408084631, + 0.005561037454754114, + 0.005414158571511507, + 0.008910519070923328, + 0.013761916197836399, + 0.0067984056659042835, + 0.01142208930104971, + 0.032876044511795044, + 0.013957498595118523, + 0.00877735111862421, + 0.006333413999527693, + 0.027974175289273262, + 0.03199390694499016, + 0.013608894310891628, + 0.020815370604395866, + 0.03049180656671524, + 0.00729144923388958, + 0.0052590277045965195, + 0.002236404921859503, + 0.005908312276005745, + 0.032699890434741974, + 0.004313377663493156, + 0.007513202261179686, + 0.010334242135286331, + 0.007119427435100079, + 0.007824400439858437, + 0.004096720367670059, + 0.01898638717830181, + 0.009254119358956814, + 0.006896668113768101, + 0.02039303071796894 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 7, + 0 + ], + "selected_reward": -0.0024880189448595047 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a vast, snow-covered landscape?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the visual style rendered with soft edges and luminous transparency resembling a vivid watercolor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the woman wearing a thick, cream-colored wool coat with a soft, painterly texture?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the northern lights visible in the deep blue sky with bleeding emerald green, ethereal violet, and soft mauve colors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot framed as a static, low-angle, medium-long shot?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.3485725522041321 + ], + "opposite_logprob_tokens": [ + -1.2235724925994873 + ], + "logprob_reward": -0.3485725522041321, + "yes_no_margin_reward": 0.8749999403953552, + "token_credit_reward": -0.3485725522041321, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08283817023038864, + 0.0871230736374855, + 0.04672715440392494, + 0.09885373711585999, + 0.08226186037063599, + 0.11223634332418442, + 0.11272887885570526, + 0.17249806225299835, + 0.0914497971534729, + 0.1132829412817955 + ], + "credit_map": [ + 0.010272416286170483, + 0.009317858144640923, + 0.007913329638540745, + 0.009486591443419456, + 0.00902076717466116, + 0.00858250167220831, + 0.016696887090802193, + 0.011547810398042202, + 0.011315280571579933, + 0.01170857809484005, + 0.00815595593303442, + 0.010928722098469734, + 0.01496743131428957, + 0.007804613094776869, + 0.010774916969239712, + 0.01146757137030363, + 0.006549885496497154, + 0.005172166042029858, + 0.004255687352269888, + 0.005177422426640987, + 0.004812254570424557, + 0.004487947095185518, + 0.005059783812612295, + 0.011212009005248547, + 0.011113378219306469, + 0.014604620635509491, + 0.011409841477870941, + 0.009822679683566093, + 0.011263910681009293, + 0.009150494821369648, + 0.01387748308479786, + 0.017611319199204445, + 0.010286213830113411, + 0.012068266980350018, + 0.00693156011402607, + 0.018127018585801125, + 0.007733718026429415, + 0.009749048389494419, + 0.009836631827056408, + 0.007529400754719973, + 0.01496397890150547, + 0.01752481982111931, + 0.008993401192128658, + 0.009507703594863415, + 0.019435305148363113, + 0.012528752908110619, + 0.013299494981765747, + 0.01598287746310234, + 0.013470571488142014, + 0.014827552251517773, + 0.00816016923636198, + 0.0074853552505373955, + 0.020395493134856224, + 0.01205579936504364, + 0.011109556071460247, + 0.02522437460720539, + 0.03241180628538132, + 0.018154971301555634, + 0.010172504000365734, + 0.01087234541773796, + 0.03152591362595558, + 0.030378306284546852, + 0.016576683148741722, + 0.02240551821887493, + 0.02969730831682682, + 0.005539024248719215, + 0.004971475340425968, + 0.002188730053603649, + 0.005833826959133148, + 0.03330248221755028, + 0.004481870215386152, + 0.0054350742138922215, + 0.013888736255466938, + 0.02167794667184353, + 0.009664438664913177, + 0.011204813607037067, + 0.019780535250902176, + 0.010110284201800823, + 0.010009047575294971, + 0.01694714091718197 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 9, + 6 + ], + "selected_reward": -0.3485725522041321 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "static, low-angle, medium-long shot frames a young woman amidst a vast, silent, snow-covered landscape, rendered with the soft edges and luminous transparency of a vivid watercolor. Her long, dark hair, gently stirred by the crisp evening air, is adorned with a distinctive, vibrant red ribbon, its saturated color creating a striking accent against the cool palette. She is wrapped in a thick, cream-colored wool coat, its texture appearing soft and painterly. Her head is tilted fully back, gazing upward, her pose completely focused. Her facial expression is one of pure wonder, her wide, sparkling eyes reflecting the celestial scene, her lips slightly parted in silent awe. Above her, the northern lights shimmer and dance across the deep blue sky—not clear rays of light, but magnificent, bleeding emerald green, ethereal violet, and soft mauve, their colors blending seamlessly as if applied with a brush to a damp canvas. This otherworldly light casts a dynamic, colorful glow on her face and the powdery snow at her feet, creating an atmosphere of tranquil, magical contemplation.", + "seed": 3760, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000101/group_4/index_7096_seed_3760_dp_8_rank_16_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 4, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.0468235066626221, + "vlm_logprob": -0.0468235066626221, + "vlm_yes_no_margin": 4.1, + "vlm_token_credit": -0.0468235066626221, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a vast, snow-covered landscape?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the visual style rendered with soft edges and luminous transparency resembling a vivid watercolor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the woman wearing a thick, cream-colored wool coat with a soft, painterly texture?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the northern lights visible in the deep blue sky with bleeding emerald green, ethereal violet, and soft mauve colors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot framed as a static, low-angle, medium-long shot?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.05714461952447891 + ], + "opposite_logprob_tokens": [ + -2.9321446418762207 + ], + "logprob_reward": -0.05714461952447891, + "yes_no_margin_reward": 2.875, + "token_credit_reward": -0.05714461952447891, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11517958343029022, + 0.06271830201148987, + 0.1938401758670807, + 0.10844063758850098, + 0.14650669693946838, + 0.0768931433558464, + 0.05255519598722458, + 0.03676518797874451, + 0.03671754524111748, + 0.1703835278749466 + ], + "credit_map": [ + 0.015598735772073269, + 0.015094693750143051, + 0.013919666409492493, + 0.019354691728949547, + 0.0158346276730299, + 0.013900361955165863, + 0.01090247929096222, + 0.01057433057576418, + 0.011053316295146942, + 0.009494028054177761, + 0.0075562517158687115, + 0.007928471080958843, + 0.007962662726640701, + 0.00686388136819005, + 0.006492153741419315, + 0.005367535166442394, + 0.033783867955207825, + 0.062141310423612595, + 0.018577372655272484, + 0.015997830778360367, + 0.012127232737839222, + 0.028432045131921768, + 0.010823813267052174, + 0.011956707574427128, + 0.01529488991945982, + 0.02286800928413868, + 0.012540435418486595, + 0.013657755218446255, + 0.01251598633825779, + 0.011437173932790756, + 0.010835420340299606, + 0.009290975518524647, + 0.02398238144814968, + 0.03346513956785202, + 0.016870731487870216, + 0.020551422610878944, + 0.013498596847057343, + 0.012105477973818779, + 0.012859410606324673, + 0.013173535466194153, + 0.014301968738436699, + 0.012626788578927517, + 0.009119478054344654, + 0.008835499174892902, + 0.008456024341285229, + 0.008712561801075935, + 0.006704134866595268, + 0.008136695250868797, + 0.01308067049831152, + 0.018134666606783867, + 0.005705629009753466, + 0.0032051587477326393, + 0.003490633564069867, + 0.003773154690861702, + 0.0025006472133100033, + 0.0026646365877240896, + 0.0060057491064071655, + 0.012179841287434101, + 0.005981334950774908, + 0.003088797675445676, + 0.0024900862481445074, + 0.0030583515763282776, + 0.002010500291362405, + 0.0019505268428474665, + 0.007984819822013378, + 0.010723419487476349, + 0.004175085108727217, + 0.002940164878964424, + 0.0025118105113506317, + 0.0037696766667068005, + 0.0021623875945806503, + 0.00245018326677382, + 0.03790130093693733, + 0.020507320761680603, + 0.025639522820711136, + 0.02571958489716053, + 0.017799517139792442, + 0.016205400228500366, + 0.013180157169699669, + 0.013430734165012836 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 9, + 4 + ], + "selected_reward": -0.05714461952447891 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a vast, snow-covered landscape?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the visual style rendered with soft edges and luminous transparency resembling a vivid watercolor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the woman wearing a thick, cream-colored wool coat with a soft, painterly texture?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the northern lights visible in the deep blue sky with bleeding emerald green, ethereal violet, and soft mauve colors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot framed as a static, low-angle, medium-long shot?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.002221261151134968 + ], + "opposite_logprob_tokens": [ + -6.12722110748291 + ], + "logprob_reward": -0.002221261151134968, + "yes_no_margin_reward": 6.125, + "token_credit_reward": -0.002221261151134968, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12325891107320786, + 0.07773534208536148, + 0.04086842015385628, + 0.06585616618394852, + 0.09322308003902435, + 0.31481069326400757, + 0.16334740817546844, + 0.036166876554489136, + 0.04404786229133606, + 0.04068518429994583 + ], + "credit_map": [ + 0.012582078576087952, + 0.021725647151470184, + 0.017234420403838158, + 0.019946664571762085, + 0.01814117096364498, + 0.011142068542540073, + 0.014085674658417702, + 0.008401203900575638, + 0.009828786365687847, + 0.012570204213261604, + 0.006884285248816013, + 0.009721565991640091, + 0.009796768426895142, + 0.01119241677224636, + 0.010189513675868511, + 0.0075518060475587845, + 0.005660097114741802, + 0.006998266093432903, + 0.0068909889087080956, + 0.004445347469300032, + 0.0041551669128239155, + 0.0038635798264294863, + 0.004372341558337212, + 0.004482634831219912, + 0.011266895569860935, + 0.02006690204143524, + 0.006576021667569876, + 0.006348102353513241, + 0.004807729739695787, + 0.007162838242948055, + 0.005266566760838032, + 0.004361118655651808, + 0.014947179704904556, + 0.018018852919340134, + 0.013030393049120903, + 0.013166163116693497, + 0.010403250344097614, + 0.009325115010142326, + 0.006902862805873156, + 0.007429269142448902, + 0.06764844059944153, + 0.07820983976125717, + 0.04463779553771019, + 0.035310547798871994, + 0.02363048680126667, + 0.02622859738767147, + 0.018212078139185905, + 0.020932912826538086, + 0.0489833690226078, + 0.06665333360433578, + 0.012734965421259403, + 0.006844528019428253, + 0.0074442289769649506, + 0.01167949941009283, + 0.004874240607023239, + 0.004133259877562523, + 0.00744350953027606, + 0.006385032553225756, + 0.0037975963205099106, + 0.003991276025772095, + 0.004546081647276878, + 0.0031138798221945763, + 0.0031279108952730894, + 0.003761593485251069, + 0.0068198502995073795, + 0.0066900188103318214, + 0.009038017131388187, + 0.007815178483724594, + 0.004204161930829287, + 0.003637974616140127, + 0.002498852787539363, + 0.003343817312270403, + 0.010427319444715977, + 0.005544512532651424, + 0.0079848887398839, + 0.005007636267691851, + 0.0034046187065541744, + 0.0030251238495111465, + 0.002305316273123026, + 0.0029857715126127005 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 6, + 0 + ], + "selected_reward": -0.002221261151134968 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a vast, snow-covered landscape?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the visual style rendered with soft edges and luminous transparency resembling a vivid watercolor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the woman wearing a thick, cream-colored wool coat with a soft, painterly texture?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the northern lights visible in the deep blue sky with bleeding emerald green, ethereal violet, and soft mauve colors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot framed as a static, low-angle, medium-long shot?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.029802024364471436 + ], + "opposite_logprob_tokens": [ + -3.529802083969116 + ], + "logprob_reward": -0.029802024364471436, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.029802024364471436, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.18688860535621643, + 0.24767173826694489, + 0.08407279849052429, + 0.04719258472323418, + 0.06033680588006973, + 0.05873076990246773, + 0.08249621838331223, + 0.07627115398645401, + 0.044977784156799316, + 0.11136158555746078 + ], + "credit_map": [ + 0.02062463015317917, + 0.018597666174173355, + 0.03526342287659645, + 0.03251050040125847, + 0.024191981181502342, + 0.014710268937051296, + 0.023716995492577553, + 0.017273135483264923, + 0.030956994742155075, + 0.03412163630127907, + 0.03040202520787716, + 0.03013790026307106, + 0.03761861100792885, + 0.030087275430560112, + 0.03152484819293022, + 0.022822432219982147, + 0.01586136780679226, + 0.012413619086146355, + 0.012371904216706753, + 0.010988002642989159, + 0.008981156162917614, + 0.009042839519679546, + 0.007251182105392218, + 0.007162729278206825, + 0.00698307016864419, + 0.014403904788196087, + 0.005830463487654924, + 0.004435645416378975, + 0.0034257874358445406, + 0.004784327000379562, + 0.003714450169354677, + 0.0036149355582892895, + 0.010449836030602455, + 0.02015163190662861, + 0.005041084717959166, + 0.0059502013027668, + 0.005176577717065811, + 0.0055280765518546104, + 0.004224784672260284, + 0.0038146169390529394, + 0.00819349940866232, + 0.008360293693840504, + 0.0048525212332606316, + 0.004562111105769873, + 0.009521380998194218, + 0.007871120236814022, + 0.00808755587786436, + 0.0072822873480618, + 0.017909416928887367, + 0.028059758245944977, + 0.00915000680834055, + 0.0064651863649487495, + 0.0056533170863986015, + 0.006199705880135298, + 0.004364512395113707, + 0.0046943179331719875, + 0.023056263104081154, + 0.019247518852353096, + 0.006329391151666641, + 0.00589837646111846, + 0.006159481592476368, + 0.0055134715512394905, + 0.0051329126581549644, + 0.004933737684041262, + 0.011463001370429993, + 0.008074847050011158, + 0.005665682256221771, + 0.0060067069716751575, + 0.0032633068040013313, + 0.004034365527331829, + 0.0034586372785270214, + 0.0030112359672784805, + 0.0327797457575798, + 0.014146558940410614, + 0.01343892514705658, + 0.012001934461295605, + 0.012064182199537754, + 0.010617087595164776, + 0.008642628788948059, + 0.007670516148209572 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 9 + ], + "selected_reward": -0.029802024364471436 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a vast, snow-covered landscape?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the visual style rendered with soft edges and luminous transparency resembling a vivid watercolor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the woman wearing a thick, cream-colored wool coat with a soft, painterly texture?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the northern lights visible in the deep blue sky with bleeding emerald green, ethereal violet, and soft mauve colors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot framed as a static, low-angle, medium-long shot?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0021972341928631067 + ], + "opposite_logprob_tokens": [ + -6.127197265625 + ], + "logprob_reward": -0.0021972341928631067, + "yes_no_margin_reward": 6.125, + "token_credit_reward": -0.0021972341928631067, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13814988732337952, + 0.052830711007118225, + 0.03673988953232765, + 0.05568825453519821, + 0.10023075342178345, + 0.08040279150009155, + 0.13259826600551605, + 0.05369030684232712, + 0.03487425670027733, + 0.31479477882385254 + ], + "credit_map": [ + 0.017636476084589958, + 0.0196797177195549, + 0.0171679574996233, + 0.022100787609815598, + 0.016362810507416725, + 0.01338122971355915, + 0.016569441184401512, + 0.01525147445499897, + 0.006924607791006565, + 0.009876064956188202, + 0.005600465927273035, + 0.006005001254379749, + 0.007648617494851351, + 0.006337759084999561, + 0.006192184519022703, + 0.004246012773364782, + 0.006403411738574505, + 0.005033900961279869, + 0.005874411668628454, + 0.006397028919309378, + 0.0037884432822465897, + 0.0035842282231897116, + 0.0031011521350592375, + 0.0025573160964995623, + 0.008034349419176579, + 0.014155575074255466, + 0.0062010097317397594, + 0.00527570815756917, + 0.005943307653069496, + 0.006045744754374027, + 0.005128504242748022, + 0.0049040610902011395, + 0.022341858595609665, + 0.0306381955742836, + 0.0108483312651515, + 0.00875259842723608, + 0.007181841880083084, + 0.008092186413705349, + 0.005846918560564518, + 0.0065288301557302475, + 0.017192138358950615, + 0.016787895932793617, + 0.009067357517778873, + 0.008037102408707142, + 0.009327956475317478, + 0.008836367167532444, + 0.005892336368560791, + 0.005261646118015051, + 0.03889157623052597, + 0.05212037265300751, + 0.01062388252466917, + 0.007808874361217022, + 0.006295480765402317, + 0.0094186095520854, + 0.004344797693192959, + 0.003094688756391406, + 0.007367249112576246, + 0.014881913550198078, + 0.008145293220877647, + 0.007273708004504442, + 0.004350882023572922, + 0.00388781214132905, + 0.003665075171738863, + 0.004118375014513731, + 0.005707214120775461, + 0.006970237009227276, + 0.00437946105375886, + 0.004373738542199135, + 0.0038822998758405447, + 0.0032024458050727844, + 0.0027405517175793648, + 0.0036183123011142015, + 0.0993027612566948, + 0.04048418626189232, + 0.04599536210298538, + 0.0381377711892128, + 0.022037524729967117, + 0.029263276606798172, + 0.019502142444252968, + 0.020071789622306824 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 0, + 6 + ], + "selected_reward": -0.0021972341928631067 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a vast, snow-covered landscape?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the visual style rendered with soft edges and luminous transparency resembling a vivid watercolor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the woman wearing a thick, cream-colored wool coat with a soft, painterly texture?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the northern lights visible in the deep blue sky with bleeding emerald green, ethereal violet, and soft mauve colors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot framed as a static, low-angle, medium-long shot?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.14275239408016205 + ], + "opposite_logprob_tokens": [ + -2.0177524089813232 + ], + "logprob_reward": -0.14275239408016205, + "yes_no_margin_reward": 1.875, + "token_credit_reward": -0.14275239408016205, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12192019820213318, + 0.07925017178058624, + 0.22213242948055267, + 0.08835053443908691, + 0.06515572965145111, + 0.07779156416654587, + 0.19449035823345184, + 0.04635937139391899, + 0.045127201825380325, + 0.05942239612340927 + ], + "credit_map": [ + 0.010408246889710426, + 0.013824062421917915, + 0.015046139247715473, + 0.018573766574263573, + 0.0195133276283741, + 0.01637570559978485, + 0.014345548115670681, + 0.013833410106599331, + 0.012261544354259968, + 0.012664665468037128, + 0.009722833521664143, + 0.010412449017167091, + 0.009071236476302147, + 0.00868974719196558, + 0.00923915859311819, + 0.007188539486378431, + 0.03420642390847206, + 0.07755322754383087, + 0.021776964887976646, + 0.0180759746581316, + 0.012899194844067097, + 0.03745238855481148, + 0.01057805772870779, + 0.009590229019522667, + 0.009479200467467308, + 0.01885617896914482, + 0.009142193011939526, + 0.008448965847492218, + 0.011304530315101147, + 0.011498378589749336, + 0.009899919852614403, + 0.009721173904836178, + 0.007317825220525265, + 0.008937539532780647, + 0.005310526583343744, + 0.008666038513183594, + 0.009451099671423435, + 0.007178789004683495, + 0.008773899637162685, + 0.009520011954009533, + 0.013254114426672459, + 0.014167062006890774, + 0.009365281090140343, + 0.008879205211997032, + 0.00910340715199709, + 0.00979491975158453, + 0.006587098818272352, + 0.006640485487878323, + 0.05464216694235802, + 0.07418061792850494, + 0.01747070625424385, + 0.010944963432848454, + 0.009517387486994267, + 0.013351386412978172, + 0.007015238981693983, + 0.0073678926564753056, + 0.009878925047814846, + 0.013813713565468788, + 0.0036128582432866096, + 0.004479507450014353, + 0.0035689272917807102, + 0.004186464007943869, + 0.003075543325394392, + 0.0037434364203363657, + 0.010696504265069962, + 0.009528150781989098, + 0.005729369353502989, + 0.004788228776305914, + 0.004610502161085606, + 0.004037277773022652, + 0.0025277086533606052, + 0.003209460759535432, + 0.014372899197041988, + 0.008616872131824493, + 0.008556650951504707, + 0.00785570777952671, + 0.006381263490766287, + 0.005335640627890825, + 0.004654949065297842, + 0.003648419864475727 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 6, + 0 + ], + "selected_reward": -0.14275239408016205 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "static, low-angle, medium-long shot frames a young woman amidst a vast, silent, snow-covered landscape, rendered with the soft edges and luminous transparency of a vivid watercolor. Her long, dark hair, gently stirred by the crisp evening air, is adorned with a distinctive, vibrant red ribbon, its saturated color creating a striking accent against the cool palette. She is wrapped in a thick, cream-colored wool coat, its texture appearing soft and painterly. Her head is tilted fully back, gazing upward, her pose completely focused. Her facial expression is one of pure wonder, her wide, sparkling eyes reflecting the celestial scene, her lips slightly parted in silent awe. Above her, the northern lights shimmer and dance across the deep blue sky—not clear rays of light, but magnificent, bleeding emerald green, ethereal violet, and soft mauve, their colors blending seamlessly as if applied with a brush to a damp canvas. This otherworldly light casts a dynamic, colorful glow on her face and the powdery snow at her feet, creating an atmosphere of tranquil, magical contemplation.", + "seed": 3761, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000101/group_4/index_7096_seed_3761_dp_9_rank_18_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 4, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.055139195872470737, + "vlm_logprob": -0.055139195872470737, + "vlm_yes_no_margin": 4.375, + "vlm_token_credit": -0.055139195872470737, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a vast, snow-covered landscape?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the visual style rendered with soft edges and luminous transparency resembling a vivid watercolor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the woman wearing a thick, cream-colored wool coat with a soft, painterly texture?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the northern lights visible in the deep blue sky with bleeding emerald green, ethereal violet, and soft mauve colors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot framed as a static, low-angle, medium-long shot?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.1027367115020752 + ], + "opposite_logprob_tokens": [ + -2.352736711502075 + ], + "logprob_reward": -0.1027367115020752, + "yes_no_margin_reward": 2.25, + "token_credit_reward": -0.1027367115020752, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1683807522058487, + 0.04891367629170418, + 0.07244600355625153, + 0.25989532470703125, + 0.04388468712568283, + 0.051277775317430496, + 0.23850472271442413, + 0.08270851522684097, + 0.019830161705613136, + 0.014158472418785095 + ], + "credit_map": [ + 0.015674615278840065, + 0.010363600216805935, + 0.03190995752811432, + 0.02270253375172615, + 0.013196368701756, + 0.019922373816370964, + 0.014881090261042118, + 0.039730217307806015, + 0.004923798143863678, + 0.003638818161562085, + 0.01197355892509222, + 0.005495535209774971, + 0.003626910736784339, + 0.006075129844248295, + 0.00475345691666007, + 0.008426466956734657, + 0.00839462224394083, + 0.006150905042886734, + 0.00953004788607359, + 0.015317544341087341, + 0.007250406313687563, + 0.006268181838095188, + 0.009476962499320507, + 0.010057325474917889, + 0.049464061856269836, + 0.01239657774567604, + 0.05248672887682915, + 0.011438373476266861, + 0.009192808531224728, + 0.015294201672077179, + 0.008652790449559689, + 0.10096976161003113, + 0.004083585925400257, + 0.004178761038929224, + 0.012823758646845818, + 0.0033565936610102654, + 0.0032466172706335783, + 0.004049855750054121, + 0.0021922271698713303, + 0.009953286498785019, + 0.0059368796646595, + 0.005991400219500065, + 0.0066154515370726585, + 0.007896696217358112, + 0.005979869980365038, + 0.005877352319657803, + 0.005735985469073057, + 0.007244143169373274, + 0.014964909292757511, + 0.008933995850384235, + 0.09755607694387436, + 0.008473186753690243, + 0.006161134224385023, + 0.012220301665365696, + 0.004970738664269447, + 0.08522437512874603, + 0.010049850679934025, + 0.010514366440474987, + 0.014555780217051506, + 0.009068300947546959, + 0.009303753264248371, + 0.01043879147619009, + 0.009840942919254303, + 0.008936726488173008, + 0.0019999584183096886, + 0.001774072414264083, + 0.0026837424375116825, + 0.0023246679920703173, + 0.00285340310074389, + 0.0025344318710267544, + 0.0024181006010621786, + 0.0032417848706245422, + 0.002679448574781418, + 0.0015324077103286982, + 0.0016956391045823693, + 0.0015277258353307843, + 0.0018228847766295075, + 0.001980721717700362, + 0.0013220300897955894, + 0.0015976150752976537 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 6, + 0 + ], + "selected_reward": -0.1027367115020752 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a vast, snow-covered landscape?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the visual style rendered with soft edges and luminous transparency resembling a vivid watercolor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the woman wearing a thick, cream-colored wool coat with a soft, painterly texture?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the northern lights visible in the deep blue sky with bleeding emerald green, ethereal violet, and soft mauve colors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot framed as a static, low-angle, medium-long shot?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0009382136631757021 + ], + "opposite_logprob_tokens": [ + -7.000938415527344 + ], + "logprob_reward": -0.0009382136631757021, + "yes_no_margin_reward": 7.0, + "token_credit_reward": -0.0009382136631757021, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11182854324579239, + 0.020909544080495834, + 0.04664205387234688, + 0.12726819515228271, + 0.02031608298420906, + 0.04090886563062668, + 0.5759655833244324, + 0.02538255788385868, + 0.016939358785748482, + 0.013839149847626686 + ], + "credit_map": [ + 0.010919484309852123, + 0.010096441954374313, + 0.02641214244067669, + 0.014172486029565334, + 0.00628441059961915, + 0.01363054383546114, + 0.006582929287105799, + 0.02373010665178299, + 0.0020983635913580656, + 0.0016098091145977378, + 0.004857122898101807, + 0.0025598674546927214, + 0.0013862357009202242, + 0.0026872234884649515, + 0.0014504275750368834, + 0.004260494373738766, + 0.006869164295494556, + 0.00519903376698494, + 0.006709363777190447, + 0.006457229144871235, + 0.004463351797312498, + 0.0063069541938602924, + 0.004938229918479919, + 0.00569872884079814, + 0.02543543092906475, + 0.005204903427511454, + 0.026758451014757156, + 0.004685844294726849, + 0.0034165941178798676, + 0.008707334287464619, + 0.0041330838575959206, + 0.04892656207084656, + 0.0024304792750626802, + 0.0014815847389400005, + 0.005040759686380625, + 0.0016945120878517628, + 0.002242565620690584, + 0.0022122887894511223, + 0.0016187706496566534, + 0.003595123067498207, + 0.0037246178835630417, + 0.002626581583172083, + 0.004824490286409855, + 0.002289452124387026, + 0.0023292629048228264, + 0.01097929198294878, + 0.0018904325552284718, + 0.012244737707078457, + 0.03325185924768448, + 0.01724068634212017, + 0.24899008870124817, + 0.01497279480099678, + 0.006089179776608944, + 0.03202870115637779, + 0.007434399798512459, + 0.215957909822464, + 0.0033281745854765177, + 0.0030021057464182377, + 0.0035284669138491154, + 0.0026414275635033846, + 0.002808394143357873, + 0.004808622412383556, + 0.002658788114786148, + 0.0026065814308822155, + 0.0016854227287694812, + 0.0014490922912955284, + 0.001985263777896762, + 0.0019044908694922924, + 0.0032544087152928114, + 0.0027164591010659933, + 0.0019557480700314045, + 0.0019884747453033924, + 0.001825655926950276, + 0.0012755647767335176, + 0.0022876800503581762, + 0.0013803454348817468, + 0.0016243079444393516, + 0.002087085973471403, + 0.0011717358138412237, + 0.002186776604503393 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 3, + 0 + ], + "selected_reward": -0.0009382136631757021 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a vast, snow-covered landscape?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the visual style rendered with soft edges and luminous transparency resembling a vivid watercolor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the woman wearing a thick, cream-colored wool coat with a soft, painterly texture?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the northern lights visible in the deep blue sky with bleeding emerald green, ethereal violet, and soft mauve colors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot framed as a static, low-angle, medium-long shot?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.009782115928828716 + ], + "opposite_logprob_tokens": [ + -4.634782314300537 + ], + "logprob_reward": -0.009782115928828716, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.009782115928828716, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.23104599118232727, + 0.0192717257887125, + 0.1933615654706955, + 0.09480330348014832, + 0.021482912823557854, + 0.046566709876060486, + 0.32465043663978577, + 0.044329240918159485, + 0.01148156076669693, + 0.013006501831114292 + ], + "credit_map": [ + 0.019265953451395035, + 0.017700767144560814, + 0.04290846362709999, + 0.042716484516859055, + 0.02550637535750866, + 0.021373344585299492, + 0.02570105344057083, + 0.035873547196388245, + 0.00168164586648345, + 0.0012940851738676429, + 0.0034406057093292475, + 0.0036077091936022043, + 0.002468247665092349, + 0.0019214045023545623, + 0.001908967737108469, + 0.0029490594752132893, + 0.026455547660589218, + 0.01955881528556347, + 0.023934004828333855, + 0.0313572883605957, + 0.01846100203692913, + 0.024167794734239578, + 0.01943298801779747, + 0.02999412827193737, + 0.016391843557357788, + 0.0040477244183421135, + 0.019035175442695618, + 0.007934543304145336, + 0.003879143390804529, + 0.0057950569316744804, + 0.004262556321918964, + 0.033457256853580475, + 0.0019362224265933037, + 0.0013863954227417707, + 0.0051808664575219154, + 0.002049992559477687, + 0.002558249980211258, + 0.0015176221495494246, + 0.0016657846281304955, + 0.005187779199331999, + 0.003975637257099152, + 0.0028205635026097298, + 0.0046474081464111805, + 0.0031510326080024242, + 0.0037752031348645687, + 0.012469282373785973, + 0.0025920632760971785, + 0.013135520741343498, + 0.020920060575008392, + 0.013638856820762157, + 0.1351540982723236, + 0.009787165559828281, + 0.004597679246217012, + 0.015466437675058842, + 0.0046525015495717525, + 0.12043366581201553, + 0.006524547468870878, + 0.005628034006804228, + 0.005754345096647739, + 0.004886337090283632, + 0.00415141461417079, + 0.008434723131358624, + 0.004649256356060505, + 0.004300578031688929, + 0.0012020448921248317, + 0.0010030108969658613, + 0.0013603054685518146, + 0.0018135573482140899, + 0.0016219643875956535, + 0.0015964774647727609, + 0.001528641558252275, + 0.0013555584009736776, + 0.00130971590988338, + 0.0012932665413245559, + 0.0014241294702515006, + 0.0015643197111785412, + 0.002142995595932007, + 0.0017244871705770493, + 0.0014383713714778423, + 0.0021092158276587725 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 0, + 2 + ], + "selected_reward": -0.009782115928828716 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a vast, snow-covered landscape?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the visual style rendered with soft edges and luminous transparency resembling a vivid watercolor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the woman wearing a thick, cream-colored wool coat with a soft, painterly texture?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the northern lights visible in the deep blue sky with bleeding emerald green, ethereal violet, and soft mauve colors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot framed as a static, low-angle, medium-long shot?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0019383705221116543 + ], + "opposite_logprob_tokens": [ + -6.251938343048096 + ], + "logprob_reward": -0.0019383705221116543, + "yes_no_margin_reward": 6.25, + "token_credit_reward": -0.0019383705221116543, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.16823843121528625, + 0.08969448506832123, + 0.028386417776346207, + 0.2823285758495331, + 0.09279683977365494, + 0.06426776200532913, + 0.12466772645711899, + 0.06861790269613266, + 0.024638241156935692, + 0.05636367201805115 + ], + "credit_map": [ + 0.012375183403491974, + 0.01081863697618246, + 0.038566458970308304, + 0.0229866411536932, + 0.012874305248260498, + 0.019005009904503822, + 0.01489298976957798, + 0.0367191918194294, + 0.009332586079835892, + 0.008946768008172512, + 0.01666606031358242, + 0.012737780809402466, + 0.00883730873465538, + 0.007486107759177685, + 0.010457241907715797, + 0.015230638906359673, + 0.0031655875500291586, + 0.0019133413443341851, + 0.004644182510674, + 0.003868688363581896, + 0.003053127322345972, + 0.001831498695537448, + 0.0029461414087563753, + 0.006963848602026701, + 0.051750220358371735, + 0.0135500468313694, + 0.06119898706674576, + 0.013489077799022198, + 0.008898509666323662, + 0.01774129830300808, + 0.009917182847857475, + 0.10578324645757675, + 0.015340973623096943, + 0.007794284261763096, + 0.011957444250583649, + 0.009965559467673302, + 0.010823777876794338, + 0.011324524879455566, + 0.007580405566841364, + 0.01800987496972084, + 0.00913966167718172, + 0.008789600804448128, + 0.008444168604910374, + 0.010235429741442204, + 0.006831980776041746, + 0.007697721011936665, + 0.006655334495007992, + 0.006473864428699017, + 0.015217488631606102, + 0.011864172294735909, + 0.02799210324883461, + 0.011522999964654446, + 0.00840742327272892, + 0.0151100168004632, + 0.009772185236215591, + 0.024781331419944763, + 0.009321412071585655, + 0.009899127297103405, + 0.0074736629612743855, + 0.006839745212346315, + 0.006802462041378021, + 0.01160991657525301, + 0.009072442539036274, + 0.007599132135510445, + 0.00390944629907608, + 0.002419356955215335, + 0.0031773678492754698, + 0.0025169069413095713, + 0.00249345856718719, + 0.0035794393625110388, + 0.0033082636073231697, + 0.00323400367051363, + 0.006194209214299917, + 0.004910229239612818, + 0.01022354606539011, + 0.0047060102224349976, + 0.005532146897166967, + 0.009025273844599724, + 0.004653282463550568, + 0.01111896988004446 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 0, + 6 + ], + "selected_reward": -0.0019383705221116543 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a vast, snow-covered landscape?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the visual style rendered with soft edges and luminous transparency resembling a vivid watercolor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the woman wearing a thick, cream-colored wool coat with a soft, painterly texture?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the northern lights visible in the deep blue sky with bleeding emerald green, ethereal violet, and soft mauve colors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot framed as a static, low-angle, medium-long shot?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.16030056774616241 + ], + "opposite_logprob_tokens": [ + -1.910300612449646 + ], + "logprob_reward": -0.16030056774616241, + "yes_no_margin_reward": 1.75, + "token_credit_reward": -0.16030056774616241, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.05680592730641365, + 0.03213798254728317, + 0.045935649424791336, + 0.017243921756744385, + 0.039838019758462906, + 0.025770118460059166, + 0.7116186618804932, + 0.020751075819134712, + 0.01995428465306759, + 0.029944363981485367 + ], + "credit_map": [ + 0.005156771745532751, + 0.005115726497024298, + 0.007812161464244127, + 0.008892156183719635, + 0.006777599919587374, + 0.006944097578525543, + 0.005993662867695093, + 0.010113755241036415, + 0.0024029724299907684, + 0.0024403135757893324, + 0.004161075688898563, + 0.006317188031971455, + 0.004811062011867762, + 0.0026800637133419514, + 0.004828191827982664, + 0.004497116897255182, + 0.006361125037074089, + 0.004333470482379198, + 0.005269328597933054, + 0.006855475716292858, + 0.004567096475511789, + 0.006370016373693943, + 0.004935598466545343, + 0.007243540603667498, + 0.0024485276080667973, + 0.0016436978476122022, + 0.0024419596884399652, + 0.002805576892569661, + 0.0019560102373361588, + 0.0017081103287637234, + 0.0014441279927268624, + 0.002795912092551589, + 0.0037780343554913998, + 0.0033835836220532656, + 0.0070236194878816605, + 0.004865582566708326, + 0.005968573968857527, + 0.0027432020287960768, + 0.003979926463216543, + 0.00809550005942583, + 0.002412489615380764, + 0.0018200591439381242, + 0.003375366097316146, + 0.0018928293138742447, + 0.0032193700317293406, + 0.004589154850691557, + 0.001541289733722806, + 0.006919560022652149, + 0.04111522436141968, + 0.021175416186451912, + 0.30697664618492126, + 0.01909770630300045, + 0.00895781721919775, + 0.03758808970451355, + 0.008815222419798374, + 0.26789259910583496, + 0.0028735543601214886, + 0.003161973785609007, + 0.002477380447089672, + 0.0021405278239399195, + 0.0026083241682499647, + 0.003103668102994561, + 0.0022127812262624502, + 0.0021728663705289364, + 0.0025736941024661064, + 0.0018680230714380741, + 0.0032092395704239607, + 0.0021298883948475122, + 0.0023517655208706856, + 0.0025208157021552324, + 0.0027458358090370893, + 0.0025550255086272955, + 0.003601415315642953, + 0.003425607457756996, + 0.006463371217250824, + 0.0028514459263533354, + 0.0027575669810175896, + 0.004109273664653301, + 0.0024464018642902374, + 0.004289284814149141 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 0, + 2 + ], + "selected_reward": -0.16030056774616241 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "static, low-angle, medium-long shot frames a young woman amidst a vast, silent, snow-covered landscape, rendered with the soft edges and luminous transparency of a vivid watercolor. Her long, dark hair, gently stirred by the crisp evening air, is adorned with a distinctive, vibrant red ribbon, its saturated color creating a striking accent against the cool palette. She is wrapped in a thick, cream-colored wool coat, its texture appearing soft and painterly. Her head is tilted fully back, gazing upward, her pose completely focused. Her facial expression is one of pure wonder, her wide, sparkling eyes reflecting the celestial scene, her lips slightly parted in silent awe. Above her, the northern lights shimmer and dance across the deep blue sky—not clear rays of light, but magnificent, bleeding emerald green, ethereal violet, and soft mauve, their colors blending seamlessly as if applied with a brush to a damp canvas. This otherworldly light casts a dynamic, colorful glow on her face and the powdery snow at her feet, creating an atmosphere of tranquil, magical contemplation.", + "seed": 3762, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000101/group_4/index_7096_seed_3762_dp_9_rank_18_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 4, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.05474302689544856, + "vlm_logprob": -0.05474302689544856, + "vlm_yes_no_margin": 4.575, + "vlm_token_credit": -0.05474302689544856, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a vast, snow-covered landscape?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the visual style rendered with soft edges and luminous transparency resembling a vivid watercolor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the woman wearing a thick, cream-colored wool coat with a soft, painterly texture?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the northern lights visible in the deep blue sky with bleeding emerald green, ethereal violet, and soft mauve colors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot framed as a static, low-angle, medium-long shot?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.006447347812354565 + ], + "opposite_logprob_tokens": [ + -5.6314473152160645 + ], + "logprob_reward": -0.006447347812354565, + "yes_no_margin_reward": 5.625, + "token_credit_reward": -0.006447347812354565, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.19678868353366852, + 0.04929523915052414, + 0.07220157980918884, + 0.045523229986429214, + 0.07134382426738739, + 0.06654941290616989, + 0.08842848986387253, + 0.07972987741231918, + 0.08283121883869171, + 0.24730835855007172 + ], + "credit_map": [ + 0.0307391919195652, + 0.037215206772089005, + 0.03051239810883999, + 0.028395486995577812, + 0.015358144417405128, + 0.01884639449417591, + 0.013867873698472977, + 0.021853988990187645, + 0.007087478414177895, + 0.006730174645781517, + 0.005410355515778065, + 0.007046999875456095, + 0.005308087449520826, + 0.005062120035290718, + 0.005739628802984953, + 0.006910392083227634, + 0.008633679710328579, + 0.010444479994475842, + 0.010510402731597424, + 0.010513541288673878, + 0.010969443246722221, + 0.007741336245089769, + 0.007069092709571123, + 0.006319609936326742, + 0.0062040346674621105, + 0.00622912310063839, + 0.010123259387910366, + 0.005374716594815254, + 0.006782866548746824, + 0.004232113249599934, + 0.0035910450387746096, + 0.0029860727954655886, + 0.008530831895768642, + 0.007732367608696222, + 0.011833513155579567, + 0.014379977248609066, + 0.009433521889150143, + 0.006324782036244869, + 0.007164609152823687, + 0.00594422547146678, + 0.008820746093988419, + 0.007079860661178827, + 0.0072177196852862835, + 0.007697180844843388, + 0.01076304167509079, + 0.0067357090301811695, + 0.012816475704312325, + 0.0054186806082725525, + 0.018714256584644318, + 0.013808119110763073, + 0.018164725974202156, + 0.01255993451923132, + 0.0075113484635949135, + 0.0073980954475700855, + 0.005706914700567722, + 0.004565100651234388, + 0.015134376473724842, + 0.011362521909177303, + 0.01345930714160204, + 0.015478698536753654, + 0.006567410659044981, + 0.008676552213728428, + 0.004745081532746553, + 0.004305931273847818, + 0.013843280263245106, + 0.011808588169515133, + 0.013513014651834965, + 0.01868855208158493, + 0.007725087460130453, + 0.005025905091315508, + 0.006483228877186775, + 0.005743551999330521, + 0.06365326792001724, + 0.03303976729512215, + 0.03856334090232849, + 0.04073575139045715, + 0.018587158992886543, + 0.019828828051686287, + 0.020114708691835403, + 0.01278554555028677 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 0, + 6 + ], + "selected_reward": -0.006447347812354565 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a vast, snow-covered landscape?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the visual style rendered with soft edges and luminous transparency resembling a vivid watercolor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the woman wearing a thick, cream-colored wool coat with a soft, painterly texture?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the northern lights visible in the deep blue sky with bleeding emerald green, ethereal violet, and soft mauve colors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot framed as a static, low-angle, medium-long shot?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0017337066819891334 + ], + "opposite_logprob_tokens": [ + -6.376733779907227 + ], + "logprob_reward": -0.0017337066819891334, + "yes_no_margin_reward": 6.375, + "token_credit_reward": -0.0017337066819891334, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.154554545879364, + 0.18723776936531067, + 0.06933724880218506, + 0.0661187693476677, + 0.11402582377195358, + 0.08413373678922653, + 0.06533516198396683, + 0.090256467461586, + 0.07359088957309723, + 0.09540962427854538 + ], + "credit_map": [ + 0.023031320422887802, + 0.02774817682802677, + 0.01952304318547249, + 0.01890351064503193, + 0.022216971963644028, + 0.01505275908857584, + 0.016126012429594994, + 0.011952743865549564, + 0.014877420850098133, + 0.03256246820092201, + 0.029490815475583076, + 0.04138806089758873, + 0.020945850759744644, + 0.01466365996748209, + 0.01699710264801979, + 0.01631239987909794, + 0.007219687569886446, + 0.007132372353225946, + 0.008882146328687668, + 0.01665174402296543, + 0.0057915085926651955, + 0.004944856744259596, + 0.008623064495623112, + 0.010091868229210377, + 0.0070315455086529255, + 0.008381715044379234, + 0.012879710644483566, + 0.013990680687129498, + 0.008364816196262836, + 0.004985773004591465, + 0.006247947923839092, + 0.004236580338329077, + 0.015612277202308178, + 0.01100711990147829, + 0.011741731315851212, + 0.015540809370577335, + 0.015683572739362717, + 0.014582913368940353, + 0.015950309112668037, + 0.013907092623412609, + 0.01674267277121544, + 0.00864679180085659, + 0.011870209127664566, + 0.015856046229600906, + 0.008013393729925156, + 0.008613976649940014, + 0.008968290872871876, + 0.005422347690910101, + 0.009841869585216045, + 0.01084311306476593, + 0.008264664560556412, + 0.011335632763803005, + 0.006998355966061354, + 0.006417021155357361, + 0.005851934663951397, + 0.005782568361610174, + 0.011133141815662384, + 0.010435906238853931, + 0.014854167588055134, + 0.013318384066224098, + 0.010003369301557541, + 0.006893057841807604, + 0.01055624708533287, + 0.013062192127108574, + 0.013033555820584297, + 0.008362575434148312, + 0.008713774383068085, + 0.011687137186527252, + 0.008274412713944912, + 0.006479666568338871, + 0.00857845414429903, + 0.00846131332218647, + 0.024356931447982788, + 0.013294286094605923, + 0.014118808321654797, + 0.015322079882025719, + 0.00760231725871563, + 0.0069269342347979546, + 0.007670470513403416, + 0.006117797922343016 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 4 + ], + "selected_reward": -0.0017337066819891334 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a vast, snow-covered landscape?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the visual style rendered with soft edges and luminous transparency resembling a vivid watercolor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the woman wearing a thick, cream-colored wool coat with a soft, painterly texture?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the northern lights visible in the deep blue sky with bleeding emerald green, ethereal violet, and soft mauve colors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot framed as a static, low-angle, medium-long shot?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03809133172035217 + ], + "opposite_logprob_tokens": [ + -3.2880914211273193 + ], + "logprob_reward": -0.03809133172035217, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.03809133172035217, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.17109964787960052, + 0.05494572967290878, + 0.04482646659016609, + 0.051474690437316895, + 0.12081345170736313, + 0.2661188542842865, + 0.02786445803940296, + 0.06923052668571472, + 0.14485789835453033, + 0.04876834526658058 + ], + "credit_map": [ + 0.018624989315867424, + 0.02033015713095665, + 0.011903559789061546, + 0.02052064798772335, + 0.023566681891679764, + 0.01765068992972374, + 0.023208022117614746, + 0.03529487922787666, + 0.004268743563443422, + 0.009699475020170212, + 0.008448024280369282, + 0.008732739835977554, + 0.005766031797975302, + 0.006566231604665518, + 0.006204947363585234, + 0.005259528756141663, + 0.004828613717108965, + 0.006835221778601408, + 0.007143108639866114, + 0.007393309380859137, + 0.00711022736504674, + 0.003539071185514331, + 0.004438622854650021, + 0.0035382884088903666, + 0.006358927581459284, + 0.006793990731239319, + 0.0073790960013866425, + 0.013007905334234238, + 0.004380594007670879, + 0.005378170404583216, + 0.004863068461418152, + 0.0033129306975752115, + 0.013644438236951828, + 0.011307477951049805, + 0.009576343931257725, + 0.01553672831505537, + 0.022294752299785614, + 0.012713339179754257, + 0.020385418087244034, + 0.015354943461716175, + 0.02362062782049179, + 0.02907031774520874, + 0.01708385720849037, + 0.02305031754076481, + 0.05528198182582855, + 0.029590941965579987, + 0.05973668769001961, + 0.028684111312031746, + 0.0036377892829477787, + 0.004192790947854519, + 0.002922200132161379, + 0.003508363850414753, + 0.0041313557885587215, + 0.002993619302287698, + 0.0035289162769913673, + 0.002949420129880309, + 0.0069371601566672325, + 0.007128648925572634, + 0.006042673718184233, + 0.008835854008793831, + 0.011600346304476261, + 0.007608200889080763, + 0.012454258278012276, + 0.008623377420008183, + 0.01896081119775772, + 0.016641173511743546, + 0.009156174026429653, + 0.015330922789871693, + 0.021633733063936234, + 0.015600677579641342, + 0.02874535508453846, + 0.018789032474160194, + 0.009134381078183651, + 0.006696976721286774, + 0.007157465443015099, + 0.0077927084639668465, + 0.004617147147655487, + 0.004460231401026249, + 0.004583665635436773, + 0.004325765650719404 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 0, + 8 + ], + "selected_reward": -0.03809133172035217 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a vast, snow-covered landscape?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the visual style rendered with soft edges and luminous transparency resembling a vivid watercolor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the woman wearing a thick, cream-colored wool coat with a soft, painterly texture?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the northern lights visible in the deep blue sky with bleeding emerald green, ethereal violet, and soft mauve colors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot framed as a static, low-angle, medium-long shot?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0019413450499996543 + ], + "opposite_logprob_tokens": [ + -6.251941204071045 + ], + "logprob_reward": -0.0019413450499996543, + "yes_no_margin_reward": 6.25, + "token_credit_reward": -0.0019413450499996543, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.20210076868534088, + 0.10455618798732758, + 0.07859449833631516, + 0.14362679421901703, + 0.0628020167350769, + 0.07692504674196243, + 0.07864750921726227, + 0.05226637050509453, + 0.07035969942808151, + 0.13012108206748962 + ], + "credit_map": [ + 0.026749232783913612, + 0.03622924163937569, + 0.024299779906868935, + 0.021812131628394127, + 0.024843744933605194, + 0.024589022621512413, + 0.025454960763454437, + 0.018122656270861626, + 0.014016261324286461, + 0.017827970907092094, + 0.011290130205452442, + 0.02268739603459835, + 0.009346523322165012, + 0.012962893582880497, + 0.008529947139322758, + 0.007895059883594513, + 0.009823094122111797, + 0.01207462977617979, + 0.0130220428109169, + 0.011034130118787289, + 0.007113783620297909, + 0.006903300993144512, + 0.00925137847661972, + 0.009372137486934662, + 0.02104446478188038, + 0.017046140506863594, + 0.033242758363485336, + 0.03267886862158775, + 0.010320044122636318, + 0.010648161172866821, + 0.010350975207984447, + 0.008295383304357529, + 0.013192574493587017, + 0.007017654832452536, + 0.013207010924816132, + 0.009968142956495285, + 0.005783668719232082, + 0.004327298142015934, + 0.005629011895507574, + 0.003676650347188115, + 0.009043347090482712, + 0.009106470271945, + 0.006318352185189724, + 0.008037657476961613, + 0.01257241703569889, + 0.008955487981438637, + 0.015157789923250675, + 0.0077335224486887455, + 0.010917119681835175, + 0.0103886928409338, + 0.01678379811346531, + 0.013077065348625183, + 0.008022826164960861, + 0.007941609248518944, + 0.006016868632286787, + 0.005499522667378187, + 0.005292797926813364, + 0.006139346398413181, + 0.009923439472913742, + 0.011663848534226418, + 0.00517728365957737, + 0.004071368370205164, + 0.005068552680313587, + 0.004929731599986553, + 0.00802415981888771, + 0.007667546160519123, + 0.006940782070159912, + 0.00856180302798748, + 0.010031508281826973, + 0.008038965985178947, + 0.011814934201538563, + 0.009280003607273102, + 0.025549206882715225, + 0.01957331970334053, + 0.021627049893140793, + 0.02413814887404442, + 0.010730578564107418, + 0.008874966762959957, + 0.011310159228742123, + 0.008317659609019756 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 9 + ], + "selected_reward": -0.0019413450499996543 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a vast, snow-covered landscape?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the visual style rendered with soft edges and luminous transparency resembling a vivid watercolor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the woman wearing a thick, cream-colored wool coat with a soft, painterly texture?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the northern lights visible in the deep blue sky with bleeding emerald green, ethereal violet, and soft mauve colors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot framed as a static, low-angle, medium-long shot?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.2255014032125473 + ], + "opposite_logprob_tokens": [ + -1.6005014181137085 + ], + "logprob_reward": -0.2255014032125473, + "yes_no_margin_reward": 1.375, + "token_credit_reward": -0.2255014032125473, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2306510955095291, + 0.10958506911993027, + 0.10176284611225128, + 0.0989709347486496, + 0.09560216963291168, + 0.09217635542154312, + 0.08411574363708496, + 0.04065919294953346, + 0.056163955479860306, + 0.09031261503696442 + ], + "credit_map": [ + 0.02868344821035862, + 0.02859172783792019, + 0.027815723791718483, + 0.039655815809965134, + 0.0278561282902956, + 0.02378198131918907, + 0.021894259378314018, + 0.03237200900912285, + 0.01063714548945427, + 0.015308257192373276, + 0.023013930767774582, + 0.012121401727199554, + 0.016170935705304146, + 0.008069337345659733, + 0.015620364807546139, + 0.008643695153295994, + 0.010112321935594082, + 0.01202979777008295, + 0.01627267152070999, + 0.013506166636943817, + 0.012934845872223377, + 0.007242968771606684, + 0.015907010063529015, + 0.013757064007222652, + 0.010611966252326965, + 0.013887776993215084, + 0.01855652406811714, + 0.013298538513481617, + 0.01823100633919239, + 0.008514108136296272, + 0.009377894923090935, + 0.006493120454251766, + 0.010724768973886967, + 0.010116159915924072, + 0.008538298308849335, + 0.009950387291610241, + 0.016741031780838966, + 0.017355622723698616, + 0.013523166067898273, + 0.00865273829549551, + 0.009539089165627956, + 0.0104330163449049, + 0.009185169823467731, + 0.011471951380372047, + 0.015221681445837021, + 0.010670355521142483, + 0.017686877399683, + 0.007968220859766006, + 0.013122071512043476, + 0.010012734681367874, + 0.012911329977214336, + 0.010659214109182358, + 0.00846407376229763, + 0.009467719122767448, + 0.007744158152490854, + 0.011734439060091972, + 0.007239922881126404, + 0.004795602057129145, + 0.0070218974724411964, + 0.004960377234965563, + 0.005239354912191629, + 0.003415387123823166, + 0.0042196596041321754, + 0.0037669946905225515, + 0.007796685677021742, + 0.006794092245399952, + 0.008141439408063889, + 0.010008113458752632, + 0.006716085132211447, + 0.005251356400549412, + 0.0063666230998933315, + 0.005089559592306614, + 0.018702201545238495, + 0.011447189375758171, + 0.012685289606451988, + 0.011085213162004948, + 0.008835031650960445, + 0.008382121101021767, + 0.010153901763260365, + 0.009021664969623089 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 2 + ], + "selected_reward": -0.2255014032125473 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000101/group_5/index_tensor([599, 599])_group_5_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000101/group_5/index_tensor([599, 599])_group_5_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..7535a000bc7e1b5796da73c04c76e7e4554c2c4f --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000101/group_5/index_tensor([599, 599])_group_5_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "Shot in a surreal documentary style, the drone footage begins by gliding low over a sun-drenched expanse of ocean, its azure waters shimmering with countless diamond-like reflections beneath a brilliant, cloudless sky. The camera smoothly pans and rises, revealing a futuristic research vessel effortlessly traversing the gentle waves, leaving behind a clean, white wake. The vessel's trimaran hull, constructed of a sleek pearlescent white composite material, gleams in the high-angle sunlight. Its deck is a marvel of functional design, featuring integrated deep blue solar panels, multiple sensor domes, and a central communications array, all arranged with aerodynamic precision. As the camera continues to sweep around the vessel, complex luminous holographic overlays materialize in the air around it, projected along the hull from nearly invisible emitters. These translucent projections, emitting cool cyan and soft green light, visualize real-time environmental data: elegantly swirling lines and arrows depict the ever-changing wind patterns above the water, while color-graded streams flowing along the hull chart the complex ocean currents below. The data updates dynamically, creating a living, breathing information bubble around the ship, establishing an atmosphere of serene, advanced scientific exploration, captured with the crisp, objective clarity of a high-definition documentary, making the futuristic scenes feel tangibly real and awe-inspiring.", + "seed": 167, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000101/group_5/index_599_seed_167_dp_10_rank_20_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 5, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.01592423180118203, + "vlm_logprob": -0.01592423180118203, + "vlm_yes_no_margin": 4.375, + "vlm_token_credit": -0.01592423180118203, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a futuristic research vessel with a trimaran hull?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the vessel's hull made of a sleek pearlescent white composite material?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are deep blue solar panels, sensor domes, and a central communications array visible on the deck?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do translucent holographic overlays with cyan and green light visualize wind patterns and ocean currents around the vessel?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera smoothly pan and rise while gliding over the sun-drenched ocean?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02858089655637741 + ], + "opposite_logprob_tokens": [ + -4.153580665588379 + ], + "logprob_reward": -0.02858089655637741, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.02858089655637741, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12687407433986664, + 0.0832328349351883, + 0.06162838637828827, + 0.0745791345834732, + 0.04236283525824547, + 0.17257031798362732, + 0.1522430181503296, + 0.12815143167972565, + 0.07875369489192963, + 0.07960431277751923 + ], + "credit_map": [ + 0.015110482461750507, + 0.012239780277013779, + 0.024428293108940125, + 0.012555418536067009, + 0.013953002169728279, + 0.01920144446194172, + 0.01389708835631609, + 0.015488571487367153, + 0.011142364703118801, + 0.007764517795294523, + 0.009616613388061523, + 0.010795678943395615, + 0.008288645185530186, + 0.008822518400847912, + 0.01443008054047823, + 0.012372414581477642, + 0.00955467950552702, + 0.0072229341603815556, + 0.00746846292167902, + 0.0069189961068332195, + 0.006582516245543957, + 0.007404041942209005, + 0.007317378651350737, + 0.009159375913441181, + 0.008773726411163807, + 0.012576071545481682, + 0.007313296664506197, + 0.006838352885097265, + 0.009017737582325935, + 0.010930673219263554, + 0.010826224461197853, + 0.008303052745759487, + 0.007239423226565123, + 0.005149838048964739, + 0.004841554444283247, + 0.0041647013276815414, + 0.005600413773208857, + 0.005151642486453056, + 0.00453664967790246, + 0.005678611341863871, + 0.024224016815423965, + 0.022172586992383003, + 0.028249619528651237, + 0.01612325757741928, + 0.01206966396421194, + 0.023124871775507927, + 0.029870295897126198, + 0.016735998913645744, + 0.018382767215371132, + 0.019595526158809662, + 0.02006847970187664, + 0.01829908788204193, + 0.01618545688688755, + 0.025494100525975227, + 0.014736865647137165, + 0.019480733200907707, + 0.016094492748379707, + 0.01695672981441021, + 0.026223720982670784, + 0.012319429777562618, + 0.014337987639009953, + 0.014400164596736431, + 0.01196405477821827, + 0.0158548466861248, + 0.009904786944389343, + 0.008971500210464, + 0.016514863818883896, + 0.008229881525039673, + 0.009167809970676899, + 0.006194958463311195, + 0.008791739121079445, + 0.0109781539067626, + 0.006511450745165348, + 0.004068834241479635, + 0.01044230442494154, + 0.004080955404788256, + 0.004866435192525387, + 0.020241141319274902, + 0.004416438285261393, + 0.02497675269842148 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 6, + 7 + ], + "selected_reward": -0.02858089655637741 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a futuristic research vessel with a trimaran hull?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the vessel's hull made of a sleek pearlescent white composite material?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are deep blue solar panels, sensor domes, and a central communications array visible on the deck?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do translucent holographic overlays with cyan and green light visualize wind patterns and ocean currents around the vessel?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera smoothly pan and rise while gliding over the sun-drenched ocean?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.007639596704393625 + ], + "opposite_logprob_tokens": [ + -4.882639408111572 + ], + "logprob_reward": -0.007639596704393625, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.007639596704393625, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12265859544277191, + 0.05851183086633682, + 0.09023167937994003, + 0.18904662132263184, + 0.0649382472038269, + 0.12018623948097229, + 0.16791000962257385, + 0.06004924699664116, + 0.04780502989888191, + 0.07866252213716507 + ], + "credit_map": [ + 0.015475094318389893, + 0.013309668749570847, + 0.018719101324677467, + 0.013350950554013252, + 0.011213866993784904, + 0.018744492903351784, + 0.01801980659365654, + 0.0138256149366498, + 0.007400271948426962, + 0.005564537364989519, + 0.006082502659410238, + 0.008525019511580467, + 0.008306579664349556, + 0.005425156094133854, + 0.008517676033079624, + 0.008690088987350464, + 0.01700194738805294, + 0.009476535953581333, + 0.012848436832427979, + 0.011577274650335312, + 0.010111212730407715, + 0.009009365923702717, + 0.010212857276201248, + 0.009994047693908215, + 0.027917804196476936, + 0.013075590133666992, + 0.0208427757024765, + 0.017841698601841927, + 0.01463273260742426, + 0.04697560518980026, + 0.02404918521642685, + 0.023711223155260086, + 0.01078940462321043, + 0.011223611421883106, + 0.006753278896212578, + 0.0076478044502437115, + 0.005646693520247936, + 0.008692465722560883, + 0.006305505055934191, + 0.007879478856921196, + 0.013760123401880264, + 0.01603008061647415, + 0.022035028785467148, + 0.013386395759880543, + 0.007418514229357243, + 0.016025053337216377, + 0.017211413010954857, + 0.014319625683128834, + 0.02104760892689228, + 0.021467706188559532, + 0.021578796207904816, + 0.023407699540257454, + 0.017826450988650322, + 0.026546038687229156, + 0.01648920401930809, + 0.01954651065170765, + 0.0071479082107543945, + 0.007229700684547424, + 0.013680598698556423, + 0.007230078801512718, + 0.006740862969309092, + 0.004967419430613518, + 0.005489130970090628, + 0.0075635481625795364, + 0.006495798472315073, + 0.0057729678228497505, + 0.008433002047240734, + 0.005471792072057724, + 0.006139097269624472, + 0.005221160128712654, + 0.004572441801428795, + 0.005698769818991423, + 0.00795571506023407, + 0.0047182184644043446, + 0.009010073728859425, + 0.004793382715433836, + 0.004849455785006285, + 0.017084887251257896, + 0.004874777514487505, + 0.025376008823513985 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 6, + 0 + ], + "selected_reward": -0.007639596704393625 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a futuristic research vessel with a trimaran hull?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the vessel's hull made of a sleek pearlescent white composite material?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are deep blue solar panels, sensor domes, and a central communications array visible on the deck?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do translucent holographic overlays with cyan and green light visualize wind patterns and ocean currents around the vessel?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera smoothly pan and rise while gliding over the sun-drenched ocean?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.008640876971185207 + ], + "opposite_logprob_tokens": [ + -4.758640766143799 + ], + "logprob_reward": -0.008640876971185207, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.008640876971185207, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11629726737737656, + 0.22852756083011627, + 0.0723351538181305, + 0.10880797356367111, + 0.07174691557884216, + 0.09038665145635605, + 0.10108572989702225, + 0.03444521874189377, + 0.0635848194360733, + 0.11278270930051804 + ], + "credit_map": [ + 0.011027798987925053, + 0.013342801481485367, + 0.017290011048316956, + 0.013200880028307438, + 0.011052442714571953, + 0.01931184157729149, + 0.013295929878950119, + 0.01777556911110878, + 0.039096713066101074, + 0.020279936492443085, + 0.02406426891684532, + 0.0422288216650486, + 0.026798903942108154, + 0.020492609590291977, + 0.025392882525920868, + 0.030173417180776596, + 0.011183952912688255, + 0.0064451745711266994, + 0.010233442299067974, + 0.012037032283842564, + 0.00610091257840395, + 0.01065625436604023, + 0.00695344153791666, + 0.008724941872060299, + 0.01346785482019186, + 0.026452716439962387, + 0.013302672654390335, + 0.011047185398638248, + 0.01188136450946331, + 0.01318893302232027, + 0.009027202613651752, + 0.010440049692988396, + 0.00908713135868311, + 0.01604851894080639, + 0.0076386756263673306, + 0.008762165904045105, + 0.005694417282938957, + 0.00910928938537836, + 0.006291660480201244, + 0.00911505427211523, + 0.01152582373470068, + 0.01145833358168602, + 0.010192585177719593, + 0.00906823854893446, + 0.006446169689297676, + 0.011069023981690407, + 0.022006329149007797, + 0.00862015038728714, + 0.012002777308225632, + 0.012529715895652771, + 0.013135568238794804, + 0.015365957282483578, + 0.01068777497857809, + 0.01674557477235794, + 0.009285086765885353, + 0.011333277449011803, + 0.004724533762782812, + 0.004152600187808275, + 0.006597894709557295, + 0.004618356470018625, + 0.00328419404104352, + 0.0041803219355642796, + 0.0027728406712412834, + 0.004114476498216391, + 0.008673586882650852, + 0.006662619300186634, + 0.014102086424827576, + 0.006647532340139151, + 0.0070727248676121235, + 0.0047561549581587315, + 0.007025416009128094, + 0.00864469911903143, + 0.009164909832179546, + 0.005617369432002306, + 0.0070450808852910995, + 0.005069618579000235, + 0.006254835985600948, + 0.035065796226263046, + 0.0055830832570791245, + 0.038982026278972626 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 9 + ], + "selected_reward": -0.008640876971185207 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a futuristic research vessel with a trimaran hull?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the vessel's hull made of a sleek pearlescent white composite material?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are deep blue solar panels, sensor domes, and a central communications array visible on the deck?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do translucent holographic overlays with cyan and green light visualize wind patterns and ocean currents around the vessel?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera smoothly pan and rise while gliding over the sun-drenched ocean?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.020575184375047684 + ], + "opposite_logprob_tokens": [ + -3.8955752849578857 + ], + "logprob_reward": -0.020575184375047684, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.020575184375047684, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13479816913604736, + 0.12066672742366791, + 0.07286921888589859, + 0.11033359915018082, + 0.14116767048835754, + 0.08698301017284393, + 0.0609685480594635, + 0.05731065571308136, + 0.04652643948793411, + 0.16837596893310547 + ], + "credit_map": [ + 0.012769353576004505, + 0.013034883886575699, + 0.023771584033966064, + 0.011579470708966255, + 0.012830832973122597, + 0.022978587076067924, + 0.017623405903577805, + 0.02021005190908909, + 0.018982430920004845, + 0.012681279331445694, + 0.014740919694304466, + 0.012547683902084827, + 0.017725499346852303, + 0.014135060831904411, + 0.011772728525102139, + 0.018081119284033775, + 0.014984529465436935, + 0.006820765323936939, + 0.0085787083953619, + 0.0063436017371714115, + 0.01136770099401474, + 0.009503400884568691, + 0.008692706935107708, + 0.006577808875590563, + 0.01178345363587141, + 0.024984033778309822, + 0.013432340696454048, + 0.013797672465443611, + 0.01138283871114254, + 0.01281406544148922, + 0.011310044676065445, + 0.0108291395008564, + 0.02785330079495907, + 0.01753045618534088, + 0.015081231482326984, + 0.012846645899116993, + 0.01805286854505539, + 0.016720019280910492, + 0.01567218452692032, + 0.01741095446050167, + 0.01204968523234129, + 0.010345881804823875, + 0.006940899882465601, + 0.00641619460657239, + 0.005127806216478348, + 0.01959696225821972, + 0.01764734648168087, + 0.008858224377036095, + 0.009185214526951313, + 0.008200453594326973, + 0.008264053612947464, + 0.008957578800618649, + 0.007391569670289755, + 0.006297937594354153, + 0.005961921531707048, + 0.0067098187282681465, + 0.007363447919487953, + 0.007541736122220755, + 0.012165441177785397, + 0.00655230600386858, + 0.006164810620248318, + 0.004221017472445965, + 0.005785690620541573, + 0.0075162071734666824, + 0.006208694539964199, + 0.004317059647291899, + 0.013178063556551933, + 0.00507727125659585, + 0.003688547294586897, + 0.0037945283111184835, + 0.003934139851480722, + 0.00632813386619091, + 0.013062594458460808, + 0.005910131148993969, + 0.009655526839196682, + 0.006713176611810923, + 0.00815560296177864, + 0.055173929780721664, + 0.00796656496822834, + 0.06173843890428543 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 4, + 0 + ], + "selected_reward": -0.020575184375047684 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a futuristic research vessel with a trimaran hull?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the vessel's hull made of a sleek pearlescent white composite material?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are deep blue solar panels, sensor domes, and a central communications array visible on the deck?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do translucent holographic overlays with cyan and green light visualize wind patterns and ocean currents around the vessel?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera smoothly pan and rise while gliding over the sun-drenched ocean?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.014184604398906231 + ], + "opposite_logprob_tokens": [ + -4.264184474945068 + ], + "logprob_reward": -0.014184604398906231, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.014184604398906231, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15634360909461975, + 0.0640537440776825, + 0.12859566509723663, + 0.06139039993286133, + 0.0380646176636219, + 0.10727797448635101, + 0.13031993806362152, + 0.08631183952093124, + 0.12018215656280518, + 0.10746002942323685 + ], + "credit_map": [ + 0.017881784588098526, + 0.01472676731646061, + 0.030958175659179688, + 0.015053740702569485, + 0.014397290535271168, + 0.02707710675895214, + 0.013540560379624367, + 0.022708185017108917, + 0.00900508463382721, + 0.008591610006988049, + 0.0072599234990775585, + 0.007557309232652187, + 0.008639814332127571, + 0.007320048287510872, + 0.006199063267558813, + 0.00948088988661766, + 0.020570609718561172, + 0.011679685674607754, + 0.014639079570770264, + 0.01215542946010828, + 0.01806625910103321, + 0.017646590247750282, + 0.015005061402916908, + 0.018832959234714508, + 0.007385655306279659, + 0.008004488423466682, + 0.007545181550085545, + 0.007238310761749744, + 0.006953007075935602, + 0.010404945351183414, + 0.007319580297917128, + 0.006539227440953255, + 0.0065725017338991165, + 0.005948963109403849, + 0.003883153200149536, + 0.003690546378493309, + 0.004599113017320633, + 0.004675216041505337, + 0.0041193887591362, + 0.004575734026730061, + 0.016513202339410782, + 0.009997411631047726, + 0.013604057021439075, + 0.009075714275240898, + 0.006152797490358353, + 0.02168663963675499, + 0.019638754427433014, + 0.010609389282763004, + 0.015861667692661285, + 0.018902109935879707, + 0.018629956990480423, + 0.018275968730449677, + 0.01624949835240841, + 0.012670965865254402, + 0.014095466583967209, + 0.015634311363101006, + 0.01171592716127634, + 0.012507820501923561, + 0.016457725316286087, + 0.007888819091022015, + 0.010106963105499744, + 0.007968774996697903, + 0.008462955243885517, + 0.011202852241694927, + 0.016293302178382874, + 0.00967853982001543, + 0.038035739213228226, + 0.014332965016365051, + 0.007971533574163914, + 0.0059921396896243095, + 0.010106963105499744, + 0.017770973965525627, + 0.007532984018325806, + 0.00412790710106492, + 0.01290050707757473, + 0.004856539890170097, + 0.00540696969255805, + 0.031179208308458328, + 0.005222593899816275, + 0.0362333282828331 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 6, + 2 + ], + "selected_reward": -0.014184604398906231 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "Shot in a surreal documentary style, the drone footage begins by gliding low over a sun-drenched expanse of ocean, its azure waters shimmering with countless diamond-like reflections beneath a brilliant, cloudless sky. The camera smoothly pans and rises, revealing a futuristic research vessel effortlessly traversing the gentle waves, leaving behind a clean, white wake. The vessel's trimaran hull, constructed of a sleek pearlescent white composite material, gleams in the high-angle sunlight. Its deck is a marvel of functional design, featuring integrated deep blue solar panels, multiple sensor domes, and a central communications array, all arranged with aerodynamic precision. As the camera continues to sweep around the vessel, complex luminous holographic overlays materialize in the air around it, projected along the hull from nearly invisible emitters. These translucent projections, emitting cool cyan and soft green light, visualize real-time environmental data: elegantly swirling lines and arrows depict the ever-changing wind patterns above the water, while color-graded streams flowing along the hull chart the complex ocean currents below. The data updates dynamically, creating a living, breathing information bubble around the ship, establishing an atmosphere of serene, advanced scientific exploration, captured with the crisp, objective clarity of a high-definition documentary, making the futuristic scenes feel tangibly real and awe-inspiring.", + "seed": 168, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000101/group_5/index_599_seed_168_dp_10_rank_20_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 5, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.012843700312077998, + "vlm_logprob": -0.012843700312077998, + "vlm_yes_no_margin": 4.624999952316284, + "vlm_token_credit": -0.012843700312077998, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a futuristic research vessel with a trimaran hull?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the vessel's hull made of a sleek pearlescent white composite material?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are deep blue solar panels, sensor domes, and a central communications array visible on the deck?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do translucent holographic overlays with cyan and green light visualize wind patterns and ocean currents around the vessel?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera smoothly pan and rise while gliding over the sun-drenched ocean?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02230101078748703 + ], + "opposite_logprob_tokens": [ + -4.397301197052002 + ], + "logprob_reward": -0.02230101078748703, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.02230101078748703, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.03995296731591225, + 0.02102106809616089, + 0.059040069580078125, + 0.03787723928689957, + 0.7192749381065369, + 0.06326141953468323, + 0.01626230962574482, + 0.026314683258533478, + 0.005925286095589399, + 0.011070086620748043 + ], + "credit_map": [ + 0.0037855945993214846, + 0.008284232579171658, + 0.003937126137316227, + 0.003891447326168418, + 0.0046635945327579975, + 0.004964534193277359, + 0.004563914146274328, + 0.0058625200763344765, + 0.002767250407487154, + 0.0033745020627975464, + 0.0015810796758159995, + 0.002527205739170313, + 0.0026252444367855787, + 0.003119939938187599, + 0.0020250144880264997, + 0.003000829368829727, + 0.008011216297745705, + 0.009966578334569931, + 0.007428486365824938, + 0.005141568370163441, + 0.00648328335955739, + 0.006005681119859219, + 0.005639036186039448, + 0.010364213027060032, + 0.007469064090400934, + 0.005780939012765884, + 0.0035222135484218597, + 0.002799408743157983, + 0.004125799983739853, + 0.009441482834517956, + 0.0017848756397143006, + 0.0029534529894590378, + 0.22128655016422272, + 0.03408446162939072, + 0.02046845108270645, + 0.010291452519595623, + 0.06434096395969391, + 0.33570775389671326, + 0.011094923131167889, + 0.022000353783369064, + 0.010035237297415733, + 0.009829330258071423, + 0.006567963398993015, + 0.006417703814804554, + 0.005068149417638779, + 0.009301051497459412, + 0.007914955727756023, + 0.008127019740641117, + 0.002866371301934123, + 0.0030675167217850685, + 0.002219861140474677, + 0.0013235934311524034, + 0.0018609247636049986, + 0.0016373780090361834, + 0.001893252250738442, + 0.0013934094458818436, + 0.004140639211982489, + 0.004237450193613768, + 0.00275319442152977, + 0.00168970029335469, + 0.0023704618215560913, + 0.0046791089698672295, + 0.002141161821782589, + 0.004302964545786381, + 0.0011014629853889346, + 0.000869161041919142, + 0.0006634027231484652, + 0.0005058337119407952, + 0.0006409367779269814, + 0.0008372629527002573, + 0.0004950715228915215, + 0.0008121534483507276, + 0.0016081988578662276, + 0.0012580911861732602, + 0.0019038371974602342, + 0.0013417191803455353, + 0.0013110943837091327, + 0.0014946219744160771, + 0.000991886598058045, + 0.0011606357293203473 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 5, + 2 + ], + "selected_reward": -0.02230101078748703 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a futuristic research vessel with a trimaran hull?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the vessel's hull made of a sleek pearlescent white composite material?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are deep blue solar panels, sensor domes, and a central communications array visible on the deck?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do translucent holographic overlays with cyan and green light visualize wind patterns and ocean currents around the vessel?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera smoothly pan and rise while gliding over the sun-drenched ocean?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0052629257552325726 + ], + "opposite_logprob_tokens": [ + -5.255262851715088 + ], + "logprob_reward": -0.0052629257552325726, + "yes_no_margin_reward": 5.25, + "token_credit_reward": -0.0052629257552325726, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.04886004701256752, + 0.09736756235361099, + 0.03656382858753204, + 0.03526177257299423, + 0.42118358612060547, + 0.2137034386396408, + 0.026611395180225372, + 0.06774476170539856, + 0.026130395010113716, + 0.026573237031698227 + ], + "credit_map": [ + 0.006038788240402937, + 0.00828785914927721, + 0.0048409211449325085, + 0.004525281488895416, + 0.005793729331344366, + 0.007119420450180769, + 0.004338713362812996, + 0.007915333844721317, + 0.014915959909558296, + 0.011470490135252476, + 0.008648914285004139, + 0.011837059631943703, + 0.012002282775938511, + 0.016161907464265823, + 0.00972608383744955, + 0.012604857794940472, + 0.004516962915658951, + 0.004231486469507217, + 0.005487494636327028, + 0.003542295889928937, + 0.00454604160040617, + 0.005180947482585907, + 0.003794216550886631, + 0.005264379549771547, + 0.004789623431861401, + 0.006342642940580845, + 0.004041342530399561, + 0.0038374229334294796, + 0.005672987084835768, + 0.005194429308176041, + 0.002436757320538163, + 0.002946564694866538, + 0.12963391840457916, + 0.02030271850526333, + 0.012078811414539814, + 0.007749101147055626, + 0.0363445058465004, + 0.19480274617671967, + 0.007643837947398424, + 0.012627911753952503, + 0.03461166471242905, + 0.03103204257786274, + 0.023366782814264297, + 0.02165866270661354, + 0.01832006499171257, + 0.03272112458944321, + 0.023298032581806183, + 0.02869504503905773, + 0.006092381663620472, + 0.006422603968530893, + 0.0033153872936964035, + 0.0015497133135795593, + 0.002236347645521164, + 0.0029961243271827698, + 0.0017984252190217376, + 0.0022004093043506145, + 0.011260122060775757, + 0.01218204665929079, + 0.007088295184075832, + 0.004743579775094986, + 0.00529242679476738, + 0.011660567484796047, + 0.006042498629540205, + 0.009475219063460827, + 0.004599353764206171, + 0.002943673636764288, + 0.003828623564913869, + 0.002224675612524152, + 0.002554659964516759, + 0.004101343918591738, + 0.002222444163635373, + 0.00365561805665493, + 0.004167736973613501, + 0.004997926764190197, + 0.002925083739683032, + 0.0030904507730156183, + 0.0028640804812312126, + 0.0039079561829566956, + 0.0016297252150252461, + 0.0029902756214141846 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 5, + 1 + ], + "selected_reward": -0.0052629257552325726 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a futuristic research vessel with a trimaran hull?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the vessel's hull made of a sleek pearlescent white composite material?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are deep blue solar panels, sensor domes, and a central communications array visible on the deck?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do translucent holographic overlays with cyan and green light visualize wind patterns and ocean currents around the vessel?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera smoothly pan and rise while gliding over the sun-drenched ocean?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005947511177510023 + ], + "opposite_logprob_tokens": [ + -5.130947589874268 + ], + "logprob_reward": -0.005947511177510023, + "yes_no_margin_reward": 5.125, + "token_credit_reward": -0.005947511177510023, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.04598749056458473, + 0.02613372914493084, + 0.02731442265212536, + 0.02408652938902378, + 0.7136455178260803, + 0.10074854642152786, + 0.013325712643563747, + 0.01856130175292492, + 0.01264185830950737, + 0.01755487360060215 + ], + "credit_map": [ + 0.004238941241055727, + 0.0076237148605287075, + 0.003549614455550909, + 0.004767471458762884, + 0.003923194482922554, + 0.005622187629342079, + 0.007029279600828886, + 0.009233084507286549, + 0.003674414474517107, + 0.0029641143046319485, + 0.00341876526363194, + 0.002956512151286006, + 0.00313763739541173, + 0.004309485200792551, + 0.002605798887088895, + 0.003067001234740019, + 0.004031457472592592, + 0.004944839049130678, + 0.003768616123124957, + 0.0033221798948943615, + 0.0029465577099472284, + 0.0026211075019091368, + 0.002487566787749529, + 0.003192097879946232, + 0.0032656185794621706, + 0.005176203325390816, + 0.0032613954972475767, + 0.002764910226687789, + 0.002144494093954563, + 0.0038151878397911787, + 0.001591492909938097, + 0.00206722691655159, + 0.220478817820549, + 0.03383094444870949, + 0.02023652195930481, + 0.010842193849384785, + 0.06296703964471817, + 0.3325408101081848, + 0.01120738685131073, + 0.02154184877872467, + 0.015911974012851715, + 0.014407229609787464, + 0.011043902486562729, + 0.010239599272608757, + 0.008592836558818817, + 0.015591533854603767, + 0.011556198820471764, + 0.013405276462435722, + 0.0025177940260618925, + 0.0016959435306489468, + 0.0024200077168643475, + 0.0010034353472292423, + 0.0012204282684251666, + 0.0017745232908055186, + 0.001317276619374752, + 0.0013763051247224212, + 0.002819332992658019, + 0.002926030196249485, + 0.0023923178669065237, + 0.0009825500892475247, + 0.002380815800279379, + 0.0031685743015259504, + 0.0013557468773797154, + 0.002535932930186391, + 0.002535827225074172, + 0.0018218016484752297, + 0.0017888217698782682, + 0.0012245967518538237, + 0.0013343703467398882, + 0.0018590689869597554, + 0.0010026399977505207, + 0.0010747319320216775, + 0.002752127358689904, + 0.0016338302521035075, + 0.003927284851670265, + 0.0018553498666733503, + 0.001554502989165485, + 0.0021064365282654762, + 0.0014506003353744745, + 0.002274740720167756 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 5, + 0 + ], + "selected_reward": -0.005947511177510023 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a futuristic research vessel with a trimaran hull?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the vessel's hull made of a sleek pearlescent white composite material?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are deep blue solar panels, sensor domes, and a central communications array visible on the deck?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do translucent holographic overlays with cyan and green light visualize wind patterns and ocean currents around the vessel?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera smoothly pan and rise while gliding over the sun-drenched ocean?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.018177704885601997 + ], + "opposite_logprob_tokens": [ + -4.018177509307861 + ], + "logprob_reward": -0.018177704885601997, + "yes_no_margin_reward": 3.999999761581421, + "token_credit_reward": -0.018177704885601997, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.05226251110434532, + 0.08095898479223251, + 0.03815946355462074, + 0.052595604211091995, + 0.5375313758850098, + 0.1173519417643547, + 0.034720245748758316, + 0.026093238964676857, + 0.034608144313097, + 0.025718513876199722 + ], + "credit_map": [ + 0.005441466346383095, + 0.006483647972345352, + 0.005961309187114239, + 0.00940791331231594, + 0.005801101680845022, + 0.006718361284583807, + 0.005812579300254583, + 0.006636132951825857, + 0.012686341069638729, + 0.009670792147517204, + 0.007083931937813759, + 0.007814938202500343, + 0.010301548056304455, + 0.013801714405417442, + 0.007796877529472113, + 0.011802831664681435, + 0.006429191213101149, + 0.004872022196650505, + 0.00483696861192584, + 0.0038946890272200108, + 0.004067655652761459, + 0.0048361713998019695, + 0.004039928317070007, + 0.00518283573910594, + 0.011906667612493038, + 0.004756844136863947, + 0.005314169451594353, + 0.0036312672309577465, + 0.0049600861966609955, + 0.015449981205165386, + 0.0027616668958216906, + 0.0038149175234138966, + 0.16599977016448975, + 0.02588477171957493, + 0.01473287120461464, + 0.008792550303041935, + 0.04702026769518852, + 0.24949119985103607, + 0.00900210440158844, + 0.016607826575636864, + 0.02134571596980095, + 0.015474209561944008, + 0.01479274034500122, + 0.011461794376373291, + 0.00935821421444416, + 0.019656425341963768, + 0.010167103260755539, + 0.015095739625394344, + 0.006187036167830229, + 0.006629798095673323, + 0.0044297147542238235, + 0.0029064358677715063, + 0.0036689131520688534, + 0.003550464054569602, + 0.003988288808614016, + 0.0033595887944102287, + 0.003592862980440259, + 0.003828237298876047, + 0.003539707977324724, + 0.0012704397086054087, + 0.0028941233176738024, + 0.0049857827834784985, + 0.0015265378169715405, + 0.004455544054508209, + 0.006086481735110283, + 0.004196218214929104, + 0.004949166905134916, + 0.002736175199970603, + 0.003295720089226961, + 0.005879419390112162, + 0.00312872976064682, + 0.004336229059845209, + 0.0040711634792387486, + 0.0023020762018859386, + 0.00458334106951952, + 0.0025189153384417295, + 0.002799906302243471, + 0.0035978928208351135, + 0.002259735483676195, + 0.0035854815505445004 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 5, + 1 + ], + "selected_reward": -0.018177704885601997 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a futuristic research vessel with a trimaran hull?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the vessel's hull made of a sleek pearlescent white composite material?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are deep blue solar panels, sensor domes, and a central communications array visible on the deck?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do translucent holographic overlays with cyan and green light visualize wind patterns and ocean currents around the vessel?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera smoothly pan and rise while gliding over the sun-drenched ocean?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.012529348954558372 + ], + "opposite_logprob_tokens": [ + -4.387529373168945 + ], + "logprob_reward": -0.012529348954558372, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.012529348954558372, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.025394313037395477, + 0.05083821713924408, + 0.02008606679737568, + 0.017587810754776, + 0.6959959268569946, + 0.059053368866443634, + 0.04703371971845627, + 0.01705317758023739, + 0.0303978081792593, + 0.03655955195426941 + ], + "credit_map": [ + 0.002799508860334754, + 0.004199150018393993, + 0.0018976260907948017, + 0.0020496719516813755, + 0.004649532027542591, + 0.0033733188174664974, + 0.0021513681858778, + 0.0042741382494568825, + 0.006785267498344183, + 0.007244254928082228, + 0.005037716589868069, + 0.0076565807685256, + 0.005404197610914707, + 0.005766589194536209, + 0.0060507492162287235, + 0.006892862264066935, + 0.003033718327060342, + 0.002488906728103757, + 0.003023577155545354, + 0.002149323932826519, + 0.0020823339000344276, + 0.002084315288811922, + 0.0019316395046189427, + 0.0032922527752816677, + 0.0027429962065070868, + 0.002575310878455639, + 0.0023545229341834784, + 0.0023461359087377787, + 0.002330745104700327, + 0.0019922670908272266, + 0.0013977342750877142, + 0.0018480995204299688, + 0.21492020785808563, + 0.03312187269330025, + 0.019457392394542694, + 0.010646856389939785, + 0.06142378970980644, + 0.3246470093727112, + 0.010792291723191738, + 0.020986586809158325, + 0.00704655097797513, + 0.011020133271813393, + 0.006292519625276327, + 0.005711839534342289, + 0.006109563168138266, + 0.007967683486640453, + 0.00633564218878746, + 0.0085694445297122, + 0.009669066406786442, + 0.009834203869104385, + 0.00600976450368762, + 0.003364491043612361, + 0.00474740331992507, + 0.0055150543339550495, + 0.004331749398261309, + 0.0035619926638901234, + 0.0029850397258996964, + 0.0030315318144857883, + 0.0021843237336724997, + 0.0013268538750708103, + 0.0017930438043549657, + 0.0024266382679343224, + 0.0012522150063887239, + 0.0020535329822450876, + 0.007281516678631306, + 0.005161578301340342, + 0.0031538265757262707, + 0.002580352360382676, + 0.002856107661500573, + 0.004051256459206343, + 0.002585788257420063, + 0.002727383514866233, + 0.0051601058803498745, + 0.0032101538963615894, + 0.0045553878881037235, + 0.0030944400932639837, + 0.0038331858813762665, + 0.008429343812167645, + 0.0026694382540881634, + 0.005607500672340393 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 5, + 1 + ], + "selected_reward": -0.012529348954558372 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "Shot in a surreal documentary style, the drone footage begins by gliding low over a sun-drenched expanse of ocean, its azure waters shimmering with countless diamond-like reflections beneath a brilliant, cloudless sky. The camera smoothly pans and rises, revealing a futuristic research vessel effortlessly traversing the gentle waves, leaving behind a clean, white wake. The vessel's trimaran hull, constructed of a sleek pearlescent white composite material, gleams in the high-angle sunlight. Its deck is a marvel of functional design, featuring integrated deep blue solar panels, multiple sensor domes, and a central communications array, all arranged with aerodynamic precision. As the camera continues to sweep around the vessel, complex luminous holographic overlays materialize in the air around it, projected along the hull from nearly invisible emitters. These translucent projections, emitting cool cyan and soft green light, visualize real-time environmental data: elegantly swirling lines and arrows depict the ever-changing wind patterns above the water, while color-graded streams flowing along the hull chart the complex ocean currents below. The data updates dynamically, creating a living, breathing information bubble around the ship, establishing an atmosphere of serene, advanced scientific exploration, captured with the crisp, objective clarity of a high-definition documentary, making the futuristic scenes feel tangibly real and awe-inspiring.", + "seed": 169, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000101/group_5/index_599_seed_169_dp_11_rank_22_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 5, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.01968338470906019, + "vlm_logprob": -0.01968338470906019, + "vlm_yes_no_margin": 4.15, + "vlm_token_credit": -0.01968338470906019, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a futuristic research vessel with a trimaran hull?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the vessel's hull made of a sleek pearlescent white composite material?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are deep blue solar panels, sensor domes, and a central communications array visible on the deck?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do translucent holographic overlays with cyan and green light visualize wind patterns and ocean currents around the vessel?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera smoothly pan and rise while gliding over the sun-drenched ocean?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.038816314190626144 + ], + "opposite_logprob_tokens": [ + -3.663816213607788 + ], + "logprob_reward": -0.038816314190626144, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.038816314190626144, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.21703706681728363, + 0.08498229831457138, + 0.046252019703388214, + 0.08602059632539749, + 0.04771694168448448, + 0.08788206428289413, + 0.08672066777944565, + 0.14462944865226746, + 0.07049510627985, + 0.12826378643512726 + ], + "credit_map": [ + 0.02140263468027115, + 0.047261543571949005, + 0.028978843241930008, + 0.036432716995477676, + 0.016280438750982285, + 0.01946488581597805, + 0.03206875920295715, + 0.015147264115512371, + 0.008563410490751266, + 0.014359698630869389, + 0.010518308728933334, + 0.019221540540456772, + 0.0070605347864329815, + 0.005707893054932356, + 0.00943180825561285, + 0.010119112208485603, + 0.003943906165659428, + 0.0062540746293962, + 0.005734225269407034, + 0.009669349528849125, + 0.005503077059984207, + 0.005524518433958292, + 0.0059984908439219, + 0.003624381497502327, + 0.012104692868888378, + 0.012267203070223331, + 0.0126260407269001, + 0.016217123717069626, + 0.006792426109313965, + 0.009864655323326588, + 0.0072836363688111305, + 0.008864822797477245, + 0.006907545495778322, + 0.006361449137330055, + 0.007370872423052788, + 0.006150457542389631, + 0.004472872242331505, + 0.0064186532981693745, + 0.004075857810676098, + 0.005959234666079283, + 0.011129419319331646, + 0.013985605910420418, + 0.011576822958886623, + 0.009109302423894405, + 0.011613043956458569, + 0.014342286624014378, + 0.006010832265019417, + 0.010114749893546104, + 0.008647882379591465, + 0.01189772505313158, + 0.00969162117689848, + 0.007108853664249182, + 0.0189655851572752, + 0.015639424324035645, + 0.005851114634424448, + 0.0089184595271945, + 0.01950922980904579, + 0.02405146323144436, + 0.00702710309997201, + 0.013341816142201424, + 0.029601000249385834, + 0.032261740416288376, + 0.010547220706939697, + 0.008289874531328678, + 0.014031964354217052, + 0.013559055514633656, + 0.0032615780364722013, + 0.004717898555099964, + 0.018574193120002747, + 0.009209638461470604, + 0.0035282240714877844, + 0.0036125548649579287, + 0.02497028559446335, + 0.02714373543858528, + 0.006153627764433622, + 0.010124310851097107, + 0.017260437831282616, + 0.020311620086431503, + 0.010653544217348099, + 0.011646229773759842 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 7, + 9 + ], + "selected_reward": -0.038816314190626144 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a futuristic research vessel with a trimaran hull?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the vessel's hull made of a sleek pearlescent white composite material?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are deep blue solar panels, sensor domes, and a central communications array visible on the deck?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do translucent holographic overlays with cyan and green light visualize wind patterns and ocean currents around the vessel?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera smoothly pan and rise while gliding over the sun-drenched ocean?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0098017118871212 + ], + "opposite_logprob_tokens": [ + -4.634801864624023 + ], + "logprob_reward": -0.0098017118871212, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.0098017118871212, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3607904314994812, + 0.05374104529619217, + 0.04969984292984009, + 0.04970003664493561, + 0.06231283023953438, + 0.07083713263273239, + 0.056183237582445145, + 0.15249691903591156, + 0.08275402337312698, + 0.06148448586463928 + ], + "credit_map": [ + 0.03733333572745323, + 0.06062446907162666, + 0.04808144271373749, + 0.06691066175699234, + 0.028483852744102478, + 0.029819943010807037, + 0.05998191237449646, + 0.029554802924394608, + 0.00613214960321784, + 0.007593401242047548, + 0.006082905922085047, + 0.01044787373393774, + 0.005923958495259285, + 0.005615826230496168, + 0.006756290793418884, + 0.005188644398003817, + 0.005715395789593458, + 0.009337915107607841, + 0.007150272838771343, + 0.007327314931899309, + 0.005598421674221754, + 0.006168970838189125, + 0.0049754418432712555, + 0.003426117356866598, + 0.0056098164059221745, + 0.00700082303956151, + 0.007572580128908157, + 0.0080637875944376, + 0.006648819427937269, + 0.005939661059528589, + 0.004099354147911072, + 0.004765200428664684, + 0.007632048334926367, + 0.008885095827281475, + 0.007728771306574345, + 0.008078213781118393, + 0.00877936277538538, + 0.008499303832650185, + 0.004070770926773548, + 0.008639267645776272, + 0.009194772690534592, + 0.01286839134991169, + 0.008443622849881649, + 0.005183653440326452, + 0.009408274665474892, + 0.014951558783650398, + 0.004111803602427244, + 0.006675059907138348, + 0.005215506535023451, + 0.007509041111916304, + 0.006173720117658377, + 0.004933158401399851, + 0.015621989034116268, + 0.008899022825062275, + 0.004072957206517458, + 0.003757843980565667, + 0.02138867974281311, + 0.025873593986034393, + 0.008245473727583885, + 0.01591782458126545, + 0.02530609630048275, + 0.03358352929353714, + 0.011620677076280117, + 0.010561070404946804, + 0.011332946829497814, + 0.012676131911575794, + 0.0031386232003569603, + 0.0048501272685825825, + 0.024659734219312668, + 0.011629131622612476, + 0.008086571469902992, + 0.006380758713930845, + 0.011717082001268864, + 0.014543557539582253, + 0.0027416690718382597, + 0.003675976302474737, + 0.01048045139759779, + 0.010267051868140697, + 0.003929208964109421, + 0.004129493143409491 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 7, + 8 + ], + "selected_reward": -0.0098017118871212 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a futuristic research vessel with a trimaran hull?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the vessel's hull made of a sleek pearlescent white composite material?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are deep blue solar panels, sensor domes, and a central communications array visible on the deck?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do translucent holographic overlays with cyan and green light visualize wind patterns and ocean currents around the vessel?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera smoothly pan and rise while gliding over the sun-drenched ocean?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.008646194823086262 + ], + "opposite_logprob_tokens": [ + -4.758646011352539 + ], + "logprob_reward": -0.008646194823086262, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.008646194823086262, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13900047540664673, + 0.07278929650783539, + 0.03985774517059326, + 0.0557328537106514, + 0.06439708173274994, + 0.07574785500764847, + 0.09123915433883667, + 0.17100563645362854, + 0.19731959700584412, + 0.0929102823138237 + ], + "credit_map": [ + 0.017881954088807106, + 0.023695113137364388, + 0.017055943608283997, + 0.02584538422524929, + 0.01115830335766077, + 0.012105495668947697, + 0.016108494251966476, + 0.015149783343076706, + 0.006280453409999609, + 0.013153942301869392, + 0.009116041474044323, + 0.012794638983905315, + 0.00821545161306858, + 0.008839216083288193, + 0.006645041052252054, + 0.0077445111237466335, + 0.004618196748197079, + 0.008701445534825325, + 0.0055646332912147045, + 0.005622383672744036, + 0.004555969964712858, + 0.0056501650251448154, + 0.002607817528769374, + 0.002537132939323783, + 0.005890964064747095, + 0.00741229671984911, + 0.0087219113484025, + 0.006480325944721699, + 0.009992971085011959, + 0.0062418775632977486, + 0.004755023401230574, + 0.006237487308681011, + 0.008520086295902729, + 0.009094665758311749, + 0.01109171099960804, + 0.01161003764718771, + 0.005281047895550728, + 0.006197646725922823, + 0.005024629645049572, + 0.00757725490257144, + 0.011608434841036797, + 0.01567133702337742, + 0.007207620423287153, + 0.005646670702844858, + 0.008793444372713566, + 0.015378488227725029, + 0.004292511846870184, + 0.007149347569793463, + 0.011493923142552376, + 0.01619185321033001, + 0.007593129761517048, + 0.009179756976664066, + 0.021488942205905914, + 0.012605883181095123, + 0.006534170359373093, + 0.0061514959670603275, + 0.024468299001455307, + 0.027554472908377647, + 0.00882299616932869, + 0.017500251531600952, + 0.029225805774331093, + 0.038516171276569366, + 0.012931414879858494, + 0.011986234225332737, + 0.04118228331208229, + 0.02971290796995163, + 0.007571266498416662, + 0.008888043463230133, + 0.04633352532982826, + 0.03339214622974396, + 0.014088848605751991, + 0.016150571405887604, + 0.01753844879567623, + 0.021879049018025398, + 0.004559405148029327, + 0.0066227661445736885, + 0.013980340212583542, + 0.01510537602007389, + 0.00735966581851244, + 0.0058652241714298725 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 7, + 0 + ], + "selected_reward": -0.008646194823086262 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a futuristic research vessel with a trimaran hull?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the vessel's hull made of a sleek pearlescent white composite material?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are deep blue solar panels, sensor domes, and a central communications array visible on the deck?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do translucent holographic overlays with cyan and green light visualize wind patterns and ocean currents around the vessel?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera smoothly pan and rise while gliding over the sun-drenched ocean?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.020579971373081207 + ], + "opposite_logprob_tokens": [ + -3.8955800533294678 + ], + "logprob_reward": -0.020579971373081207, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.020579971373081207, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.26570236682891846, + 0.06164927780628204, + 0.05900407209992409, + 0.03853997960686684, + 0.044374581426382065, + 0.047863371670246124, + 0.08952383697032928, + 0.16991302371025085, + 0.12332700192928314, + 0.10010246187448502 + ], + "credit_map": [ + 0.024309320375323296, + 0.05309975892305374, + 0.03666337579488754, + 0.047095879912376404, + 0.018684199079871178, + 0.02267361618578434, + 0.044386979192495346, + 0.018789228051900864, + 0.0056155710481107235, + 0.011322039179503918, + 0.007531507406383753, + 0.009998438879847527, + 0.006614824756979942, + 0.005217427387833595, + 0.007151960860937834, + 0.00819750502705574, + 0.004958931356668472, + 0.008723348379135132, + 0.0076218461617827415, + 0.008622078225016594, + 0.008292450569570065, + 0.007390966638922691, + 0.008552723564207554, + 0.0048417262732982635, + 0.0062775397673249245, + 0.004633765202015638, + 0.005770544987171888, + 0.005423457361757755, + 0.0057912301272153854, + 0.0043974630534648895, + 0.0029691243544220924, + 0.0032768503297120333, + 0.005722483620047569, + 0.005746858660131693, + 0.006719592027366161, + 0.007997997105121613, + 0.003944569267332554, + 0.0049153538420796394, + 0.003631911240518093, + 0.00569581426680088, + 0.008531572297215462, + 0.009407639503479004, + 0.005789947230368853, + 0.004436414688825607, + 0.004598313011229038, + 0.00791311264038086, + 0.003334302455186844, + 0.0038520677480846643, + 0.008323241025209427, + 0.012350856326520443, + 0.00873508956283331, + 0.0077000753954052925, + 0.022119488567113876, + 0.014550220221281052, + 0.006941039580851793, + 0.008803832344710827, + 0.025219347327947617, + 0.029170429334044456, + 0.008511737920343876, + 0.016691965982317924, + 0.029894430190324783, + 0.035200245678424835, + 0.013469713740050793, + 0.011755152605473995, + 0.021311823278665543, + 0.025272585451602936, + 0.0038750816602259874, + 0.006338959559798241, + 0.032974135130643845, + 0.017267651855945587, + 0.008553125895559788, + 0.0077336449176073074, + 0.0207086019217968, + 0.0216529443860054, + 0.004154070280492306, + 0.007310133893042803, + 0.016018448397517204, + 0.01655193231999874, + 0.006402822211384773, + 0.007303514517843723 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 7, + 8 + ], + "selected_reward": -0.020579971373081207 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a futuristic research vessel with a trimaran hull?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the vessel's hull made of a sleek pearlescent white composite material?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are deep blue solar panels, sensor domes, and a central communications array visible on the deck?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do translucent holographic overlays with cyan and green light visualize wind patterns and ocean currents around the vessel?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera smoothly pan and rise while gliding over the sun-drenched ocean?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.020572731271386147 + ], + "opposite_logprob_tokens": [ + -3.8955726623535156 + ], + "logprob_reward": -0.020572731271386147, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.020572731271386147, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2832884192466736, + 0.06649136543273926, + 0.03680388629436493, + 0.05772116407752037, + 0.07501066476106644, + 0.11855142563581467, + 0.16907815635204315, + 0.06537624448537827, + 0.0763281062245369, + 0.051350537687540054 + ], + "credit_map": [ + 0.025967271998524666, + 0.04587160423398018, + 0.04004444554448128, + 0.05099146440625191, + 0.0225520096719265, + 0.024465126916766167, + 0.05086676776409149, + 0.02252974919974804, + 0.0062406668439507484, + 0.01072936225682497, + 0.007790669798851013, + 0.009481985121965408, + 0.00901668332517147, + 0.007679385598748922, + 0.006560603156685829, + 0.008992011658847332, + 0.004384576342999935, + 0.005401983391493559, + 0.004644813947379589, + 0.005013954825699329, + 0.005792465526610613, + 0.0042901411652565, + 0.004999595694243908, + 0.002276358427479863, + 0.008744418621063232, + 0.010366306640207767, + 0.007377766538411379, + 0.006864427123218775, + 0.006316015031188726, + 0.006917909253388643, + 0.003975951578468084, + 0.007158374879509211, + 0.01273167971521616, + 0.011931739747524261, + 0.010093018412590027, + 0.007042123470455408, + 0.007296678144484758, + 0.013568989932537079, + 0.0044351969845592976, + 0.007911242544651031, + 0.017252566292881966, + 0.02063797414302826, + 0.01468830555677414, + 0.010023966431617737, + 0.013977416791021824, + 0.023699361830949783, + 0.008149686269462109, + 0.01012216042727232, + 0.019525300711393356, + 0.028652729466557503, + 0.023763438686728477, + 0.018794452771544456, + 0.02661488950252533, + 0.025995008647441864, + 0.010389582253992558, + 0.015342747792601585, + 0.008517426438629627, + 0.009401834569871426, + 0.003208905691280961, + 0.00600868696346879, + 0.012284019961953163, + 0.015557928010821342, + 0.005305383820086718, + 0.005092061590403318, + 0.012868944555521011, + 0.010251063853502274, + 0.0033897662069648504, + 0.004996319767087698, + 0.014349904842674732, + 0.011433444917201996, + 0.007393460255116224, + 0.011645205318927765, + 0.008835915476083755, + 0.010843564756214619, + 0.0026267576031386852, + 0.00375881209038198, + 0.007239685393869877, + 0.008474900387227535, + 0.0051262471824884415, + 0.004444657359272242 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 6, + 5 + ], + "selected_reward": -0.020572731271386147 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "Shot in a surreal documentary style, the drone footage begins by gliding low over a sun-drenched expanse of ocean, its azure waters shimmering with countless diamond-like reflections beneath a brilliant, cloudless sky. The camera smoothly pans and rises, revealing a futuristic research vessel effortlessly traversing the gentle waves, leaving behind a clean, white wake. The vessel's trimaran hull, constructed of a sleek pearlescent white composite material, gleams in the high-angle sunlight. Its deck is a marvel of functional design, featuring integrated deep blue solar panels, multiple sensor domes, and a central communications array, all arranged with aerodynamic precision. As the camera continues to sweep around the vessel, complex luminous holographic overlays materialize in the air around it, projected along the hull from nearly invisible emitters. These translucent projections, emitting cool cyan and soft green light, visualize real-time environmental data: elegantly swirling lines and arrows depict the ever-changing wind patterns above the water, while color-graded streams flowing along the hull chart the complex ocean currents below. The data updates dynamically, creating a living, breathing information bubble around the ship, establishing an atmosphere of serene, advanced scientific exploration, captured with the crisp, objective clarity of a high-definition documentary, making the futuristic scenes feel tangibly real and awe-inspiring.", + "seed": 170, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000101/group_5/index_599_seed_170_dp_11_rank_22_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 5, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.018681892659515142, + "vlm_logprob": -0.018681892659515142, + "vlm_yes_no_margin": 4.375, + "vlm_token_credit": -0.018681892659515142, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a futuristic research vessel with a trimaran hull?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the vessel's hull made of a sleek pearlescent white composite material?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are deep blue solar panels, sensor domes, and a central communications array visible on the deck?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do translucent holographic overlays with cyan and green light visualize wind patterns and ocean currents around the vessel?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera smoothly pan and rise while gliding over the sun-drenched ocean?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.055348802357912064 + ], + "opposite_logprob_tokens": [ + -3.0553488731384277 + ], + "logprob_reward": -0.055348802357912064, + "yes_no_margin_reward": 3.0, + "token_credit_reward": -0.055348802357912064, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.16539594531059265, + 0.137144535779953, + 0.06519550830125809, + 0.18760378658771515, + 0.0843377485871315, + 0.09052862226963043, + 0.04551253095269203, + 0.046486347913742065, + 0.0854777842760086, + 0.09231726825237274 + ], + "credit_map": [ + 0.017419327050447464, + 0.026281923055648804, + 0.012324891053140163, + 0.018997224047780037, + 0.03169143572449684, + 0.02303193509578705, + 0.015515181235969067, + 0.020134031772613525, + 0.019507532939314842, + 0.018645813688635826, + 0.015141697600483894, + 0.015912991017103195, + 0.01692904159426689, + 0.01603754796087742, + 0.021605098620057106, + 0.013364803977310658, + 0.007251092232763767, + 0.01279444433748722, + 0.005504297092556953, + 0.008504612371325493, + 0.009138275869190693, + 0.009475614875555038, + 0.007299317978322506, + 0.005227854009717703, + 0.018492145463824272, + 0.029218288138508797, + 0.01551225408911705, + 0.022311916574835777, + 0.02814616449177265, + 0.025243762880563736, + 0.02854332886636257, + 0.020135903730988503, + 0.010152057744562626, + 0.01198911014944315, + 0.014070682227611542, + 0.01076322142034769, + 0.008899454958736897, + 0.011985275894403458, + 0.007571832276880741, + 0.008906122297048569, + 0.009792855940759182, + 0.013485020957887173, + 0.01347760483622551, + 0.010453353635966778, + 0.008970634080469608, + 0.013928456231951714, + 0.007460955064743757, + 0.0129597382619977, + 0.00417927885428071, + 0.004440587479621172, + 0.006119340192526579, + 0.005803839303553104, + 0.005092126317322254, + 0.006613990291953087, + 0.00672884052619338, + 0.006534531246870756, + 0.0059047299437224865, + 0.005415778607130051, + 0.007356226444244385, + 0.007291872054338455, + 0.004634816199541092, + 0.0071332682855427265, + 0.0031301251146942377, + 0.005619530566036701, + 0.010204064659774303, + 0.00924637634307146, + 0.022412894293665886, + 0.010116307064890862, + 0.006573581602424383, + 0.014301575720310211, + 0.005347710568457842, + 0.007275281473994255, + 0.010001071728765965, + 0.009401332587003708, + 0.017330171540379524, + 0.0060948715545237064, + 0.019390342757105827, + 0.013690310530364513, + 0.007952410727739334, + 0.008456755429506302 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 0, + 1 + ], + "selected_reward": -0.055348802357912064 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a futuristic research vessel with a trimaran hull?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the vessel's hull made of a sleek pearlescent white composite material?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are deep blue solar panels, sensor domes, and a central communications array visible on the deck?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do translucent holographic overlays with cyan and green light visualize wind patterns and ocean currents around the vessel?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera smoothly pan and rise while gliding over the sun-drenched ocean?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.006746135652065277 + ], + "opposite_logprob_tokens": [ + -5.006746292114258 + ], + "logprob_reward": -0.006746135652065277, + "yes_no_margin_reward": 5.0, + "token_credit_reward": -0.006746135652065277, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14120520651340485, + 0.08873825520277023, + 0.11759994924068451, + 0.09194975346326828, + 0.20424604415893555, + 0.07741215080022812, + 0.0643935352563858, + 0.04414663463830948, + 0.06960690021514893, + 0.10070148855447769 + ], + "credit_map": [ + 0.015726270154118538, + 0.02076050266623497, + 0.013413279317319393, + 0.025069165974855423, + 0.018814420327544212, + 0.015006585977971554, + 0.013935980387032032, + 0.018479011952877045, + 0.01163643877953291, + 0.010845179669559002, + 0.01090036891400814, + 0.011002827435731888, + 0.010885220021009445, + 0.010535312816500664, + 0.013878632336854935, + 0.009054280817508698, + 0.011659353040158749, + 0.014949692413210869, + 0.020126042887568474, + 0.013621090911328793, + 0.013994753360748291, + 0.014911781996488571, + 0.01345934346318245, + 0.01487789861857891, + 0.0076207444071769714, + 0.012689994648098946, + 0.012172880582511425, + 0.012285491451621056, + 0.013814502395689487, + 0.012637631967663765, + 0.010742580518126488, + 0.009985940530896187, + 0.027232399210333824, + 0.0347624272108078, + 0.023518027737736702, + 0.02820216678082943, + 0.0231467317789793, + 0.028523588553071022, + 0.02082333154976368, + 0.01803738623857498, + 0.008757264353334904, + 0.01388191431760788, + 0.00748971663415432, + 0.010176652111113071, + 0.009738537482917309, + 0.011547040194272995, + 0.006654105149209499, + 0.00916692242026329, + 0.007563971448689699, + 0.00658992538228631, + 0.011272530071437359, + 0.006637634709477425, + 0.005687651690095663, + 0.009502878412604332, + 0.006769794505089521, + 0.010369155555963516, + 0.008550534956157207, + 0.00594041682779789, + 0.004544807132333517, + 0.0051658861339092255, + 0.004445632454007864, + 0.0070741428062319756, + 0.003766475012525916, + 0.004658740479499102, + 0.01071323361247778, + 0.008325176313519478, + 0.015241211280226707, + 0.006562864873558283, + 0.006283754017204046, + 0.011196883395314217, + 0.004696649499237537, + 0.006587136071175337, + 0.020205985754728317, + 0.00958721898496151, + 0.023449858650565147, + 0.006379119586199522, + 0.009784998372197151, + 0.01563795655965805, + 0.006926003377884626, + 0.008730360306799412 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 0, + 2 + ], + "selected_reward": -0.006746135652065277 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a futuristic research vessel with a trimaran hull?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the vessel's hull made of a sleek pearlescent white composite material?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are deep blue solar panels, sensor domes, and a central communications array visible on the deck?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do translucent holographic overlays with cyan and green light visualize wind patterns and ocean currents around the vessel?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera smoothly pan and rise while gliding over the sun-drenched ocean?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.007631551939994097 + ], + "opposite_logprob_tokens": [ + -4.882631778717041 + ], + "logprob_reward": -0.007631551939994097, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.007631551939994097, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09818000346422195, + 0.08928364515304565, + 0.05955392122268677, + 0.11822240799665451, + 0.20045675337314606, + 0.1535293459892273, + 0.060403112322092056, + 0.05462837591767311, + 0.07876218110322952, + 0.08698026835918427 + ], + "credit_map": [ + 0.014226057566702366, + 0.017468644306063652, + 0.007118797395378351, + 0.009429387748241425, + 0.01796678639948368, + 0.012039782479405403, + 0.008965298533439636, + 0.010965251363813877, + 0.011380499228835106, + 0.012236693874001503, + 0.009397270157933235, + 0.010757412761449814, + 0.012583741918206215, + 0.010355889797210693, + 0.01160276960581541, + 0.010969370603561401, + 0.006934343837201595, + 0.01132410392165184, + 0.008863547816872597, + 0.007403127383440733, + 0.007194416131824255, + 0.008673125877976418, + 0.003859757911413908, + 0.005301500670611858, + 0.010501966811716557, + 0.022877035662531853, + 0.009393541142344475, + 0.013720812276005745, + 0.012517609633505344, + 0.015935376286506653, + 0.01943274587392807, + 0.013843326829373837, + 0.02204112708568573, + 0.03357569873332977, + 0.02682955004274845, + 0.02781703881919384, + 0.02289409004151821, + 0.02713577263057232, + 0.018838180229067802, + 0.021325286477804184, + 0.018095580860972404, + 0.024176327511668205, + 0.02190849371254444, + 0.01667260378599167, + 0.015095934271812439, + 0.023149708285927773, + 0.011997349560260773, + 0.022433344274759293, + 0.006856240332126617, + 0.006828130688518286, + 0.01150253601372242, + 0.007005652412772179, + 0.006109942682087421, + 0.008421715348958969, + 0.0052336836233735085, + 0.008445207960903645, + 0.008555217646062374, + 0.006259177811443806, + 0.006641060579568148, + 0.007569173350930214, + 0.006993373855948448, + 0.008745400235056877, + 0.0034862752072513103, + 0.0063786981627345085, + 0.013889221474528313, + 0.008409720845520496, + 0.018151860684156418, + 0.006627470720559359, + 0.0075275348499417305, + 0.012962348759174347, + 0.004298198968172073, + 0.006895827129483223, + 0.01938563771545887, + 0.009107994846999645, + 0.018563659861683846, + 0.005695185158401728, + 0.010519139468669891, + 0.013186385855078697, + 0.004948093555867672, + 0.005574167240411043 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 5, + 3 + ], + "selected_reward": -0.007631551939994097 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a futuristic research vessel with a trimaran hull?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the vessel's hull made of a sleek pearlescent white composite material?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are deep blue solar panels, sensor domes, and a central communications array visible on the deck?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do translucent holographic overlays with cyan and green light visualize wind patterns and ocean currents around the vessel?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera smoothly pan and rise while gliding over the sun-drenched ocean?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.007625163532793522 + ], + "opposite_logprob_tokens": [ + -4.882625102996826 + ], + "logprob_reward": -0.007625163532793522, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.007625163532793522, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.16009917855262756, + 0.07378027588129044, + 0.05522705614566803, + 0.051820989698171616, + 0.11832768470048904, + 0.12227266281843185, + 0.08695200085639954, + 0.09435193985700607, + 0.13013076782226562, + 0.10703745484352112 + ], + "credit_map": [ + 0.016061225906014442, + 0.023430291563272476, + 0.014141450636088848, + 0.014703039079904556, + 0.03402015566825867, + 0.021772755309939384, + 0.01569809950888157, + 0.02027214504778385, + 0.009190170094370842, + 0.008378748781979084, + 0.010615363717079163, + 0.007548271678388119, + 0.010094930417835712, + 0.008251932449638844, + 0.012010342441499233, + 0.007690516300499439, + 0.005877814255654812, + 0.008028428070247173, + 0.011169422417879105, + 0.006388837471604347, + 0.005177482031285763, + 0.007832765579223633, + 0.005361418705433607, + 0.005390884820371866, + 0.0047239442355930805, + 0.010127285495400429, + 0.004253755323588848, + 0.006165546365082264, + 0.005725331604480743, + 0.007601048797369003, + 0.007399353198707104, + 0.005824715830385685, + 0.01643890142440796, + 0.016135744750499725, + 0.012629728764295578, + 0.02237301878631115, + 0.010957903228700161, + 0.017795942723751068, + 0.012381235137581825, + 0.009615200571715832, + 0.015995614230632782, + 0.018652353435754776, + 0.01727020926773548, + 0.012077168561518192, + 0.010789965279400349, + 0.018616272136569023, + 0.008867805823683739, + 0.020003261044621468, + 0.011855844408273697, + 0.008960933424532413, + 0.016486037522554398, + 0.01198143046349287, + 0.008483835496008396, + 0.012594383209943771, + 0.0063760532066226006, + 0.010213472880423069, + 0.01755402982234955, + 0.011000500060617924, + 0.01274244673550129, + 0.011201396584510803, + 0.00990944355726242, + 0.014823958277702332, + 0.007494635414332151, + 0.009625514037907124, + 0.02235468290746212, + 0.014647861011326313, + 0.029188500717282295, + 0.013519716449081898, + 0.010415741242468357, + 0.022014785557985306, + 0.007372286170721054, + 0.010617166757583618, + 0.025660758838057518, + 0.009496359154582024, + 0.029709672555327415, + 0.008374211378395557, + 0.005859719589352608, + 0.01762489229440689, + 0.004143641330301762, + 0.006168194580823183 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 5 + ], + "selected_reward": -0.007625163532793522 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a futuristic research vessel with a trimaran hull?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the vessel's hull made of a sleek pearlescent white composite material?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are deep blue solar panels, sensor domes, and a central communications array visible on the deck?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do translucent holographic overlays with cyan and green light visualize wind patterns and ocean currents around the vessel?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera smoothly pan and rise while gliding over the sun-drenched ocean?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.016057809814810753 + ], + "opposite_logprob_tokens": [ + -4.141057968139648 + ], + "logprob_reward": -0.016057809814810753, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.016057809814810753, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10721208155155182, + 0.13839532434940338, + 0.057846762239933014, + 0.04901544749736786, + 0.11582028120756149, + 0.09818928688764572, + 0.17299963533878326, + 0.11882468312978745, + 0.0966801792383194, + 0.045016270130872726 + ], + "credit_map": [ + 0.014332545921206474, + 0.01640717126429081, + 0.012129836715757847, + 0.01331255491822958, + 0.010822626762092113, + 0.013172301463782787, + 0.010081629268825054, + 0.016953416168689728, + 0.014811660163104534, + 0.01347900927066803, + 0.018516967073082924, + 0.013578503392636776, + 0.019948367029428482, + 0.016777273267507553, + 0.01926395297050476, + 0.022019581869244576, + 0.007495572790503502, + 0.007928864099085331, + 0.008432635106146336, + 0.007555125746876001, + 0.006447130814194679, + 0.0075670694932341576, + 0.006285350304096937, + 0.006135013420134783, + 0.005413974169641733, + 0.008143083192408085, + 0.007198132108896971, + 0.005234684329479933, + 0.006688075605779886, + 0.007269540801644325, + 0.004331162199378014, + 0.004736793227493763, + 0.01631796546280384, + 0.021416882053017616, + 0.011868583969771862, + 0.018445821478962898, + 0.011830111965537071, + 0.014873355627059937, + 0.01064476277679205, + 0.01042279414832592, + 0.010295418091118336, + 0.01197802647948265, + 0.013604295440018177, + 0.013589815236628056, + 0.00896200630813837, + 0.013650298118591309, + 0.00989375077188015, + 0.01621568202972412, + 0.023867644369602203, + 0.01820238120853901, + 0.034670278429985046, + 0.024022715166211128, + 0.01476387307047844, + 0.02550361305475235, + 0.014249633997678757, + 0.017719494178891182, + 0.014783627353608608, + 0.017733601853251457, + 0.014544537290930748, + 0.01123599149286747, + 0.014663862995803356, + 0.018188390880823135, + 0.010544986464083195, + 0.0171296838670969, + 0.01634979620575905, + 0.01211358979344368, + 0.015654267743229866, + 0.008739231154322624, + 0.011727533303201199, + 0.01366499811410904, + 0.009617330506443977, + 0.008813437074422836, + 0.005752660799771547, + 0.0044939955696463585, + 0.011697419919073582, + 0.0039642066694796085, + 0.003958858549594879, + 0.0067237745970487595, + 0.004427473992109299, + 0.003997878171503544 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 1, + 7 + ], + "selected_reward": -0.016057809814810753 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000101/group_6/index_tensor([7173, 7173])_group_6_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000101/group_6/index_tensor([7173, 7173])_group_6_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..02955d1d42d2683e5c7dd3763d6e1a40c41efd32 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000101/group_6/index_tensor([7173, 7173])_group_6_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "medium-long shot, shot from a slightly upward angle, the camera swaying precariously with the ship's violent pitching, captures an ancient wooden expedition vessel battling mountainous waves. The hull is constructed of weathered, dark wood, its surface covered in dark water stains and salt crystals. Thick, damp ropes lie scattered on the deck. In a night sky thick with dark clouds and thunder, occasional flashes of pale lightning provide the sole source of light, momentarily illuminating the inky black sea roiling with white foam. An explorer, clad in a heavy windbreaker thoroughly soaked by rain and sea spray, his dark, wet hair disheveled by the wind plastered to his cold-white forehead, clings desperately to a swaying railing, his face stern but unconcealed with fear as he tries to steady himself on the mirror-slick, rain-soaked deck. Suddenly, a huge wave hit the boat violently. His fingers that were holding the railing lost their strength due to the wetness and slipperiness. His feet slipped suddenly and his body lost balance instantly. He waved his arms in vain in the air and then fell backward helplessly. He went over the side of the boat and was finally swallowed up by the inky black waves below that were roaring and rolling with white foam. The whole picture was full of despair, chaos and epic tragedy.", + "seed": 22518, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000101/group_6/index_7173_seed_22518_dp_12_rank_24_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 6, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.019865477457642555, + "vlm_logprob": -0.019865477457642555, + "vlm_yes_no_margin": 4.3, + "vlm_token_credit": -0.019865477457642555, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set at night with a sky thick with dark clouds and occasional flashes of pale lightning?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is an ancient wooden expedition vessel visible battling mountainous waves in an inky black sea?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is an explorer visible wearing a heavy, rain-soaked windbreaker with wet hair plastered to his forehead?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the explorer slip from the railing, fall backward, and go over the side of the boat into the waves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a medium-long shot from a slightly upward angle with a camera swaying precariously?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.020967375487089157 + ], + "opposite_logprob_tokens": [ + -5.020967483520508 + ], + "logprob_reward": -0.020967375487089157, + "yes_no_margin_reward": 5.0, + "token_credit_reward": -0.020967375487089157, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11465228348970413, + 0.17482195794582367, + 0.08422522246837616, + 0.07415700703859329, + 0.04430002346634865, + 0.08890623599290848, + 0.04903141036629677, + 0.19339792430400848, + 0.13297231495380402, + 0.043535660952329636 + ], + "credit_map": [ + 0.00919684860855341, + 0.013615939766168594, + 0.019171468913555145, + 0.018016355112195015, + 0.010622255504131317, + 0.015055767260491848, + 0.014540320262312889, + 0.014433330856263638, + 0.03321654722094536, + 0.010871420614421368, + 0.013953995890915394, + 0.015143646858632565, + 0.017074929550290108, + 0.031159760430455208, + 0.04054459184408188, + 0.012857066467404366, + 0.008678464218974113, + 0.011154286563396454, + 0.021047845482826233, + 0.011712110601365566, + 0.007399826776236296, + 0.009828261099755764, + 0.006570393685251474, + 0.007834036834537983, + 0.010251589119434357, + 0.006756512448191643, + 0.010729258880019188, + 0.00873072911053896, + 0.011821920052170753, + 0.007371881045401096, + 0.010018576867878437, + 0.00847653765231371, + 0.003346286714076996, + 0.0049070208333432674, + 0.004145095124840736, + 0.005173536017537117, + 0.010862872935831547, + 0.005459198262542486, + 0.006765902973711491, + 0.0036401122342795134, + 0.007033293601125479, + 0.011633253656327724, + 0.012198634445667267, + 0.008959540165960789, + 0.017524465918540955, + 0.015203669667243958, + 0.011953158304095268, + 0.004400219768285751, + 0.0054132575169205666, + 0.004831991158425808, + 0.007951948791742325, + 0.008532802574336529, + 0.0059564379043877125, + 0.0063599455170333385, + 0.005541164893656969, + 0.004443860612809658, + 0.01222243346273899, + 0.013447807170450687, + 0.0737341046333313, + 0.055556561797857285, + 0.012654549442231655, + 0.009080027230083942, + 0.010851682163774967, + 0.005850765388458967, + 0.008754977956414223, + 0.012872176244854927, + 0.021891357377171516, + 0.04548091068863869, + 0.012639671564102173, + 0.011758235283195972, + 0.012249406427145004, + 0.0073255691677331924, + 0.005588033236563206, + 0.010248664766550064, + 0.006813172250986099, + 0.006393907591700554, + 0.003260948695242405, + 0.003563673235476017, + 0.004266015253961086, + 0.003401246154680848 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 1, + 8 + ], + "selected_reward": -0.020967375487089157 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set at night with a sky thick with dark clouds and occasional flashes of pale lightning?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is an ancient wooden expedition vessel visible battling mountainous waves in an inky black sea?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is an explorer visible wearing a heavy, rain-soaked windbreaker with wet hair plastered to his forehead?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the explorer slip from the railing, fall backward, and go over the side of the boat into the waves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a medium-long shot from a slightly upward angle with a camera swaying precariously?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005986379459500313 + ], + "opposite_logprob_tokens": [ + -5.130986213684082 + ], + "logprob_reward": -0.005986379459500313, + "yes_no_margin_reward": 5.125, + "token_credit_reward": -0.005986379459500313, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13359776139259338, + 0.10565130412578583, + 0.09912379086017609, + 0.08849698305130005, + 0.023709235712885857, + 0.04633425921201706, + 0.09637627750635147, + 0.16771236062049866, + 0.18461962044239044, + 0.05437839403748512 + ], + "credit_map": [ + 0.010893784463405609, + 0.01663401536643505, + 0.018797313794493675, + 0.028456375002861023, + 0.009114101529121399, + 0.017157377675175667, + 0.01503373496234417, + 0.017511069774627686, + 0.01820741966366768, + 0.006565928924828768, + 0.009683895856142044, + 0.009371105581521988, + 0.009436145424842834, + 0.015255648642778397, + 0.029609864577651024, + 0.007521301042288542, + 0.010271253064274788, + 0.009368113242089748, + 0.028730064630508423, + 0.014181185513734818, + 0.007385564967989922, + 0.01525209192186594, + 0.008315146900713444, + 0.0056203738786280155, + 0.01458787452429533, + 0.009286473505198956, + 0.015035833232104778, + 0.009316461160779, + 0.008514100685715675, + 0.011083338409662247, + 0.00992461945861578, + 0.010748283006250858, + 0.001734436838887632, + 0.003387215780094266, + 0.0030397328082472086, + 0.003062655683606863, + 0.004217981360852718, + 0.003714936086907983, + 0.0030071930959820747, + 0.0015450846403837204, + 0.004197349771857262, + 0.00749405100941658, + 0.007769038900732994, + 0.0050381822511553764, + 0.006577104330062866, + 0.006623850669711828, + 0.004797547589987516, + 0.0038371370173990726, + 0.015196103602647781, + 0.016211936250329018, + 0.01430138386785984, + 0.014688600786030293, + 0.010659312829375267, + 0.008214332163333893, + 0.009441056288778782, + 0.007663549855351448, + 0.012141291052103043, + 0.024267051368951797, + 0.05892033502459526, + 0.034576039761304855, + 0.008506547659635544, + 0.0135316988453269, + 0.008512144908308983, + 0.007257247343659401, + 0.010872875340282917, + 0.019141633063554764, + 0.03671959787607193, + 0.07009627670049667, + 0.014382487162947655, + 0.009291463531553745, + 0.013901611790060997, + 0.010213675908744335, + 0.004717481322586536, + 0.008918806910514832, + 0.007619029842317104, + 0.0070708151906728745, + 0.0063699823804199696, + 0.005449309013783932, + 0.009341639466583729, + 0.004891328047960997 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 7, + 0 + ], + "selected_reward": -0.005986379459500313 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set at night with a sky thick with dark clouds and occasional flashes of pale lightning?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is an ancient wooden expedition vessel visible battling mountainous waves in an inky black sea?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is an explorer visible wearing a heavy, rain-soaked windbreaker with wet hair plastered to his forehead?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the explorer slip from the railing, fall backward, and go over the side of the boat into the waves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a medium-long shot from a slightly upward angle with a camera swaying precariously?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01608467288315296 + ], + "opposite_logprob_tokens": [ + -4.141084671020508 + ], + "logprob_reward": -0.01608467288315296, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.01608467288315296, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09876687824726105, + 0.07247262448072433, + 0.0713108479976654, + 0.048761412501335144, + 0.04599353298544884, + 0.04795511066913605, + 0.05190017819404602, + 0.4178147614002228, + 0.12455644458532333, + 0.020468218252062798 + ], + "credit_map": [ + 0.007272147107869387, + 0.011565807275474072, + 0.010944576933979988, + 0.015966370701789856, + 0.011700691655278206, + 0.011085313744843006, + 0.014129448682069778, + 0.016102517023682594, + 0.003900402458384633, + 0.003273819573223591, + 0.010431579314172268, + 0.004785104189068079, + 0.004878224339336157, + 0.02580404467880726, + 0.0142137436196208, + 0.005185712594538927, + 0.007669171318411827, + 0.00905674323439598, + 0.01086627971380949, + 0.011197301559150219, + 0.01358131691813469, + 0.005407289136201143, + 0.009724954143166542, + 0.0038077873177826405, + 0.003390346886590123, + 0.006210534367710352, + 0.006935360375791788, + 0.006475918926298618, + 0.010435661301016808, + 0.0052465940825641155, + 0.0072550042532384396, + 0.0028119944036006927, + 0.002678075572475791, + 0.004392661154270172, + 0.004309660289436579, + 0.005916862282902002, + 0.010875752195715904, + 0.00622048694640398, + 0.008276059292256832, + 0.0033239705953747034, + 0.0031725026201456785, + 0.0050489711575210094, + 0.00855328794568777, + 0.004972320981323719, + 0.007598669268190861, + 0.008208313956856728, + 0.0063004083931446075, + 0.004100637510418892, + 0.0069194273091852665, + 0.008030664175748825, + 0.005785694345831871, + 0.0076356143690645695, + 0.007930340245366096, + 0.005477720871567726, + 0.0059621441178023815, + 0.004158572759479284, + 0.02267950400710106, + 0.016422614455223083, + 0.16692380607128143, + 0.12342724949121475, + 0.020339326933026314, + 0.027388373389840126, + 0.02180970087647438, + 0.01882418431341648, + 0.014584642834961414, + 0.010746991261839867, + 0.016433527693152428, + 0.016122261062264442, + 0.015514695085585117, + 0.01910870522260666, + 0.01790430024266243, + 0.014141326770186424, + 0.0024086220655590296, + 0.003023929428309202, + 0.002809253754094243, + 0.003221665509045124, + 0.0029493055772036314, + 0.001976106083020568, + 0.002572455210611224, + 0.001506880740635097 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 8, + 0 + ], + "selected_reward": -0.01608467288315296 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set at night with a sky thick with dark clouds and occasional flashes of pale lightning?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is an ancient wooden expedition vessel visible battling mountainous waves in an inky black sea?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is an explorer visible wearing a heavy, rain-soaked windbreaker with wet hair plastered to his forehead?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the explorer slip from the railing, fall backward, and go over the side of the boat into the waves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a medium-long shot from a slightly upward angle with a camera swaying precariously?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.038101889193058014 + ], + "opposite_logprob_tokens": [ + -3.2881019115448 + ], + "logprob_reward": -0.038101889193058014, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.038101889193058014, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.047066301107406616, + 0.06715105473995209, + 0.15998250246047974, + 0.03066832758486271, + 0.04862872511148453, + 0.17637628316879272, + 0.031427789479494095, + 0.14624068140983582, + 0.1275336891412735, + 0.16492460668087006 + ], + "credit_map": [ + 0.003550551366060972, + 0.003595111658796668, + 0.006794407498091459, + 0.00813471619039774, + 0.0056594074703752995, + 0.003984441515058279, + 0.006782745011150837, + 0.008564924821257591, + 0.003781573148444295, + 0.002585522597655654, + 0.005003452301025391, + 0.007812279276549816, + 0.007271127309650183, + 0.014118345454335213, + 0.02141173556447029, + 0.005167026072740555, + 0.01170266792178154, + 0.019097071141004562, + 0.023336829617619514, + 0.03733711689710617, + 0.014994990080595016, + 0.026818960905075073, + 0.012324697338044643, + 0.014370177872478962, + 0.0022886593360453844, + 0.0028973447624593973, + 0.0036507092881947756, + 0.004816013853996992, + 0.005597172770649195, + 0.004221930168569088, + 0.005235372111201286, + 0.0019611294846981764, + 0.003293558256700635, + 0.0068816011771559715, + 0.004542991053313017, + 0.005147505085915327, + 0.01082994882017374, + 0.006938679609447718, + 0.007388283498585224, + 0.0036061624996364117, + 0.013583810068666935, + 0.016436591744422913, + 0.02600092440843582, + 0.02002475969493389, + 0.030372891575098038, + 0.03262758627533913, + 0.017601240426301956, + 0.019728492945432663, + 0.0019177324138581753, + 0.00493306340649724, + 0.0034982746001333, + 0.004905649460852146, + 0.005187619477510452, + 0.00351250940002501, + 0.004882147070020437, + 0.0025907964445650578, + 0.01286135334521532, + 0.010102516040205956, + 0.050339240580797195, + 0.04011004790663719, + 0.0076559013687074184, + 0.012035726569592953, + 0.0063661024905741215, + 0.006769813597202301, + 0.012042414397001266, + 0.012478874996304512, + 0.02077937126159668, + 0.024958675727248192, + 0.013307664543390274, + 0.019529111683368683, + 0.012162658385932446, + 0.012274923734366894, + 0.015705494210124016, + 0.02779470384120941, + 0.02396358549594879, + 0.02427760511636734, + 0.014177531003952026, + 0.01547032967209816, + 0.02239922434091568, + 0.021136147901415825 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 9, + 2 + ], + "selected_reward": -0.038101889193058014 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set at night with a sky thick with dark clouds and occasional flashes of pale lightning?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is an ancient wooden expedition vessel visible battling mountainous waves in an inky black sea?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is an explorer visible wearing a heavy, rain-soaked windbreaker with wet hair plastered to his forehead?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the explorer slip from the railing, fall backward, and go over the side of the boat into the waves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a medium-long shot from a slightly upward angle with a camera swaying precariously?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01818707026541233 + ], + "opposite_logprob_tokens": [ + -4.018187046051025 + ], + "logprob_reward": -0.01818707026541233, + "yes_no_margin_reward": 4.0, + "token_credit_reward": -0.01818707026541233, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0656074583530426, + 0.06356582790613174, + 0.12663212418556213, + 0.04987282305955887, + 0.03603820502758026, + 0.05895930156111717, + 0.09397125244140625, + 0.2612171471118927, + 0.14027927815914154, + 0.1038566455245018 + ], + "credit_map": [ + 0.005719457287341356, + 0.009370764717459679, + 0.009345616213977337, + 0.008241751231253147, + 0.005069963168352842, + 0.010106903500854969, + 0.0064299451187253, + 0.011323049664497375, + 0.0058508566580712795, + 0.003156067105010152, + 0.009052429348230362, + 0.005546479485929012, + 0.006175701040774584, + 0.01972007006406784, + 0.006438913755118847, + 0.007625298108905554, + 0.011850173585116863, + 0.016817141324281693, + 0.020272336900234222, + 0.026258917525410652, + 0.0145020242780447, + 0.015021520666778088, + 0.013027634471654892, + 0.00888236053287983, + 0.005127356853336096, + 0.004997997544705868, + 0.00704581942409277, + 0.006931771989911795, + 0.008641407825052738, + 0.00557232229039073, + 0.006862118374556303, + 0.004694017115980387, + 0.0020983845461159945, + 0.003668748075142503, + 0.006399542558938265, + 0.005002527963370085, + 0.006288048345595598, + 0.0044464897364377975, + 0.0060423011891543865, + 0.0020921537652611732, + 0.0042532566003501415, + 0.006758012343198061, + 0.010914658196270466, + 0.003959016874432564, + 0.013134567067027092, + 0.0084399888291955, + 0.007480286527425051, + 0.00401950953528285, + 0.009513644501566887, + 0.01109333150088787, + 0.011685796082019806, + 0.017761988565325737, + 0.015351792797446251, + 0.009782136417925358, + 0.012804352678358555, + 0.005978195928037167, + 0.01544419676065445, + 0.020840326324105263, + 0.08580911159515381, + 0.06767462939023972, + 0.013240043073892593, + 0.027253657579421997, + 0.01271121110767126, + 0.018243934959173203, + 0.009426983073353767, + 0.015347118489444256, + 0.025341901928186417, + 0.05300082638859749, + 0.01031255628913641, + 0.010166197083890438, + 0.00810912437736988, + 0.00857454538345337, + 0.00881368387490511, + 0.018203340470790863, + 0.014930703677237034, + 0.01475611887872219, + 0.008952640928328037, + 0.008985232561826706, + 0.015989989042282104, + 0.013224921189248562 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 8, + 2 + ], + "selected_reward": -0.01818707026541233 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "medium-long shot, shot from a slightly upward angle, the camera swaying precariously with the ship's violent pitching, captures an ancient wooden expedition vessel battling mountainous waves. The hull is constructed of weathered, dark wood, its surface covered in dark water stains and salt crystals. Thick, damp ropes lie scattered on the deck. In a night sky thick with dark clouds and thunder, occasional flashes of pale lightning provide the sole source of light, momentarily illuminating the inky black sea roiling with white foam. An explorer, clad in a heavy windbreaker thoroughly soaked by rain and sea spray, his dark, wet hair disheveled by the wind plastered to his cold-white forehead, clings desperately to a swaying railing, his face stern but unconcealed with fear as he tries to steady himself on the mirror-slick, rain-soaked deck. Suddenly, a huge wave hit the boat violently. His fingers that were holding the railing lost their strength due to the wetness and slipperiness. His feet slipped suddenly and his body lost balance instantly. He waved his arms in vain in the air and then fell backward helplessly. He went over the side of the boat and was finally swallowed up by the inky black waves below that were roaring and rolling with white foam. The whole picture was full of despair, chaos and epic tragedy.", + "seed": 22519, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000101/group_6/index_7173_seed_22519_dp_12_rank_24_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 6, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.015176798962056638, + "vlm_logprob": -0.015176798962056638, + "vlm_yes_no_margin": 4.725, + "vlm_token_credit": -0.015176798962056638, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set at night with a sky thick with dark clouds and occasional flashes of pale lightning?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is an ancient wooden expedition vessel visible battling mountainous waves in an inky black sea?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is an explorer visible wearing a heavy, rain-soaked windbreaker with wet hair plastered to his forehead?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the explorer slip from the railing, fall backward, and go over the side of the boat into the waves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a medium-long shot from a slightly upward angle with a camera swaying precariously?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01639982871711254 + ], + "opposite_logprob_tokens": [ + -6.14139986038208 + ], + "logprob_reward": -0.01639982871711254, + "yes_no_margin_reward": 6.125, + "token_credit_reward": -0.01639982871711254, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1309943050146103, + 0.22344106435775757, + 0.03526689112186432, + 0.050329096615314484, + 0.2025168538093567, + 0.035597000271081924, + 0.09850852191448212, + 0.06903944164514542, + 0.08157217502593994, + 0.07273462414741516 + ], + "credit_map": [ + 0.015586582012474537, + 0.014444759115576744, + 0.010761888697743416, + 0.02651904709637165, + 0.019622137770056725, + 0.014238511212170124, + 0.01703987829387188, + 0.012781506404280663, + 0.08952870965003967, + 0.010015260428190231, + 0.006930613424628973, + 0.0061212098225951195, + 0.00942705012857914, + 0.08645310252904892, + 0.008374886587262154, + 0.006590243894606829, + 0.002968493616208434, + 0.003654472529888153, + 0.004391669761389494, + 0.005049179308116436, + 0.006015795283019543, + 0.0030533750541508198, + 0.006571526639163494, + 0.0035623833537101746, + 0.006241756957024336, + 0.003838807577267289, + 0.006952890194952488, + 0.007261293474584818, + 0.007087502162903547, + 0.005824557971209288, + 0.006864032708108425, + 0.006258259527385235, + 0.017101511359214783, + 0.009869853034615517, + 0.017462657764554024, + 0.02613745629787445, + 0.012308876030147076, + 0.01199044194072485, + 0.04562190920114517, + 0.06202416121959686, + 0.0034210507292300463, + 0.004533424042165279, + 0.004122271668165922, + 0.005682738497853279, + 0.004462153650820255, + 0.003638504073023796, + 0.004379036836326122, + 0.005357824731618166, + 0.007734138052910566, + 0.0074514178559184074, + 0.012871144339442253, + 0.009934787638485432, + 0.01755772903561592, + 0.010933139361441135, + 0.023538878187537193, + 0.008487291634082794, + 0.007187189068645239, + 0.007354398723691702, + 0.008908599615097046, + 0.011152071878314018, + 0.012240888550877571, + 0.008147998712956905, + 0.00857272744178772, + 0.0054755741730332375, + 0.005908842198550701, + 0.008044294081628323, + 0.009195227175951004, + 0.015907716006040573, + 0.021028274670243263, + 0.012658275663852692, + 0.004751242697238922, + 0.004078310914337635, + 0.008160687051713467, + 0.008881023153662682, + 0.00950284581631422, + 0.008004014380276203, + 0.010330828838050365, + 0.010455735959112644, + 0.006885993294417858, + 0.010513502173125744 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 4, + 0 + ], + "selected_reward": -0.01639982871711254 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set at night with a sky thick with dark clouds and occasional flashes of pale lightning?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is an ancient wooden expedition vessel visible battling mountainous waves in an inky black sea?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is an explorer visible wearing a heavy, rain-soaked windbreaker with wet hair plastered to his forehead?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the explorer slip from the railing, fall backward, and go over the side of the boat into the waves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a medium-long shot from a slightly upward angle with a camera swaying precariously?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004118770360946655 + ], + "opposite_logprob_tokens": [ + -5.504118919372559 + ], + "logprob_reward": -0.004118770360946655, + "yes_no_margin_reward": 5.5, + "token_credit_reward": -0.004118770360946655, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.03358996659517288, + 0.6283819675445557, + 0.027213696390390396, + 0.040442127734422684, + 0.05487961322069168, + 0.029204655438661575, + 0.0340406633913517, + 0.012989820912480354, + 0.035394057631492615, + 0.10386336594820023 + ], + "credit_map": [ + 0.004394178744405508, + 0.002842176705598831, + 0.0038066471461206675, + 0.0049403076991438866, + 0.0048492723144590855, + 0.0039136530831456184, + 0.004738477058708668, + 0.0041052536107599735, + 0.2582041025161743, + 0.028864169493317604, + 0.014657005667686462, + 0.014759136363863945, + 0.024649139493703842, + 0.24671417474746704, + 0.023537324741482735, + 0.016996920108795166, + 0.0023968347813934088, + 0.0032403746154159307, + 0.002894408069550991, + 0.0033069902565330267, + 0.005978827830404043, + 0.0024307691492140293, + 0.0042603532783687115, + 0.0027051367796957493, + 0.004512287210673094, + 0.0026894048787653446, + 0.004499771632254124, + 0.005747057497501373, + 0.006219820585101843, + 0.0038114413619041443, + 0.005868561565876007, + 0.007093783468008041, + 0.003997888881713152, + 0.006399196106940508, + 0.007788130082190037, + 0.008006259799003601, + 0.006523790769279003, + 0.0034896160941570997, + 0.013506595976650715, + 0.005168134346604347, + 0.0023116017691791058, + 0.002490744460374117, + 0.004150212276726961, + 0.0038377228192985058, + 0.005051929969340563, + 0.002888947958126664, + 0.004639218561351299, + 0.0038342762272804976, + 0.002674184273928404, + 0.0021723764948546886, + 0.004573764745146036, + 0.004060088191181421, + 0.005537758115679026, + 0.004010443575680256, + 0.008212234824895859, + 0.0027998127043247223, + 0.0017324298387393355, + 0.0011167926713824272, + 0.0016623922856524587, + 0.0021313661709427834, + 0.0019915280863642693, + 0.0016619754023849964, + 0.0018343244446441531, + 0.000859011837746948, + 0.003614237764850259, + 0.0034263632260262966, + 0.005923683289438486, + 0.003391498001292348, + 0.00820295698940754, + 0.005033308640122414, + 0.0031787066254764795, + 0.0026233044918626547, + 0.009790603071451187, + 0.013505212962627411, + 0.013037238270044327, + 0.008313924074172974, + 0.017820212990045547, + 0.014722047373652458, + 0.012001690454781055, + 0.014672440476715565 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 9, + 4 + ], + "selected_reward": -0.004118770360946655 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set at night with a sky thick with dark clouds and occasional flashes of pale lightning?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is an ancient wooden expedition vessel visible battling mountainous waves in an inky black sea?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is an explorer visible wearing a heavy, rain-soaked windbreaker with wet hair plastered to his forehead?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the explorer slip from the railing, fall backward, and go over the side of the boat into the waves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a medium-long shot from a slightly upward angle with a camera swaying precariously?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02059071511030197 + ], + "opposite_logprob_tokens": [ + -3.8955907821655273 + ], + "logprob_reward": -0.02059071511030197, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.02059071511030197, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.05042017251253128, + 0.0930841937661171, + 0.037916071712970734, + 0.12571801245212555, + 0.14998659491539001, + 0.044451016932725906, + 0.16837751865386963, + 0.09689531475305557, + 0.08559462428092957, + 0.14755649864673615 + ], + "credit_map": [ + 0.005394740030169487, + 0.0039308322593569756, + 0.005937803070992231, + 0.00890369899570942, + 0.006343784276396036, + 0.006533806212246418, + 0.007609345484524965, + 0.005766160320490599, + 0.033024825155735016, + 0.004587327130138874, + 0.004991709720343351, + 0.004750032443553209, + 0.005512777250260115, + 0.03142119199037552, + 0.004967765416949987, + 0.003828569548204541, + 0.003427269170060754, + 0.003868822008371353, + 0.004553461447358131, + 0.005817854311317205, + 0.00563039630651474, + 0.003339343937113881, + 0.007392848841845989, + 0.0038860782515257597, + 0.01633782871067524, + 0.009600026533007622, + 0.012150386348366737, + 0.01752431131899357, + 0.01636817678809166, + 0.011646248400211334, + 0.02321312017738819, + 0.0188779067248106, + 0.009022041223943233, + 0.0068422039039433, + 0.017680203542113304, + 0.02596287615597248, + 0.02130887657403946, + 0.012067656964063644, + 0.04068071395158768, + 0.016422022134065628, + 0.0033609934616833925, + 0.004357108846306801, + 0.005382840055972338, + 0.00516920443624258, + 0.007013888563960791, + 0.004803238436579704, + 0.007815863937139511, + 0.006547876168042421, + 0.012344226241111755, + 0.008766362443566322, + 0.023672685027122498, + 0.01710604690015316, + 0.02927597425878048, + 0.018541764467954636, + 0.04295625165104866, + 0.01571420580148697, + 0.00927220843732357, + 0.010498188436031342, + 0.010757752694189548, + 0.015260177664458752, + 0.021166346967220306, + 0.012454863637685776, + 0.010409881360828876, + 0.007075903937220573, + 0.0070038409903645515, + 0.009871152229607105, + 0.012238854542374611, + 0.0197385735809803, + 0.016867689788341522, + 0.01047334261238575, + 0.0058891125954687595, + 0.0035120598040521145, + 0.01613040640950203, + 0.018842898309230804, + 0.018453985452651978, + 0.013863827101886272, + 0.02610946260392666, + 0.020472615957260132, + 0.016295161098241806, + 0.01738813705742359 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 4, + 9 + ], + "selected_reward": -0.02059071511030197 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set at night with a sky thick with dark clouds and occasional flashes of pale lightning?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is an ancient wooden expedition vessel visible battling mountainous waves in an inky black sea?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is an explorer visible wearing a heavy, rain-soaked windbreaker with wet hair plastered to his forehead?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the explorer slip from the railing, fall backward, and go over the side of the boat into the waves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a medium-long shot from a slightly upward angle with a camera swaying precariously?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01419412437826395 + ], + "opposite_logprob_tokens": [ + -4.264194011688232 + ], + "logprob_reward": -0.01419412437826395, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.01419412437826395, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.04422672092914581, + 0.2130022495985031, + 0.038142673671245575, + 0.07266179472208023, + 0.08051006495952606, + 0.03476601094007492, + 0.1742255836725235, + 0.23384112119674683, + 0.049166176468133926, + 0.05945756286382675 + ], + "credit_map": [ + 0.005817676428705454, + 0.003974683582782745, + 0.0042780241928994656, + 0.008261014707386494, + 0.006222787778824568, + 0.0058000460267066956, + 0.005838820710778236, + 0.004033665172755718, + 0.08880038559436798, + 0.008832491002976894, + 0.005184743087738752, + 0.004049056209623814, + 0.00801218580454588, + 0.0840500220656395, + 0.008024424314498901, + 0.006048937793821096, + 0.004538574256002903, + 0.003405555384233594, + 0.005151223856955767, + 0.004151749890297651, + 0.006383967585861683, + 0.0041800737380981445, + 0.005024774000048637, + 0.005306754261255264, + 0.009752756915986538, + 0.004923615604639053, + 0.007922284305095673, + 0.009761905297636986, + 0.006980483420193195, + 0.0073904371820390224, + 0.012873467989265919, + 0.013056846335530281, + 0.010129696689546108, + 0.008419293910264969, + 0.005910220555961132, + 0.008625893853604794, + 0.005523455794900656, + 0.006336103659123182, + 0.013637036085128784, + 0.021928364410996437, + 0.0026710587553679943, + 0.002303760265931487, + 0.005681644659489393, + 0.004244575276970863, + 0.004441448021680117, + 0.0028641491662710905, + 0.006062688771635294, + 0.006496687885373831, + 0.01327374018728733, + 0.008090289309620857, + 0.023189086467027664, + 0.017964674159884453, + 0.02940552495419979, + 0.019320400431752205, + 0.04622689262032509, + 0.016754984855651855, + 0.023993564769625664, + 0.02538049966096878, + 0.03223322331905365, + 0.03531937301158905, + 0.04264070838689804, + 0.031353116035461426, + 0.02461916208267212, + 0.018301496282219887, + 0.004498916212469339, + 0.004696231801062822, + 0.008191859349608421, + 0.011297839693725109, + 0.0056014456786215305, + 0.006217942573130131, + 0.005356495268642902, + 0.0033054486848413944, + 0.006946053355932236, + 0.007085960824042559, + 0.006797126028686762, + 0.005797363817691803, + 0.010193432681262493, + 0.00868711993098259, + 0.006866155657917261, + 0.007084347773343325 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 1, + 6 + ], + "selected_reward": -0.01419412437826395 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set at night with a sky thick with dark clouds and occasional flashes of pale lightning?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is an ancient wooden expedition vessel visible battling mountainous waves in an inky black sea?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is an explorer visible wearing a heavy, rain-soaked windbreaker with wet hair plastered to his forehead?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the explorer slip from the railing, fall backward, and go over the side of the boat into the waves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a medium-long shot from a slightly upward angle with a camera swaying precariously?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.020580556243658066 + ], + "opposite_logprob_tokens": [ + -3.895580530166626 + ], + "logprob_reward": -0.020580556243658066, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.020580556243658066, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.05249052122235298, + 0.4692690074443817, + 0.03273705393075943, + 0.02483321540057659, + 0.1507207453250885, + 0.031054656952619553, + 0.04652637243270874, + 0.10263364017009735, + 0.03463314101099968, + 0.05510164052248001 + ], + "credit_map": [ + 0.006265438161790371, + 0.0073736258782446384, + 0.0056800199672579765, + 0.007355244364589453, + 0.008380790241062641, + 0.00663065304979682, + 0.005927092861384153, + 0.004877655301243067, + 0.19653332233428955, + 0.020023096352815628, + 0.010650086216628551, + 0.00800822302699089, + 0.016972241923213005, + 0.18739613890647888, + 0.016039712354540825, + 0.013646162115037441, + 0.003360083093866706, + 0.00370426126755774, + 0.003675157902762294, + 0.004274806007742882, + 0.005160152912139893, + 0.0037272910121828318, + 0.005472928751260042, + 0.003362372051924467, + 0.0024019840639084578, + 0.0016616993816569448, + 0.0032089438755065203, + 0.0035559122916311026, + 0.00400365237146616, + 0.002436096780002117, + 0.0044710636138916016, + 0.0030938617419451475, + 0.010748141445219517, + 0.01966908387839794, + 0.021258773282170296, + 0.02420751377940178, + 0.017739778384566307, + 0.010803543962538242, + 0.028914799913764, + 0.017379116266965866, + 0.002797577530145645, + 0.0021092549432069063, + 0.003387687262147665, + 0.004178767092525959, + 0.005509268492460251, + 0.002813724335283041, + 0.005503611173480749, + 0.004754768218845129, + 0.004315776750445366, + 0.005244842264801264, + 0.007661847863346338, + 0.005196412559598684, + 0.0059072598814964294, + 0.005431638564914465, + 0.00791951920837164, + 0.004849074874073267, + 0.011654021218419075, + 0.008702384307980537, + 0.010508403182029724, + 0.015662971884012222, + 0.020952923223376274, + 0.015104011632502079, + 0.011954490095376968, + 0.008094435557723045, + 0.0028273023199290037, + 0.004124541766941547, + 0.005228989291936159, + 0.007147128228098154, + 0.005234581883996725, + 0.004218677990138531, + 0.003058506641536951, + 0.0027934147510677576, + 0.004862652625888586, + 0.006762044038623571, + 0.006269470788538456, + 0.004775893874466419, + 0.008976622484624386, + 0.007660974282771349, + 0.006738363765180111, + 0.00905561912804842 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 4, + 7 + ], + "selected_reward": -0.020580556243658066 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "medium-long shot, shot from a slightly upward angle, the camera swaying precariously with the ship's violent pitching, captures an ancient wooden expedition vessel battling mountainous waves. The hull is constructed of weathered, dark wood, its surface covered in dark water stains and salt crystals. Thick, damp ropes lie scattered on the deck. In a night sky thick with dark clouds and thunder, occasional flashes of pale lightning provide the sole source of light, momentarily illuminating the inky black sea roiling with white foam. An explorer, clad in a heavy windbreaker thoroughly soaked by rain and sea spray, his dark, wet hair disheveled by the wind plastered to his cold-white forehead, clings desperately to a swaying railing, his face stern but unconcealed with fear as he tries to steady himself on the mirror-slick, rain-soaked deck. Suddenly, a huge wave hit the boat violently. His fingers that were holding the railing lost their strength due to the wetness and slipperiness. His feet slipped suddenly and his body lost balance instantly. He waved his arms in vain in the air and then fell backward helplessly. He went over the side of the boat and was finally swallowed up by the inky black waves below that were roaring and rolling with white foam. The whole picture was full of despair, chaos and epic tragedy.", + "seed": 22520, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000101/group_6/index_7173_seed_22520_dp_13_rank_26_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 6, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.03773127067834139, + "vlm_logprob": -0.03773127067834139, + "vlm_yes_no_margin": 3.875, + "vlm_token_credit": -0.03773127067834139, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set at night with a sky thick with dark clouds and occasional flashes of pale lightning?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is an ancient wooden expedition vessel visible battling mountainous waves in an inky black sea?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is an explorer visible wearing a heavy, rain-soaked windbreaker with wet hair plastered to his forehead?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the explorer slip from the railing, fall backward, and go over the side of the boat into the waves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a medium-long shot from a slightly upward angle with a camera swaying precariously?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.012910587713122368 + ], + "opposite_logprob_tokens": [ + -4.88791036605835 + ], + "logprob_reward": -0.012910587713122368, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.012910587713122368, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15531472861766815, + 0.06002211198210716, + 0.056267932057380676, + 0.19547194242477417, + 0.03718731552362442, + 0.04060879722237587, + 0.12282133847475052, + 0.22562797367572784, + 0.05818972736597061, + 0.048488032072782516 + ], + "credit_map": [ + 0.014018522575497627, + 0.017664652317762375, + 0.023536095395684242, + 0.027405008673667908, + 0.015266701579093933, + 0.017367271706461906, + 0.02272009663283825, + 0.017336411401629448, + 0.006233364809304476, + 0.008123490028083324, + 0.007794574368745089, + 0.010019293986260891, + 0.005888799205422401, + 0.005438452120870352, + 0.009907686151564121, + 0.006616454571485519, + 0.006813680753111839, + 0.007088965270668268, + 0.007786519825458527, + 0.007806622423231602, + 0.006578283850103617, + 0.006347839254885912, + 0.008101707324385643, + 0.005744323134422302, + 0.028955742716789246, + 0.029265768826007843, + 0.021074315533041954, + 0.02496279962360859, + 0.023313326761126518, + 0.026903053745627403, + 0.02129989117383957, + 0.01969706080853939, + 0.003535281401127577, + 0.006095763761550188, + 0.004768645390868187, + 0.004601323511451483, + 0.0047893282026052475, + 0.004319667350500822, + 0.0038497773930430412, + 0.005227529909461737, + 0.004769004415720701, + 0.007020389195531607, + 0.004865801427513361, + 0.005413321778178215, + 0.005274370778352022, + 0.004687082953751087, + 0.004486361518502235, + 0.0040924688801169395, + 0.016289344057440758, + 0.01578688621520996, + 0.009434491395950317, + 0.009238027967512608, + 0.04360180348157883, + 0.008297106251120567, + 0.010337339714169502, + 0.009836352430284023, + 0.020263560116291046, + 0.024249117821455002, + 0.021314363926649094, + 0.029608922079205513, + 0.021183790639042854, + 0.023556729778647423, + 0.05516141280531883, + 0.03029010072350502, + 0.006460119970142841, + 0.0070215435698628426, + 0.007619266398251057, + 0.007157651707530022, + 0.010614990256726742, + 0.007702284958213568, + 0.005309461150318384, + 0.0063044121488928795, + 0.004956373944878578, + 0.00588639872148633, + 0.0074057867750525475, + 0.005647170357406139, + 0.006572921760380268, + 0.0070841629058122635, + 0.005546540953218937, + 0.00538868410512805 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 3, + 0 + ], + "selected_reward": -0.012910587713122368 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set at night with a sky thick with dark clouds and occasional flashes of pale lightning?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is an ancient wooden expedition vessel visible battling mountainous waves in an inky black sea?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is an explorer visible wearing a heavy, rain-soaked windbreaker with wet hair plastered to his forehead?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the explorer slip from the railing, fall backward, and go over the side of the boat into the waves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a medium-long shot from a slightly upward angle with a camera swaying precariously?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.008666876703500748 + ], + "opposite_logprob_tokens": [ + -4.7586669921875 + ], + "logprob_reward": -0.008666876703500748, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.008666876703500748, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.17490993440151215, + 0.07792996615171432, + 0.11654887348413467, + 0.19367550313472748, + 0.026895660907030106, + 0.08781030774116516, + 0.09394964575767517, + 0.153660848736763, + 0.041941262781620026, + 0.032677993178367615 + ], + "credit_map": [ + 0.01509006880223751, + 0.020860157907009125, + 0.0232694111764431, + 0.02622189000248909, + 0.017838574945926666, + 0.019625050947070122, + 0.025544870644807816, + 0.02645989879965782, + 0.007710317149758339, + 0.010302729904651642, + 0.011184459552168846, + 0.013029935769736767, + 0.008055862039327621, + 0.008030963130295277, + 0.011980317533016205, + 0.007635373156517744, + 0.012897539883852005, + 0.015734460204839706, + 0.014373996295034885, + 0.017366480082273483, + 0.015083436854183674, + 0.013101640157401562, + 0.01834598183631897, + 0.009645339101552963, + 0.020199118182063103, + 0.023239552974700928, + 0.024240609258413315, + 0.032052259892225266, + 0.022150857374072075, + 0.01683810167014599, + 0.020828727632761, + 0.034126270562410355, + 0.0021675042808055878, + 0.0026447735726833344, + 0.003838225966319442, + 0.004045483656227589, + 0.0030178793240338564, + 0.0048412964679300785, + 0.003373719286173582, + 0.002966778352856636, + 0.007941985502839088, + 0.01856485940515995, + 0.011385226622223854, + 0.012416490353643894, + 0.008899013511836529, + 0.009944114834070206, + 0.012045454233884811, + 0.006613167002797127, + 0.01495842169970274, + 0.01066168025135994, + 0.009859275072813034, + 0.00665796734392643, + 0.02755715698003769, + 0.00764737231656909, + 0.008207005448639393, + 0.008400771766901016, + 0.012552458792924881, + 0.014510785229504108, + 0.015385939739644527, + 0.020332666113972664, + 0.014852003194391727, + 0.016340110450983047, + 0.03756733983755112, + 0.022119546309113503, + 0.00634214049205184, + 0.004628431983292103, + 0.007184338290244341, + 0.0039672693237662315, + 0.005292181391268969, + 0.004899703431874514, + 0.004615944344550371, + 0.005011256318539381, + 0.0030481265857815742, + 0.00376547290943563, + 0.006161519326269627, + 0.0027213464491069317, + 0.004699320532381535, + 0.004452710971236229, + 0.003842824138700962, + 0.003986673895269632 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 0, + 7 + ], + "selected_reward": -0.008666876703500748 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set at night with a sky thick with dark clouds and occasional flashes of pale lightning?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is an ancient wooden expedition vessel visible battling mountainous waves in an inky black sea?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is an explorer visible wearing a heavy, rain-soaked windbreaker with wet hair plastered to his forehead?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the explorer slip from the railing, fall backward, and go over the side of the boat into the waves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a medium-long shot from a slightly upward angle with a camera swaying precariously?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04306899383664131 + ], + "opposite_logprob_tokens": [ + -3.1680688858032227 + ], + "logprob_reward": -0.04306899383664131, + "yes_no_margin_reward": 3.125, + "token_credit_reward": -0.04306899383664131, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14632214605808258, + 0.06741393357515335, + 0.08323784917593002, + 0.08964065462350845, + 0.05268052592873573, + 0.08479751646518707, + 0.08826795965433121, + 0.33836260437965393, + 0.02651975490152836, + 0.022757021710276604 + ], + "credit_map": [ + 0.011887745000422001, + 0.0401928536593914, + 0.020594606176018715, + 0.018837861716747284, + 0.013885173946619034, + 0.014779099263250828, + 0.015547189861536026, + 0.010597622022032738, + 0.006882591638714075, + 0.00904038641601801, + 0.009932579472661018, + 0.00925105344504118, + 0.009550212882459164, + 0.007286109495908022, + 0.012165157124400139, + 0.0033058500848710537, + 0.0065874275751411915, + 0.0067024691961705685, + 0.01141887716948986, + 0.01432300079613924, + 0.017135897651314735, + 0.0054955254308879375, + 0.015615930780768394, + 0.005958729889243841, + 0.010085134766995907, + 0.01156350877135992, + 0.015826933085918427, + 0.016512786969542503, + 0.010289234109222889, + 0.007434884551912546, + 0.010002023540437222, + 0.007926154881715775, + 0.0059986598789691925, + 0.004808647558093071, + 0.007397940382361412, + 0.007492356933653355, + 0.007696386426687241, + 0.006100732367485762, + 0.006514550652354956, + 0.006671257317066193, + 0.00692750234156847, + 0.018963374197483063, + 0.01106555201113224, + 0.011646383441984653, + 0.00741399172693491, + 0.005005271639674902, + 0.010443225502967834, + 0.013332224451005459, + 0.015671560540795326, + 0.007895248010754585, + 0.009626271203160286, + 0.00613440852612257, + 0.008640014566481113, + 0.008212530054152012, + 0.0076148598454892635, + 0.024473078548908234, + 0.03473925217986107, + 0.04276164621114731, + 0.033071745187044144, + 0.04461463540792465, + 0.03134741634130478, + 0.03262171521782875, + 0.08195357024669647, + 0.03725261986255646, + 0.0028179676737636328, + 0.002277820836752653, + 0.004175918642431498, + 0.003564309561625123, + 0.002984581748023629, + 0.005993236787617207, + 0.002507875207811594, + 0.002198045840486884, + 0.00212497659958899, + 0.0039115361869335175, + 0.003214264987036586, + 0.0020381147041916847, + 0.0025986251421272755, + 0.0037125302478671074, + 0.002833862788975239, + 0.002323112217709422 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 0, + 3 + ], + "selected_reward": -0.04306899383664131 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set at night with a sky thick with dark clouds and occasional flashes of pale lightning?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is an ancient wooden expedition vessel visible battling mountainous waves in an inky black sea?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is an explorer visible wearing a heavy, rain-soaked windbreaker with wet hair plastered to his forehead?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the explorer slip from the railing, fall backward, and go over the side of the boat into the waves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a medium-long shot from a slightly upward angle with a camera swaying precariously?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.11293362081050873 + ], + "opposite_logprob_tokens": [ + -2.23793363571167 + ], + "logprob_reward": -0.11293362081050873, + "yes_no_margin_reward": 2.125, + "token_credit_reward": -0.11293362081050873, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09141484647989273, + 0.07905807346105576, + 0.05094520375132561, + 0.06143646314740181, + 0.03493119403719902, + 0.08460356295108795, + 0.2134687602519989, + 0.2588346302509308, + 0.05775768682360649, + 0.06754950433969498 + ], + "credit_map": [ + 0.00921569112688303, + 0.014027358032763004, + 0.014809241518378258, + 0.015915868803858757, + 0.008711169473826885, + 0.01108170859515667, + 0.011844134889543056, + 0.005809675436466932, + 0.006297494284808636, + 0.008283889852464199, + 0.01190159935504198, + 0.01144509669393301, + 0.009171937592327595, + 0.008111870847642422, + 0.018925409764051437, + 0.004920776002109051, + 0.0047330064699053764, + 0.004971511662006378, + 0.006985285319387913, + 0.008554317988455296, + 0.006542399059981108, + 0.004462409298866987, + 0.008567414246499538, + 0.006128862965852022, + 0.00688514020293951, + 0.008832545019686222, + 0.007493531331419945, + 0.010156664066016674, + 0.007970183156430721, + 0.004684715531766415, + 0.007512487005442381, + 0.007901201955974102, + 0.0026210397481918335, + 0.002907385816797614, + 0.00465018767863512, + 0.006156296003609896, + 0.005196461454033852, + 0.004770219326019287, + 0.005325625650584698, + 0.003303984645754099, + 0.0072151049971580505, + 0.018461817875504494, + 0.009514322504401207, + 0.013903224840760231, + 0.007999373599886894, + 0.005132203921675682, + 0.009544325061142445, + 0.012833193875849247, + 0.03497793525457382, + 0.02222844772040844, + 0.02164604142308235, + 0.018270516768097878, + 0.03986335173249245, + 0.019246812909841537, + 0.021364005282521248, + 0.03587166219949722, + 0.0222857054322958, + 0.026895053684711456, + 0.02394944429397583, + 0.03330017253756523, + 0.024425756186246872, + 0.029474541544914246, + 0.061016157269477844, + 0.037487830966711044, + 0.0056966557167470455, + 0.005381525028496981, + 0.00882629118859768, + 0.008069322444498539, + 0.005985249765217304, + 0.013019878417253494, + 0.005677736364305019, + 0.005101028364151716, + 0.005848835688084364, + 0.006328482646495104, + 0.012264532968401909, + 0.0060375891625881195, + 0.010504670441150665, + 0.010669391602277756, + 0.008758489042520523, + 0.00713751558214426 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 6, + 0 + ], + "selected_reward": -0.11293362081050873 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set at night with a sky thick with dark clouds and occasional flashes of pale lightning?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is an ancient wooden expedition vessel visible battling mountainous waves in an inky black sea?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is an explorer visible wearing a heavy, rain-soaked windbreaker with wet hair plastered to his forehead?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the explorer slip from the railing, fall backward, and go over the side of the boat into the waves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a medium-long shot from a slightly upward angle with a camera swaying precariously?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.011076274327933788 + ], + "opposite_logprob_tokens": [ + -4.5110764503479 + ], + "logprob_reward": -0.011076274327933788, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.011076274327933788, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09477885067462921, + 0.07518892735242844, + 0.10982813686132431, + 0.16620558500289917, + 0.04632817581295967, + 0.09849635511636734, + 0.09287731349468231, + 0.17626024782657623, + 0.07646655291318893, + 0.06356983631849289 + ], + "credit_map": [ + 0.009502900764346123, + 0.011451395228505135, + 0.013386987149715424, + 0.011404917575418949, + 0.008536292240023613, + 0.01700238510966301, + 0.011259193532168865, + 0.012234781868755817, + 0.005216219462454319, + 0.007797526195645332, + 0.01027365680783987, + 0.01067324262112379, + 0.011672505177557468, + 0.0060722981579601765, + 0.018009383231401443, + 0.00547409662976861, + 0.009054129011929035, + 0.010779112577438354, + 0.016254298388957977, + 0.022107400000095367, + 0.01585238240659237, + 0.008387571200728416, + 0.02054859697818756, + 0.006844648625701666, + 0.016463421285152435, + 0.00787730235606432, + 0.036760710179805756, + 0.02926633693277836, + 0.016356272622942924, + 0.01027245819568634, + 0.020681817084550858, + 0.02852724865078926, + 0.0032524000853300095, + 0.0044364421628415585, + 0.006600033957511187, + 0.007087315432727337, + 0.005787174683064222, + 0.006473947782069445, + 0.008259437046945095, + 0.004431422799825668, + 0.008947168476879597, + 0.010812610387802124, + 0.01864304579794407, + 0.011314290575683117, + 0.007892211899161339, + 0.010426843538880348, + 0.013902823440730572, + 0.016557365655899048, + 0.0076885404996573925, + 0.011899611912667751, + 0.007330461870878935, + 0.009035428054630756, + 0.022539522498846054, + 0.007958056405186653, + 0.007135586813092232, + 0.019290106371045113, + 0.011147914454340935, + 0.018889304250478745, + 0.017167169600725174, + 0.01578223519027233, + 0.022783907130360603, + 0.033079009503126144, + 0.030470378696918488, + 0.026940323412418365, + 0.006305406801402569, + 0.008974728174507618, + 0.012821182608604431, + 0.009347163140773773, + 0.010460362769663334, + 0.013404653407633305, + 0.008001159876585007, + 0.007151896599680185, + 0.0050890217535197735, + 0.009958920069038868, + 0.012975537218153477, + 0.0066825696267187595, + 0.0076977889984846115, + 0.008369074203073978, + 0.006161689758300781, + 0.006635237950831652 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 3, + 2 + ], + "selected_reward": -0.011076274327933788 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "medium-long shot, shot from a slightly upward angle, the camera swaying precariously with the ship's violent pitching, captures an ancient wooden expedition vessel battling mountainous waves. The hull is constructed of weathered, dark wood, its surface covered in dark water stains and salt crystals. Thick, damp ropes lie scattered on the deck. In a night sky thick with dark clouds and thunder, occasional flashes of pale lightning provide the sole source of light, momentarily illuminating the inky black sea roiling with white foam. An explorer, clad in a heavy windbreaker thoroughly soaked by rain and sea spray, his dark, wet hair disheveled by the wind plastered to his cold-white forehead, clings desperately to a swaying railing, his face stern but unconcealed with fear as he tries to steady himself on the mirror-slick, rain-soaked deck. Suddenly, a huge wave hit the boat violently. His fingers that were holding the railing lost their strength due to the wetness and slipperiness. His feet slipped suddenly and his body lost balance instantly. He waved his arms in vain in the air and then fell backward helplessly. He went over the side of the boat and was finally swallowed up by the inky black waves below that were roaring and rolling with white foam. The whole picture was full of despair, chaos and epic tragedy.", + "seed": 22521, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000101/group_6/index_7173_seed_22521_dp_13_rank_26_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 6, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.013042869418859482, + "vlm_logprob": -0.013042869418859482, + "vlm_yes_no_margin": 4.575, + "vlm_token_credit": -0.013042869418859482, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set at night with a sky thick with dark clouds and occasional flashes of pale lightning?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is an ancient wooden expedition vessel visible battling mountainous waves in an inky black sea?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is an explorer visible wearing a heavy, rain-soaked windbreaker with wet hair plastered to his forehead?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the explorer slip from the railing, fall backward, and go over the side of the boat into the waves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a medium-long shot from a slightly upward angle with a camera swaying precariously?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.015750762075185776 + ], + "opposite_logprob_tokens": [ + -5.140750885009766 + ], + "logprob_reward": -0.015750762075185776, + "yes_no_margin_reward": 5.125, + "token_credit_reward": -0.015750762075185776, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.16197878122329712, + 0.029652297496795654, + 0.033230941742658615, + 0.03948436677455902, + 0.042004674673080444, + 0.02932010404765606, + 0.4617329239845276, + 0.09158072620630264, + 0.0519452840089798, + 0.05906981974840164 + ], + "credit_map": [ + 0.008306670933961868, + 0.013475202023983002, + 0.012938633561134338, + 0.045460838824510574, + 0.022350380197167397, + 0.022204933688044548, + 0.021504899486899376, + 0.01573721505701542, + 0.0023394511081278324, + 0.004874077159911394, + 0.0023793899454176426, + 0.004637847654521465, + 0.0033495971001684666, + 0.005191763397306204, + 0.003599528456106782, + 0.003280645003542304, + 0.0024366637226194143, + 0.006591030862182379, + 0.003101265523582697, + 0.004892976023256779, + 0.004298269748687744, + 0.004491652362048626, + 0.0038066457491368055, + 0.003612438915297389, + 0.0019547026604413986, + 0.0038242340087890625, + 0.003211169270798564, + 0.012358180247247219, + 0.0031564226374030113, + 0.007435288280248642, + 0.003935660235583782, + 0.0036087108310312033, + 0.003332569031044841, + 0.003222433151677251, + 0.003841710975393653, + 0.006333114579319954, + 0.00450848788022995, + 0.0038707663770765066, + 0.003994752652943134, + 0.01290084421634674, + 0.002846202114596963, + 0.004090036265552044, + 0.0036249500699341297, + 0.0031915579456835985, + 0.004729917738586664, + 0.004078990779817104, + 0.0038800565525889397, + 0.0028783963061869144, + 0.012412255629897118, + 0.02061508037149906, + 0.004815434105694294, + 0.01915537752211094, + 0.03019590675830841, + 0.1755315661430359, + 0.04447450488805771, + 0.15453284978866577, + 0.003603423712775111, + 0.00603931350633502, + 0.00763738714158535, + 0.017417417839169502, + 0.007727256044745445, + 0.004397967830300331, + 0.03091094084084034, + 0.01384702231734991, + 0.004174969624727964, + 0.00962977297604084, + 0.0076184687204658985, + 0.005099585745483637, + 0.007107776589691639, + 0.009199341759085655, + 0.004383434075862169, + 0.004731936380267143, + 0.006087696645408869, + 0.004557027481496334, + 0.017780788242816925, + 0.004836614243686199, + 0.005359198898077011, + 0.005905711557716131, + 0.005055820569396019, + 0.009486964903771877 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 0, + 7 + ], + "selected_reward": -0.015750762075185776 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set at night with a sky thick with dark clouds and occasional flashes of pale lightning?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is an ancient wooden expedition vessel visible battling mountainous waves in an inky black sea?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is an explorer visible wearing a heavy, rain-soaked windbreaker with wet hair plastered to his forehead?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the explorer slip from the railing, fall backward, and go over the side of the boat into the waves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a medium-long shot from a slightly upward angle with a camera swaying precariously?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.008671722374856472 + ], + "opposite_logprob_tokens": [ + -4.758671760559082 + ], + "logprob_reward": -0.008671722374856472, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.008671722374856472, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0907711386680603, + 0.04497360438108444, + 0.061612021178007126, + 0.06848133355379105, + 0.04874246567487717, + 0.06094875931739807, + 0.4020439088344574, + 0.09784609079360962, + 0.02338513918220997, + 0.1011955663561821 + ], + "credit_map": [ + 0.005225350148975849, + 0.008059856481850147, + 0.006341055501252413, + 0.0233662948012352, + 0.014575791545212269, + 0.011856522411108017, + 0.013701669871807098, + 0.00764458766207099, + 0.004224239382892847, + 0.005315685179084539, + 0.003099591936916113, + 0.004623728338629007, + 0.007857599295675755, + 0.006872453726828098, + 0.0069036297500133514, + 0.006076674908399582, + 0.004413920920342207, + 0.004335313569754362, + 0.006910237483680248, + 0.009089433588087559, + 0.007157179061323404, + 0.009320693090558052, + 0.011404139921069145, + 0.008981102146208286, + 0.004122761078178883, + 0.008152468129992485, + 0.0044268649071455, + 0.010689484886825085, + 0.013787459582090378, + 0.008836813271045685, + 0.008893193677067757, + 0.009572282433509827, + 0.004368428606539965, + 0.0037091602571308613, + 0.00381481577642262, + 0.006189325824379921, + 0.007378487382084131, + 0.004154426511377096, + 0.004464380908757448, + 0.014663436450064182, + 0.008115825243294239, + 0.009302234277129173, + 0.005138232838362455, + 0.006002653390169144, + 0.01337179820984602, + 0.007197050843387842, + 0.008031625300645828, + 0.003789337817579508, + 0.010776902548968792, + 0.017857694998383522, + 0.0043190522119402885, + 0.01732001081109047, + 0.02695014327764511, + 0.15231767296791077, + 0.038134630769491196, + 0.13436776399612427, + 0.00603081239387393, + 0.006468912586569786, + 0.008428020402789116, + 0.017406998202204704, + 0.011051509529352188, + 0.006962239742279053, + 0.012849604710936546, + 0.02864799275994301, + 0.002183884847909212, + 0.0043375627137720585, + 0.0031605924014002085, + 0.002238619839772582, + 0.002939950441941619, + 0.004356317687779665, + 0.001789468340575695, + 0.002378743374720216, + 0.011130397208034992, + 0.004900000989437103, + 0.03167080879211426, + 0.01021601539105177, + 0.007559353485703468, + 0.00795547105371952, + 0.00771472929045558, + 0.020048782229423523 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 9, + 7 + ], + "selected_reward": -0.008671722374856472 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set at night with a sky thick with dark clouds and occasional flashes of pale lightning?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is an ancient wooden expedition vessel visible battling mountainous waves in an inky black sea?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is an explorer visible wearing a heavy, rain-soaked windbreaker with wet hair plastered to his forehead?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the explorer slip from the railing, fall backward, and go over the side of the boat into the waves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a medium-long shot from a slightly upward angle with a camera swaying precariously?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.008648558519780636 + ], + "opposite_logprob_tokens": [ + -4.75864839553833 + ], + "logprob_reward": -0.008648558519780636, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.008648558519780636, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.05460036173462868, + 0.021047769114375114, + 0.011917168274521828, + 0.009622073732316494, + 0.008557118475437164, + 0.02248052880167961, + 0.7756171226501465, + 0.04767993465065956, + 0.01723809726536274, + 0.031239774078130722 + ], + "credit_map": [ + 0.003127441508695483, + 0.005518458317965269, + 0.004192944150418043, + 0.0085871247574687, + 0.01437223982065916, + 0.00585767300799489, + 0.0083242729306221, + 0.004620211198925972, + 0.0014827288687229156, + 0.0017786656972020864, + 0.0010008885292336345, + 0.0023816395550966263, + 0.006159570999443531, + 0.003353728912770748, + 0.003045802703127265, + 0.001844744081608951, + 0.001040515722706914, + 0.0014439393999055028, + 0.001012579770758748, + 0.001891222782433033, + 0.0017123919678851962, + 0.0016983160749077797, + 0.0018883788725361228, + 0.0012298246147111058, + 0.0006646181573159993, + 0.001230768975801766, + 0.000807920761872083, + 0.001453014905564487, + 0.0019308143528178334, + 0.001243291306309402, + 0.0012731809401884675, + 0.001018464914523065, + 0.0009277503122575581, + 0.0010490583954378963, + 0.0008539878181181848, + 0.0010315842228010297, + 0.0011890701716765761, + 0.0008267687517218292, + 0.0011563437292352319, + 0.0015225561801344156, + 0.004107628017663956, + 0.0032143937423825264, + 0.002607146045193076, + 0.0025467134546488523, + 0.002210220554843545, + 0.00296147633343935, + 0.0022782294545322657, + 0.0025547242257744074, + 0.02098180539906025, + 0.0341731458902359, + 0.00787930004298687, + 0.03195030242204666, + 0.05183161795139313, + 0.2941162884235382, + 0.07482294738292694, + 0.2598617970943451, + 0.0012348664458841085, + 0.001570515800267458, + 0.0025380810257047415, + 0.005872601177543402, + 0.0023510842584073544, + 0.0024189893156290054, + 0.01812778040766716, + 0.013566019013524055, + 0.0013115854235365987, + 0.0030036314856261015, + 0.00217252504080534, + 0.0018331167520955205, + 0.002695648465305567, + 0.0027153976261615753, + 0.0015607289969921112, + 0.0019454644061625004, + 0.002862101187929511, + 0.0019857860170304775, + 0.00846988707780838, + 0.0023704024497419596, + 0.0028957934118807316, + 0.0034500169567763805, + 0.003108944045379758, + 0.006096845027059317 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 0, + 7 + ], + "selected_reward": -0.008648558519780636 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set at night with a sky thick with dark clouds and occasional flashes of pale lightning?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is an ancient wooden expedition vessel visible battling mountainous waves in an inky black sea?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is an explorer visible wearing a heavy, rain-soaked windbreaker with wet hair plastered to his forehead?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the explorer slip from the railing, fall backward, and go over the side of the boat into the waves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a medium-long shot from a slightly upward angle with a camera swaying precariously?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.016069892793893814 + ], + "opposite_logprob_tokens": [ + -4.1410698890686035 + ], + "logprob_reward": -0.016069892793893814, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.016069892793893814, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07835067063570023, + 0.05898910015821457, + 0.04564540460705757, + 0.03281054645776749, + 0.03399798646569252, + 0.0294897872954607, + 0.21546870470046997, + 0.19941730797290802, + 0.13235126435756683, + 0.17347924411296844 + ], + "credit_map": [ + 0.004529220517724752, + 0.010935232043266296, + 0.007421575486660004, + 0.013458139263093472, + 0.011904469691216946, + 0.015436858870089054, + 0.008453658781945705, + 0.006211514119058847, + 0.00400786567479372, + 0.00871667917817831, + 0.0029270234517753124, + 0.008295451290905476, + 0.009652464650571346, + 0.009591457433998585, + 0.007728815544396639, + 0.008069339208304882, + 0.003405337454751134, + 0.009014410898089409, + 0.0037204294931143522, + 0.006796366069465876, + 0.005590921267867088, + 0.006838494911789894, + 0.004928742069751024, + 0.005350704770535231, + 0.002137863775715232, + 0.0038868486881256104, + 0.0034016079735010862, + 0.005155178718268871, + 0.0030336789786815643, + 0.006067813839763403, + 0.00635030260309577, + 0.002777252346277237, + 0.002553743775933981, + 0.00442934688180685, + 0.004453463479876518, + 0.004669490735977888, + 0.004861999768763781, + 0.0033946966286748648, + 0.004926370922476053, + 0.004708877764642239, + 0.004510459955781698, + 0.003611193969845772, + 0.0022589671425521374, + 0.003689610632136464, + 0.004123303573578596, + 0.0032387543469667435, + 0.003984886221587658, + 0.0040726130828261375, + 0.008534658700227737, + 0.0285266675055027, + 0.006407026667147875, + 0.022458849474787712, + 0.025104541331529617, + 0.05056203156709671, + 0.021361976861953735, + 0.05251294746994972, + 0.013822581619024277, + 0.011741957627236843, + 0.015353767201304436, + 0.05141661316156387, + 0.01828838512301445, + 0.012725120410323143, + 0.05155128240585327, + 0.024517599493265152, + 0.009327908046543598, + 0.02691124565899372, + 0.014457658864557743, + 0.009862746112048626, + 0.019235163927078247, + 0.031122839078307152, + 0.011168324388563633, + 0.010265377350151539, + 0.015373745933175087, + 0.02246486023068428, + 0.02590198628604412, + 0.027323797345161438, + 0.018370406702160835, + 0.022228753194212914, + 0.026601817458868027, + 0.015213866718113422 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 7, + 9 + ], + "selected_reward": -0.016069892793893814 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set at night with a sky thick with dark clouds and occasional flashes of pale lightning?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is an ancient wooden expedition vessel visible battling mountainous waves in an inky black sea?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is an explorer visible wearing a heavy, rain-soaked windbreaker with wet hair plastered to his forehead?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the explorer slip from the railing, fall backward, and go over the side of the boat into the waves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a medium-long shot from a slightly upward angle with a camera swaying precariously?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01607341133058071 + ], + "opposite_logprob_tokens": [ + -4.141073226928711 + ], + "logprob_reward": -0.01607341133058071, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.01607341133058071, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0543447770178318, + 0.03296661004424095, + 0.03011023998260498, + 0.024600738659501076, + 0.02381082996726036, + 0.031691376119852066, + 0.4955753982067108, + 0.08390892297029495, + 0.1440320611000061, + 0.07895907759666443 + ], + "credit_map": [ + 0.00453645596280694, + 0.007936425507068634, + 0.004740619100630283, + 0.008161009289324284, + 0.006857346277683973, + 0.006880633067339659, + 0.006395161151885986, + 0.00883712712675333, + 0.002081741811707616, + 0.004019014537334442, + 0.00174937944393605, + 0.0050195311196148396, + 0.007086292374879122, + 0.005589230451732874, + 0.004675090778619051, + 0.0027463280130177736, + 0.002384419785812497, + 0.0059653655625879765, + 0.0025133732706308365, + 0.003775459947064519, + 0.003829686902463436, + 0.004489376675337553, + 0.0035501690581440926, + 0.0036023883149027824, + 0.0016891669947654009, + 0.0026914416812360287, + 0.00211262796074152, + 0.004848800599575043, + 0.0028447003569453955, + 0.0036266078241169453, + 0.0038040289655327797, + 0.002983363578096032, + 0.0031739105470478535, + 0.0030165803618729115, + 0.0028695662040263414, + 0.003195166354998946, + 0.00348899164237082, + 0.0025824042968451977, + 0.0026490462478250265, + 0.002835165709257126, + 0.0029105341527611017, + 0.0032805083319544792, + 0.0029342309571802616, + 0.00339964940212667, + 0.007421195041388273, + 0.003830968402326107, + 0.004634870681911707, + 0.0032794212456792593, + 0.013455313630402088, + 0.02211230807006359, + 0.004961851984262466, + 0.020267700776457787, + 0.033695705235004425, + 0.18841055035591125, + 0.047567810863256454, + 0.16510413587093353, + 0.004240831825882196, + 0.002894014585763216, + 0.004980023950338364, + 0.010011150501668453, + 0.003868675557896495, + 0.003808144712820649, + 0.02112300880253315, + 0.03298307582736015, + 0.008467253297567368, + 0.03517559915781021, + 0.011071269400417805, + 0.009178966283798218, + 0.015887759625911713, + 0.04547683894634247, + 0.006657571531832218, + 0.012116800993680954, + 0.008548353798687458, + 0.013166680932044983, + 0.00983893871307373, + 0.014533953741192818, + 0.008795652538537979, + 0.008930266834795475, + 0.008711275644600391, + 0.006433958187699318 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 8, + 7 + ], + "selected_reward": -0.01607341133058071 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000101/group_7/index_tensor([5650, 5650])_group_7_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000101/group_7/index_tensor([5650, 5650])_group_7_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..f7717b7680d2f3b9584662307c1c279d57e930c5 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000101/group_7/index_tensor([5650, 5650])_group_7_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "stark, almost static, documentary-style close-up shot in stark black and white captures the delicate interaction between two dahlias in full bloom. The lighting is high-contrast, as if coming from a single, focused source, sculpting the subjects in dramatic black and white. The primary flower's countless layers of spade-shaped petals form a perfect, intricate spiral of brilliant ivory, with deep, velvety charcoal shadows between each petal, emphasizing its complex geometry and texture. The background is a soft, out-of-focus bokeh composed of muted grays and deep blacks, suggesting a garden setting reduced to abstract forms. The first flower sways slowly and lazily, lifted by an imperceptible breeze. Its entire head gracefully tilts on its stem, allowing the edge of its outermost petal to gently brush the surface of the second, still flower. This micro-movement is a silent, slow dance; the petals trace a gentle, fleeting path across their partner, a contact so light as to barely deform their delicate structure, creating a profound sense of intimacy and a tranquil, natural poetry.", + "seed": 4642, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000101/group_7/index_5650_seed_4642_dp_14_rank_28_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 7, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.07065605046227574, + "vlm_logprob": -0.07065605046227574, + "vlm_yes_no_margin": 3.525, + "vlm_token_credit": -0.07065605046227574, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video presented in black and white with high-contrast lighting?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are two dahlias in full bloom visible as the primary subjects?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the first dahlia have ivory petals arranged in a spiral pattern?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the first dahlia sway and tilt to brush its petal against the second dahlia?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a soft, out-of-focus bokeh of muted grays and deep blacks?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.061588600277900696 + ], + "opposite_logprob_tokens": [ + -2.9365885257720947 + ], + "logprob_reward": -0.061588600277900696, + "yes_no_margin_reward": 2.875, + "token_credit_reward": -0.061588600277900696, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11101182550191879, + 0.3476695120334625, + 0.057063356041908264, + 0.08812510222196579, + 0.2809045612812042, + 0.03369368985295296, + 0.02698487788438797, + 0.022670965641736984, + 0.010669911280274391, + 0.02120617963373661 + ], + "credit_map": [ + 0.01146640907973051, + 0.0137846190482378, + 0.020441176369786263, + 0.01572442427277565, + 0.01125046331435442, + 0.01815369538962841, + 0.009560899809002876, + 0.010630134493112564, + 0.016908198595046997, + 0.013681017793715, + 0.1204470545053482, + 0.02327617257833481, + 0.015642140060663223, + 0.12502998113632202, + 0.012476254254579544, + 0.02020866982638836, + 0.0040325080044567585, + 0.002969075459986925, + 0.011007631197571754, + 0.008008789271116257, + 0.003342197509482503, + 0.01180369034409523, + 0.005758205428719521, + 0.010141254402697086, + 0.0065526883117854595, + 0.013373593799769878, + 0.01535947248339653, + 0.009757124818861485, + 0.007576047908514738, + 0.011052723042666912, + 0.010746432468295097, + 0.013707018457353115, + 0.02117645926773548, + 0.01966564729809761, + 0.04647991433739662, + 0.025016924366354942, + 0.01810085028409958, + 0.045517392456531525, + 0.044320814311504364, + 0.060626544058322906, + 0.002676364267244935, + 0.004231294617056847, + 0.006066479720175266, + 0.0028030420653522015, + 0.0031384755857288837, + 0.004372432362288237, + 0.0050668865442276, + 0.005338709335774183, + 0.00191568024456501, + 0.002591505413874984, + 0.0033112638629972935, + 0.0035488398279994726, + 0.00259311287663877, + 0.003145775757730007, + 0.006172433495521545, + 0.0037062647752463818, + 0.00166424794588238, + 0.0017197821289300919, + 0.002667514607310295, + 0.002870704047381878, + 0.002435246715322137, + 0.0026648682542145252, + 0.005809430964291096, + 0.0028391703963279724, + 0.0005487148300744593, + 0.0007213651551865041, + 0.0011263741180300713, + 0.001511674840003252, + 0.0011322317877784371, + 0.0015439728740602732, + 0.0028867945075035095, + 0.00119878223631531, + 0.001666670199483633, + 0.0019250490004196763, + 0.0017619599821045995, + 0.005633070599287748, + 0.002591676777228713, + 0.0028538075275719166, + 0.0032684162724763155, + 0.0015055289259180427 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 4, + 0 + ], + "selected_reward": -0.061588600277900696 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video presented in black and white with high-contrast lighting?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are two dahlias in full bloom visible as the primary subjects?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the first dahlia have ivory petals arranged in a spiral pattern?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the first dahlia sway and tilt to brush its petal against the second dahlia?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a soft, out-of-focus bokeh of muted grays and deep blacks?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07899043709039688 + ], + "opposite_logprob_tokens": [ + -2.5789904594421387 + ], + "logprob_reward": -0.07899043709039688, + "yes_no_margin_reward": 2.5, + "token_credit_reward": -0.07899043709039688, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.23795631527900696, + 0.277178019285202, + 0.029031816869974136, + 0.032101213932037354, + 0.08120749145746231, + 0.14055553078651428, + 0.034243326634168625, + 0.04146689549088478, + 0.018604114651679993, + 0.10765530914068222 + ], + "credit_map": [ + 0.020576125010848045, + 0.019636673852801323, + 0.07182330638170242, + 0.020306536927819252, + 0.017167845740914345, + 0.05202808603644371, + 0.014826135709881783, + 0.02159159444272518, + 0.01295708492398262, + 0.007041720673441887, + 0.10253579914569855, + 0.015088203363120556, + 0.009691180661320686, + 0.10579358041286469, + 0.008781095035374165, + 0.01528932061046362, + 0.002252998761832714, + 0.0027292973827570677, + 0.004531822632998228, + 0.003840399207547307, + 0.0033982261084020138, + 0.0032762300688773394, + 0.004851463716477156, + 0.004151379689574242, + 0.0025100901257246733, + 0.004068873822689056, + 0.005748324561864138, + 0.003372775623574853, + 0.003141137072816491, + 0.004858619067817926, + 0.00345165329053998, + 0.004949741996824741, + 0.008179849945008755, + 0.006854348350316286, + 0.01228802464902401, + 0.006795428693294525, + 0.00614821957424283, + 0.011316712014377117, + 0.012917006388306618, + 0.016707908362150192, + 0.01111901830881834, + 0.01759428158402443, + 0.023120352998375893, + 0.014409792609512806, + 0.01227616984397173, + 0.019173316657543182, + 0.01931406371295452, + 0.023548519238829613, + 0.0029451879672706127, + 0.0037473358679562807, + 0.004591230768710375, + 0.004808164667338133, + 0.0030999304726719856, + 0.003972673323005438, + 0.006804272066801786, + 0.004274529404938221, + 0.003048467915505171, + 0.003954233601689339, + 0.004686084575951099, + 0.004148457199335098, + 0.0041318777948617935, + 0.00965449120849371, + 0.006537573877722025, + 0.005305708851665258, + 0.0010483701480552554, + 0.0016863172641023993, + 0.0033869300968945026, + 0.0026213808450847864, + 0.002218774287030101, + 0.002490021288394928, + 0.0032952288165688515, + 0.0018570933025330305, + 0.00881686806678772, + 0.01060220506042242, + 0.01224326342344284, + 0.016869286075234413, + 0.014224084094166756, + 0.015847837552428246, + 0.01959274895489216, + 0.00945900846272707 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 5 + ], + "selected_reward": -0.07899043709039688 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video presented in black and white with high-contrast lighting?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are two dahlias in full bloom visible as the primary subjects?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the first dahlia have ivory petals arranged in a spiral pattern?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the first dahlia sway and tilt to brush its petal against the second dahlia?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a soft, out-of-focus bokeh of muted grays and deep blacks?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.008644777350127697 + ], + "opposite_logprob_tokens": [ + -4.7586445808410645 + ], + "logprob_reward": -0.008644777350127697, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.008644777350127697, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15532812476158142, + 0.22297030687332153, + 0.07831496000289917, + 0.1332188993692398, + 0.08658474683761597, + 0.1485801488161087, + 0.05681849271059036, + 0.06748384237289429, + 0.011602863669395447, + 0.03909764438867569 + ], + "credit_map": [ + 0.012750687077641487, + 0.009629541076719761, + 0.04231186583638191, + 0.019318826496601105, + 0.012767254374921322, + 0.030860373750329018, + 0.014401753433048725, + 0.013287830166518688, + 0.011499816551804543, + 0.012200203724205494, + 0.07278625667095184, + 0.01551151368767023, + 0.014482674188911915, + 0.07463351637125015, + 0.008702754974365234, + 0.013153559528291225, + 0.005942537449300289, + 0.0065641021355986595, + 0.012725302949547768, + 0.012566020712256432, + 0.007335054222494364, + 0.016176044940948486, + 0.009953223168849945, + 0.007052670698612928, + 0.011107063852250576, + 0.021546073257923126, + 0.023769011721014977, + 0.014381365850567818, + 0.01172384899109602, + 0.016261303797364235, + 0.014134328812360764, + 0.020295899361371994, + 0.006377301178872585, + 0.0050489576533436775, + 0.012683382257819176, + 0.009479689411818981, + 0.005743206944316626, + 0.011585920117795467, + 0.01889163628220558, + 0.016774650663137436, + 0.011870019137859344, + 0.01886937767267227, + 0.02661304548382759, + 0.013160783797502518, + 0.01427630614489317, + 0.018245799466967583, + 0.021176578477025032, + 0.02436824142932892, + 0.004721831530332565, + 0.005215242505073547, + 0.007580315228551626, + 0.0069924513809382915, + 0.005156963597983122, + 0.0074945781379938126, + 0.013003092259168625, + 0.006654014810919762, + 0.004259034525603056, + 0.0054230112582445145, + 0.007454826962202787, + 0.011405336670577526, + 0.006082851439714432, + 0.008988304063677788, + 0.015969522297382355, + 0.007900954224169254, + 0.0010127585846930742, + 0.0010631814366206527, + 0.0020550419576466084, + 0.0023545383010059595, + 0.0009298502700403333, + 0.0015661371871829033, + 0.0015153433196246624, + 0.0011060124961659312, + 0.00326921371743083, + 0.003560487413778901, + 0.00571301719173789, + 0.008092560805380344, + 0.0044858455657958984, + 0.005449877120554447, + 0.005711103789508343, + 0.002815539250150323 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 5 + ], + "selected_reward": -0.008644777350127697 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video presented in black and white with high-contrast lighting?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are two dahlias in full bloom visible as the primary subjects?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the first dahlia have ivory petals arranged in a spiral pattern?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the first dahlia sway and tilt to brush its petal against the second dahlia?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a soft, out-of-focus bokeh of muted grays and deep blacks?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.20156604051589966 + ], + "opposite_logprob_tokens": [ + -1.7015659809112549 + ], + "logprob_reward": -0.20156604051589966, + "yes_no_margin_reward": 1.5, + "token_credit_reward": -0.20156604051589966, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06593357026576996, + 0.4226125180721283, + 0.015872392803430557, + 0.07165838032960892, + 0.07129352539777756, + 0.21463747322559357, + 0.02560342289507389, + 0.037928126752376556, + 0.02192770503461361, + 0.0525328665971756 + ], + "credit_map": [ + 0.010003017261624336, + 0.007220113649964333, + 0.008775968104600906, + 0.012147031724452972, + 0.006655069533735514, + 0.005606926046311855, + 0.007331484463065863, + 0.008193962275981903, + 0.02118144929409027, + 0.02044759690761566, + 0.1408129632472992, + 0.029182979837059975, + 0.02390599250793457, + 0.143134206533432, + 0.01781458407640457, + 0.026132794097065926, + 0.0011384005192667246, + 0.0012640057830139995, + 0.002725838217884302, + 0.0017759129405021667, + 0.0013977327616885304, + 0.0030321122612804174, + 0.0023035891354084015, + 0.0022348021157085896, + 0.005950177554041147, + 0.010418636724352837, + 0.012460737489163876, + 0.008488035760819912, + 0.005796022247523069, + 0.009816776029765606, + 0.00739017128944397, + 0.011337834410369396, + 0.004851505160331726, + 0.006064570043236017, + 0.011430344544351101, + 0.005920379888266325, + 0.0053297970443964005, + 0.011893348768353462, + 0.012260944582521915, + 0.013542636297643185, + 0.01644318923354149, + 0.025158843025565147, + 0.03780776634812355, + 0.02166869305074215, + 0.018825270235538483, + 0.02754361554980278, + 0.030683990567922592, + 0.03650611266493797, + 0.0020338110625743866, + 0.00220692646689713, + 0.0034250784665346146, + 0.0025841728784143925, + 0.00230458308942616, + 0.004105559550225735, + 0.005824084859341383, + 0.0031192083843052387, + 0.002372748451307416, + 0.0022796366829425097, + 0.0055586714297533035, + 0.004133484326303005, + 0.0034546044189482927, + 0.005025121849030256, + 0.008744033053517342, + 0.006359829101711512, + 0.0021591330878436565, + 0.003030552063137293, + 0.003037108341231942, + 0.0034824807662516832, + 0.0018820170080289245, + 0.002721989993005991, + 0.0029109595343470573, + 0.002703465986996889, + 0.0036621959879994392, + 0.005054349545389414, + 0.003709845943376422, + 0.012658045627176762, + 0.006637936923652887, + 0.006670661736279726, + 0.009387711994349957, + 0.004752126056700945 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 5, + 3 + ], + "selected_reward": -0.20156604051589966 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video presented in black and white with high-contrast lighting?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are two dahlias in full bloom visible as the primary subjects?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the first dahlia have ivory petals arranged in a spiral pattern?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the first dahlia sway and tilt to brush its petal against the second dahlia?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a soft, out-of-focus bokeh of muted grays and deep blacks?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0024903970770537853 + ], + "opposite_logprob_tokens": [ + -6.002490520477295 + ], + "logprob_reward": -0.0024903970770537853, + "yes_no_margin_reward": 6.0, + "token_credit_reward": -0.0024903970770537853, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07594717293977737, + 0.05005820840597153, + 0.0449543297290802, + 0.325742244720459, + 0.2603370249271393, + 0.057874225080013275, + 0.036208853125572205, + 0.10932645201683044, + 0.00931953638792038, + 0.030231857672333717 + ], + "credit_map": [ + 0.008757246658205986, + 0.009734042920172215, + 0.012925483286380768, + 0.01583857089281082, + 0.008326270617544651, + 0.007444491144269705, + 0.00633666105568409, + 0.006584420334547758, + 0.0026524777058511972, + 0.0026960251852869987, + 0.013615547679364681, + 0.004858660977333784, + 0.0031865171622484922, + 0.014941280707716942, + 0.003736441256478429, + 0.00437126774340868, + 0.004273459315299988, + 0.004679420962929726, + 0.009727705270051956, + 0.004415606148540974, + 0.0036424242425709963, + 0.007048871833831072, + 0.0056902626529335976, + 0.005476584192365408, + 0.026449086144566536, + 0.04798898473381996, + 0.058961477130651474, + 0.03772405534982681, + 0.026430951431393623, + 0.045268621295690536, + 0.03370112553238869, + 0.04921797662973404, + 0.01771649532020092, + 0.02021150290966034, + 0.042475804686546326, + 0.024381158873438835, + 0.020086681470274925, + 0.0451236292719841, + 0.036602944135665894, + 0.05373886600136757, + 0.004586473107337952, + 0.007092475891113281, + 0.010503347963094711, + 0.005535291042178869, + 0.0050665996968746185, + 0.007326297927647829, + 0.008515889756381512, + 0.009247858077287674, + 0.0023472104221582413, + 0.0028590599540621042, + 0.005257750395685434, + 0.004592833109200001, + 0.0032337834127247334, + 0.004618289414793253, + 0.007947666570544243, + 0.005352267064154148, + 0.009813153184950352, + 0.00684125954285264, + 0.013911942020058632, + 0.008017529733479023, + 0.010755925439298153, + 0.016249898821115494, + 0.023175040259957314, + 0.020561715587973595, + 0.0006779605173505843, + 0.001013805391266942, + 0.001194989774376154, + 0.0016317005502060056, + 0.0007673433283343911, + 0.0015472471714019775, + 0.0014318933244794607, + 0.0010545970872044563, + 0.0026102527044713497, + 0.002471416722983122, + 0.003507317043840885, + 0.005248706322163343, + 0.003906466998159885, + 0.004384745378047228, + 0.004857360385358334, + 0.003245595609769225 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 4, + 7 + ], + "selected_reward": -0.0024903970770537853 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "stark, almost static, documentary-style close-up shot in stark black and white captures the delicate interaction between two dahlias in full bloom. The lighting is high-contrast, as if coming from a single, focused source, sculpting the subjects in dramatic black and white. The primary flower's countless layers of spade-shaped petals form a perfect, intricate spiral of brilliant ivory, with deep, velvety charcoal shadows between each petal, emphasizing its complex geometry and texture. The background is a soft, out-of-focus bokeh composed of muted grays and deep blacks, suggesting a garden setting reduced to abstract forms. The first flower sways slowly and lazily, lifted by an imperceptible breeze. Its entire head gracefully tilts on its stem, allowing the edge of its outermost petal to gently brush the surface of the second, still flower. This micro-movement is a silent, slow dance; the petals trace a gentle, fleeting path across their partner, a contact so light as to barely deform their delicate structure, creating a profound sense of intimacy and a tranquil, natural poetry.", + "seed": 4643, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000101/group_7/index_5650_seed_4643_dp_14_rank_28_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 7, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.1882819103077054, + "vlm_logprob": -0.1882819103077054, + "vlm_yes_no_margin": 2.85, + "vlm_token_credit": -0.1882819103077054, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video presented in black and white with high-contrast lighting?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are two dahlias in full bloom visible as the primary subjects?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the first dahlia have ivory petals arranged in a spiral pattern?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the first dahlia sway and tilt to brush its petal against the second dahlia?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a soft, out-of-focus bokeh of muted grays and deep blacks?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.011404914781451225 + ], + "opposite_logprob_tokens": [ + -5.011404991149902 + ], + "logprob_reward": -0.011404914781451225, + "yes_no_margin_reward": 5.0, + "token_credit_reward": -0.011404914781451225, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.185504749417305, + 0.1076713353395462, + 0.09257204830646515, + 0.06400629132986069, + 0.02564414218068123, + 0.14103548228740692, + 0.11204928904771805, + 0.2070055603981018, + 0.019144665449857712, + 0.04536635801196098 + ], + "credit_map": [ + 0.021486081182956696, + 0.011101257055997849, + 0.02697811834514141, + 0.01599614880979061, + 0.02188451588153839, + 0.025920839980244637, + 0.01824355311691761, + 0.04389426112174988, + 0.013635188341140747, + 0.0065119205974042416, + 0.013192900456488132, + 0.012806703336536884, + 0.006223995238542557, + 0.011495406739413738, + 0.024805499240756035, + 0.018999729305505753, + 0.011129249818623066, + 0.004109655506908894, + 0.013340895064175129, + 0.008599258959293365, + 0.0064460705034434795, + 0.015226265415549278, + 0.01042942889034748, + 0.023291241377592087, + 0.0072905197739601135, + 0.0048324125818908215, + 0.01039142906665802, + 0.009118273854255676, + 0.006745171267539263, + 0.008540519513189793, + 0.009979799389839172, + 0.007108176127076149, + 0.003849729197099805, + 0.0014132782816886902, + 0.004652506671845913, + 0.0019512176513671875, + 0.0019436342408880591, + 0.0029897214844822884, + 0.00455622561275959, + 0.004287833347916603, + 0.025903185829520226, + 0.0037201836239546537, + 0.05099008232355118, + 0.00980426836758852, + 0.006250636652112007, + 0.011583764106035233, + 0.00953584723174572, + 0.023247532546520233, + 0.018661560490727425, + 0.0038602908607572317, + 0.03053702972829342, + 0.006700634956359863, + 0.008758878335356712, + 0.011718818917870522, + 0.012086519971489906, + 0.019725576043128967, + 0.018206290900707245, + 0.003467854345217347, + 0.07493594288825989, + 0.008575006388127804, + 0.005415456369519234, + 0.0667414665222168, + 0.012464988976716995, + 0.017198577523231506, + 0.0024961931630969048, + 0.0011389601277187467, + 0.0033510036300867796, + 0.0027437724638730288, + 0.0015434075612574816, + 0.002411029301583767, + 0.0034332810901105404, + 0.002027021022513509, + 0.0046278201043605804, + 0.0021170673426240683, + 0.011252363212406635, + 0.0028781606815755367, + 0.0038213494699448347, + 0.008416826836764812, + 0.004867778159677982, + 0.0073849959298968315 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 0, + 5 + ], + "selected_reward": -0.011404914781451225 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video presented in black and white with high-contrast lighting?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are two dahlias in full bloom visible as the primary subjects?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the first dahlia have ivory petals arranged in a spiral pattern?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the first dahlia sway and tilt to brush its petal against the second dahlia?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a soft, out-of-focus bokeh of muted grays and deep blacks?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.5233480334281921 + ], + "opposite_logprob_tokens": [ + -0.8983480334281921 + ], + "logprob_reward": -0.5233480334281921, + "yes_no_margin_reward": 0.375, + "token_credit_reward": -0.5233480334281921, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07979070395231247, + 0.0648965835571289, + 0.09168226271867752, + 0.03838970139622688, + 0.032582540065050125, + 0.1207447275519371, + 0.05301746353507042, + 0.14347679913043976, + 0.1070822924375534, + 0.2683368921279907 + ], + "credit_map": [ + 0.008758663199841976, + 0.006154266651719809, + 0.011821739375591278, + 0.008259841240942478, + 0.0080939382314682, + 0.010524312034249306, + 0.008336192928254604, + 0.017841750755906105, + 0.009718443267047405, + 0.0038362683262676, + 0.0092667480930686, + 0.008612483739852905, + 0.003603628370910883, + 0.008273368701338768, + 0.013162737712264061, + 0.008422906510531902, + 0.012603062205016613, + 0.004827944561839104, + 0.010007691569626331, + 0.012774741277098656, + 0.007567120715975761, + 0.008165786974132061, + 0.024049844592809677, + 0.011686081066727638, + 0.005320990923792124, + 0.00223200349137187, + 0.004695029463618994, + 0.004744368139654398, + 0.0033190674148499966, + 0.004372735507786274, + 0.0072495173662900925, + 0.006455990951508284, + 0.005280118435621262, + 0.0016085276147350669, + 0.005278699565678835, + 0.0024877057876437902, + 0.0036977005656808615, + 0.0030201177578419447, + 0.005913048982620239, + 0.0052966224029660225, + 0.016603054478764534, + 0.003398875705897808, + 0.03725609555840492, + 0.009335985407233238, + 0.012252825312316418, + 0.008316651917994022, + 0.012573285028338432, + 0.021007953211665154, + 0.011122913099825382, + 0.0017412101151421666, + 0.01434202492237091, + 0.002885232213884592, + 0.0016502940561622381, + 0.005433754529803991, + 0.005579517222940922, + 0.010262515395879745, + 0.012490412220358849, + 0.0025030856486409903, + 0.05409665405750275, + 0.0058347806334495544, + 0.004030613694339991, + 0.04528975114226341, + 0.009622753597795963, + 0.009608748368918896, + 0.010395122691988945, + 0.0049313087947666645, + 0.026393886655569077, + 0.009706597775220871, + 0.007525724358856678, + 0.014185834676027298, + 0.01815962605178356, + 0.01578419841825962, + 0.028223557397723198, + 0.012070666067302227, + 0.05937740206718445, + 0.018607497215270996, + 0.021790752187371254, + 0.06244289129972458, + 0.028713146224617958, + 0.03711098060011864 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 7, + 5 + ], + "selected_reward": -0.5233480334281921 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video presented in black and white with high-contrast lighting?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are two dahlias in full bloom visible as the primary subjects?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the first dahlia have ivory petals arranged in a spiral pattern?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the first dahlia sway and tilt to brush its petal against the second dahlia?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a soft, out-of-focus bokeh of muted grays and deep blacks?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.08909235149621964 + ], + "opposite_logprob_tokens": [ + -2.464092254638672 + ], + "logprob_reward": -0.08909235149621964, + "yes_no_margin_reward": 2.375, + "token_credit_reward": -0.08909235149621964, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13188791275024414, + 0.11491801589727402, + 0.12661395967006683, + 0.05319484323263168, + 0.028531843796372414, + 0.04559694230556488, + 0.01848544366657734, + 0.39582499861717224, + 0.01666981354355812, + 0.0682762935757637 + ], + "credit_map": [ + 0.017032630741596222, + 0.009937331080436707, + 0.018752655014395714, + 0.011363061144948006, + 0.015231866389513016, + 0.012153256684541702, + 0.021775880828499794, + 0.025641225278377533, + 0.01573840342462063, + 0.007407249417155981, + 0.01101959589868784, + 0.01417467463761568, + 0.006692943163216114, + 0.015080881305038929, + 0.027996836230158806, + 0.01680743508040905, + 0.014477378688752651, + 0.006863144692033529, + 0.0142851322889328, + 0.01561801228672266, + 0.01149359904229641, + 0.01593797281384468, + 0.02430201880633831, + 0.023636698722839355, + 0.0056696487590670586, + 0.003361827926710248, + 0.00893592182546854, + 0.00668101804330945, + 0.0039430707693099976, + 0.008707650937139988, + 0.008266184478998184, + 0.00762952072545886, + 0.0042315153405070305, + 0.0013105619000270963, + 0.0043528713285923, + 0.0025700926780700684, + 0.0021919298451393843, + 0.002679771278053522, + 0.007668742910027504, + 0.003526357701048255, + 0.008364475332200527, + 0.001751169329509139, + 0.011331785470247269, + 0.0030314053874462843, + 0.002233891747891903, + 0.004165066406130791, + 0.004916498903185129, + 0.009802653454244137, + 0.0022266500163823366, + 0.0011009350419044495, + 0.002549014752730727, + 0.001323031377978623, + 0.002848840318620205, + 0.0016458133468404412, + 0.0038853834848850965, + 0.0029057750944048166, + 0.03252226114273071, + 0.006030535325407982, + 0.1476677507162094, + 0.01694108545780182, + 0.010017657652497292, + 0.1288784295320511, + 0.025203388184309006, + 0.028563885018229485, + 0.002066300017759204, + 0.0008630803786218166, + 0.004390493500977755, + 0.0013287448091432452, + 0.0010505529353395104, + 0.002544839633628726, + 0.0027445198502391577, + 0.0016812838148325682, + 0.0072191813960671425, + 0.0031403345055878162, + 0.016269199550151825, + 0.0042034643702209, + 0.005498955026268959, + 0.012852487154304981, + 0.008027225732803345, + 0.011065452359616756 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 0, + 2 + ], + "selected_reward": -0.08909235149621964 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video presented in black and white with high-contrast lighting?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are two dahlias in full bloom visible as the primary subjects?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the first dahlia have ivory petals arranged in a spiral pattern?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the first dahlia sway and tilt to brush its petal against the second dahlia?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a soft, out-of-focus bokeh of muted grays and deep blacks?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.31346791982650757 + ], + "opposite_logprob_tokens": [ + -1.3134679794311523 + ], + "logprob_reward": -0.31346791982650757, + "yes_no_margin_reward": 1.0, + "token_credit_reward": -0.31346791982650757, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0606355145573616, + 0.058846160769462585, + 0.08902281522750854, + 0.13975024223327637, + 0.051131345331668854, + 0.16502971947193146, + 0.10758121311664581, + 0.14168916642665863, + 0.05799422785639763, + 0.1283196359872818 + ], + "credit_map": [ + 0.007185860071331263, + 0.005331459455192089, + 0.008092238567769527, + 0.004113269969820976, + 0.005130430217832327, + 0.01144363172352314, + 0.006384172476828098, + 0.012954451143741608, + 0.00730055570602417, + 0.0037213508039712906, + 0.008504008874297142, + 0.006731975823640823, + 0.004484050441533327, + 0.008286602795124054, + 0.01038689911365509, + 0.00943071860820055, + 0.010920483618974686, + 0.003349598962813616, + 0.010835975408554077, + 0.0056197550147771835, + 0.006912317592650652, + 0.012946927919983864, + 0.01479413453489542, + 0.023643624037504196, + 0.016584603115916252, + 0.006078039295971394, + 0.017320331186056137, + 0.010150791145861149, + 0.0158323273062706, + 0.030751390382647514, + 0.014873198233544827, + 0.02815956249833107, + 0.006681984756141901, + 0.0032040332444012165, + 0.006728327367454767, + 0.004314481746405363, + 0.008833187632262707, + 0.005686210002750158, + 0.007828318513929844, + 0.007854803465306759, + 0.03087541088461876, + 0.004321797750890255, + 0.06336785852909088, + 0.011950887739658356, + 0.0057686446234583855, + 0.013491206802427769, + 0.015052977949380875, + 0.02020094357430935, + 0.01273816917091608, + 0.003769197501242161, + 0.02237102761864662, + 0.008283204399049282, + 0.011789112351834774, + 0.009280500933527946, + 0.017432553693652153, + 0.021917447447776794, + 0.02289925143122673, + 0.006309135816991329, + 0.023773930966854095, + 0.00963066890835762, + 0.011160774156451225, + 0.02119174599647522, + 0.02549527958035469, + 0.02122839167714119, + 0.006620038766413927, + 0.002280599670484662, + 0.018786195665597916, + 0.004559047985821962, + 0.003862757934257388, + 0.0071828896179795265, + 0.008388836868107319, + 0.006313857156783342, + 0.013780124485492706, + 0.005738904234021902, + 0.027249876409769058, + 0.010101279243826866, + 0.010377984493970871, + 0.02786168083548546, + 0.015872711315751076, + 0.017337078228592873 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 7, + 3 + ], + "selected_reward": -0.31346791982650757 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video presented in black and white with high-contrast lighting?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are two dahlias in full bloom visible as the primary subjects?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the first dahlia have ivory petals arranged in a spiral pattern?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the first dahlia sway and tilt to brush its petal against the second dahlia?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a soft, out-of-focus bokeh of muted grays and deep blacks?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0040963320061564445 + ], + "opposite_logprob_tokens": [ + -5.504096508026123 + ], + "logprob_reward": -0.0040963320061564445, + "yes_no_margin_reward": 5.5, + "token_credit_reward": -0.0040963320061564445, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.110328808426857, + 0.11833243072032928, + 0.0917663648724556, + 0.057162731885910034, + 0.1860015094280243, + 0.1319001168012619, + 0.0644490197300911, + 0.04763783514499664, + 0.07086624950170517, + 0.1215549185872078 + ], + "credit_map": [ + 0.014334951527416706, + 0.010404420085251331, + 0.013691372238099575, + 0.009624731726944447, + 0.013584830798208714, + 0.013181835412979126, + 0.01863081194460392, + 0.0168758574873209, + 0.016845010221004486, + 0.005102862138301134, + 0.01600458100438118, + 0.008057192899286747, + 0.006734068039804697, + 0.017316218465566635, + 0.022734563797712326, + 0.02553793415427208, + 0.012756554409861565, + 0.0053791445679962635, + 0.011334062553942204, + 0.01039425190538168, + 0.0074407076463103294, + 0.01033785380423069, + 0.021198267117142677, + 0.012925518676638603, + 0.008099706843495369, + 0.0034047248773276806, + 0.010851998813450336, + 0.006692373659461737, + 0.004438296426087618, + 0.007237638812512159, + 0.008555758744478226, + 0.007882234640419483, + 0.024927355349063873, + 0.007492860779166222, + 0.026037372648715973, + 0.013264982029795647, + 0.02526453696191311, + 0.016141103580594063, + 0.03678188845515251, + 0.036091405898332596, + 0.025079671293497086, + 0.0036632553674280643, + 0.05019956827163696, + 0.009777223691344261, + 0.005029154010117054, + 0.011033703573048115, + 0.01206926815211773, + 0.015048273839056492, + 0.011805886402726173, + 0.003274779999628663, + 0.006824194453656673, + 0.004709014669060707, + 0.0058493115939199924, + 0.005747088231146336, + 0.010670868679881096, + 0.01556787732988596, + 0.012016726657748222, + 0.002330464543774724, + 0.006157338619232178, + 0.0034558463376015425, + 0.0037362477742135525, + 0.006003648042678833, + 0.00683895917609334, + 0.007098602596670389, + 0.008149911649525166, + 0.002767869969829917, + 0.014675176702439785, + 0.005948367062956095, + 0.004735163878649473, + 0.00916781835258007, + 0.016023261472582817, + 0.009398678317666054, + 0.01222788356244564, + 0.0055920896120369434, + 0.027447452768683434, + 0.007955346256494522, + 0.00967162474989891, + 0.029484163969755173, + 0.012536768801510334, + 0.01663959212601185 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 5, + 9 + ], + "selected_reward": -0.0040963320061564445 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "stark, almost static, documentary-style close-up shot in stark black and white captures the delicate interaction between two dahlias in full bloom. The lighting is high-contrast, as if coming from a single, focused source, sculpting the subjects in dramatic black and white. The primary flower's countless layers of spade-shaped petals form a perfect, intricate spiral of brilliant ivory, with deep, velvety charcoal shadows between each petal, emphasizing its complex geometry and texture. The background is a soft, out-of-focus bokeh composed of muted grays and deep blacks, suggesting a garden setting reduced to abstract forms. The first flower sways slowly and lazily, lifted by an imperceptible breeze. Its entire head gracefully tilts on its stem, allowing the edge of its outermost petal to gently brush the surface of the second, still flower. This micro-movement is a silent, slow dance; the petals trace a gentle, fleeting path across their partner, a contact so light as to barely deform their delicate structure, creating a profound sense of intimacy and a tranquil, natural poetry.", + "seed": 4644, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000101/group_7/index_5650_seed_4644_dp_15_rank_30_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 7, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.42635741494596, + "vlm_logprob": -0.42635741494596, + "vlm_yes_no_margin": 2.225, + "vlm_token_credit": -0.42635741494596, + "vlm_yes_ratio": 0.8, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video presented in black and white with high-contrast lighting?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are two dahlias in full bloom visible as the primary subjects?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the first dahlia have ivory petals arranged in a spiral pattern?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the first dahlia sway and tilt to brush its petal against the second dahlia?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a soft, out-of-focus bokeh of muted grays and deep blacks?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -1.9130594730377197 + ], + "opposite_logprob_tokens": [ + -0.16305941343307495 + ], + "logprob_reward": -1.9130594730377197, + "yes_no_margin_reward": -1.75, + "token_credit_reward": -1.9130594730377197, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.18034017086029053, + 0.07077956199645996, + 0.08207239210605621, + 0.06370075047016144, + 0.1404481679201126, + 0.244697704911232, + 0.06694762408733368, + 0.08677632361650467, + 0.02610812894999981, + 0.03812914714217186 + ], + "credit_map": [ + 0.016012469306588173, + 0.014825638383626938, + 0.026395585387945175, + 0.03694130480289459, + 0.009737884625792503, + 0.02580404281616211, + 0.022340165451169014, + 0.028283096849918365, + 0.0062308222986757755, + 0.00792399886995554, + 0.0078054326586425304, + 0.007820375263690948, + 0.004540282301604748, + 0.01579490303993225, + 0.007890704087913036, + 0.012773050926625729, + 0.010834982618689537, + 0.014779482036828995, + 0.01712845079600811, + 0.007484381552785635, + 0.005434112157672644, + 0.008050806820392609, + 0.008135988377034664, + 0.010224190540611744, + 0.0068912808783352375, + 0.008561506867408752, + 0.011130556464195251, + 0.01136818528175354, + 0.0038033719174563885, + 0.00903414748609066, + 0.007745133712887764, + 0.005166571121662855, + 0.01480099093168974, + 0.017431434243917465, + 0.014905639924108982, + 0.02021414414048195, + 0.012026858516037464, + 0.01382813323289156, + 0.030024226754903793, + 0.017216747626662254, + 0.05507931113243103, + 0.026532988995313644, + 0.03097345121204853, + 0.020116377621889114, + 0.014807568863034248, + 0.05397332087159157, + 0.01797671616077423, + 0.02523798681795597, + 0.009876902215182781, + 0.004557694308459759, + 0.01096990518271923, + 0.008252182975411415, + 0.005141290370374918, + 0.014430714771151543, + 0.0076567367650568485, + 0.006062202155590057, + 0.02069598063826561, + 0.008000300265848637, + 0.01267921831458807, + 0.012263089418411255, + 0.0082614216953516, + 0.009165137074887753, + 0.008024217560887337, + 0.007686970289796591, + 0.004448247607797384, + 0.004041614010930061, + 0.0029415804892778397, + 0.004902460612356663, + 0.002587553346529603, + 0.0019522275542840362, + 0.002913961885496974, + 0.0023204837925732136, + 0.009522741660475731, + 0.0032901079393923283, + 0.005241708364337683, + 0.005434107035398483, + 0.003669188590720296, + 0.003148640040308237, + 0.004335177130997181, + 0.0034874803386628628 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 0, + 4 + ], + "selected_reward": -1.9130594730377197 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video presented in black and white with high-contrast lighting?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are two dahlias in full bloom visible as the primary subjects?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the first dahlia have ivory petals arranged in a spiral pattern?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the first dahlia sway and tilt to brush its petal against the second dahlia?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a soft, out-of-focus bokeh of muted grays and deep blacks?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.020617924630641937 + ], + "opposite_logprob_tokens": [ + -3.895617961883545 + ], + "logprob_reward": -0.020617924630641937, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.020617924630641937, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1342540830373764, + 0.06295722723007202, + 0.05497194081544876, + 0.11183814704418182, + 0.12052450329065323, + 0.3159780502319336, + 0.060320671647787094, + 0.07227309048175812, + 0.025858988985419273, + 0.04102335497736931 + ], + "credit_map": [ + 0.01135927252471447, + 0.01090591587126255, + 0.019572202116250992, + 0.02827492728829384, + 0.006521099247038364, + 0.020798539742827415, + 0.01586533710360527, + 0.020956767722964287, + 0.006494679022580385, + 0.010003275237977505, + 0.009031987749040127, + 0.008449185639619827, + 0.003615342080593109, + 0.008200671523809433, + 0.00847560167312622, + 0.008686477318406105, + 0.005628011655062437, + 0.007865817286074162, + 0.011039286851882935, + 0.004824357572942972, + 0.0034444655757397413, + 0.008801708929240704, + 0.007439754903316498, + 0.005928531754761934, + 0.009433578699827194, + 0.017210574820637703, + 0.022424286231398582, + 0.009071174077689648, + 0.008874555118381977, + 0.018441788852214813, + 0.018146444112062454, + 0.008235739544034004, + 0.012313216924667358, + 0.013865920715034008, + 0.01370116788893938, + 0.018499379977583885, + 0.00845364946871996, + 0.013486102223396301, + 0.023319346830248833, + 0.01688571646809578, + 0.0659593790769577, + 0.04245750606060028, + 0.038500092923641205, + 0.027072526514530182, + 0.022022409364581108, + 0.0669894814491272, + 0.02313350886106491, + 0.029843106865882874, + 0.009401253424584866, + 0.005821045022457838, + 0.010354612953960896, + 0.007622013799846172, + 0.005151859950274229, + 0.011321394704282284, + 0.0062588960863649845, + 0.004389592446386814, + 0.017382357269525528, + 0.007736249826848507, + 0.010652373544871807, + 0.009312145411968231, + 0.006932059768587351, + 0.006257474422454834, + 0.0069442018866539, + 0.007056224625557661, + 0.0052551040425896645, + 0.0027883376460522413, + 0.003487522015348077, + 0.004229214042425156, + 0.0025837477296590805, + 0.002512079430744052, + 0.002694926457479596, + 0.00230805529281497, + 0.011545997112989426, + 0.0032947768922895193, + 0.0059456792660057545, + 0.005318238865584135, + 0.004103983752429485, + 0.00325588951818645, + 0.003615366527810693, + 0.003943419083952904 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 0, + 4 + ], + "selected_reward": -0.020617924630641937 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video presented in black and white with high-contrast lighting?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are two dahlias in full bloom visible as the primary subjects?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the first dahlia have ivory petals arranged in a spiral pattern?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the first dahlia sway and tilt to brush its petal against the second dahlia?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a soft, out-of-focus bokeh of muted grays and deep blacks?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01608221046626568 + ], + "opposite_logprob_tokens": [ + -4.141082286834717 + ], + "logprob_reward": -0.01608221046626568, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.01608221046626568, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.18604907393455505, + 0.11650785803794861, + 0.1297973245382309, + 0.08030081540346146, + 0.13276544213294983, + 0.14423781633377075, + 0.05912758409976959, + 0.08510474860668182, + 0.023625047877430916, + 0.04248427599668503 + ], + "credit_map": [ + 0.0176529660820961, + 0.015061426907777786, + 0.026351643726229668, + 0.036933764815330505, + 0.011093473993241787, + 0.028217097744345665, + 0.02021832764148712, + 0.030520375818014145, + 0.008408328518271446, + 0.009791172109544277, + 0.01112609263509512, + 0.00793428998440504, + 0.006639300845563412, + 0.0364629328250885, + 0.007949721068143845, + 0.028196027502417564, + 0.015509686432778835, + 0.020212886855006218, + 0.024392610415816307, + 0.012616641819477081, + 0.00959155336022377, + 0.012258141301572323, + 0.0180289838463068, + 0.01718682236969471, + 0.00811469741165638, + 0.011460836045444012, + 0.013639611192047596, + 0.014510630629956722, + 0.005024029407650232, + 0.010579174384474754, + 0.011087491177022457, + 0.005884348880499601, + 0.01586379110813141, + 0.01408858597278595, + 0.014253155328333378, + 0.020058922469615936, + 0.01028185524046421, + 0.01394272968173027, + 0.02806685119867325, + 0.0162095557898283, + 0.030645236372947693, + 0.015828784555196762, + 0.01807033084332943, + 0.010992145165801048, + 0.008443924598395824, + 0.03056548722088337, + 0.013619310222566128, + 0.016072586178779602, + 0.007631584070622921, + 0.005610564723610878, + 0.009606098756194115, + 0.0061347163282334805, + 0.005261319223791361, + 0.013849820010364056, + 0.006010738201439381, + 0.005022740922868252, + 0.017802178859710693, + 0.009154951199889183, + 0.010973900556564331, + 0.011842983774840832, + 0.009974736720323563, + 0.008650131523609161, + 0.009391787461936474, + 0.0073140813037753105, + 0.003601392963901162, + 0.004712024703621864, + 0.002595743164420128, + 0.0037237079814076424, + 0.0027980946470052004, + 0.0016219762619584799, + 0.0029940747190266848, + 0.0015780329704284668, + 0.010615631937980652, + 0.003269979264587164, + 0.005740134045481682, + 0.005109898746013641, + 0.004260482732206583, + 0.00458043348044157, + 0.005269606597721577, + 0.0036381101235747337 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 5, + 4 + ], + "selected_reward": -0.01608221046626568 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video presented in black and white with high-contrast lighting?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are two dahlias in full bloom visible as the primary subjects?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the first dahlia have ivory petals arranged in a spiral pattern?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the first dahlia sway and tilt to brush its petal against the second dahlia?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a soft, out-of-focus bokeh of muted grays and deep blacks?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.12706820666790009 + ], + "opposite_logprob_tokens": [ + -2.127068281173706 + ], + "logprob_reward": -0.12706820666790009, + "yes_no_margin_reward": 2.0, + "token_credit_reward": -0.12706820666790009, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.114568330347538, + 0.11701186746358871, + 0.08746127784252167, + 0.0678546354174614, + 0.16092906892299652, + 0.295430988073349, + 0.04172755405306816, + 0.05301468074321747, + 0.027301883324980736, + 0.034699697047472 + ], + "credit_map": [ + 0.00659363716840744, + 0.009191441349685192, + 0.011296208016574383, + 0.013354611583054066, + 0.004622768145054579, + 0.033519942313432693, + 0.008882365189492702, + 0.02710735611617565, + 0.012922346591949463, + 0.017689645290374756, + 0.01582581363618374, + 0.013759292662143707, + 0.006817635148763657, + 0.01411417592316866, + 0.020412463694810867, + 0.015470494516193867, + 0.008222507312893867, + 0.01452343724668026, + 0.01167345978319645, + 0.009229750372469425, + 0.006059880368411541, + 0.013573178090155125, + 0.016031157225370407, + 0.008147905580699444, + 0.006194745190441608, + 0.0074065313674509525, + 0.012551945634186268, + 0.006957236211746931, + 0.005079794209450483, + 0.012365926057100296, + 0.012415830977261066, + 0.0048826211132109165, + 0.02126004733145237, + 0.02193661592900753, + 0.018690528348088264, + 0.01688554137945175, + 0.01792329177260399, + 0.019206823781132698, + 0.030311502516269684, + 0.01471471693366766, + 0.06302543729543686, + 0.04163876920938492, + 0.03451208397746086, + 0.020341357216238976, + 0.018337130546569824, + 0.06577472388744354, + 0.022136753425002098, + 0.029664749279618263, + 0.006338099949061871, + 0.002491453429684043, + 0.005385777913033962, + 0.0031927861273288727, + 0.0032434756867587566, + 0.011226826347410679, + 0.005430301185697317, + 0.0044188364408910275, + 0.015391762368381023, + 0.00374919013120234, + 0.009618541225790977, + 0.0064871483482420444, + 0.0036212799604982138, + 0.006723121739923954, + 0.0033903156872838736, + 0.004033319186419249, + 0.002476921072229743, + 0.005524609703570604, + 0.0026362177450209856, + 0.004720300901681185, + 0.0038674469105899334, + 0.00328668300062418, + 0.0026652670931071043, + 0.002124437363818288, + 0.00868633296340704, + 0.002772226231172681, + 0.0053740814328193665, + 0.004971479997038841, + 0.003226346103474498, + 0.0033195584546774626, + 0.0027124909684062004, + 0.003637182293459773 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 4, + 1 + ], + "selected_reward": -0.12706820666790009 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video presented in black and white with high-contrast lighting?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are two dahlias in full bloom visible as the primary subjects?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the first dahlia have ivory petals arranged in a spiral pattern?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the first dahlia sway and tilt to brush its petal against the second dahlia?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a soft, out-of-focus bokeh of muted grays and deep blacks?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0549592599272728 + ], + "opposite_logprob_tokens": [ + -2.929959297180176 + ], + "logprob_reward": -0.0549592599272728, + "yes_no_margin_reward": 2.875, + "token_credit_reward": -0.0549592599272728, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.17491550743579865, + 0.10053351521492004, + 0.07375636696815491, + 0.05568264052271843, + 0.11558198183774948, + 0.2549740672111511, + 0.08393221348524094, + 0.08090890198945999, + 0.022287853062152863, + 0.03742692992091179 + ], + "credit_map": [ + 0.01570877432823181, + 0.011398203670978546, + 0.02960287779569626, + 0.024111149832606316, + 0.007599845994263887, + 0.038979534059762955, + 0.015315122902393341, + 0.03220000118017197, + 0.008748490363359451, + 0.008066383190453053, + 0.013285873457789421, + 0.0067263091914355755, + 0.004513581283390522, + 0.029016386717557907, + 0.005767019931226969, + 0.02440946362912655, + 0.010847087949514389, + 0.012387348338961601, + 0.01681400090456009, + 0.006366618908941746, + 0.004448492079973221, + 0.0070212786085903645, + 0.005802070256322622, + 0.010069472715258598, + 0.0074003273621201515, + 0.006192390341311693, + 0.008271128870546818, + 0.010445678606629372, + 0.0033368286676704884, + 0.0074789319187402725, + 0.007552259601652622, + 0.00500509375706315, + 0.013357091695070267, + 0.014319109730422497, + 0.012130999937653542, + 0.015027224086225033, + 0.008909396827220917, + 0.011754092760384083, + 0.025377707555890083, + 0.014706369489431381, + 0.05820102244615555, + 0.02818133682012558, + 0.03195715695619583, + 0.019659215584397316, + 0.015474380925297737, + 0.05663267895579338, + 0.018379496410489082, + 0.026488782837986946, + 0.011319991201162338, + 0.005338369403034449, + 0.014479154720902443, + 0.012110994197428226, + 0.005666232202202082, + 0.017900047823786736, + 0.009827879257500172, + 0.007289547007530928, + 0.01817738078534603, + 0.00759443175047636, + 0.011623505502939224, + 0.011723929084837437, + 0.007899267598986626, + 0.009098966605961323, + 0.007639117538928986, + 0.007152301259338856, + 0.0031561178620904684, + 0.0035954772029072046, + 0.0030012798961251974, + 0.004196752328425646, + 0.0016849804669618607, + 0.0021394267678260803, + 0.0024021556600928307, + 0.002111661247909069, + 0.01035615149885416, + 0.002767972880974412, + 0.005308276508003473, + 0.005108354613184929, + 0.003310520201921463, + 0.0032469674479216337, + 0.003943478222936392, + 0.003385205753147602 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 0, + 4 + ], + "selected_reward": -0.0549592599272728 + } + ], + "vlm_decoded_outputs": [ + "No", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "stark, almost static, documentary-style close-up shot in stark black and white captures the delicate interaction between two dahlias in full bloom. The lighting is high-contrast, as if coming from a single, focused source, sculpting the subjects in dramatic black and white. The primary flower's countless layers of spade-shaped petals form a perfect, intricate spiral of brilliant ivory, with deep, velvety charcoal shadows between each petal, emphasizing its complex geometry and texture. The background is a soft, out-of-focus bokeh composed of muted grays and deep blacks, suggesting a garden setting reduced to abstract forms. The first flower sways slowly and lazily, lifted by an imperceptible breeze. Its entire head gracefully tilts on its stem, allowing the edge of its outermost petal to gently brush the surface of the second, still flower. This micro-movement is a silent, slow dance; the petals trace a gentle, fleeting path across their partner, a contact so light as to barely deform their delicate structure, creating a profound sense of intimacy and a tranquil, natural poetry.", + "seed": 4645, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000101/group_7/index_5650_seed_4645_dp_15_rank_30_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 7, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.08548816004768015, + "vlm_logprob": -0.08548816004768015, + "vlm_yes_no_margin": 3.35, + "vlm_token_credit": -0.08548816004768015, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video presented in black and white with high-contrast lighting?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are two dahlias in full bloom visible as the primary subjects?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the first dahlia have ivory petals arranged in a spiral pattern?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the first dahlia sway and tilt to brush its petal against the second dahlia?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a soft, out-of-focus bokeh of muted grays and deep blacks?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02689477801322937 + ], + "opposite_logprob_tokens": [ + -3.7768948078155518 + ], + "logprob_reward": -0.02689477801322937, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.02689477801322937, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.04401353374123573, + 0.04939465597271919, + 0.10231079906225204, + 0.3162011206150055, + 0.03511927276849747, + 0.04144177585840225, + 0.0725359320640564, + 0.0076640937477350235, + 0.17376920580863953, + 0.1575496941804886 + ], + "credit_map": [ + 0.006640779785811901, + 0.005568148102611303, + 0.008145732805132866, + 0.005088532343506813, + 0.0026859960053116083, + 0.003110276535153389, + 0.0054234168492257595, + 0.007350652478635311, + 0.006127319764345884, + 0.004502118565142155, + 0.013187732547521591, + 0.0028803683817386627, + 0.0037113893777132034, + 0.003649112768471241, + 0.0026902377139776945, + 0.012646379880607128, + 0.014362158253788948, + 0.012566396966576576, + 0.028869224712252617, + 0.009229032322764397, + 0.008649385534226894, + 0.007265771273523569, + 0.008543332107365131, + 0.012825492769479752, + 0.03594239801168442, + 0.09270523488521576, + 0.037849802523851395, + 0.029953911900520325, + 0.03884969279170036, + 0.017586782574653625, + 0.026903005316853523, + 0.036410264670848846, + 0.0057722278870642185, + 0.00420919107273221, + 0.008441351354122162, + 0.0035701601300388575, + 0.0034071628469973803, + 0.002166006015613675, + 0.0028441359754651785, + 0.004709037020802498, + 0.0044860150665044785, + 0.004053740296512842, + 0.007677091285586357, + 0.006676905322819948, + 0.005125356372445822, + 0.0029310823883861303, + 0.004911800380796194, + 0.0055797831155359745, + 0.007697261869907379, + 0.009277096018195152, + 0.013837838545441628, + 0.007623767014592886, + 0.010366830043494701, + 0.008245376870036125, + 0.006911765318363905, + 0.00857599638402462, + 0.00128855649381876, + 0.0008541036513634026, + 0.001526320120319724, + 0.0008064149296842515, + 0.000921076163649559, + 0.0006284832488745451, + 0.0006965395295992494, + 0.0009426000178791583, + 0.01576198637485504, + 0.005523603409528732, + 0.07387213408946991, + 0.011905086226761341, + 0.0052434299141168594, + 0.006837482564151287, + 0.00466556753963232, + 0.04995991662144661, + 0.015685439109802246, + 0.007563170976936817, + 0.06450514495372772, + 0.015045433305203915, + 0.003578681033104658, + 0.006669516675174236, + 0.003592402208596468, + 0.040909893810749054 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 8, + 9 + ], + "selected_reward": -0.02689477801322937 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video presented in black and white with high-contrast lighting?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are two dahlias in full bloom visible as the primary subjects?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the first dahlia have ivory petals arranged in a spiral pattern?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the first dahlia sway and tilt to brush its petal against the second dahlia?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a soft, out-of-focus bokeh of muted grays and deep blacks?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.12705613672733307 + ], + "opposite_logprob_tokens": [ + -2.127056121826172 + ], + "logprob_reward": -0.12705613672733307, + "yes_no_margin_reward": 2.0, + "token_credit_reward": -0.12705613672733307, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07524796575307846, + 0.01563771814107895, + 0.011962507851421833, + 0.038319431245326996, + 0.020827997475862503, + 0.039562053978443146, + 0.009097537025809288, + 0.023037636652588844, + 0.45034298300743103, + 0.3159642219543457 + ], + "credit_map": [ + 0.01186705194413662, + 0.013056650757789612, + 0.010290684178471565, + 0.006558450870215893, + 0.0066506932489573956, + 0.005558196920901537, + 0.009207330644130707, + 0.012058897875249386, + 0.00232597510330379, + 0.0022845470812171698, + 0.0029599068220704794, + 0.0015364655992016196, + 0.00167572358623147, + 0.0011999770067632198, + 0.001709049567580223, + 0.0019460722105577588, + 0.0014934486243873835, + 0.0028052618727087975, + 0.0021977864671498537, + 0.0012553316773846745, + 0.0011721892515197396, + 0.0006714786868542433, + 0.0010503698140382767, + 0.0013166400603950024, + 0.003769764443859458, + 0.01079784706234932, + 0.0043428633362054825, + 0.0036199032329022884, + 0.005582599900662899, + 0.002289597410708666, + 0.0037682796828448772, + 0.004148572217673063, + 0.003880125004798174, + 0.001328244456090033, + 0.004469956737011671, + 0.0026902772951871157, + 0.0014883230905979872, + 0.0014435077318921685, + 0.0025321566499769688, + 0.0029954041820019484, + 0.004651645664125681, + 0.006238352973014116, + 0.005922375712543726, + 0.004599889740347862, + 0.006833553314208984, + 0.002402405720204115, + 0.004473068751394749, + 0.004440758377313614, + 0.0012493504909798503, + 0.0012927396455779672, + 0.0013098671333864331, + 0.0013906878884881735, + 0.0013005335349589586, + 0.0006423766608349979, + 0.00077060516923666, + 0.0011413757456466556, + 0.0034410313237458467, + 0.001813969574868679, + 0.005036066751927137, + 0.0033218390308320522, + 0.0021953347604721785, + 0.0019284612499177456, + 0.0023914691992104053, + 0.0029094635974615812, + 0.041002094745635986, + 0.014637695625424385, + 0.1912689507007599, + 0.03066434897482395, + 0.013512990437448025, + 0.01775178872048855, + 0.012434592470526695, + 0.1290704756975174, + 0.0315067395567894, + 0.014440610073506832, + 0.1300186812877655, + 0.029276862740516663, + 0.007626131176948547, + 0.01416823361068964, + 0.007487012539058924, + 0.08143994212150574 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 9, + 0 + ], + "selected_reward": -0.12705613672733307 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video presented in black and white with high-contrast lighting?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are two dahlias in full bloom visible as the primary subjects?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the first dahlia have ivory petals arranged in a spiral pattern?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the first dahlia sway and tilt to brush its petal against the second dahlia?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a soft, out-of-focus bokeh of muted grays and deep blacks?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.018193509429693222 + ], + "opposite_logprob_tokens": [ + -4.01819372177124 + ], + "logprob_reward": -0.018193509429693222, + "yes_no_margin_reward": 4.0, + "token_credit_reward": -0.018193509429693222, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.045083336532115936, + 0.02757478505373001, + 0.038628850132226944, + 0.06506752222776413, + 0.02504424937069416, + 0.07086095958948135, + 0.03497638180851936, + 0.01985017955303192, + 0.03958101570606232, + 0.6333326697349548 + ], + "credit_map": [ + 0.006235682405531406, + 0.006331061944365501, + 0.006771418731659651, + 0.004994051996618509, + 0.00458729499951005, + 0.0034064496867358685, + 0.007036888971924782, + 0.005720492452383041, + 0.0036920509301126003, + 0.0031660364475101233, + 0.005743236280977726, + 0.003355066291987896, + 0.0029582383576780558, + 0.0020821355283260345, + 0.0028014229610562325, + 0.0037766005843877792, + 0.008371849544346333, + 0.003451131284236908, + 0.009499616920948029, + 0.004317405167967081, + 0.002043931046500802, + 0.002514692023396492, + 0.0026626745238900185, + 0.005767551250755787, + 0.00698772631585598, + 0.01847522333264351, + 0.009782837703824043, + 0.006122775375843048, + 0.007051697000861168, + 0.003649803576990962, + 0.005371759179979563, + 0.007625706493854523, + 0.0032503067050129175, + 0.0022962703369557858, + 0.0033298300113528967, + 0.0030721821822226048, + 0.003798597725108266, + 0.0028185865376144648, + 0.0036257717292755842, + 0.002852708101272583, + 0.0057751783169806, + 0.009636571630835533, + 0.007500488311052322, + 0.010822766460478306, + 0.012515237554907799, + 0.00457371212542057, + 0.011306220665574074, + 0.008730793371796608, + 0.005839001387357712, + 0.002802574774250388, + 0.008619877509772778, + 0.004309738054871559, + 0.00305647449567914, + 0.0024445904418826103, + 0.003544424893334508, + 0.004359703045338392, + 0.003181951120495796, + 0.001725945738144219, + 0.005887945648282766, + 0.0019266835879534483, + 0.0018644887022674084, + 0.0011861642124131322, + 0.0016916828462854028, + 0.002385318512097001, + 0.003470237599685788, + 0.0042472584173083305, + 0.009273165836930275, + 0.003871508873999119, + 0.005504234693944454, + 0.003095086431130767, + 0.003825222374871373, + 0.006294305436313152, + 0.06344149261713028, + 0.029311586171388626, + 0.26045700907707214, + 0.0587100088596344, + 0.014433901757001877, + 0.02723848633468151, + 0.015001717954874039, + 0.16473855078220367 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 5, + 3 + ], + "selected_reward": -0.018193509429693222 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video presented in black and white with high-contrast lighting?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are two dahlias in full bloom visible as the primary subjects?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the first dahlia have ivory petals arranged in a spiral pattern?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the first dahlia sway and tilt to brush its petal against the second dahlia?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a soft, out-of-focus bokeh of muted grays and deep blacks?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.2521045207977295 + ], + "opposite_logprob_tokens": [ + -1.5021045207977295 + ], + "logprob_reward": -0.2521045207977295, + "yes_no_margin_reward": 1.25, + "token_credit_reward": -0.2521045207977295, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.039396077394485474, + 0.028308972716331482, + 0.008787939324975014, + 0.007082173600792885, + 0.009116649627685547, + 0.009053236804902554, + 0.012037890963256359, + 0.007529702968895435, + 0.5268003940582275, + 0.35188695788383484 + ], + "credit_map": [ + 0.00513202790170908, + 0.006889230106025934, + 0.007682396098971367, + 0.0037158657796680927, + 0.0031216945499181747, + 0.002480229130014777, + 0.005314519628882408, + 0.0050601172260940075, + 0.0028427119832485914, + 0.002951177069917321, + 0.007798672653734684, + 0.0015734952175989747, + 0.0020444148685783148, + 0.001974803861230612, + 0.0014124477747827768, + 0.00771125266328454, + 0.0014249125961214304, + 0.0008403816027566791, + 0.0019042380154132843, + 0.0009177929605357349, + 0.0007998932851478457, + 0.0006912134122103453, + 0.0008706033695489168, + 0.0013389043742790818, + 0.0008188470965251327, + 0.0010434730211272836, + 0.0017161578871309757, + 0.0007233153446577489, + 0.0006456907722167671, + 0.0004457465838640928, + 0.0007728083292022347, + 0.000916134798899293, + 0.0016780868172645569, + 0.0004724247264675796, + 0.0023900000378489494, + 0.0008217047434300184, + 0.0006141160847619176, + 0.0007427341188304126, + 0.0010417465819045901, + 0.0013558374485000968, + 0.0009095656569115818, + 0.0008014171617105603, + 0.0016642072005197406, + 0.0015146652003750205, + 0.0010061361826956272, + 0.0008857623906806111, + 0.0010237648384645581, + 0.001247718115337193, + 0.0019720997661352158, + 0.0011069877073168755, + 0.002711130538955331, + 0.001061694580130279, + 0.0013413084670901299, + 0.0008631443488411605, + 0.0010306790936738253, + 0.0019508481491357088, + 0.001033795066177845, + 0.0008043950074352324, + 0.002019554376602173, + 0.0007229958428069949, + 0.0008392918389290571, + 0.0005803685635328293, + 0.0006289344746619463, + 0.0009003679733723402, + 0.04777601361274719, + 0.016636965796351433, + 0.22386181354522705, + 0.03601248562335968, + 0.016055108979344368, + 0.021002719178795815, + 0.014155511744320393, + 0.15129977464675903, + 0.03527893126010895, + 0.016161726787686348, + 0.14480814337730408, + 0.0328277051448822, + 0.008077790960669518, + 0.015454431064426899, + 0.00807054340839386, + 0.09120772033929825 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 9, + 0 + ], + "selected_reward": -0.2521045207977295 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video presented in black and white with high-contrast lighting?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are two dahlias in full bloom visible as the primary subjects?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the first dahlia have ivory petals arranged in a spiral pattern?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the first dahlia sway and tilt to brush its petal against the second dahlia?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a soft, out-of-focus bokeh of muted grays and deep blacks?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0031918552704155445 + ], + "opposite_logprob_tokens": [ + -5.753191947937012 + ], + "logprob_reward": -0.0031918552704155445, + "yes_no_margin_reward": 5.75, + "token_credit_reward": -0.0031918552704155445, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07613539695739746, + 0.0345301516354084, + 0.0550479032099247, + 0.0943676084280014, + 0.02784658968448639, + 0.022441094741225243, + 0.029172491282224655, + 0.0563063807785511, + 0.3301492929458618, + 0.2740030884742737 + ], + "credit_map": [ + 0.010528653860092163, + 0.008137847296893597, + 0.017405305057764053, + 0.008419044315814972, + 0.003653265768662095, + 0.005047708284109831, + 0.008700459264218807, + 0.014243118464946747, + 0.005318759474903345, + 0.004680864047259092, + 0.008150584064424038, + 0.003134209429845214, + 0.0018324867123737931, + 0.002023178618401289, + 0.002364829648286104, + 0.007025242783129215, + 0.007708655204623938, + 0.005515181925147772, + 0.01636102795600891, + 0.004961390048265457, + 0.005109015386551619, + 0.0037996554747223854, + 0.004271551035344601, + 0.007321429438889027, + 0.012636837549507618, + 0.02172176167368889, + 0.02020876482129097, + 0.007495504338294268, + 0.007792268879711628, + 0.005651008803397417, + 0.007150298450142145, + 0.011711165308952332, + 0.0035460565704852343, + 0.003188113681972027, + 0.004717560019344091, + 0.0038002918008714914, + 0.0032987233716994524, + 0.0023261478636413813, + 0.003651075530797243, + 0.0033186227083206177, + 0.0023840726353228092, + 0.0031118483748286963, + 0.004497595131397247, + 0.0032666530460119247, + 0.0032638665288686752, + 0.0011285357177257538, + 0.002084225183352828, + 0.0027042971923947334, + 0.004198749549686909, + 0.003268585307523608, + 0.005832612980157137, + 0.0030565618071705103, + 0.003922945819795132, + 0.0031261444091796875, + 0.002271376084536314, + 0.0034955155570060015, + 0.006731534376740456, + 0.007765608374029398, + 0.010286533273756504, + 0.006202960852533579, + 0.00873600598424673, + 0.003912389278411865, + 0.006552116479724646, + 0.0061192321591079235, + 0.03009694069623947, + 0.010564303025603294, + 0.1404462456703186, + 0.02250552549958229, + 0.009882166981697083, + 0.012925928458571434, + 0.00890085194259882, + 0.09482733905315399, + 0.027363166213035583, + 0.01274135708808899, + 0.112672820687294, + 0.025605803355574608, + 0.00628941785544157, + 0.011898149736225605, + 0.006129765417426825, + 0.07130259275436401 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 9, + 3 + ], + "selected_reward": -0.0031918552704155445 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000101/group_8/index_tensor([4709, 4709])_group_8_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000101/group_8/index_tensor([4709, 4709])_group_8_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..86cd1e0109f47716707d284e36d0c9146db596d5 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000101/group_8/index_tensor([4709, 4709])_group_8_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "medium shot, at eye level, shot with a shallow depth of field against a softly blurred, clean white background, frames a tall, transparent cylindrical glass filled with pristine, still water. The camera begins a slow, deliberate push-in, moving ever closer as a white, circular pencil tablet is placed at its base. Instantly, the tablet erupts with a violent hissing, releasing a turbulent, chaotic cloud of tiny bubbles that swirl upward in a frantic dance. A crisp, bright backlight illuminates the scene, making each individual bubble shimmer like a tiny crystal and casting a radiant glow through the glass. As the camera continues its steady advance into extreme macro close-up, it captures the tablet's surface disintegrating, steadily releasing a milky plume that tumbles and pours into the clear water. The action is a mesmerizing spectacle, brimming with energy, a furious, silent explosion of rebirth, creating an atmosphere of vibrant transformation and clean, powerful force.", + "seed": 14266, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000101/group_8/index_4709_seed_14266_dp_16_rank_32_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 8, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.037028145417571066, + "vlm_logprob": -0.037028145417571066, + "vlm_yes_no_margin": 3.925, + "vlm_token_credit": -0.037028145417571066, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene framed as a medium shot at eye level with a shallow depth of field against a clean white background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a tall, transparent cylindrical glass filled with still water visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a white circular pencil tablet appear at the base of the glass and erupt with a turbulent cloud of tiny bubbles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera begin with a slow push-in that continues into an extreme macro close-up?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the tablet surface shown disintegrating and releasing a milky plume into the water?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.11110520362854004 + ], + "opposite_logprob_tokens": [ + -2.36110520362854 + ], + "logprob_reward": -0.11110520362854004, + "yes_no_margin_reward": 2.25, + "token_credit_reward": -0.11110520362854004, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3281170427799225, + 0.08471742272377014, + 0.13754913210868835, + 0.06014883890748024, + 0.0579628087580204, + 0.12380369007587433, + 0.06955181062221527, + 0.01946895197033882, + 0.05663631111383438, + 0.06204402446746826 + ], + "credit_map": [ + 0.0332997627556324, + 0.06840259581804276, + 0.04205671325325966, + 0.03292786329984665, + 0.04455767199397087, + 0.04088709503412247, + 0.028888272121548653, + 0.03709707409143448, + 0.00824929028749466, + 0.01168318372219801, + 0.01254751905798912, + 0.01055370457470417, + 0.007218504790216684, + 0.012147944420576096, + 0.012644379399716854, + 0.009672898799180984, + 0.014153040014207363, + 0.023535244166851044, + 0.015221214853227139, + 0.011402115225791931, + 0.022014394402503967, + 0.01839352771639824, + 0.015272034332156181, + 0.017557572573423386, + 0.0077408067882061005, + 0.009977297857403755, + 0.005146408919245005, + 0.008107702247798443, + 0.006839997135102749, + 0.006176475901156664, + 0.009825041517615318, + 0.006335115060210228, + 0.004944360814988613, + 0.007172716781497002, + 0.008623701520264149, + 0.007736195810139179, + 0.005357447080314159, + 0.006018505431711674, + 0.010792497545480728, + 0.007317389361560345, + 0.011588957160711288, + 0.021395638585090637, + 0.014027375727891922, + 0.011057356372475624, + 0.013834725134074688, + 0.015413605608046055, + 0.014272465370595455, + 0.02221357263624668, + 0.013903733342885971, + 0.008286151103675365, + 0.007812637835741043, + 0.008322594687342644, + 0.006356773432344198, + 0.009224322624504566, + 0.0071665276773273945, + 0.008479068987071514, + 0.002700209151953459, + 0.002998883603140712, + 0.0017624034080654383, + 0.0018872544169425964, + 0.0027107896748930216, + 0.0031686367001384497, + 0.0018532180693000555, + 0.0023875581100583076, + 0.005597362294793129, + 0.00957825779914856, + 0.009243834763765335, + 0.006518892478197813, + 0.007089993916451931, + 0.004436524584889412, + 0.007934910245239735, + 0.006236537825316191, + 0.0057953596115112305, + 0.00817013531923294, + 0.006672501564025879, + 0.012592385523021221, + 0.008886454626917839, + 0.00792333297431469, + 0.006884847767651081, + 0.005119012668728828 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 5 + ], + "selected_reward": -0.11110520362854004 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene framed as a medium shot at eye level with a shallow depth of field against a clean white background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a tall, transparent cylindrical glass filled with still water visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a white circular pencil tablet appear at the base of the glass and erupt with a turbulent cloud of tiny bubbles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera begin with a slow push-in that continues into an extreme macro close-up?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the tablet surface shown disintegrating and releasing a milky plume into the water?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02983754687011242 + ], + "opposite_logprob_tokens": [ + -3.5298376083374023 + ], + "logprob_reward": -0.02983754687011242, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.02983754687011242, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3636092245578766, + 0.03993066027760506, + 0.04584475979208946, + 0.04412905499339104, + 0.1025824025273323, + 0.06281499564647675, + 0.05706549063324928, + 0.0924672782421112, + 0.1327981948852539, + 0.058757930994033813 + ], + "credit_map": [ + 0.04100403934717178, + 0.06573577225208282, + 0.05301642417907715, + 0.03663277626037598, + 0.037588272243738174, + 0.048109814524650574, + 0.037892624735832214, + 0.04362950474023819, + 0.004401274956762791, + 0.004830395337194204, + 0.004150754306465387, + 0.0059277089312672615, + 0.00532806571573019, + 0.005379989743232727, + 0.004338598810136318, + 0.005573872476816177, + 0.005294996313750744, + 0.00536772795021534, + 0.004876281600445509, + 0.005579463671892881, + 0.005017258692532778, + 0.006408296059817076, + 0.0062063587829470634, + 0.0070943767204880714, + 0.0044513582251966, + 0.005785481538623571, + 0.004536577500402927, + 0.005798857659101486, + 0.004407218191772699, + 0.006262593436986208, + 0.006802129093557596, + 0.006084841210395098, + 0.010340156964957714, + 0.01447730790823698, + 0.018704993650317192, + 0.011039404198527336, + 0.012891425751149654, + 0.011376342736184597, + 0.011954677291214466, + 0.011798089370131493, + 0.005601935088634491, + 0.010803384706377983, + 0.007827459834516048, + 0.004673394840210676, + 0.006202767137438059, + 0.005335094407200813, + 0.006880444474518299, + 0.015490520745515823, + 0.004880235996097326, + 0.006227635312825441, + 0.007314854301512241, + 0.005659318529069424, + 0.007736133877187967, + 0.006213906221091747, + 0.006708741653710604, + 0.012324667535722256, + 0.009514586068689823, + 0.014366701245307922, + 0.010322384536266327, + 0.008777895011007786, + 0.012313410639762878, + 0.011468634940683842, + 0.011995766311883926, + 0.013707895763218403, + 0.014688997529447079, + 0.02484569326043129, + 0.017457617446780205, + 0.013448708690702915, + 0.020166713744401932, + 0.010141892358660698, + 0.012357035651803017, + 0.01969153620302677, + 0.005448105279356241, + 0.007381836883723736, + 0.005205594468861818, + 0.008493885397911072, + 0.011039750650525093, + 0.007340756710618734, + 0.005195073317736387, + 0.00865293387323618 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 4 + ], + "selected_reward": -0.02983754687011242 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene framed as a medium shot at eye level with a shallow depth of field against a clean white background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a tall, transparent cylindrical glass filled with still water visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a white circular pencil tablet appear at the base of the glass and erupt with a turbulent cloud of tiny bubbles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera begin with a slow push-in that continues into an extreme macro close-up?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the tablet surface shown disintegrating and releasing a milky plume into the water?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005255098920315504 + ], + "opposite_logprob_tokens": [ + -5.255255222320557 + ], + "logprob_reward": -0.005255098920315504, + "yes_no_margin_reward": 5.25, + "token_credit_reward": -0.005255098920315504, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.34872743487358093, + 0.051227401942014694, + 0.11106955260038376, + 0.08265282958745956, + 0.08482517302036285, + 0.02553432434797287, + 0.06470602750778198, + 0.0717460960149765, + 0.10965730994939804, + 0.04985393211245537 + ], + "credit_map": [ + 0.03803356736898422, + 0.06292669475078583, + 0.04461100324988365, + 0.03745991736650467, + 0.041172198951244354, + 0.04358265921473503, + 0.03999416157603264, + 0.040947187691926956, + 0.005575346760451794, + 0.007417008280754089, + 0.006706105079501867, + 0.00555590633302927, + 0.005416039377450943, + 0.00810210406780243, + 0.006686695385724306, + 0.005768195725977421, + 0.01351214200258255, + 0.016647808253765106, + 0.010371199809014797, + 0.012238619849085808, + 0.013855615630745888, + 0.014549696817994118, + 0.015435855835676193, + 0.01445862092077732, + 0.007930192165076733, + 0.014201884157955647, + 0.008916318416595459, + 0.008929758332669735, + 0.011163361370563507, + 0.009508105926215649, + 0.010534588247537613, + 0.011468615382909775, + 0.006800771225243807, + 0.008436879143118858, + 0.01338266022503376, + 0.008584154769778252, + 0.011561699211597443, + 0.008565674535930157, + 0.014935918152332306, + 0.012557408772408962, + 0.0020911330357193947, + 0.004271144978702068, + 0.002481431234627962, + 0.004174670670181513, + 0.0030733956955373287, + 0.0016812977846711874, + 0.004205307457596064, + 0.003555943723767996, + 0.0065475995652377605, + 0.007377351634204388, + 0.008422929793596268, + 0.00999493058770895, + 0.005997504107654095, + 0.0056001367047429085, + 0.011434727348387241, + 0.009330849163234234, + 0.007027470041066408, + 0.012245381250977516, + 0.008531059138476849, + 0.007377194706350565, + 0.01213604211807251, + 0.0066175502724945545, + 0.007823198102414608, + 0.00998819898813963, + 0.010644723661243916, + 0.017702383920550346, + 0.016706865280866623, + 0.01275753602385521, + 0.013750521466135979, + 0.007484493777155876, + 0.012094387784600258, + 0.01851639524102211, + 0.004593562334775925, + 0.007269551511853933, + 0.004222438205033541, + 0.010250728577375412, + 0.007287677377462387, + 0.004161498975008726, + 0.005510633811354637, + 0.006557841319590807 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 8 + ], + "selected_reward": -0.005255098920315504 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene framed as a medium shot at eye level with a shallow depth of field against a clean white background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a tall, transparent cylindrical glass filled with still water visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a white circular pencil tablet appear at the base of the glass and erupt with a turbulent cloud of tiny bubbles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera begin with a slow push-in that continues into an extreme macro close-up?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the tablet surface shown disintegrating and releasing a milky plume into the water?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03369003161787987 + ], + "opposite_logprob_tokens": [ + -3.4086899757385254 + ], + "logprob_reward": -0.03369003161787987, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.03369003161787987, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10331235080957413, + 0.06658007204532623, + 0.03977884352207184, + 0.12123516947031021, + 0.04534216970205307, + 0.05646532028913498, + 0.1199495792388916, + 0.09465838968753815, + 0.16801726818084717, + 0.18466077744960785 + ], + "credit_map": [ + 0.009550793096423149, + 0.016671501100063324, + 0.01059199869632721, + 0.012429513968527317, + 0.021015068516135216, + 0.010775207541882992, + 0.012844312936067581, + 0.009433961473405361, + 0.004828901030123234, + 0.010191780515015125, + 0.00853108149021864, + 0.008812453597784042, + 0.007853508926928043, + 0.0076338155195117, + 0.009594671428203583, + 0.009133853018283844, + 0.006283652037382126, + 0.004342546686530113, + 0.004824383184313774, + 0.0036490580532699823, + 0.003353097243234515, + 0.006299788597971201, + 0.0055458759889006615, + 0.005480441730469465, + 0.009768083691596985, + 0.020569413900375366, + 0.011169692501425743, + 0.018948130309581757, + 0.014192393980920315, + 0.0136436577886343, + 0.019056223332881927, + 0.013887571170926094, + 0.003954185172915459, + 0.005975741893053055, + 0.005893927998840809, + 0.0042531974613666534, + 0.00879590306431055, + 0.004718845244497061, + 0.006437299773097038, + 0.005313067231327295, + 0.004334725439548492, + 0.007284011226147413, + 0.007153291720896959, + 0.008420966565608978, + 0.004401558078825474, + 0.004127636551856995, + 0.008960769511759281, + 0.011782363057136536, + 0.0117602264508605, + 0.017152106389403343, + 0.015650296583771706, + 0.013834732584655285, + 0.012422977946698666, + 0.011385634541511536, + 0.018021101132035255, + 0.019722506403923035, + 0.00984533317387104, + 0.012927306815981865, + 0.011242394335567951, + 0.011847371235489845, + 0.012956549413502216, + 0.011259419843554497, + 0.013290828093886375, + 0.011289193294942379, + 0.021880241110920906, + 0.021816663444042206, + 0.027042267844080925, + 0.021586362272500992, + 0.015998613089323044, + 0.011586344800889492, + 0.022300073876976967, + 0.025806687772274017, + 0.014826575294137001, + 0.022125422954559326, + 0.020877959206700325, + 0.04115436226129532, + 0.02185930497944355, + 0.014100546948611736, + 0.02402603253722191, + 0.02569056674838066 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 8, + 3 + ], + "selected_reward": -0.03369003161787987 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene framed as a medium shot at eye level with a shallow depth of field against a clean white background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a tall, transparent cylindrical glass filled with still water visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a white circular pencil tablet appear at the base of the glass and erupt with a turbulent cloud of tiny bubbles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera begin with a slow push-in that continues into an extreme macro close-up?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the tablet surface shown disintegrating and releasing a milky plume into the water?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005252846051007509 + ], + "opposite_logprob_tokens": [ + -5.255252838134766 + ], + "logprob_reward": -0.005252846051007509, + "yes_no_margin_reward": 5.25, + "token_credit_reward": -0.005252846051007509, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2899770438671112, + 0.05867335572838783, + 0.05758744478225708, + 0.06522607058286667, + 0.11810772866010666, + 0.06054382398724556, + 0.10223199427127838, + 0.07866981625556946, + 0.10753298550844193, + 0.06144971027970314 + ], + "credit_map": [ + 0.028861714527010918, + 0.05300242081284523, + 0.035792026668787, + 0.03289682790637016, + 0.03836707025766373, + 0.03478735685348511, + 0.03175784647464752, + 0.034511782228946686, + 0.00483653275296092, + 0.010116149671375751, + 0.007591632194817066, + 0.008439621888101101, + 0.005375917535275221, + 0.007445259019732475, + 0.007533592637628317, + 0.007334650494158268, + 0.00747775100171566, + 0.006448979023844004, + 0.007484121713787317, + 0.006643309257924557, + 0.005033440887928009, + 0.0073714242316782475, + 0.010034025646746159, + 0.007094394415616989, + 0.007117000408470631, + 0.009530344977974892, + 0.007825643755495548, + 0.00934511423110962, + 0.005898829083889723, + 0.0072915744967758656, + 0.009555613622069359, + 0.008661951869726181, + 0.012104503810405731, + 0.013571563176810741, + 0.021197235211730003, + 0.01161181926727295, + 0.015828153118491173, + 0.012520878575742245, + 0.015471912920475006, + 0.01580165885388851, + 0.005532334558665752, + 0.007860979065299034, + 0.007896823808550835, + 0.007884715683758259, + 0.005699284374713898, + 0.004508747719228268, + 0.01085266750305891, + 0.010308269411325455, + 0.017820345237851143, + 0.01395449973642826, + 0.011880958452820778, + 0.007958374917507172, + 0.009020404890179634, + 0.013566958718001842, + 0.010176216252148151, + 0.017854241654276848, + 0.010502029210329056, + 0.011731092818081379, + 0.00794231053441763, + 0.006926290690898895, + 0.014434801414608955, + 0.00852557085454464, + 0.008209748193621635, + 0.010397978127002716, + 0.013727467507123947, + 0.014070868492126465, + 0.017910953611135483, + 0.010660340078175068, + 0.009878487326204777, + 0.009275680407881737, + 0.011779668740928173, + 0.020229522138834, + 0.005033728666603565, + 0.00599575275555253, + 0.008150309324264526, + 0.01033819280564785, + 0.008256064727902412, + 0.005993619095534086, + 0.01161989476531744, + 0.0060621523298323154 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 8 + ], + "selected_reward": -0.005252846051007509 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "medium shot, at eye level, shot with a shallow depth of field against a softly blurred, clean white background, frames a tall, transparent cylindrical glass filled with pristine, still water. The camera begins a slow, deliberate push-in, moving ever closer as a white, circular pencil tablet is placed at its base. Instantly, the tablet erupts with a violent hissing, releasing a turbulent, chaotic cloud of tiny bubbles that swirl upward in a frantic dance. A crisp, bright backlight illuminates the scene, making each individual bubble shimmer like a tiny crystal and casting a radiant glow through the glass. As the camera continues its steady advance into extreme macro close-up, it captures the tablet's surface disintegrating, steadily releasing a milky plume that tumbles and pours into the clear water. The action is a mesmerizing spectacle, brimming with energy, a furious, silent explosion of rebirth, creating an atmosphere of vibrant transformation and clean, powerful force.", + "seed": 14267, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000101/group_8/index_4709_seed_14267_dp_16_rank_32_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 8, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.022365172486752273, + "vlm_logprob": -0.022365172486752273, + "vlm_yes_no_margin": 3.975, + "vlm_token_credit": -0.022365172486752273, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene framed as a medium shot at eye level with a shallow depth of field against a clean white background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a tall, transparent cylindrical glass filled with still water visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a white circular pencil tablet appear at the base of the glass and erupt with a turbulent cloud of tiny bubbles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera begin with a slow push-in that continues into an extreme macro close-up?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the tablet surface shown disintegrating and releasing a milky plume into the water?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03100270964205265 + ], + "opposite_logprob_tokens": [ + -3.6560027599334717 + ], + "logprob_reward": -0.03100270964205265, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.03100270964205265, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07781757414340973, + 0.038052335381507874, + 0.15738415718078613, + 0.0974486917257309, + 0.03836047276854515, + 0.033432476222515106, + 0.04978335276246071, + 0.08647606521844864, + 0.20762597024440765, + 0.2136189490556717 + ], + "credit_map": [ + 0.009195133112370968, + 0.0132713308557868, + 0.005706341005861759, + 0.00758026447147131, + 0.013207283802330494, + 0.009873255155980587, + 0.01165743451565504, + 0.007326530758291483, + 0.005148550495505333, + 0.0046172416768968105, + 0.005397973116487265, + 0.0034537382889539003, + 0.005058261565864086, + 0.007818775251507759, + 0.0040860590524971485, + 0.0024717352353036404, + 0.02172105573117733, + 0.02170909382402897, + 0.018566954880952835, + 0.012536386027932167, + 0.02796294540166855, + 0.01725398749113083, + 0.020036820322275162, + 0.01759689673781395, + 0.0189487524330616, + 0.011144622229039669, + 0.021625187247991562, + 0.012226372957229614, + 0.009030151180922985, + 0.007305520586669445, + 0.0075076306238770485, + 0.009660456329584122, + 0.004594413097947836, + 0.007410193327814341, + 0.003670210251584649, + 0.004686435218900442, + 0.006037887651473284, + 0.004219436086714268, + 0.004364573396742344, + 0.0033773237373679876, + 0.003064744407311082, + 0.010677507147192955, + 0.005593332927674055, + 0.002622708911076188, + 0.0033291501458734274, + 0.003739045700058341, + 0.0022700035478919744, + 0.0021359811071306467, + 0.00535811111330986, + 0.007822501473128796, + 0.0060134707018733025, + 0.006440100260078907, + 0.008510100655257702, + 0.005911422893404961, + 0.004963076673448086, + 0.004764570388942957, + 0.006997648626565933, + 0.013159729540348053, + 0.00973176583647728, + 0.010382888838648796, + 0.006485138554126024, + 0.01120621059089899, + 0.010985047556459904, + 0.017527641728520393, + 0.018070854246616364, + 0.029222900047898293, + 0.020554913207888603, + 0.02617214247584343, + 0.02266920916736126, + 0.051622580736875534, + 0.02069382555782795, + 0.01861952245235443, + 0.022603850811719894, + 0.022877177223563194, + 0.030551038682460785, + 0.021754007786512375, + 0.021199464797973633, + 0.044324491173028946, + 0.023539099842309952, + 0.026769813150167465 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 8, + 2 + ], + "selected_reward": -0.03100270964205265 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene framed as a medium shot at eye level with a shallow depth of field against a clean white background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a tall, transparent cylindrical glass filled with still water visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a white circular pencil tablet appear at the base of the glass and erupt with a turbulent cloud of tiny bubbles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera begin with a slow push-in that continues into an extreme macro close-up?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the tablet surface shown disintegrating and releasing a milky plume into the water?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.020606013014912605 + ], + "opposite_logprob_tokens": [ + -3.89560604095459 + ], + "logprob_reward": -0.020606013014912605, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.020606013014912605, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0674934834241867, + 0.05075136944651604, + 0.48593825101852417, + 0.07942160964012146, + 0.025082970038056374, + 0.012937420979142189, + 0.020088719204068184, + 0.022233080118894577, + 0.18713538348674774, + 0.04891769960522652 + ], + "credit_map": [ + 0.008094456978142262, + 0.012146295048296452, + 0.005103449337184429, + 0.004916757810860872, + 0.01076891366392374, + 0.009325901046395302, + 0.012519946321845055, + 0.00461776927113533, + 0.005291370674967766, + 0.005789659451693296, + 0.008574012666940689, + 0.004681676626205444, + 0.006258498411625624, + 0.010996771976351738, + 0.005171440541744232, + 0.003987943287938833, + 0.06328848004341125, + 0.06551870703697205, + 0.054327819496393204, + 0.046374525874853134, + 0.08218985050916672, + 0.049564994871616364, + 0.07318701595067978, + 0.05148688331246376, + 0.019671965390443802, + 0.009404858574271202, + 0.018272293731570244, + 0.005050602834671736, + 0.008086366578936577, + 0.006815018597990274, + 0.005995456129312515, + 0.0061250473372638226, + 0.0027676536701619625, + 0.00635121576488018, + 0.0021846944000571966, + 0.003527509281411767, + 0.0032784321811050177, + 0.003024814184755087, + 0.0021752670872956514, + 0.0017733839340507984, + 0.0011928942985832691, + 0.0022460082545876503, + 0.0013033495051786304, + 0.0025037836749106646, + 0.0016459011239930987, + 0.0014117099344730377, + 0.0014926108997315168, + 0.0011411645682528615, + 0.0019059974001720548, + 0.002246981021016836, + 0.0034579995553940535, + 0.0036163816694170237, + 0.002439196454361081, + 0.0027169857639819384, + 0.0017464914126321673, + 0.0019586870912462473, + 0.0016340432921424508, + 0.0034447559155523777, + 0.003652560757473111, + 0.0036864650901407003, + 0.002267887583002448, + 0.0031178088393062353, + 0.001973371021449566, + 0.0024561902973800898, + 0.013825083151459694, + 0.02210298553109169, + 0.02603388950228691, + 0.024194959551095963, + 0.020691458135843277, + 0.03664379194378853, + 0.018502606078982353, + 0.02514061890542507, + 0.004975940100848675, + 0.005323316901922226, + 0.012520786374807358, + 0.005290375556796789, + 0.00442864466458559, + 0.009479251690208912, + 0.0039528547786176205, + 0.0029465295374393463 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 8, + 3 + ], + "selected_reward": -0.020606013014912605 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene framed as a medium shot at eye level with a shallow depth of field against a clean white background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a tall, transparent cylindrical glass filled with still water visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a white circular pencil tablet appear at the base of the glass and erupt with a turbulent cloud of tiny bubbles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera begin with a slow push-in that continues into an extreme macro close-up?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the tablet surface shown disintegrating and releasing a milky plume into the water?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005949051585048437 + ], + "opposite_logprob_tokens": [ + -5.130949020385742 + ], + "logprob_reward": -0.005949051585048437, + "yes_no_margin_reward": 5.125, + "token_credit_reward": -0.005949051585048437, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07411132007837296, + 0.13265489041805267, + 0.18147556483745575, + 0.16260188817977905, + 0.0722000002861023, + 0.03478914871811867, + 0.060151975601911545, + 0.06367483735084534, + 0.09886923432350159, + 0.11947114765644073 + ], + "credit_map": [ + 0.008113413117825985, + 0.01438070833683014, + 0.007136663421988487, + 0.00793713890016079, + 0.013366392813622952, + 0.009904026053845882, + 0.006080996245145798, + 0.007191977463662624, + 0.015059641562402248, + 0.015561404637992382, + 0.026490213349461555, + 0.008407961577177048, + 0.014893542975187302, + 0.035186171531677246, + 0.00944722630083561, + 0.007608717307448387, + 0.021183842793107033, + 0.027691995725035667, + 0.024785475805401802, + 0.013491724617779255, + 0.03808562085032463, + 0.024161966517567635, + 0.014762632548809052, + 0.01731230691075325, + 0.04488004744052887, + 0.011295823380351067, + 0.04670396447181702, + 0.008899169974029064, + 0.016552800312638283, + 0.01293033454567194, + 0.01041346974670887, + 0.01092627365142107, + 0.008043690584599972, + 0.023301417008042336, + 0.004641890991479158, + 0.007918116636574268, + 0.009297157637774944, + 0.009509197436273098, + 0.004317091777920723, + 0.005171430297195911, + 0.003089515957981348, + 0.005245620850473642, + 0.0032223830930888653, + 0.004521285183727741, + 0.0038372259587049484, + 0.003741507651284337, + 0.0057873809710145, + 0.005344226490706205, + 0.00463556544855237, + 0.007980229333043098, + 0.00771817471832037, + 0.008869835175573826, + 0.007733032573014498, + 0.007291931193321943, + 0.00841954629868269, + 0.007503656670451164, + 0.006675281561911106, + 0.016112839803099632, + 0.00609622523188591, + 0.006367567460983992, + 0.00393541669473052, + 0.007816372439265251, + 0.00382463657297194, + 0.01284649595618248, + 0.00800241157412529, + 0.01474957074970007, + 0.01088692806661129, + 0.013673841953277588, + 0.011415494605898857, + 0.01705661602318287, + 0.013433963060379028, + 0.0096503971144557, + 0.01673009991645813, + 0.01633354276418686, + 0.015776440501213074, + 0.015902649611234665, + 0.01593872532248497, + 0.017190881073474884, + 0.011374874040484428, + 0.010223924182355404 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 3, + 1 + ], + "selected_reward": -0.005949051585048437 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene framed as a medium shot at eye level with a shallow depth of field against a clean white background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a tall, transparent cylindrical glass filled with still water visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a white circular pencil tablet appear at the base of the glass and erupt with a turbulent cloud of tiny bubbles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera begin with a slow push-in that continues into an extreme macro close-up?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the tablet surface shown disintegrating and releasing a milky plume into the water?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.020574599504470825 + ], + "opposite_logprob_tokens": [ + -3.8955745697021484 + ], + "logprob_reward": -0.020574599504470825, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.020574599504470825, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08907229453325272, + 0.05565834417939186, + 0.040164798498153687, + 0.05462756007909775, + 0.07215403020381927, + 0.06827547401189804, + 0.08072736114263535, + 0.1692206710577011, + 0.2242746353149414, + 0.1458248496055603 + ], + "credit_map": [ + 0.013541467487812042, + 0.016895858570933342, + 0.009091648273169994, + 0.007694852538406849, + 0.013999001123011112, + 0.013340312987565994, + 0.007309349719434977, + 0.007199807092547417, + 0.006984683219343424, + 0.005424125585705042, + 0.011046219617128372, + 0.003988970536738634, + 0.0051896595396101475, + 0.014408516697585583, + 0.005358452443033457, + 0.003257713047787547, + 0.004770401865243912, + 0.00661495141685009, + 0.003901510965079069, + 0.005249519366770983, + 0.005544512066990137, + 0.0045069302432239056, + 0.006360737141221762, + 0.003216233104467392, + 0.011781754903495312, + 0.005212326534092426, + 0.01013848464936018, + 0.005478809121996164, + 0.005522709805518389, + 0.006117870099842548, + 0.005153392441570759, + 0.005222212988883257, + 0.009173102676868439, + 0.009695703163743019, + 0.00672401487827301, + 0.010018598288297653, + 0.010914386250078678, + 0.009446959011256695, + 0.00824388861656189, + 0.00793737918138504, + 0.005996725521981716, + 0.01348373107612133, + 0.009189811535179615, + 0.007879374548792839, + 0.008629662916064262, + 0.009320924058556557, + 0.007852866314351559, + 0.005922379903495312, + 0.008098446764051914, + 0.01617693528532982, + 0.00739020761102438, + 0.011417382396757603, + 0.013011816889047623, + 0.010641887784004211, + 0.007209423463791609, + 0.006781264208257198, + 0.015716107562184334, + 0.030337641015648842, + 0.019671324640512466, + 0.021464496850967407, + 0.01617000252008438, + 0.0184346791356802, + 0.011445067822933197, + 0.035981349647045135, + 0.017362207174301147, + 0.027565550059080124, + 0.024309901520609856, + 0.031183090060949326, + 0.024461396038532257, + 0.05330934002995491, + 0.02155485562980175, + 0.024528304114937782, + 0.014638899825513363, + 0.01475426834076643, + 0.03394027799367905, + 0.020931638777256012, + 0.014077669940888882, + 0.02003316953778267, + 0.014822982251644135, + 0.012625941075384617 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 7, + 9 + ], + "selected_reward": -0.020574599504470825 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene framed as a medium shot at eye level with a shallow depth of field against a clean white background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a tall, transparent cylindrical glass filled with still water visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a white circular pencil tablet appear at the base of the glass and erupt with a turbulent cloud of tiny bubbles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera begin with a slow push-in that continues into an extreme macro close-up?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the tablet surface shown disintegrating and releasing a milky plume into the water?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03369348868727684 + ], + "opposite_logprob_tokens": [ + -3.408693552017212 + ], + "logprob_reward": -0.03369348868727684, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.03369348868727684, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07866799831390381, + 0.09156728535890579, + 0.14089611172676086, + 0.25328218936920166, + 0.03048236109316349, + 0.038621220737695694, + 0.049809493124485016, + 0.0740593746304512, + 0.1286453902721405, + 0.11396852135658264 + ], + "credit_map": [ + 0.009459331631660461, + 0.019334839656949043, + 0.006874482613056898, + 0.006697053089737892, + 0.010576315224170685, + 0.01142860297113657, + 0.00819400791078806, + 0.006103367079049349, + 0.009883418679237366, + 0.01263425312936306, + 0.01569385826587677, + 0.006068037357181311, + 0.013417995534837246, + 0.019370151683688164, + 0.009199169464409351, + 0.00530041242018342, + 0.01920083351433277, + 0.01810358837246895, + 0.015051589347422123, + 0.016684770584106445, + 0.01889003999531269, + 0.015877753496170044, + 0.02135450579226017, + 0.015733027830719948, + 0.073029063642025, + 0.017948130145668983, + 0.0763198509812355, + 0.012371483258903027, + 0.024590283632278442, + 0.021000951528549194, + 0.010505040176212788, + 0.017517419531941414, + 0.004945538472384214, + 0.00942646712064743, + 0.0016906786477193236, + 0.002841238398104906, + 0.004743087105453014, + 0.0031600408256053925, + 0.0020671882666647434, + 0.0016081256326287985, + 0.0039983633905649185, + 0.01062476634979248, + 0.004217543173581362, + 0.004170795436948538, + 0.004420643672347069, + 0.004187104757875204, + 0.0037514884024858475, + 0.0032505211420357227, + 0.004428664688020945, + 0.006682788021862507, + 0.006402172148227692, + 0.0077249412424862385, + 0.008867631666362286, + 0.006534299813210964, + 0.0043395510874688625, + 0.00482945004478097, + 0.009222768247127533, + 0.02494288980960846, + 0.005738704930990934, + 0.0066749462857842445, + 0.006026159971952438, + 0.008785230107605457, + 0.0035485581029206514, + 0.009120127186179161, + 0.010064137168228626, + 0.015209639444947243, + 0.023033883422613144, + 0.021484948694705963, + 0.013773930259048939, + 0.012109718285501003, + 0.01628400757908821, + 0.01668514683842659, + 0.011835508048534393, + 0.013085294514894485, + 0.027825534343719482, + 0.01112446840852499, + 0.012411544099450111, + 0.021755164489150047, + 0.008903730660676956, + 0.007027293089777231 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 2, + 8 + ], + "selected_reward": -0.03369348868727684 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "medium shot, at eye level, shot with a shallow depth of field against a softly blurred, clean white background, frames a tall, transparent cylindrical glass filled with pristine, still water. The camera begins a slow, deliberate push-in, moving ever closer as a white, circular pencil tablet is placed at its base. Instantly, the tablet erupts with a violent hissing, releasing a turbulent, chaotic cloud of tiny bubbles that swirl upward in a frantic dance. A crisp, bright backlight illuminates the scene, making each individual bubble shimmer like a tiny crystal and casting a radiant glow through the glass. As the camera continues its steady advance into extreme macro close-up, it captures the tablet's surface disintegrating, steadily releasing a milky plume that tumbles and pours into the clear water. The action is a mesmerizing spectacle, brimming with energy, a furious, silent explosion of rebirth, creating an atmosphere of vibrant transformation and clean, powerful force.", + "seed": 14268, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000101/group_8/index_4709_seed_14268_dp_17_rank_34_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 8, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.03200002145022154, + "vlm_logprob": -0.03200002145022154, + "vlm_yes_no_margin": 3.65, + "vlm_token_credit": -0.03200002145022154, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene framed as a medium shot at eye level with a shallow depth of field against a clean white background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a tall, transparent cylindrical glass filled with still water visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a white circular pencil tablet appear at the base of the glass and erupt with a turbulent cloud of tiny bubbles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera begin with a slow push-in that continues into an extreme macro close-up?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the tablet surface shown disintegrating and releasing a milky plume into the water?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.045728910714387894 + ], + "opposite_logprob_tokens": [ + -3.295728921890259 + ], + "logprob_reward": -0.045728910714387894, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.045728910714387894, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07727012038230896, + 0.08475159108638763, + 0.12739138305187225, + 0.035859763622283936, + 0.0912473201751709, + 0.1786026656627655, + 0.055291954427957535, + 0.10523003339767456, + 0.10105136036872864, + 0.14330382645130157 + ], + "credit_map": [ + 0.006943323649466038, + 0.009916397742927074, + 0.007156158331781626, + 0.007816579192876816, + 0.0114372493699193, + 0.013273683376610279, + 0.011011638678610325, + 0.009715085849165916, + 0.008149728178977966, + 0.010639193467795849, + 0.011501330882310867, + 0.008321096189320087, + 0.01342389453202486, + 0.013348514214158058, + 0.007370106875896454, + 0.011997733265161514, + 0.013823381625115871, + 0.017824329435825348, + 0.014137384481728077, + 0.01875738613307476, + 0.018909301608800888, + 0.019030533730983734, + 0.014546792954206467, + 0.010362273082137108, + 0.00395163893699646, + 0.0037138385232537985, + 0.0037273168563842773, + 0.0035607803147286177, + 0.006873487960547209, + 0.004717203322798014, + 0.004170975182205439, + 0.005144523456692696, + 0.030341556295752525, + 0.0054231430403888226, + 0.02873215824365616, + 0.004822771064937115, + 0.005105368327349424, + 0.005072684492915869, + 0.004897908307611942, + 0.006851725745946169, + 0.0456889271736145, + 0.012372121214866638, + 0.053544823080301285, + 0.01394896861165762, + 0.012588177807629108, + 0.015282082371413708, + 0.010928365401923656, + 0.014249203726649284, + 0.0054606785997748375, + 0.006290670949965715, + 0.0069547127932310104, + 0.010159365832805634, + 0.006800469942390919, + 0.005252175964415073, + 0.006866947747766972, + 0.0075069330632686615, + 0.008537743240594864, + 0.008859921246767044, + 0.026567915454506874, + 0.018833568319678307, + 0.010540476068854332, + 0.009000900201499462, + 0.013669717125594616, + 0.009219794534146786, + 0.012336612679064274, + 0.010683081112802029, + 0.016411734744906425, + 0.01145430188626051, + 0.014744535088539124, + 0.017499195411801338, + 0.008770220912992954, + 0.009151674807071686, + 0.009252985939383507, + 0.015118354000151157, + 0.024193761870265007, + 0.028040282428264618, + 0.01453039888292551, + 0.009578523226082325, + 0.030197089537978172, + 0.012392424046993256 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 9, + 2 + ], + "selected_reward": -0.045728910714387894 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene framed as a medium shot at eye level with a shallow depth of field against a clean white background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a tall, transparent cylindrical glass filled with still water visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a white circular pencil tablet appear at the base of the glass and erupt with a turbulent cloud of tiny bubbles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera begin with a slow push-in that continues into an extreme macro close-up?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the tablet surface shown disintegrating and releasing a milky plume into the water?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03812851384282112 + ], + "opposite_logprob_tokens": [ + -3.288128614425659 + ], + "logprob_reward": -0.03812851384282112, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.03812851384282112, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.055399488657712936, + 0.09458190202713013, + 0.07867289334535599, + 0.42802882194519043, + 0.16367436945438385, + 0.05754305049777031, + 0.019119447097182274, + 0.034192174673080444, + 0.04906722530722618, + 0.019720658659934998 + ], + "credit_map": [ + 0.005654835142195225, + 0.010387425310909748, + 0.0036560555454343557, + 0.004346710164099932, + 0.012380030937492847, + 0.008205531165003777, + 0.006711435969918966, + 0.0040574632585048676, + 0.008461395278573036, + 0.009973036125302315, + 0.010594971477985382, + 0.012385690584778786, + 0.01266707293689251, + 0.013387785293161869, + 0.008559279143810272, + 0.018552672117948532, + 0.00996413454413414, + 0.008816559799015522, + 0.01037768553942442, + 0.007928564213216305, + 0.013991914689540863, + 0.012936295010149479, + 0.007175175007432699, + 0.007482570130378008, + 0.09137691557407379, + 0.031122690066695213, + 0.10867425054311752, + 0.03721815347671509, + 0.040891434997320175, + 0.035378023982048035, + 0.03820384293794632, + 0.04516350477933884, + 0.025570658966898918, + 0.025370048359036446, + 0.018062224611639977, + 0.01988273113965988, + 0.021769458428025246, + 0.017167864367365837, + 0.018369652330875397, + 0.017481742426753044, + 0.013806965202093124, + 0.004785143304616213, + 0.01771116442978382, + 0.0033233147114515305, + 0.005221611820161343, + 0.0058158221654593945, + 0.002714933827519417, + 0.004164096433669329, + 0.002079674741253257, + 0.0021975282579660416, + 0.003257842967286706, + 0.0027842498384416103, + 0.002722739242017269, + 0.0020520398393273354, + 0.002237156266346574, + 0.0017882164102047682, + 0.0025786105543375015, + 0.0029577696695923805, + 0.007564447820186615, + 0.005255045834928751, + 0.0028804540634155273, + 0.00468347268179059, + 0.00452415132895112, + 0.0037482227198779583, + 0.003700586035847664, + 0.0055661071091890335, + 0.006054082419723272, + 0.0055394480004906654, + 0.008044244721531868, + 0.008611463941633701, + 0.0075357635505497456, + 0.004015529528260231, + 0.0016085702227428555, + 0.002915468066930771, + 0.0017344033112749457, + 0.002385837724432349, + 0.0039421566762030125, + 0.002458328614011407, + 0.0028712796047329903, + 0.0018046145560219884 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 4, + 1 + ], + "selected_reward": -0.03812851384282112 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene framed as a medium shot at eye level with a shallow depth of field against a clean white background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a tall, transparent cylindrical glass filled with still water visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a white circular pencil tablet appear at the base of the glass and erupt with a turbulent cloud of tiny bubbles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera begin with a slow push-in that continues into an extreme macro close-up?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the tablet surface shown disintegrating and releasing a milky plume into the water?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.006737728603184223 + ], + "opposite_logprob_tokens": [ + -5.00673770904541 + ], + "logprob_reward": -0.006737728603184223, + "yes_no_margin_reward": 5.0, + "token_credit_reward": -0.006737728603184223, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08682587742805481, + 0.10913241654634476, + 0.051708776503801346, + 0.12882092595100403, + 0.20783668756484985, + 0.10673735290765762, + 0.1354862004518509, + 0.05112072080373764, + 0.07860305160284042, + 0.043728020042181015 + ], + "credit_map": [ + 0.006822424009442329, + 0.010815752670168877, + 0.008100559003651142, + 0.010466177016496658, + 0.014750343747437, + 0.013181357644498348, + 0.011100461706519127, + 0.011588802561163902, + 0.008811119012534618, + 0.009585477411746979, + 0.011833431199193, + 0.012936916202306747, + 0.020832424983382225, + 0.01644403487443924, + 0.010006644763052464, + 0.018682366237044334, + 0.005458206404000521, + 0.005943886935710907, + 0.006867998279631138, + 0.005787491798400879, + 0.009246828965842724, + 0.008320124819874763, + 0.005296036135405302, + 0.004788205958902836, + 0.030143871903419495, + 0.009843379259109497, + 0.03313382714986801, + 0.0062141334637999535, + 0.01937938667833805, + 0.010026020929217339, + 0.01134544424712658, + 0.00873485952615738, + 0.04077436774969101, + 0.016775427386164665, + 0.03667399659752846, + 0.02469564415514469, + 0.027888905256986618, + 0.01869169995188713, + 0.02131500095129013, + 0.02102164924144745, + 0.014358976855874062, + 0.010470762848854065, + 0.01534617505967617, + 0.013967339880764484, + 0.016429778188467026, + 0.012566108256578445, + 0.011841473169624805, + 0.01175673771649599, + 0.010990099981427193, + 0.013523750007152557, + 0.024151306599378586, + 0.022614164277911186, + 0.02297939360141754, + 0.007394223008304834, + 0.020500950515270233, + 0.013332311064004898, + 0.004094578791409731, + 0.0032894613686949015, + 0.012340952642261982, + 0.006674020551145077, + 0.004598530009388924, + 0.006758565083146095, + 0.008348232135176659, + 0.00501638138666749, + 0.006036560051143169, + 0.00811083521693945, + 0.008941443637013435, + 0.007208387833088636, + 0.01432640291750431, + 0.008845945820212364, + 0.01851162128150463, + 0.006621850188821554, + 0.0036918523255735636, + 0.006424603518098593, + 0.003711278550326824, + 0.007263122592121363, + 0.007243784610182047, + 0.004887579474598169, + 0.005805215798318386, + 0.0047005824744701385 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 6, + 3 + ], + "selected_reward": -0.006737728603184223 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene framed as a medium shot at eye level with a shallow depth of field against a clean white background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a tall, transparent cylindrical glass filled with still water visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a white circular pencil tablet appear at the base of the glass and erupt with a turbulent cloud of tiny bubbles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera begin with a slow push-in that continues into an extreme macro close-up?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the tablet surface shown disintegrating and releasing a milky plume into the water?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.026341667398810387 + ], + "opposite_logprob_tokens": [ + -3.651341676712036 + ], + "logprob_reward": -0.026341667398810387, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.026341667398810387, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.02645229920744896, + 0.03263293206691742, + 0.08276938647031784, + 0.18594905734062195, + 0.0691266879439354, + 0.32992634177207947, + 0.05198080837726593, + 0.031627632677555084, + 0.07353625446557999, + 0.11599859595298767 + ], + "credit_map": [ + 0.00217761960811913, + 0.004083223175257444, + 0.0023520183749496937, + 0.0027904438320547342, + 0.0041871629655361176, + 0.004181715659797192, + 0.004048241768032312, + 0.002631870796903968, + 0.0034310081973671913, + 0.00360282976180315, + 0.005369279067963362, + 0.0035289209336042404, + 0.004216840956360102, + 0.00483246985822916, + 0.0026267063803970814, + 0.005024876445531845, + 0.012329645454883575, + 0.007776214275509119, + 0.00545690068975091, + 0.007393667008727789, + 0.0181877501308918, + 0.018811337649822235, + 0.008463378995656967, + 0.004350487608462572, + 0.06554918736219406, + 0.008482493460178375, + 0.0696367472410202, + 0.0067560686729848385, + 0.006874498911201954, + 0.01239908579736948, + 0.005062959622591734, + 0.011188012547791004, + 0.018197020515799522, + 0.008885027840733528, + 0.016461782157421112, + 0.004569380544126034, + 0.006887875031679869, + 0.004960027523338795, + 0.004166168160736561, + 0.004999399185180664, + 0.09995122998952866, + 0.01733420602977276, + 0.11186125129461288, + 0.020760245621204376, + 0.02019178308546543, + 0.022723881527781487, + 0.01577150635421276, + 0.021332189440727234, + 0.0039835767820477486, + 0.005860749166458845, + 0.007168714422732592, + 0.0072882757522165775, + 0.00864739716053009, + 0.004825613461434841, + 0.009001045487821102, + 0.00520543335005641, + 0.003330381354317069, + 0.0021710272412747145, + 0.007163197733461857, + 0.004053301643580198, + 0.0027852212078869343, + 0.003075433662161231, + 0.0037533019203692675, + 0.0052957674488425255, + 0.005967524368315935, + 0.008838129229843616, + 0.006904917303472757, + 0.007861727848649025, + 0.014461933635175228, + 0.012594139203429222, + 0.011556807905435562, + 0.005351073108613491, + 0.007183737587183714, + 0.013780544511973858, + 0.01575818471610546, + 0.02402500994503498, + 0.011046159081161022, + 0.010740582831203938, + 0.024687206372618675, + 0.008777165785431862 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 3, + 9 + ], + "selected_reward": -0.026341667398810387 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene framed as a medium shot at eye level with a shallow depth of field against a clean white background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a tall, transparent cylindrical glass filled with still water visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a white circular pencil tablet appear at the base of the glass and erupt with a turbulent cloud of tiny bubbles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera begin with a slow push-in that continues into an extreme macro close-up?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the tablet surface shown disintegrating and releasing a milky plume into the water?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04306328669190407 + ], + "opposite_logprob_tokens": [ + -3.1680634021759033 + ], + "logprob_reward": -0.04306328669190407, + "yes_no_margin_reward": 3.125, + "token_credit_reward": -0.04306328669190407, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06171286478638649, + 0.04471958801150322, + 0.0437488779425621, + 0.12401799112558365, + 0.09054793417453766, + 0.28742554783821106, + 0.09026496112346649, + 0.08700964599847794, + 0.05280442535877228, + 0.11774812638759613 + ], + "credit_map": [ + 0.005628274753689766, + 0.00947666261345148, + 0.005402301903814077, + 0.004421417601406574, + 0.013684693723917007, + 0.01097059715539217, + 0.007223920896649361, + 0.004904997069388628, + 0.004347939509898424, + 0.004097397904843092, + 0.005040125921368599, + 0.004548508208245039, + 0.008770317770540714, + 0.006822521332651377, + 0.004411994479596615, + 0.006680783350020647, + 0.0057508088648319244, + 0.003963342402130365, + 0.005305496975779533, + 0.004512862768024206, + 0.007107362616807222, + 0.01001717522740364, + 0.004334317985922098, + 0.002757513429969549, + 0.03706202283501625, + 0.009250693954527378, + 0.0347268283367157, + 0.007136162836104631, + 0.010740670375525951, + 0.009386696852743626, + 0.008516283705830574, + 0.007198627572506666, + 0.014864630997180939, + 0.012511986307799816, + 0.011983689852058887, + 0.009015361778438091, + 0.016583075746893883, + 0.0076169781386852264, + 0.010669665411114693, + 0.007302547805011272, + 0.07018841058015823, + 0.020823342725634575, + 0.06464836001396179, + 0.02452065981924534, + 0.04177508503198624, + 0.022850604727864265, + 0.023897169157862663, + 0.018721915781497955, + 0.007210887968540192, + 0.009805271402001381, + 0.011255700141191483, + 0.013335133902728558, + 0.016510559245944023, + 0.007939253933727741, + 0.016059352084994316, + 0.008148803375661373, + 0.0066784219816327095, + 0.006234286818653345, + 0.02492259442806244, + 0.013639310374855995, + 0.00699665816500783, + 0.006849448662251234, + 0.014092881232500076, + 0.007596042472869158, + 0.005708200391381979, + 0.006106246262788773, + 0.006832796614617109, + 0.005019021220505238, + 0.0075698839500546455, + 0.012322064489126205, + 0.005510574672371149, + 0.003735634498298168, + 0.006094601470977068, + 0.010390178300440311, + 0.026145219802856445, + 0.010030705481767654, + 0.012634743005037308, + 0.00814756564795971, + 0.03517545387148857, + 0.009129658341407776 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 3, + 9 + ], + "selected_reward": -0.04306328669190407 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "medium shot, at eye level, shot with a shallow depth of field against a softly blurred, clean white background, frames a tall, transparent cylindrical glass filled with pristine, still water. The camera begins a slow, deliberate push-in, moving ever closer as a white, circular pencil tablet is placed at its base. Instantly, the tablet erupts with a violent hissing, releasing a turbulent, chaotic cloud of tiny bubbles that swirl upward in a frantic dance. A crisp, bright backlight illuminates the scene, making each individual bubble shimmer like a tiny crystal and casting a radiant glow through the glass. As the camera continues its steady advance into extreme macro close-up, it captures the tablet's surface disintegrating, steadily releasing a milky plume that tumbles and pours into the clear water. The action is a mesmerizing spectacle, brimming with energy, a furious, silent explosion of rebirth, creating an atmosphere of vibrant transformation and clean, powerful force.", + "seed": 14269, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000101/group_8/index_4709_seed_14269_dp_17_rank_34_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 8, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.04487281385809183, + "vlm_logprob": -0.04487281385809183, + "vlm_yes_no_margin": 3.3, + "vlm_token_credit": -0.04487281385809183, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene framed as a medium shot at eye level with a shallow depth of field against a clean white background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a tall, transparent cylindrical glass filled with still water visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a white circular pencil tablet appear at the base of the glass and erupt with a turbulent cloud of tiny bubbles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera begin with a slow push-in that continues into an extreme macro close-up?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the tablet surface shown disintegrating and releasing a milky plume into the water?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07445864379405975 + ], + "opposite_logprob_tokens": [ + -2.824458599090576 + ], + "logprob_reward": -0.07445864379405975, + "yes_no_margin_reward": 2.75, + "token_credit_reward": -0.07445864379405975, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1552709937095642, + 0.059173330664634705, + 0.057930804789066315, + 0.03648122772574425, + 0.02542640082538128, + 0.09542717039585114, + 0.36226800084114075, + 0.06595443189144135, + 0.08930907398462296, + 0.05275856703519821 + ], + "credit_map": [ + 0.022100893780589104, + 0.024724552407860756, + 0.020445087924599648, + 0.01640464924275875, + 0.01797374337911606, + 0.02308536507189274, + 0.011373857036232948, + 0.01916285790503025, + 0.006356746889650822, + 0.011440729722380638, + 0.0055723791010677814, + 0.005363902542740107, + 0.010477989912033081, + 0.00915260799229145, + 0.004199206829071045, + 0.006609770003706217, + 0.009779165498912334, + 0.004868448246270418, + 0.008839639835059643, + 0.006569928023964167, + 0.00787802692502737, + 0.007882388308644295, + 0.004730090033262968, + 0.007383119780570269, + 0.003981651738286018, + 0.003719112602993846, + 0.00440949434414506, + 0.005908744875341654, + 0.003956479951739311, + 0.007219261024147272, + 0.0034640931989997625, + 0.003822387894615531, + 0.0033826788421720266, + 0.0034555760212242603, + 0.002764904173091054, + 0.0027210533153265715, + 0.004332803189754486, + 0.003216253127902746, + 0.002498409477993846, + 0.0030547238420695066, + 0.00728171830996871, + 0.015493799932301044, + 0.008711153641343117, + 0.012772932648658752, + 0.019171632826328278, + 0.010688459500670433, + 0.01264621876180172, + 0.008661255240440369, + 0.03251959756016731, + 0.05980540066957474, + 0.031072868034243584, + 0.05018892139196396, + 0.07021652907133102, + 0.04224051162600517, + 0.037635043263435364, + 0.03858914226293564, + 0.006483511999249458, + 0.006508997175842524, + 0.00788662675768137, + 0.010898793116211891, + 0.010613015852868557, + 0.00681030610576272, + 0.010191991925239563, + 0.006561188958585262, + 0.013014477677643299, + 0.019106335937976837, + 0.007814249023795128, + 0.008186743594706059, + 0.011717201210558414, + 0.008837508969008923, + 0.011265156790614128, + 0.009367401711642742, + 0.006333127152174711, + 0.010041426867246628, + 0.005776160396635532, + 0.0055984435603022575, + 0.007551704067736864, + 0.005067329853773117, + 0.005428147502243519, + 0.0069622257724404335 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 0, + 5 + ], + "selected_reward": -0.07445864379405975 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene framed as a medium shot at eye level with a shallow depth of field against a clean white background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a tall, transparent cylindrical glass filled with still water visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a white circular pencil tablet appear at the base of the glass and erupt with a turbulent cloud of tiny bubbles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera begin with a slow push-in that continues into an extreme macro close-up?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the tablet surface shown disintegrating and releasing a milky plume into the water?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04309365898370743 + ], + "opposite_logprob_tokens": [ + -3.168093681335449 + ], + "logprob_reward": -0.04309365898370743, + "yes_no_margin_reward": 3.125, + "token_credit_reward": -0.04309365898370743, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.040223266929388046, + 0.04427877813577652, + 0.02219894342124462, + 0.016793308779597282, + 0.13144941627979279, + 0.5015948414802551, + 0.0931321308016777, + 0.05500379949808121, + 0.04547516629099846, + 0.04985031485557556 + ], + "credit_map": [ + 0.0039667109958827496, + 0.006763697136193514, + 0.003484465414658189, + 0.00302483094856143, + 0.006669966969639063, + 0.006521122995764017, + 0.0036995415575802326, + 0.006092930678278208, + 0.00487494794651866, + 0.005715412087738514, + 0.005096218548715115, + 0.004207786172628403, + 0.008357329294085503, + 0.006505209021270275, + 0.0040579428896307945, + 0.005463933106511831, + 0.00225920625962317, + 0.002476536203175783, + 0.0020764200016856194, + 0.002420143224298954, + 0.005916090682148933, + 0.0030715458560734987, + 0.0018147459486499429, + 0.0021642546635121107, + 0.0017150655621662736, + 0.0017748100217431784, + 0.0018598249880596995, + 0.0018398563843220472, + 0.0032756042201071978, + 0.00257920753210783, + 0.0019856009166687727, + 0.001763337873853743, + 0.014946980401873589, + 0.020675642415881157, + 0.01335885375738144, + 0.01636543869972229, + 0.024004150182008743, + 0.016255222260951996, + 0.01376589760184288, + 0.012077221646904945, + 0.04299386590719223, + 0.07299534231424332, + 0.05258646979928017, + 0.07730607688426971, + 0.08317840099334717, + 0.05391550436615944, + 0.06833633035421371, + 0.05028289183974266, + 0.008878933265805244, + 0.016423888504505157, + 0.007872982881963253, + 0.011595883406698704, + 0.01943092793226242, + 0.009974884800612926, + 0.009865188971161842, + 0.009089434519410133, + 0.006702159997075796, + 0.005436836741864681, + 0.004922972060739994, + 0.00842025876045227, + 0.010108798742294312, + 0.00634752307087183, + 0.008227330632507801, + 0.004837917163968086, + 0.006171667017042637, + 0.009817209094762802, + 0.0047143250703811646, + 0.004816744942218065, + 0.0062614381313323975, + 0.004894996527582407, + 0.00510426377877593, + 0.0036945233587175608, + 0.006063143722712994, + 0.008781635202467442, + 0.003565204795449972, + 0.0038101531099528074, + 0.012751989997923374, + 0.005968718323856592, + 0.004400026053190231, + 0.0045094480738043785 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 4, + 6 + ], + "selected_reward": -0.04309365898370743 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene framed as a medium shot at eye level with a shallow depth of field against a clean white background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a tall, transparent cylindrical glass filled with still water visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a white circular pencil tablet appear at the base of the glass and erupt with a turbulent cloud of tiny bubbles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera begin with a slow push-in that continues into an extreme macro close-up?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the tablet surface shown disintegrating and releasing a milky plume into the water?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.011075449176132679 + ], + "opposite_logprob_tokens": [ + -4.511075496673584 + ], + "logprob_reward": -0.011075449176132679, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.011075449176132679, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.04823082312941551, + 0.03926660120487213, + 0.018889043480157852, + 0.012722824700176716, + 0.0910547524690628, + 0.3089982271194458, + 0.16628967225551605, + 0.15975184738636017, + 0.07558440417051315, + 0.07921185344457626 + ], + "credit_map": [ + 0.006618307437747717, + 0.006859285291284323, + 0.005703544244170189, + 0.004843380302190781, + 0.007210855837911367, + 0.007418177556246519, + 0.00468067079782486, + 0.004896597471088171, + 0.004834611900150776, + 0.004334917291998863, + 0.005433693528175354, + 0.003782395739108324, + 0.005581434350460768, + 0.005809211637824774, + 0.0036571796517819166, + 0.005833157338202, + 0.002513574669137597, + 0.0021473229862749577, + 0.002286678645759821, + 0.0017077865777537227, + 0.004062610678374767, + 0.002526704454794526, + 0.0020553129725158215, + 0.0015890495851635933, + 0.0011207343777641654, + 0.0013852949487045407, + 0.0014706689398735762, + 0.001428247895091772, + 0.0027943847235292196, + 0.0016602353425696492, + 0.0013735293177887797, + 0.0014897276414558291, + 0.010635661892592907, + 0.015092478133738041, + 0.009349456988275051, + 0.011501586064696312, + 0.015858251601457596, + 0.010717685334384441, + 0.008941889740526676, + 0.008957736194133759, + 0.02473781444132328, + 0.04589866101741791, + 0.030009735375642776, + 0.04646662250161171, + 0.055508825927972794, + 0.03376073017716408, + 0.04418537765741348, + 0.02843041718006134, + 0.014406170696020126, + 0.02767850086092949, + 0.01438842248171568, + 0.02187865972518921, + 0.03480661287903786, + 0.01847737468779087, + 0.01760757900774479, + 0.01704634539783001, + 0.014257505536079407, + 0.01739123836159706, + 0.017325513064861298, + 0.026549944654107094, + 0.028616366907954216, + 0.015207869000732899, + 0.02425352856516838, + 0.01614985801279545, + 0.009967917576432228, + 0.012185940518975258, + 0.006394840311259031, + 0.006591816898435354, + 0.018561920151114464, + 0.01134861446917057, + 0.005490254145115614, + 0.005043089389801025, + 0.010361605323851109, + 0.014153590425848961, + 0.009188227355480194, + 0.008072979748249054, + 0.01154530793428421, + 0.00929594598710537, + 0.009004952386021614, + 0.007589238230139017 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 6, + 7 + ], + "selected_reward": -0.011075449176132679 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene framed as a medium shot at eye level with a shallow depth of field against a clean white background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a tall, transparent cylindrical glass filled with still water visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a white circular pencil tablet appear at the base of the glass and erupt with a turbulent cloud of tiny bubbles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera begin with a slow push-in that continues into an extreme macro close-up?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the tablet surface shown disintegrating and releasing a milky plume into the water?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.06203775480389595 + ], + "opposite_logprob_tokens": [ + -2.812037706375122 + ], + "logprob_reward": -0.06203775480389595, + "yes_no_margin_reward": 2.75, + "token_credit_reward": -0.06203775480389595, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06394712626934052, + 0.02345273084938526, + 0.034028418362140656, + 0.037315353751182556, + 0.028137417510151863, + 0.057950209826231, + 0.2358548790216446, + 0.04818296432495117, + 0.3192598521709442, + 0.15187107026576996 + ], + "credit_map": [ + 0.010268229991197586, + 0.008045898750424385, + 0.007802121341228485, + 0.006515270099043846, + 0.007858329452574253, + 0.010144727304577827, + 0.005525152664631605, + 0.007787391543388367, + 0.0032684532925486565, + 0.0027059076819568872, + 0.0025387268979102373, + 0.002483041724190116, + 0.0033210618421435356, + 0.003646241966634989, + 0.0026588996406644583, + 0.0028303966391831636, + 0.003137540305033326, + 0.004542618989944458, + 0.002886982634663582, + 0.0034495671279728413, + 0.009657514281570911, + 0.005070078186690807, + 0.002799184527248144, + 0.002484929747879505, + 0.004553457722067833, + 0.0038292331155389547, + 0.004306960850954056, + 0.006617318838834763, + 0.004883668385446072, + 0.005563024431467056, + 0.003963455092161894, + 0.003598232986405492, + 0.0031638743821531534, + 0.004471563268452883, + 0.002654534298926592, + 0.0024741184897720814, + 0.005943657830357552, + 0.00324683147482574, + 0.0030564190819859505, + 0.003126416588202119, + 0.005929434671998024, + 0.008018981665372849, + 0.00584788341075182, + 0.0072204796597361565, + 0.011531168594956398, + 0.006931285373866558, + 0.006318217143416405, + 0.006152763031423092, + 0.023178938776254654, + 0.04410293698310852, + 0.018243303522467613, + 0.0316895954310894, + 0.04193592071533203, + 0.027816148474812508, + 0.02463185228407383, + 0.02425619401037693, + 0.00742315175011754, + 0.00484015978872776, + 0.004875784274190664, + 0.007186353672295809, + 0.006779890973120928, + 0.00396216893568635, + 0.006200100295245647, + 0.006915356032550335, + 0.04254451021552086, + 0.05452955886721611, + 0.023374857380986214, + 0.02842392586171627, + 0.07636264711618423, + 0.05006025731563568, + 0.025603707879781723, + 0.01836039125919342, + 0.02034682407975197, + 0.03316853195428848, + 0.014054812490940094, + 0.011913318186998367, + 0.020559297874569893, + 0.013645838014781475, + 0.020658351480960846, + 0.017524095252156258 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 6, + 9 + ], + "selected_reward": -0.06203775480389595 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene framed as a medium shot at eye level with a shallow depth of field against a clean white background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a tall, transparent cylindrical glass filled with still water visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a white circular pencil tablet appear at the base of the glass and erupt with a turbulent cloud of tiny bubbles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera begin with a slow push-in that continues into an extreme macro close-up?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the tablet surface shown disintegrating and releasing a milky plume into the water?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.033698562532663345 + ], + "opposite_logprob_tokens": [ + -3.408698558807373 + ], + "logprob_reward": -0.033698562532663345, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.033698562532663345, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.19747017323970795, + 0.04993818327784538, + 0.047517817467451096, + 0.02455938793718815, + 0.029465341940522194, + 0.11586807668209076, + 0.26851925253868103, + 0.08225946128368378, + 0.059334781020879745, + 0.12506753206253052 + ], + "credit_map": [ + 0.028753861784934998, + 0.02708633989095688, + 0.021717043593525887, + 0.022372301667928696, + 0.026106256991624832, + 0.030562760308384895, + 0.015391712076961994, + 0.025479882955551147, + 0.005496167577803135, + 0.008638636209070683, + 0.005541782360523939, + 0.004806291777640581, + 0.0088044423609972, + 0.006831507198512554, + 0.0049212719313800335, + 0.004898084793239832, + 0.004885048605501652, + 0.005445491056889296, + 0.004402091260999441, + 0.005008178297430277, + 0.012330185621976852, + 0.007517371326684952, + 0.003563676029443741, + 0.004365772008895874, + 0.0023939411621540785, + 0.0023227466735988855, + 0.0027329663280397654, + 0.0030726520344614983, + 0.003884840989485383, + 0.005109407473355532, + 0.002218654379248619, + 0.0028241779655218124, + 0.0033341089729219675, + 0.003745097666978836, + 0.003332355059683323, + 0.003045467659831047, + 0.00433733593672514, + 0.00477690901607275, + 0.0034180162474513054, + 0.0034760513808578253, + 0.011043435893952847, + 0.020220156759023666, + 0.012392677366733551, + 0.01190576609224081, + 0.020916009321808815, + 0.01376271527260542, + 0.01344392355531454, + 0.012183398939669132, + 0.023378506302833557, + 0.04676578566431999, + 0.02158341556787491, + 0.037109848111867905, + 0.05300484225153923, + 0.031125715002417564, + 0.02766488678753376, + 0.027886250987648964, + 0.008455367758870125, + 0.00919946189969778, + 0.009041747078299522, + 0.009938212111592293, + 0.014148681424558163, + 0.008913155645132065, + 0.011389385908842087, + 0.011173452250659466, + 0.005725999362766743, + 0.010699900798499584, + 0.006293098907917738, + 0.007354804314672947, + 0.007849866524338722, + 0.005578691139817238, + 0.009021342732012272, + 0.006811081897467375, + 0.014821377582848072, + 0.018740423023700714, + 0.010242260992527008, + 0.009006254374980927, + 0.028100835159420967, + 0.016014020889997482, + 0.014846720732748508, + 0.013295634649693966 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 0, + 9 + ], + "selected_reward": -0.033698562532663345 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000101/group_9/index_tensor([1667, 1667])_group_9_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000101/group_9/index_tensor([1667, 1667])_group_9_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..c4ed65b2514f8462c815cc3304040e4c66833fff --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000101/group_9/index_tensor([1667, 1667])_group_9_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "drone shot, taken from high above, slowly zooms out and pans to the right, revealing a surreal scene: a small boat, skillfully carved from a massive, deep orange pumpkin, steadily and calmly sails alone across a mirror-like emerald lake. Soft morning sunlight penetrates the mist rising from the water, illuminating the pumpkin boat's rough, natural textures and saturated colors, making it stand out against the tranquil surface. The pumpkin, hollowed out, houses a simple mast made of dry branches, supported by a white, rag-like sail that billows slightly in the breeze. The lake is surrounded by a vibrant autumn forest, its leaves shimmering in shades of crimson and gold, their reflections gently swaying in the still water. The pumpkin boat cuts through the tranquility, leaving behind a widening V-shaped trail of ripples. Its hull rises and falls rhythmically with the waves, permeating the scene with a sense of solitary, poetic tranquility, creating a magical atmosphere straight out of a fairy tale.", + "seed": 18790, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000101/group_9/index_1667_seed_18790_dp_18_rank_36_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 9, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.31193060260266064, + "vlm_logprob": -0.31193060260266064, + "vlm_yes_no_margin": 3.65, + "vlm_token_credit": -0.31193060260266064, + "vlm_yes_ratio": 0.8, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a small boat carved from a deep orange pumpkin?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the boat sailing on a mirror-like emerald lake surrounded by an autumn forest with crimson and gold leaves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the pumpkin boat feature a mast made of dry branches and a white rag-like sail?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the boat leave a widening V-shaped trail of ripples behind it while moving steadily?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start high above and slowly zoom out while panning to the right?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005637579597532749 + ], + "opposite_logprob_tokens": [ + -5.8806376457214355 + ], + "logprob_reward": -0.005637579597532749, + "yes_no_margin_reward": 5.875, + "token_credit_reward": -0.005637579597532749, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14842982590198517, + 0.16254688799381256, + 0.07725822925567627, + 0.08424321562051773, + 0.12016269564628601, + 0.16249680519104004, + 0.07025222480297089, + 0.059043534100055695, + 0.06205952167510986, + 0.053507015109062195 + ], + "credit_map": [ + 0.012349015101790428, + 0.018603594973683357, + 0.013956742361187935, + 0.017676223069429398, + 0.027729740366339684, + 0.013825030997395515, + 0.022594066336750984, + 0.02169542945921421, + 0.012726268731057644, + 0.017698902636766434, + 0.017503513023257256, + 0.0145047502592206, + 0.029077403247356415, + 0.025433089584112167, + 0.01970725506544113, + 0.025895709171891212, + 0.004692208021879196, + 0.008214062079787254, + 0.00790575984865427, + 0.01132205780595541, + 0.008489711210131645, + 0.0119016719982028, + 0.011426662094891071, + 0.013306102715432644, + 0.00686551071703434, + 0.010972581803798676, + 0.010603340342640877, + 0.014743391424417496, + 0.009104420430958271, + 0.009775237180292606, + 0.010406709276139736, + 0.011772029101848602, + 0.011621459387242794, + 0.018091939389705658, + 0.01658342033624649, + 0.0168792475014925, + 0.012763630598783493, + 0.012528295628726482, + 0.013787854462862015, + 0.017906855791807175, + 0.011388096027076244, + 0.01825215294957161, + 0.019934553653001785, + 0.026013946160674095, + 0.029691889882087708, + 0.017025217413902283, + 0.02126488648355007, + 0.018926074728369713, + 0.008858462795615196, + 0.009475314058363438, + 0.009123568423092365, + 0.007110175210982561, + 0.010998960584402084, + 0.007625984027981758, + 0.00855847168713808, + 0.008501289412379265, + 0.0058532957918941975, + 0.01210893876850605, + 0.010792816989123821, + 0.0056398529559373856, + 0.004852158948779106, + 0.005991323851048946, + 0.006683940067887306, + 0.007121209055185318, + 0.005825687665492296, + 0.007636586669832468, + 0.00932199228554964, + 0.009307376109063625, + 0.00620795926079154, + 0.00588844483718276, + 0.011291328817605972, + 0.006580146960914135, + 0.01244574598968029, + 0.011265626177191734, + 0.007060988340526819, + 0.004452077206224203, + 0.002649555914103985, + 0.0060019115917384624, + 0.003121700370684266, + 0.006509413477033377 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 5, + 0 + ], + "selected_reward": -0.005637579597532749 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a small boat carved from a deep orange pumpkin?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the boat sailing on a mirror-like emerald lake surrounded by an autumn forest with crimson and gold leaves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the pumpkin boat feature a mast made of dry branches and a white rag-like sail?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the boat leave a widening V-shaped trail of ripples behind it while moving steadily?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start high above and slowly zoom out while panning to the right?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.011079574935138226 + ], + "opposite_logprob_tokens": [ + -4.511079788208008 + ], + "logprob_reward": -0.011079574935138226, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.011079574935138226, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12378055602312088, + 0.3172483742237091, + 0.05699973553419113, + 0.08082351088523865, + 0.07677321881055832, + 0.03510824590921402, + 0.07214318215847015, + 0.0892462283372879, + 0.04910893365740776, + 0.098767951130867 + ], + "credit_map": [ + 0.0116566838696599, + 0.012012817896902561, + 0.01353994756937027, + 0.010087832808494568, + 0.028810087591409683, + 0.014102029614150524, + 0.01712305285036564, + 0.016448115929961205, + 0.019370032474398613, + 0.034793008118867874, + 0.02488802745938301, + 0.025309165939688683, + 0.07853937149047852, + 0.03804667666554451, + 0.04412371292710304, + 0.052178386598825455, + 0.0029362859204411507, + 0.0033947965130209923, + 0.003510939422994852, + 0.0049225217662751675, + 0.003929657395929098, + 0.016457345336675644, + 0.004452116321772337, + 0.017396073788404465, + 0.007736411411315203, + 0.009694676846265793, + 0.01660316064953804, + 0.010109126567840576, + 0.00676700659096241, + 0.008171113207936287, + 0.008574912324547768, + 0.013167105615139008, + 0.008321563713252544, + 0.012206610292196274, + 0.008327137678861618, + 0.008706275373697281, + 0.009813262149691582, + 0.009987207129597664, + 0.008755152113735676, + 0.010656017810106277, + 0.0027839327231049538, + 0.003347489284351468, + 0.0041237748228013515, + 0.004093296360224485, + 0.007715272251516581, + 0.004542556591331959, + 0.003914339002221823, + 0.00458758557215333, + 0.007861421443521976, + 0.01705559901893139, + 0.011429835110902786, + 0.008602478541433811, + 0.005150412209331989, + 0.006846720352768898, + 0.004436133429408073, + 0.010760593228042126, + 0.009838234633207321, + 0.015024971216917038, + 0.020464641973376274, + 0.009760319255292416, + 0.0053795683197677135, + 0.008412863127887249, + 0.007544236723333597, + 0.012821396812796593, + 0.005714122671633959, + 0.006383030209690332, + 0.009876374155282974, + 0.006722413934767246, + 0.002959561301395297, + 0.004774501081556082, + 0.00635736295953393, + 0.006321571301668882, + 0.016906915232539177, + 0.01777013950049877, + 0.0169045552611351, + 0.009804542176425457, + 0.005246033892035484, + 0.010542292147874832, + 0.009459828026592731, + 0.012133652344346046 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 9 + ], + "selected_reward": -0.011079574935138226 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a small boat carved from a deep orange pumpkin?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the boat sailing on a mirror-like emerald lake surrounded by an autumn forest with crimson and gold leaves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the pumpkin boat feature a mast made of dry branches and a white rag-like sail?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the boat leave a widening V-shaped trail of ripples behind it while moving steadily?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start high above and slowly zoom out while panning to the right?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.002817234955728054 + ], + "opposite_logprob_tokens": [ + -5.877817153930664 + ], + "logprob_reward": -0.002817234955728054, + "yes_no_margin_reward": 5.875, + "token_credit_reward": -0.002817234955728054, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1500733494758606, + 0.10851086676120758, + 0.14219197630882263, + 0.0258555356413126, + 0.1587943583726883, + 0.12579010426998138, + 0.07033997029066086, + 0.076393261551857, + 0.07700059562921524, + 0.06505001336336136 + ], + "credit_map": [ + 0.009710915386676788, + 0.016757136210799217, + 0.01519828662276268, + 0.01714864745736122, + 0.03087344951927662, + 0.01805133931338787, + 0.018797585740685463, + 0.023535987362265587, + 0.008704188279807568, + 0.013650555163621902, + 0.009653141722083092, + 0.010921315290033817, + 0.01941106468439102, + 0.014257252216339111, + 0.011339334771037102, + 0.020574023947119713, + 0.0072453683242201805, + 0.006761232390999794, + 0.007794258184731007, + 0.004872260615229607, + 0.006218542344868183, + 0.051515769213438034, + 0.005583310499787331, + 0.05220123007893562, + 0.0026553822681307793, + 0.0031801951117813587, + 0.0038918836507946253, + 0.004056539386510849, + 0.0023834507446736097, + 0.002095981501042843, + 0.00411470839753747, + 0.0034773938823491335, + 0.014512716792523861, + 0.025313235819339752, + 0.020804347470402718, + 0.02157672308385372, + 0.01853916421532631, + 0.01780327782034874, + 0.01981552317738533, + 0.02042936533689499, + 0.009002192877233028, + 0.012409956194460392, + 0.014384064823389053, + 0.020724928006529808, + 0.023731999099254608, + 0.011861834675073624, + 0.019465109333395958, + 0.014210015535354614, + 0.00809941440820694, + 0.009385383687913418, + 0.011256602592766285, + 0.01093490980565548, + 0.006876789964735508, + 0.006904091686010361, + 0.008056797087192535, + 0.008825981989502907, + 0.006995279807597399, + 0.012339736334979534, + 0.013650198467075825, + 0.009865875355899334, + 0.0072862678207457066, + 0.006679530255496502, + 0.010713246650993824, + 0.00886312872171402, + 0.006923198234289885, + 0.009129812009632587, + 0.013781601563096046, + 0.011635638773441315, + 0.007352019194513559, + 0.007925858721137047, + 0.0118600158020854, + 0.008392442017793655, + 0.014383653178811073, + 0.01166555192321539, + 0.007223397493362427, + 0.007080291863530874, + 0.0033349127043038607, + 0.006914002355188131, + 0.006048628594726324, + 0.008399573154747486 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 0, + 2 + ], + "selected_reward": -0.002817234955728054 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a small boat carved from a deep orange pumpkin?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the boat sailing on a mirror-like emerald lake surrounded by an autumn forest with crimson and gold leaves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the pumpkin boat feature a mast made of dry branches and a white rag-like sail?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the boat leave a widening V-shaped trail of ripples behind it while moving steadily?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start high above and slowly zoom out while panning to the right?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03808169439435005 + ], + "opposite_logprob_tokens": [ + -3.288081645965576 + ], + "logprob_reward": -0.03808169439435005, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.03808169439435005, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.29276663064956665, + 0.28954583406448364, + 0.03912343829870224, + 0.05827192962169647, + 0.13422100245952606, + 0.049483537673950195, + 0.04003775864839554, + 0.042264629155397415, + 0.01846885308623314, + 0.035816390067338943 + ], + "credit_map": [ + 0.019016630947589874, + 0.032168980687856674, + 0.026369405910372734, + 0.037308648228645325, + 0.06414223462343216, + 0.03040020912885666, + 0.0413140207529068, + 0.042046528309583664, + 0.01724742352962494, + 0.029521698132157326, + 0.018954014405608177, + 0.02141256257891655, + 0.0668971836566925, + 0.04301353171467781, + 0.03793858364224434, + 0.0545608252286911, + 0.0031894410494714975, + 0.00594694958999753, + 0.0033543575555086136, + 0.007155371829867363, + 0.0036996298003941774, + 0.0046533821150660515, + 0.005933020263910294, + 0.00519128842279315, + 0.006247102748602629, + 0.008332653902471066, + 0.009626706130802631, + 0.005784819833934307, + 0.0088950265198946, + 0.006166043225675821, + 0.005675440188497305, + 0.0075441342778503895, + 0.014595170505344868, + 0.022505756467580795, + 0.01767997071146965, + 0.01731221377849579, + 0.014854087494313717, + 0.015744134783744812, + 0.014344140887260437, + 0.017185520380735397, + 0.004934870172291994, + 0.0055587706156075, + 0.0074359457939863205, + 0.008186081424355507, + 0.005994734819978476, + 0.005474656354635954, + 0.005638104863464832, + 0.006260375492274761, + 0.006174624431878328, + 0.006736502051353455, + 0.006267131771892309, + 0.005061106290668249, + 0.003025867510586977, + 0.0036318902857601643, + 0.00433390773832798, + 0.004806724842637777, + 0.004089808091521263, + 0.006045166868716478, + 0.009960459545254707, + 0.004775953944772482, + 0.0026907860301434994, + 0.0038705547340214252, + 0.004173607099801302, + 0.006658293306827545, + 0.0019176921341568232, + 0.003222189610823989, + 0.003375182393938303, + 0.002343198750168085, + 0.0013809417141601443, + 0.0017910721944645047, + 0.0023943069390952587, + 0.002044268883764744, + 0.005847526248544455, + 0.005361742805689573, + 0.005420508328825235, + 0.0039018464740365744, + 0.0022470722906291485, + 0.004223991651087999, + 0.0034651237074285746, + 0.0053485785610973835 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 4 + ], + "selected_reward": -0.03808169439435005 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a small boat carved from a deep orange pumpkin?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the boat sailing on a mirror-like emerald lake surrounded by an autumn forest with crimson and gold leaves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the pumpkin boat feature a mast made of dry branches and a white rag-like sail?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the boat leave a widening V-shaped trail of ripples behind it while moving steadily?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start high above and slowly zoom out while panning to the right?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -1.5020369291305542 + ], + "opposite_logprob_tokens": [ + -0.2520368993282318 + ], + "logprob_reward": -1.5020369291305542, + "yes_no_margin_reward": -1.25, + "token_credit_reward": -1.5020369291305542, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.25823745131492615, + 0.06975378096103668, + 0.2429836541414261, + 0.05809308588504791, + 0.06017235293984413, + 0.08549492061138153, + 0.0688861832022667, + 0.09269466996192932, + 0.03640744090080261, + 0.027276532724499702 + ], + "credit_map": [ + 0.017009809613227844, + 0.028894901275634766, + 0.02480900287628174, + 0.03075144812464714, + 0.05570346862077713, + 0.027554236352443695, + 0.03541433811187744, + 0.03810020908713341, + 0.005242410581558943, + 0.006572389509528875, + 0.0066251568496227264, + 0.005701734684407711, + 0.011195859871804714, + 0.0104787303134799, + 0.008219040930271149, + 0.015718454495072365, + 0.011901436373591423, + 0.012920711189508438, + 0.011959127150475979, + 0.011159640736877918, + 0.013185152783989906, + 0.08220307528972626, + 0.011721084825694561, + 0.08793340623378754, + 0.005425218027085066, + 0.004790686070919037, + 0.011954933404922485, + 0.00615919241681695, + 0.004682102706283331, + 0.00884817261248827, + 0.006419755518436432, + 0.009813015349209309, + 0.006847372744232416, + 0.009448342956602573, + 0.009409645572304726, + 0.007564994040876627, + 0.006468175444751978, + 0.006657554768025875, + 0.005409697070717812, + 0.008366570807993412, + 0.009980255737900734, + 0.01325770653784275, + 0.012398628517985344, + 0.0144168920814991, + 0.007106670178472996, + 0.008277361281216145, + 0.009094463661313057, + 0.010962934233248234, + 0.009217332117259502, + 0.012574401684105396, + 0.011364980600774288, + 0.007279309444129467, + 0.005895441863685846, + 0.006824328098446131, + 0.007453874684870243, + 0.00827651098370552, + 0.008796020410954952, + 0.024684228003025055, + 0.01362173818051815, + 0.010142345912754536, + 0.005419598892331123, + 0.009732113219797611, + 0.007547466084361076, + 0.012751145288348198, + 0.0032187255565077066, + 0.0048473612405359745, + 0.008334281854331493, + 0.004940378945320845, + 0.0034581669606268406, + 0.0035986071452498436, + 0.0034080164041370153, + 0.004601898603141308, + 0.003986557479947805, + 0.0038531776517629623, + 0.004917501471936703, + 0.0032810773700475693, + 0.0019648910965770483, + 0.002885703230276704, + 0.0030102981254458427, + 0.0033773216418921947 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 7 + ], + "selected_reward": -1.5020369291305542 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "No" + ] + }, + { + "prompt": "drone shot, taken from high above, slowly zooms out and pans to the right, revealing a surreal scene: a small boat, skillfully carved from a massive, deep orange pumpkin, steadily and calmly sails alone across a mirror-like emerald lake. Soft morning sunlight penetrates the mist rising from the water, illuminating the pumpkin boat's rough, natural textures and saturated colors, making it stand out against the tranquil surface. The pumpkin, hollowed out, houses a simple mast made of dry branches, supported by a white, rag-like sail that billows slightly in the breeze. The lake is surrounded by a vibrant autumn forest, its leaves shimmering in shades of crimson and gold, their reflections gently swaying in the still water. The pumpkin boat cuts through the tranquility, leaving behind a widening V-shaped trail of ripples. Its hull rises and falls rhythmically with the waves, permeating the scene with a sense of solitary, poetic tranquility, creating a magical atmosphere straight out of a fairy tale.", + "seed": 18791, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000101/group_9/index_1667_seed_18791_dp_18_rank_36_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 9, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.32137394512537865, + "vlm_logprob": -0.32137394512537865, + "vlm_yes_no_margin": 3.45, + "vlm_token_credit": -0.32137394512537865, + "vlm_yes_ratio": 0.8, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a small boat carved from a deep orange pumpkin?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the boat sailing on a mirror-like emerald lake surrounded by an autumn forest with crimson and gold leaves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the pumpkin boat feature a mast made of dry branches and a white rag-like sail?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the boat leave a widening V-shaped trail of ripples behind it while moving steadily?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start high above and slowly zoom out while panning to the right?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005305733531713486 + ], + "opposite_logprob_tokens": [ + -6.005305767059326 + ], + "logprob_reward": -0.005305733531713486, + "yes_no_margin_reward": 6.0, + "token_credit_reward": -0.005305733531713486, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10656055808067322, + 0.11298957467079163, + 0.2151353359222412, + 0.07092083990573883, + 0.057332638651132584, + 0.0698942020535469, + 0.09685896337032318, + 0.09572593867778778, + 0.11285647004842758, + 0.061725467443466187 + ], + "credit_map": [ + 0.011764207854866982, + 0.011853774078190327, + 0.017025304958224297, + 0.011926813051104546, + 0.014217695221304893, + 0.012301343493163586, + 0.014819719828665257, + 0.012651708908379078, + 0.0148028414696455, + 0.004740273579955101, + 0.017619246616959572, + 0.009983101859688759, + 0.007671490777283907, + 0.02876984141767025, + 0.011216738261282444, + 0.01818605326116085, + 0.029280681163072586, + 0.010905693285167217, + 0.032619692385196686, + 0.011954456567764282, + 0.013646857813000679, + 0.05661638826131821, + 0.0240537878125906, + 0.036057788878679276, + 0.00966249406337738, + 0.004593680612742901, + 0.011545278131961823, + 0.00539134256541729, + 0.004472022410482168, + 0.016707688570022583, + 0.007281576283276081, + 0.01126676332205534, + 0.004115501884371042, + 0.006302408874034882, + 0.00665357057005167, + 0.005893234629184008, + 0.004921513143926859, + 0.014064053073525429, + 0.005516777280718088, + 0.009865584783256054, + 0.00617198133841157, + 0.0076901549473404884, + 0.010183583945035934, + 0.005562476348131895, + 0.01017787866294384, + 0.01227138377726078, + 0.008011729456484318, + 0.009825014509260654, + 0.00960517581552267, + 0.01408159639686346, + 0.01419643871486187, + 0.00908566638827324, + 0.013822647742927074, + 0.012749622575938702, + 0.009047183208167553, + 0.014270637184381485, + 0.013051644898951054, + 0.012430647388100624, + 0.009564786218106747, + 0.009068677201867104, + 0.009247314184904099, + 0.018941503018140793, + 0.008279169909656048, + 0.015142202377319336, + 0.014210795052349567, + 0.008744771592319012, + 0.012809003703296185, + 0.015667743980884552, + 0.010420992970466614, + 0.01890043541789055, + 0.014288472011685371, + 0.017814267426729202, + 0.011539392173290253, + 0.006033024284988642, + 0.007648789323866367, + 0.007700337562710047, + 0.004917467944324017, + 0.006666113156825304, + 0.008299589157104492, + 0.0089207598939538 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 1, + 8 + ], + "selected_reward": -0.005305733531713486 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a small boat carved from a deep orange pumpkin?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the boat sailing on a mirror-like emerald lake surrounded by an autumn forest with crimson and gold leaves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the pumpkin boat feature a mast made of dry branches and a white rag-like sail?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the boat leave a widening V-shaped trail of ripples behind it while moving steadily?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start high above and slowly zoom out while panning to the right?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.05496479198336601 + ], + "opposite_logprob_tokens": [ + -2.929964780807495 + ], + "logprob_reward": -0.05496479198336601, + "yes_no_margin_reward": 2.875, + "token_credit_reward": -0.05496479198336601, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.16816700994968414, + 0.051077552139759064, + 0.29937437176704407, + 0.03790735825896263, + 0.08082844316959381, + 0.09918414056301117, + 0.03704803064465523, + 0.151892751455307, + 0.05696386098861694, + 0.01755654811859131 + ], + "credit_map": [ + 0.018782513216137886, + 0.013177808374166489, + 0.024324260652065277, + 0.01568598672747612, + 0.026307735592126846, + 0.02536308579146862, + 0.024977020919322968, + 0.019548581913113594, + 0.00785766914486885, + 0.0045481412671506405, + 0.007358503062278032, + 0.0032464424148201942, + 0.006220768205821514, + 0.010435555130243301, + 0.003953324630856514, + 0.007457145489752293, + 0.03781859204173088, + 0.013431761413812637, + 0.03658115863800049, + 0.020062927156686783, + 0.017355574294924736, + 0.08600831031799316, + 0.03566750884056091, + 0.05244853347539902, + 0.004236429929733276, + 0.002900073304772377, + 0.006206234451383352, + 0.004783123731613159, + 0.003252273891121149, + 0.006835601292550564, + 0.004445658065378666, + 0.005247961729764938, + 0.008317246101796627, + 0.007472052704542875, + 0.011352716945111752, + 0.009475640021264553, + 0.007780752144753933, + 0.014673277735710144, + 0.010033903643488884, + 0.011722852475941181, + 0.010837716050446033, + 0.012749364599585533, + 0.012169919908046722, + 0.00743479048833251, + 0.008784444071352482, + 0.02333739586174488, + 0.007355728652328253, + 0.016514774411916733, + 0.0031652352772653103, + 0.004296473227441311, + 0.006159021053463221, + 0.003746072528883815, + 0.0039015677757561207, + 0.006324520334601402, + 0.004086023662239313, + 0.0053691137582063675, + 0.019243167713284492, + 0.01554324571043253, + 0.01475792471319437, + 0.021547038108110428, + 0.014796254225075245, + 0.022550418972969055, + 0.02158394828438759, + 0.02187073789536953, + 0.011190280318260193, + 0.004105180036276579, + 0.005089546088129282, + 0.006352710071951151, + 0.003571051172912121, + 0.00993669219315052, + 0.007775024510920048, + 0.00894337147474289, + 0.0027900340501219034, + 0.0025002530310302973, + 0.0024686239194124937, + 0.0017170083010569215, + 0.0019843485206365585, + 0.0017074665520340204, + 0.001734151504933834, + 0.002654661890119314 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 0, + 7 + ], + "selected_reward": -0.05496479198336601 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a small boat carved from a deep orange pumpkin?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the boat sailing on a mirror-like emerald lake surrounded by an autumn forest with crimson and gold leaves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the pumpkin boat feature a mast made of dry branches and a white rag-like sail?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the boat leave a widening V-shaped trail of ripples behind it while moving steadily?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start high above and slowly zoom out while panning to the right?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.001513170194812119 + ], + "opposite_logprob_tokens": [ + -6.5015130043029785 + ], + "logprob_reward": -0.001513170194812119, + "yes_no_margin_reward": 6.5, + "token_credit_reward": -0.001513170194812119, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11522474139928818, + 0.08608303219079971, + 0.1895417869091034, + 0.07009975612163544, + 0.0469278059899807, + 0.02930920571088791, + 0.2763846814632416, + 0.09460669755935669, + 0.05773991346359253, + 0.0340823233127594 + ], + "credit_map": [ + 0.018531160429120064, + 0.009154639206826687, + 0.012290430255234241, + 0.00998541060835123, + 0.013131341896951199, + 0.020483795553445816, + 0.013105865567922592, + 0.018542088568210602, + 0.007785663474351168, + 0.004070755559951067, + 0.015035896562039852, + 0.006369592621922493, + 0.005743301473557949, + 0.023474421352148056, + 0.008030102588236332, + 0.01557330321520567, + 0.024610327556729317, + 0.008774691261351109, + 0.02100248634815216, + 0.01896720565855503, + 0.0108275031670928, + 0.05228841304779053, + 0.021351397037506104, + 0.031719766557216644, + 0.009087251499295235, + 0.0068747252225875854, + 0.008325254544615746, + 0.005722608417272568, + 0.005579828750342131, + 0.01617332175374031, + 0.006126397289335728, + 0.012210370041429996, + 0.0042145391926169395, + 0.004186895210295916, + 0.005960110109299421, + 0.003089745296165347, + 0.00394944753497839, + 0.013470824807882309, + 0.0033934395760297775, + 0.008662804029881954, + 0.002950370777398348, + 0.002172876847907901, + 0.005042788572609425, + 0.002871694043278694, + 0.0027390376199036837, + 0.005067512392997742, + 0.0042716688476502895, + 0.004193257540464401, + 0.02464129403233528, + 0.0393557995557785, + 0.032629936933517456, + 0.02669278159737587, + 0.04106219857931137, + 0.043854713439941406, + 0.028548477217555046, + 0.0395994633436203, + 0.01443041954189539, + 0.009719104506075382, + 0.008255859836935997, + 0.008075110614299774, + 0.007795592304319143, + 0.01983114518225193, + 0.01130114309489727, + 0.015198329463601112, + 0.007299001794308424, + 0.004030395764857531, + 0.009130195714533329, + 0.006649911869317293, + 0.005228545516729355, + 0.008845114149153233, + 0.007704032119363546, + 0.008852717466652393, + 0.006073863245546818, + 0.004192330874502659, + 0.003977015148848295, + 0.0040384456515312195, + 0.002630565781146288, + 0.004209728445857763, + 0.003704204922541976, + 0.00525616854429245 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 2, + 0 + ], + "selected_reward": -0.001513170194812119 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a small boat carved from a deep orange pumpkin?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the boat sailing on a mirror-like emerald lake surrounded by an autumn forest with crimson and gold leaves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the pumpkin boat feature a mast made of dry branches and a white rag-like sail?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the boat leave a widening V-shaped trail of ripples behind it while moving steadily?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start high above and slowly zoom out while panning to the right?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.043041590601205826 + ], + "opposite_logprob_tokens": [ + -3.168041706085205 + ], + "logprob_reward": -0.043041590601205826, + "yes_no_margin_reward": 3.125, + "token_credit_reward": -0.043041590601205826, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06798623502254486, + 0.1542988270521164, + 0.25127270817756653, + 0.12312573939561844, + 0.06232815608382225, + 0.05261116847395897, + 0.046535540372133255, + 0.15949957072734833, + 0.056097421795129776, + 0.026244698092341423 + ], + "credit_map": [ + 0.010120421648025513, + 0.0058002835139632225, + 0.009415095672011375, + 0.005876635201275349, + 0.011745389550924301, + 0.007741371635347605, + 0.009052377194166183, + 0.008234660141170025, + 0.017287418246269226, + 0.005938346032053232, + 0.024128712713718414, + 0.008242269977927208, + 0.009900354780256748, + 0.04947054758667946, + 0.010490681044757366, + 0.028840502724051476, + 0.03740769252181053, + 0.010870425961911678, + 0.028696266934275627, + 0.01850302144885063, + 0.013484288938343525, + 0.06307324767112732, + 0.037173543125391006, + 0.04206418991088867, + 0.020658643916249275, + 0.011275685392320156, + 0.019605007022619247, + 0.014250712469220161, + 0.009499410167336464, + 0.01855943538248539, + 0.010648342780768871, + 0.01862848550081253, + 0.007091611623764038, + 0.004502108786255121, + 0.010280485264956951, + 0.003853044006973505, + 0.0046551842242479324, + 0.01579202525317669, + 0.005946287885308266, + 0.01020740531384945, + 0.0050734481774270535, + 0.005007928237318993, + 0.010525966063141823, + 0.00371701386757195, + 0.004708433989435434, + 0.009559785947203636, + 0.006478792056441307, + 0.007539799436926842, + 0.003824728075414896, + 0.004977384116500616, + 0.00529444869607687, + 0.006509365513920784, + 0.0035912259481847286, + 0.006078135222196579, + 0.008425986394286156, + 0.007834264077246189, + 0.02089550718665123, + 0.013897123746573925, + 0.01625736989080906, + 0.016866274178028107, + 0.017606135457754135, + 0.029073674231767654, + 0.02001378685235977, + 0.024889709427952766, + 0.009617095813155174, + 0.0038464991375803947, + 0.006427662912756205, + 0.005260215140879154, + 0.004625608213245869, + 0.010593743063509464, + 0.006081621162593365, + 0.009644976817071438, + 0.0030345655977725983, + 0.0034138604532927275, + 0.0024893945083022118, + 0.003556532319635153, + 0.002831100020557642, + 0.0040046400390565395, + 0.003206628607586026, + 0.0037079760804772377 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 7, + 1 + ], + "selected_reward": -0.043041590601205826 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a small boat carved from a deep orange pumpkin?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the boat sailing on a mirror-like emerald lake surrounded by an autumn forest with crimson and gold leaves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the pumpkin boat feature a mast made of dry branches and a white rag-like sail?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the boat leave a widening V-shaped trail of ripples behind it while moving steadily?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start high above and slowly zoom out while panning to the right?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -1.502044439315796 + ], + "opposite_logprob_tokens": [ + -0.2520444989204407 + ], + "logprob_reward": -1.502044439315796, + "yes_no_margin_reward": -1.25, + "token_credit_reward": -1.502044439315796, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.15425363183021545, + 0.11005373299121857, + 0.22485288977622986, + 0.06615030765533447, + 0.045923858880996704, + 0.03180364891886711, + 0.06639192253351212, + 0.1336800754070282, + 0.14178436994552612, + 0.025105535984039307 + ], + "credit_map": [ + 0.01994691975414753, + 0.010677220299839973, + 0.023859461769461632, + 0.014176281169056892, + 0.019216861575841904, + 0.02131117694079876, + 0.02223295345902443, + 0.022832751274108887, + 0.01569456420838833, + 0.009439386427402496, + 0.01368087437003851, + 0.011699210852384567, + 0.01793586276471615, + 0.016346978023648262, + 0.011438067071139812, + 0.013818785548210144, + 0.033403925597667694, + 0.011282164603471756, + 0.03180774673819542, + 0.019594557583332062, + 0.013947568833827972, + 0.033671218901872635, + 0.04478602111339569, + 0.03635967895388603, + 0.006326886359602213, + 0.004613855388015509, + 0.008697777055203915, + 0.010079090483486652, + 0.005594195332378149, + 0.007305453531444073, + 0.014030372723937035, + 0.009502680040895939, + 0.004114074166864157, + 0.004736094735562801, + 0.006042662542313337, + 0.005867971573024988, + 0.004879178944975138, + 0.0071121095679700375, + 0.006213643122464418, + 0.006958125624805689, + 0.0037626950070261955, + 0.0024057049304246902, + 0.005944294389337301, + 0.0024365452118217945, + 0.0027660070918500423, + 0.0056155165657401085, + 0.0043803611770272255, + 0.004492524545639753, + 0.005418102256953716, + 0.008551131933927536, + 0.01094010192900896, + 0.005919511429965496, + 0.009060429409146309, + 0.009630792774260044, + 0.0066906544379889965, + 0.010181200690567493, + 0.021099545061588287, + 0.007824997417628765, + 0.014308588579297066, + 0.01196257583796978, + 0.009343230165541172, + 0.03017370030283928, + 0.016642572358250618, + 0.022324873134493828, + 0.025129957124590874, + 0.012217385694384575, + 0.016135655343532562, + 0.014323156327009201, + 0.009750723838806152, + 0.025955310091376305, + 0.015228698961436749, + 0.023043472319841385, + 0.005164428614079952, + 0.0018026934703812003, + 0.0031267039012163877, + 0.003304601414129138, + 0.001902771764434874, + 0.0030239373445510864, + 0.002967594424262643, + 0.0038128059823065996 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 0, + 8 + ], + "selected_reward": -1.502044439315796 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "No" + ] + }, + { + "prompt": "drone shot, taken from high above, slowly zooms out and pans to the right, revealing a surreal scene: a small boat, skillfully carved from a massive, deep orange pumpkin, steadily and calmly sails alone across a mirror-like emerald lake. Soft morning sunlight penetrates the mist rising from the water, illuminating the pumpkin boat's rough, natural textures and saturated colors, making it stand out against the tranquil surface. The pumpkin, hollowed out, houses a simple mast made of dry branches, supported by a white, rag-like sail that billows slightly in the breeze. The lake is surrounded by a vibrant autumn forest, its leaves shimmering in shades of crimson and gold, their reflections gently swaying in the still water. The pumpkin boat cuts through the tranquility, leaving behind a widening V-shaped trail of ripples. Its hull rises and falls rhythmically with the waves, permeating the scene with a sense of solitary, poetic tranquility, creating a magical atmosphere straight out of a fairy tale.", + "seed": 18792, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000101/group_9/index_1667_seed_18792_dp_19_rank_38_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 9, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.3724730284884572, + "vlm_logprob": -0.3724730284884572, + "vlm_yes_no_margin": 2.625, + "vlm_token_credit": -0.3724730284884572, + "vlm_yes_ratio": 0.8, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a small boat carved from a deep orange pumpkin?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the boat sailing on a mirror-like emerald lake surrounded by an autumn forest with crimson and gold leaves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the pumpkin boat feature a mast made of dry branches and a white rag-like sail?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the boat leave a widening V-shaped trail of ripples behind it while moving steadily?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start high above and slowly zoom out while panning to the right?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.012682733125984669 + ], + "opposite_logprob_tokens": [ + -5.012682914733887 + ], + "logprob_reward": -0.012682733125984669, + "yes_no_margin_reward": 5.0, + "token_credit_reward": -0.012682733125984669, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08180491626262665, + 0.032564617693424225, + 0.08834203332662582, + 0.05946586653590202, + 0.20902857184410095, + 0.28889036178588867, + 0.1495894342660904, + 0.026361284777522087, + 0.03070661798119545, + 0.03324636444449425 + ], + "credit_map": [ + 0.006965945474803448, + 0.0045165447518229485, + 0.006390290800482035, + 0.011312293820083141, + 0.013457749970257282, + 0.01426050066947937, + 0.01248238142579794, + 0.012419203296303749, + 0.0032822987996041775, + 0.0033043299335986376, + 0.0016635775100439787, + 0.002754196524620056, + 0.004809683654457331, + 0.004869759548455477, + 0.005550961475819349, + 0.006329808849841356, + 0.005305121187120676, + 0.005276484880596399, + 0.005211907904595137, + 0.01004644576460123, + 0.019309207797050476, + 0.013842941261827946, + 0.014386419206857681, + 0.014963502064347267, + 0.00356684485450387, + 0.0059377518482506275, + 0.0038409822154790163, + 0.005929131526499987, + 0.010489117354154587, + 0.010068741627037525, + 0.010179674252867699, + 0.009453623555600643, + 0.021402493119239807, + 0.03937641531229019, + 0.022134553641080856, + 0.023319873958826065, + 0.03190717473626137, + 0.024152101948857307, + 0.02319665625691414, + 0.02353929728269577, + 0.029768411070108414, + 0.05395479127764702, + 0.021938243880867958, + 0.03797983005642891, + 0.03888639435172081, + 0.03296031430363655, + 0.027472296729683876, + 0.04593006521463394, + 0.011913919821381569, + 0.024590054526925087, + 0.013456366956233978, + 0.01797574572265148, + 0.02218230813741684, + 0.023876653984189034, + 0.01667795330286026, + 0.018916429951786995, + 0.0029639853164553642, + 0.0015233047306537628, + 0.004340789280831814, + 0.0020813364535570145, + 0.0044405097141861916, + 0.0044880458153784275, + 0.003132479963824153, + 0.003390830708667636, + 0.002323088003322482, + 0.0024247211404144764, + 0.0053147803992033005, + 0.0020082457922399044, + 0.004813588224351406, + 0.004535744432359934, + 0.005025158170610666, + 0.0042612915858626366, + 0.0027941984590142965, + 0.0029498895164579153, + 0.003563209669664502, + 0.004938744008541107, + 0.004798737354576588, + 0.005696464795619249, + 0.004557502456009388, + 0.00394761748611927 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 4, + 6 + ], + "selected_reward": -0.012682733125984669 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a small boat carved from a deep orange pumpkin?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the boat sailing on a mirror-like emerald lake surrounded by an autumn forest with crimson and gold leaves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the pumpkin boat feature a mast made of dry branches and a white rag-like sail?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the boat leave a widening V-shaped trail of ripples behind it while moving steadily?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start high above and slowly zoom out while panning to the right?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.08904065936803818 + ], + "opposite_logprob_tokens": [ + -2.464040756225586 + ], + "logprob_reward": -0.08904065936803818, + "yes_no_margin_reward": 2.375, + "token_credit_reward": -0.08904065936803818, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15508249402046204, + 0.04375125840306282, + 0.03783687204122543, + 0.10635276883840561, + 0.1511353999376297, + 0.2000211626291275, + 0.09426230937242508, + 0.14903798699378967, + 0.043610379099845886, + 0.018909290432929993 + ], + "credit_map": [ + 0.008997290395200253, + 0.00629038829356432, + 0.007666073273867369, + 0.018987372517585754, + 0.01516604796051979, + 0.013239676132798195, + 0.04153726249933243, + 0.04319838434457779, + 0.002205443102866411, + 0.0016743220621719956, + 0.002391334157437086, + 0.00426408601924777, + 0.008514790795743465, + 0.010750126093626022, + 0.006089887116104364, + 0.007861276157200336, + 0.002226446755230427, + 0.003362904768437147, + 0.0033070920035243034, + 0.004587006289511919, + 0.006719536148011684, + 0.005143804010003805, + 0.006827223114669323, + 0.005662864074110985, + 0.004864756017923355, + 0.005623739678412676, + 0.005200882907956839, + 0.008505298756062984, + 0.017033301293849945, + 0.013138138689100742, + 0.03156013786792755, + 0.020426524803042412, + 0.012890755198895931, + 0.01889163814485073, + 0.008986568078398705, + 0.011875467374920845, + 0.029157666489481926, + 0.03253394737839699, + 0.014667578041553497, + 0.022131795063614845, + 0.0174716804176569, + 0.03744484484195709, + 0.01410706527531147, + 0.02273641899228096, + 0.02585665136575699, + 0.029496045783162117, + 0.017769502475857735, + 0.03513896465301514, + 0.00771904643625021, + 0.01401106733828783, + 0.008520074188709259, + 0.009436242282390594, + 0.017723020166158676, + 0.017500344663858414, + 0.00883491337299347, + 0.010517610237002373, + 0.00866632629185915, + 0.009208223782479763, + 0.01081333588808775, + 0.009495330043137074, + 0.041332583874464035, + 0.0318797267973423, + 0.01792972907423973, + 0.019712740555405617, + 0.004444635473191738, + 0.0024399172980338335, + 0.008395331911742687, + 0.0029625638853758574, + 0.004298402927815914, + 0.008611474186182022, + 0.006544826086610556, + 0.005913235247135162, + 0.0018653004663065076, + 0.0015019337879493833, + 0.0022122343070805073, + 0.0017846814589574933, + 0.003029915736988187, + 0.003299910807982087, + 0.002785559045150876, + 0.0024297568015754223 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 0, + 4 + ], + "selected_reward": -0.08904065936803818 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a small boat carved from a deep orange pumpkin?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the boat sailing on a mirror-like emerald lake surrounded by an autumn forest with crimson and gold leaves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the pumpkin boat feature a mast made of dry branches and a white rag-like sail?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the boat leave a widening V-shaped trail of ripples behind it while moving steadily?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start high above and slowly zoom out while panning to the right?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01606801524758339 + ], + "opposite_logprob_tokens": [ + -4.141067981719971 + ], + "logprob_reward": -0.01606801524758339, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.01606801524758339, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11168183386325836, + 0.048463065177202225, + 0.06257592141628265, + 0.10692838579416275, + 0.31612828373908997, + 0.18807293474674225, + 0.054973527789115906, + 0.02062838152050972, + 0.014135149307549, + 0.07641252875328064 + ], + "credit_map": [ + 0.006424780935049057, + 0.005559919402003288, + 0.006931512616574764, + 0.013936598785221577, + 0.014990264549851418, + 0.012735571712255478, + 0.02418908290565014, + 0.026914097368717194, + 0.0034436697605997324, + 0.004263192880898714, + 0.0038728404324501753, + 0.00974956713616848, + 0.004851513542234898, + 0.005757653620094061, + 0.006929948925971985, + 0.009594682604074478, + 0.0035912268795073032, + 0.005134175065904856, + 0.004133670125156641, + 0.009144662879407406, + 0.013562319800257683, + 0.008831404149532318, + 0.009308197535574436, + 0.008870271034538746, + 0.005016579758375883, + 0.004602155648171902, + 0.005714504048228264, + 0.00805292371660471, + 0.017711589112877846, + 0.014046060852706432, + 0.031115608289837837, + 0.020668966695666313, + 0.03178461268544197, + 0.060209307819604874, + 0.028414994478225708, + 0.03317048028111458, + 0.050855886191129684, + 0.035680145025253296, + 0.03715509548783302, + 0.038857754319906235, + 0.01928388699889183, + 0.03138606995344162, + 0.01683187112212181, + 0.027666619047522545, + 0.0264524444937706, + 0.024598238989710808, + 0.019419094547629356, + 0.022434711456298828, + 0.004815860651433468, + 0.01013962272554636, + 0.0043000755831599236, + 0.006203014403581619, + 0.008406243287026882, + 0.007903313264250755, + 0.005645242519676685, + 0.007560158148407936, + 0.001452249358408153, + 0.0016881240298971534, + 0.002245932351797819, + 0.0015918751014396548, + 0.003705728566274047, + 0.005125440191477537, + 0.0022000453900545835, + 0.002618987113237381, + 0.0012603739742189646, + 0.0010087062837556005, + 0.0014998242259025574, + 0.0009659812785685062, + 0.0015010142233222723, + 0.003970785532146692, + 0.0017678698059171438, + 0.0021605947986245155, + 0.010168282315135002, + 0.007324436213821173, + 0.008027383126318455, + 0.0066951606422662735, + 0.013511154800653458, + 0.014050456695258617, + 0.00745497178286314, + 0.009180685505270958 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 5, + 0 + ], + "selected_reward": -0.01606801524758339 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a small boat carved from a deep orange pumpkin?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the boat sailing on a mirror-like emerald lake surrounded by an autumn forest with crimson and gold leaves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the pumpkin boat feature a mast made of dry branches and a white rag-like sail?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the boat leave a widening V-shaped trail of ripples behind it while moving steadily?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start high above and slowly zoom out while panning to the right?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.043054722249507904 + ], + "opposite_logprob_tokens": [ + -3.1680548191070557 + ], + "logprob_reward": -0.043054722249507904, + "yes_no_margin_reward": 3.125, + "token_credit_reward": -0.043054722249507904, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13154292106628418, + 0.056839555501937866, + 0.0740571916103363, + 0.09736238420009613, + 0.1702669858932495, + 0.1418783813714981, + 0.03992138430476189, + 0.15468870103359222, + 0.07754027098417282, + 0.05590220168232918 + ], + "credit_map": [ + 0.010674936696887016, + 0.006106466054916382, + 0.009411538951098919, + 0.02166581340134144, + 0.01636977680027485, + 0.016307387501001358, + 0.024552341550588608, + 0.026454653590917587, + 0.004320505540817976, + 0.004554183688014746, + 0.002192434621974826, + 0.003949358593672514, + 0.00967702828347683, + 0.009243215434253216, + 0.011508302763104439, + 0.011394521221518517, + 0.004808263387531042, + 0.004345814697444439, + 0.003285726299509406, + 0.0070524681359529495, + 0.01740751601755619, + 0.014380010776221752, + 0.010424832813441753, + 0.01235256064683199, + 0.00427114637568593, + 0.004072247538715601, + 0.004806639160960913, + 0.006573998369276524, + 0.01877426914870739, + 0.02061353623867035, + 0.021324925124645233, + 0.016925625503063202, + 0.01568993739783764, + 0.03219902887940407, + 0.011090772226452827, + 0.016548583284020424, + 0.02696283534169197, + 0.023129640147089958, + 0.018980935215950012, + 0.025665244087576866, + 0.0120677025988698, + 0.02510102465748787, + 0.013300358317792416, + 0.017817150801420212, + 0.020580390468239784, + 0.020245863124728203, + 0.015309443697333336, + 0.01745644584298134, + 0.0033840041141957045, + 0.0026181605644524097, + 0.004117599688470364, + 0.004142181482166052, + 0.005274726077914238, + 0.010046911425888538, + 0.004515737295150757, + 0.005822064820677042, + 0.01637641154229641, + 0.009360901080071926, + 0.016178257763385773, + 0.01753394678235054, + 0.026974795386195183, + 0.029974184930324554, + 0.019286926835775375, + 0.019003279507160187, + 0.006632449105381966, + 0.005293288268148899, + 0.01637706719338894, + 0.0055923666805028915, + 0.00825094897300005, + 0.011115004308521748, + 0.013890404254198074, + 0.010388746857643127, + 0.0070701418444514275, + 0.005583800841122866, + 0.005958396941423416, + 0.006162021309137344, + 0.008999020792543888, + 0.009809360839426517, + 0.005505244247615337, + 0.0068142167292535305 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 7, + 5 + ], + "selected_reward": -0.043054722249507904 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a small boat carved from a deep orange pumpkin?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the boat sailing on a mirror-like emerald lake surrounded by an autumn forest with crimson and gold leaves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the pumpkin boat feature a mast made of dry branches and a white rag-like sail?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the boat leave a widening V-shaped trail of ripples behind it while moving steadily?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start high above and slowly zoom out while panning to the right?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -1.7015190124511719 + ], + "opposite_logprob_tokens": [ + -0.20151907205581665 + ], + "logprob_reward": -1.7015190124511719, + "yes_no_margin_reward": -1.5, + "token_credit_reward": -1.7015190124511719, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.1697973906993866, + 0.08878875523805618, + 0.07182742655277252, + 0.2118351310491562, + 0.11827049404382706, + 0.06466474384069443, + 0.06658986210823059, + 0.086746945977211, + 0.05506862327456474, + 0.06641065329313278 + ], + "credit_map": [ + 0.014919071458280087, + 0.008628008887171745, + 0.009025202132761478, + 0.014607216231524944, + 0.03427256643772125, + 0.026535416021943092, + 0.0338219553232193, + 0.027987955138087273, + 0.004753637593239546, + 0.005501383449882269, + 0.003840835066512227, + 0.006573282647877932, + 0.01801905781030655, + 0.023033788427710533, + 0.013038881123065948, + 0.01402788981795311, + 0.004322889260947704, + 0.005016191862523556, + 0.006074888631701469, + 0.011084462516009808, + 0.015117607079446316, + 0.008355447091162205, + 0.00997057743370533, + 0.011885368265211582, + 0.01003138069063425, + 0.011968237347900867, + 0.011068902909755707, + 0.012990850955247879, + 0.0352889783680439, + 0.03906075283885002, + 0.05075770616531372, + 0.04066832363605499, + 0.010205266065895557, + 0.021673839539289474, + 0.007585863582789898, + 0.014115134254097939, + 0.015505319461226463, + 0.017188800498843193, + 0.01388104259967804, + 0.01811523362994194, + 0.0058248089626431465, + 0.008530565537512302, + 0.004886538255959749, + 0.009232811629772186, + 0.009798846207559109, + 0.011842823587357998, + 0.006577338557690382, + 0.007971005514264107, + 0.006867087911814451, + 0.009736844338476658, + 0.005254589486867189, + 0.00796368345618248, + 0.008108883164823055, + 0.013034161180257797, + 0.006280943285673857, + 0.009343665093183517, + 0.007846791297197342, + 0.007390874903649092, + 0.0075838277116417885, + 0.008229510858654976, + 0.01767451874911785, + 0.018734030425548553, + 0.009172076359391212, + 0.010115310549736023, + 0.005105963442474604, + 0.002620916347950697, + 0.004376216791570187, + 0.0036976069677621126, + 0.009204496629536152, + 0.01631930284202099, + 0.005641066003590822, + 0.008103054948151112, + 0.00695804925635457, + 0.005923357792198658, + 0.005956085864454508, + 0.012499130330979824, + 0.010761067271232605, + 0.009402059949934483, + 0.007539662066847086, + 0.007371242623776197 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 0, + 4 + ], + "selected_reward": -1.7015190124511719 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "No" + ] + }, + { + "prompt": "drone shot, taken from high above, slowly zooms out and pans to the right, revealing a surreal scene: a small boat, skillfully carved from a massive, deep orange pumpkin, steadily and calmly sails alone across a mirror-like emerald lake. Soft morning sunlight penetrates the mist rising from the water, illuminating the pumpkin boat's rough, natural textures and saturated colors, making it stand out against the tranquil surface. The pumpkin, hollowed out, houses a simple mast made of dry branches, supported by a white, rag-like sail that billows slightly in the breeze. The lake is surrounded by a vibrant autumn forest, its leaves shimmering in shades of crimson and gold, their reflections gently swaying in the still water. The pumpkin boat cuts through the tranquility, leaving behind a widening V-shaped trail of ripples. Its hull rises and falls rhythmically with the waves, permeating the scene with a sense of solitary, poetic tranquility, creating a magical atmosphere straight out of a fairy tale.", + "seed": 18793, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000101/group_9/index_1667_seed_18793_dp_19_rank_38_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 9, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.33608450228348374, + "vlm_logprob": -0.33608450228348374, + "vlm_yes_no_margin": 3.325, + "vlm_token_credit": -0.33608450228348374, + "vlm_yes_ratio": 0.8, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a small boat carved from a deep orange pumpkin?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the boat sailing on a mirror-like emerald lake surrounded by an autumn forest with crimson and gold leaves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the pumpkin boat feature a mast made of dry branches and a white rag-like sail?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the boat leave a widening V-shaped trail of ripples behind it while moving steadily?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start high above and slowly zoom out while panning to the right?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.007830873131752014 + ], + "opposite_logprob_tokens": [ + -5.75783109664917 + ], + "logprob_reward": -0.007830873131752014, + "yes_no_margin_reward": 5.75, + "token_credit_reward": -0.007830873131752014, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2643057703971863, + 0.08959823101758957, + 0.0729629322886467, + 0.11253278702497482, + 0.05249829962849617, + 0.08403802663087845, + 0.08458595722913742, + 0.09327392280101776, + 0.06745094060897827, + 0.07875312119722366 + ], + "credit_map": [ + 0.04663233458995819, + 0.054209668189287186, + 0.035155173391103745, + 0.03649364784359932, + 0.016889717429876328, + 0.02028139866888523, + 0.03032861277461052, + 0.02431519702076912, + 0.011475203558802605, + 0.021291639655828476, + 0.011663154698908329, + 0.012779615819454193, + 0.008821914903819561, + 0.005403666291385889, + 0.009981608018279076, + 0.008181431330740452, + 0.006001131609082222, + 0.008265724405646324, + 0.01109411008656025, + 0.010687596164643764, + 0.00906092394143343, + 0.006721031852066517, + 0.011891377158463001, + 0.009241044521331787, + 0.012354160659015179, + 0.010035838931798935, + 0.017996424809098244, + 0.029018200933933258, + 0.01152948010712862, + 0.008860007859766483, + 0.011717906221747398, + 0.011020763777196407, + 0.008675050921738148, + 0.008616229519248009, + 0.009710567072033882, + 0.0073950886726379395, + 0.005030693020671606, + 0.00474551972001791, + 0.0038388220127671957, + 0.004486327059566975, + 0.010715734213590622, + 0.013641138561069965, + 0.010321984998881817, + 0.012820711359381676, + 0.007414727937430143, + 0.011549558490514755, + 0.007668481674045324, + 0.009905691258609295, + 0.01220758631825447, + 0.011878290213644505, + 0.009495099075138569, + 0.013663777150213718, + 0.008937113918364048, + 0.007047784049063921, + 0.011933153495192528, + 0.009423143230378628, + 0.017341535538434982, + 0.010985719040036201, + 0.01189473271369934, + 0.019757412374019623, + 0.008174157701432705, + 0.01008366048336029, + 0.006901053246110678, + 0.008135644719004631, + 0.009385575540363789, + 0.008978457190096378, + 0.007002844940871, + 0.008260439150035381, + 0.01198289543390274, + 0.008056350983679295, + 0.006468291860073805, + 0.007316084578633308, + 0.007429271470755339, + 0.013510555028915405, + 0.009437788277864456, + 0.017095897346735, + 0.007445525377988815, + 0.009744281880557537, + 0.006185571197420359, + 0.007904229685664177 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 7 + ], + "selected_reward": -0.007830873131752014 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a small boat carved from a deep orange pumpkin?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the boat sailing on a mirror-like emerald lake surrounded by an autumn forest with crimson and gold leaves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the pumpkin boat feature a mast made of dry branches and a white rag-like sail?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the boat leave a widening V-shaped trail of ripples behind it while moving steadily?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start high above and slowly zoom out while panning to the right?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01819128356873989 + ], + "opposite_logprob_tokens": [ + -4.018191337585449 + ], + "logprob_reward": -0.01819128356873989, + "yes_no_margin_reward": 4.0, + "token_credit_reward": -0.01819128356873989, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.19368740916252136, + 0.13685168325901031, + 0.1261511743068695, + 0.07272729277610779, + 0.024465611204504967, + 0.06010642275214195, + 0.1945631355047226, + 0.08602169156074524, + 0.07059178501367569, + 0.03483382239937782 + ], + "credit_map": [ + 0.039558324962854385, + 0.03610921651124954, + 0.02440379001200199, + 0.03233957663178444, + 0.01499650627374649, + 0.013391469605267048, + 0.01482048723846674, + 0.018068037927150726, + 0.023438256233930588, + 0.025614649057388306, + 0.02831505797803402, + 0.028293458744883537, + 0.011809390969574451, + 0.006398939993232489, + 0.006107437890022993, + 0.006874492857605219, + 0.01488495897501707, + 0.021687274798750877, + 0.018905576318502426, + 0.02799542434513569, + 0.012677524238824844, + 0.008552063256502151, + 0.010132047347724438, + 0.011316312476992607, + 0.007725339382886887, + 0.011897430755198002, + 0.010040310211479664, + 0.016723262146115303, + 0.005530145950615406, + 0.007017209660261869, + 0.007135873194783926, + 0.006657727062702179, + 0.00432074349373579, + 0.0036791565362364054, + 0.0026609867345541716, + 0.0036948889028280973, + 0.002475376706570387, + 0.0020692951511591673, + 0.0017220061272382736, + 0.0038431575521826744, + 0.00821793731302023, + 0.007831389084458351, + 0.00855469424277544, + 0.014676998369395733, + 0.003898250637575984, + 0.004120126366615295, + 0.005888462997972965, + 0.006918560713529587, + 0.029531987383961678, + 0.0349048376083374, + 0.02675716206431389, + 0.040720995515584946, + 0.012578283436596394, + 0.012746059335768223, + 0.018515311181545258, + 0.01880849525332451, + 0.009233447723090649, + 0.01276617031544447, + 0.012610499747097492, + 0.025614723563194275, + 0.00345910107716918, + 0.005703862756490707, + 0.006836795713752508, + 0.009797093458473682, + 0.010945108719170094, + 0.008634144440293312, + 0.010834147222340107, + 0.01017198245972395, + 0.00828336551785469, + 0.009479913860559464, + 0.005433633457869291, + 0.006809484213590622, + 0.005847069434821606, + 0.00693053612485528, + 0.0035879251081496477, + 0.004244593903422356, + 0.0039479536935687065, + 0.004704403690993786, + 0.002509593265131116, + 0.003061750903725624 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 0, + 1 + ], + "selected_reward": -0.01819128356873989 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a small boat carved from a deep orange pumpkin?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the boat sailing on a mirror-like emerald lake surrounded by an autumn forest with crimson and gold leaves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the pumpkin boat feature a mast made of dry branches and a white rag-like sail?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the boat leave a widening V-shaped trail of ripples behind it while moving steadily?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start high above and slowly zoom out while panning to the right?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005249644163995981 + ], + "opposite_logprob_tokens": [ + -5.255249500274658 + ], + "logprob_reward": -0.005249644163995981, + "yes_no_margin_reward": 5.25, + "token_credit_reward": -0.005249644163995981, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.20254464447498322, + 0.07607395201921463, + 0.10147205740213394, + 0.08681154251098633, + 0.08659569174051285, + 0.09485684335231781, + 0.12063179910182953, + 0.05872887372970581, + 0.09414694458246231, + 0.07813762128353119 + ], + "credit_map": [ + 0.03029886819422245, + 0.037791818380355835, + 0.027826515957713127, + 0.03302178531885147, + 0.015931468456983566, + 0.017689039930701256, + 0.018219593912363052, + 0.021765561774373055, + 0.009774891659617424, + 0.015589391812682152, + 0.01147022657096386, + 0.012729387730360031, + 0.007246074266731739, + 0.0050182524137198925, + 0.008660392835736275, + 0.005585336592048407, + 0.011243559420108795, + 0.015117308124899864, + 0.01191423274576664, + 0.018725693225860596, + 0.01007495354861021, + 0.00794983096420765, + 0.016034837812185287, + 0.010411640629172325, + 0.008859037421643734, + 0.01036056224256754, + 0.012232134118676186, + 0.0166955403983593, + 0.012744407169520855, + 0.008144229650497437, + 0.010302381590008736, + 0.0074732485227286816, + 0.011708274483680725, + 0.013013659045100212, + 0.014405271969735622, + 0.015809936448931694, + 0.009441127069294453, + 0.007530721370130777, + 0.007933949120342731, + 0.006752748042345047, + 0.009022168815135956, + 0.010975448414683342, + 0.011507165618240833, + 0.020464029163122177, + 0.010007485747337341, + 0.012305952608585358, + 0.00918843224644661, + 0.011386161670088768, + 0.013575448654592037, + 0.014826848171651363, + 0.017558878287672997, + 0.031103244051337242, + 0.00873502716422081, + 0.00865557324141264, + 0.012449427507817745, + 0.013727345503866673, + 0.008367639034986496, + 0.006942675914615393, + 0.006407006178051233, + 0.01216187234967947, + 0.004709410946816206, + 0.006422858219593763, + 0.006791635416448116, + 0.006925774738192558, + 0.010546628385782242, + 0.013691160827875137, + 0.01084249559789896, + 0.008372638374567032, + 0.02098691277205944, + 0.01170383207499981, + 0.011867927387356758, + 0.006135346367955208, + 0.01057597529143095, + 0.019338443875312805, + 0.008626139722764492, + 0.009530985727906227, + 0.0075814928859472275, + 0.009542461484670639, + 0.005915062502026558, + 0.007027055602520704 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 6, + 2 + ], + "selected_reward": -0.005249644163995981 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a small boat carved from a deep orange pumpkin?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the boat sailing on a mirror-like emerald lake surrounded by an autumn forest with crimson and gold leaves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the pumpkin boat feature a mast made of dry branches and a white rag-like sail?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the boat leave a widening V-shaped trail of ripples behind it while moving steadily?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start high above and slowly zoom out while panning to the right?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.048630934208631516 + ], + "opposite_logprob_tokens": [ + -3.048630952835083 + ], + "logprob_reward": -0.048630934208631516, + "yes_no_margin_reward": 3.0, + "token_credit_reward": -0.048630934208631516, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.23050841689109802, + 0.11398576945066452, + 0.07416914403438568, + 0.1020442470908165, + 0.0716019943356514, + 0.07091110944747925, + 0.14947429299354553, + 0.05576889589428902, + 0.0796072855591774, + 0.05192883312702179 + ], + "credit_map": [ + 0.03890541195869446, + 0.03799975663423538, + 0.030066894367337227, + 0.043500080704689026, + 0.019588783383369446, + 0.01787855476140976, + 0.022045986726880074, + 0.020522931590676308, + 0.012469659559428692, + 0.015584850683808327, + 0.020632170140743256, + 0.018728800117969513, + 0.012318403460085392, + 0.006871505174785852, + 0.01702623814344406, + 0.010354137979447842, + 0.012515231035649776, + 0.01645267941057682, + 0.009187351912260056, + 0.012798018753528595, + 0.0056725479662418365, + 0.005636310670524836, + 0.005517254117876291, + 0.00638975016772747, + 0.012016978114843369, + 0.012718473561108112, + 0.016096677631139755, + 0.02885853312909603, + 0.006746799219399691, + 0.007496102247387171, + 0.008426107466220856, + 0.009684580378234386, + 0.013237062841653824, + 0.014161329716444016, + 0.012201792560517788, + 0.009781046770513058, + 0.006750073749572039, + 0.005323456600308418, + 0.0050522624514997005, + 0.005094969179481268, + 0.010251987725496292, + 0.014788703061640263, + 0.0081099234521389, + 0.014925161376595497, + 0.004743106663227081, + 0.005158435087651014, + 0.0075803278014063835, + 0.005353460554033518, + 0.013914519920945168, + 0.014147762209177017, + 0.026444759219884872, + 0.04966070502996445, + 0.006187583785504103, + 0.009745057672262192, + 0.010803480632603168, + 0.018570421263575554, + 0.008918755687773228, + 0.008796833455562592, + 0.00736235361546278, + 0.010393209755420685, + 0.004328843206167221, + 0.0057379333302378654, + 0.005840913392603397, + 0.004390053451061249, + 0.011430586688220501, + 0.011089171282947063, + 0.010247319005429745, + 0.00842664297670126, + 0.01198135782033205, + 0.01174223329871893, + 0.007931740954518318, + 0.006758235860615969, + 0.0070900083519518375, + 0.009131846018135548, + 0.004842331632971764, + 0.007681561168283224, + 0.007409083191305399, + 0.006804746575653553, + 0.0047244904562830925, + 0.004244762938469648 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 6, + 1 + ], + "selected_reward": -0.048630934208631516 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a small boat carved from a deep orange pumpkin?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the boat sailing on a mirror-like emerald lake surrounded by an autumn forest with crimson and gold leaves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the pumpkin boat feature a mast made of dry branches and a white rag-like sail?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the boat leave a widening V-shaped trail of ripples behind it while moving steadily?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start high above and slowly zoom out while panning to the right?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -1.6005197763442993 + ], + "opposite_logprob_tokens": [ + -0.22551976144313812 + ], + "logprob_reward": -1.6005197763442993, + "yes_no_margin_reward": -1.375, + "token_credit_reward": -1.6005197763442993, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.31365785002708435, + 0.12398354709148407, + 0.06371451914310455, + 0.08093748986721039, + 0.05550602078437805, + 0.07738423347473145, + 0.053107816725969315, + 0.06920177489519119, + 0.10870129615068436, + 0.053805503994226456 + ], + "credit_map": [ + 0.05213196203112602, + 0.06065429747104645, + 0.043404076248407364, + 0.05298304930329323, + 0.027419721707701683, + 0.027841517701745033, + 0.02548586204648018, + 0.02373734675347805, + 0.017834771424531937, + 0.02832077629864216, + 0.018426064401865005, + 0.021077513694763184, + 0.008898313157260418, + 0.005881874356418848, + 0.014600293710827827, + 0.008943939581513405, + 0.008800488896667957, + 0.010794701054692268, + 0.009613468311727047, + 0.011995885521173477, + 0.0060440851375460625, + 0.0039053200744092464, + 0.006193174049258232, + 0.0063673933036625385, + 0.009090040810406208, + 0.016874827444553375, + 0.0116355549544096, + 0.01999315619468689, + 0.004518674220889807, + 0.005070584826171398, + 0.007333936635404825, + 0.006420713383704424, + 0.006356918253004551, + 0.00803440622985363, + 0.009363677352666855, + 0.008545893244445324, + 0.004646924324333668, + 0.004988390952348709, + 0.004283103626221418, + 0.009286709129810333, + 0.007998734712600708, + 0.0069604311138391495, + 0.009004182182252407, + 0.012865917757153511, + 0.008579699322581291, + 0.008831310085952282, + 0.012298204004764557, + 0.010845750570297241, + 0.006532111670821905, + 0.006041347980499268, + 0.007937914691865444, + 0.011469447985291481, + 0.005036575254052877, + 0.004406968131661415, + 0.007197139319032431, + 0.00448630889877677, + 0.008907589130103588, + 0.007173300255089998, + 0.008402853272855282, + 0.016545670107007027, + 0.0061180479824543, + 0.007291101384907961, + 0.00657591363415122, + 0.008187301456928253, + 0.015262139961123466, + 0.012457293458282948, + 0.01602979563176632, + 0.017033740878105164, + 0.01577676832675934, + 0.013663604855537415, + 0.010364883579313755, + 0.008113074116408825, + 0.006764991674572229, + 0.007190071977674961, + 0.00550776906311512, + 0.008782045915722847, + 0.0070135341957211494, + 0.007162466645240784, + 0.004166654776781797, + 0.007217966951429844 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 8 + ], + "selected_reward": -1.6005197763442993 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "No" + ] + } +] \ No newline at end of file